text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Best Friends Forever (BFF): Finding Lasting Dense Subgraphs
Konstantinos Semertzidis 1 , Evaggelia Pitoura 1 , Evimaria Terzi 2 , Panayiotis Tsaparas 1
1
Dept. of Computer Science and Engineering, University of Ioannina, Greece
2
Dept. of Computer Science, Boston University, USA
arXiv:1612.05440v3 [cs.SI] 3 Oct 2017
1
{ksemer,pitoura,tsap}@cs.uoi.gr
2
[email protected]
Abstract—Graphs form a natural model for relationships and
interactions between entities, for example, between people in
social and cooperation networks, servers in computer networks,
or tags and words in documents and tweets. But, which of these
relationships or interactions are the most lasting ones? In this
paper, we study the following problem: given a set of graph
snapshots, which may correspond to the state of an evolving
graph at different time instances, identify the set of nodes that
are the most densely connected in all snapshots. We call this
problem the Best Friends For Ever (B FF) problem. We provide
definitions for density over multiple graph snapshots, that capture
different semantics of connectedness over time, and we study the
corresponding variants of the B FF problem. We then look at the
On-Off B FF (O2 B FF) problem that relaxes the requirement of
nodes being connected in all snapshots, and asks for the densest
set of nodes in at least k of a given set of graph snapshots.
We show that this problem is NP-complete for all definitions of
density, and we propose a set of efficient algorithms. Finally, we
present experiments with synthetic and real datasets that show
both the efficiency of our algorithms and the usefulness of the
B FF and the O2 B FF problems.
I. I NTRODUCTION
Graphs offer a natural model for capturing the interactions
and relationships among entities. Oftentimes, multiple snapshots of a graph are available; for example, these snapshots
may correspond to the states of a dynamic graph at different
time instances, or the states of a complex system at different
conditions. We call such sets of graph snapshots, a graph
history. Analysis of the graph history finds a large spectrum of
applications, ranging from social-network marketing, to virus
propagation and digital forensics. A central question in this
context is: which interactions, or relationships in a graph
history are the most lasting ones? In this paper, we formalize
this question and we design algorithms that effectively identify
such relationships.
In particular, given a graph history, we introduce the problem of efficiently finding the set of nodes, that remains the
most tightly connected through history. We call this problem
the Best Friends For Ever (B FF) problem. We formulate the
B FF problem as the problem of locating the set of nodes that
have the maximum aggregate density in the graph history.
We provide different definitions for the aggregate density that
capture different notions of connectedness over time, and result
in four variants of the B FF problem.
We then extend the B FF problem to capture the cases where
subsets of nodes are densely connected for only a subset of
the snapshots. Consider for example, a set of collaborators
that work intensely together for some years and then they
drift apart, or, a set of friends in a social network that stop
interacting for a few snapshots and then, they reconnect with
each other. To identify such subsets of nodes, we define the
On-Off B FF problem, or O2 B FF for short. In the O2 B FF
problem, we ask for a set of nodes and a set of k snapshots
such that the aggregate density of the nodes over these
snapshots is maximized.
Identifying B FF nodes finds many applications. For example, in collaboration and social networks such nodes correspond to well-acquainted individuals, and they can be chosen
to form teams, or organize successful professional or social
events. In a protein-interaction network, we can locate protein
complexes that are densely interacting at different states, thus
indicating a possible underlying regulatory mechanism. In a
network where nodes are words or tags and edges correspond
to their co-occurrences in documents or tweets published in a
specific period of time, identifying B FF nodes may serve as a
first step in topic identification, tag recommendation and other
types of analysis. In a computer network, locating servers that
communicate heavily over time may be useful in identifying
potential attacks, or bottlenecks.
The problem of identifying a dense subgraph in a static (i.e.,
single-snapshot) graph has received a lot of attention (e.g., [1],
[2], [3]). There has been also work on finding dense subgraphs
in dynamic graphs (e.g., [4]). However, in this line of work,
the goal is to efficiently locate the densest subgraph in the
current graph snapshot, whereas we are interested in locating
subgraphs that remain dense in the whole graph history. To
the best of our knowledge, we are the first to systematically
introduce and study density in a graph history, and define the
B FF and O2 B FF problems. The most related work to ours is
[5] where the authors study just one of the four variants of the
B FF problem in the context of graph databases. We compare
the performance of our algorithms for this variant with the
algorithm proposed in [5] experimentally.
We study the complexity of the different variants of the
B FF and O2 B FF problems. Two of the B FF variants can be
solved optimally, while the O2 B FF is NP-hard. We propose
a generic algorithmic framework for solving our problems,
that works in linear time. Experimental results with real and
synthetic datasets show the efficiency and effectiveness of our
algorithms in discovering lasting dense subgraphs. Two case
studies on bibliographic collaboration networks, and hashtag
co-occurrence networks in Twitter validate our approach.
To summarize, the main contributions of this work are the
following:
while the minimum density, dm (G), of the graph is the
minimum degree of any node in V :
• We introduce the novel B FF and O2 B FF problems of
identifying a subset of nodes that define dense subgraphs
in a graph history. To this end, we extend the notion of
density for graph histories, and provide definitions that
capture different semantics of density over time leading
to four variants of our problems.
dm (G) = min degree(u, G).
• We study the complexity of the variants of the B FF and
O2 B FF problems and propose appropriate algorithms. We
prove the optimality, or the approximation factor of our
algorithms whenever possible.
• We extend our definitions and algorithms to identify the
B FFs of an input set of query nodes.
• We perform experiments with both real and synthetic
datasets and demonstrate that our problem definitions are
meaningful, and that our algorithms work well in identifying dense subgraphs in practice.
Roadmap: In Section II, we provide definitions of aggregate
density. We introduce the B FF problem and its algorithms in
Section III, and the O2 B FF problem and its algorithms in
Section IV, while in Section V we study extensions to the
original problem. Our experimental evaluation is presented in
Section VI and comparison with related work in Section VII.
Section VIII concludes the paper.
II. AGGREGATE DENSITY
We assume that we are given as input multiple graph snapshots over the same set of nodes. Snapshots may be ordered,
for example, when the snapshots correspond to the states of a
dynamic graph. We may also have an unordered collection of
graphs, for example, when the snapshots correspond to graphs
collected as a result of some scientific experiments. We refer
to such graph collections as a graph history.
Definition 1 (G RAPH H ISTORY ). A graph history G = {G1 ,
G2 , . . . , Gτ } is a collection of τ graph snapshots, where each
snapshot Gt = (V, Et ), t ∈ [1, τ ], is defined over the same set
of nodes V .
An example of a graph history with four snapshots is shown
in Figure 1. Note that our definition is applicable to graph
snapshots with different set of nodes by considering V as their
union.
We will now define the notion of density for a graph history.
We start by reviewing two basic definitions of graph density
for a single graph snapshot. Given an undirected graph G =
(V, E) and a node u in V , we use degree(u, G) to denote the
degree of u in G.
The average density, da (G), of the graph G is the average
degree of the nodes in V :
da (G) =
2|E|
1 X
degree(u, G) =
|V |
|V |
u∈V
u∈V
Intuitively, for a given graph, dm is defined by a single node,
the one with the minimum degree, while da accounts for the
degrees and thus the connectivity of all nodes. For example,
in Figure (1a), dm (G1 ) = 2, while da (G1 ) = 10/3. Clearly,
dm is a lower bound for da . From now on, when the subscript
of d is ignored, density can be either da or dm .
We also define the density of a subset of nodes S ⊆ V
in the graph G = (V, E). To this end, we use the induced
subgraph G[S] = (S, E(S)) in G, where E(S) = {(u, v) ∈
E : u ∈ S, v ∈ S}. We define the density d(S, G) of S in G
as d(G[S]). For example, again for snapshot G1 in Figure 1,
for Sx = {x1 , x2 , x3 , x4 }, dm (Sx , G1 ) = da (Sx , G1 ) = 3,
while for Sy = {y1 , y2 , y3 , y4 , y5 }, dm (Sy , G1 ) = 2 and
da (Sy , G1 ) = 16/5. Between Sx and Sy , Sx has the highest
minimum density, whereas Sy the highest average density.
We now define the density of a set of nodes S on a graph
history. For this, we need a way to aggregate the density of a
set of nodes over multiple graph snapshots.
Aggregating density sequences: Given a graph history G = {G1 , . . . , Gτ }, we will use d(S, G) =
{d(S, G1 ), . . . , d(S, Gτ )} to denote the sequence of density
values for the graph induced by the set S in the graph snapshots. We consider two definitions for an aggregation function
g(d(S, G)) that aggregates the densities over snapshots: the
first, gm , computes the minimum density over all snapshots:
gm (d(S, G)) = min d(S, Gt ),
Gt ∈G
while the second, ga , computes the average density over all
snapshots:
1 X
d(S, Gt ).
ga (d(S, G)) =
|G|
Gt ∈G
Intuitively, the minimum aggregation function requires high
density in each and every snapshot, while the average aggregation function looks at the snapshots as a whole. Again, we
use g to collectively refer to gm or ga . We can now define the
aggregate density f .
Definition 2 (AGGREGATE D ENSITY ). Given a graph history G = {G1 , . . . , Gτ } defined over a set of nodes V
and S ⊆ V , we define the aggregate density f (S, G) to
be f (S, G) = g(d(S, G)). Depending on the choice of the
density function d and the aggregation function g, we have the
following four versions of f : (a) fmm (S, G) = gm (dm (S, G)),
(b) fma (S, G) = gm (da (S, G)), (c) fam (S, G) = ga (dm (S, G)),
and (d) faa (S, G) = ga (da (S, G)).
Each density definition associates different semantics with
the meaning of density among nodes in a graph history. Large
values of fmm (S, G) correspond to groups of nodes S where
each member of the group is connected with a large number of
other members of the group at each snapshot. A node ceases
y5
y5
y5
y5
x1
x4
y1
y4
x1
x4
y1
y4
x1
x4
y1
y4
x1
x4
y1
y4
x2
x3
y2
y3
x2
x3
y2
y3
x2
x3
y2
y3
x2
x3
y2
y3
(a) G1
(b) G2
(c) G3
(d) G4
Fig. 1: A graph history G = {G1 , . . . , G4 } consisting of four snapshots.
to be considered a member of the group, if it loses touch with
the other members even in a single snapshot.
Large values of fma (S, G) are achieved for groups with
high average density at each snapshot G ∈ G. As opposed to
fmm (S, G), where the requirement is placed at each member
of the group, large values of fma (S, G) are indicative that the
group S has persistently high density as a whole.
The faa (S, G) metric takes large values when the group S
has many connections on average; thus, faa is more “loose”
both in terms of consistency over time and in terms of
requirements at the individual group member level.
Lastly, fam (S, G) takes the average of the minimum degree
node at each snapshot, thus is less sensitive to the density of
S at a single instance.
For example, in the graph history G in Figure 1, all
aggregate densities for Sx are equal to 3. However, for Sy
faa (Sy , G) = 31/10, while fma (Sy , G) = 12/5. That is, while
faa (Sy , G) > faa (Sx , G), fma (Sy , G) < fma (Sx , G) due to the
last instance. Note also that fmm (Sy , G) = 1 and that this value
is determined by just one node in just one snapshot, i.e., node
y4 in the last snapshot, while fam (Sy , G) = 2.
The average graph: Finally, let us define the average graph
of a graph history G which is an edge-weighted graph where
the weight of an edge is equal to the fraction of snapshots in
G where the edge appears.
Definition 3 (AVERAGE G RAPH ). Given a graph history G =
bG =
{G1 , . . . , Gτ } on a set of nodes V , the average graph H
b
(V, E, w)
b is a weighted, undirected graph on the set of nodes
b = V × V , and for each (u, v) ∈ E,
b w(u,
V , where E
b v) =
|Gt =(V,Et )∈G|(u,v)∈Et |
.
|G|
As usual, the degree of a node
u in a weighted graph is
bG ) = P
defined as: degree(u, H
b v). The average
b w(u,
(u,v)∈E
graph performs aggregation on a per-node basis, in that, the
b G is the average degree of u in
degree of each node u in H
time.
With the average graph. we lose information regarding
density at individual snapshots. With some algebraic manipulation, we can prove the following lemma that shows a
connection between the average graph and the faa density
function:
Lemma 1. Let G = {G1 , . . . , Gτ } be a graph history over
a set of nodes V and
S a subset of nodes in V , it holds:
b
faa (S, G) = da HG [S] .
III. T HE
B FF P ROBLEM
In this section, we introduce the B FF problem, we study its
hardness and propose appropriate algorithms.
A. Problem definition
Given the snapshots of a graph history G, our goal is to
locate the Best Friends For Ever (B FF), that is, to identify a
subset of nodes of V such that these nodes remain densely
connected with each other in all snapshots of G. Formally:
Problem 1 (The Best Friends Forever (B FF) Problem). Given
a graph history G and an aggregate density function f , find a
subset of nodes S ⊆ V , such that f (S, G) is maximized.
By considering the four choices for the aggregate density
function f , we have four variants of the B FF problem. Specifically, fmm , fma , fam and faa give rise to problems: B FF - MM,
B FF - MA, B FF - AM and B FF - AA respectively.
B.
B FF algorithms
We now introduce a generic algorithm for the B FF problem.
The algorithm (shown in Algorithm 1) is a “greedy-like”
algorithm inspired by a popular algorithm for the densest
subgraph problem on a static graph [6], [1]. We use G[S]
= {G1 [S], . . . , Gτ [S]} to denote the sequence of the induced
subgraphs of the set of nodes S. The algorithm starts with a set
of nodes S0 consisting of all nodes V , and then it performs
n − 1 steps, where at each step i it produces a set Si by
removing one of the nodes in the set Si−1 . It then finds the
set Si with the maximum aggregate density f (S, G).
Algorithm 1 The F IND B FF algorithm.
Input: Graph history G = {G1 , . . . , Gτ }; aggregate density
function f
Output: A subset of nodes S
1: S0 = V
2: for i = 1, . . . , n − 1 do
3:
vi = arg min score (v, G [Si−1 ])
v∈Si−1
Si = Si−1 \ {vi }
5: return arg max f (Si , G)
4:
i=0...n−1
The F IND B FF algorithm forms the basis for the algorithms we propose for the four variants of the B FF problem. Interestingly, by defining appropriate scoring functions,
score (v, G [S]), (used in line 3 to select which node to remove), we can get efficient algorithms for each of the variants.
Algorithm 2 The scorem algorithm.
Algorithm 3 The scorea algorithm.
Input: Graph history G = {G1 , . . . , Gτ }
Output: Node with the minimum scorem
1: Lt [d] ← list of nodes with degree d in Gt
2:
3: procedure S CORE A ND U PDATE()
4:
for t = 1, . . . , τ do
5:
dmint ← smallest d s.t. Lt [d] 6= ∅
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
scorem =
0
min dmint
t=1,...,τ
t = arg min dmint
t=1,...,τ
u = Lt0 [scorem ].get()
for each Gt ∈ G do
Lt [scorem ].remove(u)
for each (u, v) ∈ Et do
Lt [degree(v, Gt )].remove(v)
Et = Et − (u, v) // update degreev∈V (v, Gt )
Lt [degree(v, Gt )].add(v)
V = V \ {u}
return u
1) Solving B FF - MM : For the B FF - MM problem, we define
the score for a node v in S, scorem , as the minimum degree
of v in the sequence G [S]. That is,
scorem (v, G [S]) = min degree (v, Gt [S]) .
Gt ∈G
At the i-th iteration of the F IND B FF algorithm, we select
the node vi with the minimum scorem value. We call this
instantiation of the F IND B FF algorithm F IND B FF M . Below we
prove that F IND B FF M provides the optimal solution to the
B FF - MM problem.
Proposition 1. The B FF - MM problem can be solved optimally
in polynomial time using the F IND B FF M algorithm.
Proof: Let i be the iteration of the F IND B FF M algorithm,
where for the first time, a node that belongs to an optimal
solution S ∗ is selected to be removed. Let vi be this node.
Then clearly, S ∗ ⊆ Si−1 and by the fact that at every iteration
we remove edges from the graphs we have that
scorem (vi , G [St−1 ]) ≥ scorem (vi , G [S ∗ ]) .
Since vi is the node we pick at iteration i, every node u ∈ Si−1
satisfies:
minGt ∈G degree(u, Gt [Si−1 ]) = scorem (u, G [Si−1 ]) ≥
scorem (vi , G [Si−1 ]) ≥ scorem (vi , G [S ∗ ]).
Since this is true for every node u, this means that Si−1 is
indeed optimal and that our algorithm will find it.
The running time of F IND B FF M is O(nτ + M ), where
n = |V |, τ the number of snapshots in the history graph and
M = m1 +m2 +. . .+mτ the total number of edges that appear
in all snapshots. The node with the minimum scorem value
is computed by the procedure S CORE A ND U PDATE shown in
Algorithm 2, which also removes the node and its edges from
all snapshots. For each snapshot Gt , we keep the list of nodes
with degree d (line 1 in Algorithm 2); these lists can be
constructed in time O(nτ ). Given these lists, the time required
Input: Graph history G = {G1 , . . . Gτ }
Output: Node with the minimum scorea
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
b G ← construct the average graph of G
H
bG
L[d] ← list of nodes with degree d in H
procedure S CORE A ND U PDATE()
scorea ← smallest d s.t. L[d] 6= ∅
u = L[scorea ].get()
L[scorea ].remove(u)
b do
for each (u, v) ∈ E
b G )].remove(v)
L[degree(v, H
b=E
b − (u, v) // update degreev∈V (v, H
bG )
E
b
L[degree(v, HG )].add(v)
V = V \ {u}
return u
to find the node with the minimum scorem is O(τ ) (lines 4–8).
Now in all snapshots, the neighbors of the removed node need
to be moved from their position in the τ lists (lines 9–14); the
degree of every neighbor of the removed node is decreased
by one. Throughout the execution of the algorithm at most
O(M ) such moves can happen. Therefore, the total running
time of F IND B FF M is O(nτ + M ). Note that an algorithm that
iteratively removes from a graph G the node with the minimum
degree was first studied in [6] and shown to compute a 2approximation of the densest subgraph problem for the da (G)
density in [1] and the optimal for the dm (G) density in [7].
2) Solving B FF - AA: To solve the B FF - AA problem, we
b
shall use the average
HG of G. Lemma 1 shows that
graph
b G [S] . Thus, based on the results of
faa (S, G) = da H
Charikar [1] and Goldberg [2], we conclude that:
Proposition 2. The B FF - AA problem can be solved optimally
in polynomial time.
Although there exists a polynomial-time optimal algorithms
for B FF - AA, the computational complexity of these algorithms
b 2 ) for the case of the max-flow algorithm
(e..g., O(|V ||E|
in [2]), makes them hard to use for large-scale real graphs.
Therefore, instead of these algorithm we use the F IND B FF
algorithm, where we define the score of a node v in S, scorea ,
to be equal to its average degree of v in graph history G[S].
That is,
1 X
scorea (v, G [S]) =
degree (v, Gt [S]) .
|G|
Gt ∈G
At the i-th iteration, we select the node vi with the minimum
average degree in G. We will refer to this instantiation of the
F IND B FF, as F IND B FF A . Using Lemma 1 and the results of
Charikar [1] we have the following:
Proposition 3. F IND B FF A is a 21 -approximation algorithm for
the B FF - AA problem.
Proof: It is easy to see that F IND B FF A removes the node
b G . Charikar [1] has shown that
with the minimum density in H
an algorithm that iteratively removes from a graph the node
with minimum density provides a 12 -approximation for finding
the subset of nodes that maximizes the average density on a
single (weighted) graph snapshot. Given the equivalence we
established in Lemma 1, F IND B FF A is also a 12 -approximation
algorithm for B FF - AA.
We show the steps for finding the node with the minimum
scorea value in Algorithm 3 that uses lists of nodes with degree
d in the average graph to achieve an O(nτ + M ) total running
time for F IND B FF A .
3) Solving B FF - MA and B FF - AM: We consider the application of F IND B FF M and F IND B FF A algorithms for the two
problems. In the following propositions, we prove that the two
algorithms give a poor approximation ratio for both problems.
Recall that all our problems are maximization problems, and,
therefore, the lower the approximation ratio, the worse the
performance of the algorithm.
Proposition 4. The approximation ratio of algorithm
F IND B FF M for the B FF - AM problem is O n1 where n is the
number of nodes.
f . We refer to this algorithm as F IND B FF G . F IND B FF G
requires to check all nodes when choosing which node to
remove at each step (shown in Algorithm 4), thus leading to
complexity O(n2 τ + nM ).
Algorithm 4 The scoreg algorithm.
Input: Graph history G = {G1 , . . . Gτ }; an aggregate density
function f
Output: Node with the minimum scoreg
1: procedure S CORE A ND U PDATE()
2:
scoreg [u] = ∅ for u ∈ V
3:
for each u ∈ V do
4:
V 0 = V \ {u}
5:
scoreg [u] = f (V 0 , G)
6:
7:
8:
9:
10:
11:
u = arg min scoreg [v]
v∈V
for each Gt ∈ G do
for each (u, v) ∈ Et do
Et = Et − (u, v)
V = V \ {u}
return u
Proof: In the Appendix.
Proposition 5. The approximation ratio of algorithm
F IND B FF A for the B FF - AM problem is O n1 where n is the
number of nodes.
Proof: In the Appendix.
Proposition 6. The approximation ratio
algorithm
of
F IND B FF M for the B FF - MA problem is O √1n where n is
the number of nodes.
IV. T HE
O2 B FF P ROBLEM
In this section, we relax the requirement that the nodes are
connected in all snapshots of a graph history. Instead, we ask to
find the subset of nodes with the maximum aggregate density
in at least k of the snapshots. We call this problem On-Off
B FF (O2 B FF) problem. We formally define O2 B FF, we show
that it is NP-hard and develop two general types of algorithms
for efficiently solving it in practice.
A. Problem definition
Proof: In the Appendix.
We also consider applying the F IND B FF A algorithm that
selects to remove the node with the minimum average degree.
We can show that F IND B FF A has a poor approximation ratio
for the B FF - AM problem.
In the O2 B FF problem, we seek to find a collection Ck of
k graph snapshots, and a set of nodes S ⊆ V , such that the
subgraphs induced by S in Ck have high aggregate density.
Formally, the O2 B FF problem is defined as follows:
Proposition 7. The approximation ratio
of
algorithm
1
√
F IND B FF A for the B FF - MA problem is O
where n is
n
the number of nodes.
Problem 2 (The On-Off B FF (O2 B FF) Problem). Given a
graph history G = {G1 , G2 , . . . , Gτ }, an aggregate density
function f , and an integer k, find a subset of nodes S ⊆ V , and
a subset Ck of G of size k, such that f (S, Ck ) is maximized.
Proof: In the Appendix.
The complexity of B FF - MA and B FF - AM is an open problem. Jethava and Beerenwinkel [5] conjecture that the B FF - MA
problem is NP-hard, yet they do not provide a proof.
Given that F IND B FF A and F IND B FF M have no theoretical
guarantees, we also investigate a greedy approach, which
selects which node to remove based on the objective function
of the problem at hand. This greedy approach is again an
instance of the iterative algorithm shown in Algorithm 1. More
specifically, for a target function f (either fam or fma ), given a
set Si−1 , we define the score scoreg (v, G[Si ]) of node v ∈ Si
as follows:
As for Problem 1, depending on the choice of the aggregate
density function f , we have four variants of O2 B FF. Thus,
fmm , fma , fam and faa give rise to problems O2 B FF - MM,
O2 B FF - MA, O2 B FF - AM and O2 B FF - AA respectively.
Note that the subcollection of graphs Ck ⊂ G does not
need to consist of contiguous graph snapshots. If this were the
case, then the problem could be solved easily by considering
all possible contiguous subsets of [1, τ ] and outputting the
one with the highest density. However, all the four variants
of the O2 B FF become NP-hard if we drop the constraint for
consecutive graph snapshots.
scoreg (v, G[Si−1 ]) = f (Si−1 , G) − f (Si−1 \ {v}, G) .
Theorem 1. Problem 2 is NP-hard for any definition of the
aggregate density function f .
At iteration i, the algorithm selects the node vi that causes the
smallest decrease, or the largest increase in the target function
We will prove that there exists a clique of size at least k
in graph G if and only if there exists a set of nodes S and
Algorithm 5 The Iterative (ITR) F IND O2 B FF algorithm.
Input: Graph history G = {G1 , . . . Gτ }; an aggregate-density
function f ; integer k
Output: A subset of nodes S and a subset of snapshots Ck ⊆ G.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
converged = False
(Ck0 , S 0 ) = I NITIALIZE (G, f )
ds0 = 0
while not converged do
Ck = B EST S NAPSHOTS(S 0 , f )
S = F IND B FF(Ck , f )
ds = f (S, Ck )
if ds < ds0 then
Converged = True
else ds0 = ds, S 0 = S
return S, Ck
a subset Ck ⊆ G of k snapshots, with f (S, Ck ) ≥ 1. The
forward direction is easy; if there exists a subset of nodes S
in G, with |S| ≥ k, that form a clique, then selecting this set
of nodes S, and a subset Ck of k snapshots that correspond
to nodes in S will wield fmm (S, Ck ) = fam (S, Ck ) = 1. This
follows from the fact that every snapshot is a complete star
where dm (S, Gi ) = 1 for all Gi ∈ Ck . To prove the other
direction, we observe that all our snapshots consist of a star
graph, and a collection of disconnected nodes. Given a set S,
dm (S, Gi ) = 1, if i ∈ S and all nodes in S are connected to the
center node i, and zero otherwise. Therefore, if fmm (S, Ck ) = 1
or fam (S, Ck ) = 1, then this implies that dm (S, Gi ) = 1 for
all Gi ∈ Ck , which means that the k centers of the graph
snapshots in Ck are connected to all nodes in S, and hence
to each other. Therefore, they form a clique of size k in the
graph G.
In the case of faa and fma the construction proceeds as
follows: given the graph G = (V, E), with |E| = m edges,
we construct a graph history G = {G1 , . . . , Gτ } with τ = m
snapshots. All snapshots are defined over the vertex set V .
There is a snapshot Ge for each edge e ∈ E, consisting of the
single edge e. We can prove that there exists a clique of size
at least k in graph G if and only if there exists a set of nodes
S and a subset CK ⊆ G of K = k(k − 1)/2 snapshots, with
f (S, CK ) ≥ 1/k.
We will prove that there exists a clique of size at least
k in graph G if and only if there exists a set of nodes S
and a subset CK ⊆ G of K = k(k − 1)/2 snapshots, with
f (S, CK ) ≥ 1/k. The forward direction is easy. If there exists
a subset of nodes S in G, with |S| = k, that form a clique,
then selecting this set of nodes S, and the k2 snapshots CK
in G that correspond to the edges between the nodes in S will
yield faa (S, CK ) = fma (S, CK ) = 1/k.
To prove the other direction, assume that there is no clique
of size greater or equal to k in G. Let CK be any subset of K =
k(k − 1)/2 snapshots, and let S be the union of the endpoints
of the edges in CK . Since S cannot be a clique, it follows that
|S| = ` > k. Therefore, faa (S, CK ) = fma (S, CK ) = 1/` <
1/k.
B.
O2 B FF algorithms
We consider two general types of algorithms: iterative and
incremental algorithms. The iterative algorithm starts with an
initial size k collection Ck of graph snapshots and improves
it, whereas the incremental algorithm builds the collection incrementally, adding one snapshot at a time. Next, we describe
these two types of algorithms in detail.
Note that depending on whether we are solving the
O2 B FF - MM, O2 B FF - MA, O2 B FF - AM or O2 B FF - AA problem,
we use the appropriate version of the F IND B FF algorithm in
each of these algorithms.
1) Iterative Algorithm: The iterative (ITR) algorithm
(shown in Algorithm 5) starts with an initial collection of
snapshots Ck0 and set of nodes S 0 (routine I NITIALIZE). At
each iteration, given a set S, it finds the k graph snapshots with
the highest d(S, Gi ) score; this is done by B EST S NAPSHOTS.
B EST S NAPSHOTS computes the density d(S, Gi ) of S in each
snapshot Gi ∈ G and outputs the k snapshots Ck with the
highest density. Given Ck , the algorithm then finds the set
S ⊆ V such that f (S, Ck ) is maximized. This step essentially
solves Problem 1 on input Ck for aggregate density function
f using the F IND B FF algorithm. The ITR algorithm keeps
iterating between collections Ck and dense sets of nodes S
until no further iterations can improve the score f (S, Ck ).
An important step of the Iterative F IND O2 B FF is the initialization of Ck0 and S 0 . We consider three different alternatives
for this initialization: random, contiguous, and at least-k.
Random initialization (ITRR ): In this initialization, we randomly pick k snapshots Ck0 from G. These snapshots are then
used for solving the corresponding B FF problem on input Ck0
and produce S 0 = F IND B FF(Ck0 , f ).
Contiguous initialization (ITRC ): In this initialization, we
first find an S 0 that consists of the best k contiguous graph
snapshots. Given G = {G1 , . . . , Gτ }, we go over all the O(τ )
contiguous sets of k snapshots from G, and find the set of k
snapshots Ck0 and corresponding set of nodes S 0 that maximize
f (S 0 , Ck0 ). The intuition behind this initialization technique is
that it assumes that the best k snapshots of G are going to be
contiguous. Our experiments demonstrate that in practice this
is true in many datasets – e.g., in collaboration networks that
evolve over time and we expect to see some temporal locality.
At least-k initialization (ITRK ): In this initialization, we solve
the B FF problem independently in each snapshot Gi ∈ G. This
results in τ different sets Si ⊆ V , one for each solution of B FF
on Gi . S 0 includes the nodes that appear in at least k of the
τ sets Si . The intuition behind this initialization is to include
in the initial solution those nodes that appear to be densely
connected in many snapshots. We also experimented with other
natural alternatives, such as the union: S 0 = ∪i=1...τ Si and the
intersection: S 0 = ∩i=1...τ Si ; the at least-k approach seems
to strike a balance between the two.
The running time of the iterative F IND O2 B FF algorithm is
O (I (nτ + M )), where I is the number of iterations required
until convergence, and the O(nτ +M ) comes from the running
time of F IND B FF. In practice, we observed that the algorithm
converges in at most 6 iterations.
Algorithm 6 The Incremental Density (INCD ) F IND O2 B FF algorithm.
Input: Graph history G = {G1 , . . . Gτ }; aggregate-density
function f ; integer k
Output: A subset of nodes S and a subset of snapshots Ck ⊆ G.
1: Sij = F IND B FF({Gi , Gj }, f ), ∀Gi , Gj ∈ G
2: C2 = arg max f (Sij , {Gi , Gj })
Gi ,Gj ∈G
3: for i = 3 ; i ≤ k do
4:
for each Gt ∈ G \ Ci−1 do
5:
St = F IND B FF(Ci−1 ∪ {Gt }, f )
6:
Gm = arg max f (St , Ci−1 ∪ {Gt })
Gt
Ci = Ci−1 ∪ {Gm }
8: S = F IND B FF(Ck , f )
9: return S, Ck
7:
Algorithm 7 The Incremental Overlap (INCO ) F IND O2 B FF algorithm.
1:
2:
3:
4:
5:
6:
7:
8:
Input: Graph history G = {G1 , . . . Gτ }; aggregate-density
function f ; integer k
Output: A subset of nodes S and a subset of snapshots Ck ⊆ G.
Si = F IND B FF(Gi , f ), ∀Gi ∈ G
|Si ∩ Sj |
C2 = arg max
Gi ,Gj ∈G |Si ∪ Sj |
for i = 3 ; i ≤ k do
SC = F IND B FF(Ci−1 , f )
|St ∩ SC |
Gm = arg max
Gt |St ∪ SC |
Ci = Ci−1 ∪ {Gm }
S = F IND B FF(Ck , f )
return S, Ck
2) Incremental Algorithm: The incremental algorithm starts
with a collection C2 with two snapshots and incrementally adds
snapshots to it until a collection Ck with k snapshots is formed.
Then, the appropriate F IND B FF algorithm is used to compute
the most dense subset of nodes S in Ck .
We use two different policies for selecting snapshots.
The first one, termed incremental density (INCD ) algorithm
(shown in Algorithm 6), selects graph snapshots so as to
maximize density, whereas the second one, termed incremental
overlap (INCO ) algorithm (shown in Algorithm 7), selects
graph snapshots so as to maximize the overlap among nodes
in the dense subsets.
Incremental density (INCD ): To select the pair of snapshots
to form the initial collection C2 , we solve the B FF problem
independently
for each pair of snapshots Gi , Gj ∈ G. This
gives us τ2 dense sets Sij as solutions. We select the pair of
snapshots whose dense subgraph Sij has the largest density
(lines 1–2). The algorithm then builds the solution incrementally in iterations. In iteration i, we construct the solution Ci by
adding to solution Ci−1 the graph snapshot Gm that maximizes
the density function f . That is if St is the densest subset in
the sequence Ci−1 ∪ {Gt }, Gm = arg max f (St , Ci−1 ∪ {Gt })
Gt
(lines 3–6). The running time of the INCD algorithm is
O τ 2 (n + M ) + kτ (knτ + M ) . The first term is due to the
initialization step in line 1, where we look for the best pair
of snapshots. If efficiency is important we can initialize the
algorithm with a random pair to save time.
Incremental overlap (INCO ): To form the initial collection C2 ,
we first solve the B FF problem independently in each snapshot
Gi ∈ G. This gives us τ different sets Si ⊆ V , where Si is the
most dense subgraph in Gi . The algorithm selects from these
τ sets the two most similar ones, Si and Sj , and initializes C2
with the corresponding snapshots Gi and Gj (lines 1–2). For
defining similarity between sets of nodes, we use the Jaccard
similarity. To form Ci from Ci−1 , the algorithm first solves
the B FF problem in Ci−1 . Let SC be the solution. Then, it
selects from the remaining snapshots and adds to Ci−1 the
snapshot Gm whose dense set St is the most similar with
SC (lines 3–6). The running time of the INCO algorithm is
O(τ 2 n + k (nτ + M )), where the first term is the time for the
initialization and the second term for the for-loop.
V.
B FF P ROBLEM E XTENSIONS
The definitions of the B FF and O2 B FF problem focus on
the identification of a set of nodes S such that their aggregate
density is maximized. We now consider natural extensions of
the B FF problem by placing additional constraints on the dense
subgraphs.
Query-node constraint: An interesting extension is introducing a set Q of seed query nodes and requiring that the output
set of nodes S has high density and also contains the input
seed nodes. A similar extension was introduced for static (e.g.,
single snapshots) graphs in [7]. In practice, this variant of B FF
identifies the lasting “best friends” of the query nodes. We call
this the Q R -B FF problem.
We can modify the F IND B FF algorithms appropriately so
that they take into consideration this additional constraint.
In particular, F IND B FF M stops when a query node in Q is
selected to be removed. Let us call this modified algorithm,
Q R -F IND B FF M . We can prove the following proposition. (We
omit the proof due to space constraints.)
Proposition 8. Q R -F IND B FF M solves the Q R -B FF - MM problem optimally in polynomial time.
We also modify F IND B FF A so that it does not remove seed
nodes as follows: If at any step, the node with the minimum
average degree happens to be a seed node, the algorithm
selects to remove the node with the next smallest degree that is
not a seed node. The algorithm stops when the only remaining
nodes are seed nodes. Let us call this modified algorithm,
Q R -F IND B FF A .
Proposition 9. Let S ∗ be an optimal solution for the
Q R -B FF - AA problem and SA be the solution of
the
(S ∗ )+2 ω
,
Q R -F IND B FF A algorithm. It holds: faa (SA ) ≥ s faa2(s+q)
P
where q = |Q|, s = |S ∗ \ Q| and ω = u∈Q degree(u, S ∗ ).
Proof: By Lemma 1, it suffices to show that the
Q R -F IND B FF A algorithm provides an approximation of the
average density of a single graph G. Let S ∗ be the optimal
solution for G. Let G0 be the graph that results from G
when we delete all edges between two query nodes in G.
Clearly, S ∗ is also an optimal solution for G0 . Assume that
we assign each edge (u, v) to either u or v. For each node u,
let a(u) be the number of edges assigned to it and let amax
= maxu {a(u)}. It is easy to see that faa (S ∗ ) ≤ 21 amax ,
since each edge in the optimal solution must be assigned to
a node in it. Now assume that the assignment of edges to
nodes is performed as the Q R -F IND B FF A algorithm proceeds.
Initially, all edges are unassigned. When at step i, a node u
is deleted, we assign to u all the edges that go from Si−1
to u. Note that this assignment maintains the invariant that
at each step, all edges between two nodes in the current set
S are unassigned, while all other edges are assigned. When
the algorithm stops, all edges have been assigned. Consider a
single iteration of the algorithm when a node umin is selected
to be removed and let S be the current set. Let s be the
number of non-query nodes in S, and q = |Q|
Pbe the number
1
of query nodes. It holds: faa (S) = s+q
u∈S degree(u)
P
P
1
1
= s+q
degree(u)+
degree(u).
Let Ω =
u∈Q
s+q
P v∈S\Q
1
degree(u).
Since
u
has
the
smallest
degree
min
u∈Q
s+q
among all nodes in S but the seed nodes, we have faa (S) ≥
1
s+q s a(umin ) + Ω. Since all edges are assigned and edges are
assigned to a node only when this node is removed, at some
step of the execution of the algorithm a(umin ) = amax . Thus,
s
s
1
∗
for some S, faa (S) ≥ s+q
amax +Ω ≥ s+q
2 faa (S )+Ω.
Connectivity constraint: Another meaningful extension is to
impose restrictions on the connectivity of S. The connectivity
of S in a graph history G = {G1 , . . . , Gτ } may have many
different interpretations. One may consider a version where all
the induced subgraphs Gt [S] for t ∈ {1, . . . , τ } are connected.
Another alternative is that at least m > 0 of the τ Gt [S]’s are
connected. Here, we assume that a definition of connectivity
for S is given in the form of a predicate connected(S, G)
which is true if S is connected and false otherwise. Our
problem now becomes: given a graph history G, a set of query
nodes Q ⊂ V and an aggregate density function f , find a
subset of nodes S ⊆ V , such that (1) f (S, G) is maximized,
(2) Q ⊆ S, and (3) connected(S, G) is true.
To solve this problem, we can modify F IND B FF so that
it tests for the connectivity predicate connected(S, G) and
stops when the connectivity constraint no longer holds. In
our experiments, we apply a simplest test, just running the
algorithms on the connected components of the query nodes.
Size constraint: Finally, note that the definition of B FF does
not impose a constraint on the size of the output set of nodes S.
In that respect the problem is parameter-free. If necessary, one
can add an additional constraint to the problem definition by
imposing a cardinality constraint on the output S. However the
cardinality constraint makes the subgraph-discovery problem
computationally hard [7]. This also holds for the B FF problem;
simply consider a graph history with replicas of the same
single-snapshot graph.
VI. E XPERIMENTAL E VALUATION
The goal of our experimental evaluation is threefold. First,
we want to evaluate the performance of our algorithms for
the B FF and the O2 B FF problems in terms of the quality of
the solutions and running time. Second, we want to compare
the different variants of the aggregate density functions. Third,
we want to show the usefulness of the problem, by presenting
results of B FF’s and O2 B FF’s in two real datasets, namely
research collaborators in DBLP and hashtags in Twitter.
TABLE I: Real dataset characteristics
Dataset
DBLP10
Oregon1
Oregon2
Caida
Twitter
AS
# Nodes
2,625
11,492
11,806
31,379
849
7,716
# Edges (aver. per snapshot)
1,143
22,569
31,559
45,833
100
7,783
# Snapshots
10
9
9
122
15
733
Datasets and setting. To evaluate our algorithms, we use a
number of real graph histories, where the snapshots correspond
to collaboration, computer, and concept networks.
• The DBLP10 1 dataset contains yearly snapshots of the coauthorship graph in the 2006-2015 interval, for 11 top
database and data mining conferences. There is an edge
between two authors in a graph snapshot, if they coauthored a paper in the corresponding year and more than
two papers in the corresponding interval.
• The Oregon1 2 dataset consists of nine graph snapshots of
AS peering information inferred from Oregon route-views
between March 31 2001 and May 26 2001 (one snapshot
per week).
• The Oregon2 3 dataset consists of nine weekly snapshots of
AS graphs, between March 31, 2001 and May 26, 2001.
• The Caida4 dataset, contains 122 CAIDA autonomous
systems (AS) graphs, derived from a set of route views
BGP-table instances.
• In the Twitter dataset [8], nodes are hashtags of tweets and
edges represent the co-appearance of hashtags in a tweet.
The dataset contains 15 daily snapshots from October 27,
2013 to November 10, 2013.
• The AS5 dataset represents a communication network of
who-talks-to-whom from the BGP (Border Gateway Protocol) logs. The dataset contains 733 daily snapshots which
span an interval of 785 days from November 8, 1997 to
January 2, 2000.
The dataset characteristics are summarized in Table I.
Since we do not have any ground truth information for the
real datasets, we also use synthetic datasets. In particular, we
create graph snapshots using the forest fire model [9], a wellknown model for creating evolving networks, using the default
forward and backward burning probabilities of 0.35. Then,
we plant dense subgraphs in these snapshots, by randomly
selecting a set X ⊂ V of the nodes and creating additional
edges between them, different at each snapshot.
1 http://dblp.uni-trier.de/
2 https://snap.stanford.edu/data/oregon1.html
3 https://snap.stanford.edu/data/oregon2.html
4 http://www.caida.org/data/as-relationships/
5 https://snap.stanford.edu/data/as.html
A.
B FF evaluation
Since, as shown in Section III-B, F IND B FF M and F IND B FF A
are provably good for the B FF - MM and B FF - AA problems
respectively, we only consider these algorithms for these
problems. For the B FF - MA and B FF - AM problems, we use all
three algorithms, i.e., F IND B FF M , F IND B FF A , and F IND B FF G .
For the B FF - MA problem, we also use the DCS algorithm
proposed in [5] for a problem similar to B FF - MA. The DCS
algorithm is also an iterative algorithm that removes nodes,
one at a time. At each step, DCS finds the subgraphs with
the largest average density for each of the snapshots. Then,
it identifies the subgraph with the smallest average density
among them and removes the node that has the smallest degree
in this subgraph.
Accuracy of F IND B FF and comparison of the density
definitions: We start by an evaluation of the accuracy of our
algorithms along with a comparison of the different aggregate
densities. Since we do not have any ground truth information
for the real data, we use first the synthetic datasets.
First, we create 10 graph snapshots with 4, 000 nodes each
using the forest fire model [9]. Then, in each one of the 10
snapshots we plant a dense random subgraph A with 100 nodes
by inserting extra edges with probability pA . We vary the edge
probabilities from pA = 0.1 to pA = 0.9, and in Fig. 2(a), we
report the F measure achieved for the four density definitions,
when trying to recover subgraph A. Recall that the F takes
values in [0, 1] and the larger the value the better the recall
and precision of the solution with respect to the ground truth
(in this case A). B FF - MM is the most sensitive measure, since
it reports A as a dense subgraph even for the smallest edge
probability. B FF - MA and B FF - AM achieve a perfect F value,
for an edge probability larger than pA = 0.1 and B FF - AA
for an edge probability at least pA = 0.3. For smaller values,
these three density definitions locate supersets of A, due to
averaging. All variations of the F IND B FF algorithms produce
the same results. We now study how the various density
definitions behave when there is a second dense subgraph.
In this case, we plant a subgraph A with edge probability
pA = 0.5 in all snapshots and a second dense subgraph B
with the same number of nodes as A and edge probability
pB = 0.9 in a percentage ` of the snapshots, for different
values of `. Fig. 2(b) depicts which of two graphs, graph A
(shown in blue), or graph B (shown in red), is output by
the F IND B FF algorithms for the different density definitions.
B FF - MM and B FF - MA report A as the densest subgraph, since
these measures ask for high density at each and every snapshot.
However, B FF - AM and B FF - AA report B, when the very dense
subgraph B appears in a sufficient number (more than half) of
the snapshots. All density definitions and algorithms, recover
the exact set A, or B, at each case.
We also run all algorithms using the real datasets and
present the results in Table II, where we report the value of the
BFF-MM
BFF-MA
BFF-AM
BFF-AA
BFF-AA
1
0.8
F-measure
We ran our experiments on a system with a quad-core Intel
Core i7-3820 3.6 GHz processor, with 64 GB memory. We
only used one core in all experiments.
BFF-AM
A
B
0.6
BFF-MA
0.4
0.2
BFF-MM
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
PA
(a) F -measure for the BF F problems
0%
20%
40%
60%
80%
Fraction of planted snapshots for B
100%
(b) Dense graph reported by each of the
BF F problems
Fig. 2: Accuracy and density definition comparison for B FF
objective function and the size of the solution. A first observation is that as expected, the value of the aggregate density of
the reported solution (independently of the problem variant)
increases with the density of the graphs. For B FF - MM problem
we observe that the solutions usually have small cardinality
compared to the solutions for other problems, since the fmm
objective is rather strict (the solution for Twitter was empty).
The solutions for B FF - MM problem in the autonomous-system
datasets appear to have higher fmm scores. This may be due
to the fact that there are larger groups of nodes with lasting
connections in these datasets, e.g., nodes that communicate
intensely between each other during the observation period.
Comparison of F IND B FF alternatives for B FF - MA and
B FF - AM: As shown in Table II, for the B FF - MA problem, F IND B FF G and F IND B FF A perform overall the best
in all datasets producing subgraphs with large fma values.
F IND B FF A performs slightly worse than F IND B FF G only in
the Caida dataset. In the Caida dataset, due probably to the
large number of snapshots, F IND B FF A – which is based on
the average degree – returns a set with the smallest density.
F IND B FF M and DCS have comparable performance, since they
both remove nodes with small degrees in individual snapshots.
They are both outperformed by F IND B FF A and F IND B FF G .
For the B FF - AM problem, F IND B FF A outperforms both
F IND B FF M and F IND B FF G . Our deeper analysis of the inferior
performance of F IND B FF G for this problem revealed that
F IND B FF G often gets trapped in local maxima after removing
just a few nodes of the graph and it cannot find good solutions.
Running time: In Table III, we report execution times. As
expected, the response time of F IND B FF G algorithm is the
slowest in all datasets, due to its quadratic complexity. For the
B FF - MA problem, F IND B FF A is in general faster than DCS.
The difference in times in F IND B FF M algorithms are due to
differences in the computation of the density functions. Additional experiments including ones with synthetic datasets with
larger graphs and more intervals that show similar behavior
are depicted in Figs. 3(a)(b). In particular the Fig. 3(a) show
the execution time of the different algorithms for the B FF - MA
problem for varying nodes with τ = 10, whereas Fig. 3(b)
shows the execution time for varying snapshots.
Summary: In conclusion, our algorithms successfully discovered the planted dense subgraphs even when their density
is small, with B FF - MM being the most sensitive measure.
Minimum aggregation over densities (i.e., B FF - MM, B FF - MA)
requires a dense subgraph to be present at all snapshots,
TABLE II: Results of the different algorithms for the B FF problem on the real datasets.
B FF - MM
F IND B FF M
Size fmm
Datasets
11
33
75
17
15
DBLP10
Oregon1
Oregon2
Caida
Twitter
AS
1.0
14.0
23.0
8.0
0.0
4.0
F IND B FF M
Size fma
B FF - MA
F IND B FF A
F IND B FF G
Size fma
Size fma
DCS
Size fma
3
80
140
33
836
19
8
73
131
29
7
18
14
77
116
57
720
16
1.33
23.7
44.33
13.76
0.04
8.53
1.75
23.86
45.24
12.76
0.29
6.67
61
80
132
60
13
20
1.7
24.05
45.95
15.43
0.62
9.0
F IND B FF M
Size fam
1.29
24.05
44.91
15.05
0.05
8.75
11
33
63
20
12
1.0
14.22
24.44
12.72
0.0
7.44
B FF - AM
F IND B FF A
Size fam
4
35
44
36
3
14
1.7
15.0
23.22
18.11
1.0
9.05
F IND B FF G
Size fam
B FF - AA
F IND B FF A
Size
faa
4
20
461
311
3
14
8
59
147
96
5
38
1.0
2.0
3.22
3.43
1.0
3.14
2.75
25.73
47.89
33.21
1.38
16.38
TABLE III: Execution time (sec) of the different algorithms for the B FF problem on the real datasets.
Datasets
B FF - MM
F IND B FF M
F IND B FF M
DBLP10
Oregon1
Oregon2
Caida
Twitter
AS
0.08
0,27
0.36
2.24
0.37
3.49
0.05
0.24
0.29
2.51
0.57
2.82
F IND B FF M
F IND B FF A
1,000
F IND B FF A
2.04
48
52
2,519
2.81
738
DCS
F IND B FF G
15
Time (sec)
Time (sec)
0.03
0.21
0.47
2.30
0.24
2.16
F IND B FF M
DCS
F IND B FF G
B FF - MA
F IND B FF A F IND B FF G
800
600
400
10
5
200
0
0
8
16
24
32
# nodes in thousands
(a)
40
10
20
30
40
50
# snapshots
(b)
Fig. 3: Synthetic dataset (pA = 0.5): execution time of
the different algorithms for the B FF - MA problem for varying
number of (a) nodes, and (b) snapshots.
whereas average aggregation over densities (i.e., B FF - AM,
B FF - AA) asks that the nodes are sufficiently connected with
each other on average. For the B FF - MA and B FF - AM problems,
F IND B FF A returns in general more dense subgraphs than the
alternatives (including DCS). Both F IND B FF A and F IND B FF M
scale well. They perform similarly for the different density
functions with the differences in running time attributed to
the complexity of calculating the respective functions.
B.
O2 B FF evaluation
In this set of experiments, we evaluate the performance of
the iterative and incremental F IND O2 B FF algorithms.
Comparison of the algorithms in terms of solution quality:
Similar to before, we plant a dense random graph A in
k snapshots. We then run the F IND O2 B FF algorithms with
the same value of k. In Fig. 4, we report the F measure
for the different values of k expressed as a percentage of
the total number of snapshots. For the iterative F IND O2 B FF
algorithm, the at-least-k initialization (ITRK ) outperforms
the other two, and it successfully locates A for all four
density definitions, when A appears in a sufficient number
of snapshots. Non-surprisingly, all initializations work equally
well for average aggregation over time (i.e., O2 B FF - AM and
O2 B FF - AA). For the incremental F IND O2 B FF algorithm, den-
DCS
F IND B FF M
B FF - AM
F IND B FF A
F IND B FF G
B FF - AA
F IND B FF A
0.34
0.83
1.03
11.22
0.47
17.37
0.05
0.48
0.58
6.31
0.85
9.29
0.08
0.57
0.65
5.97
0.28
10.43
1.58
131
57
1,652
2.65
470
0.04
0.28
0.48
2.14
0.52
2.64
sity (INCD ) slightly outperforms overlap (INCO ). Overall,
the incremental algorithms achieve highest F , when compared
with the iterative ones.
We also conduct a second experiment in which we plant
a dense random graph A with edge probability pA = 0.5
in all snapshots and a dense random graph B with edge
probability pB = 0.9 in k snapshots. In Fig. 5, we report the F
measure assuming that B is the correct output for the O2 B FF
problem for different values of k expressed as a percentage
of the total number of snapshots. Again, by comparing the
different initializations for the iterative F IND O2 B FF algorithm,
we observe that among the iterative algorithms, ITRK successfully locates B for all four density definitions, when B
appears in a sufficient number of snapshots. As in the previous
experiment, all initializations work equally well for average
aggregation over time. The incremental algorithms outperform
the iterative ones with INCD being the champion, since they
achieve higher F measure values even when B appears in a
few snapshots.
We also apply the F IND O2 B FF algorithms on all real
datasets for various values of k. In Figs. 6 – 8, we report
the value of the aggregate density for DBLP10 , Oregon1 ,
and Oregon2 for different values of k, again expressed as a
percentage of the total number of snapshots of the input graph
history. Overall, we observed that, in contradistinction to the
experiments with real datasets, the contiguous initialization
(ITRC ) of the iterative O2 B FF - AA algorithm emerges as the
best algorithm in many cases, slightly outperforming INCD .
This is indicative of temporal locality of dense subgraphs
in these datasets, i.e., in these datasets dense subgraphs are
usually alive in a few contiguous snapshots. This is especially
evident in datasets from collaboration networks such as the
DBLP datasets. We also notice that the incremental algorithms
find solutions with density very close to that of the iterative
algorithms. Finally, we also observe that as k increases the
aggregate density of the solutions decrease. This again is
explained by the fact that often dense subgraphs are only
“alive” in a few snapshots.
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
IT RC
20
40
60
80
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
IT RC
20
k (in % of snapshots)
40
60
80
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
IT RC
F-measure for B FF - AA
IT RR
F-measure for B FF - AM
IT RK
F-measure for B FF - MA
F-measure for B FF - MM
IT RC
20
k (in % of snapshots)
40
60
80
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
20
k (in % of snapshots)
40
60
80
k (in % of snapshots)
Fig. 4: Synthetic dataset (pA = 0.9): F -measure values for the O2 B FF problems
IN CD
IN CO
IT RC
1
0.8
0.6
0.4
0.2
0
20
40
60
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
80
IT RC
20
k (in % of snapshots)
40
60
80
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
IT RC
F-measure for B FF - AA
IT RR
F-measure for B FF - AM
IT RK
F-measure for B FF - MA
F-measure for B FF - MM
IT RC
20
k (in % of snapshots)
40
60
80
IT RK
IT RR
IN CD
IN CO
1
0.8
0.6
0.4
0.2
0
20
k (in % of snapshots)
40
60
80
k (in % of snapshots)
Fig. 5: Synthetic dataset (pA = 0.5, pB = 0.9): F -measure values for the O2 B FF problems
IT RK
IT RR
IN CD
IT RC
IN CO
IT RK
IT RR
IN CD
IN CO
IT RC
IT RK
IT RR
IN CD
IN CO
IT RC
5
5
4
4
4
4
3
2
2
1
3
20
40
60
1
80
faa
5
fam
5
fma
fmm
IT RC
3
2
20
40
60
1
80
IT RR
IN CD
IN CO
3
2
20
k (in % of snapshots)
k (in % of snapshots)
IT RK
40
60
1
80
20
k (in % of snapshots)
40
60
80
k (in % of snapshots)
Fig. 6: DBLP10 dataset: scores of aggregate density functions f
IT RC
IT RK
IT RR
IN CD
IT RC
IN CO
IT RK
IT RR
IN CD
IN CO
IT RC
17
IT RK
IT RR
IN CD
IT RC
IN CO
IT RK
IT RR
IN CD
IN CO
17
26
26
25
faa
fam
fma
fmm
16
16
24
15
24
14
15
20
40
60
20
80
k (in % of snapshots)
40
60
80
20
k (in % of snapshots)
40
60
22
80
20
k (in % of snapshots)
40
60
80
k (in % of snapshots)
Fig. 7: Oregon1 dataset: scores of aggregate density functions f
IT RC
IT RK
IT RR
IN CD
IN CO
IT RC
IT RK
IT RR
IN CD
IN CO
IT RC
30
IT RK
IT RR
IN CD
IN CO
IT RC
IT RK
IT RR
IN CD
IN CO
30
50
50
48
fam
26
faa
28
fma
fmm
28
26
46
24
22
46
24
20
40
60
k (in % of snapshots)
80
44
20
40
60
k (in % of snapshots)
80
22
48
20
40
60
80
k (in % of snapshots)
Fig. 8: Oregon2 dataset: scores of aggregate density functions f
44
20
40
60
k (in % of snapshots)
80
IT RC
IT RK
IT RR
IN CD
IN CO
IT RC
IT RK
IT RR
IN CD
IN CO
600
TABLE IV: The B FF solutions for DBLP10 (in parenthesis
dense author subgroups).
Time (sec)
Time (sec)
60
400
200
B FF - MM
(Wei Fan, Philip S. Yu, Jiawei Han, Charu C. Aggarwal), (Lu Qin,
Jeffrey Xu Yu, Xuemin Lin), (Guoliang Li, Jianhua Feng), (Craig
Macdonald, Iadh Ounis)
40
20
B FF - MA
0
(Wei Fan, Jing Gao, Philip S. Yu, Jiawei Han, Charu C. Aggarwal),
(Jeffrey Xu Yu, Xuemin Lin, Ying Zhang)
0
20
40
60
80
20
k (in % of snapshots)
40
60
80
k (in % of snapshots)
(a)
B FF - AM
(b)
(Wei Fan, Jing Gao, Philip S. Yu, Jiawei Han)
Fig. 9: Execution time of the different algorithms for the
O2 B FF - MM problem in (a) Synthetic (pA = 0.5, pB = 0.9),
and (b) Oregon2 datasets.
Convergence and running time: In terms of convergence,
iterative F IND O2 B FF requires 2-6 iterations to converge in all
datasets. In Fig. 9 we report the execution time of O2 B FF
algorithms for the B FF - MM problem in synthetic (pA = 0.5,
pB = 0.9), and (b) Oregon2 datasets. As we observed,
iterative and incremental INCO algorithms scale well with
k. Comparing incremental algorithms, INCO is up to 6x and
3.5x faster than INCD in synthetic and Oregon2 datasets
respectively due to the quadratic complexity of the latter.
Additional experiments including ones with synthetic datasets
with larger graphs and more intervals are depicted in Fig. 10(a)
and Fig. 10(b) respectively. In particular, Fig. 10(a) shows the
execution time of the different algorithms for the O2 B FF - MM
problem for varying number of nodes, with τ = 10 and
k = 6 whereas Fig. 10(b) shows the execution time for varying
number of snapshots with k = 16 τ .
Summary: In conclusion, all algorithms successfully discovered the planted dense subgraphs that lasted a sufficient
percentage (much less than half) of the snapshots with the incremental ones being more sensitive. Among the F IND O2 B FF
algorithms, incremental algorithms outperform the iterative
ones in most cases. Among the incremental algorithms, INCD
is slightly better than INCO . However, given the slow running
time of INCD , INCO is a more preferable choice. Finally, in
datasets consisting of dense subgraphs with temporal locality,
ITRC is a good choice for detecting such graphs.
IT RC
IT RK
IT RR
IN CD
IT RC
IN CO
IT RK
IT RR
IN CD
IN CO
213
213
Time (sec)
Time (sec)
211
29
27
27
24
25
23
210
21
8
16
24
32
# nodes in thousands
(a)
40
10
20
30
40
50
# snapshots
(b)
Fig. 10: Synthetic dataset (pA = 0.5): execution time (log
scale) of the different algorithms for the O2 B FF - MM problem
for varying number of (a) nodes, and (b) snapshots.
B FF - AA
(Wei Fan, Jing Gao, Philip S. Yu, Jiawei Han, Charu C. Aggarwal,
Mohammad M. Masud, Latifur Khan, Bhavani M. Thuraisingham)
C. Case studies
In this section, we report indicative results we obtained
using the DBLP10 and the Twitter datasets. These results
identify lasting dense author collaborations and hashtag cooccurrences respectively.
Lasting dense co-authorships in DBLP10 : In Table IV, we
report the set of nodes output as solutions to the different B FF
problem variants, on the DBLP10 dataset.
First, observe that three authors “Wei Fan”, “Philip S. Yu”,
and “Jiawei Han” are part of all four solutions. These three
authors have co-authored only two papers together in our
dataset, but pairs of them have collaborated very frequently
over the last decade. The solutions for B FF - AM and B FF - AA
contain additional collaborators of these authors. For B FF - AA
we obtain a solution of 8 authors. Although, this group has no
paper in which they are all co-authors, subsets of the authors
have collaborated with each other in many snapshots, resulting
in high value of faa . The solutions for B FF - MM and B FF - MA
contain the aforementioned three authors and some of their
collaborators, but also some new names. These are authors
that have scarce or no collaborations with the former group.
Thus, in this case, the solutions consist of more than one dense
subgroups of authors (grouped in parentheses), that are densely
connected within themselves, but sparsely or not connected
with others, while this is not the case for B FF - AM and B FF - AA.
In Table V, we report results for O2 B FF - MM, O2 B FF - MA,
2
O B FF - AM and O2 B FF - AA on the same dataset. These authors
are the most dense collaborators for k = 2, 4, 6, and 8
(recall there are 10 years in the dataset). We also report the
corresponding years of their dense collaborations. Many new
groups of authors appear. For example, we have new groups of
collaborators from Tsinghua University, CMU and RPI among
others. The authors appeared in the solutions of B FF also
appear here for large values of k.
We also studied experimentally the Q R -B FF problem. In
Table VI, we show indicative results for three of the authors
of this paper as seed nodes. For E. Pitoura, we retrieve a
group of ex-graduate students with whom she had a lasting
and prolific collaboration; for E. Terzi close collaborators from
BU University, and for P. Tsaparas, a group of collaborators
from his time at Microsoft Research. Note that in the last
case, the selected set consists of researchers with whom P.
TABLE V: The authors output as solutions to the O2 B FF problem on DBLP10 .
k=2
Years:
k=4
Years:
k=6
Years:
k=8
Years:
B FF - MM, B FF - MA, B FF - AM, B FF - AA
Christos Faloutsos, Leman Akoglu, Lei Li, Keith Henderson, Hanghang Tong, Tina Eliassi-Rad
2010 - 2011
B FF - MM, B FF - MA, B FF - AM
B FF - AA
Mo Liu, Chetan Gupta, Song Wang, Ismail Ari, Elke
A. Rundensteiner
2010 - 2013
Yong Yu, Dingyi Han, Zhong Su, Lichun Yang, Shengliang Xu,
Shenghua Bao)
B FF - MM, B FF - MA, B FF - AM
2007, 2009 - 2011
B FF - AA
Liyun Ru, Min Zhang, Yiqun Liu, Shaoping Ma
Min Zhang, Liyun Ru Bhavani, Yiqun Liu, Shaoping Ma), Latifur
Khan, M. Thuraisingham, Mohammad M. Masud, Wei Fan, Jing Gao,
Philip S. Yu, Jiawei Han
2007 - 2012
B FF - MM
2007 - 2012
B FF - AM
B FF - MA
B FF - AA
Min Zhang, Yiqun Liu,
Shaoping Ma
Wei Fan, Jing Gao, Philip
S. Yu, Jiawei Han, Charu C.
Aggarwal
Liyun Ru, Min Zhang,
Yiqun Liu, Shaoping Ma
2007 - 2014
2007 - 2008, 2010 - 2015
2007 - 2014
TABLE VI: An example of authors output as solutions to the
Q R -B FF problem on DBLP10 .
Q R -B FF
E. Pitoura: G. Koloniari, M. Drosou, K. Stefanidis
E.Terzi: V. Ishakian, D. Erdos, A. Bestavros
P. Tsaparas: A. Fuxman, A. Kannan, R. Agrawal
C. Faloutsos, D. Koutra: Chris H. Q. Ding, L. Akoglu, H. Huang,
Lei Li, Tao Li, H. Tong
Tsaparas has co-authored several papers in the period recorded
in our dataset, but these authors are also collaborating amongst
themselves. Finally, we use one of the authors appearing in the
dense subgraphs of the O2 B FF, namely C. Faloutsos as seed
node. In this case, we obtain a dense subgraph similar to the
one we have reported in Table V. Finally, we consider a query
with two authors: C. Faloutsos and his student D. Koutra.
Adding D. Koutra to the query set changes the consistency of
the result, focusing more on authors that are collaborators of
both query nodes.
Lasting dense hashtag appearances in Twitter: In Table VII,
we report results of the O2 B FF problem on the Twitter dataset.
Note that the results of the B FF problem on this dataset (as
shown in Table II) are very small graphs, since very few
hashtags appear together in all 15 days of the dataset. As
seen in Table VII, we were able to discover interesting dense
subgraphs of hashtags appearing in k = 3, 6, and 9 of these
days. These hashtags correspond to actual events (including
f1 races and wikileaks) that were trending during that period.
For each solution, we also report the selected snapshot dates.
As expected there is time-contiguity in the selected dates, but
our approach also captures the interest fluctuation over time.
For example, for the wikileaks topic that is captured in the
dense hashtag set {“wikileaks”, “snowden”, “nsa”, “prism”},
the best snapshots are collections of contiguous intervals,
rather than a single contiguous interval.
Note also, that for large values of k, we do not get interesting results which is a fact consistent with the ephemeral nature
of Twitter, where hashtags are short-lived. This is especially
true for fmm and fma that impose strict density constraints and
as a result the solutions consist of disconnected edges.
Latifur Khan, Bhavani M. Thuraisingham, Mohammad M. Masud, Wei Fan,
Jing Gao, Philip S. Yu, Jiawei Han, Charu
C. Aggarwal
2007 - 2012, 2014 - 2015
VII. R ELATED W ORK
To the best of our knowledge, we are the first to systematically study all the variants of the B FF, and O2 B FF problems.
The research most related to ours is the recent work of
Jethava and Beerenwinkel [5] and Rozenshtein et al. [10]. To
the best of our understanding, the authors of [5] introduce
one of the four variants of the B FF problem we studied here,
namely, B FF - MA. In their paper, the authors conjecture that the
problem is NP-hard and they propose a heuristic algorithm.
Our work performs a rigorous and systematic study of the
general B FF problem for multiple variants of the aggregate
density function. Additionally, we introduce and study the
O2 B FF problem, which is not studied in [5]. The authors of
[10] study a problem that can be considered a special case
of the O2 B FF problem. In particular, their goal is to identify
a subset of nodes that are dense in the graph consisting of
the union of edges appearing in the selected snapshots, which
is a weak definition of aggregate density. Furthermore, they
focus on finding collections of contiguous intervals, rather than
arbitrary snapshots. They propose an algorithm similar to the
iterative algorithm we consider, which we have shown to be
outperformed by the incremental algorithms.
There is a huge literature on extracting “dense” subgraphs
from a single graph snapshot. Most formulations for finding
subgraphs that define near-cliques are often NP-hard and often
hard to approximate due to their connection to the maximumclique problem [11], [12], [13], [14], [15]. As a result, the
problem of finding the subgraph with the maximum average
or minimum degree has become particularly popular, due to its
computational tractability. Specifically, the problem of finding
a subgraph with the maximum average degree can be solved
optimally in polynomial time [1], [2], [3], and there exists
a practical greedy algorithm that gives a 2-approximation
guarantee in time linear to the number of edges and nodes
of the input graph [1]. The problem of identifying a subgraph
with the maximum minimum degree, can be solved optimally
in polynomial time [7], using again the greedy algorithm
proposed by Charikar [1]. In our work, we use the average
and minimum degree to quantify the density of the subgraph
TABLE VII: The hashtags and the chosen snapshot dates output as solutions to the O2 B FF problem on Twitter.
k=3
Dates:
k=6
Dates:
k=9
Dates:
k = 12
B FF - MM, B FF - MA
B FF - AM
B FF - AA
kimi, abudhabigp, f1, allowin
ozpol, nz, mexico, malaysia, signapore, vietnam,
chile, peru, tpp, japan, canada
Oct 27-28, Nov 7
abudhabigp, fp1, abudhabi, guti, f1, pushpush, skyf1,
hulk, allowin, bottas, kimi, fp3, fp2
Oct 31-Nov 2
Oct 31-Nov 2
B FF - MM, B FF - MA
B FF - AM
B FF - AA
abudhabigp, f1, skyf1
wikileaks, snowden, nsa, prism
Oct 28-Nov 2
Oct 27-28, Nov 3,5,7
abudhabigp, fp1, abudhabi, guti, f1, pushpush, skyf1,
hulk, allowin, bottas, kimi, fp3, fp2
Oct 28, Oct 30-Nov 1, Nov 9
B FF - MM, B FF - MA
B FF - AM
B FF - AA
(Too many tags to report)
wikileaks, snowden, nsa, prism
Oct 27-31, Nov 3,5-7
assange, wikileaks, snowden, nsa, prism
Oct 27-29,31, Nov 3,5-7,10
B FF - MM, B FF - MA
B FF - AM
B FF - AA
(Too many tags to report)
wikileaks, snowden, nsa
Oct 27-Nov 1, Nov 3-7,10
assange, wikileaks, snowden, nsa, prism
Oct 27-31, Nov 2-7, 10
Dates:
in a single graph snapshot, and we extend these definitions
to sets of snapshots. The algorithmic techniques we use for
the B FF problem are inspired by the techniques proposed by
Charikar [1], and by Sozio and Gionis [7]; however, adapting
them to handle multiple snapshots is non-trivial.
Existing work also studies the problem of identifying a
dense subgraph on time-evolving graphs [4], [16], [17]; these
are graphs where new nodes and edges may appear over time
and existing ones may disappear. The goal in this line of work
is to devise a streaming algorithm that at any point in time it
reports the densest subgraph for the current version of the
graph. In our work, we are not interested in the dynamic
version of the problem and thus the algorithmic challenges
that our problem raises are orthogonal to those faced by the
work on streaming algorithms. Other recent work [18] focuses
on detecting dense subgraphs in a special class of temporal
weighted networks with fixed nodes and edges, where edge
weights change over time and may take both positive and
negative values. This is a different problem, since we consider
graphs with changing edge sets. Furthermore, density in the
presence of edges with negative weights is different than
density when edges have only positive weights.
Finally, another line of research focuses on processing
queries e.g., reachability, path distance, graph matching, etc.
over multiple graph snapshots [19], [20], [21], [22], [23]. The
main goal of this work is to devise effective storage, indexing
and retrieving techniques so that queries over such sequences
of graphs are answered efficiently. In this paper, we propose
a novel problem that of finding dense subgraphs.
VIII. S UMMARY
In this paper, we introduced and systematically studied the
problem of identifying dense subgraphs in a collection of
graph snapshots defining a graph history. We showed that for
many definitions of aggregate density functions the problem
of identifying a subset of nodes that are densely-connected
in all snapshots (i.e., the B FF problem) can be solved in
linear time. We also demonstrated that other versions of the
B FF problem (i.e., B FF - MA and B FF - AM) cannot be solved
with the same algorithm. To identify dense subgraphs that
occur in k, yet not all, the snapshots of a graph history
we also defined the O2 B FF problem. For all variants of this
problem we showed that they are NP-hard and we devised an
iterative and an incremental algorithm for solving them. Our
extensive experimental evaluation with datasets from diverse
domains demonstrated the effectiveness and the efficiency of
our algorithms.
R EFERENCES
[1] M. Charikar, “Greedy approximation algorithms for finding dense components in a graph,” in Approximation Algorithms for Combinatorial
Optimization, 2000, pp. 84–95.
[2] A. V. Goldberg, Finding a maximum density subgraph. University of
California Berkeley, CA, 1984.
[3] S. Khuller and B. Saha, “On finding dense subgraphs,” in ICALP, 2009,
pp. 597–608.
[4] A. Epasto, S. Lattanzi, and M. Sozio, “Efficient densest subgraph
computation in evolving graphs,” in WWW, 2015, pp. 300–310.
[5] V. Jethava and N. Beerenwinkel, “Finding dense subgraphs in relational
graphs,” in ECML PKDD, 2015, pp. 641–654.
[6] Y. Asahiro, K. Iwama, H. Tamaki, and T. Tokuyama, “Greedily finding
a dense subgraph,” in SWAT, 1996, pp. 136–148.
[7] M. Sozio and A. Gionis, “The community-search problem and how to
plan a successful cocktail party,” in ACM SIGKDD, 2010, pp. 939–948.
[8] P. Tsantarliotis and E. Pitoura, “Topic detection using a critical term
graph on news-related tweets,” in Proceedings of the Workshops of the
EDBT/ICDT, 2015, pp. 177–182.
[9] J. Leskovec, J. M. Kleinberg, and C. Faloutsos, “Graph evolution:
Densification and shrinking diameters,” TKDD, vol. 1, no. 1, 2007.
[10] P. Rozenshtein, N. Tatti, and A. Gionis, “Discovering dynamic communities in interaction networks,” in ECML PKDD, 2014, pp. 678–693.
[11] J. I. Alvarez-Hamelin, L. Dall’Asta, A. Barrat, and A. Vespignani,
“Large scale networks fingerprinting and visualization using the k-core
decomposition,” in NIPS, 2005, pp. 41–50.
[12] J. Bourjolly, G. Laporte, and G. Pesant, “An exact algorithm for the
maximum k-club problem in an undirected graph,” European Journal of
Operational Research, vol. 138, no. 1, pp. 21–28, 2002.
[13] K. Makino and T. Uno, “New algorithms for enumerating all maximal
cliques,” in SWAT, 2004, pp. 260–272.
[14] B. McClosky and I. V. Hicks, “Combinatorial algorithms for the maximum k-plex problem,” J. Comb. Optim., vol. 23, no. 1, pp. 29–49, 2012.
[15] C. E. Tsourakakis, F. Bonchi, A. Gionis, F. Gullo, and M. A. Tsiarli,
“Denser than the densest subgraph: extracting optimal quasi-cliques with
quality guarantees,” in ACM SIGKDD, 2013, pp. 104–112.
[16] B. Bahmani, R. Kumar, and S. Vassilvitskii, “Densest subgraph in
streaming and mapreduce,” PVLDB, vol. 5, no. 5, pp. 454–465, 2012.
[17] S. Bhattacharya, M. Henzinger, D. Nanongkai, and C. E. Tsourakakis,
“Space- and time-efficient algorithm for maintaining dense subgraphs
on one-pass dynamic streams,” in STOC, 2015, pp. 173–182.
[18] S. Ma, R. Hu, L. Wang, X. Lin, and J. Huai, “Fast computation of dense
temporal subgraphs,” in ICDE, 2017, pp. 361–372.
[19] K. Semertzidis, E. Pitoura, and K. Lillis, “Timereach: Historical reachability queries on evolving graphs,” in EDBT, 2015, pp. 121–132.
[20] V. Z. Moffitt and J. Stoyanovich, “Towards a distributed infrastructure
for evolving graph analytics,” in WWW, 2016, pp. 843–848.
[21] K. Semertzidis and E. Pitoura, “Durable graph pattern queries on
historical graphs,” in ICDE, 2016, pp. 541–552.
[22] U. Khurana and A. Deshpande, “Efficient snapshot retrieval over historical graph data,” in ICDE, 2013, pp. 997–1008.
[23] C. Ren, E. Lo, B. Kao, X. Zhu, and R. Cheng, “On querying historical
evolving graph sequences,” PVLDB, vol. 4, no. 11, pp. 726–737, 2011.
IX. A PPENDIX
In this section we present counter-examples that demonstrate that the F IND B FF M and F IND B FF A when applied to
the B FF - MA and B FF - AA yield a solution that is a poor
approximation of the optimal solution. For the following, we
use n = |V | to denote the number of nodes in the different
snapshots, and τ = |G| to denote the number of snapshots.
A. Proof of Proposition 4
Proof: In order to prove our claim we need to construct
an instance of the B FF - AM problem where the F IND B FFM
algorithm produces a solution with approximation ratio O n1 .
We construct the graph history G = {G1 , ..., Gτ } as follows.
The first τ − 1 snapshots consist of a full clique with n − 1
nodes, plus an additional node v that is connected to a single
node u from the clique. The last snapshot Gτ consists of just
the edge (v, u).
In the first n − 2 iterations of the F IND B FF M algorithm, the
node with the minimum minimum degree is one of the nodes
in the clique (other than the node u). Thus the nodes in the
clique will be iteratively removed, until we are left with the
edge (u, v). Since node v is present in all intermediate subsets
Si , the minimum degree in all snapshots Gt is 1. Therefore,
the solution S of the F IND B FF M algorithm has fam (S) = 1.
On the other hand clearly the optimal solution S ∗ consists of
the nodes in the clique, where we have minimum degree n−2,
except of the last instance where the minimum degree is zero.
which proves our claim.
Therefore, f (S ∗ ) = (n − 2) τ −1
τ
B. Proof of Proposition 5
Proof: In order to prove our claim we need to construct
an instance of the B FF - AM problem where the F IND B FFA
algorithm produces a solution with approximation ratio O n1 .
We construct the graph history G = {G1 , ..., Gτ }, where τ is
even, as follows. Each snapshot Gt contains n = 2b+3 nodes.
The 2b of these nodes form a complete b × b bipartite graph.
Let u, v, and s denote the additional three nodes. Node s is
connected to all nodes in the graph, in all snapshots, except for
the last snapshot where s is connected only to u and v. Nodes
u and v are connected to each other in all snapshots, and node
u is connected to all 2b nodes of the bipartite graph in the
first τ /2 snapshots, while node v is connected to all 2b nodes
of the bipartite graph in the last τ /2 snapshots. Throughout
assume that τ ≥ 2. Note that the optimal set S ∗ for this
history graph consists of the 2b nodes in the bipartite graph,
with fam (S ∗ , G) = b = Θ(n).
The score scorea for every node w of the 2b nodes in the
bipartite graph is scorea (w, G) = b + 1 + τ −1
τ . For the nodes
u and v, we have scorea (u, G) = scorea (v, G) = 2bτ /2+2τ
=
τ
b + 2. Node s has score scorea (s, G) = 2b τ −1
+
2.
τ
Therefore, in the first iteration, the algorithm will remove
one of the nodes of the bipartite graph. Without loss of
generality assume that it removes one of the nodes in the
left partition. Now, for a node w in the left partition, we
still have that scorea (w, G[S1 ]) = b + 1 + τ −1
τ . For a node
w in the right partition we have that scorea (w, G[S1 ]) =
b + τ −1
τ . For nodes u and v we have scorea (u, G[S1 ]) =
scorea (v, G[S1 ]) = (2b−1)ττ /2+2τ = b + 23 . For node s we
have that scorea (s, G[S1 ]) = (2b − 1) τ −1
τ + 2.
Therefore, in the second iteration the algorithm will select
to remove one of the nodes in the right partition. Note that
the resulting graph G[S2 ] is identical in structure with G, with
n = 2(b − 1) + 3 nodes. Therefore, the same procedure will
be repeated until all the nodes from the bipartite graph are
removed, while nodes u and v will be kept in the set until the
last iterations. As a result, the set S returned by F IND B FF A
has fam (S, G) = 2 (the degree
of the nodes u and v), yielding
approximation ratio O n1 .
C. Proof of Proposition 6
Proof: In order to prove our claim we need to construct an
instance of the B FF - MA problem where the F IND B FF M algo
rithm produces a solution with approximation ratio O √1n .
We construct the graph history G = {G1 , ..., Gτ } as follows.
We have τ = m snapshots that are all identical. They consist of
two sets of nodes A and B of size m and m2 respectively. The
nodes in B form a cycle. The nodes in A in graph snapshot Gt
form a clique with all nodes except for one node vt , different
for each snapshot. The optimal set S ∗ consists of the nodes
in A, that have average degree (m−1)(m−2)
= Θ(m).
m
The F IND B FF M starts with the set of all nodes. The average
2
= Θ(1), which is
degree of any snapshot is 2m +(m−1)(m−2)
m2 +m
also the value of the fma (V ) function. In the first m iterations
of the algorithm, the nodes in A have scorem (v, Si ) = 0, so
these are the ones to be removed first. Then the nodes in B are
removed. In all iterations the average degree in each snapshot
remains O(1). Therefore, the set S returned by the F IND B FFM
1
has fma (S)√= Θ(1), and the approximation ratio is Θ m
.
Since m = n, this proves our claim.
D. Proof of Proposition 7
Proof: In order to prove our claim we need to construct an
instance of the B FF - MA problem where the F IND B FF A algo
rithm produces a solution with approximation ratio O √1n .
The construction of the proof is very similar to before. We
construct the graph history G = {G1 , ..., Gτ } as follows. We
have τ = m snapshots that are all identical, except for the last
snapshot Gm . The snapshots G1 , ..., Gm−1 consist of two sets
of nodes A and B that form two complete cliques of size m
and m2 respectively. In the last snapshot the nodes in B are
all disconnected. The optimal set S ∗ consists of the nodes in
A, that have fma (A) = m(m−1)
= Θ(m − 1).
m
The F IND B FF A starts with the set of all nodes. The value
of fma (V ) is determined by the last snapshot Gm that has
average degree m(m−1)
m2 +m = Θ(1). The nodes in A have average
degree (over time) m(m−1)
= Θ(m), while the nodes in
m
2
−1)
B have average degree (m−1)(m
= Θ(m2 ). Therefore,
m
the algorithm will iteratively remove all nodes in A. In each
iteration the resulting set Si has fma (Si ) = O(1). When all the
nodes in A are removed, we have that fma (Si ) = 0. Therefore,
1
the approximation ratio for this instance is Θ( m
). Our claim
2
follows from the fact that n = m + m.
| 8cs.DS
|
Dimensionality reduction of SDPs through
sketching
Andreas Bluhm∗
arXiv:1707.09863v1 [math.OC] 31 Jul 2017
1
1
and Daniel Stilck França†
1
Department of Mathematics, Technical University of Munich, 85748 Garching, Germany
August 1, 2017
We show how to sketch semidefinite programs (SDPs) using positive maps
in order to reduce their dimension. More precisely, we use Johnson-Lindenstrauss transforms to produce a smaller SDP whose solution preserves
feasibility or approximates the value of the original problem with high probability. These techniques allow to improve both complexity and storage space
requirements. They apply to problems in which the Schatten 1-norm of the
matrices specifying the SDP and of a solution to the problem is constant in
the problem size. Furthermore, we provide some no-go results which clarify
the limitations of positive, linear sketches in this setting. Finally, we discuss
numerical examples to benchmark our methods.
Contents
1. Introduction
2
2. Preliminaries
3
3. Sketching the Hilbert-Schmidt product with positive maps
6
4. Sketching linear matrix inequality feasibility problems
8
5. Approximating the value of semidefinite programs through sketching
10
6. Complexity and memory gains
16
∗
†
[email protected]
[email protected]
1
7. Applications and numerical examples
17
7.1. Estimating the value of a semidefinite packing problem . . . . . . . . . . . 17
7.2. Linear matrix inequality feasibility problems . . . . . . . . . . . . . . . . . 18
8. Conclusion
19
A. Complexifying Johnson-Lindenstrauss transforms
23
B. Proof of Theorem 3.2
24
C. Lower bound on the value of SDPs through sketching
26
D. Random feasibility problems
29
1. Introduction
Semidefinite programs (SDPs) are a prominent class of optimization problems [LA16].
They have applications across different areas of science and mathematics, such as discrete
optimization [WA02] or control theory [BEFB94].
However, although there are many different algorithms that solve an SDP up to an
error ǫ in a time that scales polynomially with the dimension and logarithmically with
ǫ−1 [Bub15], solving large instances of SDPs still remains a challenge. This is not only
due to the fact that the number and cost of the iterations scale superquadratically with
the dimension for most algorithms to solve SDPs, but also due to the fact that the memory required to solve large instances is beyond current capabilities. This has therefore
motivated research on algorithms that can solve SDPs, or at least obtain an approximate
solution, with less memory requirements. One such example is the recent [YUAC17],
where ideas similar to ours were applied to achieve optimal storage requirements necessary to solve a certain class of SDPs. While their work proposes a new way to solve an
SDP using linear sketches, our approach relies on standard convex optimization methods.
In this work, we develop algorithms to estimate the value of an SDP with linear
inequality constraints and to determine if a given linear matrix inequality (LMI) is
feasible or not. These algorithms convert the original problem to one of the same type,
but of smaller dimension, which we call the sketched problem. Subsequently, this new
problem can be solved with the same techniques as the original one, but potentially
using less memory and achieving a smaller runtime. Therefore, we call this a black box
algorithm. With high probability an optimal solution to the sketched problem allows us
to infer something about the original problem.
In the case of LMIs, if the sketched problem is infeasibile, we obtain a certificate
that the original problem is also infeasibile. If the sketched problem is feasible, we are
able to infer that the original problem is either “close to feasible” or feasible with high
probability, under some technical assumptions.
In the case of estimating the value of SDPs, we are able to give an upper bound that
holds with high probability and a lower bound on the value of the SDP from the value
2
of the sketched problem, again under some technical assumptions. For a certain class of
SDPs, which includes the so-called semidefinite packing problems [IPS05], we are able
to find a feasible point of the original problem which is close to the optimal point and
most technical aspects simplify significantly. For this class it can be checked whether
this feasible point is indeed optimal.
Our algorithms work by conjugating the matrices that define the constraints of the
SDP with Johnson-Lindenstrauss transforms [Woo14], thereby preserving the structure
of the problem. Similar ideas have been proposed to reduce the memory usage and
complexity of solving linear programs [VPL15]. While those techniques aim to reduce
the number of constraints, our goal is to reduce the dimension of the matrices involved.
Unfortunately, the dimension of the sketch needed to have a fixed error with high
probability scales with the Schatten 1-norm of the constraints and that of an optimal
solution to the SDP, which significantly restricts the class of problems to which these
methods can be applied. We are able to show that one cannot significantly improve this
scaling and that one cannot sketch general SDPs using linear maps.
This paper is organized as follows: in Section 2, we fix our notation and recall some
basic notions from matrix analysis, Johnson-Lindenstrauss transforms, semidefinite programs and convex analysis which we will need throughout the paper. We then proceed to
show how to sketch the Hilbert-Schmidt scalar product with positive maps in Section 3.
We apply these techniques in Section 4 to show how to certify that certain LMIs are
infeasible by showing the infeasibility of an LMI of smaller dimension. In Section 5, we
apply similar ideas to estimate the value of an SDP with linear inequality constraints
by solving an SDP of lower dimension. This is followed by a discussion of the possible
gains in the complexity of solving these problems and for the memory requirements in
Section 6. Furthermore, we make some numerical simulations in Section 7 to benchmark
our findings by applying our techniques to a problem from the field of optimal designs
of experiments and to a random LMI with matrices sampled from the Gaussian unitary
ensemble.
2. Preliminaries
We begin by fixing our notation. For brevity, we will write the set {1, . . . , d} as [d].
The set of d × D matrices over some field K ∈ {R, C} will be written as Md,D (K) and
just Md (K) if d = D. We will often omit the underlying field if it is not relevant for
the statement. We will denote by Msym
the set of symmetric d × d matrices in the
d
real case and the set of Hermitian matrices in the complex case. For A ∈ Md , AT will
denote the transpose of A in the real case and the Hermitian conjugate in the complex
case. To avoid cumbersome notation and redundant theorems, we will prove most of the
statements only for real matrices. However, note that all statements translate to the
complex case in a straightforward fashion. We will state most of the definitions just for
real matrices, but it should be clear how to generalize them to the complex case. For
A ∈ Msym
we will write A ≥ 0 if A is positive semidefinite. We will denote the cone of
d
d × d positive semidefinite matrices by Sd+ and its interior, the positive definite matrices,
3
by Sd++ .
Definition 2.1 (Schatten p-norm and Hilbert-Schmidt scalar product). Let A ∈ Md
and s1 , . . . , sd be its singular values. We define the Schatten p-norm of A for p ≥ 1,
denoted by kAkp , to be given by
d
X
p
spi
kAkp =
i=1
and for p = ∞ by kAk∞ = max si The Schatten 2−norm is induced by the Hilbert1≤i≤d
Schmidt scalar product, which is given by
hA, BiHS = Tr AT B
for A, B ∈ Md .
We will sometimes refer to the case p = 2 as the Hilbert-Schmidt (HS) norm and
p = ∞ as the operator norm.
Definition 2.2 (Positive Map). A linear map Φ : MD → Md is called positive if
+
) ⊆ Sd+ .
Φ(SD
The structure of the set of positive maps is still not very well understood [Stø13].
For our purposes, however, we will only need maps of the form Φ(X) = SXS T with
S ∈ Md,D , which are easily seen to be positive.
We will adopt the standard Big O notation for the asymptotic behavior of functions.
That is, for two functions f, g : R → R, we will write g = O(f ) if there exists a
constant M > 0 such that for all x > x0 we have |g(x)| ≤ M |f (x)|. Analogously, we
write g = Ω(f ) if there exists a constant M > 0 such that for all x > x0 we have
|g(x)| ≥ M |f (x)|.
The following families of matrices will play a crucial role for our purposes:
Definition 2.3 (Johnson-Lindenstrauss transform). A random matrix S ∈ Md,D (K) is
a Johnson-Lindenstrauss transform (JLT) with parameters (ǫ, δ, k) if with probability at
least 1 − δ, for any k-element subset V ⊆ KD , for all v, w ∈ V it holds that
|hSv, Swi − hv, wi| ≤ ǫkvk2 kwk2 .
Note that one usually only demands that the norm of the vectors involved is distorted
by at most ǫ in the definition of JLTs, but this is equivalent to the definition we chose by
the polarization identity. There are many different examples of JLTs in the literature and
we refer to [Woo14] and references therein for more details. Most of the constructions of
JLTs focus on real matrices, but in Section A we show how to lift some of these results
to cover complex matrices. The most prominent JLT is probably the following:
√1 R
d
∈ Md,D (R), where the entries of R
are i.i.d. standard Gaussian random variables. If d = Ω(ǫ−2 log kδ−1 ), then S is an
(ǫ, δ, k)-JLT.
Theorem 2.4. Let 0 < ǫ, δ < 1 and S =
4
Proof. We refer to [Woo14, Lemma 2.12] for a proof.
The main drawback of using Gaussian JLTs is that these are dense matrices. We will
denote by nnz(X) the number of nonzero elements of a matrix X ∈ MD . As we later
want to compute products of the form SXS T , it will be advantageous to have a sparse
S to speed up the computation of this product. The computational cost of forming this
product will most often be the bottleneck of our algorithms. Fortunately, there has been
a lot of recent work on sparse JLTs. In particular, we have the following almost optimal
result.
Theorem 2.5 (Sparse JLT
1.1]). There
is an (ǫ, δ, k)-JLT S ∈ Md,D
[KN14, Section
−2
−1
−1
−1
with d = O ǫ log kδ
and s = O(ǫ log kδ
) nonzero entries per column.
Proof. We refer to [KN14, Section 1.1] for a proof and remark that the proof is constructive.
Given some JLT S ∈ Md,D , the positive map Φ : MD → Md , X 7→ SXS T will be
called the sketching map and d the sketching dimension.
We will now fix our notation for semidefinite programs. Semidefinite programs are
a class of optimization problems in which a linear functional is optimized under linear
constraints over the set of positive semidefinite matrices. We refer to [LA16] for an
introduction to the topic. There are many equivalent ways of formulating SDPs. In this
work, we will assume w.l.o.g. that the SDPs are given in the following form:
Definition 2.6 (Sketchable SDP). Let A, B1 , . . . , Bm ∈ Msym
and γ1 , . . . , γm ∈ R. We
D
will call the constrained optimization problem
maximize
Tr (AX)
subject to
Tr (Bi X) ≤ γi ,
X ≥ 0,
i ∈ [m]
(1)
a sketchable SDP.
Sometimes we will also refer to a sketchable SDP as the original problem. We will
see later how to approximate the value of these SDPs. SDPs have a rich duality theory,
in which, instead of optimizing over positive semidefinite matrices that satisfy certain
constraints, one optimizes over the points that satisfy a linear matrix inequality (LMI).
The dual problem of a sketchable SDP is given by the following:
minimize
subject to
hc, γi
m
X
ci Bi − A ≥ 0
(2)
i=1
c ∈ Rm
+,
m
where γ ∈ Rm is the vector with coefficients γi . Here, Rm
+ = { x ∈ R : xi ≥ 0 }. SDPs
and LMIs will be called feasible if there is at least one point satisfying all the constraints,
5
otherwise we will call it infeasible. A sketchable SDP will be called strictly feasible if
there is a point X ≥ 0 such that all the constraints in (1) are satisfied with strict
inequality.
Under some conditions the primal problem (1) and the dual problem (2) have the same
value. One widely used sufficient condition is Slater’s condition [LA16], which asserts
that if we have a strictly feasible point of full rank for the primal problem and if the
dual problem is feasible, then the primal and the dual have the same value and there is
an optimal solution to the dual problem.
We will need some standard concepts from convex analysis. Given a1 , . . . , an ∈ V
for a vector space V , we denote by conv{a1 , . . . , an } the convex hull of the points. By
cone{a1 , . . . , an } we will denote the cone generated by these elements and a convex cone
C will be called pointed if C ∩ −C = {0}.
3. Sketching the Hilbert-Schmidt product with positive maps
One of our main ingredients to sketch an SDP or LMI will be a random positive map
Φ : MD → Md that preserves the Hilbert-Schmidt scalar product with high probability.
We demand positivity to assure that the structure of the SDP or LMI is preserved.
Below, we first consider the example Φ(X) = SXS T with S a JLT.
Lemma 3.1. Let B1 , . . . , Bm ∈ Msym
and S ∈ Md,D be an (ǫ, δ, k)-JLT with ǫ ≤ 1 and
D
k such that
m
X
rank Bi .
k≥
i=1
Then
P ∀i, j ∈ [m] : |Tr SBi S T SBj S T − Tr (Bi Bj ) | ≤ 3ǫkBi k1 kBj k1 ≥ 1 − δ.
(3)
Proof. Observe that the eigenvectors of the Bi corresponding to nonzero eigenvalues of
the Bi form a subset of cardinality at most k of KD . Let A, B ∈ {B1 , . . . , Bm }. As S is
an (ǫ, δ, k)-JLT, with probability at least 1 − δ we have for all normalized eigenvectors
ai of A and bj of B that
|hSai , Sbj i| − |hai , bj i| ≤ ǫ
by the reverse triangle inequality. We also have that for any ai , bj
√
kSai k2 ; kSbj k2 ≤ 1 + ǫ,
again by the fact that S is a JLT. As ǫ ≤ 1 and by the Cauchy-Schwarz inequality, it
follows that
|hSai , Sbj i| + |hai , bj i| ≤ 3
and hence
|hSai , Sbj i|2 − |hai , bj i|2 ≤ 3ǫ.
6
(4)
Now let λi and µj be the eigenvalues of A and B, respectively. We have:
T
Tr SAS SBS
T
− Tr (AB) =
D
X
i,j=1
λi µj (|hSai , Sbj i|2 − |hai , bj i|2 )
≤ 3ǫ
D
X
i,j=1
|λi ||µj | = 3ǫkAk1 kBk1
with probability at least 1 − δ. As A, B were arbitrary, the claim follows.
The scaling of the error with the Schatten 1-norm of the matrices involved in Lemma 3.1
is highly undesirable and the estimates used to prove it are admittedly crude. We
note that a similar estimate was proved in [SH15]. Moreover, just using the fact that
D(D+1)
D(D+1)
2
as a Hilbert space, we could use an (ǫ, δ, k)-JLT L for R 2
Msym
D (R) ≃ R
and isometrically embed the resulting vector into a symmetric matrix. Denoting this
transformation by L̃, we obtain
h
i
P ∀i, j ∈ [n] : Tr L̃(Bi )L̃(Bj ) − Tr (Bi Bj ) ≤ 3ǫkBi k2 kBj k2 ≥ 1 − δ.
That is, if only demand the sketching map to map symmetric matrices to symmetric
matrices, we clearly obtain a better scaling of the error with this procedure. Note,
however, that the map L may not be positive, one of the requirements to later sketch
SDPs. The next theorem shows that a scaling of the error with the Schatten 2−norm
of the matrices involved is not possible with positive maps if we want to achieve a
non-trivial compression.
Theorem 3.2. Let Φ : MD → Md be a random positive map such that with strictly
positive probability for any Y1 , . . . YD+1 ∈ MD and 0 < ǫ < 14 we have
Then d = Ω(D).
|Tr Φ(Yi )T Φ(Yj ) − Tr YiT Yj | ≤ ǫkYi k2 kYj k2 .
(5)
Proof. We refer to Appendix B for a proof.
One could hope to achieve a better bound for low rank matrices, but we note that
this does not significantly improve our bound,
as for A ∈ MD of rank r we have kAk1 ≤
√
rkAk2 . That is, by choosing an rǫ , δ, k -JLT, we may ensure that inequality (3) holds
with the HS norm if the rank of the matrices involved is bounded by r ≪ d. This
just increases the dimension of the involved JLT matrices by a factor of r 2 if we have
the usual ǫ−2 dependence on the dimension for the JLTs. It remains open if one could
achieve a better compression for a sublinear number of matrices.
7
4. Sketching linear matrix inequality feasibility problems
In this section we will show how to use JLTs to certify that certain linear matrix inequalities (LMI) are infeasible by showing that an LMI of smaller dimension is infeasible. We
will consider inequalities like the ones in the following lemma:
Lemma 4.1. Let A, B1 , . . . , Bm ∈ Msym
D \ { 0 } such that
m
X
i=1
ci Bi − A 6≥ 0
(6)
for all c ∈ Rm
+ . Suppose further that
Λ = cone{B1 , . . . , Bm }
+
+
such that for all i ∈ [m]
= {0}. Then there exists a ρ ∈ SD
is pointed and Λ ∩ SD
Tr (ρBi ) < 0,
Tr (−Aρ) < 0
and
Tr (ρ) = 1.
+
∩ E = ∅. Suppose there
Proof. Let E = conv{−A, B1 , . . . , Bm }. We will show that SD
m
P
+
exists an X = −p0 A +
∩ E with p ∈ [0, 1]m+1 . If p0 > 0, we could rescale
p i Bi ∈ S D
i=1
X by p−1
0 and obtain a feasible point for (6), a contradiction. If p0 = 0 and X 6= 0,
+
= {0}. And if X = 0, the cone Λ would not be
this would in turn contradict Λ ∩ SD
pointed. From these arguments it follows that 0 6∈ E. The set E is therefore closed,
+
. We may thus find a
convex, compact and disjoint from the convex and closed set SD
+
such that w.l.o.g.
hyperplane that strictly separates SD from E. That is, a ρ ∈ Msym
D
+
+
, and Tr (Y ρ) < 0 for all Y ∈ E. As Tr (ρX) ≥ 0
, as 0 ∈ SD
Tr (ρX) ≥ 0 for all X ∈ SD
for all X ≥ 0, it follows that ρ is positive semidefinite and it is clear that by normalizing
ρ we may choose ρ with Tr (ρ) = 1.
The main idea is now to show that under these conditions we may sketch the hyperplane in a way that it still separates
the set of positive semidefinite matrices and the
P
sketched version of the set { m
γ
B
i=1 i i − A|γi ≥ 0}.
Theorem 4.2. Let A, B1 , . . . , Bm ∈ Msym
D \ { 0 } such that
m
X
i=1
ci Bi − A 6≥ 0
for all c ∈ Rm
+ . Suppose further that
Λ = cone{B1 , . . . , Bm }
+
+
be such that for all i ∈ [m]
= {0}. Moreover, let ρ ∈ SD
is pointed and Λ ∩ SD
Tr (ρBi ) < 0,
Tr (−Aρ) < 0
8
and
Tr (ρ) = 1.
Set
1
ǫ = min
2
Tr (ρB1 )
Tr (ρBm ) Tr (ρA)
,...,
,
kB1 k1
kBm k1
kAk1
and take S ∈ Md,D to be an (ǫ, δ, k)-JLT. Here,
k ≥ rank A + rank ρ +
m
X
rank Bi .
i=1
Then
m
X
i=1
ci SBi S T − SAS T 6≥ 0
(7)
for all c ∈ Rm
+ , with probability at least 1 − δ.
Proof. It should first be noted that ρ exists and ǫ > 0 by Lemma 4.1. The matrix ρ
defines a hyperplane that strictly separates the set
)
( m
X
ci Bi − A c ∈ Rm
E=
+
i=1
+
. We will now show that SρS T strictly separates the sets
and SD
)
( m
X
T
T
m
ci SBi S − SAS
ES =
c ∈ R+
i=1
and Sd+ with probability at least 1 − δ, from which the claim follows. Note that by our
choice of ρ and ǫ, it follows from Lemma 3.1 that we have
Tr SρS T SBi S T ≤ Tr (ρBi ) + ǫkBi k1 < 0
with probability at least 1 − δ and similarly for −A instead of Bi . Therefore, it follows
that Tr(ZSρS T ) < 0 for all Z ∈ ES . As SρS T is a positive semidefinite matrix, it follows
that Tr Y SρS T ≥ 0 for all Y ∈ Sd+ . We have therefore found a strictly separating
+
and the LMI (7) is infeasible.
hyperplane for ES and SD
Theorem 4.2 suggests a way of sketching feasibility problems of the form
m
X
i=1
ci Bi − A ≥ 0,
9
c ∈ Rm
+.
(8)
If we are interested in the case in which it is infeasible, we investigate if the LMI
m
X
i=1
ci SBi S T − SAS T ≥ 0,
c ∈ Rm
+
(9)
is feasible, for S a suitably chosen JLT as in Theorem 4.2. If Equation (9) is infeasible,
we know that Equation (8) is infeasible, as any choice of a feasible c leads to a feasible
c for Equation (9). Moreover, it follows from Theorem 4.2 that if the cone spanned by
the Bi is well-behaved enough and the JLT is suitably chosen, it only happens with very
low probability that Equation (9) is feasible and Equation (8) is not. To obtain more
concrete bounds on the probability that the original problem is feasible, one would need
to know the parameter ǫ, which is not possible in most applications. We emphasize that
this is a black box algorithm. That is, we can decide whether a large instance of an LMI
is infeasible by showing that a smaller instance of an LMI is infeasible. In Section 6 we
will discuss the implications for the complexity and memory usage of the last theorems.
5. Approximating the value of semidefinite programs through
sketching
We will now show how to approximate with high probability the value of a sketchable
SDP by conjugating the target matrix and the matrices that describe the constraints
with JLTs and subsequently solving a smaller instance of an SDP. The next theorem
shows that in general it is not possible to approximate the value of a sketchable SDP
using linear sketches with high probability and that we need to make further assumptions
on the problem to achieve a non-trivial compression using linear maps.
Theorem 5.1. Let Φ : M2D → Rd be a random linear map such that for all sketchable
SDPs there exists an algorithm which allows us to estimate the value of an SDP up to a
constant factor 1 ≤ τ < √23 given the sketch {Φ(A), Φ(B1 ), . . . , Φ(Bm )} with probability
at least 9/10. Then d = Ω(D 2 ).
Proof. By the duality relations for Schatten p-norms, it is easy to see that the value of
the SDP
maximize
Tr (AX)
subject to
Tr (X) ≤ 1
(10)
X≥0
with
A=
0
GT
G
0
is twice the operator norm of the matrix G ∈ MD . In [Woo14, Theorem 6.5] it was shown
that any algorithm that estimates the operator norm of a matrix from a linear sketch
10
with probability larger than 9/10 must have sketch dimension Ω(D 2 ). As the sketch
{ Φ(A), Φ(1) } would thus allow to sketch the operator norm, the assertion follows.
The above result remains true even if we restrict to SDPs that have optimal points
with small Schatten 1-norm and low rank. This follows from the fact that the SDP given
in Equation (10) has an optimal solution with rank 1 and trace 1.
We may even restrict to SDPs whose value scales sublinearly. To see that, notice that
to show that the operator norm cannot be sketched,
√ [Woo14] constructs two families of
random matrices whose operator norm is of order D (cf. [Woo14, Lemma 6.3]). As the
class of SDPs considered here covers this problem, we obtain the claim.
As we will see soon, the main hurdle to sketch SDPs and thus overcome the last no-go
theorem is that we also need to suppose that the matrices that define the constraints
and the target function have a small Schatten 1-norm, not only one optimal solution.
To this end, we define:
Definition 5.2 (Sketched SDP). Let A, B1 , . . . , Bm ∈ Msym
D , η, γ1 , . . . , γm ∈ R and
+
∗
ǫ > 0. Given that an optimal point X ∈ SD of the sketchable SDP defined through
these matrices satisfies Tr (X ∗ ) ≤ η and given a random matrix S ∈ Md,D , we call the
optimization problem
maximize
Tr SAS T Y
subject to
Tr SBi S T Y ≤ γi + µkBi k1 ,
Y ≥0
i ∈ [m]
(11)
with µ = 3ǫη the sketched SDP.
The motivation for defining the sketched SDP is given by the following theorem.
Theorem 5.3. Let A, B1 , . . . , Bm ∈ Msym
D , η, γ1 , . . . , γm ∈ R and ǫ > 0. Denote by α
the value of the sketchable SDP and assume it is attained at an optimal point X ∗ which
satisfies Tr (X ∗ ) ≤ η. Moreover, let S ∈ Md,D be an (ǫ, δ, k)-JLT, with
∗
k ≥ rank X + rank A +
m
X
rank Bi .
i=1
Let αS be the value of the sketched SDP defined by A, Bi and S. Then
αS + 3ǫηkAk1 ≥ α
with probability at least 1 − δ.
Proof. It follows from Lemma 3.1 that SX ∗ S T is a feasible point of the sketched SDP
with probability at least 1 − δ. Again by Lemma 3.1, we have that
Tr SAS T SX ∗ S T ≥ Tr (AX ∗ ) − 3ǫηkAk1 .
11
It then follows that αS + 3ǫηkAk1 ≥ α.
In Section 6, we will discuss the implications for memory usage and complexity of
approximating the value of an SDP.
Note that the optimal value of an SDP is not necessarily attained. We could also
demand X ∗ to be only close to optimality which would slightly increase the error made
by the sketch. Since this makes notation more cumbersome, we assume the existence of
such an optimal point.
Although it is not customary to assume a bound on the Schatten 1-norm of an optimal
solution to an SDP, it is common to assume that for example the solution lies in a given
ellipsoid when using the ellipsoid method to solve SDPs [LA16]. From such assumptions
it is straightforward to derive bounds on the HS norm of the solution. If we are also
given that an optimal solution to the SDP is low rank, the HS norm gives a good upper
bound on the Schatten 1-norm, as remarked after the proof of Lemma 3.1. Solutions of
low rank of SDPs have been extensively studied over the past years and there are many
results available in the literature which guarantee that the optimal solution to an SDP
has low rank. In general, it has been shown [Bar95, Pat98] that if we have m constraints
√
⌋.
and the SDP is feasible, there is an optimal solution with rank at most r = ⌊ 8m+1−1
2
Notice that Theorem 5.3 does not rule out the possibility that the value of the sketched
problem is much larger than that of the sketchable SDP. To investigate this issue, we
introduce the following:
Definition 5.4 (Relaxed SDP). Let A, B1 , . . . , Bm ∈ Msym
D , η, γ1 , . . . , γm ∈ R and
ǫ > 0. Given that an optimal point X ∗ of the sketchable SDP defined through these
matrices satisfies Tr (X ∗ ) ≤ η, we call the optimization problem
maximize
Tr (AX)
subject to
Tr (Bi X) ≤ γi + µkBi k1 ,
X≥0
i ∈ [m]
(12)
with µ = 3ǫη the relaxed SDP.
Notice that, given a feasible point Y to the sketched SDP, S T Y S is a feasible point
for the relaxed problem by the cyclicity of the trace. It follows that if the values of the
original and the relaxed are close, the values of the original and the sketched problem
are close as well. We formalize this intuition and prove the following bound in Appendix
C.
Theorem 5.5. We are in the setting of Definition 2.6. Assume that there exists an
X0 > 0 such that all the constraints of the sketchable SDP are strictly satisfied and that
the dual problem is feasible. Then the value of the sketched SDP αS is bounded by
αS ≤ α(0) + ǫC1 (α(0) − Tr (AX0 )) /C2 .
12
Here
C1 = max { 3ηkBi k1 | i ∈ [m] } ,
C2 = min { (γi − Tr (Bi X0 )) | i ∈ [m] } ,
where η ≥ Tr (X ∗ ) for an optimal point X ∗ of the sketchable SDP.
Proof. We refer to Appendix C for the proof.
Note that this statement is not probabilistic and holds regardless of the choice of the
sketching map. It could also be the case that Tr (AX0 ) itself gives a better lower bound
on the value than the one in Theorem 5.5. One can therefore say that Theorem 5.5
guarantees that in general the value of the sketched SDP cannot differ significantly
from the value of the original one if we have feasible points which are not too close to
the boundary. Combining Theorem 5.3 and Theorem 5.5 it is possible to pick ǫ small
enough to have an arbitrarily small additive error under some structural assumptions
on the SDP. That is, we need bounds on the Schatten 1-norms of A and Bi , be given a
strictly feasible point for the relaxed SDP and a bound on the Schatten 1-norm of an
optimal solution to the sketchable SDP.
In the case that all the γi > 0 for a sketchable SDP we may obtain a bound on the
value and an approximate solution to it in a much simpler way. This class includes the
so-called semidefinite packing problems [IPS05], where we have in addition that Bi ≥ 0.
Note that we may set all γi = 1 w.l.o.g. by dividing Bi by γi . We then obtain:
Theorem 5.6. For a sketchable SDP with γi = 1 and κ = max kBi k1 , we have that
i∈[m]
αS
≤ α,
1+ν
(13)
where ν = 3ǫηκ. Moreover, denoting by XS∗ an optimal point of the sketched SDP, we
1
have that 1+ν
S T XS∗ S is a feasible point of the sketchable SDP that attains this lower
bound. Furthermore, if kBi k1 = κ for all i ∈ [m] it can be checked if this lower bound is
the optimal value.
Proof. The lower bound in Equation (13) follows immediately from the cyclicity of the
1
S T XS∗ S is a feasible point of the sketchable SDP. Given an optimal solution
trace, as 1+ν
{ci }i∈[m] to the dual of the sketched SDP and that kBi k1 = κ for all i ∈ [m], it is
possible to check if the lower bound given by the sketched SDP is indeed optimal as
follows. Slater’s condition holds for the sketched and sketchable SDP, as ϑ1 is a strictly
feasible point for ϑ > 0 small enough. If we have
m
X
i=1
ci Bi − A ≥ 0,
then the obtained feasible point is indeed optimal for the sketchable SDP by strong
duality.
13
It is possible to relax the condition that all the Schatten 1-norms of the matrices that
define the constraints are the same and still obtain a lower bound for which it can be
checked whether it is indeed optimal. To achieve this, it is necessary to modify all the
constraints of the sketched SDP, as in Equation (11), to 1 + µκ instead of 1 + µkBi k1 .
In the primal picture, we will find an optimal point of the sketchable SDP through the
+
such that S T Y S = X ∗ , for X ∗ an optimal
sketched SDP whenever there is a Y ∈ SD
point of the sketchable SDP. Note that for semidefinite packing problems it is possible
to derive a bound on the Schatten 1-norm of an optimal solution in a straightforward
way.
Lemma 5.7. Let B1 , . . . , Bm ∈ MD be positive semidefinite matrices such that the
m
P
smallest strictly positive eigenvalue of
Bi is given by λ. Then for a sketchable SDP
i=1
with constraints Bi , γ1 , . . . , γm ∈ R+ , A ≥ 0 and finite value there exists an optimal
point X ∗ such that
m
1X
∗
γi .
(14)
Tr (X ) ≤
λ
i=1
Proof. As we assume that the SDP has a finite value, we may restrict to solutions whose
m
P
support is contained in the support of
Bi . Denote by P the projection onto the
support of
m
P
i=1
Bi . We then have
i=1
λP ≤
1
m
X
Bi .
(15)
i=1
Conjugating both sides with (X ∗ ) 2 and taking the trace we obtain
λTr (X ∗ ) ≤
m
X
Tr (X ∗ Bi ) ,
i=1
as we supposed w.l.o.g. that the support of X ∗ is contained in the support of
X ∗ is a feasible point, we have Tr (X ∗ Bi ) ≤ γi and we obtain the claim.
m
P
Bi . As
i=1
Of all the assumptions we needed for the results of Theorem 5.3, the bound on the
Schatten 1-norm of an optimal solution to the SDP is arguably the most difficult to
show, as bounds of this form are not readily available in the literature. Moreover, some
SDPs have Tr (X) ≤ η as constraint and would be natural candidates to apply these
methods to. As Tr (1) = D we will not be able to obtain any non-trivial compression
with the scheme discussed so far. However, if one is only interested in obtaining an
upper bound on the value of the SDP, it is still possible to have 1 as a constraint or as
the target matrix and achieve a non-trivial compression.
Theorem 5.8. Let A, B1 , . . . , Bm−1 ∈ Msym
D , Bm = 1. Further, let γ1 , . . . , γm−1 ∈ R,
γm = η and ǫ > 0. Denote the value of the sketchable SDP by α and assume it is attained
14
at an optimal point X ∗ . Moreover, let S ∈ Md,D be an (ǫ, δ, k)-JLT, with
∗
k ≥ rank X + rank A +
m
X
rank Bi .
i=1
Let α′S be the value of the modifed sketched SDP defined by A, Bi and S, given by
maximize
Tr SAS T X
subject to
Tr SBi S T X ≤ γi + µkBi k1 , i ∈ [m − 1]
(16)
Tr (X) ≤ (1 + ǫ)η,
X≥0
with µ = 3ǫη. Then
α′S + 3ǫηkAk1 ≥ α
with probability at least 1 − δ.
Proof. The proof is essentially the same as the one of Theorem 5.3, as we have that
|Tr SX ∗ S T − Tr (X ∗ ) | ≤ 3ǫTr (X ∗ ) and so SX ∗ S T is a feasible point with probability
at least 1 − δ.
The main difference between the modified sketched SDP and the sketched SDP is that
here we do not conjugate the identity with S, only the other constraints. With this, we
do not have that S T X ∗ S is a feasible point of the relaxed SDP, but we do not need the
assumption Tr (X ∗ ) ≤ η to obtain an upper bound. It should be clear from Theorem 5.8
that we may also optimize over the trace, i.e. A = 1, without conjugating with S and
still have an upper bound and non-trivial compression.
We may therefore summarize the results of this section as follows. If we want to
obtain an upper bound on the value of the sketchable SDP with our techniques, it is
necessary to have upper bounds on the Schatten 1-norm of all the matrices that define
the constraints, the target matrix and of an optimal solution. We may then choose a
JLT of suitable dimension to solve the sketched SDP, whose value will allow us to infer
an upper bound to the original problem with high probability. If we are addtionaly given
a strictly feasible point of the sketchable SDP or if we are solving a semidefinite packing
problem, we also obtain a lower bound on the value of the sketchable SDP in terms of
the sketched one. In the case of semidefinite packing problems, we even obtain a feasible
point of the sketchable SDP whose value is close to the sketched value. If we are not
given a bound on the Schatten 1-norm of an optimal solution, we may impose it as a
constraint as in Theorem 5.8 and obtain an upper bound on the value of the sketchable
SDP constrained to points which have their Schatten 1-norm bounded by η. Although
we are able to drop the assumption on the Schatten 1-norm of an optimal solution, we
are not able to prove that this upper bound cannot differ significantly from the true
value in this case.
15
6. Complexity and memory gains
In this section, we will discuss how much we gain by considering the sketched SDP
instead of the sketchable SDP. We focus on the results of Section 5, but the discussion
carries over to the results of Section 4. Throughout this section we will assume that
we are guaranteed that the Schatten 1−norm of an optimal solution to our SDP and of
the matrices that define the constraints is O(1). It is therefore Theorem 5.3 for which
we need a sketch of appropriate size. The theorem states that we need an (ǫ, δ, k)JLT for the upper bound on α to hold with probability at least 1 − δ. As stated in
Theorem 5.3, we can choose
a sketching matrix S ∈ Md,D with d = O(ǫ−2 log kδ−1 )
and s = O(ǫ−1 log kδ−1 ) nonzero entries per column. Here k is as in Theorem 5.3. The
cost of generating S is at most O(dD), which will be of smaller order than the necessary
matrix multiplications. We will therefore not take this cost into account for the rest of
the analysis.
One could argue that one needs to know the Schatten 1-norm of the different matrices
that define our constraints for estimating the value or obtaining more concrete bounds
for the feasibility problems. We will, however, suppose that an upper bound on the
Schatten 1-norm of an optimal solution and the constraints is given or that this can be
computed in a time which is O(D 2 ). This is the case if for example we have a semidefinite
packing problem, where the matrices are positive semidefinite and we can compute their
Schatten 1-norm in O(D) time.
To generate the sketched SDP, we need to compute m + 1 matrices of the form SBS T ,
where B ∈ MD . Each of this computations needs O(max { nnz(B), Dd } ǫ−1 log kδ−1 )
operations. In the worst case, when all matrices { A, B1 , . . . , Bm } are dense and have
full rank, this becomes O(mD 2 log(mD)) operations to generate the sketched SDP for
fixed ǫ and δ.
Let us collect these considerations in a proposition:
Proposition 6.1. Let A, B1 , . . . , Bm ∈ Msym
D , γ1 , . . . , γm ∈ R of a sketchable SDP be
given. Furthermore, let z := max { nnz(A), nnz(B1 ), . . . , nnz(Bm ) } and SDP(m, d, ζ) be
the complexity of solving a sketchable SDP (up to accuracy ζ) of dimension d. Then a
number of
O(max z, Dǫ−2 log kδ−1 ǫ−1 m log kδ−1 + SDP(m, ǫ−2 log kδ−1 , ζ))
operations is needed to generate and solve the sketched SDP, where k is defined as in
Theorem 5.3.
It is easy to see that we can parallelize computing the matrices SBi S T . Typically,
the costs of forming the sketched matrices SBi S T dominates the overall complexity.
For example, using the ellipsoid method [GS88,
Chapter 3], the complexity of solving
6
an SDP becomes SDP(m, D, ζ) = O(max m, D 2 D
(cf. [Bub15, p.250]).
log(1/ζ))
Assuming that ǫ, δ and ζ are fixed, we need O(max m, D 2 D 6 ) operations to solve
the sketchable SDP, compared to O(mD 2 log(mD)) operations to obtain an approximate
solution via first forming the sketched problem and then solving it. Admittedly, the
16
ellipsoid method is not used
but using interior point methods, we still need
3in practice,
2
2
SDP(m, D, ζ) = O(max m , D m , mD ω D 0.5 log(D/ζ)) operations [dK02, Chapter
5], where ω is the exponent of matrix multiplication. The best known algorithms achieve
ω ≈ 2.37 [LG14]. If the SDP can be sketched, doing so gives a speedup as long as the
complexity of solving the SDP directly is Ω(mD 2+µ ), where µ > 0.
A great advantage is that for the sketched problem, we only need to store m + 1
matrices of size d × d instead of D × D. We collect this in a proposition.
Proposition 6.2. Let A, B1 , . . . , Bm ∈ Msym
D , γ1 , . . . , γm ∈ R be a sketchable SDP.
2
−4
Then we need only store O(mǫ log(mk/δ) ) entries for the sketched problem, where k
is defined as in Theorem 5.3.
7. Applications and numerical examples
7.1. Estimating the value of a semidefinite packing problem
Inspired by [Sag11] we will test our techniques on an SDP stemming from the field of
optimal design of experiments. The problem is the following: an experimenter wishes
to estimate the quantity hc, θi, where θ ∈ RD is an unknown D-dimensional parameter
and c ∈ RD is given. To this end, one is given linear measurements of the parameter
yi = Ai θ, up to a (centered) measurement noise for Ai ∈ MD . We refer to [Puk06] for
more details on the topic. To find the amount of effort to spend on the i−th experiment
to minimize the variance is given by the SDP
maximize
Tr ccT X
subject to
Tr (Mi X) ≤ 1,
X≥0
i ∈ [m]
(17)
with Mi = ATi Ai . This problem always admits optimal solutions of rank 1 [Sag11]. We
generated random instances of this SDP in the following way:
1. We sampled four matrices Ai distributed as follows: the first three rows of Ai are
sampled independently from the uniform distribution on the unit sphere in RD .
The other D − 3 rows of Ai are set to 0.
2. Given the Ai , we generate c by getting four samples k1 , k2 , k3 , k4 from the uniform
distribution on { 1, 2, 3 } and four samples x1 , x2 , x3 , x4 from the standard normal
4
P
distribution. c is then given by
xi (Ai )ki , where (Ai )ki is the ki -th row of Ai .
i=1
This gives matrices Mi of rank 3 almost surely and Schatten 1-norm equal to 3. The fact
that c is a linear combination of the rows of Ai ensures that the problem is bounded, as
can be easily seen by looking at the dual problem [Sag11]. Note that this is a semidefinite
packing problem, so we are able to use the results of Theorem 5.6 to obtain a lower bound.
There exists an optimal solution whose Schatten 1-norm is bounded by 8 with very high
17
D
100
100
200
200
500
500
700
700
d
10
20
20
40
50
100
70
140
Value
2.52
2.50
2.69
2.53
2.55
2.66
2.57
2.49
Error L.B.
0.0880
0.00
0.00
0.00
0.00
0.00
0.00
0.00
Error U.B.
0.156
0.250
0.269
0.00102
0.255
0.266
0.257
0.249
M.R.T. Sketchable [s]
1.27
1.17
6.50
6.82
98.0
97.6
557
548
M.R.T Sketch [s]
0.324
0.305
0.299
0.375
0.453
1.23
1.38
3.53
Table 1: For each combination of the sketchable dimension (D) and dimension of the
sketch (d) we have generated 40 instances of the SDP in Equation (17). Here
“M.R.T.” stands for mean running time, “L.B.” stands for lower bound and
“U.B.” for upper bound and each column shows the mean of the sample. The
column “Value” stands for the optimal value of the sketchable SDP.
probability. This easily follows from Theorem 5.7 and the fact that i.i.d. unit vectors are
almost orthogonal. To obtain the upper bound we have used the results of Theorem 5.3
and for the lower bound Theorem 5.6. We have chosen η = 0.1 to generate these results.
We used sparse JLTs with sparsity parameter s = 1 [KN14] to obtain faster matrix
multiplications to form the sketches. We define the error of the lower bound to be given
1
αS and of the upper to be αS − α. To solve the SDP given in Equation (17)
by α − 1+η
we used cvx, a package for specifying and solving convex programs [GB14, GB08].
As we can see, the results of Table 1 show that, excluding the case where the sketching
dimension was 10, we were able to find feasible points which were numerically indistinguishable from being optimal by using our sketching methods. Moreover, the time needed
to find an optimal solution was smaller by 1 or 2 orders of magnitude.
7.2. Linear matrix inequality feasibility problems
We will now apply our techniques to an LMI feasibility problem. Let Gi ∈ Md′ , i ∈ [m],
be random matrices sampled independently from the Gaussian unitary ensemble (GUE).
See for example [AGZ10, Section 2.2] or Section D for their definition. Consider the
rescaled and shifted matrices G̃i (α) = √1d′ Gi + α1 for some α ∈ R. For V : Md′ → MD
a random isometry, we will test our techniques on the feasibility of the LMI
m
X
i=1
ti V G̃i (α)V ∗ − V V ∗ ≥ 0,
t ∈ Rm
+
(18)
depending on α. As V is an isometry, Equation (18) is clearly feasible if and only if the
LMI
m
X
ti G̃i (α) − 1 ≥ 0, t ∈ Rm
(19)
+
i=1
18
is feasible. Using standard techniques from random matrix theory, we show that Equation (19), and so Equation (18), is feasible with high probability for α > √2m and
infeasible with high probability for α < √2m in case m ≪ d′ .
We refer to Appendix Section D for a proof of this claim. This therefore allows us to
quantify “how close to feasible” the LMI inequality is in terms of how close α is to √2m
and to know whether the LMI was feasible or not. We will choose d′ ≪ D, as this way
we avoid having a Schatten 1-norm of the matrices that define the LMI which is of the
same order as the dimension. The technique we used to solve Equation (18) is the same
as discussed in Section 4. That is, we will check for the feasibility of
m
X
i=1
ti SV G̃i (α)V ∗ S T − SV V ∗ S T ≥ 0,
t ∈ Rm
+
(20)
for a complex Gaussian JLT S. That is, S = √12 (S1 + iS2 ) with S1 and S2 independent
Gaussian JLTs. We refer to Section A for a proof that this choice of random matrices
indeed gives a JLT with the same scaling of the parameters as the real one. We refer to
Theorem D.4 for a proof that the LMI defined in Equation (20) satisfies the assumptions
of Theorem 4.2 with high probability. To solve the SDP given in Equation (20) we used
cvx, a package for specifying and solving convex programs [GB14, GB08]. The results
are summarized in Table 2. We can observe that by using our methods we were able
to show that the LMI is infeasibile in a much smaller running time or even show that
certain LMI were infeasible when a direct computation was not possible due to memory
constraints in most choices of the parameters. In some cases it was, however, necessary
to increase the sketch dimension to show that the inequality was infeasibile.
8. Conclusion
We have shown how to obtain sketches of the HS product using positive maps obtained
from JLTs, how to apply these to show that certain LMI are infeasible and to obtain
approximations of the value of certain SDPs. In some cases, these techniques can lead to
significant improvements in the runtime necessary to solve the instances of the SDPs and
significant gains in the memory needed to solve them. However, the class of problems
to which these techniques can be applied is significantly restricted by the fact that the
matrices that define the constraints of the problems and a solution must have Schatten
1-norms which do not scale with the dimension for them to be advantageous. Moreover,
the no-go theorems proved here show that one cannot significantly improve our results
using positive linear maps to sketch the HS norm or to approximate the value of SDPs.
Acknowledgements
We would like to thank Ion Nechita for helpful discussions. A.B. acknowledges support
from the ISAM Graduate Center at Technical University of Munich. D.S.F. acknowl-
19
D
200
200
200
200
200
200
200
200
400
400
400
400
400
400
400
400
d
50
50
50
50
100
100
100
100
50
50
50
50
100
100
100
100
d′
100
100
100
100
100
100
100
100
200
200
200
200
200
200
200
200
α
0.3
0.4
0.5
0.6
0.3
0.4
0.5
0.6
0.3
0.4
0.5
0.6
0.3
0.4
0.5
0.6
M.R.T. Original [s]
452
407
383
407
449
344
393
457
-
M.R.T. Sketch [s]
4.65
4.31
5.71
6.46
66.3
86.6
102
91.4
4.02
7.04
3.39
2.35
114
122
118
115
Error Rate
0
0
0
0
0
0
0
0
0
0
0.975
1.0
0
0
0
0
Table 2: For each combination of the dimension of the image of the random isometry (D),
dimension of the domain of the random isometry (d′ ), dimension of the sketch
(d) and α we have generated 40 instances of the random LMI in Equation (18)
with m = 9. Here “M.R.T.” stands for mean running time. The error rate
gives the ratio of infeasible problems that were not detected to be infeasible by
sketching. A dash in the running time means that we were not able to solve
the LMI because we ran out of memory.
.
20
edges support from the graduate program TopMath of the Elite Network of Bavaria, the
TopMath Graduate Center of TUM Graduate School at Technical University of Munich.
D.S.F. is supported by the Technical University of Munich – Institute for Advanced
Study, funded by the German Excellence Initiative and the European Union Seventh
Framework Programme under grant agreement no. 291763.
References
[AGZ10] G. W. Anderson, A. Guionnet, and O. Zeitouni. An Introduction to Random
Matrices. Cambridge University Press, 2010.
[Bar95] A. I. Barvinok. Problems of distance geometry and convex properties of
quadratic maps. Discrete & Computational Geometry, 13(2):189–202, 1995.
[BEFB94] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix Inequalities in System and Control Theory, volume 15 of Studies in Applied
Mathematics. SIAM, June 1994.
[Bha97] R. Bhatia. Matrix analysis, volume 169 of Graduate Texts in Mathematics.
Springer, 1997.
[Bub15] S. Bubeck. Convex optimization: Algorithms and complexity. Found. Trends
Mach. Learn., 8(3-4):231–357, November 2015.
[BV04] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University
Press, 2004.
[dK02] E. de Klerk. Aspects of Semidefinite Programming: Interior Point Algorithms
and Selected Applications. Applied Optimization. Springer US, 2002.
[GB08] M. Grant and S. Boyd. Graph implementations for nonsmooth convex programs.
In V. Blondel, S. Boyd, and H. Kimura, editors, Recent Advances in Learning and Control, Lecture Notes in
Control and Information Sciences, pages 95–110. Springer, 2008.
http://stanford.edu/~boyd/graph_dcp.html.
[GB14] M. Grant and S. Boyd. CVX: Matlab software for disciplined convex programming, version 2.1. http://cvxr.com/cvx, March 2014.
[GS88] Lovász L. Grötschel, M. and A. Schrijver. Geometric algorithms and combinatorial optimization. Springer, 1988.
[IPS05] G. Iyengar, D. J. Phillips, and C. Stein. Approximation Algorithms for
Semidefinite Packing Problems with Applications to Maxcut and Graph Coloring, pages 152–166. Springer, 2005.
[KN14] D. M. Kane and J. Nelson. Sparser Johnson-Lindenstrauss transforms. J.
ACM, 61(1):4:1–4:23, January 2014.
21
[LA16] J.B. Lasserre and M.F. Anjos. Handbook on Semidefinite, Conic and Polynomial Optimization. International Series in Operations Research & Management Science Series. Springer, 2016.
[LG14] F. Le Gall. Powers of tensors and fast matrix multiplication. In Proceedings of
the 39th International Symposium on Symbolic and Algebraic Computation,
ISSAC ’14, pages 296–303. ACM, 2014.
[LR10] M. Ledoux and B. Rider. Small deviations for beta ensembles. Electronic
Journal of Probability, 15(41):1319 – 1343, 2010.
[Pat98] G. Pataki. On the rank of extreme matrices in semidefinite programs and
the multiplicity of optimal eigenvalues. Mathematics of Operations Research,
23(2):339–358, 1998.
[Puk06] F. Pukelsheim. Optimal Design of Experiments. Classics in Applied Mathematics. Society for Industrial and Applied Mathematics, 2006.
[RV13] M. Rudelson and R. Vershynin. Hanson-Wright inequality and sub-gaussian
concentration. Electronic Journal of Probability, 18(82):1–9, 2013.
[Sag11] G. Sagnol. A class of semidefinite programs with rank-one solutions. Linear
Algebra and its Applications, 435(6):1446 – 1463, 2011.
[SH15] C. J. Stark and A. W. Harrow. Compressibility of positive semidefinite factorizations and quantum models. In ISIT, pages 2777–2781. IEEE, 2015.
[Stø13] E. Størmer. Positive Linear Maps of Operator Algebras. Springer Monographs in Mathematics. Springer, 2013.
[Ver12] R. Vershynin. Compressed sensing, chapter Introduction to the nonasymptotic analysis of random matrices, pages 210–268. Cambridge University Press, 2012.
[VPL15] K. Vu, P.-L. Poirion, and L. Liberti. Using the Johnson-Lindenstrauss lemma
in linear and integer programming. ArXiv 1507.00990, July 2015.
[WA02] H. Wolkowicz and M. F. Anjos. Semidefinite programming for discrete optimization and matrix completion problems. Discrete Appl. Math., 123(13):513–577, November 2002.
[Wol12] M.
M.
Wolf.
Quantum
channels
and
operations:
Guided
tour.
Lecture
notes
available
at
http://www-m5.ma.tum.de/foswiki/pub/M5/Allgemeines/MichaelWolf/QChannelLecture.
2012.
[Woo14] D. P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations
and Trends in Theoretical Computer Science, 10(1–2):1–157, 2014.
22
[YUAC17] A. Yurtsever, M. Udell, Tropp J. A., and V. Cevher. Sketchy Decisions: Convex Low-Rank Matrix Optimization with Optimal Storage. ArXiv
1702.06838, February 2017.
A. Complexifying Johnson-Lindenstrauss transforms
In this Appendix we will generalize some of the results we need concerning JLTs to
complex vector spaces. We will see that, up to a constant, most of the statements that
hold in the real case also hold in the complex case. We will consider matrices of the
form √1 (S + iT ), with S, T independent and with i.i.d. entries that are sub-gaussian
2d
and show that they give us complex JLTs. Note that these constructions clearly give
sparse JLTs if the real JLTs used are sparse.
Definition A.1 (Sub-Gaussian Distribution). The probability distribution of a random
variable X is called sub-gaussian if there exist C, v > 0 such that ∀t > 0
2
P(|X| > t) ≤ Ce−vt .
p
A random variable is sub-gaussian if and only if for p ≥ 1, E(|X|p ) = O(p) 2 and the
sub-gaussian norm of X is defined as:
1
1
kXkψ2 = sup p− 2 (E(|X|p ) p .
(21)
p≥1
See for example [Ver12, Section 5.2.3] for more details on this. The main ingredient to
show how to generalize JLTs to the complex case will be the following theorem.
Theorem A.2 ([RV13, Theorem 2.1]). Let A ∈ Md,D (R) be fixed. Consider a random
vector X = (X1 , . . . , XD ), where Xi are independent random variables satisfying E[Xi ] =
0, E[Xi2 ] = 1 and kXi kψ2 ≤ K. Then for any t ≥ 0, we have
"
P [|kAXk2 − kAkHS | > t] ≤ 2 exp −
ct2
K 4 kAk2∞
#
.
Using this, we have a different way of proving the Johnson-Lindenstrauss lemma which
generalizes to the complex case. We follow the proof of [RV13, Theorem 3.1].
Lemma A.3. Let S, T ∈ Md,D (R) have independent sub-gaussian entries with E[Xij ] =
2 ] = 1 and kX k
−2 ln(2/δ) , we
0, E[Xij
ij ψ2 ≤ K, for all X ∈ { S, T }. Then for d = O ǫ
have
S + iT
√
x ∈ (1 ± ǫ)kxk2 ≥ 1 − δ
P
2d
2
for any fixed x ∈ CD .
23
Proof. Define the linear operator Φ : Md,D (C) → Cd , G 7→ Gx, where x ∈ Cd is a
fixed vector. We use the standard isomorphisms Md,D (C) ≃ R2dD and Cd ≃ R2d and
denote by Φ̃ the map Φ composed with these isomorphisms, which is now a linear map
from a real vector space to another real vector space. Moreover, observe that a matrix
of the form X + iY ∈ Md,D (C) with X, Y ∈ Md,D (R) is mapped to (X, Y ) under the
isomorphism. The map Φ̃ will play the role of A in the statement of Theorem A.2. It is
straightforward to compute the norms involved in the statement, as explained in [RV13,
Section 3.1]. We have
Φ̃
2
2
= 2dkxk22 ,
Φ̃
2
∞
= kxk22 ,
Φ̃((S, T )T )
2
2
= 2k(S + iT )xk22 .
As S, T have sub-gaussian entries, the vector (S, T ) satisfies the assumptions of Theorem
A.2 and the statement follows.
Unfortunately, the sparse JLTs discussed in Theorem 2.5 are not of this form. The
entries of these JLTs are not independent from each other, one of the assumptions of
Lemma A.3.
B. Proof of Theorem 3.2
Theorem 3.2. Let Φ : MD → Md be a random positive map such that with positive
probability for any Y1 , . . . , YD+1 ∈ MD and 0 < ǫ < 14 we have
Then d = Ω(D).
|Tr Φ(Yi )T Φ(Yj ) − Tr YiT Yj | ≤ ǫkYi k2 kYj k2 .
(22)
Proof. Let {ei }1≤i≤D be an orthonormal basis of CD and define Xi = ei eTi . As Equation (22) is satisfied with positive probability, there must exist a positive map Φ : MD →
Md such that Equation (22) is satisfied for Yi = Xi , i ∈ [D], and YD+1 = 1. As the Xi
are orthonormal w.r.t. the Hilbert-Schmidt scalar product and by the positivity of Φ we
have for i, j ∈ [D]
(
[0, ǫ], for i 6= j
(23)
Tr (Φ(Xi )Φ(Xj )) ∈
[1 − ǫ, 1 + ǫ], for i = j.
Define the matrix A ∈ MD with (A)ij = Tr (Φ(Xi )Φ(Xj )) for i, j ∈ [D]. It is clear
that A is Hermitian and that its entries are positive. We have
X
Aij = Tr (Φ(1)Φ(1)) ∈ [(1 − ǫ)D, (1 + ǫ)D] .
i,j∈[D]
As Aii ≥ (1 − ǫ), it follows that
X
i6=j
Aij ≤ 2ǫD.
24
(24)
Let
J = { (i, j) ∈ [D] × [D] | i 6= j, Aij ≤
1
}.
D
It follows from Equation (24) that | { (i, j) ∈ [D] × [D] | i 6= j, (i, j) ∈
/ J } | ≤ 2D 2 ǫ and
so
|J| ≥ (1 − 2ǫ)D 2 − D .
Since for (i, j) ∈ J also (j, i) ∈ J, we can write J = (I × I)\{(i, i)|i ∈ I} for I ⊆ [D].
Thus, we infer for D ≥ 2
1
2
− 2ǫ D 2 .
|J| = |I|(|I| − 1) ≥ ((1 − 2ǫ)D − D) ≥
2
From this it follows that
1
|I|2 ≥ |I|(|I| − 1) ≥ ( − 2ǫ)D 2 ,
2
and we finally obtain
|I| ≥
p
1/2 − 2ǫD.
(25)
′
Notice that it follows from Equation (23) that we may rescale all
the Xi to Xi such that
1
Tr Φ(Xi′ )2 = 1 and the pairwise scalar product still satisfies Tr Φ(Xi′ )Φ(Xj′ ) ≤ D(1−ǫ)
p
for (i, j) ∈ J. If there is an N ∈ N such that d > p 1/2 − 2ǫD for all D ≥ N , the
claim follows. We therefore now suppose that d ≤ 1/2 − 2ǫD. Hence, d ≤ |I| by
Equation (25). By the positivity of Φ and the fact that the Xi′ are positive semidefinite,
we have that Φ(Xi′ ) is positive semidefinite. In [Wol12, Proposition 2.7] it is shownthat
for any set {Pi }i∈I of |I| ≥ d positive semidefinite matrices in Md such that Tr Pi2 = 1
we have that
X
(|I| − d)2 |I|
.
Tr (Pi Pj )2 ≥
(|I| − 1)d2
i6=j
By the definition of the set J, we have that
X
(i,j)∈J
Tr Xi′ Xj′
2
≤
|J|
1
≤
,
2
2
(1 − ǫ) D
(1 − ǫ)2
as |J| ≤ D 2 . From Equation (25) it follows that
1
≥
(1 − ǫ)2
p
!2
1/2 − 2ǫD
−1
d
and after some elementary computations we finally obtain
p
(1 − ǫ) 1/2 − 2ǫ
D.
d≥
2−ǫ
25
C. Lower bound on the value of SDPs through sketching
We will obtain lower bounds on the value of the sketchable SDP in terms of the value
of the sketched SDP through continuity bounds on the relaxed SDP. As the continuity
bound we use is for SDPs given in equality form, we begin by giving an equivalent
formulation of a sketchable SDP with equality constraints. The method of using duality
to derive perturbation bounds on a convex optimization problem used here is standard
and we refer to [BV04, Section 5.6] for a similar derivation. Given a sketchable SDP,
define the maps Φ : MD → Mm
Φ(X) =
m
X
Tr (Bj X) ej eTj
j=1
m
for { ej }m
j=1 an orthonormal basis of R and Ψ : MD+m → Mm
Ψ
X ∗
∗ Z
= Φ(X) + [Zjj ]j .
With the help of the matrix
G=
m
X
γj ej eTj ,
(26)
j=1
the sketchable SDP can be written in equality form as
A 0 X ∗
maximize
Tr
0 0 ∗ Z
X ∗
subject to
Ψ
=G
∗ Z
X ∗
≥ 0,
∗ Z
(27)
where ∗ are submatrices which are not relevant for our discussion. The dual problem of
Equation (27) may be written as
minimize
subject to
m
X
j=1
m
X
j=1
γj Yjj
Bj Yjj ≥ A
Yjj ≥ 0,
26
∀j ∈ [m].
(28)
Lemma C.1. If there is an X > 0 such that the sketchable SDP is satisfied with strict
inequality and the dual problem is feasible, then both the primal problem given in Equation
(27) and dual in Equation (28) are feasible, there is no duality gap and there is a dual
solution which attains the optimal value. Furthermore, this condition is equivalent to
Slater’s condition.
Proof. If we can show equivalence with Slater’s condition, the first statement follows
automatically.
Slater’s condition for the primal problem is the following. If there is a
X ∗
> 0 which satisfies the constraints, then the duality gap is zero and there is a
∗ Z
Y which attains the optimal value. For such a block matrix, we also need X > 0, Z > 0.
Hence, we can formulate Slater’s condition with off-diagonal entries ∗ = 0. We observe
γ1 − Tr (B1 X)
X ∗
..
Ψ
=G⇔
= Z > 0.
.
∗ Z
γm − Tr (Bm X)
Hence X > 0 satisfies the constraints with strict inequality. The converse is clear.
Now we can bound the optimal solution to Equation (11). We denote by A(ǫ) the
feasible set of the relaxed SDP as in Definition 5.4 for some ǫ > 0. With this notation,
A(0) is the feasible set for the primal problem. Analogously, we denote by α(ǫ) the
optimal value of the relaxed problem, by α(0) the optimal value of the sketchable SDP.
Lemma C.2. Assume that there exists X0 ∈ A(0) such that X0 > 0 and the constraints
are strictly satisfied. Then
α(0) ≤ α(ǫ) ≤ α(0) + hǫ̃, y ∗ i,
where y ∗ is an optimal solution to the dual problem to the sketchable SDP and ǫ̃ ∈ Rm
with ǫ̃i = 3ǫηkBi k.
Proof. The first inequality is obvious, since any X ∈ A(0) is also in A(ǫ). By Lemma C.1,
strong duality holds and there is a y ∗ ≥ 0 which achieves the optimal value. Hence
α(0) =
m
X
yj∗ γj
j=1
≥
m
X
j=1
yj∗ γj − Tr
= Tr (AX) −
m
X
i=1
"
m
X
i=1
#
yi∗ Bi − A X
!
yi∗ [Tr (Bi X) − γi ] .
27
,
X≥0
(29)
The first line holds by duality. If we take the supremum over X ∈ A(ǫ), we infer
α(0) ≥ α(ǫ) −
m
X
yi∗ ǫ̃i ,
i=1
since yi∗ ≥ 0.
Corollary C.3. Assume that there exists X0 ∈ A(0) such that X0 > 0 and the constraints are strictly satisfied. Then
α(ǫ) ≤ α(0) + max ǫ̃i (α(0) − Tr (AX0 )) / min (γk − Tr (Bk X0 )) ,
k
i∈[m]
where ǫ̃ ∈ Rm is defined as in Lemma C.2.
Proof. By Equation (29), we have that
α(0) ≥ Tr (AX0 ) −
m
X
i=1
yi∗ [Tr (Bi X0 ) − γi ] .
Since [Tr (Bi X0 ) − γi ] < 0, it follows that
m
X
i=1
yi∗ ≤ (α(0) − Tr (AX0 ))/ min [γi − Tr (Bi X0 )] .
With hǫ̃, y ∗ i ≤ [max ǫi ]
i∈[m]
i∈[m]
Pm
∗
i=1 yi
for y ∗ ≥ 0, the Corollary follows.
Theorem C.4. Assume that there exists X0 ∈ A(0) such that X0 > 0 and the constraints
are strictly satisfied. Then the value of the sketched SDP αS is bounded by
αS ≤ α(0) + ǫC1 (α(0) − Tr (AX0 )) /C2 .
Here
C1 = max { 3ηkBi k1 | i ∈ [m] } ,
C2 = min { (γi − Tr (Bi X0 )) | i ∈ [m] } ,
where η = Tr (X ∗ ) for an optimal point X ∗ of the sketchable SDP.
Proof. The key step is to recognize that Tr SCS T X is equal to Tr CS T XS by the
cyclicity of the trace. Thus, the relaxed SDP gives an upper bound for the sketched
SDP. The theorem then follows by Corollary C.3.
Note that this result is not probabilistic and holds regardless of the sketching matrix
S used.
28
D. Random feasibility problems
In this section, we investigate under which conditions the convex hull of m random
GUE [AGZ10, Section 2.2] matrices shifted by a multiple of the identity both contains
a positive semidefinite matrix and the cone they define is pointed. This is used in
Section 7.2.
Let Gi ∈ Md′ , i ∈ [m], be random matrices sampled independently from the GUE.
This means that Gi is Hermitian and that
(
(Gi )kl = Yk for k = l
,
(Gi )kl =
(Gi )kl = Zkl for k > l
where Yk is a real normal random variable with mean 0 and variance 1 and Zkl is a
complex normal random variable with mean 0 and variance 1. Since we will need similar
matrices with different variance, we will call this distribution GUE(0, 1). Consider the
rescaled and shifted matrices G̃i (α) = √1d′ Gi + α1 for some α ∈ R. We call the convex
hull of these matrices
n
o
Xα := conv
.
G̃i (α) i ∈ [m]
Lemma D.1. Let α ≥
√2 (1
m
+ ǫ). Then
′ 3/2
P Xα ∩ Sd++
6= ∅ ≥ 1 − Ce−2d ǫ /C ,
′
where C is a numerical constant independent of m, d′ , ǫ.
Pm
Proof. Let t ∈ Rm
+ such that
i=1 ti = 1. By the definition of the GUE,
G(t) :=
m
X
ti Gi
i=1
Pn
is again in GUE(0, i=1 t2i ). By [LR10, Theorem 1] and the fact that the GUE is
invariant under unitary transformations, it holds that
i
h
√
′ 3/2
P λmin (G(t)/ d′ ) ≤ −2ktk2 (1 + ǫ) ≤ Ce−2d ǫ /C .
The expression −2ktk2 (1 + ǫ) is maximized by tmin = (1/m, . . . , 1/m), for which we
√
√
obtain the value −2/ m (see [Bha97, Remark II.3.7]). Hence, for α ≥ 2(1 + ǫ)/ m, we
infer that
h
i
h
i
√
√
P λmin (G(tmin )/ d′ + α1) ≤ 0 = P λmin (G(tmin )/ d′ ) ≤ −α
√
2
′
≤ P λmin (G(tmin )/ d ) ≤ − √ (1 + ǫ)
m
′ 3/2 /C
≤ Ce−2d ǫ
29
.
√
√
As λmin (G(tmin )/ d′ + α1) > 0 implies that G(tmin )/ d′ + α1 is positive definite, the
assertions follows.
Lemma D.2. Let α ≤ √2m (1 − ǫ) with ǫ ∈ 0, 12 . Then
P Xα ∩
Sd+′
√ m
1
8 m
′ 3/2 3/2 ′
C 4 e− 4C (8+d ǫ )ǫ d ,
=∅ ≥1−m 1+
ǫ
where C is a numerical constant independent of m, d′ , ǫ.
√
Proof. Take an ǫ/(4 m)-net N for the ℓ1 -sphere S1m−1 in Rm . This means that for all
√
t ∈ S1m−1 , there is an s ∈ N such that kt − sk1 < ǫ/(4 m). It can be shown as in [Ver12,
√
Section 5.2.2] that we can choose N such that |N | ≤ (1 + 8 m/ǫ)m . Now assume that
√
√
λmax (Gi / d′ ) ≤ 2(1 + ǫ) ∧ λmin (Gi / d′ ) ≥ −2(1 + ǫ)
∀i ∈ [m].
This implies that
√
Gi / d′
∞
≤ 2(1 + ǫ). By Weyl’s perturbation theorem [Bha97,
Theorem II.2.6], it follows that for t ∈ S1m−1 ∩ Rm
+.
√
√
√
√
|λmin (G(t)/ d′ ) − λmin (G(s)/ d′ )| ≤ G(t)/ d′ − G(s)/ d′
≤
m
X
i=1
√
|ti − si | Gi / d′
∞
∞
ǫ
≤ 2(1 + ǫ)kt − sk1 ≤ √ .
m
√
√
Now assume further that λmin (G(s)/ d′ + α1) ≤ −ǫ/ m for all s ∈ N . Then clearly
Xα ∩ Sd+′ = ∅ by the above. We thus have to estimate the probability with which our
assumptions are met. Using a union bound, we obtain
P Xα ∩ Sd+′ = ∅
Gi
ǫ
Gi
√
√
√
∀s ∈ N ∧
≤ 2(1 + ǫ) ∀i ∈ [m]
≤−
≥P
λmin α1 +
m
d′
d′ ∞
Y
Gi
ǫ Y
G(s)
2
P λmin √
1−
≥1 −
P λmin √
≥ −√
≤ −2(1 + ǫ)
′
m
2
d′
d
s∈N
i∈[m]
Gi
P λmax √
≥ 2(1 + ǫ) .
d′
30
Using again [LR10, Theorem 1] we infer that for all i ∈ [m]
′ 3/2
Gi
≤ −2(1 + ǫ) ≤ Ce−2d ǫ /C ,
P λmin √
′
d
Gi
′ 3/2
P λmax √
≥ 2(1 + ǫ) ≤ Ce−2d ǫ /C ,
′
d
ǫ
2
ǫ
G(s)
G(s)
≥ −√
≥ −2ksk2 1 −
≤ P λmin √
1−
P λmin √
2
2
m
d′
d′
′2 (ǫ/2)3 /C
≤ C 2 e−2d
.
For the last estimate, we have used that G(s) is again a GUE element with different
variance (cf. proof of Lemma D.1). Combining this with the estimates concerning |N |,
the assertion follows.
We obtain as a corollary that the cone generated by the G̃(α) is pointed with high
probability if m << d′ .
Corollary D.3. Let α ≤ √2m (1−ǫ) for ǫ ∈ 0, 12 and denote by Cα = cone { G̃1 (α), . . . , G̃m (α) }.
Then
√ m
1
8 m
′ 3/2 3/2 ′
P [Cα ∩ −Cα = { 0 }] ≥ 1 − m 1 +
C 4 e− 4C (8+d ǫ )ǫ d ,
ǫ
where C is a numerical constant independent of m, d′ , ǫ.
Proof. We know from Lemma D.2 that we have the same lower bound for the probability
of the event A = { Xα ∩ Sd+′ = ∅ }. But note that the event A implies that the cone is
pointed. That is because if the cone was not pointed, there would exist γ, µ ∈ Rm
+\ { 0 }
such that
m
m
X
X
µi G̃i (α)
γi G̃i (α) = −
i=1
i=1
and so
1
m
P
(γi + µi )
i=1
m
X
(µi + γi )Gi (α) = 0,
i=1
which implies that Xα ∩ Sd+′ 6= ∅.
Theorem D.4. Consider the LMI
m
X
i=1
Let ǫ ∈ 0, 12 . Then for α ≥
√2 (1
m
ti G̃i (α) − 1.
+ ǫ), this LMI is feasible with probability at least
′ 3/2 /C
1 − Ce−2d ǫ
31
.
(30)
Moreover, for α ≤
√2 (1
m
+ ǫ), this LMI is infeasible, the cone generated by the G̃(α) is
pointed and we have Xα ∩ Sd+′ = ∅ with probability at least
√ m
1
8 m
′ 3/2 3/2 ′
1−m 1+
C 4 e− 4C (8+d ǫ )ǫ d .
ǫ
(31)
Proof.
Pm For the first assertion, by Lemma D.1 Xα contains a positive definite element
i=1 ri G̃i (α) with probability lower bounded by the expression in Equation (30). Then
G(µr) is feasible for µ ∈ R+ large enough. For the second assertion, we note that
the LMI being infeasible is equivalent to Xα not containing a positive definite element.
From Lemma D.2 the lower bound in (31) for the probability that the LMI is infeasible.
Moreover, the fact that the same lower bound holds for the probability that the cone is
pointed follows from Corollary D.3.
32
| 8cs.DS
|
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
arXiv:1509.05385v2 [math.RA] 15 Jan 2016
CHRIS FRASER
Abstract. We introduce quasi-homomorphisms of cluster algebras, a flexible notion
of a map between cluster algebras of the same type (but with different coefficients).
The definition is given in terms of seed orbits, the smallest equivalence classes of seeds
on which the mutation rules for non-normalized seeds are unambiguous. We present
examples of quasi-homomorphisms involving familiar cluster algebras, such as cluster
structures on Grassmannians, and those associated with marked surfaces with boundary.
We explore the related notion of a quasi-automorphism, and compare the resulting group
with other groups of symmetries of cluster structures. For cluster algebras from surfaces,
we determine the subgroup of quasi-automorphisms inside the tagged mapping class
group of the surface.
Introduction
The general structural theory of cluster algebras has been well developed during the 15
years since their inception [15]. Despite this, there does not seem to be a consensus on
what the “right” notion of a homomorphism between cluster algebras should be– several
such notions have arisen in different mathematical settings, see e.g. [1, 2, 6, 7, 26, 27].
From our perspective, the key difficulty in defining homomorphisms of cluster algebras
is rooted in the fact that the construction of a cluster algebra involves three operations:
the addition, the multiplication, and the auxiliary addition used in the normalization
condition. Most preexisting notions of a “cluster homomorphism” are designed to respect
all three of these operations, a rather restrictive requirement. We suggest instead that
even in the ordinary (i.e., normalized) setting, it is fruitful to consider maps that only
preserve the structures intrinsic to non-normalized cluster algebras, ignoring the auxiliary
addition. This leads us to the concept of seed orbits and to the mutation patterns these
orbits form. The morphisms between such mutation patterns are the main object of our
interest; we call them quasi-homomorphisms. This paper is devoted to a systematic study
of quasi-homomorphisms and related algebraic constructs.
A cluster algebra is defined by specifying a distinguished set of generators (called cluster
variables) inside an ambient field of rational functions in n variables. Starting from
an initial cluster of n cluster variables, the remaining cluster variables are obtained by
iterating algebraic steps called mutations. Each mutation produces a new cluster from
a current one by exchanging one cluster variable for a new one. The specific rules for
computing the latter are encoded by two additional ingredients, an n×n exchange matrix B
2010 Mathematics Subject Classification. 13F60.
Key words and phrases. Cluster algebra, seed orbit, quasi-homomorphism, cluster modular group,
tagged mapping class group.
This work was supported by a graduate fellowship from the National Physical Science Consortium and
NSF grant DMS-1361789.
1
2
CHRIS FRASER
and a coefficient tuple p consisting of elements of some fixed coefficient group. The triple
consisting of the cluster, the exchange matrix, and the coefficient tuple, is called a seed.
When a cluster mutates, the ingredients B and p also do: the new matrix B ′ is given
explicitly in terms of B, and the new tuple p′ satisfies a constraint involving p and B.
A collection of seeds related to each other by mutations in all possible directions forms a
seed pattern.
In the most general cluster algebra setup – that of non-normalized seed patterns [3, 13,
15], the mutation recipe does not uniquely specify the new coefficient tuple p′ from p
and B. This ambiguity propagates through iterated mutations, and consequently the set
of cluster variables is not uniquely determined by the initial seed.
The usual way to remove this ambiguity is to impose the additional assumption that the
coefficient group is endowed with an additional operation of “auxiliary addition” (making
it into a semifield ), and then require the corresponding normalization condition to hold
at every seed. This assumption is satisfied for the most important examples of cluster
algebras arising in representation theory. In this paper, we make use of another way
of removing the ambiguity by considering seed orbits, the smallest equivalence classes of
seeds on which the mutation rules are unambiguous. This gives rise to to the concept of a
mutation pattern of seed orbits. Such a pattern is determined uniquely by any one of its
constituent seed orbits. The natural notion of a homomorphism between two mutation
patterns of seed orbits brings us to the definition of a quasi-homomorphism, a rational
map (more precisely, a semifield homomorphism) that respects the seed orbit structure
and commutes with mutations.
Though the appropriate context for defining quasi-homomorphisms is that of nonnormalized seed patterns, we see two ways in which quasi-homomorphisms are useful
in the structural theory of ordinary (normalized) seed patterns. First, it is important to
understand the relationships between cluster algebras with the same underlying pattern
of exchange matrices but with different choices of coefficients. One celebrated result of
this kind is the separation of additions formula ([17, Theorem 3.7]). For a given mutation
pattern of exchange matrices, this formula expresses the cluster variables in a cluster algebra with any choice of coefficients in terms of those in a cluster algebra with a special
choice of principal coefficients. Our Proposition 4.3 puts this formula in a wider context,
in which every quasi-homomorphism between a pair of normalized seed patterns witnesses
its own separation of additions. This idea can be used to construct a new cluster algebra starting from a known one. (More precisely, using a known cluster structure on an
algebra R, one can produce a cluster structure on another algebra R′ by describing an
appropriate map from R to R′ .)
Second, the naturally defined concept of a quasi-automorphism gives rise to the quasiautomorphism group of a seed pattern. This group interpolates between previously defined
groups that are either too sensitive to coefficients (these groups are too small) or don’t
refer to coefficients at all (these groups are too large). Most cluster algebras arising in
applications have nontrivial coefficients, and these cluster algebras often afford nontrivial
self-maps that are quasi-automorphisms of the cluster structure. The twist map on the
Grassmannian [24] is one important example, cf. Remark 6.5. In a forthcoming companion
paper we construct a large group of quasi-automorphisms of the Grassmannian cluster
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
3
algebras [28] whose action on cluster variables has a simple description. Much of the
abstract setup in this paper was developed with that application in mind.
The paper is organized as follows. Section 1 presents background on non-normalized
seed patterns. This is mostly standard and taken from [13, 17], but with emphasis on the
notion of the ambient semifield, cf. Definition 1.4. The section ends with a motivating
example: a pair of seed patterns which will illustrate the various notions in subsequent
sections. A reader familiar with cluster algebras can skim this section and head directly
to Example 1.8.
Sections 2 and 3 are the conceptual core of the paper. We define seed orbits as the
smallest equivalence classes of non-normalized seeds on which the mutation rule is unambiguous. In Proposition 2.3 we give a more explicit characterization of seed orbits as orbits
with respect to a rescaling action on seeds. Section 3 introduces quasi-homomorphisms
of seed patterns and their basic properties. We end this section by describing the key differences between quasi-homomorphisms and some preexisting notions, specifically rooted
cluster morphisms [1] and coefficient specializations [16, 26, 27].
In Section 4 we discuss quasi-homomorphisms between normalized seed patterns. For
seed patterns of geometric type, we relate quasi-homomorphisms to linear combinations
of the rows of an extended exchange matrix, making connections to the separation of
additions formula and to gradings on cluster algebras. Section 5 introduces the easiest
way of specifying a quasi-homomorphism in practice, by checking that a given semifield
map sends cluster variables to rescaled cluster variables on a nerve. In Section 6 we define
the quasi-automorphism group of a seed pattern and compare it with the cluster modular
group [9] and the group of cluster automorphisms [2].
Sections 7 and 8 focus on cluster algebras associated with bordered marked surfaces [12,
13]. The main result is Theorem 7.5 describing the quasi-automorphism group of such a
cluster algebra as a subgroup of the tagged mapping class group (excluding a few exceptional surfaces). In particular, it establishes that regardless of the choice of coefficients in
such a cluster algebra, the quasi-automorphism group is always a finite index subgroup
of the cluster modular group.
The concept of a nerve introduced in Section 5 is new and includes as a special case
the star neighborhood of a vertex. Star neighborhoods show up in the algebraic Hartogs’ principle argument used to establish that a given cluster algebra is contained in
another algebra [11, Proposition 3.6]. In Appendix 9 we extend this argument from a star
neighborhood to an arbitrary nerve.
Section 10 illustrates the techniques in Section 4. We generalize Example 1.8 by describing a quasi-isomorphism between the Grassmannian cluster algebras [28] and polynomial
rings arising as coordinate rings of band matrices.
Acknowledgements
I would like to thank Ian Le, Greg Muller, and Gregg Musiker for helpful conversations.
I especially thank Sergey Fomin for many conversations and suggestions. This work was
supported by a graduate fellowship from the National Physical Science Consortium and
NSF grant DMS-1361789. While in the midst of carrying out this work, I learned that
Thomas Lam and David Speyer had independently obtained results similar to Corollary
4.5 and Remark 4.6.
4
CHRIS FRASER
1. Preliminaries on seed patterns
A (non-normalized) cluster algebra is constructed from a set of data called a nonnormalized seed pattern. We define this data now while fixing standard notation. For a
number x we let [x]+ ∶= max(x, 0). We let sign(x) equal either −1, 0 or 1 according to
whether x is negative, zero, or positive. We denote {1, . . . , n} by [1, n].
The setup begins with a choice of ambient field of rational functions F with coefficients
in a coefficient group P. The coefficient group is an abelian multiplicative group without
torsion. The ambient field is a field of rational functions in n variables with coefficients
in P: it is the set of expressions that can be made out of n elements x1 , . . . , xn and the
elements of of P, using the standard arithmetic operations +, −, × and ÷, under the usual
notion of equivalence of such rational expressions. The integer n is called the rank.
Definition 1.1 (Non-normalized seed, [15, 13]). Let P and F be as above. A nonnormalized seed in F is a triple Σ = (B, p, x), consisting of the following three ingredients:
● a skew-symmetrizable n × n matrix B = (bij ),
● a coefficient tuple p = (p±1 , . . . , p±n ) consisting of 2n elements in P,
● a cluster x = (x1 , . . . , xn ) in F , whose elements (called cluster variables) are algebraically independent and freely generate F over QP.
The more restrictive notion of normalized seed is given in Definition 4.1. Normalized
seeds are much more studied in the literature, where they are usually simply called seeds.
Thus, we persistently use the adjective non-normalized in our setting, although this is a
little clumsy.
Definition 1.2. A labeled n-regular tree, Tn , is an n-regular tree with edges labeled by
k
integers so that the set of labels emanating from each vertex is [1, n]. We write t Ð
→ t′ to
indicate that vertices t, t′ are joined by an edge with label k. An isomorphism Tn → Tn of
labeled trees Tn and Tn sends vertices to vertices and edges to edges, preserving incidences
of edges and the edge labels. Such an isomorhism is uniquely determined by its value at
a single vertex t ∈ Tn .
Definition 1.3 (Non-normalized seed pattern, [13, 15]). Let P and F be as above. A
collection of non-normalized seeds in F , with one seed Σ(t) = (B(t), p(t), x(t)) for each t ∈
k
Tn , is called a non-normalized seed pattern if for each edge t Ð
→ t′ , the seeds Σ(t) and
Σ(t′ ) are related by a mutation in direction k:
● The matrices B(t) and B(t′ ) are related by a matrix mutation
(1.1)
(1.2)
−bij (t)
bij (t′ ) = {
bij (t) + sign(bik (t))[bik (t)bkj (t)]+
● the coefficient tuples p(t) and p(t′ ) are related by
(1.3)
when j ≠ k,
p±k (t′ ) = p∓k (t) and
p+
⎧
j (t) +
bkj(t)
⎪
⎪
⎪ p−j (t pk (t)
= ⎨ p+j (t)
p−j (t′ ) ⎪
⎪
p− (t)bkj(t)
−
⎪
⎩ pj (t) k
p+j (t′ )
if i = k or j = k
otherwise,
if bkj ≥ 0
if bkj ≤ 0
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
(1.4)
5
● and the clusters x(t) and x(t′ ) are related by
xj (t′ ) = xj (t) for j ≠ k, and
xk (t)xk (t′ ) = p+k ∏ xj (t)[bjk ]+ + p−k ∏ xj (t)[−bjk ]+ ,
(1.5)
the latter of which is called an exchange relation.
The rules (1.1) through (1.5) are ambiguous, meaning Σ(t′ ) is not determined uniquely
p+ (t′ )
from Σ(t). Indeed, since (1.3) only mentions the ratio p−j (t′ ) , for each j ≠ k one can
j
rescale both of p+j (t′ ) and p−j (t′ ) by a common element of P while preserving (1.3). We
µk
write Σ ↭ Σ′ to indicate that two seeds Σ and Σ are related by a mutation in direction k;
this condition is symmetric in Σ and Σ′ .
Thinking of (1.5) as a recipe for computing xk (t′ ) from Σ(t), we crucially observe that
the computation is subtraction-free: the only operations needed are +, × and ÷ in F . This
motivates the following definition:
Definition 1.4 (Ambient semifield). Let E be a non-normalized seed pattern, and x(t)
one of its clusters. The ambient semifield, F>0 = F>0 (E) ⊂ F is the subset of all elements
which can be given as a subtraction-free rational expression in the elements of x(t),
with coefficients in P. Thus, it is the set of rational functions which can be built out
of x1 (t), . . . , xn (t) and the elements of P using the operations +, × and ÷ in F .
Since (1.5) is subtraction-free, F>0 is independent of the choice of t (it only depends
on E), and every cluster variable for E lies in F>0 . Recall that a semifield is an abelian
multiplicative group, with an additional binary operation (called the auxiliary addition)
that is commutative and associative, and distributes over multiplication. The ambient
semifield is a semifield with respect to the multiplication and addition operations in F ,
justifying its name. Homomorphisms between semifields are defined in the obvious way.
The ambient semifield has the following universality property.
Lemma 1.5 ([17, Definition 2.1]). Let E be a non-normalized seed pattern with coefficient
group P and ambient semifield F>0 . Fix a cluster x(t) in E. Let S be any semifield. Then
given a multiplicative group homomorphism P → S, and a function x(t) → S, there exists
a unique semifield homomorphism F>0 → S agreeing with the given maps on P ∪ x(t).
The following elements of F>0 will play a prominent role in Section 2.
Definition 1.6 (Hatted variables). Let E be a non-normalized seed pattern. Let ŷ(t) =
(ŷ1 (t), . . . , ŷn (t)) denote the n-tuple of hatted variables
(1.6)
ŷj (t) =
p+j (t)
p−j (t)
∏ xi (t)bij (t) ,
i
obtained by taking the ratio of the two terms on the right hand side of (1.5).
The hatted variables in adjacent seeds determine each other as follows:
Proposition 1.7 ([13, Proposition 2.9]). Let E = (B(t), p(t), x(t)) be a non-normalized
k
seed pattern with hatted variables ŷ(t). For each edge t Ð
→ t′ , the n-tuples ŷ(t) and ŷ(t′ )
6
CHRIS FRASER
satisfy
(1.7)
ŷj (t)−1
ŷj (t′ ) = {
ŷj (t)ŷk (t)[bkj (t)]+ (ŷk (t) + 1)−bkj (t)
if j = k
if j ≠ k.
The propagation rule (1.7) takes place in F>0 , and only depends on the B matrix.
The preceding discussion is what we will need for Section 2. We briefly recall a few
more definitions which will be useful in presenting our examples. First, the exchange
graph E associated with a seed pattern E is the graph whose vertices are the unlabeled
seeds in E, and whose edges correspond to mutations between these seeds. More precisely,
permuting the indices [1, n] in a non-normalized seed commutes with the mutation rules
(1.1) through (1.5). The exchange graph is the n-regular graph obtained by identifying
vertices t1 , t2 ∈ Tn if the seeds Σ(t1 ) and Σ(t2 ) are permutations of each other. The star
neighborhood star(t) of a vertex t ∈ E is the set of n edges adjacent to it. Rather than
being indexed by [1, n], the data in an unlabeled seed Σ(t) for t ∈ E is indexed by the n
seeds adjacent to Σ(t), i.e. by the elements of star(t).
Second, in the concrete examples in this paper, we have chosen a distinguished finite
set of elements called frozen variables, and the coefficient group P is the free abelian
multiplicative group of Laurent monomials in these frozen variables. The cluster algebra A
associated with the seed pattern E is the Z-algebra generated by the frozen variables and
all of the cluster variables arising in the seeds of E.
Example 1.8. We now introduce a pair of affine algebraic varieties X and Y and a
pair of seed patterns in their respective fields of rational functions. The cluster algebras
associated with these seed patterns are the coordinate rings C[X] and C[Y]. Both cluster
algebras are of finite Dynkin type A2 .
Let X = Ĝr(3, 5) be the affine cone over the Grassmann manifold of 3-dimensional planes
in C5 . The points in X are the decomposable tensors {x ∧ y ∧ z∶ x, y, z ∈ C5 } ⊂ Λ3 (C5 ).
Its coordinate ring is generated by the Plücker coordinates ∆ijk for 1 ≤ i < j < k ≤ 5,
extracting the coefficient of ei ∧ ej ∧ ek in x ∧ y ∧ z, where e1 , . . . , e5 is the standard basis
for C5 . Representing a given x ∈ Ĝr(3, 5) by a 3 × 5 matrix, ∆ijk (x) is the maximal minor
of this matrix in columns i, j, and k.
There is a well known cluster structure on C[X] [15, 16]. It is a special case of a cluster
structure for arbitrary Grassmannians constructed by Scott [28]. The frozen variables are
the Plücker coordinates consisting of cyclically consecutive columns
(1.8)
∆123 , ∆234 , ∆345 , ∆145 , ∆125 .
There are five cluster variables, listed in (1.9) with cyclically adjacent pairs of cluster
variables forming clusters
(1.9)
∆245 , ∆235 , ∆135 , ∆134 , ∆124 .
The clusters and exchange relations are given in Figure 1. All of the other data in the seed
pattern can be determined from these. For example, focusing on the seed whose cluster
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
7
is (x1 , x2 ) = (∆235 , ∆245 ), from the first and fifth exchange relations in Figure 1 follows
(p+1 , p−1 , p+2 , p−2 ) = (∆125 ∆234 , ∆123 , ∆145 , ∆345 ∆125 )
∆125 ∆234 ∆145 ∆235
(ŷ1 , ŷ2 ) = (
(1.11)
,
).
∆123 ∆245 ∆345 ∆125
The exchange relations are written so that mutating is moving clockwise in the exchange
graph. If a mutation moves counterclockwise, one should swap the order of the two terms
in the exchange relation.
(1.10)
(∆235 , ∆245 )
●
(∆245 , ∆124 )
(∆135 , ∆235 )
●
●
●
(∆124 , ∆134 )
●
(∆134 , ∆135 )
∆245 ∆135 = ∆145 ∆235 + ∆125 ∆345
∆235 ∆134 = ∆234 ∆135 + ∆123 ∆345
∆135 ∆124 = ∆125 ∆134 + ∆123 ∆145
∆134 ∆245 = ∆345 ∆124 + ∆123 ∆345
∆124 ∆235 = ∆123 ∆245 + ∆125 ∆234
Figure 1. The exchange graph for C[X]. The vertices are clusters and
edges between vertices are mutations. Each mutation exchanges two cluster
variables via an exchange relation listed in the table at top right. The extra
data in each seed can be inferred from these exchange relations.
Second, let Y ≅ C9 the affine space of band matrices of the form
(1.12)
0 ⎞
⎛y1,1 y1,2 y1,3 0
y = ⎜ 0 y2,2 y2,3 y2,4 0 ⎟ .
⎝ 0
0 y3,3 y3,4 y3,5 ⎠
Its coordinate ring C[Y] contains the minors YI,J . Evaluating YI,J on y ∈ Y returns the
minor of y occupying rows I and columns J, e.g. Yi,j (y) = yi,j and Y12,23 (y) = y1,2y2,3 −
y1,3 y2,2 . Some of these minors factor, e.g. Y12,13 = Y1,1 Y2,3 .
Figure 2 shows a seed pattern whose cluster algebra is C[Y]. The frozen variables are
the following minors
(1.13)
Y1,1 , Y2,2 , Y3,3 , Y1,3 , Y2,4 , Y3,5 , Y123,234 .
The cluster variables are listed in (1.14), with cyclically adjacent pairs forming clusters
(1.14)
Y1,2 , Y12,23 , Y2,3 , Y23,34 , Y3,4 .
2. Seed orbits
We introduce seed orbits by first describing them as equivalence classes under a certain
equivalence relation on seeds. Proposition 2.3 gives another characterization as orbits
under an explicit rescaling action.
8
CHRIS FRASER
(Y1,2 , Y3,4 )
●
(Y12,23 , Y1,2 )
●
●
(Y3,4 , Y23,34 )
(Y2,3 , Y12,23 )
●
Y1,2 Y2,3 = Y12,23 + Y2,2 Y1,3
Y12,23 Y23,34 = Y123,234 Y2,3 + Y2,2 Y3,3 Y1,3 Y2,4
●
(Y23,34 , Y2,3 )
Y2,3 Y3,4 = Y23,34 + Y3,3 Y2,4
Y23,34 Y1,2 = Y2,2Y1,3 Y3,4 + Y123,234
Y3,4 Y12,23 = Y3,3Y2,4 Y1,2 + Y123,234
Figure 2. The exchange graph and exchange relations for C[Y], mirroring
Figure 1.
Definition 2.1. Let k⃗ = (k1 , . . . , kℓ ) be a sequence of elements of [1, n]. Choosing a base
kℓ
k1
k2
point t0 ∈ Tn , such a sequence determines a walk t0 Ð→ t1 Ð→ ⋯ Ð→ tℓ in Tn . We say that k⃗
is contractible if this walk starts and ends at the same vertex of Tn , i.e. tℓ = t0 .
Given non-normalized seeds Σ and Σ∗ , we write Σ ∼ Σ∗ if there is a contractible
sequence of mutations from Σ to Σ∗ , i.e. a contractible sequence k⃗ and non-normalized
seeds Σ1 , . . . , Σℓ−1 such that
(2.1)
µk1
µk2
µkℓ
Σ = Σ0 ↭ Σ1 ↭ Σ2 ⋯Σℓ−1 ↭ Σℓ = Σ∗ .
Clearly, ∼ is an equivalence relation on non-normalized seeds. Furthermore, it removes
the ambiguity present in mutation of non-normalized seeds:
µk
Lemma 2.2. The mutation rule ↭ becomes unambiguous and involutive once it is thought
of as a rule on equivalence classes of seeds under ∼. That is, fixing a ∼-equivalence class
S and a direction k ∈ [n], the set of seeds
(2.2)
{Σ′ ∶ Σ′ ↭ Σ for some Σ ∈ S}
µk
is again a ∼-equivalence class of seeds.
We now characterize ∼-equivalence classes explicitly. We say two elements z, x ∈ F are
proportional, written z ≍ x, if xz ∈ P. We emphasize that P does not include constants,
e.g. −1, 2 ∉ P, and thus x is not proportional to −x, 2x, etc.
Proposition 2.3 (Seed orbits). Let Σ = (B, p, x), Σ∗ = (B ∗ , p∗ , x∗ ) be non-normalized
seeds in F , of rank n ≥ 2, with x = (xi ), p = (p±i ), x∗ = (x∗i ), p∗ = ((p∗ )±i ). Then the
following are equivalent:
(1) Σ ∼ Σ∗ .
(2) B = B ∗ , ŷ(Σ) = ŷ(Σ∗ ), and xi ≍ x∗i for all i.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
9
(3) B = B ∗ , and there exist scalars c1 , . . . , cn , d1 , . . . , dn ∈ P, such that
xj
(2.3)
x∗j =
cj
p±j
[±bij ]+
∗ ±
(p )j =
(2.4)
.
∏ ci
dj
Equations (2.3) and (2.4) define a rescaling action of Pn × Pn on non-normalized seeds,
⃗ ⋅ Σ where (⃗
⃗ ∈ Pn × Pn and Σ is a non-normalized seed. Proposition
denoted by (⃗
c, d)
c, d)
2.3 says that a ∼-equivalence class of non-normalized seeds is precisely a Pn × Pn orbit
under this action; we henceforth refer to these equivalence classes as seed orbits.
Proof. Conditions (2) and (3) are a re-translation of each other by immediate calculation.
We show (1) implies (2). Defining a seed orbit by (2), this implication follows from the
⃗ are
fact that seed orbits are “closed under mutation.” More precisely, if Σ and Σ† = (⃗
c, d)⋅Σ
µk
µk
in the same seed orbit and Σ′ and (Σ† )′ are two seeds satisfying Σ ↭ Σ′ and Σ† ↭ (Σ† )′ ,
then Σ′ and (Σ† )′ are in the same seed orbit. By (1.1) and Proposition 1.7, we know that
B ′ = (B † )′ and ŷ′ = (ŷ† )′ , so the claim will follow if we check (x† )′j ≍ x′j for all j. This is
obvious when j ≠ k from (1.4). When j = k, (1.5) for the mutation Σ† ↭ (Σ† )′ says that
µk
(2.5)
(2.6)
(2.7)
(2.8)
(x† )′k = (x†k )−1 ((p† )+k ∏(x†j )[bjk ]+ + (p† )−k ∏(x†j )[−bjk ]+ )
= (x†k )−1 (p† )−k (1 + ŷk (Σ† )) ∏(x†j )[−bjk ]+
ck
[−b ]
= (xk )−1 p−k (1 + ŷk (Σ)) ∏ xj jk +
dk
ck
= x′k ,
dk
as desired. Returning to the implication (1) ⇒ (2), from the symmetry of ↭ it follows
⃗ Since seed orbits are closed
that Σ is related to itself along any contractible sequence k.
∗
under mutation, any seed Σ related to Σ by a contractible sequence of mutations is
therefore in the same seed orbit as Σ.
Now we show (3) implies (1). Let ĉj (a) ∈ Pn × Pn denote the vector with cj = a and
all other entries equal to 1, and define similarly dˆj (a). Clearly, it suffices to show that
Σ ∼ ĉj (a) ⋅ Σ and Σ ∼ dˆj (a) ⋅ Σ, since rescalings of this type generate Pn × Pn .
Seeds of the form dˆj (a) ⋅ Σ are equivalent to Σ, as follows by mutating twice in any
µj
direction k ≠ j. For seeds of the form ĉj (a) ⋅ Σ, let Σ′ be any seed satisfying Σ ↭ Σ′ :
µk
(2.9)
(2.10)
(2.11)
(2.12)
Σ ∼ dˆj (a−1 ) ⋅ Σ
µj
↭ (ĉj (a−1 )dˆj (a−1 )) ⋅ Σ′
∼ (ĉj (a−1 )) ⋅ Σ′
↭ ĉj (a) ⋅ Σ,
µj
where (2.10) and (2.12) follow from the calculation in (2.8), and (2.9) and (2.11) are
admissible since we already know rescaling by dˆj (a) preserves equivalence of seeds. Since
10
CHRIS FRASER
(2.9) through (2.12) amounts to mutating in direction j twice on seed equivalence classes,
it follows that Σ ∼ ĉj (a) ⋅ Σ as desired.
3. Quasi-homomorphisms
We will now give the definition of a quasi-homomorphisms from a seed pattern E to
another seed pattern E. We retain the notation of Section 1 for all the data in E, and we
use bars to denote the analogous quantities in the second pattern E. Thus E has coefficient
group P, ambient field F, seeds Σ(t) = (B(t), p(t), x(t)), hatted variables ŷ j (t), and so
on. It is built on a second copy of the n-regular tree, Tn .
The motivating observation is the following: since the mutation rules (1.3) through
(1.5) are certain algebraic relations in in F>0 , they are preserved by a homomorphism of
semifields.
Definition 3.1 (Quasi-homomorphism). Let E and E be non-normalized seed patterns.
Let Ψ∶ F>0 → F >0 be a semifield homomorphism satisfying Ψ(P) ⊂ P (in this case we say Ψ
preserves coefficients). We say Ψ is a quasi-homomorphism from E to E if it maps each
seed in E to a seed that is ∼-equivalent to a seed in E, in a way that is compatible with
mutation. More precisely, let t ↦ t be an isomorphism of the labeled trees Tn and Tn .
Then Ψ is a quasi-homomorphism if and only if
(3.1)
Ψ(Σ(t)) ∼ Σ(t)
for all t ∈ Tn , where Ψ(Σ(t)) = (B(t), Ψ(p), Ψ(x)) is the triple obtained by evaluating Ψ
on Σ(t).
As motivation for this definition, we imagine a situation where E is well understood
combinatorially, and we would like to understand another seed pattern E by comparing it
with E. The requirement (3.1) says that the seeds Ψ(Σ(t)) mutate “in parallel” with the
seeds in E, in the sense that their corresponding seeds only differ by the rescalings (2.3)
and (2.4).
The following Propositions 3.2 and 3.3 show two ways in which quasi-homomorphisms
are well-behaved. Both of their proofs follow immediately from the observation that
applying a semifield homomorphism commutes with mutation.
Proposition 3.2. Let Ψ∶ F>0 → F >0 be a semifield homomorphism satisfying (3.1) for
some t ∈ Tn . Then Ψ is a quasi-homomorphism.
That is, rather than checking that (3.1) holds at every t ∈ Tn , it suffices to check this
at a single t ∈ Tn .
Proposition 3.3. Let Ψ be a quasi-homomorphism from E to E. Let Σ be a seed in E,
and let Σ∗ be a non-normalized seed satisfying Σ ∼ Σ∗ . Then Ψ(Σ) ∼ Ψ(Σ∗ ).
Proposition 3.3 says that quasi-homomorphism preserves ∼-equivalence of seeds. Thus,
if S(t) denotes the seed orbit of Σ(t) and ditto for S(t) and Σ(t), then Ψ maps S(t) inside S(t) for all t. A quasi-homomorphism is therefore a natural notion of homomorphism
between the respective seed orbit patterns (t, S(t)) and (t, S(t)).
Now we describe a quasi-homomorphism between the pair of seed patterns in Example 1.8.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
11
Example 3.4. Given Y ∈ Y, let Y [1], Y [2], Y [3] ∈ C5 denote its rows. There is a
F
surjective map of varieties F ∶ Y → X sending Y ↦ Y [1] ∧ Y [2] ∧ Y [3]. It determines a
map on cluster algebras F ∗ ∶ C[X] → C[Y] sending ∆ijk ↦ Y123,ijk . Figure 3 shows the
non-normalized seed pattern that arises from applying F ∗ to Figure 1 and factoring the
cluster variables inside C[Y]. The seeds in Figure 3 are in the same seed orbit as the
corresponding seeds in Figure 2, and thus F ∗ is a quasi-homomorphism from C[X] to
C[Y].
(Y3,5 Y12,23 , Y2,4 Y3,5 Y1,2 )
●
(Y2,4 Y3,5 Y1,2 , Y1,1 Y2,2 Y3,4)
●
(Y1,1 Y2,2 Y3,4 , Y1,1 Y23,34 ) ●
(Y1,1 Y3,5 Y2,3 , Y3,5 Y12,23 )
●
● (Y1,1 Y23,34 , Y1,1 Y3,5 Y2,3 )
2
Y2,4 (Y12,23 + Y2,2 Y1,3 )
Y2,4 Y3,5 Y1,2 ⋅ Y1,1 Y3,5 Y2,3 = Y1,1 Y3,5
Y3.5 Y12,23 ⋅ Y1,1 Y23,34 = Y1,1 Y3,5(Y123,234 Y2,3 + Y2,2 Y3,3 Y1,3 Y2,4)
2
Y2,2 Y35 (Y23,34 + Y3,3 Y24 )
Y1.1 Y3.5 Y2,3 ⋅ Y1.1 Y2.2 Y3,4 = Y1.,1
Y1.1 Y23,34 ⋅ Y24 Y3.5 Y1,2 = Y1.1 Y24 Y3,5 (Y2,2 Y1,3 Y3,4 + Y123,234 )
Y1.1 Y2.2Y3,4 ⋅ Y35 Y12,23 = Y1.1 Y22 Y3,5 (Y3,3 Y2,4 Y1,2 + Y123,234 )
Figure 3. The non-normalized seed pattern obtained by applying F ∗
to the seed pattern in Figure 1. The clusters agree with the clusters in
Figure 2 up to the frozen variables listed in (1.13). Cancelling the common
frozen variable factors from both sides of the exchange relations yields the
exchange relations in Figure 2. It follows that the ŷ values are the same in
both figures.
Definition 3.5. Two quasi-homomorphisms Ψ1 , Ψ2 from E to E are called proportional
if Ψ1 (Σ) ∼ Ψ2 (Σ) for all seeds Σ in E. We say a quasi-homomorphism Ψ from E to E is
a quasi-isomorphism if there is a quasi-homomorphism Φ from E to E such that Φ ○ Ψ is
proportional to the identity map on F>0 . We say that Ψ and Φ are quasi-inverses of one
another.
Once we have a quasi-isomorphism between two seed patterns, we think of them as
being essentially “the same.” Up to coefficients, the maps in both directions allows us
to write the cluster variables in one seed pattern in terms of the cluster variables in the
other one.
12
CHRIS FRASER
Remark 3.6. The set of seed patterns with quasi-homomorphisms as morphisms is a
category. Proportionality is an equivalence relation on the morphisms in this category,
and this equivalence relation respects composition of quasi-homomorphisms. This yields
a quotient category whose objects are seed patterns and whose morphisms are proportionality classes of quasi-homomorphisms. A morphism in this quotient category is an
isomorphism if and only if one (hence any) of its constituent quasi-homomorphisms is a
quasi-isomorphism.
The following lemma provides a simple method for checking a candidate map is a
quasi-inverse of a given quasi-homomorphism.
Lemma 3.7. Let Ψ be a quasi-homomorphism from E to E, and Φ∶ F >0 → F>0 a semifield
map that preserves coefficients and for which Φ ○ Ψ(x) ≍ x for all cluster variables x in E.
Then Ψ and ϕ are quasi-inverse quasi-isomorphisms.
Lemma 3.7 follows from the more general Proposition 5.2 below. In fact, it will suffice
to merely check that ϕ ○ Ψ(x) ≍ x for all x lying on a nerve (cf. Definition 5.1).
Example 3.8. Using Lemma 3.7 we describe a quasi-inverse G∗ for the quasi- homomorphism F ∗ from Example 3.4. Let G∶ X → Y be the morphism sending X ∈ X to the the
band matrix
0
0
⎞
⎛∆145 (X) ∆245 (X) ∆345 (X)
0
∆125 (X) ∆135 (X) ∆145 (X)
0
⎟∈Y
G(X) = ⎜
⎝
0
0
∆123 (X) ∆124 (X) ∆125 (X)⎠
all of whose entries are Plücker coordinates of X. The coordinate ring map C[Y] → C[X]
sends Yi,j to the Plücker coordinate in the (i, j) entry of G(X), e.g. G∗ (Y1,2 ) = ∆245 .
The matrix G(X) has an interesting property: all of its minors are monomials in the
Plücker coordinates of X. In particular, its maximal minors agree with those of X, up to
a multiplicative factor:
(3.2)
∆ijk (G(X)) = ∆145 (X)∆125 (X)∆ijk (X).
Thus, G∗ ○ F ∗ (∆ijk ) = ∆145 ∆125 ∆ijk ≍ ∆ijk for each cluster variable ∆ijk . Since G∗
preserves coefficients (the only nontrivial check is G∗ (Y123,234 ) = ∆125 ∆145 ∆234 ), from
Lemma 3.7 it follows that G∗ is a quasi-inverse of F ∗ .
Remark 3.9. A quasi-homomorphism is defined as a map on ambient semifields since
these maps transparently preserve the mutation rules (1.1)–(1.5). This should be suitable
for most purposes, since one is mostly interested in evaluating a quasi-homomorphism on
cluster variables or coefficients. However, the cluster algebra A is the more familiar algebraic object associated to a seed pattern. If one wants to think of a quasi-homomorphism
Ψ as an algebra map of cluster algebras A → A, one will sometimes need to first localize
at the frozen variables in A.
We close this section by explaining the differences between quasi-homomorphisms and
preexisting notions of homomorphisms between cluster algebras. Specifically, we consider
the notion of a rooted cluster morphism in the category of cluster algebras described
by Assem, Dupont, and Shiffler [1], and also of a coefficient specialization defined by
Fomin and Zelevinsky [17] and studied by Reading [26, 27]. The key difference between
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
13
these notions and quasi-homomorphisms is that a quasi-homomorphism allows for cluster
variables to be rescaled by an element of P. This extra flexibility provides more freedom in
constructing new cluster algebras from old ones (cf. Section 4) or in finding nice self-maps
of cluster algebras giving rise to elements of the cluster modular group (cf. Section 6).
In a little more detail, a coefficient specialization is a map whose underlying map on
coefficients can be any group homomorphism P → P, but that must send each cluster variable to a cluster variable. Thus, each coefficient specialization is a quasi-homomorphism,
but a very special one since cluster variables are not allowed to be rescaled by elements
of P. Rooted cluster morphisms require choosing a pair of initial seeds in E and E (this
is the sense in which the morphism is rooted). Between this pair of seeds, a morphism is
an algebra map that sends each cluster variable to either a cluster variable or an integer,
and sends each frozen variable to either a frozen variable, a cluster variable, or an integer.
Hence, while quasi-homomorphisms are more flexible in allowing for cluster variables to
be rescaled and for frozen variables to be sent to monomials in the frozen variables, they
are also less flexible as they do not allow for unfreezing frozen variables or specializing
variables to integers. It probably would not be hard to combine these two notions.
One more technicality: to streamline the discussion, we have formulated Definition 3.1
so that quasi-homomorphisms preserve exchange matrices, whereas rooted cluster morphisms allow for B ↦ −B. To make Definition 3.1 more consonant with these preexisting
notions, one could modify (3.1) to say that either Ψ(Σ(t)) ∼ Σ(t) or Ψ(Σ(t)) ∼ Σ(t)opp
(see the definition of opposite seed in Section 5 below) without any significant changes.
4. Normalized seed patterns
In this section, we recall the definition of normalized seed patterns and apply the results
of Section 2 in the case that E and E are normalized.
Definition 4.1 (Normalized seed pattern). A seed pattern E as in Definition 1.3 is called
normalized if the coefficient group P is a semifield, and each coefficient tuple p(t) satisfies
p+j (t) ⊕ p−j (t) = 1 for all j,
(4.1)
where ⊕ is the addition in P.
The advantage of this normalization condition is that it makes the mutation rule (1.3),
and therefore mutation of normalized seeds, unambiguous. Indeed, (1.3) specifies the
p+ (t′ )
ratio yj (t′ ) = pj− (t′ ) in terms of B(t) and p(t), and there is a unique choice of p±j (t′ ) ∈ P
j
with this ratio and satisfying the normalization condition, namely the pair
(4.2)
p+j (t′ ) =
yj (t′ )
1
and p−j (t′ ) =
.
′
1 ⊕ yj (t )
1 ⊕ yj (t′ )
Furthermore, mutating twice in a given direction is the identity.
At the same time, the disadvantage is that computing a cluster algebra now involves
three operations, the two operations present in F>0 along with ⊕ in P. The definition of
quasi-homomorphism prioritizes these first two operations. Proposition 4.3 says that in
the case of a quasi-homomorphism between normalized seed patterns, there is a “separation of additions” phenomenon, separating the addition in F>0 from the one in P.
14
CHRIS FRASER
Before stating Proposition 4.3, we say a little more about normalized seed patterns
and Y -patterns. In a normalized seed pattern, the tuple of ratios (y1 (t), . . . , yn (t)) determines the coefficient tuple by (4.2). Accordingly, for normalized seed patterns one keeps
track of yj (t) rather than p±j (t). Rewriting (1.2) and (1.3) in terms of yj (t) determines a
Y -pattern recurrence in the semifield P:
(4.3)
yj (t)−1
yj (t′ ) = {
yj (t)yk (t)[bkj (t)]+ (yk (t) ⊕ 1)−bkj (t)
if j = k
if j ≠ k.
A collection of quantities (B(t), y(t))t∈Tn satisfying (1.1) and (4.3), with the y(t) lying
in some semifield S, is called a Y-pattern in the semfield S. Notice that the concept of
semifield is now playing two different roles, either as the ambient semifield F>0 in which
the exchange relation calculations take place, or as the coefficient semifield P used to
remove the ambiguity in mutation of seeds. The surprising connection between these
two roles is Lemma 1.7, which we now recognize as saying that the (B(t), ŷ(t)) form
a Y -pattern in the ambient semfield F>0 .
The most important example of a coefficient semifield arising in applications is the
tropical semifield.
Definition 4.2. A tropical semifield is a free abelian multiplicative group in some generators u1 , . . . , um , with auxiliary addition ⊕ given by
a
min(aj ,bj )
b
j
j
∏ uj ⊕ ∏ uj = ∏ uj
j
j
.
j
A normalized seed pattern over a tropical semifield is said to be of geometric type. When
this is the case, denoting the frozen variables by xn+1 , . . . , xn+m , the data of (B(t), y(t))
is entirely described by an (n+m)×n matrix B̃(t) = (bij (t)), called the extended exchange
matrix. Its top n × n submatrix is B(t) and is called the principal part. Its bottom m
n+i,j
rows are called coefficient rows. They are specified from the equality yj (t) = ∏ xn+i
m
b
(t)
.
i=1
The mutation rule (4.3) translates into the rule (1.1) on B̃.
The seed pattern in Figure 1 is of geometric type over the tropical semifield in the
frozen variables in (1.8). The same holds for the seed pattern in Figure 2 over the frozen
variables in (1.13). On the other hand, the seed pattern in Figure 3 is not normalized,
e.g. the first exchange relation there satisfies p+2 ⊕ p−2 = Y1,1 Y2,4 Y3,5 .
Now we state Proposition 4.3 describing quasi-homomorphisms between normalized
seed patterns E and E. It arose during the process of writing a forthcoming book on
cluster algebras [14], in proving one direction of the finite type classification (namely,
that a cluster algebra with a quiver whose principal part is an orientation of a Dynkin
quiver necessarily has only finitely many seeds). We will state it as a recipe for constructing a normalized seed pattern from a given one, since we envision this being useful in
applications.
Proposition 4.3. Let E be a non-normalized seed pattern, with the usual notation. Let
(xi ) = (xi (t0 )) be a fixed initial cluster in E. Let P be a semifield, and F >0 the semifield of
subtraction-free rational expressions in algebraically independent elements x1 , . . . , xn with
coefficients in P.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
15
Let Ψ∶ F>0 → F >0 be a semifield map satisfying Ψ(xi ) ≍ xi , and let c∶ F>0 → P be
xi ↦1
Ψ
the composition of semifield maps F>0 Ð
→ F >0 ÐÐ→ P where the second map in this
composition specializes all xi to 1 and is the identity on P.
Then there is a normalized seed pattern E in F >0 with seeds (B(t), p(t), x(t)) satisfying
(4.4)
B(t) = B(t)
(4.5)
xi (t) =
(4.6)
(4.7)
Ψ(xi (t))
c(xi (t))
ŷ i (t) = Ψ(ŷi (t))
y i (t) = c(ŷi (t)).
Clearly, Ψ is a quasi-homomorphism from E to E.
Proof. Formulas (4.5) through (4.7) follow by applying [13, Proposition 3.4]) to Ψ(E),
renormalizing by the scalars c(xi ), and massaging the formulas given there. Alternatively,
it is also straightforward to check the right hand sides of (4.5) through (4.7) satisfy the
required recurrences directly (this is carried out in [14]).
Example 4.4. Beginning with the seed pattern in Figure 1, one can construct the normalized seed pattern in Figure 2 by first applying the semifield map F ∗ – obtaining the nonnormalized seed pattern in Figure 3– and then normalizing by a semifield map c∶ F>0 → P.
This map c agrees with F ∗ on frozen variables and sends a cluster variable x to the frozen
variable monomial dividing F ∗ (x), e.g. c(∆235 ) = Y3,5 and c(∆245 ) = Y2,4 Y3,5.
When both E and E are of geometric type, constructing a quasi-homomorphism that
sends one seed into (the seed orbit of) another seed is a matter of linear algebra:
Corollary 4.5. Let Σ = (B̃, {x1 , . . . , xn }) and Σ = (B̃, {x1 , . . . , xn }) be seeds of geometric
type, with frozen variables xn+1 , . . . , xn+m and xn+1 , . . . , xn+m respectively. Let E and E be
the respective seed patterns.
Let Ψ be a quasi-homomorphism from E to E such that Ψ(Σ) ∼ Σ. It determines a
monomial map from the xi to the xi . Let MΨ denote the matrix of exponents of this
monomial map, thus MΨ is an (n + m) × (n + m) matrix satisfying Ψ(xk ) = ∏ xi
n+m
(MΨ )ik
.
i=1
Then the extended exchange matrices B̃, B̃ are related by
(4.8)
B̃ = MΨ B̃.
In particular, such a a quasi-homomorphism Ψ exists if and only if the principal parts
of B̃, B̃ agree, and the (integer) row span of B̃ contains the (integer) row span of B̃.
Proof. Indeed, the (i, j) entry of the left hand side of (4.8) encodes the exponent of xi in
ŷ j , while the (i, j) entry of the right hand side encodes the exponent of xi in Ψ(ŷj ). So
(4.8) now follows from (4.6).
The final statement follows by studying (4.8): the “interesting” rows of MΨ are its
bottom m rows. Each of these rows determines a particular linear combination of the
rows of B̃, and these linear combinations can be prescribed arbitrarily by prescribing the
exponent of xi in Ψ(xj ) for 1 ≤ i ≤ m, 1 ≤ j ≤ n + m using Lemma 1.5.
16
CHRIS FRASER
Remark 4.6 (Exchange graphs and separation of additions). The formulas (4.4),(4.5)
and (4.7) show that if there is a quasi-homomorphism from E to E, then the exchange
graph of E covers that of E. In particular, by Corollary 4.5, if the rows of B̃ span Zn ,
then the exchange graph for the corresponding cluster algebra A(B̃) covers the exchange
graph of every other cluster algebra A with the same underlying exchange matrix.
This is a natural generalization of the separation of additions formula [17, Theorem 3.7]
from the case of a quiver with principal coefficients to any B̃-matrix whose rows span Zn .
Namely, let Σ0 = (B0 , y, x) and Σ0 = (B0 , y, x) be a pair of normalized seeds with the
same exchange matrix, and suppose Σ0 has principal coefficients, i.e. yi = xn+i . There is a
natural choice of maps Ψ mapping Σ0 to Σ0 as in Proposition 4.3, defined by Ψ(xi ) = xi
and Ψ(xn+i ) = y i . For this choice of Ψ, formula (4.5) becomes separation of additions:
the numerator of [17, Theorem (3.7)] (evaluating the “X polynomial” in F) is applying
the semifield homomorphism Ψ, while the denominator (specializing the cluster variables
to 1 and evaluating the X polynomial in P) is applying the semifield map c.
Remark 4.7 (Proportionality and gradings). Let E be a seed pattern of geometric type.
We recall briefly the concept of a Zr -grading on E cf. [21, 22]. Choosing an initial seed
(B̃, {xi }) in E, such a choice of grading is determined by a r × (n + m) grading matrix G
satisfying GB̃ = 0. The ith column of G determines the grading of xi as a vector in Zr ,
for 1 ≤ i ≤ n + m. The condition GB̃ = 0 guarantees that every exchange relation (1.5)
is homogeneous with respect to this Zr -grading; this in turn defines the multi-grading
of each adjacent cluster variable and thereby each adjacent grading matrix. It can be
seen that these adjacent grading matrices again satisfy the left kernel condition, so that
the grading propagates to a Zr -grading on the entire cluster algebra in which the cluster
variables and coefficients are homogeneous.
Now we suppose we are given two seeds E and E of geometric type with notation as
in Corollary 4.5. Let Ψ1 and Ψ2 be a pair of proportional quasi-homomorphisms of E
and E. We obtain as in (4.8) matrices MΨ1 and MΨ2 such that MΨ1 B̃ = MΨ2 B̃ = B̃, which
implies that MΨ1 −MΨ2 defines a Zm -grading G on E (the first n rows of MΨ1 −MΨ2 define
the trivial grading). Conversely, fixing a quasi-homomorphism Ψ1 with matrix MΨ1 , any
choice of Zm -grading matrix G on B̃ provides a quasi-homomorphism Ψ2 , proportional
to Ψ1 , whose matrix is MΨ2 = MΨ1 + G.
Remark 4.8. For simplicity, we stated Corollary 4.5 in terms of Z row spans, but a similar
statement holds for Q row spans. To do this, one enlarges the tropical semifield P to the
Puiseux tropical semifield consisting of Puiseux monomials with rational exponents in the
frozen variables. This is unpleasant from the perspective of cluster algebras as coordinate
rings, but is perfectly fine if one is only interested in writing algebraic formulas for cluster
variables, etc.
This is foreshadowed in the work of Sherman and Zelevinsky [29, Section 6], which
0 a
).
discusses the coefficient-free rank 2 cluster algebra A(b, c) with exchange matrix (
−b 0
The authors write the cluster variables in any cluster algebra with this B matrix in terms
of the cluster variables for A(b, c). Their formulas involve Puiseux monomials in the
frozen variables.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
17
5. Nerves
By Proposition 3.2, to check that a given semifield map Ψ is a quasi-homomorphism
from E to E, it suffices to check that Ψ(Σ(t)) ∼ Σ(t) for some pair of seeds Σ(t) in E and
Σ(t). By Proposition 2.3, this means checking that B(t) = B(t) and ŷ(t) = ŷ(t), and
furthermore Ψ(xj (t)) ≍ xj (t) holds for all j. We envision applications where checking the
proportionality condition on cluster variables is easy and can be done in many seeds t, but
checking the equality of exchange matrices or ŷ’s is inconvenient. The goal of this section
is to give a criterion that guarantees Ψ is a quasi-homomorphism by only checking these
proportionality conditions. The relevant concept is that of a nerve for a seed pattern.
Definition 5.1. Let E be a seed pattern. A nerve N for Tn , is a connected subgraph
of Tn such that every edge label k ∈ [1, n] arises at least once in N .
The basic example of a nerve is the star neighborhood of a vertex. We believe that
there are many theorems of the form, “if a property holds on a nerve, then it holds on the
entire seed pattern.” We give an example of such a theorem in the appendix, generalizing
the “Starfish Lemma” [11, Proposition 3.6] from a star neighborhood to a nerve.
Before stating the result of this section, we need to address the (mostly unimportant)
difference between a seed and its opposite seed. We say a seed is indecomposable if the
underlying graph described by its exchange matrix (the vertex set is [1, n] and vertices
i, j are joined by an edge if bi,j ≠ 0) is connected. For a seed Σ = (B, p, x), the opposite
= xi .
seed Σopp = (B opp , popp , xopp ) is the seed defined by B opp = −B, (popp )±j = p∓j , and xopp
i
opp
1
It satisfies ŷj = ŷj . The operations of restricting to an indecomposable component and
replacing a seed by its opposite seed both commute with mutation.
Proposition 5.2. Let E and E be non-normalized seed patterns, with respective ambient
semifields F>0 and F >0 . Suppose the seeds in E are indecomposable. Let Ψ∶ F>0 → F >0 be
a semifield homomorphism that preserves coefficients and satisfies Ψ(xj (t)) ≍ xj (t) for
j
every vertex t and label j such that t Ð
→ t′ is in N . Then Ψ is a quasi-homomorphism
opp
from E to E or from E to E .
In particular applying the proposition when N is the star neighborhood of a vertex t,
to check that Ψ is a quasi-homomorphism, it suffices to check that Ψ(xj (t)) ≍ xj (t) for
→ t′ . Lemma
all j ∈ [1, n], as well as checking Ψ(xj (t′ )) ≍ xj (t′ ) for each adjacent edge t Ð
3.7 now follows.
j
Proof. Choose a vertex t ∈ N . By hypothesis for all j, Ψ(xj (t)) = cj (t)xj (t) for some
cj (t) ∈ P, so we are left checking that B(t) = B(t) and ŷ(t) = ŷ(t). Suppose t Ð
→ t′ is an
′
′
′
′
edge in N , then there is a scalar ck (t ) such that Ψ(xk (t )) = ck (t )xk (t ).
The exchange relation defining xk (t′ ) in E is
k
(5.1)
xk (t)xk (t′ ) = p+k (t) ∏ xj (t)[bjk (t)]+ + p−k (t) ∏ xk (t)[−bjk (t)]+ .
On the other hand, applying Ψ to the relation defining xk (t′ ) in E and rearranging
yields
(5.2)
1
xk (t)xk (t′ ) =
(Ψ(p+k (t)) ∏ Ψ(xj (t))[bjk (t)]+ + Ψ(p−k (t)) ∏(Ψ(xj (t)))[−bjk (t)]+ ).
′
ck (t )ck (t)
18
CHRIS FRASER
Abbreviating the two terms on the right hand side of (5.1) as X + Y , and the two
terms in (5.2) as Z + W , we see by algebraic independence in the seed at t that either
X = Z, Y = W , or X = W, Y = Z. Refer to these as Case 1 or Case 2 respectively. By
Z
inspection, we see that ŷ k (t) is the ratio X
Y , while Ψ(ŷk (t)) is W . Thus in Case 1 we
deduce that Ψ(ŷk (t)) = ŷ k (t) and the matrices B(t) and B(t) have the same k th column.
opp
In Case 2 we deduce the same thing once we replace E by E .
Now apply Lemma 5.3.
Lemma 5.3. Let Y = {y(t), B(t)} and Y = {y(t), B(t)} be two Y -patterns whose matrices
B(t) are indecomposable. Let N be a nerve for Tn . Suppose for every vertex t ∈ N and
k
label k such that the edge t Ð
→ t′ is in N , one of the following holds
(5.3)
(5.4)
yk (t) = y k (t) and bjk (t) = bjk (t) for all j ∈ [1, n], or
ykopp (t)
= y k (t) and bopp
jk (t) = bjk (t) for all j ∈ [1, n],
then Y = Y or Y opp = Y accordingly.
Roughly, there are two issues here: first the question of whether Y -patterns can be
checked on a nerve (they can), and second whether we are dealing with Y or Y opp (this
relies on indecomposability).
Proof. In any Y -pattern, for a given (k, t) pair (not necessarily in N ), we will refer to
yk (t) and the k th column of B(t) as the k-part of the seed at t. The equations in (5.3)
k
→ t′ ∈ N .
say that Y, Y have either the same k-parts, or opposite k-parts, for any edge t Ð
Pick a vertex t0 ∈ N , and an edge k ∈ N incident to t0 . If necessary, replace Y by Y opp
so that the given Y -patterns have the same k-part at t0 . We seek to prove Y, Y have the
same j-part at t0 , for all j ∈ [n].
k
Let t0 Ð
→ t1 ∈ N be an edge in the nerve incident to t0 . The mutation rules (1.1), (4.3)
are involutive and have the property that for any j, the j-part of the seed at t1 depends
only on the j-part and k-part of the seed at t0 . Since the given Y -patterns agree at k,
we see that their j-parts agree at t1 if and only if they agree at t0 . Repeatedly apply
this observation, mutating in all possible directions in the nerve, while preserving the fact
that the j-parts at t ∈ N coincide if and only if they coincide at t0 . Since the nerve is
connected and every edge label shows up at least once in N , we conclude that for all j,
the j-parts at Y and Y are either the same or opposite. The connectedness hypothesis
assures they are all in fact the same.
6. Quasi-automorphisms and the cluster modular group
A quasi-automorphism is a quasi-isomorphism from a given seed pattern E to itself,
cf. Definition 3.5. One can think of a quasi-automorphism as a choice of a map describing an automorphism of the pattern of seed orbits associated to E. We will use
quasi-automorphisms to define a variant of a group of automorphisms of E, generalizing
the group of cluster automorphisms defined for seed patterns with trivial coefficients in
[2] while retaining many of the properties of cluster automorphisms (e.g. Proposition 3.2,
Corollary 4.5 and Proposition 5.2).
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
19
The following example illustrates that the notion of quasi-automorphism is more general
than the “naive” notion of a semifield automorphism preserving the seed orbit pattern.
Example 6.1. A quasi-automorphism does not have to be an automorphism of semifields.
Consider the composition G∗ ○ F ∗ from Example 3.8, which is a quasi-automorphism
of C[X] proportional to the identity map. It rescales each Plücker variable by a product
of frozens: G∗ ○ F ∗ (∆S ) = ∆145 ∆125 ∆S . The ambient semifield of C[X] has a grading for
which every Plücker variable is degree one, and every homogeneous element in the image
of G∗ ○ F ∗ has degree a multiple of 3. Thus G∗ ○ F ∗ cannot be surjective.
Definition 6.2. The quasi-automorphism group QAut0 (E) is the set of proportionality
classes of quasi-automorphisms of E. This is the automorphism group of E in the quotient
category discussed in Remark 3.6.
Remark 6.3. Let us call a quasi-automorphism trivial if it is proportional to the identity
map. The set of trivial quasi-automorphisms is a monoid (but not usually a group) under
composition; the composition G∗ ○ F ∗ from Example 3.8 bears witness to this. One way
to construct quasi-automorphisms proportional to a given Ψ is to compose ǫ1 ○ Ψ ○ ǫ2
with ǫ1 and ǫ2 trivial. It is tempting to try and define QAut0 (E) purely in terms of thse
trivial quasi-automorphisms, without mentioning proportionality. However the relation ≡
defined by Ψ1 ≡ Ψ2 if Ψ2 = ǫ1 ○ Ψ ○ ǫ2 is neither symmetric nor transitive, so one cannot
form a quotient category using this relation.
We write QAut0 (E) = QAut0 (B̃) when E is of geometric type and specified by an initial
matrix B̃. By Remark 4.7, two quasi-automorphisms are proportional to each other if
and only if their ratio defines a Zm -grading on E (taking exponents of elements of P to
obtain elements of Zm ). Fixing a particular quasi-automorphism Ψ, the number of degrees
of freedom in specifying another quasi-automorphism proportional to Ψ is therefore the
corank of B̃.
Lemma 6.4. Let E be a seed pattern of geometric type and Ψ a quasi-homomorphism
from E to itself. Then Ψ is a quasi-automorphism.
Thus when E is of geometric type, every quasi-homomorphism Ψ from E to itself determines an element of QAut0 (E), i.e. any such Ψ has a quasi-inverse.
Proof. By [3, Lemma 3.2], if two B̃-matrices B̃(t0 ) and B̃(t0 ) are in the same mutation
class, they are related by a pair of unimodular integer matrices: B̃(t0 ) = M B̃(t0 )N,
for M ∈ GLm+n (Z), and N ∈ GLn (Z).
By Corollary 4.5, for a pair of vertices t0 , t0 ∈ Tn , there is a quasi-homomorphism Ψ
sending the seed orbit at t0 to the seed orbit at t0 if and only if the principal parts of
B̃(t0 ) and B̃(t0 ) agree, and the row span of B̃(t0 ) contains the row span of B̃(t0 ). By
the unimodularity of mutation, this criterion is preserved under swapping the roles of t0
and t0 – if the row span of B̃(t0 ) contains the row span of B̃(t0 ) then in fact the two row
spans are equal submodules of Zn .
Remark 6.5. Marsh and Scott [24] described a version of the twist for the Grassmannian
cluster algebras. One can show that it is a quasi-automorphism using [24, Corollary 8.6].
We will now recall the definitions of some preexisting groups of automorphisms associated a seed pattern E. Namely:
20
CHRIS FRASER
● the cluster modular group CMG(E) of Fock and Goncharov [9], and
● the group Aut(E) of automorphisms in the category of (rooted) cluster algebras
defined by Assem, Dupont and Schiffler [1].
We first present these definitions and then discuss a particular example where all the
groups are computed and compared to each other and to the quasi-automorphism group.
Definition 6.6 (Cluster modular group [9, Definition 2.14]). Let E be a seed pattern with
exchange graph E. The cluster modular group CMG(E) is the group of graph automorphisms g ∈ Aut(E) that preserve the exchange matrices. More precisely, recall that the
unlabeled seed at vertex t ∈ E is indexed not by [1, n] but by the elements of star(t). Then
an element of the cluster modular group is a graph automorphism g ∈ Aut(E) satisfying
B(t)t′ ,t′′ = B(g(t))g(t′ ),g(t′′ ) for all t ∈ E and t′ , t′′ ∈ star(t). Such a graph automorphism
can be determined by choosing a pair of vertices t0 , t0 ∈ E and an identification of star(t0 )
with star(t0 ) under which B(t0 ) = B(t0 ).
Remark 6.7. Because Definition 6.6 is in terms of automorphisms of the exchange graph,
the cluster modular group appears to depend on the entire seed pattern E, and not just
the underlying exchange matrices in E. However, it is widely believed that the exchange
graph – and therefore the cluster modular group – is in fact independent of the choice of
coefficients (i.e., it only depends on the exchange matrices, and therefore can be prescribed
by giving a single such matrix). This has been proven for skew-symmetric exchange
matrices [23].
The quasi-automorphism group is a subgroup of the cluster modular group. Indeed,
each quasi-automorphism Ψ determines a cluster modular group element g via Ψ(Σ(t)) ∼
Ψ(g(t)), and proportional quasi-automorphisms determine the same g. Since Ψ preserves
exchange matrices and evaluating Ψ commutes with permuting the cluster variables in a
seed, the element g produced this way is indeed an element of the cluster modular group.
One can also consider automorphisms in the category of cluster algebras defined in [1].
We reproduce a version of the definition for the sake of convenience.
Definition 6.8. Let E be a seed pattern. We say two seeds Σ1 and Σ2 in E are similar if
Σ2 coincides with Σ1 after first permuting the frozen variables, and then permuting the
indices [1, n] appropriately.
Suppose the exchange matrices in E are indecomposable. Let A be its cluster algebra.
A Z-algebra map f ∶ A → A is an automorphism of E if for every (equivalently, for any) seed
Σ in E, f (Σ) or f (Σ)opp is similar to a seed in E. We denote the group of automorphisms
of E by Aut(E).
The elements of Aut(E) are similar to strong isomorphisms from [16] but slightly more
general since one is allowed to permute the frozen variables.
We say f as in Definition 6.8 is a direct automorphism or inverse automorphism according to whether f (Σ) or f (Σ)opp is a seed in E. Let Aut+ (E) ⊂ Aut(E) denote the
subgroup of direct automorphisms. By similar reasoning to [2, Theorem 2.11], this subgroup has index two in Aut(E) if each seed Σ in E is mutation-equivalent to Σopp ; otherwise
Aut+ (E) = Aut(E).
An important special case of Definition 6.8 is when E has trivial coefficients in which
case the group Aut(E) is the group of cluster automorphisms [2]. When E has trivial
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
21
coefficients, we have Aut+ (E) = CMG(E). Furthermore, a direct cluster automorphism is
the same as a quasi-automorphism in this case.
We can summarize the containments between the preceding groups as
(6.1)
Aut+ (E) ⊂ QAut0 (E) ⊂ CMG(E) = Aut+ (Etriv ).
?
where E is a seed pattern and Etriv is the seed pattern obtained from E by trivializing its
?
coefficients. The equality CMG(E) = Aut+ (Etriv ) depends on the belief that CMG(E) =
CMG(Etriv ), cf. Remark 6.3. The group Aut(Etriv ) contains all of the groups in (6.1), and
the group Aut(E) doesn’t sit nicely with the rest of the containments when Aut+ (E) ⊊
Aut(E).
We next illustrate the differences between the groups in (6.1) using a particular cluster
algebra associated with a bordered marked surface. Basic notions and references concerning this class of cluster algebras are given in Section 7.
Example 6.9. Let (S, M) be an annulus with two marked points on each boundary
component cf. Figure 4. We have colored the marked points either black or white to aid
in describing the automorphism groups below.
v1
v2
v3
v4
v1
∧
v3
v4
∧
v2
v3
v∨4
Figure 4. An annulus with two marked points on each boundary component. At right, we show a “flat form” of this annulus obtained by cutting
along the dashed line.
The cluster modular group CMG(S, M) for a cluster algebra associated with this annulus coincides with the mapping class group of the annulus (see Proposition 7.2 below).
This group has the following explicit description: let ρ be the (isotopy class of) the homeomorphism of S that rotates the inner boundary of the annulus clockwise by a half-turn.
Let τ be the clockwise half-turn of the outer boundary. Let σ be the homeomorphism represented by a 180 degree turn of the flat form of the annulus; it swaps the inner and outer
boundary components. Then the elements ρ, τ, and σ generate the cluster modular group.
The group has a presentation CMG(S, M) = ⟨ρ, τ, σ∶ (ρτ )2 = σ 2 = 1, ρτ = τ ρ, σρ = τ σ⟩ with
respect to these generators. It is a central extension 1 ↦ Z/2Z ↦ CMG ↦ Dih∞ ↦ 1 of
the infinite dihedral group Dih∞ = ⟨r, s∶ s2 = (sr)2 = 1⟩ by Z/2Z = ⟨ρτ ⟩, using the map
σ ↦ s, ρ ↦ r, τ ↦ r −1 .
Figure 5 gives a choice of lamination L and triangulation T , as well as the quivers
B̃(T ), B̃(ρ(T )) and B̃(ρ2 (T )). Let A be the corresponding cluster algebra with frozen
variable xL , E its seed pattern, and F>0 its ambient semifield.
The cluster modular group element σρτ permutes the arcs in T . It induces an automorphism of the quiver B̃(T ), and therefore an element of Aut(E).
22
CHRIS FRASER
T
a
c
L
xd
B̃(T )
B̃(ρ(T ))
B̃(ρ2 (T ))
xb
xd
xf
xL
xc
xf
xL
xe
xh
xL
xg
d
b
xa
xc
xe
Figure 5. A lamination L consisting of two copies of the same curve
on the annulus, determining a single frozen variable xL . We have also
drawn a triangulation T of this annulus by the arcs a, b, c, d. The quivers
B̃(T ), B̃(ρ(T )), and B̃(ρ2 (T )) are shown at right, where the extra arcs
are e = ρ2 (a), f = ρ2 (b), g = ρ2 (c), h = ρ2 (d). The values of ŷ in each quiver
are read off as the Laurent monomial “incoming variables divided by outgoing variables.”
The quivers B̃(T ) and B̃(ρ2 (T )) are neither isomorphic nor opposite, so there is no
strong automorphism sending the seed at T to the seed at ρ2 (T ). Likewise, there is no
strong automorphism between T and ρ±4 (T ), ρ±6 (T ), ρ±8 (T ), and so on. However, there
is a quasi-automorphism relating these seeds, which we describe now. It is the semifield
map Ψ∶ F>0 → F>0 defined by Ψ(xL ) = xL as well as Ψ(xγ ) = x−1
L ⋅ xρ2 (γ) for γ = a, b, c, d. It
sends each ŷ for Σ(T ) to the corresponding ŷ for Σ(ρ2 ⋅T ), defining a quasi-automorphism
of E whose map on seed orbits is ρ2 . It has a simple global description on cluster variables
which can be checked inductively by performing appropriate mutations away from T .
Namely, for each arc γ let ι(γ, L) denote the number of times γ crosses the two curves in
L. For example, ι(a, L) = 0 and ι(c, L) = 1. Then
(6.2)
Ψ(xγ ) = xL
ι(γ,L)−ι(ρ2 (γ),L)
⋅ xρ2 (γ)
for all arcs γ in the annulus. The power of xL on the right hand side of (6.2) is always
equal to 0, 1, or −1. It is also simple to describe quasi-automorphisms realizing σ and ρτ .
Perhaps surprisingly, the seeds at T and ρ(T ) are not related by a quasi-automorphism.
Indeed, the values of ŷ are equal at the top and bottom vertices of B̃(T ) in Figure 5, but
they are not equal in B̃(ρ(T )). The same holds for T and τ (T ).
Putting all of this together, there is only one nontrivial strong automorphism of E,
namely the element σρτ . On the other hand, the quasi-automorphism group is infinite,
generated by ρ2 , σ and ρτ . It is a direct product Dih∞ ×Z/2Z. It is an index two subgroup
of CMG(S, M), namely the kernel of the map CMG ↠ Z/2Z that computes the parity of
the number of black marked points sent to a white marked point.
Example 6.9 suggests that although the cluster modular group CMG(E) may be strictly
larger than the quasi-automorphism group QAut0 (E), the gap between these groups is
not so large. Indeed, Section 7 establishes that for seed patterns associated with surfaces,
QAut0 (E) is always a finite index subgroup of the cluster modular group.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
23
7. Quasi-automorphisms of cluster algebras from surfaces
In this section we place Example 6.9 in context via results valid for any cluster algebra associated to a marked bordered surface as in [10, 12, 13, 20]. We describe quasiautomorphisms of these cluster algebras in terms of the tagged mapping class group of
the marked surface.
We follow the setup and notation in [13]. Let A(S, M, L) denote the cluster algebra
of geometric type determined by the triple (S, M, L). Here S is an oriented bordered
surface with a nonempty set M of marked points. The marked points reside either in
the interior of S (we call these punctures) or in ∂S (we call these cilia). The set of
punctures is M. We disallow a few possibilities for (S, M), namely a sphere with three
or fewer punctures, an n-gon when n < 4, and a once-punctured monogon. The choice of
coefficients is specified by a multi-lamination L = (L1 , . . . , Lm ), an m-tuple of (integral
unbounded measured) laminations on (S, M). Each lamination Li consists of a finite
number of curves in (S, M).
The cluster variables in A(S, M) are indexed by tagged arcs γ, the set of which we
denote by A⋈ (S, M). The seeds in A(S, M, L) are indexed by tagged triangulations T
of (S, M). The extended exchange matrix B̃(T ) for a seed has the signed adjacency
matrix B(T ) as its principal part, and has the shear coordinate vector ⃗b(T, Li ) of the
lamination Li with respect to T as its ith row of coefficients.
The exchange graph of the resulting cluster algebra is independent of the choice of
coefficients [13, Corollary 6.2]. We let CMG(S, M) denote the corresponding cluster
modular group. It is closely related to the following geometrically defined group.
Definition 7.1 (Tagged mapping class group [2]). Let (S, M) be a bordered marked
surface that is not a closed surface with exactly one puncture. A tagged mapping class
for (S, M) is a pair g = (f, ψ), where
● f is an element of the mapping class group of (S, M) – i.e. f is an orientiationpreserving homeomorphism of S mapping M to itself setwise, considered up to
isotopies of S that fix M pointwise, and
● ψ∶ M → {±1} is a function from the set of punctures to {±1}.
When (S, M) is a closed surface with one puncture p, we make the same definition but
impose ψ(p) = 1 since tagged versions of arcs are not in the cluster algebra. The tagged
mapping classes comprise the tagged mapping class group, denoted MG⋈ (S, M).
We understand MG⋈ (S, M) by its action on tagged arcs γ ∈ A⋈ (S, M). A tagged
mapping class g = (f, ψ) acts on γ by first performing the homeomorphism f to γ, and
then changing the tag of any end of γ incident to a puncture p for which ψ(p) = −1. The
resulting action on tagged triangulations preserves the signed adjacency matrices, and
embeds MG⋈ (S, M) as a subgroup of CMG(S, M), cf. [2]. In fact the following is true:
Proposition 7.2 (Bridgeland-Smith). The tagged mapping class group MG⋈ (S, M) coincides with the cluster modular group CMG(S, M), unless (S, M) is a sphere with four
punctures, a once-punctured square, or a digon with one or two punctures.
Thus barring these exceptional cases, two tagged triangulations of (S, M) have isomorphic quivers precisely when they are related by an element of the tagged mapping class
group (see [4, Proposition 8.5] and the subsequent discussion; see also [2, Conjecture 1]).
24
CHRIS FRASER
In the exceptional cases listed in Proposition 7.2, the tagged mapping class group is a
proper finite index subgroup of the cluster modular group.
Motivated by Proposition 7.2, we set out to describe, for various choices of coefficients L,
the quasi-automorphism group QAut0 (S, M, L) from Definition 6.2 as a subgroup of the
tagged mapping class group. The main ingredient in our answer is a black-white coloring
similar to one in the examples from Section 6.
Definition 7.3. The even components of (S, M) are the punctures C ∈ M as well the
as boundary components C ⊂ ∂S having an even number of cilia. We let r denote the
number of even components, and label the even components C1 , . . . , Cr . For each even
boundary component C ⊂ ∂S, we color the cilia on C black or white so that the colors
alternate, i.e. adjacent cilia have opposite colors.
Using the black-white coloring in Definition 7.3, each tagged mapping class g = (f, ψ)
determines an r × r signed permutation matrix πg whose entries are indexed by the even
components. The (i, j) entry of πg is 0 unless f (Ci ) = Cj . If Cj ⊂ ∂S is a boundary
component and f (Ci ) = Cj , then the (i, j) entry is +1 if f sends black cilia on Ci to black
cilia on Cj , and is −1 if f sends black cilia on Ci to white cilia on Cj . When Ci and Cj
are punctures, the sign of the (i, j) entry is the sign ψ(Cj ). Not all signed permutation
matrices will arise in this way since f can only permute components that have the same
number of cilia.
Definition 7.4. Let L be a lamination. For each curve α in L, Figure 6 shows how to
assign a sign to an end of α that either lands on even boundary component or spirals
around a puncture. At a puncture, the sign is chosen according to whether α spirals
counterclockwise or clockwise into the puncture. At a boundary component, the sign
is chosen according to whether the nearest neighboring cilium in the clockwise direction
along C is black or white. An end on an odd component has zero sign. The pairing p(L; C)
of a lamination L with the even component C is the sum of all the signs associated to L,
i.e. the sum over all curves α in L of the signs of the two ends of α. We let p⃗(L) =
(p(L; Ci ))i=1,...,r ∈ Zr denote the vector of pairings of L with the even components.
Example 7.8 works out these signs for the annulus from Example 6.9.
+1
-1
●
+1
●
-1
Figure 6. The conventions for assigning signs to each end of a curve that
lands on an even boundary component (in this case, a boundary component
with 4 cilia) or spirals around a puncture. The pairing p(L, C) is obtained
by adding up all of these signs.
In addition to acting on tagged arcs, MG⋈ (S, M) also acts on laminations L. A tagged
mapping class g = (f, ψ) acts by first performing the homeomorphism f to L, and then
changing the direction of spiral at each puncture p for which ψ(p) = −1. This action
preserves sheard coordinates in the sense that ⃗b(T, L) = ⃗b(g(T ), g(L)) for a triangulation
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
25
T and lamination L. It is easy to see that g acts on the vector of pairings by the matrix
πg , i.e. p⃗(g(L)) = πg ⋅ p⃗(L) for a lamination L.
The next theorem is the main result of this section, describing QAut(S, M, L) inside
the marked mapping class group in very concrete terms.
Theorem 7.5. Suppose (S, M) is not one of the four exceptional surfaces in Proposition
7.2. Let L be a multi-lamination. Let VL = span({⃗
p(L)∶ L ∈ L}) ⊂ Zr be the submodule
spanned by the vectors of pairings associated to the laminations L in L. Then
(7.1)
QAut0 (S, M, L) = {g ∈ MG⋈ (S, M)∶ πg (VL ) = VL }.
We prove Theorem 7.5 in Section 8. The subgroup of MG⋈ (S, M) described in (7.1) only
depends on the endpoint behavior of laminations – it doesn’t mention the topology of the
surface, or how much curves wrap around the holes and handles of the surface. The map
g ↦ πg is a group homomorphism from MG⋈ (S, M) to the group of signed permutation
matrices. The subgroup in (7.1) is an inverse image of the subgroup of signed permutation
matrices that fix VL and therefore is always finite index in MG⋈ (S, M).
Corollary 7.6. Let g be a tagged mapping class. If πg is plus or minus the identity
matrix, then g ∈ QAut0 (S, M, L) for any choice of multi-lamination L. Otherwise, g ∉
QAut0 (S, M, L) for some choice of L.
Remark 7.7. The tagged mapping classes in Corollary 7.6 are those that fix all even
components setwise, and furthermore either preserve the black-white coloring of ends of
curves, or simultaneously swap all colors. This group is generated by the following four
types of elements (see [8] for generators of the mapping class group): Dehn twists about
simple closed curves, homeomorphisms that permute odd components, fractional Dehn
twists rotating the cilia on a given boundary component by two units, and the tagged
rotation. This last element is the one that simultaneously changes tags at all punctures
and rotates all boundary components by one unit. It was studied in [5], where it was
shown to coincide with the shift functor of a 2-Calabi-Yau cluster category associated
with the surface.
Proof. If πg = ±1, then πg clearly preserves VL regardless of the choice of L and by Theorem
7.5 g is in QAut0 (S, M, L) for any L.
If πg ≠ ±1, think of πg as a signed permutation σ of {±1, . . . , ±r} in the usual way. If
there is any index i ∈ [1, r] such that σ(i) ≠ ±i, then let L be a lamination consisting of
a curve with two black ends on Ci , satisfying p(L; C) = 2. If there is no such index i, we
can choose a pair of indices i, j ∈ [1, r] such that σ(i) = −i but σ(j) = j. In this case we
let L be a lamination consisting of a curve connecting the even components Ci and Cj by
a curve that is black at both ends. In both of these two cases, we see that πg (⃗
p(L)) is
not in the span of p⃗(L) and by Theorem 7.5, g ∉ QAut0 (S, M, L).
Example 7.8. We order the boundary components in Figure 5 so that the inner boundary
is first. The vector of pairings for the lamination L in Figure 5 is p⃗(L) = (−2, −2). Then ρ
and τ act on the vector of parings by swapping the sign of the first or second component
respectively, and σ acts by permuting the first and second component. The description
of QAut0 (S, M, L) in Example 6.9 matches the one in Theorem 7.5. The subgroup of
elements described in Corollary 7.6 is a a direct product Z × Z/2Z generated by ρ2 and
ρτ . It has index 4 in the cluster modular group.
26
CHRIS FRASER
Remark 7.9. Theorem 7.5 can be modified in the case that (S, M) is one of the exceptional surfaces in Proposition 7.2. Namely, the left hand side of (7.1) merely describes
the subgroup of QAut0 (S, M, L) consisting of tagged mapping classes (that is, ignoring
the exotic symmetries). For particular choices of coefficients, the “extra” elements of the
cluster modular group might also be inside QAut0 .
8. Proofs for Section 7
The key result of this section is Proposition 8.4 describing quasi-homomorphisms of
cluster algebras from surfaces. Theorem 7.5 follows from it as a special case.
Let B(S, M) denote the set of boundary segments connecting adjacent cilia in ∂S. There
is an especially natural choice of multi-lamination Lboundary = (Lβ )β∈B(S,M) with one frozen
variable for each boundary segment (see e.g. in [13, Remark 15.8]). Lemma 8.1 expresses
the shear coordinates of certain laminations in terms of the extended exchange matrix
determined by Lboundary . It is patterned after [13, Lemma 14.3].
Lemma 8.1. Let L be a lamination none of whose curves has an end that spirals at a
puncture. Given an arc γ the transverse measure of γ in L is the minimal number of
intersections of γ with the curves in L. We denote it by l(γ, L). For a boundary segment
β ∈ B(S, M), we similarly let l(β, L) denote the number of ends of the curves in L on β.
We let ⃗l(T, L) = (l(⋆, L))⋆∈T ∪B(S,M) be the row vector containing all of these transverse
measures. Then
(8.1)
− 2⃗b(T, L) = ⃗l(T, L)B̃(T, Lboundary ).
Proof. We check that the γ0 components of the left and right hand sides of (8.1) are equal,
where γ0 ∈ T . Let γ1 , . . . , γ4 be the quadrilateral containing γ0 (number in clockwise order).
Some of the γi may be boundary segments. Each time α shears across the quadrilateral
in an ‘S’ crossing, it contributes +1 to the left hand side, while contributing − 21 (−1 + −1)
to the right hand side. And so on.
This argument is like [13, Lemma 14.13] but simpler because our we are not dealing
with spirals at the puncture, for which l(T, L) = ∞. Comparing (8.1) with (4.8), we see
that if L is any multi-lamination none of whose curves spiral at punctures, then there is
a quasi-homomorphism from A(S, M, Lboundary ) to A(S, M, L). A version of Lemma 8.1
allowing for spirals at punctures would involve the extended exchange matrix B̃(T , L)
on the fully opened surface (S, M), where L and T are lifts of L and T to the opened
surface (see [13, Sections 9,14] for details). The corresponding version of (8.1) determines
a quasi-homomorphism from A(S, M, L) to A(S, M, L).
Our next step is to to describe the row span of the signed adjacency matrices B(T ). It
strengthens [12, Theorem 14.3] which states that the corank of B(T ) is the number of even
components. The description requires associating a sign to the ends of arcs γ ∈ A⋈ (S, M),
in a similar fashion as was done for the ends of curves in Definition 7.4. Namely if γ has an
endpoint on a boundary component C ⊂ ∂S, the endpoint gets sign ±1 if its endpoint is a
black or white cilium respectively. If the endpoint is on a puncture C ∈ M , the sign is ±1 if
the end is plain or tagged respectively. An endpoint on an odd component gets sign 0. The
pairing p(γ; C) of γ with C is the sum of the signs of the ends of γ that reside on C, and
the vector of pairings is p⃗(γ) = (p(γ; Ci ))i=1,...,r . This pairing satisfies p(γ; C) = p(Lγ ; C)
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
27
where Lγ is the elementary lamination determined by γ (cf. [13, Definition 17.2], see also
[27, Section 5]).
Lemma 8.2. For a tagged triangulation T , let Q(T ) = Q(γ∶ γ ∈ T ) be Q-vector space of
row vectors with entries indexed by γ ∈ T . Let Q(T )∗ = Q(γ ∗ ∶ γ ∈ T ) be the dual space of
column vectors with entries indexed by the dual basis {γ ∗ ∶ γ ∈ T }.
For each even component C, consider the column vector
(8.2)
RC = ∑ p(γ; C)γ ∗ ∈ Q(T )∗ .
γ∈T
Then a vector a
⃗ ∈ Q(T ) is in the the row span of B(T ) if and only if the dot product
a
⃗ ⋅ RC vanishes for all C.
Said differently, the map γ ↦ (γ, Ci ) for Ci an even component determines a Z-grading
on the coefficient-free cluster algebra A(S, M). These gradings form a standard Zr -grading
as i varies from 1, . . . , r (a standard grading is one that spans the kernel of the B-matrices,
see [21]). We will not rely on gradings in what follows. Lemma 8.2 is proved at the end
of this section.
For a vector a
⃗ ∈ Q(T ), the residue of a
⃗ around C in T is the dot product a⃗ ⋅ RC =
⃗ as the shear coordinate of a lamination L, the residue has the
∑γ∈T (γ, C)aγ . Writing a
following simple description.
Lemma 8.3. Let L be a lamination and C an even component. Then the residue of
⃗b(T, L) around C is the pairing p(L; C) from Definition 7.4.
Proof. The residue is computed in terms of the shear coordinates of arcs adjacent to C.
To compute these shear coordinates, rather than considering the entire surface, we can
focus on the set of triangles having at least one vertex on C. Lifting to a finite cover of S
perhaps (in order to remove interesting topology nearby C that is irrelevant to computing
the residue) this union of triangles will either be a triangulated annulus (when C ⊂ ∂S
is a boundary component) or a once-punctured n-gon for some n (if C is a puncture).
We call this set of triangles the annnular neighborhood of C. Even when L consists of a
single curve, the intersection of L with this annular neighborhood might consist of several
curves. By the linearity of shear coordinates and residues, it suffices to consider the case
that L consists of a single curve in the annular neighborhood.
When C is a puncture, its annular neighborhood is a punctured disc with a triangulation
all of whose arcs are radii joining the puncture to the boundary of the disc. By inspection,
a curve L contributes nonzero residue at C if and only it spirals at C, and the value of this
residue is ±1 according to whether it spirals counterclockwise or clockwise respectively as
claimed.
When C is a even boundary component, we compute the residue of ⃗b(T, L) using the
right hand side of (8.1). We split up this right hand side into two terms by splitting
up ⃗l(T, L) as a concatenation of (l(γ, L)γ∈T and (l(β, L)β∈∂S , and performing the matrix
multiplication with B̃ in block form. The first term in this expression has zero residue
around C since it is a linear combination of the rows of B(T ). What’s left over is a sum
1
(8.3)
−
p(γ; C)l(β, L)Bβ,γ .
∑
2 γ∈T,β∈B(S,M)
28
CHRIS FRASER
We claim the sum above evaluates to p(L; C). For the sum to be nonzero, L must
have a nonzero end at some segment β = [vi−1 , vi ] with vi−1 , vi in clockwise order. This
segment β is contained in a unique triangle in T . Call the other two sides in this triangle
γi−1 and γi , whose endpoint on C is vi−1 and vi respectively. There are cases according to
whether either of these sides is a boundary segment. If neither is, then p(γi−1 ; C)Bβ,γi−1 =
p(γi ; C)Bβ,γi is ±1 according to whether vi is white or black. The total contribution
to (8.3) is p(C; L). In the degenerate case that γi is a boundary segment, it does not
contribute to (8.3), but p(γi−1 ; C) = 2 and this effect is cancelled out, and so on.
Proposition 8.4. Suppose (S, M) is not among the four listed exceptions in Proposition 7.2. Let L, L′ be multi-laminations on (S, M) and recall the submodules VL and VL′
from Theorem 7.5. Let g ∈ MG⋈ (S, M) be a tagged mapping class and πg the corresponding
signed permutation matrix. The following are equivalent:
● there is a quasi-homomorphism Ψ from A(S, M, L) to A(S, M, L′ ) whose map on
tagged triangulations is T ↦ g(T ) (that is, Ψ(Σ(T )) ∼ Σ(g(T ))),
● VL′ ⊂ πg (VL ).
Proof of Proposition 8.4. A quasi-homomorphism Ψ from A(S, M, L) to A(S, M, L′ ) is
determined by a pair of tagged triangulations T and T ′ , such that B̃(T, L) and B̃(T ′ , L′ )
are related as in (4.8). Since the principal parts of these matrices agree, by Proposition 7.2 there is a tagged mapping class g such that g(T ) = T ′ . Furthermore, for
each lamination L′ ∈ L′ , the vector ⃗b(T ′ , L′ ) must be in span({⃗b(T, L)∶ L ∈ L}). Since
⃗b(T ′ , L′ ) = ⃗b(T, g −1 (L′ )), by Lemma 8.2, it is equivalent to find a linear combination of
⃗b(T, g −1 (L′ )) and {⃗b(T, L)∶ L ∈ L} that has zero residue around every even component.
Proposition 8.4 follows now from Lemma (8.3) and the fact that g acts on a vector of
pairings by the matrix πg .
Proof of Lemma 8.2. Restating the Lemma, we seek to show that the RC form a basis for
the dual space to the row span.We begin by verifying each of these vectors pair to zero
with the row span.
First, we check this when C is a puncture. We begin with the case that all of the
arcs in T are untagged at C. We need to check that ∑γ∈T p(γ; C)B(T )γ ′ ,γ vanishes for
each γ ′ ∈ T . Indeed, letting L be the lamination consisting of a tiny simple closed curve
contractible to C, the shear coordinate vector ⃗b(T, L) is clearly 0. Now we apply (8.1)
for this choice of L: the γ ′th component of (8.1) says 0 = ∑γ∈T p(γ; C)B(T )γ,γ ′ as desired,
using the fact that l(γ, L) = 0 if γ is a boundary segment. The argument when all arcs
are tagged at C is identical. If C is incident to exactly two arcs, namely the plain and
tagged version of the same arc, then RC follows from [12, Definition 9.6] (or a calculation
in a once-punctured digon).
Second we check this when C ⊂ ∂S is a boundary component. Number the cilia on C
by v1 , . . . , v2m . For each i ∈ [1, 2m], let Li be a tiny lamination contractible to vi – its
two endpoints are on the two boundary segments adjacent to vi . Again, ⃗b(T, Li ) is clearly
0 and in particular ∑vi black ⃗b(T, Li ) = ∑vi white ⃗b(T, Li ). Summing over the corresponding
right hand sides of (8.1), again performing the matrix multiplication in (8.1) in block form
as in the argument for Lemma 8.3, the terms corresponding to boundary segments are
present in both the sum over black vi and the sum over white vi . Canceling these common
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
29
terms, we get the equality ∑vi black, γ ∈ T l(γ, Li )B(T )γ,γ ′ = ∑vi white l(γ, Li )B(T )γ,γ ′ for all
γ ′ , which says ∑γ∈T p(γ, C)B(T )γ,γ ′ = 0 for all γ ′ as desired.
Thus all of the RC pair to zero with the row span of B(T ). We will now show that they
are linearly independent, which completes the proof since they have the expected size by
[12, Theorem 14.3].
Consider a linear relation of the form
(8.4)
∑ aC RC = 0.
We define scalars av for all marked points v ∈ M as follows: if v is a puncture C, then
bv = aC . If v is a cilium residing on an even component C, then bv = ±aC , with ± sign
consistent with the black-white coloring on C. If v is cilium on an odd component, we
set av = 0.
Now consider any vertices v1 , v2 forming an edge in the triangulation T . We claim
(8.5)
av1 + av2 = 0.
Indeed, if v1 , v2 are the endpoints of an arc γ ∈ T , the γ th component of the relation
(8.4) is av1 + av2 by construction, and (8.5) holds. If they are the endpoints of a boundary
segment, then (8.5) clearly holds.
However, in any given triangle in T with vertices v1 , v2 , v2 , the only way for (8.5) to
hold for all 3 of the pairs (v1 , v2 ), (v1 , v3 ), (v2 , v3 ) is if av1 = av2 = av3 = 0. Varying the
vertex and triangle containing it, this establishes that all av = 0 for all v ∈ M, and hence
all aC = 0, as desired.
9. Appendix: The starfish lemma on a nerve
We give the appropriate generalization of the Starfish Lemma [11, Proposition 3.6] from
a star neighborhood to a nerve. Our proof follows the proof of the Starfish Lemma in
[14], with appropriate modifications.
Let R be a domain. We say two elements r, r ′ ∈ R are coprime if they are not contained
in the same prime ideal of height 1. When R is a unique factorization domain, every pair
of non-associate irreducible elements are coprime.
Proposition 9.1. Let N be a nerve in Tn . Let R be a C-algebra and a Noetherian normal
domain. Let E be a seed pattern of geometric type, satisfying the following:
● all frozen variables are in R
● for each vertex t ∈ N , the cluster x(t) ⊂ R, and the cluster variables x ∈ x(t) are
pairwise coprime elements of R;
k
● for each edge t Ð
→ t′ in N , the cluster variables xk (t) and xk (t′ ) are pairwise
coprime.
Then the cluster algebra A defined by E satisfies A ⊂ R.
The proof relies on the following two lemmas, the first of which is a standard fact from
commutative algebra. For a prime ideal P , let RP = R[(R/P )−1 ] denote the localization
of R away from P .
Lemma 9.2 ([25, Theorem 11.5]). For a normal Noetherian domain R, the natural inclusion R ⊂ ∩ht P = 1 RP (intersection over height one primes) is an equality.
30
CHRIS FRASER
Lemma 9.3. With hypotheses as in Proposition 9.1, let P be a height one prime ideal in
R. Then at least one of the products
(9.1)
∏
x
x∈x(t),t∈N
is not in P .
Proof. By the coprimeness in each cluster t ∈ N , at most one of the cluster variables x in
a product (9.1) satisfies x ∈ P . We will show that for at least one t, none of the cluster
variables is in P , establishing our claim since P is prime. Pick any vertex t0 ∈ N , and
j
suppose the cluster variable xi ∈ P . Given an edge t0 Ð
→ t′0 ⊂ N where j ≠ i, the cluster
variable xj (t′0 ) ∉ P by the coprimality assumption in the cluster at t′0 . Repeatedly applying
this assumption while mutating along the nerve, using the connectedness hypothesis and
the fact that every edge label shows up in the nerve, we finally arrive at a vertex t ∈ N
i
such that the edge t Ð
→ t′ ⊂ N , and all of the extended cluster variables xj ∈ x̃(t) with
j ≠ i are not in P . By the coprimeness assumption along edge i, we see xi (t′ ) ∉ P , and
the cluster at t′ is one where the product (9.1) is not in P .
Proof of Proposition 9.1. We need to prove each cluster variable z is in R. By Lemma
9.2, it suffices to show z ∈ RP for any height one prime P . Indeed, by Lemma 9.3 there
is a cluster t ∈ N such that ∏x∈x(t) x ∉ P . By the Laurent Phenomenon, z is a Laurent
polynomial in the elements of x(t), with coefficients in C[xn+1 , . . . , xn+m ]. In particular,
z ∈ RP , as desired.
10. Appendix: Grassmannians and Band Matrices
As an illustration of Proposition 4.3, we extend the constructions in Example 3.4 and
Example 3.8 from the case (k, n) = (2, 5) to general (k, n). Let
(10.1)
X = Ĝr(n − k, n)
(10.2)
Y ≅ C(n−k)(k+1)
be the affine cone over the Grassmannian of (n − k)-dimensional subspaces of Cn . Its
points are the decomposable tensors in Λn−k Cn . Let
be the affine space of (n − k) × n band matrices of width k + 1, i.e. the set of matrices Y
whose entries yi,j are zero unless i ≤ j ≤ i + k.
We will describe a quasi-isomorphism of the coordinate rings C[X] and C[Y], and in
particular a cluster structure on C[Y] which appears to be new.
The coordinate ring C[X] is the ring generated by the Plücker coordinates ∆S as S
ranges over (n − k)-subsets of n. It has a cluster structure of geometric type cf. [28] in
which the frozen variables are those Plücker coordinates consisting of cyclically consecutive
columns. The non-frozen Plücker coordinates are all cluster variables. We will introduce
a useful sign convention: if S is any set of (n − k) natural numbers, we let ∆S denote the
Plücker coordinate obtained by first reducing all the elements of S to their least positive
residue modulo n, sorting these residues, and then taking the corresponding Plücker
coordinate. If there are fewer than (n − k) distinct elements in S modulo n, then ∆S is
identically zero.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
31
The coordinate ring C[Y] contains minors YI,J for I ⊂ [1, n − k], J ⊂ [1, n] subsets of
the same size denoting row and column indices respectively. It is a polynomial ring in the
coordinate functions Yi,j , 1 ≤ i ≤ j ≤ i + k ≤ n. The following elements will serve as frozen
variables in C[Y]:
(10.3)
Y1,1 , Y2,2 , . . . , Yn−k,n−k and Y1,k+1 , Y2,k+1 , . . . , Yn−k,n , and
(10.4)
Y[1,n−k],[2,n−k+1] , Y[1,n−k],[3,n−k+2] , . . . , Y[1,n−k],[k,n−1] ;
we let P denote the corresponding tropical semifield in these frozen variables.
Just as in Example 3.4, there is a morphism of varieties F ∶ Y → X sending Y ∈ Y to
the decomposable tensor Y [1] ∧ ⋯ ∧ Y [n − k] ∈ X, where the Y [i] are the rows of Y . The
map on coordinate rings is
(10.5)
F ∗ ∶ C[X] → C[Y], defined by F ∗ (∆S ) = Y[1,n−k],S .
Letting ∆S be any non-frozen Plücker coordinate, one sees that
(10.6)
F ∗ (∆S ) = c(S) ⋅ YI(S),J(S)
where c(S) ∈ P and YI(S),J(S) is a non-frozen irreducible row-solid minor in C[Y]. The
map ∆S ↦ YI(S),J(S) is a bijection between the non-frozen Plücker coordinates in C[X]
and the non-frozen irreducible row-solid minors in C[Y].
Just as in Example 3.8, there is a morphism of varieties G∶ X → Y sending X ∈ X to
the band matrix whose whose (i, j) entry is a certain Plücker coordinate evaluated on X:
(10.7)
G(X)i,j = ∆[i+k+1,n+i−1]∪j (X).
Since the Plücker coordinate on the right hand side of (10.7) is only nonzero when
unless i ≤ j ≤ i + k, G(X) is indeed a point in Y. The map on coordinate rings is
(10.8)
G∗ (Yi,j ) = ∆[i+k+1,n+i−1]∪j .
Theorem 10.1. Let X and Y be the varieties in (10.1) and (10.2). Let F >0 be the
semfield of subtraction-free expressions in the Yi,j . Abusing notation, let F ∗ ∶ F>0 → F >0
be the semfield map determined by the map F ∗ from (10.5), and let G∗ ∶ F >0 → F>0 be the
map determined by (10.8). Then C[Y] is a cluster algebra of geometric type, and the
maps F ∗ and G∗ are quasi-inverses. All of the irreducible row-solid minors in C[Y] are
cluster or frozen variables.
Proof. The proof relies on the following well-known properties of the cluster structure on
C[X]:
(1) There exist clusters in C[X] consisting entirely of Plücker coordinates (called
Plücker clusters). Every non-frozen Plücker coordinate shows up in at least one
of these clusters.
(2) The set of Plücker clusters are connected to one another by mutations whose
exchange relations are short Plücker relations, of the form ∆S∪ik ∆Sjℓ = ∆Sij ∆Skℓ +
∆Sjk ∆Siℓ , for i < j < k < l and Sij denotes the union S ∪ {i, j}).
(3) For certain Plücker clusters, every neighboring cluster is again a Plücker cluster.
The first two of these facts are consequences of the technology of plabic graphs and
square moves, the third fact follows by considering a particular explicit Plücker seed for
the Grassmannian (one whose quiver is a “grid quiver”).
32
CHRIS FRASER
We can now state Theorem 3.4 more carefully. Let Σ be any Plücker cluster in C[X].
For each Plücker coordinate xi ∈ Σ, let xi be the irreducible row-solid minor in C[Y]
related to xi as in (10.6). We will see below that {xi ∶ xi ∈ Σ} are algebraically independent
generators for F >0 over P. Assuming this has been proved, applying the construction in
Proposition 4.3, we obtain a semifield map cΣ ∶ F>0 → P and a normalized seed F ∗ (Σ)
∗ (x )
i
whose cluster variables are xi = FcΣ (x
as in (4.5). The semifield map cΣ satisfies cΣ (∆S ) =
i)
c(S) for all S ∈ Σ, where c(S) is defined by (10.6). The key claim is that in fact cΣ (∆S ) =
c(S) holds for all ∆S , and therefore the semifield map cΣ does not depend on Σ.
From this key claim, it follows that the seeds F ∗ (Σ) are all related to each other
by mutation using Proposition 4.3. Furthermore, every non-frozen irreducible row-solid
minor is a cluster variable in E by (4.5). Since this includes all of the Yi,j , this shows
that the F ∗ (Σ) are indeed seeds – each seed has the expected size necessary to form a
transcendence basis for C(Y), the seeds are all related to each other by mutation, and
their union clearly contains a generating set for the field of fractions. The cluster algebra
for the resulting seed pattern clearly contains C[Y]. The opposite containment follows
from the Algebraic Hartogs’ argument on a starfish cf. Section 9, using Fact (3) above.
Thus it remains to check the key claim that cΣ (∆S ) = c(S) for all non-frozen S. By
Fact (2), it suffices to check that c preserves the short Plücker relations, i.e.
(10.9)
c(Sik)c(Sjℓ) = c(Sij)c(Skℓ) ⊕ c(Sjk)c(Siℓ).
Verifying (10.9) is a direct piecewise check: the exponent of Ya,a in the left hand side of
(10.9) is 0, 1, 2 according to whether neither, one of, or both of Sik and Sjℓ contain the
interval [1, a]. Performing the similar computation for Sij and Skℓ, as well as Sjk and
Siℓ, and taking the minimum of their respective answers, gives the exponent of Ya,a in
the right hand side, and we claim these left and right hand exponents are always equal.
This can be done by a case analysis: let E be the largest number such that [1, E] ⊂ Sijkl.
If E < i, then both sides return a 2 if a ≤ E, and 0 otherwise. If i ≤ E < j, then both
sides return a 2 if a < i, return a 1 if i ≤ a ≤ E, and return a 0 otherwise. If j ≤ E then
both sides return a 2 if a < i, return a 1 if i ≤ a < j, and return a 0 otherwise. A similar
calculation checks that the exponents of Ya,a+k match up in both sides of (10.9).
Finally we check that G∗ is a quasi-inverse to F ∗ . By Lemma 3.7, we only need to see
that G∗ preserves coefficients and that G∗ ○F ∗ is proportional to the identity. It suffices to
check that G∗ ○ F ∗ (∆S ) ≍ ∆S for every ∆S . This follows from the determinantal identity
Lemma 10.2 below, applied to G∗ ○ F ∗ (∆S ) = G∗ (Y[1,n−k],S ).
Lemma 10.2. Let I = [a, a + s − 1] be some consecutive subset of [n − k], and J a subset
of [a, a + s − 1 + k] of size s. Then for X ∈ C[X],
YI,J (G(X)) = ( ∏ ∆[i+k+1,n+i] (X)) ⋅ ∆[a+k+s,n+a−1]∪J (X)
a+s−2
(10.10)
i=a
Notice that the first product on the right hand side of (10.10) is a monomial in the
frozen Plücker coordinates.
QUASI-HOMOMORPHISMS OF CLUSTER ALGEBRAS
33
Proof. Proceed by induction on s. It’s clear when s = 1. For s > 1, we will need a Plücker
relation
(10.11)
∆[a+k+1,n+a] (X)∆[a+k+s,n+a−1]∪J (X) = ∑(−1)ℓ+1 ∆[a+1+k,n+a−1]∪jℓ (X)∆[a+k+s,n+a]∪(J−jℓ ) (X),
s
ℓ=1
see e.g. [19, Section 9.1, Exercise 1]. Let J = {j1 , . . . , js } with j1 < j2 < ⋅ ⋅ ⋅ < js . Assuming
(10.10) holds for smaller values of s, we expand along the first row to see
YI,J (G(X)) = ∑(−1)ℓ+1 Ya,jℓ (G(X))Y(I−a),(J−jℓ ) (G(X))
s
ℓ=1
s
= ∑(−1)ℓ+1 ∆[a+k+1,n+a−1] (X)( ∏ ∆i+k+1,n+i (X)) ⋅ ∆[a+k+s,n+a]∪(J−jℓ )
ℓ=1
a+s−2
a+s−2
i=a+1
= ( ∏ ∆i+k+1,n+i (X)) ∑(−1)ℓ+1 ∆[a+1+k,n+a−1]∪jℓ (X)∆[a+k+s,n+a]∪(J−jℓ ) (X),
i=a+1
s
ℓ=1
and the result follows using (10.11).
Remark 10.3. In the case k = 2, our construction is the “motivating example” considerd
by Yang and Zelevinsky [30]. They establish that the homogeneous coordinate ring of
a certain SLn+1 -double Bruhat cell is a Dynkin type An cluster algebra with principal
coefficients. The elements of this double Bruhat cell are (n + 1) × (n + 1) band matrices
of width 3. Their example follows from ours by setting certain frozen variables equal to
1, and setting Y1,1 and Yn−k,n equal to 0 (this latter operation is permissible because both
of these frozen variables are isolated vertices in the quivers for C[Y]).
We also remark that it is already known that the Grassmannian cluster algebras are
quasi-isomorphic to a polynomial ring, by a fairly uninteresting quasi-isomorphism. Indeed, we can realize the affine space of (n − k) × k matrices as the closed subvariety of
Ĝr(n − k, n) defined by specializing the frozen variable ∆[1,n−k] to 1, and this specialization is a quasi-isomorphism. The resulting cluster structure on the polynomial ring is
unrelated to the one we have given in this section.
References
[1] I. Assem, G. Dupont, R. Schiffler. On a category of cluster algebras. J. Pure Appl. Algebra 218
(2014), no. 3, 553–582.
[2] I. Assem, R. Schiffler, V. Shramchenko. Cluster Automorphisms, Proc. Lond. Math. Soc. (3) 104
(2012), no. 6, 1271–1302.
[3] A. Berenstein, S. Fomin, and A. Zelevinsky, Cluster algebras. III. Upper bounds and double Bruhat
cells, Duke Math. J. 126 (2005), no. 1, 1–52.
[4] T. Bridgeland, I. Smith, Quadratic differentials as stability conditions, Publ. Math. Inst. Hautes tudes
Sci. 121 (2015), 155278.
[5] T. Brüstle, Y. Qiu, Tagged mapping class groups: Auslander-Reiten translation, Math. Z. 279 (2015),
no. 3-4, 1103–1120.
[6] W. Chang, B. Zhu Cluster automorphism groups and automorphism groups of exchange graphs,
arXiv:1506.02029.
[7] W. Chang, B. Zhu Cluster automorphism groups of cluster algebras with coefficients,
arXiv:1506.01950.
[8] B. Farb and D. Margalit, A Primer on Mapping Class Groups, Princeton Mathematical Series, 49.
Princeton University Press, Princeton, NJ, 2012.
34
CHRIS FRASER
[9] V. Fock and A. Goncharov, Cluster ensembles, quantization and the dilogarithm, math.AG/0311245.
Ann. Sci. Éc. Norm. Supér. (4) 42 (2009), no. 6, 86
[10] V. V. Fock and A. B. Goncharov, Moduli spaces of local systems and higher Teichmüller Theory,
Publ. Math. Inst. Hautes. Etudes. Sci. no. 103 (2006) 1-211.
[11] S. Fomin and P. Pylyavskyy, Tensor diagrams and cluster algebras,237 arXiv:1210.1888, to appear
in Adv. in. Math.
[12] S. Fomin, M. Shapiro, and D. Thurston, Cluster algebras and triangulated surfaces. Part I: Cluster
complexes, Acta Math. 201 (2008), no. 1, 83–146.
[13] S. Fomin and D. Thurston, Cluster algebras and triangulated surfaces. Part II: Lambda lengths,
arXiv:1210.5569.
[14] S. Fomin, L. Williams, and A. Zelevinsky, Introduction to cluster algebras, in preparation.
[15] S. Fomin and A. Zelevinsky, Cluster algebras I: Foundations, J. Amer. Math. Soc. 15 (2002), no. 2,
497–529
[16] S. Fomin and A. Zelevinsky, Cluster algebras II: Finite type classification, Invent. Math. 154 (2003),
no. 1, 63–121.
[17] S. Fomin and A. Zelevinsky, Cluster algebras IV: Coefficients, Compos. Math. 143 (2007), no. 1,
112–164.
[18] C. Fraser, in preparation.
[19] W. Fulton, Young tableaux. With applications to representation theory and geometry. London Mathematical Society Student Texts, 35. Cambridge University Press, Cambridge, 1997.
[20] M. Gehktman, M. Shapiro, and A. Vainshtein, Cluster algebras and Poisson geometry, Mosc. Math. J.
3 (2003), no. 3, 899-934.
[21] J. E. Grabowski, Graded cluster algebras, (2013), arXiv:1309.6170.
[22] J. E. Grabowski, S. Launois, Graded quantum cluster algebras and an application to quantum
Grassmannians, Proc. Lond. Math. Soc. 109 (2014), no. 3, 697–732.
[23] G. Cerulli Irelli, B. Keller, D. Labardini-Fragoso, P.–G. Plamondon, Linear independence of cluster
monomials for skew-symmetric cluster algebras, Compos. Math. 149 (2013), no. 10, 1753–1764.
[24] R. J. Marsh, J. Scott, Twists of Plücker coordinates as dimer partition functions, arXiv:1309.6630.
[25] H. Matsumura, Commutative ring theory, second ed., vol. 8 of Cambridge Studies in Advanced
Mathematics. Cambridge University Press, Cambridge, 1989. Trans- lated from the Japanese by M.
Reid.
[26] N. Reading, Universal Geometric Cluster Algebras, Math. Z. 277 (2014), no. 1-2, 499–547.
[27] N. Reading, Universal geometric cluster algebras from surfaces, Trans. Amer. Math. Soc. 366 (2014),
no. 12, 6647–6685.
[28] J. Scott, Grassmannians and cluster algebras, Proc. London Math. Soc. 92 (2006) 345-380.
[29] P. Sherman, A. Zelevinsky, Positivity and canonical bases in rank 2 cluster algebras of finite and
affine types Moscow Math. J. 4 (2004) 947-974.
[30] S.W. Yang and A. Zelevinsky, Cluster algebras of finite type via Coxeter elements and principal
minors, Transform. Groups 13 (2008) 855-895,
Department of Mathematics, University of Michigan, Ann Arbor, MI, 48109, USA
E-mail address: [email protected]
| 0math.AC
|
ENUMERATING GRADED IDEALS IN GRADED RINGS
ASSOCIATED TO FREE NILPOTENT LIE RINGS
arXiv:1606.04515v1 [math.RA] 14 Jun 2016
SEUNGJAI LEE AND CHRISTOPHER VOLL
Abstract. We compute the zeta functions enumerating graded ideals in the graded
Lie rings associated with the free d-generator Lie rings fc,d of nilpotency class c for all
c ď 2 and for pc, dq P tp3, 3q, p3, 2q, p4, 2qu. We apply our computations to obtain information about p-adic, reduced, and topological zeta functions, in particular pertaining
to their degrees and some special values.
1. Introduction
1.1. Enumerating graded ideals in graded Lie rings. Let R be the ring of integers
of a number field or the completion of such a ring at a nonzero prime ideal. Let L be
a nilpotent R-Lie algebra of nilpotency class c, free of finite rank over R, with lower
central series pγi pLqqci“1 . For i P t1, . . . , cu, set Lpiq :“ γi pLq{γi`1 pLq. The associated
À
graded R-Lie algebra is grL “ ci“1 Lpiq . An R-ideal I of grL (of finite index in grL)
is graded or homogeneous if it is generated by homogeneous elements or, equivalently,
À
if I “ ci“1 pI X Lpiq q. In this case we write I Ÿgr grL. We define the graded ideal zeta
function of L as the Dirichlet generating series
ÿ
Ÿ
|grL : I|´s ,
(1.1)
ζL gr psq “
IŸgr grL
enumerating graded ideals in grL of finite index in grL. Here s is a complex variable;
Ÿ
our assumptions on L guarantee that ζL gr psq converges on a complex half-plane. (For a
similar definition in a slightly more general setting, see [11, Section 3.1]).
Assume now that R “ O is the ring of integers of a number field. For a (nonzero)
prime ideal p P SpecpOq we write Op for the completion of O at p, a complete discrete
valuation ring of characteristic zero and residue field O{p of cardinality q, say. Primary
decomposition yields the Euler product
ź
Ÿgr
Ÿ
psq,
ζLpO
ζL gr psq “
pq
pPSpecpOq
Ÿ
ζL gr psq
as an infinite product of “local”, or p-adic
expressing the “global” zeta function
ones. By slight abuse of notation we denote here by SpecpOq the set of nonzero prime
ideals of O. Each individual Euler factor is a rational function in the parameter q ´s . In
fact, (1.1) is an Euler product of cone integrals in the sense of [4] (cf. [11, Theorem 3.3]),
and the far-reaching results of this paper – regarding both the factors and the product’s
Date: June 15, 2016.
2000 Mathematics Subject Classification. 17B70, 11M41, 11S40.
Key words and phrases. Graded ideal zeta functions, free nilpotent Lie rings, local functional
equations.
2
SEUNGJAI LEE AND CHRISTOPHER VOLL
analytic properties – apply. The fact that the analysis of [4] is restricted to the case
R “ Z is insubstantial for this conclusion.
1.2. Main results. In the present paper we are concerned with graded ideal zeta functions of free nilpotent Lie rings of finite rank. Given c P N and d P Ně2 , let fc,d be the free
nilpotent Lie ring of nilpotency class c on d Lie generators. One may identify fc,d with
the quotient of the free Z-Lie algebra fd on d generators by the c ` 1-th term γc`1 pfd q
of its lower central series. For i P t1, . . . , cu, the Z-rank of the i-th lower-central-series
quotient γi pfc,d q{γi`1 pfc,d q is given by the Witt function
1ÿ
µpjqdi{j ,
(1.2)
Wd piq :“
i
j|i
ř
where µ denotes the Möbius function; cf. [18, Satz 3]. Hence rkZ pfc,d q “ ci“1 Wd piq.
Given a commutative ring R – in this paper always of the form O or Op as above – we
write fc,d pRq “ fc,d bZ R, considered as an R-Lie algebra.
For c “ 1, the (graded ideal) zeta function of the free abelian Lie ring f1,d pOq, enumerating all finite index O-sublattices of Od , is well known to be equal to
Ÿ
ζf1,dgrpOq psq “
(1.3)
d
ź
ζK ps ´ i ` 1q,
i“1
where ζK denotes the Dedekind zeta function of K; cf. [7, Proposition 1.1]. The Euler
ś
product (1.1) reflects the well-known Euler product ζK psq “ pPSpecpOq p1 ´ |O{p|´s q´1 :
(1.4)
Ÿ
ζf1,dgrpOq psq “
ź
pPSpecpOq
d
ź
1
“
1 ´ |O{p|i´1´s
i“1
ź
ζOdp psq, say.
pPSpecpOq
For c “ 2, the second author computed in [15] the ideal zeta functions enumerating
(O-)ideals of finite index in the rings f2,d pOq. (The paper only discusses the case R “ Z,
but its computations carry over – mutatis mutandis – to the case of general number
Ÿ
rings.) We compute the graded ideal zeta functions ζf2,dgrpOq psq for all d ě 2 in Section 4
and those for pc, dq P tp3, 2q, p4, 2qu in Section 5 of the current paper.
The paper’s most involved result is the computation, in Section 3, of the graded ideal
Ÿ
zeta function of f3,3 pOq. To this end we compute an explicit formula for ζf3,3grpoq psq, valid
for all finite extension o of the p-adic integers Zp , where p is a prime, viz. a local ring of
the form o “ Op for a nonzero prime ideal p of O lying above p.
Ÿ
Theorem 1.1. There exists an explicitly determined rational function W3,3gr P QpX, Y q
such that, for all primes p and all finite extensions o of Zp , with residue cardinality q,
Ÿ
Ÿ
ζf3,3grpoq psq “ W3,3gr pq, q ´s q.
Ÿ
It may be written as W3,3gr “ N3,3 {D3,3 where
D3,3 pX, Y q “p1 ´ Y qp1 ´ XY qp1 ´ X 2 Y qp1 ´ X 3 Y 4 qp1 ´ XY 5 qp1 ´ X 2 Y 5 qp1 ´ Y 6 q
p1 ´ X 7 Y 7 qp1 ´ X 8 Y 7 qp1 ´ X 8 Y 8 qp1 ´ X 12 Y 8 qp1 ´ X 6 Y 9 qp1 ´ X 15 Y 9 q
p1 ´ X 16 Y 10 qp1 ´ X 15 Y 11 qp1 ´ X 12 Y 12 qp1 ´ X 7 Y 13 qp1 ´ Y 14 q,
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
3
a polynomial of degree 115 in X and 131 in Y , and N3,3 P QrX, Y s is a polynomial of
degree 81 in X and 108 in Y .
Ÿ
The rational function W3,3gr satisfies the functional equation
Ÿ
Ÿ
W3,3gr pX ´1 , Y ´1 q “ X 34 Y 23 W3,3gr pX, Y q.
(1.5)
We note that the Witt function W
`33˘ (cf.`3(1.2))
˘ `8˘takes the values pW3 p1q, W3 p2q, W3 p3qq “
p3, 3, 8q and that 115 ´ 81 “ 34 “ 2 ` 2 ` 2 and 131 ´ 108 “ 23 “ 3 ¨ 3 ` 2 ¨ 3 ` 1 ¨ 8;
cf. Conjecture 6.2.
Ÿ
Our proof of Theorem 1.1 yields W3,3gr as a sum of 15 explicitly given summands, listed
essentially in Section 3.2. We do not reproduce the “final” outcome of this summation
Ÿ
here, as the numerator N3,3 of W3,3gr fills several pages. We do record, however, several
Ÿ
corollaries of the explicit formula for W3,3gr .
Ÿ
The first corollary concerns analytic properties of the global zeta function ζf3,3grpOq psq.
Ÿ
Corollary 1.2. The global graded ideal zeta function ζf3,3grpOq psq converges on ts P C |
Repsq ą 3u and may be continued meromorphically to ts P C | Repsq ą 14{9u.
ś
This follows from two observations. Firstly, the product pPSpecpOq D3,3 pq, q ´s q´1 – a
product of finitely many translates of the Dedekind zeta function ζK psq – has abscissa
of convergence 3 and may be continued meromorphically to the whole complex plain.
ś
Secondly, the product pPSpecpOq N3,3 pq, q ´s q may be continued meromorphically to ts P
ř
C | Repsq ą 14{9u. Indeed, if N3,3 pX, Y q “ 1 ` iPI αi X ai Y bi for a finite index set
ś
´s
I and αi P Zzt0u, ai P N0 , bi P N, then
pPSpecpOq N3,3 pq, q q may be continued
ai
meromorphically to ts P C | Repsq ą βu, where β :“ maxt bi | i P Iu; cf. [5, Lemma 5.5].
That β “ 14{9 follows from inspection of N3,3 .
The second corollary concerns the reduced graded ideal zeta function
Ÿ
Ÿ
ζf3,3gr,red pY q :“ W3,3gr p1, Y q P QpY q;
cf. Section 2.4. This concept was introduced in [6], albeit not in the context of graded
ideal zeta functions. We expect that our ad hoc definition will fit into a general definition
of reduced graded ideal zeta functions along the lines of [6].
Ÿ
Corollary 1.3. The reduced graded ideal zeta function ζf3,3gr,red pY q satisfies
(1.6)
Ÿ
ζf3,3gr,red pY q “
N3,3,red pY q
p1 ´ Y q3 p1 ´ Y 3 qp1 ´ Y 4 q2 p1 ´ Y 5 q
ś14
i“8 p1
´ Y iq
,
4
SEUNGJAI LEE AND CHRISTOPHER VOLL
where N3,3,red pY q P ZrY s is equal to
1 ` 2Y 3 ` 4Y 4 ` 5Y 5 ` 16Y 6 ` 34Y 7 ` 53Y 8 ` 77Y 9 ` 98Y 10 ` 121Y 11 ` 182Y 12 `
302Y 13 ` 483Y 14 ` 712Y 15 ` 953Y 16 ` 1187Y 17 ` 1425Y 18 ` 1689Y 19 ` 2046Y 20 `
2579Y 21 ` 3298Y 22 ` 4162Y 23 ` 5059Y 24 ` 5826Y 25 ` 6398Y 26 ` 6894Y 27 `
7475Y 28 ` 8270Y 29 ` 9265Y 30 ` 10260Y 31 ` 11041Y 32 ` 11529Y 33 ` 11745Y 34 `
11798Y 35 ` 11811Y 36 ` 11811Y 37 ` 11798Y 38 ` 11745Y 39 ` 11529Y 40 ` 11041Y 41 `
10260Y 42 ` 9265Y 43 ` 8270Y 44 ` 7475Y 45 ` 6894Y 46 ` 6398Y 47 ` 5826Y 48 `
5059Y 49 ` 4162Y 50 ` 3298Y 51 ` 2579Y 52 ` 2046Y 53 ` 1689Y 54 ` 1425Y 55 `
1187Y 56 ` 953Y 57 ` 712Y 58 ` 483Y 59 ` 302Y 60 ` 182Y 61 ` 121Y 62 ` 98Y 63 `
77Y 64 ` 53Y 65 ` 34Y 66 ` 16Y 67 ` 5Y 68 ` 4Y 69 ` 2Y 70 ` Y 73 .
Ÿ
It seems remarkable that W3,3gr p1, Y q has a simple pole at Y “ 1 of order 14 “ 3`3`8,
the Z-rank of f3,3 , and that N3,3,red has nonnegative, unimodal coefficients. This is
Ÿ
consistent with the speculation that W3,3gr p1, Y q is in fact the Hilbert-Poincaré series of
a 14-dimensional graded algebra, associated to f3,3 in a natural way; cf. Remark 6.5.
Note that the right hand side of (1.6) is not in lowest terms. The palindromic symmetry
of the coefficients of N3,3,red is implied by the functional equation (1.5).
The third corollary concerns another limiting object of the local graded ideal zeta
Ÿ
Ÿ
functions ζf3,3grpoq psq, viz. the topological graded ideal zeta function ζf3,3gr,top psq; cf. Section 2.5. In [9], Rossmann introduced and studied so-called topological zeta functions
associated to a range of Dirichlet generating series. Informally speaking, topological zeta
functions may be viewed as suitably defined limits of local zeta functions. Whilst topoŸ
logical versions of zeta functions such as ζf3,3grpoq psq have not yet been studied specifically,
Ÿ
Ÿ
one may define ζf3,3gr,top psq as the coefficient of pq ´ 1q´14 in the expansion of ζf3,3grpoq psq
in q ´ 1; cf. [9, Definition 5.13].
Ÿ
Corollary 1.4. The topological graded ideal zeta function ζf3,3gr,top psq satisfies
Ÿ
ζf3,3gr,top psq “
p33250s4 ´ 81537s3 ` 66573s2 ´ 20800s ` 1920q{56
D3,3,top psq,
where D3,3,top psq P Zrss is defined to be
s3 ps´1q4 ps´2qp3s´2qp4s´3qp5s´1qp5s´2qp7s´8qp2s´3qp3s´5qp5s´8qp11s´15qp13s´7q
A number of further features of the p-adic, topological, and reduced graded ideal zeta
functions associated to f3,3 considered above seem remarkable. Some of the numerical
values in the following corollary are presented so as to illustrate the general conjectures
in Section 6 which we extracted from the paper’s explicit computations.
By the degree of a rational function f “ P {Q P QpY q in Y we mean degY f “
degY P ´ degY Q.
Corollary 1.5.
(1)
¯
´
Ÿ
degs ζf3,3gr,top psq “ ´14 “ ´p3 ` 3 ` 8q “ ´ rkZ pf3,3 q,
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
5
(2)
Ÿ
s´14 ζf3,3gr,top ps´1 q|s“0 “
19
Ÿ
“ p1 ´ Y q14 ζf3,3gr,red pY q|Y “1 P Qą0 ,
288, 288
(3)
Ÿ
s3 ζf3,3gr,top psq|s“0 “
(4)
´1
p´1q2`2`7 3 ¨ 3 ¨ 8
“
,
70, 560
3 ¨ 6 ¨ 14 ¨ 2! ¨ 2! ¨ 7!
ˇ
ˇ
ˇ
ˇ
ζo3 psqζo3 psqζo8 psq ˇ
Ÿ
ζf3,3grpoq psq
s“0
“
2
3¨3¨8
“
.
7
3 ¨ 6 ¨ 14
If the above-mentioned connection between the reduced graded ideal zeta function
and Hilbert-Poincaré series of graded algebras were to hold, the rational number in part
(2) – for which we currently do not have an interpretation – were to be interpreted in
terms of the multiplicity of the associated graded algebra; cf., for instance, [1, Section 4].
1.3. Background, motivation, and methodology. Recall that L is a nilpotent Lie
algebra over the ring of integers O of a number field. The ideal zeta function of L is the
Dirichlet generating series enumerating the O-ideals of L of finite index, viz.
ÿ
ź
Ÿ
|L : I|´s “
ζLpO
ζLŸ psq “
psq,
pq
IŸL
pPSpecpOq
where s is a complex variable and, for a (nonzero) prime ideal p of O,
ÿ
Ÿ
ζLpO
psq “
|LpOp q : I|´s
pq
IŸLpOp q
enumerates Op -ideals of LpOp q. Ideal zeta functions of nilpotent Lie rings are well-studied
relatives of the graded ideal zeta functions studied in the present paper. One of the main
results of [7], which introduced the former zeta functions, establishes the rationality of
Ÿ
each of the Euler factors ζLpO
psq in q ´s , where q “ |O : p|; cf. [7, Theorem 3.5]. For
pq
numerous examples of ideal zeta functions of nilpotent Lie rings, see [5].
By the Mal’cev correspondence, ideal zeta functions of nilpotent Lie rings are closely
related to the normal subgroup zeta functions enumerating finite index normal subgroups
of finitely generated nilpotent groups. In particular, given c, d P N, almost all (i.e. all but
finitely many) of the Euler factors of ζfŸc,d psq coincide with those of the normal subgroup
zeta function of the free class-c-nilpotent d-generator group Fc,d ; cf. [7, Section 4]. The
study of the normal subgroup growth of free nilpotent groups is connected with the
enumeration of finite p-groups up to isomorphism; cf. [3].
Informally speaking, graded ideal zeta functions of nilpotent Lie rings may be seen as
Ÿgr
psq
“approximations” of their ideal zeta functions. Indeed, almost all Euler factors ζLpO
pq
Ÿ
actually enumerate a sublattice of the lattice of ideals enumerated by ζLpO
psq.
In
pq
general, this approximation is quite coarse. In nilpotency class c ď 2, however, the
problems of computing ideal zeta functions and graded ideal zeta functions are closely
related, as the following example shows.
6
SEUNGJAI LEE AND CHRISTOPHER VOLL
Example 1.6. Assume that L is nilpotent of class 2, with isolated commutator ideal L1
such that rkO pL{L1 q “ d. Then, essentially by [7, Lemma 6.1],
ÿ
ÿ
(1.7)
ζLŸ psq “
|L{L1 : Λ1 |´s
|L1 : Λ2 |d´s ,
Λ1 ďL{L1
whereas
(1.8)
Ÿ
ζL gr psq “
ÿ
Λ1 ďL{L1
rΛ1 ,LsďΛ2 ďL1
|L{L1 : Λ1 |´s
ÿ
|L1 : Λ2 |´s .
rΛ1 ,LsďΛ2 ďL1
In the special cases L “ f2,d , the proximity between (1.7) and (1.8) explains the proximity
between the explicit formulae recorded in Theorems 4.1 and 4.2 of the current paper.
In higher nilpotency classes, we are not aware of any such simple parallels. In the
realm of free nilpotent Lie rings of class greater than two, explicit computations of ideal
zeta functions seem all but unfeasible. In particular, we do not know of formulae for the
ideal zeta functions of the Lie rings f3,3 pOq and f4,2 pOq.
Numerous questions regarding ideal zeta functions have analogues regarding the “approximating” graded ideal zeta functions, and one may speculate that the latter are
easier to answer than the former. On p. 188 of [7], Grunewald, Segal, and Smith formulate, for example, a conjecture which would imply that, for any c, d P N there exists
Ÿ pX, Y q P QpX, Y q such that, for almost all p and all finite exa rational function Wc,d
Ÿ
Ÿ pq, q ´s q. This consequence is known to hold for c ď 2
tensions o of Zp , ζfc,d poq psq “ Wc,d
and pc, dq “ p3, 2q but wide open in general, including the cases pc, dq P tp3, 3q, p4, 2qu.
Our Conjecture 6.1 – which is verified by the explicit computations in the current paper
in particular for c “ 2 and pc, dq P tp3, 3q, p3, 2q, p4, 2qu – may be viewed as a “graded”
version of this conjecture.
In [9], Rossmann formulates a number of conjectures on certain special values of padic and topological zeta functions, pertaining in particular to ideal zeta functions of
nilpotent Lie rings. Our Conjectures 6.7, 6.9 and 6.11 are “graded” counterparts.
In [17, Theorem 4.4], the second author proved a local functional equation for the
generic Euler factors of the ideal zeta functions ζfŸc,d pOq psq upon inversion of the prime
for all c P N and d P Ně2 . Our results suggest that this phenomenon also appears for
graded ideal zeta functions of free nilpotent Lie rings; cf. Conjecture 6.2.
All our computations owe their feasibility to the fact that, for the parameter values
considered, viz. c “ 2 and pc, dq P tp3, 3q, p3, 2q, p4, 2qu, the enumeration of graded ideals
is equivalent to the enumeration of various flags of lattices in free o-modules which
depend only, and in a linear fashion, on the lattices’ elementary divisor types. Our
computations rely on a simple polynomial formula, due to Birkhoff, for the numbers of
o-submodules of given type in a finite o-module of given type; cf. Proposition 2.2. To
obtain closed formulae for the relevant graded ideal zeta functions we need to organize
the enumeration of infinitely many values of Birkhoff’s formula in a manageable way. We
meet this challenge by organizing pairs of partitions, encoding two lattices’ elementary
divisor types, by their overlap type, formally one of finitely many multiset permutations,
viz. words in two letters (each with multiplicity); cf. Section 3.1. For pc, dq “ p3, 3q, for
instance, we are led to consider 15 specific words of length 11 in the alphabet t1, 2u; cf.
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
7
Table 3.1. Restricting the relevant enumerations to a fixed word yields formulae which
are products of q-multinomial coefficients and so-called Igusa functions; cf. Section 2.2.
In Theorem 3.6 we compute such formulae for each of the 15 words in question. Similar
methodology has been applied before to compute ideal zeta functions of Heisenberg Lie
rings over number rings ([12]) and free nilpotent Lie rings of class 2 ([15]).
That we are able to reduce our computations to combinatorial considerations with
partitions is essentially owed to the fact that, for the parameters pc, dq considered, the
automorphism groups of the free Lie rings fc,d act transitively on the lattices of given
elementary divisor types in relevant sections of fc,d . This allows us to assume that the
relevant lattices are generated by multiples of elements of Hall bases; cf. Section 2.3.
That a lattice may always be generated by multiples of some linear basis follows, of
course, from the elementary divisor theorem; that (subsets of) Hall bases may be used
is a lucky consequence of the existence of “many” automorphisms, as we now explain.
The automorphism group of fc,d poq contains a copy of GLd poq which allows us to perform
arbitrary invertible o-linear transformations of Lie generators x1 , . . . , xd of fc,d poq. For
c “ 2, this observation – which was already exploited in the proof of [7, Theorem 2] –
in conjunction with Birkhoff’s formula is sufficient to compute the (graded) ideal zeta
function of fc,d poq; cf. Section 4. On the Lie commutators rxi , xj s, 1 ď i ă j ď n,
these linear transformations act via the exterior square representation. In general, the
Ź
image of the natural map 2 GLd poq Ñ GLpdq poq is rather small. The map is surjective,
2
however, if d ď 3, and an isomorphism if d “ 3. This facilitates our computations for
pc, dq P tp3, 3q, p3, 2qu. If d “ 2, then the above-mentioned copy of GL2 poq even induces
the full automorphism group of the o-module generated by the weight-3-commutators
rrx1 , x2 s, x1 s and rrx1 , x2 s, x2 s, which we exploit for pc, dq “ p4, 2q.
1.4. Notation. Given n P N “ t1, 2, . . . u, we write rns for t1, 2, . . . , nu. We write Ně2
for t2, 3, . . . , u Ď N. Given a subset I Ď N, we write I0 for I Y t0u. If I Ď rn ´ 1s, we
denote n ´ I “ tn ´ i | i P Iu. Given a, b P N0 , we denote the interval ta ` 1, . . . , bu
by sa, bs and the interval ta ` 1, . . . , b ´ 1u by sa, br. We write 2S for the power set
of a set S. The notation I “ ti1 , . . . , ih uă for a subset of I Ă N0 indicates that
i1 ă i2 ă ¨ ¨ ¨ ă ih . Similarly, pλ1 , . . . , λn qě P Nn0 denotes a partition with non-ascending
parts λ1 ě ¨ ¨ ¨ ě λn ě 0.
We write o for a compact discrete valuation ring of characteristic zero, i.e. a finite
extension of the ring Zp of p-adic integers or, equivalently, a ring of the form Op , the
completion of O, the ring of integers of a number field K, at a nonzero prime ideal p of O.
We write q for the cardinality of the residue field of o and p for its residue characteristic.
We set t “ q ´s , where s is a complex variable. ζK is the Dedekind zeta function of K.
2. General preliminaries
Let c P N, d P Ně2 , and set f “ fc,d be as above. The following is analogous to [7,
Lemma 6.1].
Lemma 2.1. Let p be a prime and o be a finite extension of Zp . For each Λ2 ď fpoqp2q
let
(2.1)
pXpΛ2 q ‘ fpoqp2q q{Λ2 “ Zppfpoqp1q ‘ fpoqp2q q{Λ2 q.
8
SEUNGJAI LEE AND CHRISTOPHER VOLL
Then
c
ź
ÿ
Ÿ
gr
psq “
ζfpoq
|fpoqpiq : Λi |´s
Λ1 ďfpoqp1q ,..., Λc ďfpoqpcq i“1
@iP s1,cs: rΛi´1 ,fpoqp1q sďΛi
(2.2)
ÿ
“ ζod psq
|fpoqp1q : XpΛ2 q|´s
c
ź
|fpoqpiq : Λi |´s .
i“2
Λ2 ďfpoqp2q ,..., Λc ďfpoqpcq
@iP s2,cs: rΛi´1 ,fpoqp1q sďΛi
Proof. A graded additive sublattice Λ ď gr fpoq determines and is determined by the
sequence pΛ1 , . . . , Λc q, where Λi :“ Λ X fpoqpiq for all i P rcs. We have Λ Ÿgr fpoq if and
only if rΛi´1 , fpoqp1q s ď Λi for all i P s1, cs. This proves the first equality. The second
follows from the definition of XpΛ2 q given in (2.1), noting that rΛ1 , fpoqp1q s ď Λ2 if and
only if Λ1 ď XpΛ2 q and that, for each Λ2 ď fpoqp2q ,
ÿ
|fpoqp1q : Λ1 |´s “ ζod psq|fpoqp1q : XpΛ2 q|´s .
Λ1 ďXpΛ2 q
2.1. Birkhoff ’s formula. Given a pair of partitions pσ, τ q, let αpτ, σ; qq denote the
number of torsion o-modules of type τ contained in a fixed torsion o-module of type σ.
Clearly αpσ, τ ; qq “ 0 unless τ ď σ. Notice that, for τ “ pτ1 , . . . , τn qě ,
pnq
αpτ1 , τ ; qq “ #tΛ ď on | on {Λ – ‘nj“1 o{pτj u.
The following explicit general formula for αpσ, τ ; qq is attributed to Birkhoff in [2].
Proposition 2.2 (Birkhoff). Let τ ď σ be partitions, with dual partitions τ 1 ď σ 1 . Then
ź 1 1 1 ˆσ 1 ´ τ 1 ˙
αpσ, τ ; qq “
P Zrqs.
q τk pσk ´τk q k 1 k`1
σk ´ τk1 q´1
kě1
2.2. Igusa functions and their functional equations.
Definition 2.3. ([12, Definition 2.5]) Let h P N. Given variables X “ pX1 , . . . , Xh q
and Y , we set
ÿ ˆh˙ ź Xi
1
P QpY, X1 , . . . , Xh q,
Ih pY ; Xq “
I Y iPI 1 ´ Xi
1 ´ Xh
IĎrh´1s
ÿ ˆh˙ ź Xi
Xh
˝
Ih pY ; Xq “
P QpY, X1 , . . . , Xh q.
I Y iPI 1 ´ Xi
1 ´ Xh
IĎrh´1s
We set I0 pY q “ I0˝ pY q “ 1 P QpY q.
1
Note that I1 pY ; Xq “ 1´X
and I1˝ pY ; Xq “
functional equations ([12, Proposition 4.2])
X
1´X .
We will make repeated use of the
h
Ih pY ´1 ; X´1 q “ p´1qh Xh Y ´p2q Ih pY ; Xq,
(2.3)
h
Ih˝ pY ´1 ; X´1 q “ p´1qh Xh´1 Y ´p2 q Ih˝ pY ; Xq.
h
h
Remark 2.4. The “symmetry centres” p´1qh Xh Y ´p2 q resp. p´1qh Xh´1 Y ´p2q in the functional equations (2.3) depend on Y and Xh , but not on X1 , . . . , Xh´1 .
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
9
Throughout this paper, we always substitute q ´1 for Y and hence write Ih pXq instead
of Ih pq ´1 ; Xq and Ih˝ pXq instead of Ih˝ pq ´1 ; Xq.
Igusa functions are ubiquitous in the theory of zeta functions of groups and rings.
Note, for instance, that
ζod psq “ Id ppq pd´jqj tj qjPrds q “ śd´1
1
j“0 p1
cf. (1.4).
´ q j tq
;
2.3. Hall bases for free nilpotent Lie rings. Let x1 , . . . , xd be Lie generators for
f “ fc,d . A Hall basis for fc,d (on x1 , . . . , xd ) is a Z-basis Hc,d for fc,d which may be
constructed by selecting inductively certain basic commutators in the xi ; see [8] for
ř
details. Note that |Hc,d | “ ci“1 Wd piq, where Wd denotes the Witt function (1.2).
In Table 2.1 we record some Hall bases which are relevant for the current paper.
Here we adopt the standard abbreviation xi1 xi2 xi3 . . . xir for left-normed commutators
r. . . rrxi1 , xi2 s, xi3 s, . . . , xir s. In the case that d “ 2 (resp. d “ 3), we write x, y (and z)
instead of x1 , x2 (and x3 q.
pc, dq
p1, dq
p2, dq
p3, 3q
p3, 2q
p4, 2q
Hall basis Hc,d
tx1 , . . . , xd u
tx1 , . . . , xd , xi xj | 1 ď i ă j ď du
tx, y, z, xy, xz, yz, xyx, xzx, yzx, xyy, xzy, yzy, xzz, yzzu
tx, y, xy, xyx, xyyu
tx, y, xy, xyx, xyy, xyxx, xyyx, xyyyu
pWd piqqiPrcs
pdq
` ˘
pd, d2 q
p3, 3, 8q
p2, 1, 2q
p2, 1, 2, 3q
Table 2.1. Some Hall bases
2.4. Reduced zeta functions. In [6], Evseev introduced a certain “limit as p Ñ 1” of
various local zeta functions. Informally speaking, the idea is to exploit the fact that the
coefficients of these generating functions enumerate Fq -rational points of constructible
sets. The desired limit is obtained by replacing these constructible sets by their EulerPoincaré characteristics. The resulting rational functions are called reduced zeta functions; cf. [6, Section 3]. This paper does not define reduced graded ideal zeta functions
explicitly, but provides a theoretical framework that allows this quite readily. If c, d are
Ÿ
Ÿ
Ÿ
such that ζfc,dgrpoq psq “ Wc,dgr pq, q ´s q for a rational function Wc,dgr pX, Y q P QpX, Y q, for alŸ
Ÿ
most all primes p and all finite extensions o of Zp , then ζfc,dgr,red psq “ Wc,dgr p1, Y q P QpY q.
Evseev proved in [6, Section 4] that reduced ideal zeta functions of Lie rings with socalled nice and simple bases are Hilbert-Poincaré series enumerating integral points in
rational polyhedral cones. The Hall bases H2,d given in Table 2.1, for instance, have
this property. Comparing (1.7) and (1.8) one sees that, for nilpotent Lie rings L of
Ÿgr
Ÿ
pY q “ ζL,red
pY q; cf. also Remark 5.2.
nilpotency class 2, ζL,red
2.5. Topological zeta functions. Another means of defining limits of p-adic zeta funcŸ
Ÿ
tions are topological zeta functions. If c, d are such that ζf grpoq psq “ Wc,dgr pq, q ´s q for
c,d
10
SEUNGJAI LEE AND CHRISTOPHER VOLL
Ÿ
a rational function Wc,dgr pX, Y q P QpX, Y q, for almost all primes p and all finite extenŸ
sions o of Zp , then ζfc,dgr,top psq is simply the coefficient of pq ´ 1q´r in the expansion of
Ÿ
Wc,dgr pq, q ´s q in q ´ 1, where r “ rkZ pfc,d q.
Example 2.5. For a P N0 and b P N,
1
1´q a´bs
“
1
bs´a pq
´ 1q´1 ` Op1q. Thus
1
pq ´ 1q´d ` Oppq ´ 1q´d`1 q,
sps ´ 1q . . . ps ´ d ` 1q
Ÿ
ζf1,dgrpoq psq “ ζod psq “
whence
Ÿ
ζf1,dgr,top psq “
1
.
sps ´ 1q . . . ps ´ d ` 1q
More generally, [9, Definition 5.13] applies to any system of local zeta functions
of Denef type over a number field, in the sense of [9, Definition 5.7]. Examples of
such systems are families of p-adic zeta functions arising from families of the form
pW pq, q ´s qqpPSpecpOq , for suitable W pX, Y q P QpX, Y q (and, as usual, q “ |O{p| for
p P SpecpOq); cf. also [10]. All the local graded ideal zeta functions considered in the
current paper fit into such “uniform” families. We expect this phenomenon to be universal in the context of free nilpotent Lie rings; cf. Conjecture 6.1.
For a formal and far more general definition of topological zeta functions we refer to
[9, Section 5]. In [9, Section 8], Rossmann collects a number of intriguing conjectures
about analytic properties of topological zeta functions associated to various group- and
ring-theoretic counting problems. We expect most of these conjectures to have analogues
in the realm of topological graded ideal zeta functions. Motivated by the computation of
various topological graded ideal zeta functions of free nilpotent Lie rings made throughout the current paper, we make a number of such conjectures in Section 6.
3. Proof of Theorem 1.1 (pc, dq “ p3, 3q)
Let
Let p
fpoqp1q
fpoqp3q
f “ f3,3 be the free nilpotent Lie ring on 3 generators of nilpotency class 3.
be a prime and o be a finite extension of Zp , with uniformizer π. Note that
– o3 – fpoqp2q and fpoqp3q – o8 . In order to parameterize lattices in fpoqp2q and
we denote by
µ “ pµ1 , µ2 , µ3 qě
and
ν “ pν1 , . . . , ν8 qě
integer partitions of at most 3 and 8 parts respectively, and set
p2q
p3q
p3q
µ “ pµ1 , µ2 , µ3 q “ pµ1 , µ1 , µ2 , µ2 , µ2 , µ3 , µ3 , µ3 qě .
Proposition 3.1.
(3.1)
Ÿ
ζf3,3grpoq psq “ ζo3 psq
ÿ
µ
p3q
αpµ1 , µ; qqq ´sp3µ1 `2µ2 `µ3 q
ÿ
αpµ, ν; qqq ´s
νďµ
ř8
k“1
νk
.
Proof. Our starting point is (2.2) in Lemma 2.1, which in this case reads
ÿ
ÿ
Ÿ
|fpoqp3q : Λ3 |´s .
ζf3,3grpoq psq “ ζo3 psq
|fpoqp1q : XpΛ2 q|´s |fpoqp2q : Λ2 |´s
Λ2 ďfpoqp2q
Λ3 ďfpoqp3q ,
rΛ2 ,fpoqp1q sďΛ3
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
11
p3q
Let µ “ pµ1 , µ2 , µ3 qě be a partition. There are αpµ1 , µ; qq lattices Λ2 ď fpoqp2q whose
elementary
divisor type with respect to fpoqp2q is given by µ. Clearly |fpoqp2q : Λ2 | “
ř3
q j“1 µj . We claim that |fpoqp1q : XpΛ2 q| “ q 2µ1 `µ2 . Indeed, lattices such as Λ2 may be
parametrized by their elementary divisor type and a coset αΓµ P SL3 poq{Γµ of a certain
stabilizer subgroup Γµ ď SL3 poq. Then |fpoqp1q : XpΛ2 q| is the index of the lattice of
solutions to the simultaneous congruences
˛
¨
0
α1j α2j
px1 , x2 , x3 q ˝´α1j
0
α3j ‚ ” 0 mod pµj
´α2j ´α3j 0
for j “ 1, 2, 3; cf. [14, Section 2.2] for details. This index clearly is q 2µ1 `µ2 .
After a change of generators for fpoq if necessary, we may assume that
Λ2 “ pµ1 xy ‘ pµ2 xz ‘ pµ3 yz.
At this point we crucially use the fact that d “ 3, as GLd poq –
if d “ 3. We obtain
Ź2
GLd poq if and only
(3.2) rΛ2 , fpoqp1q s “
xπ µ1 xyx, π µ1 xyy, π µ1 xyz, π µ2 xzx, π µ2 xzy, π µ2 xzz, π µ3 yzx, π µ3 yzy, π µ3 yzzyo .
The Jacobi identity involving the three underlined terms is the only nontrivial relation
between these terms. Indeed, the relation xyz ` zxy ` yzx “ 0 implies that
xπ µ1 xyz, π µ2 xzy, π µ3 yzxyo “ xπ µ1 pxzy ´ yzxq, π µ2 xzy, π µ3 yzxyo “ pµ2 xzy ‘ pµ3 yzx.
Hence (3.2) implies that
rΛ2 , fpoqp1q s “
pµ1 xyx ‘ pµ1 xyy ‘ pµ2 xzx ‘ pµ2 xzy ‘ pµ2 xzz ‘ pµ3 yzx ‘ pµ3 yzy ‘ pµ3 yzz,
whence rΛ2 , fpoqp1q s has type µ with respect to fpoqp3q . This shows that the number of
lattices Λ3 ď fpoqp3q whose elementary divisor type with respect to fpoqp3q is given by
a partition ν “ pν1 , . . . , ν8 qě and which satisfy
rΛ2 , fpoqp1q s ď Λ3 is equal to αpµ, ν; qq.
ř8
p3q
´s
k“1 νk .
Each such lattice satisfies |fpoq : Λ3 | “ q
3.1. Overlap types and 2-dimensional words. Our approach to computing the right
hand side of (3.1) is similar to the one taken in [12] to compute local factors of the (Z-)
ideal zeta functions of Lie rings of the form f2,2 pOq.
To compute the right hand side of (3.1) we carry out a case distinction with respect
to the finitely many ways in which the partitions µ and ν may “overlap”. To be precise,
let µ and ν be partitions of 3 and 8 parts, respectively, satisfying ν ď µ. There are
uniquely determined numbers r P N0 and Mi , Ni P N (i P rr ´ 1s), such that
µ 1 ě ¨ ¨ ¨ ě µ M 1 ě ν 1 ě ¨ ¨ ¨ ě ν N1 ą
µM1 `1 ě ¨ ¨ ¨ ě µM2 ě νN1 `1 ě ¨ ¨ ¨ ě νN2 ą . . .
µMr´1`1 ě ¨ ¨ ¨ ě µ3 ě νNr´1 `1 ě ¨ ¨ ¨ ě ν8 .
12
SEUNGJAI LEE AND CHRISTOPHER VOLL
Define Mr “ 3, Nr “ 8, and M0 “ N0 “ 0. We call the integer sequence pMi , Ni qiPrr´1s
arising the overlap type of the pair pµ, νq. Set
p : N0 Ñ N0 ,
n ÞÑ n ¨ p2, 3, 3q :“ 2δně1 ` 3δně2 ` 3δně3 ,
An overlap type determines and is determined by the 2D-word
v “ 1M1 2N1 1M2 ´M1 2N2 ´N1 . . . 13´Mr´1 28´Nr´1 ,
i.e. a word on the alphabet t1, 2u of length eleven such that 1 occurs three times and 2
xi ě Ni for i P rr ´ 1s. Observe that the latter condition is equivalent
eight times, and M
to ν ď µ. In this case we write vpµ, νq “ v. We denote by Dp2q the set of 15 2D-words
arising in this way.
Remark 3.2. Our notion of 2D-word is related to the classical one of (2-dimensional)
Dyck word of length 2n. The latter are words in t1, 2u, featuring n occurrences of each
letter. They may be used to model the overlap types of two partitions, each of at most
n parts; cf. [12, Section 2.4]. 2D-words, in contrast, model the overlaps of two partitions
ν and µ; whilst ν has genuinely at most 8 parts, µ has at most 3 distinct parts, with
ties in blocks of respective sizes 2, 3, and 3.
We chose to phrase our results and the supporting notation as similar to the relevant
material in [12] as possible, leaving the reader free to concentrate on the crucial technical
differences. Our Lemmata 3.4 and 3.5, for instance, are similar to, but subtly different
from Lemmata 2.16 and 2.17 in [12]. Theorem 3.6 is analogous to [12, Theorem 3.1].
For v P Dp2q we set
(3.3)
Dv pq, tq :“
ÿ
p3q
αpµ1 , µ; qqαpµ, ν; qqq ´sp3µ1 `2µ2 `µ3 `
pµ,νq
vpµ,νq“v
Proposition 3.1 allows us to write
(3.4)
Ÿ
ζf3,3grpoq psq “ ζo3 psq
ÿ
ř8
k“1 νk q
.
Dv pq, tq.
vPDp2q
This section’s main result is Theorem 3.6, giving a general formula for the functions Dv . Table 3.1 lists the 15 words in Dp2q together with their overlap types and an
indication where in Section 3.2 an explicit formula for Dv may be found.
In order to obtain such formulae, it is useful for us to have notation for the successive
differences of the parts of µ and ν. For j P r8s we set
#
νj ´ νj`1
if j R tN1 , . . . , Nr u ,
(3.5)
rj “
νNi ´ µMi `1 if j “ Ni ,
where we define µ4 “ 0. Similarly, for j P r3s we set
#
µj ´ µj`1
if j R tM1 , . . . , Mr u ,
(3.6)
sj “
µMi ´ νNi´1 `1 if j “ Mi .
By definition, rj , sj ě 0 for all j. Note also that rj ą 0 if j P tN1 , . . . , Nr´1 u and
observe that µMi ą µMi `1 and νNi ą νNi `1 for each i P rr ´ 1s. Set
q : N0 Ñ N0 ,
n ÞÑ n ¨ p3, 2, 1q :“ 3δně1 ` 2δně2 ` 1δně3 .
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
13
Table 3.1. Overlap types and 2D-words
r pMi , Ni qiPrr´1s
v P Dp2q
1
2
2
2
2
2
2
3
3
3
3
2
3
3
3
13 28
–
p2, 1q
p2, 2q
p2, 3q
p2, 4q
p2, 5q
p1, 1q
p1, 1; 2, 2q
p1, 1; 2, 3q
p1, 1; 2, 4q
p1, 1; 2, 5q
p1, 2q
p1, 2; 2, 3q
p1, 2; 2, 4q
p1, 2; 2, 5q
Then
3µ1 ` 2µ2 ` µ3 “
3
ÿ
j“1
ν1 ` ¨ ¨ ¨ ` ν8 “
8
ÿ
j“1
Finally, for each i P rrs we define
Section
3.2.1
12 2127
12 22 126
12 23 125
12 24 124
12 25 123
1212 27
1212126
12122 125
12123 124
12124 123
122 12 26
122 12125
122 122 124
122 123 123
q
jsj `
jrj `
r
ÿ
i“1
r´1
ÿ
i“1
3.2.2
3.2.3
3.2.4
˘
`
|i rN `1 ` ¨ ¨ ¨ ` rN ,
M
i´1
i
˘
`
Ni sMi `1 ` ¨ ¨ ¨ ` sMi`1 .
(3.7)
Jiµ “ tj P rMi ´ Mi´1 ´ 1s | µMi ´j ą µMi ´j`1 u ,
(3.8)
Jiν “ tj P rNi ´ Ni´1 ´ 1s | νNi ´j ą νNi ´j`1 u .
We start with an immediate consequence of Proposition 2.2.
Lemma 3.3. Let µ and ν be partitions of 3 and 8 parts, respectively, satisfying ν ď µ.
Then
˙
˙ ˆ x1
ˆ
1
1
ÿ
ź 1
µk ´ νk`1
x1 ´ν 1 q 3 ´ µk`1
p3q
µk p3´µ1k q`νk1 pµ
k
k
.
αpµ1 , µ; qqαpµ, ν; qq “
q
1
x1 ´ ν 1
3
´
µ
µ
´1
´1
k
q
µ
kě1
k
k
q
Let µ and ν be partitions as in Lemma 3.3. Let pMi , Ni qiPrr´1s be their overlap type
and µ1 resp. ν 1 be their dual partitions. For k P N, set
Also, set
x1 ´ ν 1 q.
dk :“ µ1k p3 ´ µ1k q ` νk1 pµ
k
k
´
¯
xi ´ j ,
aj :“ Mi p3 ´ Mi q ` j M
j P sNi´1 , Ni s ,
´
¯
bj :“ j p3 ´ jq ` Ni´1 p
j ´ Ni´1 , j P sMi´1 , Mi s .
14
SEUNGJAI LEE AND CHRISTOPHER VOLL
Note that j determines a unique i.
Lemma 3.4. For i P rrs,
µMi´1 `1
ź
k“µMi
q
dk
ˆ
3 ´ µ1k`1
3 ´ µ1k
˙
q
ˆ x1
˙
1
µk ´ νk`1
x1 ´ ν 1
µ
´1
k
“
k
Mi ´M
źi´1
q
q ´1
bMi´1 `j sMi´1 `j
j“1
˙ ˆ
ˆ
˙
Mi ´ Mi´1
3 ´ Mi´1
.
Jiµ
3 ´ Mi q´1
q ´1
Proof. Since all the indices k appearing in the product on the left hand side satisfy
1
“ Ni´1 for all k in the
νNi´1 ą µMi´1`1 ě k ě µMi ě νNi´1 `1 , we have νk1 “ νk`1
interval rµMi , µMi´1 `1 s. Thus
µMi´1 `1 ˆ x
ź
µ1k
k“µMi
1
´ νk`1
x1 ´ ν 1
µ
k
k
˙
“ 1.
q ´1
For j P rMi ´ Mi´1 s we have µ1k “ Mi´1 ` j when µMi´1 `j`1 ă k ď µMi´1 `j ; observe
that it may be the case for some j that no index k satisfies this condition. As a result,
we see that for each j P rMi ´Mi´1 s, there are exactly sMi´1`j elements k of the segment
rµMi , µMi´1 `1 s for which µ1k “ Mi´1 ` j. Now, note that
˙
ˆ
3 ´ µ1k`1
‰1
3 ´ µ1k q´1
if and only if µ1k`1 ‰ µ1k . This is the case exactly when there exists an i such that
µi “ k. It follows that if Jiµ “ tji,1 , . . . , ji,γi u, with ji,1 ă . . . ă ji, γi , ji, 0 :“ 0,
ji, γi `1 :“ Mi ´ Mi´1 , then
(3.9)
µMi´1 `1 ˆ
ź
k“µMi
3 ´ µ1k`1
3 ´ µ1k
˙
q ´1
˙
γi ˆ
ź
3 ´ Mi ` ji, m`1
.
“
3 ´ Mi ` ji,m q´1
m“0
We make use of the well-known identity
ˆ ˙
ˆ
˙
α
1 ´ Xα α ´ 1
“
β X
1 ´ X a´β
β
X
for Gaussian binomial coefficients. Applying it inductively, we see that for all m P rγi ´1s,
ˆ
ˆ
˙ `3´Mi `ji, m`1 ˘
˙
3´Mi
ji,m`1
3 ´ Mi ` ji, m`1
q ´1
.
“
`3´Mi `ji, m ˘
ji,m q´1
3 ´ Mi ` ji,m q´1
´1
3´Mi
Hence the right-hand side of (3.9) is equal to
˙ ˆ
ˆ
˙
Mi ´ Mi´1
3 ´ Mi´1
.
Jiλ
3 ´ Mi q´1
q ´1
q
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
15
Thus
µMi´1 `1
ź
q
dk
k“µMi
ˆ
3 ´ µ1k`1
3 ´ µ1k
˙
q
ˆ x1
˙
1
µk ´ νk`1
x1 ´ ν 1
µ
´1
k
k
Mi ´M
źi´1
“
q
q ´1
bMi´1 `j sMi´1 `j
j“1
as required.
ˆ
Mi ´ Mi´1
Jiµ
˙
q ´1
ˆ
3 ´ Mi´1
3 ´ Mi
˙
q ´1
Lemma 3.5. For i P rrs,
µMi ´1
ź
q
k“µMi `1 `1
dk
ˆ
3 ´ µ1k`1
3 ´ µ1k
˙
q
“
˙
ˆ x1
1
µk ´ νk`1
x1 ´ ν 1
µ
´1
k
k
Ni ´N
źi´1
q
q ´1
aNi´1 `j rNi´1 `j
j“1
ˆ
Ni ´ Ni´1
Jiν
˙
ˆx
˙
Mi ´ Ni´1
.
xi ´ Ni q´1
M
q ´1
Proof. Note that the product on the left hand side may be empty; this happens in
the case µMi “ νNi´1 `1 “ ¨ ¨ ¨ “ νNi “ µMi `1 ` 1. Since µMi ą k ą µMi `1 , we have
µ1k “ µ1k`1 “ Mi for all k in the interval sµMi `1 , µMi ´1 s. Finally, observe that for
j P rNi ´ Ni´1 s we have νk1 “ Ni´1 ` j when νNi´1 `j`1 ă k ď νNi´1 `j . The claim
follows as in the proof of the previous lemma.
With these preparations in place we now give a formula for the functions Dv .
˘
ś `
Theorem 3.6. Let v “ ri“1 1Mi ´Mi´1 2Ni ´Ni´1 P Dp2q . Then
˙ ˆx
˙ ź
r ˆ
r
ź
˘
`
Mi ´ Ni´1
3 ´ Mi´1
Dv pq, tq “
IMi ´Mi´1 YMi´1 `1 , . . . , YMi
xi ´ Ni q´1
3 ´ Mi q´1 M
i“1
i“1
¸
˜
r´1
ź
˘
`
˘
`
˝
,
X
,
.
.
.
,
X
¨
I
¨
IN
X
,
.
.
.
,
X
N
`1
8
8´N
N
`1
N
´N
r´1
r´1
i´1
i
i
i´1
i“1
with numerical data
(3.10)
(3.11)
|
Xj “ q aj tMi `j
j P sNi´1 , Ni s ,
q
j`Ni´1
Yj “ q bj t
j P sMi´1 , Mi s .
Proof. Given v “ pv1 , v2 , v3 q P N30 and v1 “ pv11 , . . . , v81 q P N80 , we set, for each i P rrs,
(
suppM
i pvq “ j P rMi ´ Mi´1 ´ 1s | vMi´1 `j ą 0 ,
!
)
1
1
suppN
pv
q
“
j
P
rN
´
N
´
1s
|
v
ą
0
.
i
i´1
i
Ni´1 `j
In practice, v will be one of the vectors of successive differences s “ ps1 , s2 , s3 q (cf. (3.6))
and v 1 will be one of the vectors of successive differences r “ pr1 , . . . , r8 q (cf. (3.5)).
Given a pair pµ, νq of partitions satisfying ν ď µ, recall our definitions (3.7) and (3.8).
It is easy to see that, for every i P rrs,
µ
suppM
i pvq “ Mi ´ Mi´1 ´ Ji
and
1
ν
suppN
i pv q “ Ni ´ Ni´1 ´ Ji .
16
SEUNGJAI LEE AND CHRISTOPHER VOLL
Thus it follows that
˙
ˆ
ˆ
˙
˙
ˆ
ˆ
˙
Mi ´ Mi´1
Mi ´ Mi´1
Ni ´ Ni´1
Ni ´ Ni´1
“
and
“
.
1
Jiµ
suppM
Jiν
suppN
i pvq q ´1
i pv q q ´1
q ´1
q ´1
Let δij be the usual Kronecker delta function. Substituting Lemma 3.4 and 3.5, rewriting
the expressions in terms of rj and sj and using (3.1), we find that formula (3.3) for
Dv pq, tq splits into a product as follows:
˙ ˆx
˙ ź
r ˆ
r
ź
Mi ´ Ni´1
3 ´ Mi´1
Ai Bi ,
Dv pq, tq “
xi ´ Ni q´1
3 ´ Mi q´1 M
i“1
i“1
where, for i P rrs,
Ai “
8
ÿ
¨¨¨
rNi ´1“0 rN “1´δir
i
rNi´1 `1 “0
Bi “
8
ÿ
¨¨¨
sMi´1 `1 “0
8
ÿ
8
ÿ
8
ÿ
sMi ´1“0
ˆ
˙
Ni
¯
´
ź
Ni ´ Ni´1
|i `j rj
aj M
q
t
,
1
suppN
i pv q q ´1 j“N
`1
i´1
˙
Mi
8 ˆ
¯s
´
ÿ
ź
Mi ´ Mi´1
bj Ni´1 `q
j j
q
t
.
suppM
i pvq q ´1 j“M
s “0
`1
Mi
i´1
We now show that all of the factors Ai and Bi are products of Igusa functions and
Gaussian binomial coefficients. Given i P rrs and I Ď rMi ´ Mi´1 ´ 1s, we define Si pIq
˘
`
M ´M
to be the set of vectors si “ sMi´1 `1 , . . . , sMi P N0 i i´1 such that sj “ 0 unless
j P tMi´1 ` k | k P Iu Y tMi u. With the numerical data defined in (3.11), we have
Bi
“
ÿ
IĎrMi ´Mi´1 ´1s
ˆ
ˆ
Mi ´ Mi´1
I
˙
˙
ÿ
ź
q ´1 si PSi pIq jPpI`Mi´1 qYtMi u
¨
¨
´
¯
q sj
q bj tNi´1 `j
˛˛
8
ÿ
˘s
`
Mi ´ Mi´1
˝ ˝
pYMi qsMi
YMi´1 `ι Mi´1 `ι ‚‚
I
´1
q
sMi “0
sMi´1 `ι “1
ιPI
IĎrMi ´Mi´1 ´1s
ˆ
˙
ź YMi´1 `ι
ÿ
1
Mi ´ Mi´1
“
I
1 ´ YMi
q ´1 ιPI 1 ´ YMi´1 `ι
IĎrMi ´Mi´1 ´1s
˘
`
“IMi ´Mi´1 YMi´1 `1 , . . . , YMi .
“
ÿ
ź
8
ÿ
Analogously one shows that, with the numerical data defined in (3.10),
#
˘
`
˝
IN
for i ă r,
X
,
.
.
.
,
X
N
`1
N
i´1
i
´N
i
i´1
Ai “
˘
`
for i “ r.
I8´Nr´1 XNr´1 `1 , . . . , X8
This completes the proof of Theorem 3.6.
3.2. Explicit formulae for the functions Dv .
3.2.1. v “ 13 28 .
`
˘
D13 28 pq, tq “ I3 q 2 t3 , q 2 t5 , t6 I8 ppq ip8´iqq t6`i qi“1,...,8 q.
`
˘
Ÿ
Remark 3.7. Note that ζf2,3grpoq psq “ ζo3 psqI3 q 2 t3 , q 2 t5 , t6 ; cf. Proposition 4.10.
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
17
3.2.2. v “ 12 2j 128´j , j P t1, . . . , 5u.
ˆ ˙ ˆ ˙
`
˘
5
3
D12 2j 128´j pq, tq “
I2 q 2 t 3 , q 2 t 5
j q´1 1 q´1
¨ Ij˝ ppq 2`ip5´iq t5`i qi“1,...,j qI1 pq jp8´jq t6`j qI8´j ppq ip8´iq t6`i qi“j`1,...,8 q.
3.2.3. v “ 1212j´1 128´j , j P t1, . . . , 5u. For j “ 1,
ˆ ˙ ˆ ˙
3
2
D1212 27 pq, tq “
I1 pq 2 t3 qI1˝ pq 3 t4 qI2 pq 6 t6 , q 7 t7 qI7 ppq ip8´iq t6`i qi“2,...,8 q.
1 q´1 1 q´1
For j P t2, . . . , 5u,
D1212j´1128´j pq, tq “
ˆ
˙ ˆ ˙ ˆ ˙2
4
3
2
I1 pq 2 t3 qI1˝ pq 3 t4 qI1 pq 6 t6 q
j ´ 1 q´1 1 q´1 1 q´1
˝
¨ Ij´1
ppq 2`ip5´iq t5`i qi“2,...,j qI1 pq jp8´jq t6`j qI8´j ppq ip8´iq t6`i qi“j`1,...,8 q.
3.2.4. v “ 122 12j´1 127´j , j P t1, . . . , 4u. For j “ 1,
ˆ ˙
3
D122 12 26 pq, tq “
I1 pq 2 t3 qI2˝ pq 3 t4 , q 2 t5 qI2 pq 8 t7 , q 12 t8 qI6 ppq ip8´iq t6`i qi“3,...,8 q.
1 q´1
For j P t2, 3, 4u,
D122 12j´1 127´j pq, tq “
ˆ
3
j ´1
˙
ˆ ˙ ˆ ˙
3
2
I1 pq 2 t3 qI2˝ pq 3 t4 , q 2 t5 qI1 pq 8 t7 q
1
1
´1
´1
´1
q
q
q
˝
¨ Ij´1
ppq 2`ip5´iq t5`i qi“3,...,j`1 qI1 pq pj`1qp8´pj`1qq t7`j qI8´pj`1q ppq ip8´iq t6`i qi“j`2,...,8 q.
3.3. Completion of the proof. The first two claims of Theorem 1.1 follows from (3.4)
and the explicit formulae for Dv , v P Dp2q , given in Section 3.2. To deduce the local
functional equation (1.5), one checks, repeatedly using the functional equations (2.3)
the Igusa `functions
Ih resp. Ih˝ and the well-known fact that, for a, b P N0 , a ě b,
`for
˘
˘
a
bpa´bq a
b q “q
b q ´1 , that each of the functions Dv satisfies
Dv pq ´1 t´1 q “ ´q 31 t20 Dv pq, tq.
The functional equation (1.5) follows from (3.4) as
ζo3 psq|qÑq´1 “ ´q 3 t3 ζo3 psq.
4. Nilpotency class two (c “ 2)
Ÿ
Let d P Ně2 . In this section we compute the local graded ideal zeta functions ζfd,2grpoq psq.
We prove functional equations for these functions and establish their behaviour at s “ 0.
We use them to determine the abscissae of convergence of the global graded ideal zeta
Ÿ
functions ζf grpOq psq and some properties of the associated topological and reduced graded
d,2
`˘
ideal zeta functions. Throughout we write d1 for d2 “ W2 pdq; cf. (1.2).
18
SEUNGJAI LEE AND CHRISTOPHER VOLL
4.1. p-Adic formulae. The paper [15] determines the normal subgroup zeta functions
of the free class-2-nilpotent d-generator groups F2,d , enumerating these groups’ normal
subgroups of finite index. By the Mal’cev correspondence, these are the ideal zeta
functions of the free nilpotent Z-Lie rings f2,d . The computations generalize to the case
of general number rings in a straightforward manner. To recall the paper’s main result
define the function
ˆ
˙
i`1
1
ϕ : rd ´ 1s0 Ñ rd s, i ÞÑ id ´
.
2
1
Given pI, Jq P 2rd´1s0 ˆ 2rd s , the paper [15, p. 71] defines a total order ăϕpIq,J on the
disjoint union I Y J. Without loss of generality, we may assume that the sets I and J
have the same cardinality h. For i P I “ ti1 , . . . , ih uă and j P J “ tj1 , . . . , jh uă , set
(
(
jpiq :“ min j P J Y d1 | ϕpiq ăϕpIq,J , j ,
(
ipjq :“ max i P I Y t0u | ϕpiq ăϕpIq,J , j ,
and j0 :“ i0 :“ 0 ăϕpIq,J ϕ prd ´ 2sq Z rd1 ´ 1s ăϕpIq,J jh`1 :“ d1 . The following is
essentially the main result of [15], in notation compatible with the current paper. (The
underlining of terms of the form d` in (4.3) is meant to facilitate comparison with the
“graded” numerical data (4.5) in Theorem 4.2, and may be ignored.)
Theorem 4.1. [15, Theorem 4] For all primes p and all finite extensions o of Zp ,
ÿ
(4.1)
ζfŸ2,d poq psq “ ζod psq
AŸ
I,J pq, tq,
IĎrd´2s,JĎrd1 ´1s
|I|“|J|,ϕpIqďJ
where
(4.2)
AŸ
I,J pq, tq
´1
“ ζoih psq
Ij1 ppXα qαPrj1 ´1s , X0 q ¨
˙
h ˆ
ź
jr`1 ´ ϕpir q
r“1
jr ´ ϕpir q
q ´1
ˆ
d ´ ir´1
d ´ ir
˙
q ´1
¨ Ii˝r ´ir´1 ppYβ qβPsir´1 ,ir r , Yi1r qIjr`1 ´jr ppXα qαPsjr ,jr`1r , Xjr q,
with numerical data
“
‰
j P d1 ´ 1 0 ,
Xj “ q ipjqpd´ipjqq`pd ´jqpd`j´ϕpipjqqq td´ipjq`d ´j ,
1
(4.3)
1
Yi “ q ipd´iq`pd ´jpiqqpd`jpiq´ϕpiqq td´i`d ´jpiq ,
1
1
i P rd ´ 2s ,
Yi1r “ q ir´1 pd´ir´1 q`pd ´jpir qqpd`jpir q´ϕpir´1 qq td´ir´1`d ´jpir q ,
1
1
The following is a graded analogue of Theorem 4.1.
Theorem 4.2. For all primes p and all finite extensions o of Zp ,
ÿ
Ÿ
Ÿ
(4.4)
ζf2,dgrpoq psq “ ζod psq
AI,Jgr pq, tq,
IĎrd´2s,JĎrd1 ´1s
|I|“|J|,ϕpIqďJ
r P rhs .
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
19
Ÿ
where AI,Jgr pq, tq is defined as AŸ
I,J pq, tq in (4.2), but with numerical data
“
‰
1
1
Xj “ q ipjqpd´ipjqq`pd ´jqpj´ϕpipjqqq td´ipjq`d ´j ,
j P d1 ´ 1 0 ,
1
1
Yi “ q ipd´iq`pd ´jpiqqpjpiq´ϕpiqq td´i`d ´jpiq ,
(4.5)
Yi1r “ q ir´1 pd´ir´1
q`pd1 ´jpi
r qqpjpir q´ϕpir´1 qq
i P rd ´ 2s ,
td´ir´1
`d1 ´jpi
rq
,
r P rhs .
Proof. The difference between (4.3) and (4.5) reflects the difference between (1.7) and
(1.8).
Corollary 4.3.
(4.6)
(4.7)
ˇ
1
ˇ
d`d1 pd`d
2d`d1 Ÿ
2 qt
“
p´1q
q
ζfŸ2,d poq psqˇ
ζf2,d poq psq,
´1
qÑq
ˇ
1
ˇ
Ÿ
d`d1 pd2q`pd2 q 2d`d1 Ÿgr
“p´1q
q
t
ζf2,d poq psq.
ζf2,dgrpoq psqˇ
´1
qÑq
Proof. Eq. (4.6) is [15, Theorem 3(a)]. The proof proceeds by establishing the symmetry
in question for each of the summands in (4.1), using (2.3). The formulae (4.1) and (4.4)
only differ in their numerical data. By Remark 2.4, it suffices to control the value of
ś
X
X0 ¨ hr“1 Y j1r in both cases. Inspection of (4.3) and (4.5), respectively, reveals that in
ir
1
the ideal case this term is by a factor q dd larger than in the graded ideal case. Thus
ˇ
1
ˇ
Ÿ
´dd1 2d`d1 Ÿgr
d`d1 pd`d
2 q
ζf2,d poq psq.
t
q
“
p´1q
ζf2,dgrpoq psqˇ
´1
As
`d`d1 ˘
2
´
dd1
“
`d˘
2
qÑq
`
`d1 ˘
2
, this establishes (4.7).
Ÿ
4.2. p-Adic behaviour at zero. We describe the behaviours of ζfŸ2,d poq psq and ζf grpoq psq
2,d
at s “ 0.
Theorem 4.4. For all primes p and all finite extensions o of Zp ,
ˇ
ζfŸ2,d poq psq ˇˇ
(4.8)
“ 1,
ˇ
ζod`d1 psq ˇ
s“0
ˇ
Ÿgr
ˇ
ζf poq psq ˇ
d´1
d ¨ d1
2,d
(4.9)
“
.
“
ˇ
1
ζod psqζod1 psq ˇ
d ¨ pd ` d q
d`1
s“0
Proof. To prove (4.8) we note that, by (4.1),
¨
1
ζfŸ2,d poq psq d`d
ź´1
“
p1 ´ q j tq ˝AŸ
∅,∅ pq, tq `
ζod`d1 psq
j“d
ÿ
pI,Jq‰p∅,∅q
1
˛
‚
AŸ
I,J pq, tq .
pd`pd ´jqqj td`j q
By inspection of (4.3) we see that AŸ
jPrd1 s q and that
∅,∅ pq, tq “ Id1 ppq
#
0
if pI, Jq ‰ p∅, ∅q,
AŸ
śd`d1 ´1
1 ´jqqj
I,J pq, 1q “
pd`pd
j
qjPrd1 s q “ j“d p1 ´ q q if pI, Jq “ p∅, ∅q.
Id1 ppq
To prove (4.9) we note that
Ÿ
ζf2,dgrpoq psq
ζod psqζod1 psq
“
1 ´1
dź
j“0
¨
Ÿ
gr
p1 ´ q j tq ˝A∅,∅
pq, tq `
ÿ
pI,Jq‰p∅,∅q
˛
Ÿ
AI,Jgr pq, tq‚.
20
SEUNGJAI LEE AND CHRISTOPHER VOLL
Ÿ
1
gr
By inspection of (4.5) we see that A∅,∅
pq, tq “ Id1 ppq pd ´jqj td`j qjPrd1 s q, which has a simple
Ÿ
pole at t “ 1. In contrast, AI,Jgr pq, tq has no pole at t “ 1 when pI, Jq ‰ p∅, ∅q. Hence
ˇ
ˇ
ˇ
Ÿ
1
1
ζf2,dgrpoq psq ˇˇ
Id1 ppq pd ´jqj td`j qjPrd1 s q ˇˇ
1 ´ td ˇˇ
d1
“
“
.
“
ˇ
ˇ
ˇ
1
ζod psqζod1 psq ˇ
1 ´ td`d1 ˇ
d ` d1
Id1 ppq pd ´jqj tj qjPrd1 s q ˇ
t“1
s“0
t“1
Remark 4.5. In the pertinent special cases, (4.8) confirms [9, Conjecture IV (P-adic
form)].
4.3. Global analytic properties. The following result compares some of the known
Ÿ
analytic properties of ζfŸ2,d pOq psq (cf. [15, Theorem 3]) with those of ζf grpOq psq.
2,d
ζfŸ2,d pOq psq
Ÿ
ζf2,dgrpOq psq
are
Theorem 4.6. The abscissae of convergence of
resp.
+
# ` ˘
ˆ ˙
p d2 ´ jqpd ` jq ` 1
d
Ÿ
|jPr
(4.10)
´ 1s resp. ,
α pf2,d q “ max d,
`d`1˘
2
´j
2
# ` ˘
+
ˆ ˙
p d2 ´ jqj ` 1
d
Ÿgr
|jPr
α pf2,d q “ max d, `d`1˘
(4.11)
´ 1s .
2
´
j
2
The respective meromorphic continuations of both zeta functions beyond their abscissae
of convergence have simple poles at s “ αŸ pf2,d q resp. s “ αŸgr pf2,d q.
Proof. Eq. (4.10) and the ensuing claim about the meromorphic continuation is essentially [15, Theorem 3]. Eq. (4.11) and the analogous claim are proved analogously.
4.4. Topological zeta functions – degree and behaviour at zero.
Theorem 4.7.
Moreover,
¯
´
¯
´
Ÿ
degs ζfŸ2,d ,top psq “ ´pd ` d1 q “ degs ζf2,dgr,top psq .
ˇ
ˇ
sζfŸ2,d ,top psqˇ
(4.12)
ˇ
ˇ
Ÿ
s2 ζf2,dgr,top psqˇ
(4.13)
1
s“0
p´1qd`d ´1
,
“
pd ` d1 ´ 1q!
s“0
p´1qpd´1q`pd ´1q d1
“
.
pd ` d1 q pd ´ 1q! pd1 ´ 1q!
1
Proof. Note that, given h P N and, for i P rhs, Xi “ q ai tbi for ai P N0 , bi P N,
h!
pq ´ 1q´h ` Oppq ´ 1q´h`1 q.
Ih pX1 , . . . , Xh q “ śh
i“1 pbi s ´ ai q
The summands in the formula (4.1) for ζfŸ2,d poq psq are all products of Igusa functions and
Ÿ
Ÿ
Gaussian binomial coefficients. Hence there exist aŸ
I,J,i P N0 and bI,J,i , cI,J P N such that
ζfŸ2,d poq psq
“
ÿ
I,J
śd`d1
cŸ
I,J
Ÿ
i“1 pbI,J,i s
1
´
aŸ
I,J,i q
1
pq ´ 1q´pd`d q ` Oppq ´ 1q´pd`d q`1 q.
(Here and in the sequel, the sums are over pairs pI, Jq as in (4.1).) Hence
ÿ
cŸ
I,J
(4.14)
ζfŸ2,d ,top psq “
śd`d1 Ÿ
Ÿ
I,J
i“1 pbI,J,i s ´ aI,J,i q
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
21
is a rational function in s of degree ´pd ` d1 q, confirming the first claim.
Ÿ
The second claim, on the degree of ζf2,dgr,top psq, follows from analogous considerations
Ÿ
Ÿ
Ÿ
gr
gr
, cI,Jgr P N such that
P N0 and bI,J,i
based on (4.4). Indeed, there exist aI,J,i
Ÿ
ζf2,dgrpoq psq
“
ÿ
I,J
Hence
Ÿ
śd`d1
cI,Jgr
Ÿgr
i“1 pbI,J,i s
Ÿ
ζf2,dgr,top psq
(4.15)
´
“
Ÿgr pq
q
aI,J,i
ÿ
I,J
1
1
´ 1q´pd`d q ` Oppq ´ 1q´pd`d q`1 q.
Ÿ
śd`d1
i“1
cI,Jgr
Ÿ
Ÿ
gr
gr
q
s ´ aI,J,i
pbI,J,i
is a rational function in s of degree ´pd ` d1 q, too.
Turning to the behaviour at s “ 0, we start with the observation that ζod psqAŸ
I,J pq, tq
has no pole at t “ 1 unless pI, Jq “ p∅, ∅q, in which case it is simple. Thus the summand
cŸ
I,J
Ÿ
Ÿ
i“1 pbI,J,i s´aI,J,i q
in (4.14) has no pole at s “ 0 unless pI, Jq “ p∅, ∅q, in which case it
śd`d1
is simple. Specifically,
1
ipd´iq i
t qiPrds qId1 ppq pd`d ´jqj td`j qjPrd1 s q,
ζod psqAŸ
∅,∅ pq, tq “ Id ppq
whence
śd`d1
i“1
and thus
cŸ
∅,∅
pd1 q!
1
¨ śd1
“ śd´1
Ÿ
1
pbŸ
i“0 ps ´ iq
j“1 ppd ` jqs ´ pd ` d ´ jqjq
∅,∅,i s ´ a∅,∅,i q
ˇ
ˇ
sζfŸ2,d ,top psqˇ
1
p´1qd`d ´1
“
,
“ śd´1
śd1
1
pd ` d1 ´ 1q!
i“1 p´iq
j“1 p´pd ` d ´ jqjq
pd1 q!
s“0
which establishes (4.12).
Ÿ
The proof of (4.13) goes along similar lines. We observe that ζod psqAI,Jgr pq, tq has
a simple pole at t “ 1 unless pI, Jq “ p∅, ∅q, in which case it is double. Thus the
summand
Ÿgr
cI,J
Ÿgr
Ÿgr
i“1 pbI,J,i s´aI,J,i q
śd`d1
in (4.15) has a simple pole at s “ 0 unless pI, Jq “ p∅, ∅q,
in which case it is double. Specifically,
Ÿ
1
gr
ζod psqA∅,∅
pq, tq “ Id ppq ipd´iq ti qiPrds qId1 ppq pd ´jqj td`j qjPrd1 s q,
whence
Ÿ
śd`d1
gr
c∅,∅
Ÿgr
i“1 pb∅,∅,i s
and thus
ˇ
ˇ
Ÿ
s2 ζf2,dgr,top psqˇ
s“0
´
Ÿgr
q
a∅,∅,i
pd1 q!
1
¨ śd1
“ śd´1
1
i“0 ps ´ iq
j“1 ppd ` jqs ´ pd ´ jqjq
1
pd1 q!
p´1qpd´1q`pd ´1q d1
,
“
“
ś
ś
d1 ´1
1
pd ` d1 qpd ´ 1q!pd1 ´ 1q!
pd ` d1 q d´1
i“1 p´iq
j“1 p´pd ´ jqjq
which proves (4.13).
Remark 4.8. In the pertinent special cases, the theorem’s first statement confirms [9,
Conjecture 1], whereas (4.12) confirms [9, Conjecture IV (topological form)].
22
SEUNGJAI LEE AND CHRISTOPHER VOLL
4.5. Explicit examples. We record the following consequences of Theorem 4.2.
Proposition 4.9 (pc, dq “ p2, 2q – Heisenberg).
Ÿ
Ÿ
ζf2,2grpoq psq “ ζo2 psqI1 pt3 q “: W2,2gr pq, tq,
1
Ÿ
Ÿ
ζf2,2gr,red pY q “ W2,2gr p1, Y q “
p1 ´
1
Ÿ
ζf2,2gr,top psq “
Y q2 p1
3s2 ps
´ 1q
´ Y 3q
,
.
Ÿ
The global graded ideal zeta function ζf2,2grpOq psq “ ζK psqζK ps ´ 1qζK p3sq has abscissa of
convergence 2 and meromorphic continuation to the whole complex plane.
Proposition 4.10 (pc, dq “ p2, 3q).
Ÿ
Ÿ
ζf2,3grpoq psq “ ζo3 psqI3 pq 2 t3 , q 2 t5 , t6 q “: W2,3gr pq, tq,
Ÿ
1 ` 2Y 3 ` 2Y 5 ` Y 8
,
p1 ´ Y q3 p1 ´ Y 3 qp1 ´ Y 4 qp1 ´ Y 5 qp1 ´ Y 6 q
1
Ÿ
.
ζf2,3gr,top psq “ 2
s ps ´ 1qps ´ 2qp3s ´ 2qp5s ´ 2q
Ÿ
ζf2,3gr,red pY q “ W2,3gr p1, Y q “
Ÿ
The global graded ideal zeta function ζf2,3grpOq psq has abscissa of convergence 3 and may
be continued meromorphically to ts P C | Repsq ą 1{3u.
Ÿ
Ÿ
Remark 4.11. The formulae for ζf2,2grpoq resp. ζf2,3grpoq (for almost all p) are also given in
[11, Table 2] under the labels m3 2 resp. m6 2 1.
Ÿ
We omit the (largish) formula for ζf2,4grpoq psq, but do note the following consequences.
Proposition 4.12 (pc, dq “ p2, 4q). There exists a rational function W2,4 pX, Y q P
QrX, Y s such that
Ÿ
Ÿ
ζf2,4grpoq psq “ W2,4gr pq, q ´s q.
Setting
N2,4 pY q “ Y 33 `5Y 33 `12Y 28 `6Y 27 `15Y 26 `26Y 25 `11Y 24 `39Y 23 `40Y 22 `43Y 21 `
62Y 20 ` 45Y 19 ` 66Y 18 ` 61pY 17 ` Y 16 q ` 66Y 15 ` 45Y 14 ` 62Y 13 `
43Y 12 ` 40Y 11 ` 39Y 10 ` 11Y 9 ` 26Y 8 ` 15Y 7 ` 6Y 6 ` 12Y 5 ` 5Y 3 ` 1 P QrY s
we have
Ÿ
Ÿ
ζf2,4gr,red pY q “ W2,4gr p1, Y q “
Ÿ
ζf2,4gr,top psq “
p1 ´ Y
q4 p1
N2,4 pY q
,
ś
i
´ Y 3 q 10
i“6 p1 ´ Y q
36s2 ´ 43s ` 12
3
.
10 s2 ps ´ 1q3 ps ´ 2qps ´ 3qp3s ´ 4q2 p2s ´ 1qp7s ´ 9qp9s ´ 5q
The global graded ideal zeta function ζfŸ2,4 pOq psq has abscissa of convergence 4 and may
be continued meromorphically to ts P C | Repsq ą 7{6u.
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
23
5. Two generators (d “ 2)
Ÿ
We compute the graded ideal zeta functions ζfc,2grpoq psq for c P t3, 4u as well as their
reduced and topological counterparts. (For c “ 1 see (1.3), for c “ 2 see Proposition 4.9.)
Proposition 5.1. ppc, dq “ p3, 2qq
1 ` t4
Ÿ
“: W3,2gr pq, tq,
p1 ´ t3 qp1 ´ qt4 qp1 ´ t5 q
1`Y4
Ÿ
Ÿ
ζf3,2gr,red pY q “ W3,2gr p1, Y q “
,
p1 ´ Y q2 p1 ´ Y 3 qp1 ´ Y 4 qp1 ´ Y 5 q
2
Ÿ
.
ζf3,2gr,top psq “
3
15s ps ´ 1qp4s ´ 1q
Ÿ
ζf3,2grpoq psq “ ζo2 psq
The global graded ideal zeta function
Ÿ
ζf3,2grpOq psq “
ζK psqζK ps ´ 1qζK p3sqζK p4sqζK p4s ´ 1qζK p5sq
ζK p8sq
has abscissa of convergence 2 and meromorphic continuation to the whole complex plane.
Proof. (sketch) Let f “ f3,2 and recall the Hall basis H3,2 for f from Table 2.1. Recall
formula (2.2) for the zeta function of fpoq in terms of pairs pΛ2 , Λ3 q of sublattices of
fpoqp2q – o and fpoqp3q – o2 , respectively. Note that |fpoqp2q : Λ2 | “ q µ for some µ P N0 ,
|fpoqp1q : XpΛ2 q| “ q 2µ , and pΛ2 , Λ3 q satisfies the condition in (2.2) if and only if
pµ fpoqp3q ď Λ3 .
(5.1)
This proves the first claim; the others are trivial consequences.
Ÿ
Remark 5.2. Comparison with ζfŸ3,2 poq psq (cf. [5, Theorem 2.35]) yields ζf3,2gr,red pY q “
ζfŸ3,2 ,red pY q. Note that the Hall basis H3,2 is nice and simple in the sense of [6]. The
Ÿ
formula for ζf3,2grpoq (for almost all p) is also given in [11, Table 2] under the label m5 3 1.
Proposition 5.3. ppc, dq “ p4, 2qq Set
N4,2 pX, Y q “ ´X 2 pY 22 ` Y 18 ` Y 17 ` Y 16 ` Y 15 ` Y 11 q
´ XpY 16 ` Y 15 ´ Y 13 ` Y 9 ´ Y 7 ´ Y 6 q ` Y 11 ` Y 7 ` Y 6 ` Y 5 ` Y 4 ` 1 P QrX, Y s.
Then
N4,2 pq, tq
p1 ´
´
´
´ qt5 qp1 ´ q 2 t6 qp1 ´ q 2 t7 qp1 ´ t8 q
Y 17 ` Y 13 ` 2pY 12 ` Y 11 ` Y 10 ` Y 7 ` Y 6 ` Y 5 q ` Y 4 ` 1
Ÿ
,
ζf4,2gr,red pY q “
ś
p1 ´ Y q2 8i“3 p1 ´ Y i q
1
p20s ´ 3q
Ÿ
ζf4,2gr,top psq “
.
4
60 s ps ´ 1qp5s ´ 1qp4s ´ 1qp7s ´ 2qp3s ´ 1q
Ÿ
ζf4,2grpoq psq “ ζo2 psq
t3 qp1
qt4 qp1
t5 qp1
The global graded ideal zeta function ζfŸ4,2 pOq psq has abscissa of convergence 2 and may
be continued meromorphically to ts P C | Repsq ą 2{11u.
24
SEUNGJAI LEE AND CHRISTOPHER VOLL
Proof. (sketch) Let f “ f4,2 and recall the Hall basis H4,2 for f from Table 2.1. Recall
formula (2.2) for the zeta function of fpoq in terms of triples pΛ2 , Λ3 , Λ4 q of sublattices
of fpoqp2q – o, fpoqp3q – o2 , and fpoqp4q – o3 , respectively. Note that |fpoqp2q : Λ2 | “ q µ
for some µ P N0 , |fpoqp1q : XpΛ2 q| “ q 2µ , and pΛ2 , Λ3 , Λ4 q satisfies the condition in (2.2)
if and only if pΛ2 , Λ3 q satisfies (5.1) and rΛ3 , fpoqp1q s ď Λ4 . To investigate the latter
condition, we may assume that Λ3 “ xπ ν1 xyx, π ν2 xyyyo for a partition ν “ pν1 , ν2 qě .
Then
rΛ3 , fpoqp1q s “ xπ ν1 xyxx, π ν1 xyxy, π ν2 xyyx, π ν2 xyyyyo
As xyxy “ xyyx by the Jacobi identity, the underlined term may be omitted, whence
rΛ3 , fpoqp1q s “ π ν1 xyxx ‘ π ν2 xyyx ‘ π ν2 xyyy.
To enumerate the lattices Λ4 ď fpoqp4q satisfying rΛ3 , fpoqp1q s ď Λ4 , we distinguish according to the “overlap type” such a lattice may have with rΛ3 , fpoqp1q s. Indeed, if Λ4
has elementary divisor type given by a partition pξ1 , ξ2 , ξ3 qě , then either
ν1 ě ν2 ě ξ 1 ě ξ 2 ě ξ 3
ν1 ě ξ 1 ą ν2 ě ξ 2 ě ξ 3 .
or
Each of the two cases is covered by a formula similar to the one established in Theorem 3.6. Adding them yields a formula for the zeta function, viz.
Ÿ
ζf4,2grpoq psq “ ζo2 psqI1 pt3 q
˜
¨
¸
ˆ ˙
2
4 ˝
5
2 6
2 7 8
I2 pqt , t qI3 pq t , q t , t q `
I1 pqt qI1 pqt qI1 pq t qI2 pq t , t q .
1 q´1
4
5
2 6
2 7
8
This proves the first claim; the others are trivial consequences. The claim about the
meromorphic continuation follows from [5, Lemma 5.5].
6. General conjectures
Let c P N and d P Ně2 . We record a number of general conjectures regarding the
Ÿ
graded ideal zeta functions ζfc,dgrpoq psq as well as their topological and reduced counterparts. We write W for the Witt function Wd (cf. (1.2)) and set r “ rkZ pfc,d q “
řc
i“1 W piq. With the exception of Conjecture 6.7 for c “ 2 and d ě 5, all conjectures in
this section are confirmed, in the relevant special cases, by the results in this paper.
Ÿ
Conjecture 6.1 (Uniformity). There exists Wc,dgr pX, Y q P QpX, Y q such that, for almost all primes p and all finite extensions o of Zp ,
Ÿ
Ÿ
ζfc,dgrpoq psq “ Wc,dgr pq, q ´s q.
Conjecture 6.2 (Local functional equations). For almost all primes p and all finite
extensions o of Zp ,
Ÿ
ζfc,dgrpoq psq|qÑq´1 “ p´1qr q
řc
i“1
pW2piqq´
řc
i“1 pc`1´iqW piqs
Ÿ
ζfc,dgrpoq psq.
Remark 6.3. Local functional equations are a universal feature for quite general subring
zeta functions ([16, Theorem A]) and for ideal zeta functions of nilpotent Lie rings of
nilpotency class 2 ([16, Theorem C]). In fact, we expect that the proof of the latter
result should carry over to the graded setting, making the functional equations (4.7) an
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
25
instance of a general phenomenon in class 2; cf. Example 1.6. If Conjecture 6.1 holds,
then Conjecture 6.2 states that
Ÿ
Wc,dgr pX ´1 , Y ´1 q “ p´1qr X
řc
i“1
pW2piqq Y
řc
i“1 pc`1´iqW piq
Ÿ
Wc,dgr pX, Y q.
In general, the operation q Ñ q ´1 is defined rather more delicately in terms of inversion
of Frobenius eigenvalues.
In nilpotency class greater than 2 local functional equations are not to be expected
in general for ideal zeta functions of nilpotent Lie rings: see [5, Section 1.2.3] for counterexamples in the ungraded setting and [11, Table 2] in the graded setting.
A sufficient criterion for local functional equations in an enumerative setup generalizing ideal zeta functions of nilpotent Lie rings is given in [17, Theorem 1.2]. It applies
to the ideal zeta functions of the free nilpotent Lie rings fc,d ; [17, Theorem 4.4]. Note
ř
řc
that, in the notation of this result, N0 “ r and c´1
i“0 Ni “
i“1 pc ` 1 ´ iqW piq.
For the following conjecture, assume that Conjecture 6.1 holds. In this case, we may
define the reduced graded ideal zeta function
Ÿ
Ÿ
ζfc,dgr,red pY q “ Wc,dgr p1, Y q P QpY q.
We expect that the technology in [6] may be adapted to define the reduced graded ideal
zeta function even without the assumption of Conjecture 6.1.
Ÿ
Conjecture 6.4 (Reduced zeta function). The reduced graded ideal zeta function ζfc,dgr,red pY q
has a pole of order r at 1 and there exist ei P N, i P rrs, as well as nonnegative integers
ř
ř
aj P N0 , j P rN s0 , where N :“ p ri“1 ei q ´ cj“1 pc ` 1 ´ jqWd pjq, with a0 “ 1 and
(6.1)
aj “ aN ´j for all j P rN s0 and
řN
j
Ÿgr
j“0 aj Y
ζfc,d ,red pY q “ śr
.
ei
i“1 p1 ´ Y q
Remark 6.5. The palindromic property (6.1) holds, of course, in particular if Conjecture 6.2 holds. What Conjecture 6.4 suggests is that the reduced graded ideal zeta funcŸ
tions ζfc,dgr,red pY q share key properties with the Hilbert-Poincaré series of graded CohenMacaulay algebras. If there were such an algebra for each c, d, its Cohen-Macaulayness
would explain the nonnegativity of the coefficients aj , whereas their palindromy would
reflect its Gorensteinness (assuming that it is a domain; cf. [13]).
Ÿgr
pY q “
For c “ 2, the validity of Conjecture 6.4 may be seen from the facts that ζf2,d,red
Ÿ
ζf2,d,red pY q and the Hall basis H2,d for f2,d in Table 2.1 is nice and simple in the sense
of [6]. Hence [6, Proposition 4.1] is applicable.
The following is a graded analogue of [9, Conjecture I].
Conjecture 6.6 (Degree of topological zeta function).
¯
´
Ÿ
degs ζfc,dgr,top psq “ ´r.
Conjecture 6.7 (Behaviour of topological zeta function at 8).
ˇ
ˇ
ˇ
ˇ
Ÿ
Ÿ
P Qą0 .
“ p1 ´ Y qr ζfc,dgr,red pY qˇ
s´r ζfc,dgr,top ps´1 qˇ
s“0
Y “1
26
SEUNGJAI LEE AND CHRISTOPHER VOLL
Remark 6.8. We have no interpretation in terms of c and d for the numbers featuring in Conjecture 6.7. If Conjecture 6.4 were to hold for the reasons we explained in
Remark 6.5, they were related to the “multiplicity” of the associated graded algebra.
We close with two conjectures on the behaviour of p-adic and topological graded ideal
zeta functions at s “ 0.
Conjecture 6.9 (Behaviour of topological zeta function at zero). The topological graded
Ÿ
ideal zeta function ζfc,dgr,top psq has a pole of order c at s “ 0 with leading coefficient
řc
ś
ˇ
p´1q i“1 pW piq´1q ci“1 W piq
ˇ
c Ÿgr
´
¯
“ś
s ζfc,d ,top psqˇ
.
ři
c
s“0
j“1 W pjq pW piq ´ 1q!
i“1
Remark 6.10. [9, Conjecture IV (topological form)] asserts that ζfŸc,d ,top psq has a simple
pole at s “ 0 with residue
ˇ
p´1qr´1
ˇ
“
.
sζfŸc,d ,top psqˇ
s“0
pr ´ 1q!
Conjecture 6.11 (Behaviour of p-adic zeta function at zero).
ˇ
Ÿ
śc
ζf grpoq psq ˇˇ
W piq
c,d
¯.
śc
“ ś ´i“1
(6.2)
ˇ
ř
i
c
i“1 ζoW piq psq ˇ
W pjq
s“0
i“1
j“1
Remark 6.12. [9, Conjecture IV (P-adic form)] asserts that
ˇ
ζfŸc,d poq psq ˇˇ
“ 1.
(6.3)
ˇ
ζor psq ˇ
s“0
It seems remarkable that the right hand sides of both (6.2) and (6.3) are independent
of o. Note that s “ 0 is outside the domain of convergence of the series defining ζfŸc,d poq psq
Ÿ
resp. ζfc,dgrpoq psq.
Remark 6.13. Simply replacing the numbers W piq by the ranks of the successive quotients of the upper central series does not extend Conjectures 6.9 and 6.11 to graded
ideal zeta functions of general (non-free) nilpotent Lie rings. Consider, for instance, the
direct product L :“ f2,2 ˆ f2,2 . Then, for all primes p and all finite extensions o of Zp ,
1 ´ t5
.
p1 ´ t3 q2 p1 ´ qt5 q
Indeed, L is the “fundamental graded Lie ring” m6 2 3, whose local graded ideal zeta
functions are recorded (for almost all p) in [11, Section 10, Table 2]. Formula (6.4) may
also easily be derived directly from formulae for the ideal zeta functions of L (cf. [5,
Theorem 2.4]), using the comparison identities recorded in Example 1.6. In any case,
ˇ
p´1q3`1 4 ¨ 2
1
5
ˇ
2 Ÿgr
‰
“ .
“
s ζL,top psqˇ
54
4 ¨ 6 ¨ 3! ¨ 1!
18
s“0
Likewise,
ˇ
Ÿgr
psq ˇˇ
ζLpoq
5
4¨2
1
“ ‰
“ .
ˇ
ˇ
ζo4 psqζo2 psq
9
4¨6
3
(6.4)
Ÿ
gr
psq “ ζo4 psq
ζLpoq
s“0
ENUMERATING GRADED IDEALS IN FREE NILPOTENT LIE RINGS
27
That this value, however, is a nonnegative rational number independent of o, as well as
the coincidence
ˇ
ˇ
1
ˇ
ˇ
Ÿgr
Ÿgr
P Qą0 ,
pY qˇ
“ “ p1 ´ Y q6 ζL,red
ps´1 qˇ
s´6 ζL,top
Y “1
9
s“0
seem to indicate that some of the conjectures made in this section may have generalizations to more general graded ideal zeta functions.
Acknowledgments
Our work was supported by DFG Sonderforschungsbereich 701 “Spectral Structures
and Topological Methods in Mathematics” at Bielefeld University. The first author
was also supported by A23200000 fund from the National Institute for Mathematical
Sciences. We acknowledge numerous helpful conversations with Tobias Rossmann.
References
[1] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics,
vol. 39, Cambridge University Press, Cambridge, 1993.
[2] L. M. Butler, A unimodality result in the enumeration of subgroups of a finite abelian group, Proc.
Amer. Math. Soc. 101 (1987), no. 4, 771–775.
[3] M. P. F. du Sautoy, Counting p-groups and nilpotent groups, Publ. Math. I.H.E.S. 92 (2000), 63–112.
[4] M. P. F. du Sautoy and F. J. Grunewald, Analytic properties of zeta functions and subgroup growth,
Ann. of Math. (2) 152 (2000), 793–833.
[5] M. P. F. du Sautoy and L. Woodward, Zeta functions of groups and rings, Lecture Notes in Mathematics, vol. 1925, Springer-Verlag, Berlin, 2008.
[6] A. Evseev, Reduced zeta functions of Lie algebras, J. Reine Angew. Math. 633 (2009), 197–211.
[7] F. J. Grunewald, D. Segal, and G. C. Smith, Subgroups of finite index in nilpotent groups, Invent.
Math. 93 (1988), 185–223.
[8] M. Hall Jr., A basis for free Lie rings and higher commutators in free groups, Proc. Amer. Math.
Soc. 1 (1950), 575–581.
[9] T. Rossmann, Computing topological zeta functions of groups, algebras, and modules, I, Proc. Lond.
Math. Soc. (3) 110 (2015), no. 5, 1099–1134.
, Stability results for local zeta functions of groups and related structures, arXiv:1504.04164,
[10]
2015.
[11]
, Computing local zeta functions of groups, algebras, and modules, preprint,
arXiv:1602.00919, 2016.
[12] M. M. Schein and C. Voll, Normal zeta functions of the Heisenberg groups over number rings I –
the unramified case, J. Lond. Math. Soc. (2) 91 (2015), no. 1, 19–46.
[13] R. P. Stanley, Hilbert functions of graded algebras, Adv. Math. 28 (1978), 57–83.
[14] C. Voll, Zeta functions of groups and enumeration in Bruhat-Tits buildings, Amer. J. Math. 126
(2004), 1005–1032.
, Normal subgroup growth in free class-2-nilpotent groups, Math. Ann. 332 (2005), 67–79.
[15]
[16]
, Functional equations for zeta functions of groups and rings, Ann. of Math. (2) 172 (2010),
no. 2, 1181–1218.
[17]
, Local functional equations for submodule zeta functions associated to nilpotent algebras of
endomorphisms, preprint, arXiv:1602.07025, 2016.
[18] E. Witt, Treue Darstellung Liescher Ringe, J. Reine Angew. Math. 177 (1937), 152–160.
28
SEUNGJAI LEE AND CHRISTOPHER VOLL
Mathematical Institute, Oxford University, OX2 6GG, United Kingdom
Current address: National Institute for Mathematical Sciences, Daejeon 305-811, South Korea
E-mail address: [email protected]
Fakultät für Mathematik, Universität Bielefeld, D-33501 Bielefeld, Germany
E-mail address: [email protected]
| 4math.GR
|
Differential Privacy on Finite Computers
arXiv:1709.05396v1 [cs.DS] 15 Sep 2017
Victor Balcer∗
Salil Vadhan†
Center for Research on Computation & Society
School of Engineering & Applied Sciences
Harvard University
[email protected], salil [email protected]
September 19, 2017
Abstract
We consider the problem of designing and analyzing differentially private algorithms that
can be implemented on discrete models of computation in strict polynomial time, motivated
by known attacks on floating point implementations of real-arithmetic differentially private
algorithms (Mironov, CCS 2012) and the potential for timing attacks on expected polynomialtime algorithms. We use a case study the basic problem of approximating the histogram of
a categorical dataset over a possibly large data universe X . The classic Laplace Mechanism
(Dwork, McSherry, Nissim, Smith, TCC 2006 and J. Privacy & Confidentiality 2017) does not
satisfy our requirements, as it is based on real arithmetic, and natural discrete analogues, such
as the Geometric Mechanism (Ghosh, Roughgarden, Sundarajan, STOC 2009 and SICOMP
2012), take time at least linear in |X |, which can be exponential in the bit length of the input.
In this paper, we provide strict polynomial-time discrete algorithms for approximate histograms whose simultaneous accuracy (the maximum error over all bins) matches that of the
Laplace Mechanism up to constant factors, while retaining the same (pure) differential privacy
guarantee. One of our algorithms produces a sparse histogram as output. Its “per-bin accuracy” (the error on individual bins) is worse than that of the Laplace Mechanism by a factor
of log |X |, but we prove a lower bound showing that this is necessary for any algorithm that
produces a sparse histogram. A second algorithm avoids this lower bound, and matches the
per-bin accuracy of the Laplace Mechanism, by producing a compact and efficiently computable
representation of a dense histogram; it is based on an (n + 1)-wise independent implementation
of an appropriately clamped version of the Discrete Geometric Mechanism.
∗
Supported by NSF grant CNS-1237235 and CNS-1565387.
http://seas.harvard.edu/~ salil. Supported by NSF grant CNS-1237235, a Simons Investigator Award, and
a grant from the Sloan Foundation.
†
1
1
Introduction
Differential Privacy [DMNS06] is by now a well-established framework for privacy-protective statistical analysis of sensitive datasets. Much work on differential privacy involves an interplay between
statistics and computer science. Statistics provides many of the (non-private) analyses that we
wish to approximate with differentially private algorithms, as well as probabilistic tools that are
useful in analyzing such algorithms, which are necessarily randomized. From computer science,
differential privacy draws upon a tradition of adversarial modeling and strong security definitions,
techniques for designing and analyzing randomized algorithms, and considerations of algorithmic
resource constraints (such as time and memory).
Because of its connection to statistics, it is very natural that much of the literature on differential privacy considers the estimation of real-valued functions on real-valued data (e.g. the
sample mean) and introduces noise from continuous probability distributions (e.g. the Laplace
distribution) to obtain privacy. However, these choices are incompatible with standard computer
science models for algorithms (like the Turing machine or RAM model) as well as implementation
on physical computers (which use only finite approximations to real arithmetic, e.g. via floating
point numbers). This discrepancy is not just a theoretical concern; Mironov [Mir12] strikingly
demonstrated that common floating-point implementations of the most basic differentially private
algorithm (the Laplace Mechanism) are vulnerable to real attacks. Mironov shows how to prevent
his attack with a simple modification to the implementation, but this solution is specific to a single
differentially private mechanism and particular floating-point arithmetic standard. His solution increases the error by a constant factor and is most likely more efficient in practice than the algorithm
we will use to replace the Laplace Mechanism. However, he provides no bounds on asymptotic running time. Gazeau, Miller and Palamidessi [GMP16] provide more general conditions for which
an implementation of real numbers and a mechanism that perturbs the correct answer with noise
maintains differential privacy. However, they do not provide an explicit construction with bounds
on accuracy and running time.
From a theoretical point of view, a more appealing approach to resolving these issues is to avoid
real or floating-point arithmetic entirely and only consider differentially private computations that
involve discrete inputs and outputs, and rational probabilities. Such algorithms are realizable in
standard discrete models of computation. However, some such algorithms have running times that
are only bounded in expectation (e.g. due to sampling from an exponential distribution supported
on the natural numbers), and this raises a potential vulnerability to timing attacks. If an adversary
can observe the running time of the algorithm, it learns something about the algorithm’s coin tosses,
which are assumed to be secret in the definition of differential privacy. (Even if the time cannot
be directly observed, in practice an adversary can determine an upper bound on the running time,
which again is information that is implicitly assumed to be secret in the privacy definition.)
Because of these considerations, we advocate the following principle:
Differential Privacy for Finite Computers:
We should describe how to implement differentially private algorithms on discrete models of computation with strict bounds on running time (ideally polynomial in the bit
length of the input) and analyze the effects of those constraints on both privacy and
accuracy.
Note that a strict bound on running time does not in itself prevent timing attacks, but once we have
such a bound, we can pad all executions to take the same amount of time. Also, while standard
discrete models of computation (e.g. randomized Turing machines) are defined in terms of countable
2
rather than finite resources (e.g. the infinite tape), if we have a strict bound on running time, then
once we fix an upper bound on input length, they can indeed be implemented on a truly finite
computer (e.g. like a randomized Boolean circuit).
In many cases, the above goal can be achieved by appropriate discretizations and truncations
applied to a standard, real-arithmetic differentially private algorithm. However, such modifications
can have a nontrivial price in accuracy or privacy, and thus we also call for a rigorous analysis of
these effects.
In this paper, we carry out a case study of achieving “differential privacy for finite computers”
for one of the first tasks studied in differential privacy, namely approximating a histogram of a
categorical dataset. Even this basic problem turns out to require some nontrivial effort, particularly
to maintain strict polynomial time, optimal accuracy and pure differential privacy when the data
universe is large.
We recall the definition of differential privacy.
Definition 1.1. [DMNS06] Let M : X n → R be a randomized algorithm. We say M is (ε, δ)differentially private if for every two datasets D and D ′ that differ on one row and every subset
S⊆R
Pr[M(D) ∈ S] ≤ eε · Pr[M(D ′ ) ∈ S] + δ
We say an (ε, δ)-differentially private algorithm satisfies pure differential privacy when δ = 0
and say it satisfies approximate differential privacy when δ > 0.
In this paper, we study the problem of estimating the histogram of a dataset D ∈ X n , which is
the vector c = c(D) ∈ NX , where cx is the number of rows in D that have value x. Histograms can
be approximated while satisfying differential privacy using the Laplace Mechanism, introduced in
the original paper of Dwork, McSherry, Nissim and Smith [DMNS06]. Specifically, to obtain (ε, 0)differential privacy, we can add independent noise distributed according to a Laplace distribution,
specifically Lap(2/ε), to each component of c and output the resulting vector c̃. Here Lap(2/ε) is the
continuous, real-valued random variable with probability density function f (z) that is proportional
to exp(−ε · |z|/2). The Laplace Mechanism also achieves very high accuracy in two respects:
Per-Query Error: For each bin x, with high probability we have |c̃x − cx | ≤ O(1/ε).
Simultaneous Error: With high probability, we have maxx |c̃x − cx | ≤ O(log(|X |)/ε).
Note that both of the bounds are independent of the number n of rows in the dataset, and so the
fractional error vanishes linearly as n grows.
Simultaneous error is the more well-studied notion in the differential privacy literature, but
we consider per-query error to be an equally natural concept: if we think of the approximate
histogram c̃ as containing approximate answers to the |X | different counting queries corresponding
to the bins of X , then per-query error captures the error as experienced by an analyst who may be
only interested in one or a few of the bins of c̃. The advantage of considering per-query error is that
it can be significantly smaller than the simultaneous error, as is the case in the Laplace Mechanism
when the data universe X is very large. It is known that both of the error bounds achieved by the
Laplace Mechanism are optimal up to constant factors; no (ε, 0)-differentially private algorithm for
histograms can achieve smaller per-query error or simultaneous error [HT10, BBKN14].
Unfortunately, the Laplace Mechanism uses real arithmetic and thus cannot be implemented on
a finite computer. To avoid real arithmetic, we could use the Geometric Mechanism [GRS12], which
adds noise to each component of c according to the 2-sided geometric distribution, Geo(2/ε), which
is supported on the integers and has probability mass function f (z) ∝ exp(−ε·|z|/2). However, this
3
mechanism uses integers of unbounded size and thus cannot be implemented on a finite computer.
Indeed, while the algorithm can be implemented with a running time that is bounded in expectation
(after reducing ε so that eε/2 and hence all the probabilities are rational numbers), truncating long
executions or allowing an adversary to observe the actual running time can lead to a violation of
differential privacy. Thus, it is better to work with the Truncated Geometric Mechanism of Ghosh,
Roughgarden and Soundararajan [GRS12], where we clamp each noisy count c̃x to the interval
[0, n]. We observe that the resulting probability distribution of c̃x , supported on {0, 1, . . . , n}, can
be described explicitly in terms of cx , ε and n, and it can be sampled in polynomial time using
only integer arithmetic (after ensuring eε/2 is rational). Thus, we obtain:
Theorem 1.2 (Bounded Geometric Mechanism, informal statement of Thm. 3.7). For every finite
X , n and ε ∈ (0, 1], there is an (ε, 0)-differentially private algorithm M : X n → {0, 1, . . . , n}X for
histograms achieving:
• Per-query error O(1/ε).
• Simultaneous error O(log |X |)/ε.
• Strict running time |X | · poly(N ), where N is the bit length of the input (n, ε and a dataset
D ∈ X n ).
We note that while we only consider our particular definition of per-query accuracy, namely
that with high probability |c̃x − cx | ≤ O(1/ε), Ghosh et al. [GRS12] proved that the output of the
Bounded Geometric Mechanism can be used (with post-processing) to get optimal expected loss
with respect to an extremely general class of loss functions and arbitrary priors. The same result
applies to each individual noisy count c̃x output by our mechanism, since each bin is distributed
according to the Bounded Geometric Mechanism (up to a modification of ε to ensure rational
probabilities).
The Bounded Geometric Mechanism is not polynomial time for large data universes X . Indeed,
its running time (and output length) is linear in |X |, rather than polynomial in the bit length
of data elements, which is log |X |. To achieve truly polynomial time, we can similarly discretize
and truncate a variant of the Stability-Based Histogram of Bun, Nissim and Stemmer [BNS16].
This mechanism only adds Lap(2/ε) noise to the nonzero components of cx and then retains only
the noisy values c̃x that are larger than a threshold t = Θ(log(1/δ)/ε). Thus, the algorithm only
outputs a partial histogram, i.e. counts c̃x for a subset of the bins x, with the rest of the counts
being treated as zero. By replacing the use of the Laplace Mechanism with the (rational) Bounded
Geometric Mechanism as above, we can implement this algorithm in strict polynomial time:
Theorem 1.3 (Stability-Based Histogram, informal statement of Thm. 5.2). For every finite
X , n, ε ∈ (0, 1] and δ ∈ (0, 1/n), there is an (ε, δ)-differentially private algorithm M : X n →
{0, 1, . . . , n}⊆X for histograms achieving:
• Per-query error O(1/ε) on bins with true count at least O(log(1/δ)/ε).
• Simultaneous error O(log(1/δ)/ε).
• Strict running time poly(N ), where N is the bit length of the input (n, ε and a dataset
D ∈ X n ).
Notice that the simultaneous error bound of O(log(1/δ)/ε) is better than what is achieved by
the Laplace Mechanism when δ > 1/|X |, and is known to be optimal up to constant factors in this
4
range of parameters (see Theorem 6.1). The fact that this error bound is independent of the data
universe size |X | makes it tempting to apply even for infinite data domains X . However, we note
that when X is infinite, it is impossible for the algorithm to have a strict bound on running time
(as it needs time to read arbitrarily long data elements) and thus is vulnerable to timing attacks
and is not implementable on a finite computer.
Note also that the per-query error bound only holds on bins with large enough true count
(namely, those larger than our threshold t); we will discuss this point further below.
A disadvantage of the Stability-based Histogram is that it sacrifices pure differential privacy. It
is natural to ask whether we can achieve polynomial running time while retaining pure differential
privacy. A step in this direction was made by Cormode, Procopiuc, Srivastava and Tran [CPST11].
They observe that for an appropriate threshold t = Θ(log(|X |)/ε), if we run the Bounded Geometric
Mechanism and only retain the noisy counts c̃x that are larger than t, then the expected number of
bins that remain is less than n + 1. Indeed, the expected number of bins we retain whose true count
is zero (“empty bins”) is less than 1. They describe a method to directly sample the distribution
of the empty bins that are retained, without actually adding noise to all |X | bins. This yields
an algorithm whose output length is polynomial in expectation. However, the output length is
not strictly polynomial, as there is a nonzero probability of outputting all |X | bins. And it is not
clear how to implement the algorithm in expected polynomial time, because even after making the
probabilities rational, they have denominators of bit length linear in |X |.
To address these issues, we consider a slightly different algorithm. Instead of trying to retain
all noisy counts c̃x that are larger than some fixed threshold t, we retain the n largest noisy
counts (since there are at most n nonzero true counts). This results in a mechanism whose output
length is always polynomial, rather than only in expectation. However, the probabilities still have
denominators of bit length linear in |X |. Thus, we show how to approximately sample from this
distribution, to within an arbitrarily small statistical distance δ, at the price of a poly(log(1/δ))
increase in running time. Naively, this would result only in (ε, O(δ))-differential privacy. However,
when δ is significantly smaller than 1/|R|, where R is the range of the mechanism, we can convert
an (ε, δ)-differentially private mechanism to an (ε, 0)-differentially private mechanism by simply
outputting a uniformly random element of R with small probability. (A similar idea for the case
that |R| = 2 has been used in [KLN+ 11, CDK17].) Since our range is of at most exponential size
(indeed at most polynomial in bit length), the cost in our runtime for taking δ ≪ 1/|R| is at most
polynomial. With these ideas we obtain:
Theorem 1.4 (Pure DP Histogram in Polynomial Time, informal statement of Thm. 4.14). For
every finite X , n and ε ∈ (0, 2], there is an (ε, 0)-differentially private algorithm M : X n →
{0, 1, . . . , n}⊆X for histograms achieving:
• Per-query error O(1/ε) on bins with true count at least O(log(|X |)/ε).
• Simultaneous error O(log(|X |)/ε).
• Strict running time poly(N ), where N is the bit length of the input (n, ε and a dataset
D ∈ X n ).
Both Theorems 1.3 and 1.4 only retain per-query error O(1/ε) on bins with a large enough true
count. We also prove a lower bound showing that this limitation is inherent in any algorithm that
outputs a sparse histogram (as both of these algorithms do).
Theorem 1.5 (Lower Bound on Per-Query Error for Sparse Histograms, Theorem 6.2). Suppose
that there is an (ε, δ)-differentially private algorithm M : X n → {0, 1, . . . , n}X for histograms that
5
always outputs histograms with at most n′ nonempty bins and has per-query error at most E on all
bins. Then
min{log |X |, log(1/δ)}
E≥Ω
,
ε
provided that ε > 0, ε2 > δ > 0 and |X | ≥ (n′ )2 .
This lower bound is similar in spirit to a lower bound of [BBKN14], which shows that no (ε, 0)differentially private PAC learner for “point functions” (functions that are 1 on exactly one element
of the domain) can produce sparse functions as hypotheses.
To bypass this lower bound, we can consider algorithms that produce succinct descriptions of
dense histograms. That is, the algorithm can output a polynomial-length description of a function
c̃ : X → [0, n] that can be evaluated in polynomial time, even though X may be of exponential size.
We show that this relaxation allows us to regain per-query error O(1/ε).
Theorem 1.6 (Polynomial-Time DP Histograms with Optimal Per-Query Accuracy, informal
statement of Thm. 7.3). For every finite X , n and ε ∈ (0, 1], there is an (ε, 0)-differentially private
algorithm M : X n → H for histograms (where H is an appropriate class of succinct descriptions of
histograms) achieving:
• Per-query error O(1/ε).
• Simultaneous error O(log(|X |)/ε).
• Strict running time poly(N ), where N is the bit length of the input (n, ε and a dataset
D ∈ X n ) for both producing the description of a noisy histogram c̃ ← M(D) and for evaluating
c̃(x) at any point x ∈ X .
The algorithm is essentially an (n+1)-wise independent instantiation of the Bounded Geometric
Mechanism. Specifically, we release a function h : X → {0, 1}r selected from an (n + 1)-wise
independent family of hash functions, and for each x ∈ X , we view h(x) as coin tosses specifying
a sample from the Bounded Geometric Distribution. That is, we let S : {0, 1}r → [0, n] be an
efficient sampling algorithm for the Bounded Geometric Distribution, and then c̃x = S(h(x)) is our
noisy count for x. The hash function is chosen randomly from the family conditioned on values c̃x
for the nonempty bins x, which we obtain by running the actual Bounded Geometric Mechanism
on those bins. The (n + 1)-wise independence ensures that the behavior on any two neighboring
datasets (which together involve at most n + 1 distinct elements of X ) are indistinguishable in the
same way as in the ordinary Bounded Geometric Mechanism. The per-query accuracy comes from
the fact that the marginal distributions of each of the noisy counts are the same as in the Bounded
Geometric Mechanism. (Actually, we incur a small approximation error in matching the domain of
the sampling procedure to the range of a family of hash functions.)
As far as we know, the only other use of limited independence in constructing differentially
private algorithms is a use of pairwise independence by [BBKN14] in differentially private PAC
learning algorithms for the class of point functions. Although that problem is related to the one we
consider (releasing a histogram amounts to doing “query release” for the class of point functions, as
discussed below), the design and analysis of our algorithm appears quite different. (In particular,
our analysis seems to rely on (n + 1)-wise independence in an essential way.)
Another potential interest in our technique is as another method for bypassing limitations of
synthetic data for query release. Here, we have a large family of predicates Q = {q : X → {0, 1}},
and are interested in differentially private algorithms that, given a dataset D = (x1 , . . . , xn ) ∈ X n ,
output a “summary” M(D) that allows one to approximate the answers to all of the counting queries
6
P
q(D) = i q(xi ) associated with predicates q ∈ Q. For example, if Q is the family of point functions
consisting of all predicates that evaluate to 1 on exactly one point in the data universe X , then this
query release problem amounts to approximating the histogram of D. The fundamental result of
Blum, Ligett, and Roth [BLR13] and successors show that this is possible even for families Q and
data universes X that are of size exponential in n. Moreover, the summaries produced by these
algorithms has the form of a synthetic dataset — a dataset D̂ ∈ X n̂ such that for every query q ∈ Q,
we have q(D̂) ≈ q(D). Unfortunately, it was shown in [UV11] that even for very simple families Q
of queries, such correlations between pairs of binary attributes, constructing such a differentially
private synthetic dataset requires time exponential in the bitlength log |X | of data universe elements.
Thus, it is important to find other ways of representing approximate answers to natural families Q
of counting queries, which can bypass the inherent limitations of synthetic data, and progress along
these lines was made in a variety of works [GRU12, CKKL12, HRS12, TUV12, CTUW14, DNT15].
Our algorithm, and its use of (n + 1)-wise independence, can be seen as yet another representation
that bypasses a limitation of synthetic data (albeit a statistical rather than computational one).
Indeed, a sparse histogram is simply a synthetic dataset that approximates answers to all point
functions, and by Theorem 1.5, our algorithm achieves provably better per-query accuracy than
is possible with synthetic datasets. This raises the question of whether similar ideas can also be
useful in bypassing the computational limitations of synthetic data for more complex families of
counting queries.
2
Preliminaries
Throughout this paper, let N be the set {0, 1, . . .}, N+ be the set {1, 2, . . .}. For n ∈ N, let [n] be
the nonstandard set {0, . . . , n}. Notice that |[n]| = n + 1. Given a set A and finite set B, we define
AB to be the set of length |B| vectors over A indexed by the elements of B.
2.1
Differential Privacy
We define a dataset D ∈ X n to be an ordered tuple of n ≥ 1 rows where each row is drawn from a
discrete data universe X with each row corresponding to an individual. Two datasets D, D ′ ∈ X n
are considered neighbors if they differ in exactly one row.
Definition 2.1. [DMNS06] Let M : X n → R be a randomized algorithm. We say M is (ε, δ)differentially private if for every pair of neighboring datasets D and D ′ and every subset S ⊆ R
Pr[M(D) ∈ S] ≤ eε · Pr[M(D ′ ) ∈ S] + δ
We say an (ε, δ)-differentially private algorithm satisfies pure differential privacy when δ = 0
and say it satisfies approximate differential privacy when δ > 0. Intuitively, the ε captures an
upper bound on an adversary’s ability to determine whether a particular individual is in the dataset.
And the δ parameter represents an upper bound of the probability of a catastrophic privacy breach
(e.g. the entire dataset is released). The common setting of parameters takes ε ∈ (0, 1] to be a
small constant and δ to be negligible in n.
The following properties of differentially private algorithms will be used in some of our proofs.
Lemma 2.2 (post-processing [DMNS06]). Let M : X n → Y be (ε, δ)-differentially private and
f : Y → Z be any randomized function. Then f ◦ M : X n → Z is (ε, δ)-differentially private.
7
Lemma 2.3 (group privacy [DMNS06]). Let M : X n → Y be (ε, δ)-differentially private. Let
D1 , D2 ⊆ X n be datasets such that D2 can be obtained by changing at most m rows of D1 . Then
for all S ⊆ Y
Pr[M(D1 ) ∈ S] ≤ emε · Pr[M(D2 ) ∈ S] + emε · δ/ε
Lemma 2.4 (composition [DL09]). Let M1 : X n → Y1 be (ε1 , δ1 )-differentially private and M2 :
X n → Y2 be (ε2 , δ2 )-differentially private. Define M : X n → Y1 ×Y2 as M(D) = (M1 (D), M2 (D))
for all D ∈ X n . Then M is (ε1 + ε2 , δ1 + δ2 )-differentially private.
2.2
Histograms
For x ∈ X , the point function cx : X n → N is defined to count the number of occurrences of x in
a given dataset, i.e. for D ∈ X n
cx (D) = |{i ∈ {1, . . . n} : Di = x}|
In this paper we focus on algorithms for privately releasing approximations to the values of all point
functions, also known as a histogram. A histogram is collection of bins, one for each element x
in the data universe, with the xth bin consisting of its label x and a count cx ∈ N.
2.2.1
Representations
The input to our algorithms is always a dataset (i.e. an element D ∈ X n ) and the outputs represent
approximate histograms. We consider the following histogram representations as our algorithms’
outputs:
• A vector in NX . We use {c̃x }x∈X to denote a histogram where c̃x ∈ N is the approximate
count for the element x.
• A partial vector h ∈ (X × N)∗ such that each element x ∈ X appears at most once in h with
each pair (x, c̃x ) ∈ X × N interpreted as element x having approximate count c̃x . Elements
x not listed in the partial vector are assumed to have count c̃x = 0. Implicitly an algorithm
can return a partial vector by releasing bins for a subset of X .
• A data structure, encoded as a string, which defines a function h : X → N where h(x),
denoted hx , is the approximate count for x ∈ X and hx is efficiently computable given this
data structure (e.g. time polynomial in the length of the data structure). In Section 7, this
data structure consists of the coefficients of a polynomial, along with some parameters.
Each representation is able to express any histogram over X . The difference between them is the
memory used and the efficiency of computing a count. For example, computing the approximate
count for x ∈ X , when using the data structure representation is bounded by the time it takes to
compute the associated function. But when using partial vectors, one only needs to iterate through
the vector to determine the approximate count.
We define the following class of histograms. Let Hn,n′ (X ) ⊆ NX be the set of all histograms
over X with integer counts in [0, n] (or N when n = ∞) and at most n′ of them nonzero. By using
partial vectors each element of Hn,n′ (X ) can be stored in O(n′ · (log n + log |X |)) bits, which is
shorter than the vector representation when n′ = o(|X |/ log |X |).
8
2.2.2
Accuracy
In order to preserve privacy, our algorithms return histograms with noise added to the counts.
Therefore, it is crucial to understand their accuracy guarantees. So given a dataset D ∈ X n we
compare the noisy count c̃x = M(D)x of x ∈ X (the count released by algorithm M) to its true
count, cx (D). We focus on the following two metrics:
Definition 2.5. A histogram algorithm M : X n → NX has (a, β)-per-query accuracy if
∀D ∈ X n ∀x ∈ X
Pr[|M(D)x − cx (D)| ≤ a] ≥ 1 − β
Definition 2.6. A histogram algorithm M : X n → NX has (a, β)-simultaneous accuracy if
∀D ∈ X n
Pr[∀x ∈ X |M(D)x − cx (D)| ≤ a] ≥ 1 − β
Respectively, these metrics capture the maximum error for any one bin and the maximum error
simultaneously over all bins. Even though simultaneous accuracy is commonly used in differential
privacy, per-query accuracy has several advantages:
• For histograms, one can provable achieve a smaller per-query error than is possible for simultaneous error. Indeed, the optimal simultaneous error for (ε, 0)-differentially private histograms
is a = Θ (log(|X |/β)/ε) whereas the optimal per-query error is a = Θ (log(1/β)/ε), which is
independent of |X | [HT10, BBKN14].
• Per-query accuracy may be easier to convey to an end user of differential privacy. For example,
it is the common interpretation of error bars shown on a graphical depiction of a histogram.
Figure 1: A histogram with error bars
• For many algorithms (such as ours), per-query accuracy is good enough to imply optimal
simultaneous accuracy. Indeed, an algorithm with (a, β)-per-query accuracy also achieves
(a, β · |X |)-simultaneous accuracy (by a union bound).
However, we may not always be able to achieve as good per-query accuracy as we want. So we will
also use the following relaxation which bounds the error only on bins with large enough true count.
Definition 2.7. A histogram algorithm M : X n → NX has (a, β)-per-query accuracy on
counts larger than t if
∀D ∈ X n ∀x ∈ X s.t. cx (D) > t
2.3
Pr[|M(D)x − cx (D)| ≤ a] ≥ 1 − β
Probability Terminology
Definition 2.8. Let Z be an integer-valued random variable.
1. The probability mass function of Z, denoted fZ , is the function fZ (z) = Pr[Z = z] for
all z ∈ Z.
9
2. The cumulative distribution function of Z, denoted FZ , is the function FZ (z) = Pr[Z ≤
z] for all z ∈ Z.
3. The support of Z, denoted supp(Z), is the set of elements for which f (z) 6= 0.
Definition 2.9. Let Y and Z be random variables taking values in discrete range R. The total
variation distance between Y and Z is defined as
∆(Y, Z) = max Pr[Y ∈ A] − Pr[Z ∈ A]
A⊆R
=
1 X
Pr[Z = a] − Pr[Y = a]
·
2
a∈R
Lemma 2.10. Let Y and Z be random variables over discrete range R. Total variation distance
has the following properties:
1. Y and Z are identically distributed, denoted Y ∼ Z, if and only if ∆(Y, Z) = 0.
2. Let T : R → R′ be any function with R′ discrete. Then
∆(T (Y ), T (Z)) ≤ ∆(Y, Z)
3. Let Y1 , Y2 , Z1 and Z2 be random variables over discrete range R. Then
∆((Y1 , Y2 ), (Z1 , Z2 )) ≤ ∆(Y1 , Z1 ) +
2.3.1
max
a∈R,A⊆R
Pr [Y2 ∈ A | Y1 = a] − Pr[Z2 ∈ A | Z1 = a]
Sampling
Because we are interested in the computational efficiency of our algorithms we need to consider the
efficiency of sampling from various distributions.
A standard method for sampling a random variable is via inverse transform sampling.
Lemma 2.11. Let U be uniformly distributed on (0, 1]. Then for any integer-valued random variable
Z we have FZ−1 (U ) ∼ Z where FZ−1 (u) is defined as min{z ∈ supp(Z) : FZ (z) ≥ u}.
If Z, the random variable we wish to sample, has finite support we can compute the inverse
cumulative distribution by performing binary search on supp(Z) to find the minimum. This method
removes the need to compute the inverse function of the cumulative distribution function and is
used in some of our algorithms.
2.3.2
Order Statistics
Definition 2.12. Let Z1 , . . . , Zℓ be integer-valued random variables. The i-th order statistic of
Z1 , . . . , Zℓ denoted Z(i) is the i-th smallest value among Z1 , . . . , Zℓ .
Lemma 2.13. Let Z1 , . . . , Zℓ be i.i.d. integer-valued random variables with cumulative distribution
function F . Then
FZ(ℓ) (z) = (F (z))ℓ
and
(
1
i
FZ(i) |Z(i+1) =vi+1 ,...Z(ℓ) =vℓ (z) = FZ(i) |Z(i+1) =vi+1 (z) =
F (z)/F (vi+1)
for all 1 ≤ i < ℓ and vi+1 ≤ vi+2 ≤ . . . ≤ vℓ all in the support of Z1 .
10
if z > vi+1
otherwise
From this lemma, we can iteratively sample random variables distributed identically to Z(ℓ) ,
Z(ℓ−1) , . . . , Z(i) without having to sample all ℓ of the original random variables. The inverse cumulative distributions for the order statistics are
−1
1/i
−1
1/ℓ
−1
u
·
F
(v
)
(u)
=
F
F
u
(u)
=
F
FZ−1
i+1
Z |Z
=vi+1 ,...Z =vℓ
(ℓ)
(i)
2.4
(i+1)
(ℓ)
Two-Sided Geometric Distribution
A common technique for creating differentially private algorithms is to perturb the desired output
with appropriately scaled Laplace noise. Because our algorithms’ outputs are counts, we focus on
a discrete analogue of the Laplace distribution as in [GRS12].
We say an integer-valued random variable Z follows a two-sided geometric distribution
with scale parameter s centered at c ∈ Z (denoted Z ∼ c + Geo(s)) if its probability mass
function fZ (z) is proportional to e−|z−c|/s . It can be verified that fZ and its cumulative distribution
function FZ are
!
( 1/s
e
· e−(c−z)/s
if z ≤ c
e1/s − 1
−|z−c|/s
e1/s +1
·
e
F
(z)
=
fZ (z) =
Z
1
1/s
−(z−c)/s
e +1
1 − 1/s · e
otherwise
e
+1
for all z ∈ Z. When c is not specified, it is assumed to be 0. The inverse cumulative distribution
of Z is
(
s ln (u) + s ln e1/s + 1 − 1
if u ≤ 1/2
−1
FZ (u) = c +
1/s
−s ln (1 − u) − s ln e + 1
otherwise
or, equivalently,
&
FZ−1 (u) = c + s · sign(1/2 − u) ln(1 − |2u − 1|) + ln
2.5
e1/s + 1
2
!!'
+ ⌊2u⌋ − 1
Model of Computation
We analyze the running time of our algorithms with respect to the w-bit word RAM model
taking w = O(log N ) where N is the bit length of our algorithms’ inputs (D ∈ X n , ε and possibly
some additional parameters). In this model, memory accesses and basic operations (arithmetic,
comparisons and logical) on w-bit words are constant time. In addition, we assume the data
universe X = [m] for some m ∈ N. Some parameters to our algorithms are rational. We represent
rationals by pairs of integers.
Because our algorithms require randomness, we assume that they have access to an oracle that
when given a number d ∈ N+ returns a uniformly random integer between 1 and d inclusively.
3
The Geometric Mechanism
In this section we show how to construct a differentially private histogram using the Laplace Mechanism only requiring integer computations of bounded length.
As shown by Dwork, McSherry, Nissim and Smith [DMNS06], we can privately release a histogram by adding independent and appropriately scaled Laplace noise to each bin. Below we state
a variant that uses discrete noise, formally studied in [GRS12].
11
Algorithm 3.1. GeometricMechanism(D, ε) for D ∈ X n and ε > 0
1. For each x ∈ X , do the following:
(a) Set c̃x to cx (D) + Geo(2/ε) clamped to the interval [0, n]. i.e.
if Zx ≤ 0
0
where Zx = cx (D) + Geo(2/ε).
c̃x = n
if Zx ≥ n
Zx otherwise
(b) Release (x, c̃x ).
Note that the output of this algorithm is a collection of bins (x, c̃x ) which represents a partial
vector, but in this case we have a count for each x ∈ X so it defines a complete vector in Hn,|X |(X ) ⊆
NX . The privacy and accuracy properties of the algorithm are similar to those of the Laplace
Mechanism.
Theorem 3.2. GeometricMechanism(D, ε) has the following properties:
i. GeometricMechanism(D, ε) is (ε, 0)-differentially private [GRS12].
ii. GeometricMechanism(D, ε) has (a, β)-per-query accuracy for
2 1
a=
ln
ε β
iii. GeometricMechanism(D, ε) has (a, β)-simultaneous accuracy for
1
2 |X |
2
ln
ln
≤
a=
ε
ε
β
1 − (1 − β)1/|X |
Proof of ii-iii. Let Z ∼ Geo(2/ε). For any x ∈ X
Pr[|c̃x − cx (D)| ≤ a] ≥ Pr [|Z| ≤ ⌊a⌋]
= 1 − 2 · Pr[Z ≤ −⌊a⌋ − 1]
=1−2·
Now, for a =
l
2
ε
ln β1
m
Pr[|c̃x − cx (D)| ≤ a] ≥ 1 −
e−⌊a⌋·ε/2
eε/2 + 1
2·β
≥1−β
+1
eε/2
Part iii follows similarly after noting by independence of the counts that
Pr[∀x ∈ X |c̃x − cx (D)| ≤ a] ≥ Pr[|Z| ≤ ⌊a⌋]|X |
12
The accuracy bounds up to constant factors match the lower bounds for releasing a differentially
private histogram [HT10, BBKN14].
As presented above, this algorithm needs to store integers of unbounded size since Geo(2/ε)
is unbounded in magnitude. As noted in [GRS12], by restricting the generated noise to a fixed
range we can avoid this problem. However, even when the generated noise is restricted to a fixed
range, generating this noise via inverse transform sampling may require infinite precision. By
appropriately choosing ε, the probabilities of this noise’s cumulative distribution function can be
represented with finite precision, and therefore generating this noise via inverse transform sampling
only requires finite precision.
Proposition 3.3. For k ∈ N, n ∈ N+ and c ∈ [n], the algorithm GeoSample(k, n, c) has output
identically distributed to a two-sided geometric random variable with
scale parameter 2/ε̃ centered
−k
at c clamped to the range [0, n] where we define ε̃ = 2 · ln 1 + 2
. Morever, GeoSample(k, n, c)
has running time poly(k, n).
We have chosen ε̃ so that the cumulative distribution function of a two-sided geometric random
variable with scale parameter 2/ε̃ clamped to [0, n] takes on only rational values with a common
denominator d. Therefore, to implement inverse transform sampling on this distribution we only
need to choose a uniformly random integer in {1, . . . , d} rather than a uniformly random variable
over (0, 1].
Algorithm 3.4. GeoSample(k, n, c) for k ∈ N, n ∈ N+ and c ∈ [n]
1. Let d = (2k+1 + 1)(2k + 1)n−1 .
2. Define the function
k(c−z) 2k + 1 n+z−c
2
n−1−z+c
F (z) = d − 2k(z−c+1) 2k + 1
d
if 0 ≤ z ≤ c
if c < z < n
if z = n
3. Sample U uniformly at random from {1, . . . , d}.
4. Using binary search find the smallest z ∈ [n] such that F (z) ≥ U .
5. Return z.
The function F is obtained by clearing denominators in the cumulative distribution function of
c + Geo(2/ε̃) clamped to [0, n].
Lemma 3.5. Let F (z) be defined as in Algorithm 3.4. Then for z ∈ [n], F (z) ∈ [d] and F (z)/d
equals the cumulative distribution function of c + Geo(2/ε̃) clamped to [0, n].
We prove this lemma after seeing how it implies Proposition 3.3.
Proof of Proposition 3.3. Let U be drawn uniformly at random from {1, . . . , d}. By construction,
for all z ∈ [n]
Pr[GeoSample(k, n, c) ≤ z] = Pr[U ≤ F (z)] = F (z)/d
13
implying GeoSample(k, n, c) ∼ c + Geo(2/ε̃) by Lemma 3.5.
We now bound the running time. Binary search takes O(log n) rounds. The largest number
used is d with bit length O(nk) and all operations are polynomial in the bit length of these numbers.
Therefore, GeoSample(k, n, c) has running time poly(k, n).
Proof of Lemma 3.5. The cumulative distribution function
0
eε̃/2 · e−(c−z)·ε̃/2
ε̃/2
FZ (z) = e +1 1
1 − ε̃/2 · e(c−z)·ε̃/2
e +1
1
of Z ∼ c + Geo(2/ε̃) is
if z < 0
if 0 ≤ z ≤ c
if c < z < n
if z ≥ n
Consider the case when 0 ≤ z ≤ c.
FZ (z) =
−(c−z)
1 + 2−k
eε̃/2
−(c−z)ε̃/2
−k
·
e
=
·
1
+
2
2 + 2−k
eε̃/2 + 1
k c−z
2k + 1
2
= k+1
·
2
+1
2k + 1
2k(c−z)
=
c−z−1
(2k+1 + 1) (2k + 1)
n+z−c
2k(c−z) 2k + 1
=
d
F (z)
=
d
·
2k + 1
2k + 1
n−(c−z)
A similar argument holds when c < z < n and FZ (n) = 1 = F (n)/d. So FZ (z) = F (z)/d for all
z ∈ [n].
Using this algorithm we are ready to construct a private histogram algorithm with bounded
time complexity whose accuracy is identical to that of GeometricMechanism up to constant factors.
Algorithm 3.6. BoundedGeometricMechanism(D, ε) for D ∈ X n and rational ε ∈ (0, 1]
1. Let k = ⌈log(2/ε)⌉.
2. For each x ∈ X , do the following:
(a) Let c̃x = GeoSample(k, n, cx (D)).
(b) Release (x, c̃x ).
Theorem 3.7. Let rational ε ∈ (0, 1] and ε̃ = 2 · ln 1 + 2−⌈log(2/ε)⌉ ∈ (4/9 · ε, ε]. Then
BoundedGeometricMechanism(D, ε) has the following properties:
i. BoundedGeometricMechanism(D, ε) is (ε, 0)-differentially private.
ii. BoundedGeometricMechanism(D, ε) has (a, β)-per-query accuracy for
1
2 1
9
ln
ln
a=
≤
ε̃ β
2ε β
14
iii. BoundedGeometricMechanism(D, ε) has (a, β)-simultaneous accuracy for
|X |
9
2 |X |
ln
ln
≤
a=
ε̃
β
2ε
β
iv. BoundedGeometricMechanism(D, ε) has running time
|X | · poly(N )
where N is the bit length of the algorithm’s input (D ∈ X n and ε).
Proof of i-iii. By Proposition 3.3, GeoSample(k, n, cx(D)) generates a two-sided geometric random
variable with scale parameter 2/ε̃ centered at cx (D) clamped to [0, n]. So this algorithm is identically distributed to GeometricMechanism(D, ε̃). As ε̃ ∈ (4/9 · ε, ε], parts i-iii follows from Theorem
3.2.
Proof of iv. For each x ∈ X , computing cx (D) takes time O(n log |X |) and by Proposition 3.3
GeoSample(k, n, cx (D)) takes times poly(n, log(1/ε)) as k = O(log(1/ε)).
4
Improving the Running Time
For datasets over large domains X , the linear in |X | running time of Algorithm 3.6 can be prohibitive. We present an algorithm that reduces the running time’s dependence on the universe
size from nearly linear to poly-logarithmic based on the observation that most counts are 0 when
n ≪ |X |; this is the same observation made by Cormode, Procopiuc, Srivastava and Tran [CPST11]
to output sparse histograms.
4.1
Sparse Histograms
We start by reducing the output length of GeometricMechanism to release only the bins with the
heaviest (or largest) counts (interpreted as a partial vector).
Algorithm 4.1. KeepHeavy(D, ε) for D ∈ X n and ε > 0
1. For each x ∈ X , set c̃x to cx (D) + Geo(2/ε) clamped to [0, n].
2. Let x1 , . . . , xn+1 be the elements of X with the largest counts in sorted order, i.e.
c̃x1 ≥ c̃x2 ≥ . . . ≥ c̃xn+1 ≥
max
x∈X \{x1 ,...,xn+1 }
c̃x
3. Release h = {(x, c̃x ) : c̃x > c̃xn+1 } ∈ Hn,n (X ).
Observe that the output length has been improved to O(n · (log n + log |X |)) bits compared to
the O(|X | · log n) bits needed to represent the outputs of GeometricMechanism.
Theorem 4.2. KeepHeavy(D, ε) has the following properties:
i. KeepHeavy(D, ε) is (ε, 0)-differentially private.
15
ii. KeepHeavy(D, ε) has (a, β)-per-query accuracy on counts larger than t for
2 2
2 2|X |
ln
ln
and a =
t=2
ε
β
ε β
iii. KeepHeavy(D, ε) has (a, β)-simultaneous accuracy for
2 |X |
a=2
ln
ε
β
Note that, unlike GeometricMechanism, this algorithm only has (O(log(1/β)/ε), β)-per-query
accuracy on counts larger than t = O(log(|X |/β)/ε). This loss is necessary for any algorithm that
outputs a sparse histogram as we will show in Theorem 6.2.
Proof of i. Privacy follows from the (ε, 0)-differential privacy of GeometricMechanism (part i of
Theorem 3.2) along with differential privacy’s closure under post-processing (Lemma 2.2).
To prove the remaining parts, we start with the following lemma.
l
m
|
Lemma 4.3. For any β ′ ∈ (0, 1), let t′ = 2 2ε ln |X
and define the event
β′
Et′ = {∀x ∈ X |c̃x − cx (D)| ≤ t′ /2}
Pr[Et′ ] ≥ 1 − β ′ and Et′ implies that for all x ∈ X such that cx (D) > t′ we have c̃x > c̃xn+1 .
Proof. The probability of Et′ occurring follows from part iii of Theorem 3.2 as the {c̃x }x∈X are
identically distributed to the output of GeometricMechanism(D, ε).
Assume the event Et′ . Then for all x ∈ X such that cx (D) > t′ , we have c̃x > t′ /2 and for all
x ∈ X such that cx (D) = 0, we have c̃x ≤ t′ /2. Because there are at most n distinct elements in
D, we have c̃x > c̃xn+1 for all x ∈ X such that cx (D) > t′ .
Proof of part ii of Theorem 4.2. Let x ∈ X such that cx (D) > t. We have
Pr (|hx − cx (D)| > a) ≤ Pr c̃x ≤ c̃xn+1 + Pr (|c̃x − cx (D)| > a)
≤ β/2 + β/2
by Lemma 4.3 with β ′ = β/2 and t′ = t, and part ii of Theorem 3.2.
Proof of part iii of Theorem 4.2. Let t = 2 ⌈(2/ε) · ln(|X |/β)⌉. The event Et in Lemma 4.3 occurs
with probability at least 1 − β. Assume Et . By Lemma 4.3, for all x ∈ X such that cx (D) > t
we have c̃x > c̃xn+1 . This implies |hx − cx (D)| ≤ t/2. For the remaining x ∈ X we trivially have
|hx − cx (D)| ≤ t as hx = 0.
However, as described KeepHeavy still requires adding noise to the count of every bin. The
following algorithm M1 : X n × (0, 1] → Hn,n (X ) simulates KeepHeavy by generating a candidate
set of heavy bins from which only the heaviest are released. This candidate set is constructed from
all bins with nonzero true count and a sample representing the bins with a true count of 0 that
have the heaviest noisy counts.
16
Algorithm 4.4. M1 (D, ε) for D ∈ X n , ε > 0 and |X | ≥ 2n + 1
1
1. Let A = {x ∈ X : cx (D) > 0} and w = |X \ A|.
2. For each x ∈ A, set c̃x to cx (D) + Geo(2/ε) clamped to [0, n].
3. Pick a uniformly random sequence (q0 , . . . , qn ) of distinct elements from X \ A.
4. Sample (c̃q0 , . . . , c̃qn ) from the joint distribution of the order statistics (Z(w) , . . . , Z(w−n) ) where
Z1 , . . . , Zw are i.i.d. and distributed as Geo(2/ε) clamped to [0, n].
5. Sort the elements of A ∪ {q0 , . . . , qn } as x1 , . . . , x|A|+n+1 such that c̃x1 ≥ . . . ≥ c̃x|A|+n+1 .
6. Release h = {(x, c̃x ) : x ∈ {x1 , . . . , xn } and c̃x > c̃xn+1 } ∈ Hn,n (X ).2
Proposition 4.5. M1 (D, ε) is identically distributed to KeepHeavy(D, ε).
Proof. Let {ĉx }x∈X be the noisy counts set by KeepHeavy(D, ε) and let x̂1 , . . . , x̂n+1 be the sorted
ordering defined by these counts. We have c̃x ∼ ĉx for all x ∈ A and the Zi ’s are identically
distributed to {ĉx }x∈X \A .
{(qi , c̃qi )}ni=0 is identically distributed to the n + 1 bins with heaviest counts of {(x, ĉx )}x∈X \A .
Let the random variable B be the set of the labels of the n + 1 bins with heaviest counts of
{(x, ĉx )}x∈X \A . Therefore,
h = {(x, c̃x ) : x ∈ {x1 , . . . , xn } and c̃x > c̃xn+1 }
= {(x, c̃x ) : x ∈ A ∪ {q0 , . . . , qn } and c̃x ≥ c̃xn+1 }
∼ {(x, ĉx ) : x ∈ A ∪ B and ĉx ≥ ĉx̂n+1 }
= {(x, ĉx ) : x ∈ X and ĉx ≥ ĉx̂n+1 }
which shows that M1 (D, ε) is identically distributed to KeepHeavy(D, ε).
In order to sample from the order statistics used by M1 we construct the following algorithm
similar to GeoSample (Algorithm 3.4) from the previous section.
Proposition 4.6. Let k ∈ N and n, w ∈ N+ . Let v ∈ [n] and i ∈ {1, . . . , w}. Define the i.i.d.
random variables Z1 , . . . , Zw with each identically distributed to Geo(2/ε̃) clamped to [0, n] where
ε̃ = 2 · ln(1 + 2−k ). The following subroutine OrdSample(k, n, v, i) is identically distributed to the
i-th order statistic Z(i) conditioned on Z(i+1) = v. Also, OrdSample(k, n, v, i) has running time
poly(n, k, i).
1
|X | ≥ 2n + 1 ensures that |X \ A| ≥ n + 1. One can use GeometricMechanism(D, ε) when |X | ≤ 2n.
If instead we used continuous noise this last step is equivalent to releasing the n heaviest bins. However, in the
discrete case, where ties can occur, from the set A ∪ {x1 , . . . , xn } we cannot determine all bins with a count tied for
the n-th heaviest as there may be many other noisy counts tied with c̃xn . As a result, we only output the bins with
a strictly heavier count than c̃xn+1 .
2
17
As in BoundedGeometricMechanism (Algorithm 3.6), we have chosen ε̃ so that the cumulative
distribution function of Geo(2/ε̃) takes rational values with a common denominator of d. Therefore,
the cumulative distribution function of Z(i) conditioned on Z(i+1) = v is also rational and we can
sample from it with finite precision.
Algorithm 4.7. OrdSample(k, n, v, i) for k ∈ N; n, i ∈ N+ and v ∈ [n]
1. Let d = (2k+1 + 1)(2k + 1)n−1 .
2. Define the function
(
n−1−z
d − 2k(z+1) 2k + 1
F (z) =
d
if 0 ≤ z < n
if z = n
3. Sample U uniformly at random from {1, . . . , F (v)i }.
4. Using binary search find the smallest z ∈ [v] such that F (z)i ≥ U .
5. Return z.
Proof. By Lemma 3.5, F (z)/d is the cumulative distribution of Geo(2/ε̃) clamped to [0, n]. Therefore, by Lemma 2.13,
FZ(i) |Z(i+1) =v (z) =
F (z)/d
F (v)/d
i
=
F (z)
F (v)
i
for z ∈ [v]. Let U be drawn uniformly at random from {1, . . . , F (v)i }. Then, by construction,
i
Pr[OrdSample(k, n, v, i) ≤ z] = Pr[U ≤ F (z) ] =
F (z)
F (v)
i
implying OrdSample(k, n, v, i) ∼ Z(i) | Z(i+1) = v.
The binary search takes O(log n) iterations. Each iteration has running time polynomial in
the bit length of the numbers used. Therefore, this algorithm has running time poly(log d, i) =
poly(n, k, i).
Now from M1 we replace sampling from the joint distribution of the order statistics with
iterative calls to OrdSample to get the following algorithm.
18
Algorithm 4.8. M2 (D, ε) for D ∈ X n , rational ε ∈ (0, 1] and |X | ≥ 2n + 1
1. Let k = ⌈log(2/ε)⌉.
2. Let A = {x ∈ X : cx (D) > 0} and w = |X \ A|.
3. For each x ∈ A, let c̃x = GeoSample(k, n, cx (D)).
4. Pick a uniformly random sequence (q0 , . . . , qn ) of distinct elements from X \ A.
5. Let c̃q0 = OrdSample(k, n, n, w).
6. For each i ∈ {1, . . . , n}, let c̃qi = OrdSample(k, n, c̃qi−1 , w − i).
7. Sort the elements of A ∪ {q0 , . . . , qn } as x1 , . . . , x|A|+n+1 such that c̃x1 ≥ . . . ≥ c̃x|A|+n+1 .
8. Release h = {(x, c̃x ) : x ∈ {x1 , . . . , xn } and c̃x > c̃xn+1 } ∈ Hn,n (X ).
Theorem 4.9. Let rational ε ∈ (0, 1] and ε̃ = 2 · ln(1 + 2−⌈log(2/ε)⌉ ) ∈ (4/9 · ε, ε]. M2 (D, ε) is
identically distributed to KeepHeavy(D, ε̃). Therefore,
i. M2 (D, ε) is (ε, 0)-differentially private.
ii. M2 (D, ε) has (a, β)-per-query accuracy on counts larger than t for
2 2
2 2|X |
ln
ln
and a =
t=2
ε̃
β
ε̃ β
iii. M2 (D, ε) has (a, β)-simultaneous accuracy for
2 |X |
ln
a=2
ε̃
β
Proof. By Proposition 4.6, we have (c̃q0 , . . . , c̃qn ) is identically distributed to the joint distribution
(Z(w) , . . . , Z(w−n) ) used by M1 . Therefore, this algorithm is identically distributed to M1 (D, ε̃)
and, then by Proposition 4.5, identically distributed to KeepHeavy(D, ε̃). Parts i to iii follow from
Theorem 4.2.
This algorithm only has an output of length O(n · (log n + log |X |)). However, its running time
depends polynomially on |X | since sampling the wth order statistic, c̃q0 , using OrdSample takes time
polynomial in w ≥ |X | − n. Indeed, this is necessary since the distribution of the order statistic
Z(w) has probabilities that are exponentially small in w. 3
4.2
An Efficient Approximation
To remedy the inefficiency of M2 we consider an efficient algorithm that approximates the output
distribution of M2 .
3
Notice Pr[Z(w) = 0] = Pr[Z = 0]w where Z ∼ Geo(2/ε̃). And Pr[Z = 0]w =
coins to sample from this distribution.
19
eε̃/2 −1
eε̃/2 +1
w
. So we need to toss Ω(w)
Theorem 4.10. There exists an algorithm NonzeroGeometric : X n × (0, 1] × (0, 1) → Hn,n (X )
such that for all input datasets D ∈ X n , rational ε ∈ (0, 1] and δ ∈ (0, 1) such that 1/δ ∈ N
i. ∆ (M2 (D, ε), NonzeroGeometric(D, ε, δ)) ≤ δ.
ii. NonzeroGeometric(D, ε, δ) is (ε, (eε + 1) · δ)-differentially private.
iii. Moreover, the running time of NonzeroGeometric(D, ε, δ) is
poly(N )
where N is the bit length of this algorithm’s input (D, ε and δ).
Note that this algorithm only achieves (ε, δ)-differential privacy. By reducing δ, the algorithm
better approximates M2 , improving accuracy, at the cost of increasing running time (polynomial
in log(1/δ)). This is in contrast to most (ε, δ)-differentially private algorithms such as the stability
based algorithm of Section 5, where one needs n ≥ Ω(log(1/δ)/ε) to get any meaningful accuracy.
Now, we convert NonzeroGeometric to a pure differentially private algorithm by mixing it with
a uniformly random output inspired by a similar technique in [KLN+ 11, CDK17].
Algorithm 4.11. M∗ (D, γ) for D ∈ X n and rational γ ∈ (0, 1)
1. With probability 1 − γ release M′ (D).
2. Otherwise release a uniformly random element of R.
Lemma 4.12. Let M : X n → R be (ε, 0)-differentially private with discrete range R. Suppose
algorithm M′ : X n → R satisfies ∆ (M(D), M′ (D)) ≤ δ for all input datasets D ∈ X n with
parameter δ ∈ [0, 1). Then the algorithm M∗ has the following properties:
i. M∗ (D, γ) is (ε, 0)-differentially private whenever
δ≤
eε − 1
γ
1
·
·
eε + 1 1 − γ |R|
ii. M∗ (D, γ) has running time upper bounded by the sum of the bit length to represent γ, the
running time of M′ (D) and the time required to sample a uniformly random element of R.
By taking γ and δ small enough and satisfying the constraint in part i, the algorithm M∗
satisfies pure differential privacy and has nearly the same utility as M (due to having a statistical
distance at most γ + δ from M) while allowing for a possibly more efficient implementation since
we only need to approximately sample from the output distribution of M.
Proof of i. For all neighboring datasets D, D ′ ∈ X n and all r ∈ R
1
|R|
1
≤γ·
|R|
1
≤γ·
|R|
1
≤γ·
|R|
Pr[M∗ (D, γ) = r] = γ ·
+ (1 − γ) · Pr[M′ (D) = r]
+ (1 − γ) · (Pr[M(D) = r] + δ)
+ (1 − γ) eε · Pr[M(D ′ ) = r] + δ
+ (1 − γ) eε · Pr[M′ (D ′ ) = r] + δ + δ
20
Rearranging terms and using the upper bound on δ yields
1
+ (eε + 1) (1 − γ) · δ
|R|
1
1
ε
′
′
ε
≤ e (1 − γ) · Pr[M (D ) = r] + γ ·
+ (e − 1) γ ·
|R|
|R|
1
= eε (1 − γ) · Pr[M′ (D ′ ) = r] + γ ·
|R|
ε
∗
′
= e · Pr[M (D , γ) = r]
Pr[M∗ (D, γ) = r] = eε (1 − γ) · Pr[M′ (D ′ ) = r] + γ ·
Proof of ii. This follows directly from the construction of M∗ .
We can apply this lemma to NonzeroGeometric and under reasonable settings of parameters
get accuracy bounds identical to M2 up to constant factors.
Algorithm 4.13. PureNonzeroGeometric(D, ε, β) for D ∈ X n , rational ε ∈ (0, 1] and 1/β ∈ N
1. With probability 1 − β/3 release NonzeroGeometric(D, ε, δ) with
δ=
2−⌈log(1/ε)⌉ β
1
· ·
3
3 |Hn,n (X )|
2. Otherwise release a uniformly random element of Hn,n (X ).
Theorem 4.14. Let rational ε ∈ (0, 1], ε̃ = 2 · ln 1 + 2−⌈log(2/ε)⌉ ∈ (4/9 · ε, ε] and 1/β ∈ N.
PureNonzeroGeometric(D, ε, β) has the following properties:
i. PureNonzeroGeometric(D, ε, β) is (ε, 0)-differentially private.
ii. PureNonzeroGeometric(D, ε, β) has (a, β)-per-query accuracy on counts larger than t for
2 6
2 6|X |
ln
ln
and a =
t=2
ε̃
β
ε̃ β
iii. PureNonzeroGeometric(D, ε, β) has (a, β)-simultaneous accuracy for
2 3|X |
a=2
ln
ε̃
β
iv. PureNonzeroGeometric(D, ε, β) has running time
poly(N )
where N is the bit length of this algorithm’s input (D ∈ X n , ε and β).
Proof of i. Notice that 1/δ ∈ N (a constraint needed for NonzeroGeometric). Privacy follows from
Lemma 4.12 by taking M = M2 , M′ = NonzeroGeometric, γ = β/3 and R = Hn,n (X ) as
δ=
1
eε − 1
β/3
1
2−⌈log(1/ε)⌉ β
· ·
≤ ε
·
·
3
3 |Hn,n (X )|
e + 1 1 − β/3 |Hn,n (X )|
21
Proof of ii-iii. For any D ∈ X n and x ∈ X such that cx (D) > t define the set G = {h ∈ Hn,n (X ) :
|hx − cx (D)| ≤ a}. By construction,
Pr[PureNonzeroGeometric(D, ε, β) ∈ G] ≥ Pr[NonzeroGeometric(D, ε, δ) ∈ G] − β/3
where δ is defined in Algorithm 4.13. Notice that δ ≤ β/3. So by Theorem 4.10,
Pr[NonzeroGeometric(D, ε, δ) ∈ G] − β/3 ≥ Pr[M2 (D, ε) ∈ G] − 2 · β/3
And by part ii of Theorem 4.9, we have
Pr[M2 (D, ε) ∈ G] − 2 · β/3 ≥ 1 − β
Similarly, we can bound the simultaneous accuracy by using part iii of Theorem 4.9.
Proof of iv. By Lemma A1.1, it takes poly(n, log |X |) time to sample a uniformly random element
of Hn,n (X ) and compute |Hn,n (X )|. And by Theorem 4.10, NonzeroGeometric(D, ε, δ) with
δ=
2−⌈log(1/ε)⌉ β
1
· ·
3
3 |Hn,n (X )|
has running time poly(N ). So by Lemma 4.12, PureNonzeroGeometric has the desired running
time.
4.3
Construction of NonzeroGeometric
We finish this section with the construction of NonzeroGeometric. Notice that M2 passes arguments to OrdSample that result in OrdSample exponentiating an integer, which represents the
numerator of a fraction a/b = F (z)/F (v), to a power i ≥ |X | − n. We want to ensure that numbers
used by OrdSample do not exceed some maximum s.
The following algorithm will approximate s · (a/b)i by using repeated squaring and truncating
each intermediate result to keep the bit length manageable. The following lemma provides a bound
on the error and on the running time.
Proposition 4.15. There is an algorithm ExpApprox(a, b, i, s) such that for all b, i, s ∈ N+ and
a ∈ [b]:
i. ExpApprox(a, b, i, s) is a nondecreasing function in a and ExpApprox (a, b, i, s) = s when a = b.
ii. ExpApprox(a, b, i, s) satisfies the accuracy bound of
i
ExpApprox (a, b, i, s) a i
≤2·
−
s
b
s
iii. ExpApprox(a, b, i, s) has running time poly(log b, log i, log s).
Proof. The algorithm is defined as follows.
Algorithm 4.16. ExpApprox(a, b, i, s) for b, i, s ∈ N+ and a ∈ [b]
1. If i = 1 return as
b .
(
r 2 /s
2. Otherwise let r = ExpApprox (a, b, ⌊i/2⌋, s) and return
(a/b) · (r 2 /s)
22
if i is even
if i is odd
Proof of i. We proceed by induction. The case when i = 1 is trivial. Let i > 1 and assume that
ExpApprox(a, b, j, s) is a nondecreasing function in a for all j < i. Consider a ≤ a′ ≤ b. If i is odd,
then
a ExpApprox(a, b, ⌊i/2⌋, s)2
·
ExpApprox(a, b, i, s) =
b
s
′
a ExpApprox(a′ , b, ⌊i/2⌋, s)2
·
= ExpApprox(a′ , b, i, s)
≤
b
s
Likewise, the result holds when i is even. Therefore, ExpApprox(a, b, i, s) is a nondecreasing function
in a. It is trivial that ExpApprox (a, b, i, s) = s when a = b.
Proof of ii. For ease of notation define EA(i) = ExpApprox (a, b, i, s) as the other parameters do not
change in our analysis. By construction
EA (i) =
(
1
where mi =
a/b
mi · EA (⌊i/2⌋)2
+ ωi
s
if i is even
and |ωi | ≤ 1. We can now bound the error as
if i is odd
EA (i) a i
EA (i) mi · EA (⌊i/2⌋)2
mi · EA (⌊i/2⌋)2 a i
≤
−
−
−
+
s
b
s
s2
s2
b
EA (⌊i/2⌋) a ⌊i/2⌋
EA (⌊i/2⌋) a ⌊i/2⌋
|ωi |
+ mi ·
+
·
−
≤
s
s
b
s
b
⌊i/2⌋
EA (⌊i/2⌋)
a
1
−
≤ +2·
s
s
b
When i = 1, we have
EA(1) a
=
−
s
b
as
b
+ ω1 a
1
≤
−
s
b
s
So solving the recurrence gives
i
EA (i) a i
≤2·
−
s
b
s
Proof of iii. The running time follows from the observation that a call to ExpApprox (a, b, i, s) makes
at most O(log i) recursive calls with the remaining operations polynomial in the bit lengths of the
numbers used.
Now we can modify OrdSample using ExpApprox to keep the bit lengths of its numbers from
becoming too large, yielding an efficient algorithm whose output distribution is close to that of
OrdSample.
23
Algorithm 4.17. ApproxOrdSample(k, n, v, i, s) for k, n, i, s ∈ N+ and v ∈ [n]
1. Let d = (2k+1 + 1)(2k + 1)n−1 .
2. Define the function
(
n−1−z
d − 2k(z+1) 2k + 1
F (z) =
d
if 0 ≤ z < n
if z = n
3. Sample U uniformly from {1, . . . , s}.
4. Using binary search find the smallest z ∈ [v] such that
ExpApprox (F (z), F (v), i, s) ≥ U
5. Return z.
Proposition 4.18. Let k, n, i, s ∈ N+ and v ∈ [n]. Then
i
∆(OrdSample(k, n, v, i), ApproxOrdSample(k, n, v, i, s)) ≤ 2 · (n + 1) ·
s
In addition, ApproxOrdSample(k, n, v, i, s) has a running time of poly(k, n, log i, log s).
Proof. Let Y ∼ OrdSample(k, n, v, i) and Ye ∼ ApproxOrdSample(k, n, v, i, s). Let z ∈ [v]. Then
Pr[Y = z] =
F (z)
F (v)
i
−
F (z − 1)
F (v)
i
where we define F (−1) = 0. By part i of Proposition 4.15, ExpApprox(F (z), F (v), i, s) is an
increasing function in z. Therefore,
Pr[Ye = z] =
ExpApprox(F (z), F (v), i, s) ExpApprox(F (z − 1), F (v), i, s)
−
s
s
Therefore, by the triangle inequality and part ii of Proposition 4.15
| Pr[Y = z] − Pr[Ye = z]| ≤
z
X
z ′ =z−1
ExpApprox(F (z ′ ), F (v), i, s)
−
s
F (z ′ )
F (v)
i
i
≤4·
s
Summing over z ∈ [v] yields the desired bound on total variation distance (Definition 2.9).
The running time is dominated by the O(log n) calls to ExpApprox(F (z), F (v), i, s). By part iii
of Proposition 4.15, each call takes time
poly(log F (v), log i, log s) = poly(k, n, log i, log s)
as log F (v) ≤ log d = O(nk). So overall ApproxOrdSample has the desired running time.
Because M2 samples from a joint distribution obtained by iterated calls to OrdSample we must
also consider the accumulated distance between iterated calls to OrdSample and iterated calls to
ApproxOrdSample.
24
Corollary 4.19. Let k, n, w, s ∈ N+ with w > n. Consider the following random variables:
• Y0 ∼ OrdSample(k, n, n, w)
• Yj ∼ OrdSample(k, n, Yj−1 , w − j) for each j ∈ {1, . . . , n}
• Ye0 ∼ ApproxOrdSample(k, n, n, w, s)
• Yej ∼ ApproxOrdSample(k, n, Yej−1 , w − j, s) for each j ∈ {1, . . . , n}
Then
w
∆ (Y0 , . . . , Yn ), (Ye0 , . . . , Yen ) ≤ 2 · (n + 1)2 ·
s
Proof. By part iii of Lemma 2.10 and Proposition 4.18,
n
w
X
e
e
e
∆ Yj |Yj−1 , Yej |Yej−1 ≤ 2 · (n + 1)2 ·
∆ (Y0 , . . . , Yn ), (Y0 , . . . , Yn ) ≤ ∆ Y0 , Y0 +
s
j=1
where
∆ Yj |Yj−1 , Yej |Yej−1 =
max
x∈[n],S⊆[n]
n
Pr[Yj ∈ S|Yj−1 = x] − Pr[Yej ∈ S|Yej−1 = x]}
o
We are ready to state the mechanism NonzeroGeometric and show it satisfies Theorem 4.10.
It is identical to M2 except we replace calls to OrdSample with calls to ApproxOrdSample.
Algorithm 4.20. NonzeroGeometric(D, ε, δ) for D ∈ X n , rational ε ∈ (0, 1], 1/δ ∈ N and
|X | ≥ 2n + 1
1. Let k = ⌈log(2/ε)⌉ and s = 2 · (n + 1)2 · |X |/δ.
2. Let A = {x ∈ X : cx (D) > 0} and let w = |X \ A|.
3. For each x ∈ A, let c̃x = GeoSample(k, n, cx (D)).
4. Pick a uniformly random sequence (q0 , . . . , qn ) of distinct elements from X \ A.
5. Let c̃q0 = ApproxOrdSample(k, n, n, w, s).
6. For each i ∈ {1, . . . , n}, let c̃qi = ApproxOrdSample(k, n, c̃qi−1 , w − i, s).
7. Sort the elements of A ∪ {q0 , . . . , qn } as x1 , . . . , x|A|+n+1 such that c̃x1 ≥ . . . ≥ c̃x|A|+n+1 .
8. Release h = {(x, c̃x ) : x ∈ {x1 , . . . , xn } and c̃x > c̃xn+1 } ∈ Hn,n (X ).
Theorem 4.10 (restated). The algorithm NonzeroGeometric(D, ε, δ) satisfies for all input datasets
D ∈ X n , rational ε ∈ (0, 1] and δ ∈ (0, 1) such that 1/δ ∈ N
i. ∆ (M2 (D, ε), NonzeroGeometric(D, ε, δ)) ≤ δ.
ii. NonzeroGeometric(D, ε, δ) is (ε, (eε + 1) · δ)-differentially private.
25
iii. Moreover, the running time of NonzeroGeometric(D, ε, δ) is
poly(N )
where N is the bit length of this algorithm’s input (D, ε and δ).
Proof of i. Let M∗2 : X n × (0, 1] → Hn,2n+1 (X ) be the algorithm M2 except, instead of releasing
the heaviest bins, M∗2 releases the bins for all elements of A ∪ {x0 , . . . , xn } (i.e. M∗2 releases (y, c̃y )
for all y ∈ A and (xi , c̃xi ) for all i ∈ [n]). Similarly, we define NonzeroGeometric∗ with respect to
NonzeroGeometric.
Notice that M∗2 and NonzeroGeometric∗ have the same distribution overs the bins with nonzero
true count. Only on the bins with counts sampled using OrdSample and ApproxOrdSample respectively do their output distributions differ. As a result, we can apply Corollary 4.19 to the output
distributions of M∗2 and NonzeroGeometric∗ . So for all D ∈ X n
w
∆ (M∗2 (D, ε), NonzeroGeometric∗ (D, ε, δ)) ≤ 2 · (n + 1)2 ·
≤δ
s
Now we consider the effect of keeping the heaviest counts. Define T : Hn,2n+1 → Hn,n (X ) to be
the function that sets counts not strictly larger than the (n + 1)-heaviest count of its input to 0.
Notice that T ◦ M∗2 ∼ M2 and T ◦ NonzeroGeometric∗ ∼ NonzeroGeometric. So for all D ∈ X n ,
by part ii of Lemma 2.10,
∆(M2 (D, ε), NonzeroGeometric(D, ε, δ)) = ∆(T (M∗2 (D, ε)), T (NonzeroGeometric∗ (D, ε, δ)))
≤ ∆(M∗2 (D, ε), NonzeroGeometric∗ (D, ε, δ))
≤δ
Proof of ii. Let D and D ′ be neighboring datasets. Let S ⊆ Hn,n (X ). By the previous part and
part i of Theorem 4.9,
Pr[NonzeroGeometric(D, ε, δ) ∈ S] ≤ Pr[M2 (D, ε) ∈ S] + δ
≤ eε · Pr[M2 (D ′ , ε) ∈ S] + δ
≤ eε · Pr[NonzeroGeometric(D ′ , ε, δ) ∈ S] + δ + δ
Therefore, NonzeroGeometric(D, ε, δ) is (ε, (eε + 1) · δ)-differentially private.
Proof of iii. We consider the running time at each step. Construction of the true histogram takes
O(n log |X |) time. And by Proposition 3.3, the at most n calls to GeoSample take poly(log(1/ε), n)
time. Sampling n random bin labels from X takes O(n log |X |) time.
Now, NonzeroGeometric makes n + 1 calls to ApproxOrdSample with no argument exceeding
each term of (k, n, n, |X |, s) respectively. So by Proposition 4.18, these calls take time
poly(log(1/ε), n, log |X |, log s) ≤ poly(log(1/ε), n, log |X |, log(1/δ))
Sorting the at most 2n + 1 elements of A ∪ {q0 , . . . , qn } and then releasing the heaviest counts
takes O(n log n + n log |X |) time.
Therefore, overall NonzeroGeometric has the desired running time.
We have constructed algorithms for releasing a differentially private histogram, both pure and
approximate, with running time polynomial in log |X | and simultaneous accuracy matching that of
GeometricMechanism up to constant factors.
26
5
Removing the Dependence on Universe Size
When we would like to have accuracy independent of |X |, we can use an approximate differentially private algorithm based on stability techniques [BNS16] (Proposition 2.20). We present a
reformulation of their algorithm using two-sided geometric noise instead of Laplace noise.
Algorithm 5.1. StabilityHistogram(D, ε, b) for D ∈ X n , ε > 0 and b ∈ [n]
1. Let A = {x ∈ X : cx (D) > 0}.
2. For each x ∈ A, set c̃x to cx (D) + Geo(2/ε) clamped to [0, n].
3. Release h = {(x, c̃x ) : x ∈ A and c̃x > b} ∈ Hn,n (X ).
Note that we only release counts for x ∈ X whose true count is nonzero, namely elements in
the set A. Thus, the output length is O(n · (log n + log |X |)). However, releasing the set A is not
(ε, 0)-differentially private because this would distinguish between neighboring datasets: one with
a count of 0 and the other with a count of 1 for some element x ∈ X . Thus, we only release noisy
counts c̃x that exceed a threshold b. If b is large enough, then a count of 1 will only be kept with
small probability, yielding approximate differential privacy.
Theorem 5.2. StabilityHistogram(D, ε, b) has the following properties:
i. StabilityHistogram(D, ε, b) is (ε, δ)-differentially private provided that
b≥1+
2 1
ln
ε δ
ii. StabilityHistogram(D, ε, b) has (a, β)-per-query accuracy on counts larger than t for
2 1
2 1
t=b+
ln
ln
and a =
ε β
ε β
iii. StabilityHistogram(D, ε, b) has (a, β)-simultaneous accuracy for
2 n
a=b+
ln
ε β
Proof of i. Let D and D ′ be neighboring datasets. Let x ∈ X such that cx (D) 6= cx (D ′ ) and
S ⊆ [n]. There are 3 cases to consider:
• cx (D) ≥ 1 and cx (D ′ ) ≥ 1. Then c̃x ∼ cx (D) + Geo(2/ε) and similarly on the neighboring
database we have a noisy count c̃′x ∼ cx (D ′ ) + Geo(2/ε). So by the differential privacy of
GeometricMechanism (part i of Theorem 3.2) we have
Pr[M(D)x ∈ S] ≤ eε/2 · Pr[M(D ′ )x ∈ S]
27
• cx (D) = 1 and cx (D ′ ) = 0. Notice Pr[M(D ′ )x 6= 0] = 0. c̃x is distributed as 1 + Geo(2/ε)
clamped to [0, n]. Thus,
2 1
Pr[M(D)x 6= 0] = Pr[c̃x > b] ≤ Pr c̃x > 1 + ln
ε δ
δ
2 1
δ
≤
≤ Pr Z > ln
= ε/2
ε δ
2
e +1
where Z ∼ Geo(2/ε). Therefore,
Pr[M(D)x ∈ S] ≤ Pr[M(D ′ )x ∈ S] + δ/2
• cx (D) = 0 and cx (D ′ ) = 1. This case follows similarly to the previous case by considering
Pr[M(D)x = 0].
Then overall
Pr[M(D)x ∈ S] ≤ eε/2 · Pr[M(D ′ )x ∈ S] + δ/2
Because there are at most two bins on which D and D ′ have differing counts and each count c̃x is
computed independently, by Lemma 2.4, this algorithm is (ε, δ)-differentially private.
Proof of ii. Let x ∈ X such that cx (D) > t. Then x ∈ A. So by part ii of Theorem 3.2, we have
Pr [|c̃x − cx (D)| ≤ a] ≥ 1 − β. |c̃x − cx (D)| ≤ a implies c̃x > t − a = b. Thus, (x, c̃x ) is included in
the output of StabilityHistogram(D, ε, b) giving the desired accuracy.
Proof of iii. Notice that the counts of elements not in A are trivially accurate. Therefore, we only
need to consider the counts of elements in A. By part iii of Theorem 3.2,
2 |A|
Pr ∀x ∈ A |c̃x − cx (D)| ≤
ln
≥1−β
ε
β
The final step can increase the error additively by at most b. Also, |A| ≤ n. Therefore,
2 n
ln
≥1−β
Pr ∀x ∈ X |hx − cx (D)| ≤ b +
ε β
By using GeoSample (see Proposition 3.3), we can construct a computationally efficient algorithm for releasing a histogram that is (ε, δ)-differentially private with the accuracies matching
Algorithm 5.1 up to constant factors.
Algorithm 5.3. BoundedStabilityHistogram(D, ε, b) for D ∈ X n , rational ε ∈ (0, 1] and b ∈ [n]
1. Let k = ⌈log(2/ε)⌉.
2. Let A = {x ∈ X : cx (D) > 0}.
3. For each x ∈ A, let c̃x = GeoSample(k, n, cx (D)).
4. Release h = {(x, c̃x ) : x ∈ A and c̃x > b} ∈ Hn,n (X ).
28
Theorem 5.4. Let rational ε ∈ (0, 1], ε̃ = 2 · ln 1 + 2−⌈log(2/ε)⌉ ∈ (4/9 · ε, ε] and b ∈ [n]. Then
BoundedStabilityHistogram(D, ε, b) satisfies the following properties:
i. BoundedStabilityHistogram(D, ε, b) is (ε, δ)-differentially private provided that
b ≥1+
2
ln(1/δ)
ε̃
ii. BoundedStabilityHistogram(D, ε, b) has (a, β)-per-query accuracy on counts larger than t
for
2 1
2 1
ln
ln
and a =
t=b+
ε̃ β
ε̃ β
iii. BoundedStabilityHistogram(D, ε, b) has (a, β)-simultaneous accuracy for
2 n
a=b+
ln
ε̃ β
iv. BoundedStabilityHistogram(D, ε, b) has running time
poly(N )
where N is the bit length of this algorithm’s input (D ∈ X n , ε and b).
Proof of i-iii. By Proposition 3.3, GeoSample(k, n, cx(D)) generates a two-sided geometric random
variable with scale parameter 2/ε̃ centered at cx (D) clamped to [0, n]. Therefore, this algorithm is
identically distributed to StabilityHistogram(D, ε̃, b). Parts i-iii follow from Theorem 5.2.
Proof of iv. Construction of the true histogram takes O(n log |X |) time. And by Proposition 3.3,
the at most n calls to GeoSample take poly(log(1/ε), n) time. Because the counts do not exceed n,
the final step takes poly(n, log |X |) time.
Therefore, we have constructed an efficient algorithm for releasing a sparse histogram with
approximate differential privacy.
6
Lower Bounds
In this section, we prove a lower bound on the per-query accuracy of histogram algorithms whose
outputs are restricted to H∞,n′ (X ) (i.e. sparse histograms) using a packing argument [HT10,
BBKN14]. First, for completeness we state and reprove existing lower bounds for per-query accuracy and simultaneous accuracy as well as generalize them to the case of δ > 0.
Theorem 6.1 (following [HT10, BBKN14]). Let M : X n → H∞,|X |(X ) be (ε, δ)-differentially
private and β ∈ (0, 1/2].
i. If M has (a, β)-per-query accuracy and δ/ε ≤ β, then
1
1
1
ln
− 1, n
a ≥ · min
2
ε
4β
29
ii. If M has (a, β)-simultaneous accuracy, then
ε
1
|X | − 1
1
1
− 1, n
log
− 1, log
a ≥ · min
2
ε
4β
ε
4δ
Proof of i. Assume a < n/2. Let x, x0 ∈ X such that x 6= x0 . Define the dataset D0 ∈ X n such
that all rows are x0 . And define the dataset D such that the first m = ⌊2a⌋ + 1 rows are x and the
remaining n − m rows are x0 . Notice that Pr[|M(D)x − cx (D)| > a] ≤ β by the (a, β)-per-query
accuracy of M. By Lemma 2.3 and the fact that cx (D) > 2a while cx (D ′ ) = 0,
Pr[M(D)x − cx (D)| > a] ≥ e−mε · Pr[M(D ′ )x − cx (D)| > a] − δ/ε
≥ e−mε · Pr[M(D ′ )x − cx (D ′ )| ≤ a] − δ/ε
≥ e−mε · (1 − β) − δ/ε
Therefore,
−(2a+1)·ε
e
1
δ
≤
· β+
≤ 4β
1−β
ǫ
Proof of ii. Assume a < n/2. Let x0 ∈ X . For each x ∈ X define the dataset D (x) ∈ X n such that
the first m = ⌊2a⌋ + 1 rows are x and the remaining n − m rows are x0 . For all x ∈ X , let
Gx = {h ∈ H∞,|X |(X ) : ∀x′ ∈ X |hx′ − cx′ (D (x) )| ≤ a}
By Lemma 2.3, for all x ∈ X
Pr[M(D (x0 ) ) ∈ Gx ] ≥ e−mε · Pr[M(D (x) ) ∈ Gx ] − δ/ε
≥ e−mε · (1 − β) − δ/ε
Notice that Pr[M(D (x0 ) ) ∈
/ Gx0 ] ≤ β and {Gx }x∈X is a collection of disjoint sets. Then
X
Pr[M(D (x0 ) ) ∈
/ Gx0 ] ≥
Pr[M(D (x0 ) ) ∈ Gx ]
x∈X :x6=x0
≥ (|X | − 1) · e−mε · (1 − β) − δ/ε
Therefore,
−(2a+1)·ε
e
1
·
≤
1−β
β
δ
+
|X | − 1 ε
which implies the desired lower bound.
We now state our lower bound over sparse histograms.
Theorem 6.2. Let M : X n → H∞,n′ (X ) be (ε, δ)-differentially private with (a, β)-per-query accuracy with β ∈ (0, 1/2] and δ/ε ≤ β. Then
1
|X |
ε
1
1
ln
ln
−
1,
−
1,
n
a ≥ · min
2
2ε
16βn′
2ε
16βδ
30
The histogram algorithms of Sections 4 and 5 achieve (O(log(1/β)/ε), β)-per-query accuracy on
large enough counts. However, on smaller counts we can only guarantee (a, β)-per-query accuracy
with a = O(log(|X |/β)/ε) and a = O(log(1/(βδ))/ε) (by taking threshold b = O(log(1/δ)/ε))
for algorithms from Sections 4 and 5 respectively. Theorem 6.2 shows these bounds are the best
possible when |X | ≥ poly(n′ ) and δ ≤ poly(1/ε).
Proof. Assume a < n/2. Let x0 ∈ X . For each x ∈ X define the dataset D (x) ∈ X n such that the
first m = ⌈2a⌉ rows are x and the remaining n − m rows are x0 . By definition of (a, β)-per-query
accuracy and the fact that cx (D (x) ) ≥ 2a, we have
h
i
h
i
Pr M(D (x) )x ≥ a ≥ Pr M(D (x) )x − cx (D (x) ) ≤ a ≥ 1 − β
Then, by Lemma 2.3 and that D (x) is at distance at most m from D (x0 ) , we have
h
i
Pr M(D (x0 ) )x ≥ a ≥ (1 − β)e−mε − δ/ε
Thus, by linearity of expectations
hn
oi
E
x ∈ X : M(D (x0 ) )x ≥ a
≥ |X | · (1 − β)e−mε − δ/ε
On the other hand, as M(D (x0 ) ) ∈ H∞,n′ (X ) we have
hn
oi
E
x ∈ X : M(D (x0 ) )x ≥ a
≤ n′
Therefore,
e−⌈2a⌉·ε ≤
1
·
1−β
n′
δ
+
|X | ε
which along with ⌈2a⌉ ≤ 2a + 1 implies the lower bound of
1
1
|X |
1 ε
a ≥ · min
− 1, n
ln
− 1, ln
2
ε
4n′
ε
4δ
Therefore, along with part i of Theorem 6.1, we have
1
1
|X |
1
1
1 ε
a ≥ · min max min
− 1 , ln
ln
− 1, ln
−1 , n
2
ε
4n′
ε
4δ
ε
4β
1
1
1
1
|X |
1
1 ε
≥ · min
− 1 + ln
· min
ln
− 1, ln
−1 , n
2
2
ε
4n′
ε
4δ
ε
4β
1
1
|X |
ε
1
≥ · min
ln
ln
−
1,
−
1,
n
2
2ε
16βn′
2ε
16βδ
7
Better Per-Query Accuracy via Compact, Non-Sparse Representations
In this section, we present a histogram algorithm whose running time is poly-logarithmic in |X |, but,
unlike Algorithm 4.13, is able to achieve (a, β)-per query accuracy with a = O(log(1/β)/ε). Our
histogram algorithm will output a histogram from a properly chosen family. This family necessarily
contains histograms that have many nonzero counts to avoid the lower bound of Theorem 6.2.
31
Lemma 7.1. Let k ∈ N, n ∈ N+ and Z ∼ GeoSample(k, n, 0). There exists a multiset of histograms
Gk,n (X ) satisfying:
i. Let h be drawn uniformly at random from Gk,n (X ). For all x ∈ X and c ∈ [n]
−k+1
e−(2/3)·2
−k+1
· Pr[Z = c] ≤ Pr[hx = c] ≤ e(1/3)·2
· Pr[Z = c]
ii. Let h be drawn uniformly at random from Gk,n (X ). For all x ∈ X and a ∈ [n]
Pr[hx ≤ a] ≥ Pr[Z ≤ a]
iii. Let h be drawn uniformly at random from Gk,n (X ). For all B ⊆ X such that |B| ≤ n + 1 and
c ∈ [n]B
Y
Pr[∀x ∈ B hx = cx ] =
Pr[hx = cx ]
x∈B
iv. For all h ∈ Gk,n , the histogram h can be represented by a string of length poly(k, n, log |X |)
and given this representation for all x ∈ X the count hx can be evaluated in time
poly (k, n, log |X |)
v. For all A ⊆ X such that |A| ≤ n and c ∈ [n]A sampling a histogram h uniformly at random
from {h′ ∈ Gk,n (X ) : ∀x ∈ A h′x = cx } can be done in time
poly (k, n, log |X |)
Parts i and ii state when a histogram is sampled uniformly at random from Gn,k (X ) the marginal
distribution on each count is closely distributed to a two-sided geometric distribution centered at
0 clamped to [0, n].
Proof. (Construction) Let d = (2k+1 + 1)(2k + 1)n−1 . We would like to construct a (n + 1)-wise
independent hash family consisting of functions p : X → {1, . . . , d}, i.e. for all x ∈ X , p(x) is
uniformly distributed over {1, . . . , d} and for all distinct x1 , . . . , xn+1 ∈ X , the random variables
p(x1 ), . . . , p(xn+1 ) are independent.
Given any function p in this family we can construct a histogram by using p(x) as the randomness
for evaluating the noisy count of x via inverse transform sampling in a similar manner to Algorithm
3.4 as the marginal distribution of p(x) is uniformly distributed over {1, . . . , d}.
The set of all degree at most n polynomial over a finite field Fq is a (n + 1)-wise independent
hash family. Ideally, we would have |X | ≤ d and take q = d. In this case, we can map X to a subset
of Fq and use a bijection between {1, . . . d} and Fq to get the desired family of functions.
However, |X | may be larger than d and d is not a prime power. Therefore, we must pick a large
enough field Fq so that |X | ≤ q and that when mapping Fq to {1, . . . , d}, the resulting marginal
distributions are approximately uniform. So let
n
l
mo
m = max ⌈log |X |⌉ , log(3 · 2k−1 · d)
′ (X ) as the family of polynomials over the finite field F m with degree at most n. We
and define Gk,n
2
′ (X ) by taking
construct a histogram h ∈ Gk,n (X ) for each polynomial ph ∈ Gk,n
hx = min z ∈ [n] : F (z) ≥ (T (ph (T −1 (x))) mod d) + 1
32
for all x ∈ X where T : F2m → [2m − 1] is a bijection such that T and T −1 have running time
poly(m) and
if z < 0
0
n−1−z
k(z+1)
k
F (z) = d − 2
2 +1
if 0 ≤ z < n
d
if z = n
as defined in Algorithm 3.4 for c = 0. Notice that if h is drawn uniformly at random from Gk,n (X ),
′ (X ).
then ph is drawn uniformly at random from Gk,n
Proof of i. Let Z ∼ GeoSample(k, n, 0). For all x ∈ X and v ∈ [n], we have
Pr[hx = v] = Pr[min z ∈ [n] : F (z) ≥ (T (ph (T −1 (x))) mod d) + 1 = v]
= Pr[F (v − 1) < (T (ph (T −1 (x))) mod d) + 1 ≤ F (v)]
1
= m · |{y ∈ F2m : F (v − 1) < (T (y) mod d) + 1 ≤ F (v)}|
2
m
2
1
+ 1 (F (v) − F (v − 1))
≤ m·
2
d
d
= 1 + m · Pr[Z = v]
2
(by Lemma 3.5)
m
≤ ed/2 · Pr[Z = v]
Similarly, because d/2m ≤ 2/3
m
2
1
− 1 (F (v) − F (v − 1))
Pr[hx = v] ≥ m ·
2
d
d
≥ 1 − m · Pr[Z = v]
2
m
≥ e−2·d/2 · Pr[Z = v]
Now, part i follows as d/2m ≤ (1/3) · 2−k+1 by choice of m.
Proof of ii. Pick q ∈ N and r ∈ [d − 1] such that 2m = q · d + r. We proceed by splitting the
probability of Pr[hx ≤ a] based on whether or not the integer value of its randomness lies within
one of the q intervals of length d. Let U be drawn uniformly at random from [2m ], D be drawn
uniformly at random from {1, . . . , d} and Z ∼ GeoSample(k, n, 0). For all x ∈ X and a ∈ [n]
Pr[hx ≤ a] = Pr[min z ∈ [n] : F (z) ≥ (T (ph (T −1 (x))) mod d) + 1 ≤ a]
= Pr[(T (ph (T −1 (x))) mod d) + 1 ≤ F (a)]
= Pr[(U mod d) + 1 ≤ F (a)]
q·d
r
= m · Pr[D ≤ F (a)] + m · Pr[D ≤ F (a) | D ≤ r]
2
2
q·d
r
≥ m · Pr[D ≤ F (a)] + m · Pr[D ≤ F (a)]
2
2
= Pr[Z ≤ a]
33
(by monotonicity of F )
′ (X ) is a (n + 1)-wise independent hash family, for all B ⊆ X such that
Proof of iii. Because Gk,n
|B| ≤ n + 1 and c ∈ [n]B
Pr[∀x ∈ B hx = cx ] = Pr ∀x ∈ B F (cx − 1) < (T (ph (T −1 (x))) mod d) + 1 ≤ F (cx )
Y
=
Pr F (cx − 1) < (T (ph (T −1 (x))) mod d) + 1 ≤ F (cx )
x∈B
=
Y
Pr[hx = vx ]
x∈B
Proof of iv. F2m can be represented by an irreducible polynomial of degree m over F2 encoded as
a binary string of length m. Likewise, all elements of F2m can be represented by a polynomial
of degree at most m − 1 over F2 (requires m bits). This encoding defines an efficient bijection T
between F2m and [2m − 1] by also interpreting the string as the binary representation of an element
in [2m − 1].
For all h ∈ Gk,n (X ), h can be represented by k, n, the coefficients of ph and a description of the
field F2m . This representation can be encoded in poly(k, n, log |X |) bits.
Now, given this encoding, evaluation of ph (x) can be done in poly(m) = poly (k, n, log |X |) time.
And computing
hx = min z ∈ [n] : F (z) ≥ (T (ph (T −1 (x))) mod d) + 1
to get an approximate count takes poly (k, n, log |X |) time.
Proof of v. Let A ⊆ X such that |A| ≤ n and c ∈ [n]A . We can sample h given by the coefficients
a0 , . . . , an ∈ F2m uniformly at random from Gk,n (X ) such that hx = cx for all x ∈ A with the
following steps.
1. Construct F2m by finding an irreducible polynomial of degree m over F2 .
2. For each x ∈ X , sample ũx uniformly at random from the set
Sx = {y ∈ F2m : cx = min {z ∈ [n] : F (z) ≥ (T (y) mod d) + 1}}
We can sample from this set by observing that
Sx = {y ∈ F2m : F (cx − 1) < (T (y) mod d) + 1 ≤ F (cx )}
= {y ∈ F2m : ∃i, r ∈ N such that T (y) = i · d + r and F (cx − 1) < r + 1 ≤ F (cx )}
= T −1 ({i · d + r ∈ [2m − 1] : i ∈ N and F (cx − 1) < r + 1 ≤ F (cx )})
m
⌊2 /d⌋−1
[
{i · d + r ∈ [2m − 1] : F (cx − 1) < r + 1 ≤ F (cx )}
= T −1
i=0
3. Let B ⊆ X such that A ⊆ B and |B| = n + 1. For all x ∈ B \ A, sample ũx uniformly at
random from F2m .
4. Take the coefficients a0 , . . . , an ∈ F2m to be the coefficients of the interpolating polynomial
over F2m given the set of points (x, ũx ) for all x ∈ B. This polynomial exists and is unique.
34
We first prove correctness. Notice this procedure can only return a histogram h ∈ Gk,n (X ) such
that hx = vx for all x ∈ A. Let h be any such histogram. Then
Pr[Sampling h] = Pr[(a0 , . . . , an ) are the coefficients of ph ]
= Pr[∀x ∈ B ũx = ph (x)]
Y
=
Pr[ũx = ph (x)]
x∈B
=
Y 1
|Sx |
x∈A
!
1 |B\A|
·
2m
Therefore, these steps output h ∈ Gk,n (X ) uniformly at random such that hx = vx for all x ∈ A.
Construction of F2m can be done in O(m4 ) time [Sho90]. Integer operations are limited to
numbers not exceeding 2m . And polynomial interpolation takes O(n2 ) · poly(m) time [BP70].
Therefore, this procedure runs in time poly(k, n, log |X |).
Algorithm 3.6 (BoundedGeometricMechanism) has on each bin a count with marginal distribution Geosample(k, n, cx (D)) and its counts are independent. By using the previously defined
family, the following algorithm has essentially the same marginal distributions as Algorithm 3.6,
but the counts are only (n + 1)-wise independent. This yields an efficient algorithm as we only
need a small number of random bits (polynomial in log |X |) compared to the amount required for
all counts to be independent (linear in |X |).
Algorithm 7.2. CompactHistogram(D, ε) for D ∈ X n and rational ε ∈ (0, 1]
1. Let k = ⌈log(4/ε)⌉.
2. Let A = {x ∈ X : cx (D) > 0}.
3. For each x ∈ A, let c̃x = GeoSample(k, n, cx (D)).
4. Release h drawn uniformly at random from {h′ ∈ Gk,n (X ) : ∀x ∈ A h′x = c̃x }.
Theorem 7.3. Let rational ε ∈ (0, 1] and ε̃ = 2 · ln 1 + 2−⌈log(4/ε)⌉ ∈ (2ε/9, ε/2]. Then
CompactHistogram(D, ε) has the following properties:
i. CompactHistogram(D, ε) is (ε, 0)-differentially private.
ii. CompactHistogram(D, ε) has (a, β)-per-query accuracy for
2 1
ln
a=
ε̃ β
iii. CompactHistogram(D, ε) has (a, β)-simultaneous accuracy for
2 |X |
ln
a=
ε̃
β
35
iv. CompactHistogram(D, ε) has running time
poly(N )
where N is the bit length of this algorithm’s input (D ∈ X n and ε)
Proof of i. Let D, D ′ ∈ X n be neighboring datasets. Let A = {x ∈ X : cx (D) > 0}. Similarly,
define A′ = {x ∈ X : cx (D ′ ) > 0}. Let B = A ∪ A′ . Notice |B| ≤ n + 1. Let g ∈ Gk,n (X ). Let h be
drawn uniformly at random from Gk,n (X ). Then
Pr[CompactHistogram(D, ε) = g]
= Pr[∀x ∈ B CompactHistogram(D, ε)x = gx ] · Pr[h = g | ∀x ∈ B hx = gx ]
!
Y
Y
Pr[Zx (D) = gx ]
Pr[hx = gx ] · Pr[h = g | ∀x ∈ B hx = gx ]
=
x∈A
x∈B\A
where Zx (D) ∼ GeoSample(k, n, cx (D)). Now, because |B \ A| = 1 and and |B \ A′ | = 1 along with
Proposition 3.3 and part i of Lemma 7.1
Pr[CompactHistogram(D, ε) = g]
Pr[h = g | ∀x ∈ B hx = gx ]
!
Y
Y
Pr[Zx (D) = gx ]
Pr[hx = gx ]
=
x∈A
(1/3)·2−k+1
≤e
·
Y
x∈B\A
Pr[Zx (D) = gx ]
x∈B
−k+1
≤ eε̃+(1/3)·2
·
Y
Pr[Zx (D ′ ) = gx ]
x∈B
≤
ε̃+(1/3)·2−k+1
e
·
Y
x∈A′
≤ eε ·
!
Y
−k+1
Pr[Zx (D ′ ) = gx ] e(2/3)·2
·
Pr[hx = gx ]
x∈B\A′
Pr[CompactHistogram(D ′ , ε)
= g]
Pr[h = g | ∀x ∈ B hx = gx ]
as ε̃ ≤ ε/2 and 2−k+1 ≤ ε/2. Therefore, CompactHistogram(D, ε) is (ε, 0)-differentially private.
Proof of ii. Let h ∼ CompactHistogram(D, ε) and A = {x ∈ X : cx (D) > 0}. Let x ∈ A. Then,
by construction, hx ∼ GeoSample(k, n, cx (D)) and (a, β)-per-query accuracy follows from part ii of
Theorem 3.7.
Let x ∈ X \ A and h′ be drawn uniformly at random from Gk,n (X ). Notice cx (D) = 0 and
|A| ≤ n. By parts ii and iii of Lemma 7.1
Pr[|hx | ≤ a] = Pr[h′x ≤ a | ∀x ∈ A h′x = hx ]
= Pr[h′x ≤ a]
≥ Pr[Z ≤ a]
where Z ∼ GeoSample(k, n, 0). Thus, (a, β)-per-query accuracy follows from part ii of Theorem
3.7.
36
Proof of iii. A union bound over each x ∈ X along with the previous part gives a bound on the
(a, β)-simultaneous accuracy.
Proof of iv. CompactHistogram(D, ε) makes at most n calls to GeoSample. Therefore, by Proposition 3.3 and by part v of Lemma 7.1, we get the desired bound on running time.
Therefore, we have constructed a histogram algorithm with running time polynomial in and
accuracy matching lower bounds for releasing private histograms up to constant factors.
Acknowledgments
We thank the Harvard Privacy Tools differential privacy research group, particularly Mark Bun
and Kobbi Nissim, for informative discussions and feedback, and the anonymous TPDP reviewers
for helpful comments.
References
[BBKN14] Amos Beimel, Hai Brenner, Shiva Prasad Kasiviswanathan, and Kobbi Nissim. Bounds
on the sample complexity for private learning and private data release. Machine learning, 94(3):401–437, 2014.
[BLR13]
Avrim Blum, Katrina Ligett, and Aaron Roth. A learning theory approach to noninteractive database privacy. Journal of the ACM (JACM), 60(2):12, 2013. URL
https://arxiv.org/pdf/1109.2229.pdf.
[BNS16]
Mark Bun, Kobbi Nissim, and Uri Stemmer.
Simultaneous private learning
of multiple concepts.
In Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science, pages 369–380. ACM, 2016.
URL
https://arxiv.org/pdf/1511.08552.pdf.
[BP70]
Åke Björck and Victor Pereyra. Solution of vandermonde systems of equations. Mathematics of Computation, 24(112):893–903, 1970.
[CDK17]
Bryan Cai, Constantinos Daskalakis, and Gautam Kamath. Priv’it: Private and
sample efficient identity testing. arXiv preprint arXiv:1703.10127, 2017. URL
https://arxiv.org/pdf/1703.10127.pdf.
[CKKL12] Mahdi Cheraghchi, Adam Klivans, Pravesh Kothari, and Homin K Lee. Submodular
functions are noise stable. In Proceedings of the twenty-third annual ACM-SIAM symposium on Discrete Algorithms, pages 1586–1592. Society for Industrial and Applied
Mathematics, 2012. URL https://arxiv.org/pdf/1106.0518.pdf.
[CPST11]
Graham Cormode, Magda Procopiuc, Divesh Srivastava, and Thanh TL Tran. Differentially private publication of sparse data. arXiv preprint arXiv:1103.0825, 2011. URL
https://arxiv.org/pdf/1103.0825.pdf.
[CTUW14] Karthekeyan Chandrasekaran, Justin Thaler, Jonathan Ullman, and Andrew Wan.
Faster private release of marginals on small databases. In Proceedings of the 5th conference on Innovations in theoretical computer science, pages 387–402. ACM, 2014.
URL https://arxiv.org/pdf/1304.3754.pdf.
37
[DL09]
Cynthia Dwork and Jing Lei. Differential privacy and robust statistics. In Proceedings
of the forty-first annual ACM symposium on Theory of computing, pages 371–380.
ACM, 2009.
[DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise
to sensitivity in private data analysis. In Theory of Cryptography Conference, pages
265–284. Springer, 2006.
[DNT15]
Cynthia Dwork, Aleksandar Nikolov, and Kunal Talwar. Efficient algorithms for privately releasing marginals via convex relaxations. Discrete & Computational Geometry,
53(3):650–673, 2015. URL https://arxiv.org/pdf/1308.1385.pdf.
[GMP16]
Ivan Gazeau, Dale Miller, and Catuscia Palamidessi. Preserving differential privacy
under finite-precision semantics. Theoretical Computer Science, 655:92–108, 2016. URL
https://arxiv.org/pdf/1306.2691.pdf.
[GRS12]
Arpita Ghosh, Tim Roughgarden, and Mukund Sundararajan. Universally utilitymaximizing privacy mechanisms. SIAM Journal on Computing, 41(6):1673–1693, 2012.
[GRU12]
Anupam Gupta, Aaron Roth, and Jonathan Ullman.
Iterative constructions
and private data release. Theory of Cryptography, pages 339–356, 2012. URL
https://arxiv.org/pdf/1107.3731.pdf.
[HRS12]
Moritz Hardt, Guy N Rothblum, and Rocco A Servedio. Private data release via learning thresholds. In Proceedings of the twenty-third annual ACM-SIAM symposium on
Discrete Algorithms, pages 168–187. Society for Industrial and Applied Mathematics,
2012. URL https://arxiv.org/pdf/1107.2444.pdf.
[HT10]
Moritz Hardt and Kunal Talwar. On the geometry of differential privacy. In Proceedings
of the forty-second ACM symposium on Theory of computing, pages 705–714. ACM,
2010. URL https://arxiv.org/pdf/0907.3754.pdf.
[KLN+ 11] Shiva Prasad Kasiviswanathan, Homin K Lee, Kobbi Nissim, Sofya Raskhodnikova, and
Adam Smith. What can we learn privately? SIAM Journal on Computing, 40(3):793–
826, 2011. URL https://arxiv.org/pdf/0803.0924.pdf.
[Knu05]
Donald E Knuth. Combinatorial Algorithms, Part 1, volume 4 of The Art of Computer
Programming. Addison-Wesley Professional, 2005.
[Mir12]
Ilya Mironov. On significance of the least significant bits for differential privacy. In
Proceedings of the 2012 ACM conference on Computer and communications security,
pages 650–661. ACM, 2012.
[Sho90]
Victor Shoup. New algorithms for finding irreducible polynomials over finite fields.
Mathematics of Computation, 54(189):435–447, 1990.
[TUV12]
Justin Thaler, Jonathan Ullman, and Salil Vadhan. Faster algorithms for privately releasing marginals. In International Colloquium on Automata, Languages, and Programming, pages 810–821. Springer, 2012. URL https://arxiv.org/pdf/1205.1758.pdf.
[UV11]
Jonathan Ullman and Salil P Vadhan. PCPs and the hardness of generating private synthetic data. In TCC, volume 6597, pages 400–416. Springer, 2011. URL
https://pdfs.semanticscholar.org/bf6f/e68b91283787b9ad28494a78b4cead10fa12.pdf.
38
A1
Generating a Sparse Histogram Uniformly at Random
Being able to efficiently compute |Hn,n′ (X )| and sample a uniformly random element from Hn,n′ (X )
is needed for an efficient implementation of Algorithm 4.13.
Lemma A1.1. |Hn,n′ (X )| can be calculated and a uniformly random element of Hn,n′ (X ) can be
sampled in time
poly(n′ , log n, log |X |)
Proof. We show HistSample(X , n, n′ ) defined below efficiently samples a uniformly random element
of Hn,n′ (X ) by using a bijection between Hn,n′ (X ) and {1, . . . , |Hn,n′ (X )|}.
Algorithm A1.2. HistSample(X , n, n′ ) for n, n′ ∈ N+ with n′ ≤ |X |
P ′
1. Pick U uniformly at random from {1, . . . , |Hn,n′ (X )|} = {1, . . . , ni=0
P
2. Find the smallest ℓ ∈ [n′ ] such that ℓi=0 |Xi | ni ≥ U .
i
Pℓ−1 |X | i h |X | ℓ
3. Let u′ = U − 1 − i=0
n
∈
·
n
−
1
.
i
ℓ
4. Use integer division to find q ∈
h
|X |
ℓ −
|X | i
i n }.
i
1 and r ∈ nℓ − 1 such that u′ = qnℓ + r.
5. Map q to a corresponding subset of X of size ℓ specified by sorted elements q1 < . . . < qℓ .
Specifically, we can let q1 , . . . , qℓ ∈ X be the sequence representing q in the combinatorial
number system of degree ℓ, i.e. the unique sequence satisfying
qℓ
q1
q=
+ ... +
ℓ
1
This sequence can be found P
greedily; for each j decreasing from ℓ to 1, using binary search,
find the largest qj such that ℓi=j qii ≤ q [Knu05].
6. Let r0 , . . . , rℓ−1 ∈ [n − 1] be the digits of the base n representation of r, i.e.
r = rℓ−1 nℓ−1 + . . . + r1 n + r0
7. For each i ∈ {1, . . . , ℓ}, release (qi , ri−1 + 1).
Steps 3 and 4 define a bijection mapping u to (ℓ, u′ ). Step 5 defines a bijection mapping (ℓ, u′ )
to (ℓ, q, r). In steps 6 and 7, the decompositions of q and r into their respective number systems are
bijections. Likewise, there is a bijection mapping (q1 , . . . , qℓ , r0 , . . . , rℓ−1 ) to an element of Hn,n′ (X ).
Therefore, our mapping from {1, . . . , |Hn,n′ (X )|} to Hn,n′ (X ) is bijective proving correctness.
P ′
Computing |Hn,n′ (X )| = ni=0 |Xi | ni takes poly(n′ , log n, log |X |) time. The remaining steps
can be done in poly(n′ , log n, log |X |) time as no number used exceeds |Hn,n′ (X )|, all numbers used
are expressible as the sum of at most n′ other numbers and steps 5-7 each consist of at most n′
iterations.
39
| 8cs.DS
|
arXiv:1803.06315v1 [cs.SY] 16 Mar 2018
Benchmarks for cyber-physical systems:
A modular model library
for building automation systems
(Extended version)
Nathalie Cauchi, Alessandro Abate
Department of Computer Science, University of Oxford, Oxford, U.K
[email protected]
Abstract
Building Automation Systems (BAS) are exemplars of Cyber-Physical
Systems (CPS), incorporating digital control architectures over underlying
continuous physical processes. We provide a modular model library for
BAS drawn from expertise developed on a real BAS setup. The library
allows to build models comprising of either physical quantities or digital
control modules.The structure, operation, and dynamics of the model
can be complex, incorporating (i) stochasticity, (ii) non-linearities, (iii)
numerous continuous variables or discrete states, (iv) various input and
output signals, and (v) a large number of possible discrete configurations.
The modular composition of BAS components can generate useful CPS
benchmarks. We display this use by means of three realistic case studies,
where corresponding models are built and engaged with different analysis
goals. The benchmarks, the model library and data collected from the
BAS setup at the University of Oxford, are kept on-line at https://
github.com/natchi92/BASBenchmarks.
Keywords 1 cyber-physical systems, building automation systems, thermal modelling, hybrid models, simulation, reachability analysis, probabilistic safety, control synthesis
1
Introduction
This paper describes a library of models for Building Automation Systems (BAS), which can
be employed to create benchmarks for verification, control synthesis, or simulation purposes of
Cyber-Physical Systems (CPS). The models are inspired by and built around an experimental
setup within the Department of Computer Science at the University of Oxford, which is part of
on-going research in collaboration with service engineers and industrial partners in the sector.
This library allows to create numerous meaningful models for BAS, which are examples of
CPS integrating continuous dynamics and discrete modes.
Interest in BAS, also colloquially known as smart buildings, is gaining rapid momentum, in particular as a means for ensuring thermal comfort ([1]), minimising energy consumption ([16, 18]), and ascertaining reliability ([3, 19]). Quantitative models are needed to
1
evaluate system performance, to verify correct behaviour, and to develop specific control algorithms. An overview of the different BAS modelling techniques used in literature is presented
in [15]. Several simulation tools (see [5]) have been devised to aide in the development and
analysis of models for BAS. Attempting a multi-dimensional characterisation of the broad
spectrum of existing BAS models, we can find either deterministic or stochastic ones, low- to
high-dimensional ones, with discrete or continuous inputs and states. The choice of a model is
an art and a craft ([14]): one must select simplifying assumptions that accurately reflect the
operational performance of the BAS in specific real-world environments, and introduce uncertainty to represent un-modelled components, unknown parameters or random occupants. We
therefore aim to simplify the modelling process such that simulation, verification or strategy
synthesis can be carried out seamlessly. Different verification and policy synthesis tools exist
in literature ([6, 10, 11]). They are typically specific to a particular type of model structure
and in the case of stochastic or hybrid models are oftenlimited to a small number of continuous
variables. The use of such tools also requires expert knowledge on the specific formalism the
tool makes use of.
In order to display the versatility of the library of BAS models, we present three case
studies that are built from its components. We focus on modelling temperature dynamics, a
key element for ensuring thermal comfort. We employ the three generated models for different
analysis goals, comprising simulation, reachability, and control synthesis. The models and
delineation of the case-studies are kept on-line at
https://github.com/natchi92/BASBenchmarks.
This is to allow their use or modification for different applications and for comparison with
other modelling approaches in BAS. The repository also contains real data gathered from the
BAS lab at Oxford, which can be employed for further modelling studies. This article has the
following structure: Section 2 introduces the BAS modelling framework for CPS. We identify
three modelling trade-offs that introduce different complexities on the model dynamics. Based
on these trade-offs, we develop and analyse three case studies in Section 3.
2
2.1
Building Automation Systems
BAS: structure and components
BAS models clearly depend on the size and topology of the building ([12]), and on its climate
control setup. In this work, we consider the BAS setup in the Department of Computer
Science, at the University of Oxford. A graphical depiction is shown in Figure 1a. The BAS
consists of two teaching rooms that are connected to a boiler-heated system. The boiler
supplies heat to the heating coil within the AHU and to two radiators. Valves control the rate
of water flow in the heating coil and in radiators. The AHU supplies air to the two zones,
which are connected back to back, and adjacent both to the outside and to an interior hall
(cf. Figure 1a). The zone air of both rooms can mix with the outside air and exchanges
circulating air with the AHU. Return water from the AHU heating coils and radiators is
collected and pumped back to the boiler. Figure 1b presents the Resistor Capacitance (RC)
network circuit of the two zones ([8]), which underpins the dynamics for temperature in the
zone component - corresponding equations are in Table 3. The heat level in each room is
modified by (i) radiative solar energy absorbed through the walls, (ii) occupants, (iii) AHU
input supply air, (iv) radiators, and (v) AHU return water. The effect of heat stored in the
walls and in rooms is depicted with capacitors, whereas thermal resistance to heat transfer by
the walls is depicted by resistor elements.
2
(a) 2-zone boiler-based heating system with air handling unit and radiators
(b) Resistance-capacitance circuit for the internal thermal dynamics within
the two zones
Figure 1: Building automation system setup
3
2.2
BAS: dynamics and configurations
We define models for the individual components in the BAS system based on the expertise
developed on the BAS setup at Oxford. Single components are intended as separate physical structures within the BAS. Their models are built from the underlying physics and are
improved via industrial feedback and from existing literature ([8]). We obtain models with a
number of unknown parameters: these are estimated and validated using data collected from
the BAS setup [13]. We list indices in Table 1, while all the quantities (variables, parameters,
inputs) are listed in Table 2. Table 3 presents all the relations among variables in the model:
algebraic relations define static couplings, whereas differential relations define the dynamics
for the corresponding variables. The structure in Figure 1a, the quantities in Table 2, and the
variables (with associated dynamics) in Table 3, together allow to construct global models for
the complete BAS setup. We refer to the set of models describing the individual components
(cf. Table 3) as a library of models: one can select the individual components and models
from the library, and build different BAS configurations.
Table 1: Indices
Index
a
adj, out
d
i = {1, 2}
jw = {5, 6}
l ∈ {1, 2}
out
ref
sa
sw
w
h
Reference
AHU
adjacent exterior zones
mixer
individual zones
zone walls with windows
adjacent interior zone
outside
reference
supply air
supply water
wall
water
Index
adj
b
hall
jn = {2, 3, 7}
j = jn ∪ jw
occ
r
rw
solar
v
z
ar
Reference
adjacent zone
boiler
hallway
zone walls with no windows
all zone walls
occupants
radiator
return water
solar energy
collector
zone
air
A global model of the BAS set-up can be complex, comprising both algebraic and differential relations that may be further affected by process noise. A model also contains a number
of inputs which can either be construed as control signals or as exogenous signals. Some of
the dynamics are non-linear in view of continuous variables that are bi-linearly coupled (cf.
AHU air duct model in Table 3). The number of continuous variables also increase substantially when considering a BAS setup with multiple zones: employing the zone component for
a configuration with n zones would result in a model with (2n + 1) + n continuous state variables. Furthermore, the model features multiple components that present switching discrete
behaviours, affecting the dynamics of the continuous variables: these discrete modes are listed
in Table 4 and can result in up to 144 discrete configurations.
In order to tackle the complexity of global BAS models and to add a level of flexibility
to the modelling framework, we consider each BAS component as a separate module, characterised by inputs and output elements, and internal variables. We make use of individual
modules describing component type, and then connect different modules based on possible
physical couplings. Coupling between different modules is also achieved via input-output
relationships: e.g., in the zone module we have coupling between two zones through the continuous variable Tadj,l corresponding to the adjacent zones, which for the wall separating the
two zones (cf. W7 in Figure 1b) corresponds to the individual zone temperatures of the two
zone modules (cf. Table 3 zone equations). Having such a modular structure for the individual
components provides an added level of versatility, since we can connect different components
to create various new models. Modularisation also allows (i) to perform analysis of the whole
setup by executing analysis of individual modules and (ii) to extend the library of models by
defining new modules that connect to existing modules via their input-output relations.
4
Symbol
Ai
Ben
C
Cpa , Cpw
CO2i
kb
m
n
Pout
Q
R
T
u
(U A)
V
w
wmax
X
{α, β, µ}
σ
ρ
τ
Quantity
area of windows of each zone
boiler switched off
capacitance of medium
specific heat capacity of air and water
carbon-dioxide measurements in each zone
steady-state of the boiler
mass air flow rate
number of zones
radiator rated power output
associated heat gain
thermal resistance to heat from walls to medium
temperature of associated medium
mixing ratio
overall transmittance factor of medium
volume of medium
water flow rate of associated medium
maximum water-flow permitted by the valve
valve position of associated component
de-rating and offset factors
the associated process noise
density of medium
time constant of the medium
Type
constant
discrete
constant
constant
input
constant
input
constant
constant\input
input
constant
state\input
input
constant
constant
input
constant
input
constants
constant
constant
constant
Table 2: List of variables, inputs, and parameters
2.3
BAS: description of model library
The library of BAS components comes in the form of MATLAB scripts. Each script represents
an individual BAS component. The models are in state-space form and are of two types linear
or bilinear depending on the component they represent. They are defined using the symbolic
toolbox, are parametrised and can be described both in discrete and continuous time. We
provide the parameters which we estimated from data gathered from the BAS lab to construct
the individual models. However, users can easily make use of their own parameters and
construct their own model. Different components can be connected together based on their
input-output relations by cascading the different symbolic models for each component. Once
this is done, the provided scripts allow you to simulate the models and plots for the defined
output variables are presented.
3
Case studies
Next we set up three case studies and present the trade off between the discussed elements
of complexity. For each of the case studies, we (i) establish the dynamics of the models, (ii)
how they are constructed from the library of components and (iii) and describe the results
obtained
3.1
Two-zone heating setup with deterministic or stochastic dynamics
We consider two zones, each heated by one radiator and with a common supply air, as portrayed in Figure 2. From Table 3, we select two components and corresponding models: the
5
Component: Boiler
Continuous(variables
0
dTsw,b (t) =
−1
(τsw ) [(−Tsw,b (t) + kb )dt] + σsw dW
Relation
Ben (t) = 0
Ben (t) = 1
Component: Valve
Continuous variables
−1
w(t) = (τ ) [exp(ln(τ )X(t))wmax ]
differential
Relation
algebraic
Component: Mixer
Continuous variables
X
Td (t) = ud Tout (t) + (1 − ud )(
Tzi (t))(n)−1
Relation
algebraic
i
Component: AHU heating coil
Continuous variables
−1
dTrw,a (t) = (Cpw ρh Va ) [(Cpw wa (t)(Tsw,b (t) − Trw,a (t)) + (U A)a (Td (t) − Trw,a (t)))dt] + σrw,a dW
Relation
differential
Component: AHU air duct
Continuous variables
−1
dTsai (t) = (Ca ρa Va ) [ma (t)Cpa (Td (t) − Tsai (t)) + (U A)a (Tzi (t) − Tsai (t))]dt + σsai dW
Relation
differential
Component: Radiator
Continuous variables
−1
dTrw,ri (t) = (Cpw ρh Vri ) [(Cpw wri (t)(Tsw,b (t) − Trw,ri (t)) + (U A)ri (Tzi (t) − Trw,ri (t)))dt] + σrw,ri dW
Relation
differential
Component: Zone
Continuous variables
−1 Twjn (t) − Tzi (t)
dTzi (t) = (Czi )
+ Qrw,ri (t) + Qocci (t) + Qsai (t) dt + σzi dW
Rij
"
#
−1 Tadj,out (t) − Tzi (t) X Tadjl (t) − Twjn (t)
dTwjn (t) = Cwjn
+
+ Qrw,ajn (t) dt + σwjn dW
Rout
Rlj
l
"
#
−1 Tadj,out (t) − Tzi (t) X Tadjl (t) − Twjw (t)
dTwjw (t) = Cwjw
+
+ Qsolarjw (t) + Qrw,ajw (t) dt + σwjw dW
Rout
Rljw
Relation
differential
l
Qrw,ri (t) = Pradi (α2 (Trw,ri (t) − Tzi (t)) + α1 ), Qocci (t) = µi (CO2i (t)) + β1i , Qsai (t) = ma (t)Cpa (Tsai (t) − Tzi (t)))
Qrw,aj (t) = α3 (Trw,a (t) − Twj (t)), Qsolarjw (t) = (α0 Ai Tout (t) + β2 )
Component: Collector
Continuous variables
X
Trw,b (t) = uv Trw,a (t) + (1 − uv )(
Trw,ri (t))(n)−1
Relation
algebraic
i
Table 3: Components dynamics and functional relations
among variables
Component
Boiler
AHU air duct
Mixer
AHU heating coil
Radiator heating coil 1
Radiator heating coil 1
Discrete Modes
Boiler on, off (Ben )
Fan off, medium, high (ma )
Open, closed (ud )
Valve healthy, faulty (Xa )
Valve healthy, faulty (Xr1 )
Valve fully open, half open, closed (Xr1 )
Table 4: Discrete operational modes
radiator and the zone. We simplify these models with the following assumptions: (i) the
wall temperature is constant across the zones and is a fixed value (Tw,ss ); (ii) the boiler is
switched ON providing a supply temperature Tsw,bss ; (iii) we fix both the mass air flow rate
ma and the radiator water flow rate wr ; and (iv) we do not include the heat gain from the
windows and the AHU heating coils (Twss ) in each zone. We obtain a model with the four
T
state variables xT = Tz1 Tz2 Trw,r1 Trw,r2
and with a common supply temperature
u = Tsa as an input. For this setup we further consider three different dynamics: (i) a purely
6
Figure 2: BAS setup for the first case study
deterministic one; (ii) a deterministic model with additive disturbance; and (iii) a stochastic
model. For (i) and (ii), we thus remove the process noise in the template model, while for
models (i) and (iii) we do not include the occupancy heat gain. We also discretise the dynamics by a Forward-Euler scheme (for the deterministic models) and a Euler-Maruyama scheme
(for the stochastic model), using a uniform sampling time ∆ = 15 minutes, and obtain a set of
linear discrete-time models. One should note that the models being considered are not fully
observable since for all the individual zone temperatures (variables of interest) are the only
variables provided as outputs. The dynamics of the deterministic model (i) are described by
Ax[k] + Bu[k]#+ Qd
x[k + 1] = "
Md :
(1)
1 0 0 0
=
x[k],
yd [k]
0 1 0 0
where again the matrices are properly sized and constructed based on the models in Table 3
and where,
T
Twss ∆ Twss ∆
Cpw wr1 ∆
Cpw wr2 ∆
Tsw,bss
Tsw,bss
Qd =
Cz1 R1 Cz2 R2
Cpw ρh Vr1 ,b
Cpw ρh Vr2 ,b
Here Ri is the mean resistance offered by the walls. The deterministic model with additive
disturbances is
Ax[k] + Bu[k] + Fda dda [k]
x[k + 1] =
" + Qda
#
Mda :
(2)
1 0 0 0
=
x[k].
yda [k]
0 1 0 0
T
CO22
We have extended (1) with additive noise vector dT
(cf. Qocci in
da = CO21
Table 3) representing the different CO2 levels in each zone. Qda is defined as Qda =
iT
hβ ∆
β12 ∆
1
0 0
and Fda is a properly sized matrix.The stochastic model is
Qd + C1z
Cz
1
2
expressed by extending (1) to include process noise, as
Ax[k] + Bu[k]#+ Qd + ΣW [k]
x[k + 1] = "
Ms :
(3)
1 0 0 0
=
x[k],
ys [k]
0 1 0 0
√
√
√
√
where Σ = diag([( ∆σz1 )2 ( ∆σz2 )2 ( ∆σrw,r1 )2 ( ∆σrw,r2 )2 ])encompasses the vari
T
ances of the process noise for each state. W = w1 w2 w3 w4
are independent Gaussian random variables, which are also independent of the initial condition of the process. A
simulation run of all three models is depicted in Figure 3.
7
21.5
Tz1 (o C)
Md
Mda
Ms
21
20.5
20
0
4
8
12
16
20
0
4
8
12
16
20
Time (hours)
Figure 3: First case study: simulations over two days
3.1.1
Reachability analysis
For this case study we would like to perform the following verification task: to decide whether
traces generated by the models remain within a specified safe set for a given time period.
This is achieved by reachability analysis, which takes a probabilistic flavour for the stochastic
model. The safe set is described as an interval around the temperature set-point TSP = 20o C
±0.5o C. We constrain the input u to lie within the set {Tsa ∈ R|15 ≤ Tsa ≤ 22} for all models
and employ a fixed time horizon N = 6×∆ = 1.5 hours. We perform reachability analysis with
Axelerator ([2]), while we use FAUST2 ([17]) to perform probabilistic reachability analysis of
Ms .
In order to perform reachability analysis using Axelerator, for each of the models we set the
T
T
initial condition as Tz1 Tz2 Trw,r1 Trw,r2
= 18 18 35 35 . The reach tube
for model Md over the whole time horizon is shown in Figure 4: it encompasses the union of
reachable states over that horizon. The results obtained using Axelerator (cf. Figure 4) are
conservative results and confirm that the model can indeed stay in the required safe set for
some initial states, but can also exit it.One can note the coupling between the two zones and
that zone 1 tends to stay at higher zone temperatures then zone 2. A similar reach tube is
obtained for model Mda .
Similarly, we perform probabilistic reachability analysis on model Ms by defining the
same safe set and assuming an input set of [15 22]. The resulting adaptive partition of the
safe set along with the optimal safety probability for each partition set is depicted in Figure 5.
When performing probabilistic reachability analysis using model Ms (cf. Figure 5), we deduce
that the models have a high probability of being within the required safe set, specifically to
have Tz1 ∈ [19.5 20] and Tz2 ∈ [19.5 20.5].
3.2
Two-zone heating setup with large number of
continuous variables
In this second case study we focus on the dynamics of the zone component from Table 3 and
consider the two zones shown in Figure 1b. We assume that (i) a central fan pumps in air
in both rooms with a common supply temperature 15o C ≤ Tsa ≤ 30o C, (ii) the input mass
airflow ma is fixed to 10 m3 /hour and (iii) the return water temperature of the AHU heating
coils is fixed (Trw,ass = 35o C). As in the previous case study, the selected model is discretised
using Forward-Euler, with a sampling time ∆ = 15 minutes, to obtain the discrete-time model
(
xc [k + 1] = A
h c xc [k] + Bc uc [k] + Fc dci[k] + Qc
Mc :
(4)
yc [k]
= 1 0 0 0 0 0 0 xc [k].
Here the variables are
xc = Tz1
Tz2
Tw5
Tw6
8
Tw2
Tw3
Tw7
T
,
Tz2 (o C)
20
15
10
5
10
12
14
16
18
20
22
o
Tz1 ( C)
Figure 4: First case study: reach tube of Md over whole
time horizon
Prob.
Tz2 (o C)
20.5
0.8
20
0.6
19.5
19.5
20
Tz1 (o C)
20.5
Figure 5: First case study: partition of the safe set for model
Ms , along with optimal safety probability for each partition
set
and a common fan supplies the two zones with a supply rate uc = Tsa , whereas
T
dc = Tout Thall CO21 CO22 Trw,r1 Trw,r2 ,
and
Qc =
h
qc0 ∆ qc0 ∆ qc2 ∆ qc2 ∆ qc1 ∆ qc1 ∆ qc1 ∆
Cz1
Cz2
Cw5
Cw6
Cw2
Cw3
Cw7
iT
,
where qc0 = β1i + α1 Pradi , qc1 = α3 Trw,ass and qc2 = α0 A2 β2 + qc1 . Matrices Ac , Bc , Fc
are properly sized. Recall that the vector dc corresponds to the disturbance signals, while Qc
represents constant additive terms within the model. We finally model the disturbances as
random external effects.
3.2.1
Policy synthesis and refinement
For Mc we would like to synthesise a policy ensuring that the temperature within zone 1
does not deviate from the set point by more then 0.5o C over a time horizon equal to four
hours (i.e N = 16). This requirement can be translated into the following PCTL property
9
Φ := P=p [≤N =16 |Tz1 − TSP | ≤ 0.5]. We then aim at synthesising a policy maximising the
safety probability p. This synthesis goal can be computationally hard due to the number
of continuous variables making up Mc . To mitigate this limitation, we perform policy synthesis via abstractions ([9]). We simplify (4) into four abstract models using the technique
in ([8]). The abstract models are labelled as Mca={4,...,1} , where a represents the number
of continuous variables of the corresponding abstract model. The models take the form of
Markov decision processes ([8]). We can quantify the error in the output variable, which has
been introduced by the different levels of abstractions, through the use of (ε, δ)-approximate
simulation relations ([9]). The pair (ε, δ) represents the deviation in the output trajectories
between complex and abstract models and the differences in probability distribution of the
processes, respectively. Such metrics allows the designer to select which of the considered abstract models provides the best trade off in precision: it is desirable to achieve little deviation
in both the output trajectories (small ε) and in the probability distributions (small δ).
δ
Mc4
Mc3
Mc2
Mc1
1
0.0008
0.0006
0.0011
0.0010
−1
10 2
0.1754
0.1933
0.1950
0.1953
10−1
0.2084
0.2312
0.2373
0.2371
−3
10 2
0.2339
0.2598
0.2681
0.2595
10−2
0.2555
0.2831
0.2928
0.2854
−5
10 2
0.2745
0.3065
0.3155
0.3103
10−3
0.2910
0.3241
0.3278
0.3254
Table 5: Second case study: error metrics (ε, δ) for
concrete and abstract models.
We compute (ε, δ)-approximate simulation relations between Mc and the set of abstract
models Mca={4,...,1} , as presented in Table 5. The (ε, δ) pair providing the optimal trade off
is obtained with the abstract model Mc1 and corresponds to (0.2854, 10−2 ). Next, we use
FAUST2 to perform a grid-based computation of the safety probability for Mc1 and obtain a
model of size 14893 with an overall accuracy of 0.005. Over this approximation we synthesise
the optimal policy for the abstract model which results in a safety probability of p0 = 0.9257.
We refine the obtained policy ([8]), which results in one that can be used with Mc . The overall
process results in Φ being satisfied with a safety probability of p = p0 − η − N δ = 0.7657,
where η is the abstraction error introduced by FAUST2 . The results obtained further highlight
that by trading off the complexity in the number of continuous variables and computing
(ε, δ)-simulation relations, we can synthesise policies using simpler models, yet achieve high
performance still when the refined policy is applied to the original model.
3.3
Single zone heating with multiple switching controls
In this third and last case study, we focus on mixer, AHU air duct, and zone components
from Table 3. We select the AHU as the only source of heat within the zone (the boiler is
disconnected). A pictorial description of this setup is in Figure 6. The mixer operates in either
of two modes: open (Op) or closed (Cl). The AHU air duct recirculates air from either the
internal zone (when ud = 0 and the mixer is in mode Op) or from the outside (when ud = 1 and
the mixer is in mode Cl) via the continuous variable Td (output of the mixer component). The
rate of air being pumped into the zone (ma ) is controlled by the fan, which has three operating
speeds (off O, medium M , and high H). The mixer position and the fan settings can be used to
maintain a comfortable temperature within the zone. This setup can be described by a hybrid
model. The discrete modes q are in the set {(O, −), (M, Op), (M, Cl), (H, Op), (H, Cl)}, and
describe the possible configurations of fan operating speeds {O, M, H} and mixer position
{Op, Cl}. (When the fan is switched off, the mixer can be in any position as no air is pumped
into the zone.) Continuous variables model the zone temperature (Tz1 ) together with the
supply air temperature (Tsa ) being pumped into the zone. Transitions between discrete modes
are triggered by continuous dynamics crossing spatial guards: guards denote deviations from
temperature set-point as δ, δk , k = {2, . . . , 5}, δ < δ2 < δ3 < δ4 < δ5 . A graphical description
10
Figure 6: BAS setup for third case study
of the overall hybrid model, together with the different guard conditions, is shown in Figure 7.
The continuous dynamics are built from Table 3 and follow
Twss − Tz1 (t)
+ ma (t)Cpa (Tsa (t) − Tz1 (t) + µ1 CO21ss + β11 ],
R
−1
= (Ca ρa Va ) [ma (t)Cpa (Td (t) − Tsa (t)) + (U A)a (Tz1 (t) − Tsa (t)].
Ṫz1 = (Cz1 )−1 [
Ṫsa
Figure 7: Hybrid model for the third case study, showing discrete states and guard conditions; the initial discrete
state is (O, −)
11
The variables ma and Td take values according to the discrete mode as
q(t) = (O, −),
0
ma (t) = ma,med q(t) = (M, Op) ∨ q(t) = (M, Cl),
ma,high q(t) = (H, Op) ∨ q(t) = (H, Cl),
and
(
Td (t) =
Tout
Tz1 (t)
q(t) = (M, Op) ∨ q(t) = (H, Op),
else.
Here, ma,med , ma,high correspond to the air flow rates when the fan is operating at medium
and high speeds.
3.3.1
Reachability analysis
We are interested in performing reachability analysis of the hybrid model, which we run using
SpaceEx ([7]). Notice that in this case we do not discretise time and consider a continuous
time horizon of 2 hours. We consider two different initial conditions: in the first experiment
we select an initial condition equal to Tz1 = 15o C and Tsa = 15o C, while in the second we
set Tz1 = 20o C and Tsa = 20o C. The resulting reach tube for the both experiments is shown
in Figure 8a.
22
1
Tz (oC)
21.5
21
20.5
20
19.5
20
20.5
21
21.5
Tsa (oC)
(a) Initial condition:
Tsa = 15o C
Tz1 = 15o C, (b) Initial condition:
Tsa = 20o C
Tz1 = 20o C,
Figure 8: Third case study: reach tubes obtained from two
different initial conditions
In Figure 8a we can see that the model initially is in state (O, −) and jumps to a new
state (H, Op) such that warm outside air is pumped into the zone (due to the low temperature
of the initial conditions). From (H, Op) it switches to (M, Op) (notice the reduction in the
gradient between the variables Tsa ∈ [15 18], Tz1 ∈ [17 19]) and eventually switches back to
(O, −) in order to maintain the temperature within the comfort region. For Figure 8b, the
reach tube shows that the system remains within the initial state (O, −) over the whole time
horizon.
4
Conclusions
This paper has presented a library of CPS models for BAS. The BAS modelling framework
comprises three different types of complexities (stochasticity, number of continuous variables,
12
number of discrete modes) and is modular such that various BAS components can be composed. We illustrate the use of this BAS components library via three case studies, each
of which highlights a different side of the complexity trade off and solves a different problem
(simulation, (probabilistic) reachability analysis, and strategy synthesis respectively). Current
work is developed towards obtaining a compositional tool that is able to allow for easy construction of BAS models and for interfacing with different (i) verification tools for performing
analysis and (ii) synthesis tools for computation of optimal strategies.
Acknowledgements
This work has been funded by the European Commission in the Seventh Framework Programme project AMBI under Grant No.: 324432 and is in part supported by the Alan Turing
Institute, UK and Malta’s ENDEAVOUR Scholarships Scheme. The authors would also like to
thank Dario Cattaruzza, Sofie Haesaert and Honeywell Laboratories, Prague for their fruitful
feedback.
References
[1] Hamza Belkhouane, Jan Hensen, and Shady Attia. Thermal comfort models for net zero
energy buildings in hot climates. In Second International Conference on Energy and
Indoor Environment for Hot Climates. Doha, 2017.
[2] Dario Cattaruzza, Alessandro Abate, Peter Schrammel, and Daniel Kroening.
Unbounded-time analysis of guarded LTI systems with inputs by abstract acceleration.
In International On Static Analysis, pages 312–331. Springer, 2015.
[3] Nathalie Cauchi, Khaza Anuarul Hoque, Alessandro Abate, and Mariëlle Stoelinga. Efficient probabilistic model checking of smart building maintenance using fault maintenance
trees. arXiv preprint arXiv:1801.04263, 2018.
[4] Scott Cotton, Goran Frehse, and Olivier Lebeltel. The spaceex modeling language, 2010.
[5] Drury B Crawley, Jon W Hand, Michaël Kummert, and Brent T Griffith. Contrasting the capabilities of building energy performance simulation programs. Building and
environment, 43(4):661–673, 2008.
[6] Iulia Dragomir, Viorel Preoteasa, and Stavros Tripakis. The refinement calculus of reactive systems toolset. arXiv preprint arXiv:1710.08195, 2017.
[7] Goran Frehse, Colas Le Guernic, Alexandre Donzé, Scott Cotton, Rajarshi Ray, Olivier
Lebeltel, Rodolfo Ripado, Antoine Girard, Thao Dang, and Oded Maler. SpaceEx: Scalable verification of hybrid systems. In Computer Aided Verification, pages 379–395.
Springer, 2011.
[8] Sofie Haesaert, Nathalie Cauchi, and Alessandro Abate. Certified policy synthesis for
general markov decision processes: An application in building automation systems. Performance Evaluation, 117:75–103, 2017.
[9] Sofie Haesaert, Sadegh Esmaeil Zadeh Soudjani, and Alessandro Abate. Verification of
general Markov decision processes by approximate similarity relations and policy refinement. SIAM Journal on Control and Optimization, 55(4):2333–2367, 2017.
[10] Ernst Moritz Hahn, Arnd Hartmanns, Holger Hermanns, and Joost-Pieter Katoen. A
compositional modelling and analysis framework for stochastic hybrid systems. Formal
Methods in System Design, 43(2):191–232, 2013.
[11] Ondrej Holub, Majid Zamani, and Alessandro Abate. Efficient hvac controls: A symbolic
approach. In Control Conference (ECC), 2016 European, pages 1159–1164. IEEE, 2016.
13
[12] Woohyun Kim and Srinivas Katipamula. A review of fault detection and diagnostics
methods for building systems. Science and Technology for the Built Environment, pages
1–18, 2017.
[13] Niels Rode Kristensen, Henrik Madsen, and Sten Bay Jørgensen. Parameter estimation
in stochastic grey-box models. Automatica, 40(2):225–237, February 2004.
[14] Y. Ma, A. Kelman, A. Daly, and F. Borrelli. Predictive control for energy efficient buildings with thermal storage: modeling, stimulation, and experiments. Control Systems,
IEEE, 32(1):44–64, February 2012.
[15] S. Privara, J. Cigler, Z. Vana, F. Oldewurtel, C. Sagerschnig, and E. Zacekova. Building
modeling as a crucial part for building predictive control. Energy and Buildings, 56:8–22,
2013.
[16] S. Esmaeil Zadeh Soudjani and A. Abate. Aggregation and Control of Populations of
Thermostatically Controlled Loads by Formal Abstractions. Control Systems Technology,
IEEE Transactions on, 23(3):975–990, May 2015.
[17] Sadegh Esmaeil Zadeh Soudjani, Caspar Gevaerts, and Alessandro Abate. Faust 2: Formal Abstractions of Uncountable-STate STochastic processes. In TACAS, volume 15,
pages 272–286, 2015.
[18] Wei Yu, Baizhan Li, Hongyuan Jia, Ming Zhang, and Di Wang. Application of multiobjective genetic algorithm to optimize energy efficiency and thermal comfort in building
design. Energy and Buildings, 88:135–143, 2015.
[19] Rongpeng Zhang and Tianzhen Hong. Modeling of {HVAC} operational faults in building
performance simulation. Applied Energy, 202:178 – 188, 2017.
14
Appendix
Discrete-time models for heating system of two zones with
deterministic or stochastic dynamics
We present the corresponding system matrices of the three models Md , Mda , Ms described
in Sec. 3.1. Md is given by (1) and is characterised by the following system matrices:
0.6682 0
0.1320
3.4364
0.02632
0
0
0.6830
0
0.02096
2.9272 .
, Qd = 13.0207
A = 1.0005
, B = 0.1402
0
−0.000499
0
0
0
0.8004
0
0
0.1996
10.4166
Mda is given by (2) and is characterised by the same system matrices used for Md together
with,
8.760e−06
3.3378
0
0
2.704e−07 , Q
2.9106
Fda =
da = 13.0207 .
0
0
0
10.4166
0
Ms is given by (3) and is characterised by the same system matrices used for Md together
with,
0.0774 0
0
0
0
0.0774
0
0
.
Σ=
0
0
0.3872
0
0
0
0
0.3098
In all cases we use,
Twss = 18o C, TSP = 20o C, Tsw,bss = 75o C, Tz1ss = 20o C,
Tz2ss = 20o C, Trw,r1ss = 35o C, Trw,r2ss = 35o C.
We perform a simulation run of all three models over a two-day period, with
(
20o C k is between 8am - 12pm, 1pm - 6pm,
u[k] =
18o C otherwise,
and depict the resulting simulation runs for the temperature within zone 1 in Figure 3.
Reachability analysis using Axelarator
To perform reachability analysis using Axelarator we need to use the following commands in
command line,
../aa/axelerator Dynamics.txt -init
"[]" -sguard "[]" -inc -mpi -u
where the file Dynamics.txt contains the corresponding system matrices and defines the dimensions of both the state space and control inputs of the underlying model. The initial
conditions are defined using -init "[]" and -sguard "[]" will force the tool to stay inside
the safe region (if possible) defined within "[]". The commands -inc -mpi -u set the conditions of the problem such that Axelarator makes use of incremental search with multiple
precision intervals and to process unsound results. We also provide Axelerator with the safe
set defined in the form of a polyhedral,
Tz1
19.5
Tz1
20.5
>
∧
<
.
Tz2
19.5
Tz2
20.5
This safe set is used as a reference once the resulting reach tube is computed. The results
obtained using Axelarator are in the form of polyhedral sets which define the reach tube of
15
the underlying system over the N time steps. For Md the polyhedral sets are given as,
T
< 22.2282,
T
<
22,
z1
z2
Trw,r1
−Tz1
−Trw,r1
Tz1 +Tz2
Tz1 +Trw,r2
Tz1 −Trw,r1
−Tz1 +Tz2
−Tz1 +Trw,r2
−Tz1 −Trw,r1
Tz2 +Trw,r1
Tz2 −Trw,r1
−Tz2 +Trw,r1
−Tz2 −Trw,r1
Trw,r1 +Trw,r2
−Trw,r1 +Trw,r2
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
40,
Trw,r2
−10.0899,
−Tz2
105.958,
−Trw,r2
44,
Tz1 +Trw,r1
62,
Tz1 −Tz2
116.097,
Tz1 −Trw,r2
7,
−Tz1 +Trw,r1
25,
−Tz1 −Tz2
95.8203,
−Tz1 −Trw,r2
62,
Tz2 +Trw,r2
114.105,
Tz2 −Trw,r2
25,
−Tz2 +Trw,r2
100.555,
−Tz2 −Trw,r2
80,
Trw,r1 −Trw,r2
126.441, −Trw,r1 −Trw,r2
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
40,
−5.40334,
−11.1481,
62,
9.40198,
1.12788,
25,
−15.4932, ,
−21.238,
62,
−5.50237,
25,
−16.5515,
10,
94.8101
Tz1
Trw,r1
−Tz1
−Trw,r1
Tz1 +Tz2
Tz1 +Trw,r2
Tz1 −Trw,r1
−Tz1 +Tz2
−Tz1 +Trw,r2
−Tz1 −Trw,r1
Tz2 +Trw,r1
Tz2 −Trw,r1
−Tz2 +Trw,r1
−Tz2 −Trw,r1
Trw,r1 +Trw,r2
−Trw,r1 +Trw,r2
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
22.3242,
T z2
40,
Trw,r2
−10.1225,
−Tz2
109.275,
−Trw,r2
44,
Tz1 +Trw,r1
62,
Tz1 −Tz2
119.446,
Tz1 −Trw,r2
7,
−Tz1 +Trw,r1
25,
−Tz1 −Tz2
99.1041,
−Tz1 −Trw,r2
62,
Tz2 +Trw,r2
117.336,
Tz2 −Trw,r2
25,
−Tz2 +Trw,r2
103.886,
−Tz2 −Trw,r2
80,
Trw,r1 −Trw,r2
129.684, −Trw,r1 −Trw,r2
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
22,
40,
−5.38875,
−11.1512,
62,
9.50273,
1.20718,
25,
−15.5113,
−21.2737,
62,
−5.51993,
25,
−16.5399,
10,
98.1239.
whereas for Mda
Probabilistic reachability analysis using FAUST2
To solve this problem and achieve low errors within a computationally feasible time frame we
abstract the model Ms such that
xT = [Tz1 Tz2 ]T
(the states of interest and we fix the rest of the state to steady-state). Based on this model, we
construct a stochastic kernel in the form of a Gaussian conditional distribution, N (f (x, u), Σ)
where
f (x, u) = Ax + Bu + Qd
(cf. (3)) and
Σ = diag(∆[σz21 , σz22 ])
. The input space of the process is set to [15 22], while the safe-set is defined as
19.5 19.5
.
20.5 20.5
We select the PCTL Safety option in FAUST2 to compute the analysis. The model is
abstracted based on adaptive partitioning, with a partitioning error of = 10−5 , to obtain the
discretised transition kernel. The maximal safety probability for each partition is computed,
based on the transition kernel, recursively over the whole time horizon N = 6.
16
Discrete-time models for two-zone heating setup with large
number of continuous variables
We present the corresponding system matrices of the model Mc described in Subsection 3.2
as:
0.9998 6.54e−9 2.23e−5 2.23e−5 2.23e−5 4.88e−14 4.88e−14
0.000122
5.739e−9 0.9998 4.27e−14 4.27e−14 2.23e−5 2.23e−5 2.23e−5
0.0005 1.27e−12 0.9989 6.54e−9 6.54e−9 7.13e−18 7.13e−18
0.0005 1.27e−12 6.54e−9 0.9989 6.54e−9 7.13e−18 7.13e−18 ,
0.00051 0.00058 5.73e−9 5.73e−9 0.9989 6.54e−9 6.54e−9
1.11e−12 0.00058 6.25e−18 6.25e−18 6.54e−9 0.9989 6.54e−9
1.11e−12 0.00058 6.25e−18 6.25e−18 6.54e−9 6.54e−9 0.9980
1.027e−8 5.734e−9 7.31e−9 2.71e−15 0.0013 0.0014
1.91e−7 5.73e−9 1.39e−17 1.24e−6 0.0021 0.0022
2.00e−12 0.0005 2.13e−12 3.96e−19 3.84e−7 3.84e−7
0.0009 1.11e−12 2.13e−12 3.96e−19 3.84e−7 3.84e−7 ,
3.90e−11 2.09e−12 1.87e−12 3.63e−10 9.78e−7 9.78e−7
3.72e−11 0.00051 2.042e−21 3.63e−10 6.41e−7 6.41e−7
0.01708 1.11e−12 2.04e−21 3.63e−10 6.40e−7 6.41e−7
0.000122
3.58e−8
3.58e−8
Bc =
6.72e−8 ,
Ac =
Fc =
3.58e−8
3.58e−8
Qc =
0.2482
−0.0055
0.1270
0.0201 .
0.0145
0.0144
0.0145
We model the disturbances as random external effects affecting the room temperature dynamics as Tout [k] ∼ N (9, 1), Thall [k] ∼ N (15, 1), CO2i [k] ∼ N (500, 100), i ∈ {1, 2}, Trw,ri [k] ∼
N (35, 5), i ∈ {1, 2}.
Next, we present the abstract models Mca , a = {4, . . . , 1} taking the same form as (4)
with
xTc
[Tz1 Tw5 Tw2 Tw7 ]T
[Tz1 Tw5 Tw2 ]T
[Tz1 Tw2 ]T
[Tz1 ]T
Model
Mc4
Mc3
Mc2
Mc1
dTc
[Tout
[Tout
[Tout
[Tout
and, Tz2 is ∼ N (20, 1). For Mc4 we have
"
0.9998
0.00058
0.00058
0.00051
Ac4 =
"
Fc4 =
2.23e−5
0.9989
6.54e−9
5.73e−9
2.23e−5
6.54e−9
0.9989
5.73e−9
2.23e−5
6.54e−9
6.54e−9
0.9989
Thall CO21 Trw,r1 Tz2 ]T
Thall CO21 Trw,r1 ]T
CO21 Trw,r1 ]T
CO21 Trw,r1 ]T
#
"
, B c4 =
1.02e−8 5.73e−9 7.31e−9 0.0013 6.54e−9
2.00e−12 0.0005 2.13e−12 3.84e−7 1.27e−12
0.0009 1.11e−12 2.13e−12 3.84e−7 1.27e−12
1.75e−12 9.79e−13 1.87e−12 3.37e−7 0.00058
0.00012
3.5859e−8
3.5859e−8
3.1424e−8
#
#
, Qc4 =
,
0.2482
0.1270 ,
0.0145
0.0145
while for Mc3 ,
0.9998 2.23e−5 2.23e−5
0.00058 0.9989 6.54e−9
0.00058 6.54e−9 0.9980
6.29e−9 5.73e−9
1.22e−12 0.00051
0.00056 1.11e−12 2.13e−12 3.84e−7
Ac 3 =
Fc3 =
h 0.000122 i
, Bc3 = 0.000122 ,
3.58e−8
h 0.2482 i
7.31e−9 0.0013
2.13e−12 3.84e−7 , Qc3 = 0.1270 .
0.0145
For Mc2 the system matrices are
0.9998 2.237e−5
0.00012
Ac2 = 0.00058
, Bc2 = 3.58e−8
,
0.9989
h
i
0.2482
1.027e−8 7.31e−9 0.0013
Fc2 = 0.00091 2.13e−12 3.84e−7 , Qc2 = 0.1270 .
and Mc1 is described using
Ac1 = [ 0.9998 ], Bc1 = [ 0.000122 ],
Fc1 = [ 6.31e−5 7.31e−9 0.0013 ], Qc1 = [ 0.2482 ].
17
Models for heating system setup with multiple switching
controls
To construct the hybrid model we use Twss = 18o C, Tout = 25o C, TSP = 20o C,
δ = 1o C, δ2 = 2o C, δ3 = 3o C, δ4 = 4o C, δ5 = 5o C, ma,med = 10m3 /hr, ma,high =
15m3 /hr CO2ss = 500ppm. The continuous dynamics are composed via the following simplifying assumptions: (i) the outside air temperature is fixed Tout = 25o C and is used to warm
up the zone such the temperature set-point TSP = 20o C is maintained; (ii) we fix CO21ss ;
(iii) there is a constant wall temperature Twss ; (iv) there are no radiators within the zone;
and (v) there is no process noise. The resulting continuous state models for each discrete
mode using the corresponding models presented in Table ??.
Discrete state
q = (O, −)
q = (M, Op)
q = (M, Cl)
q = (H, Op)
q = (H, Cl)
Continuous space dynamics
Ṫz1 = −0.0116Tz1 (t) + 0.2565
Ṫsa = 0.0183Tz1 (t) − 0.0183Tsa (t)
Ṫz1 = −0.0292Tz1 (t) + 0.0176Tsa (t) + 0.2565
Ṫsa = 0.0183Tz1 (t) − 0.0185Tsa (t) + 0.005
Ṫz1 = −0.0292Tz1 (t) + 0.0176Tsa (t) + 0.2565
Ṫsa = 0.0183Tz1 (t) − 0.0183Tsa (t)
Ṫz1 = −0.038Tz1 (t) + 0.0264T sa(t) + 0.2565
Ṫsa = 0.0183Tz1 (t) − 0.0186Tsa (t) + 0.0076
Ṫz1 = −0.038Tz1 (t) + 0.0264T sa(t) + 0.2565
Ṫsa = 0.0186Tz1 (t) − 0.0186Tsa (t).
Reachability analysis using SpaceEx
We first implement the hybrid system delineated using Figure 7 within SpaceEx using SX
(the SpaceEx modelling language [4]). We further bound the zone temperature to lie between
Tz1 ∈ [10 30], while supply temperature to lie between Tsa ∈ [15 30]: corresponding to
physically feasible states when BAS is operating in a good condition. This model file is loaded
into SpaceEx and the input model is converted into a flat hybrid automaton representation
over which reachability analysis can be performed. Next, we set-up the configuration file
defining (i) the initial states in the form of "Tz1==15&Tsa==20&loc()==off", where "off"
refers to the initial state label, (ii) any forbidden states (in our case study there is none), (iii)
the time-horizon, and (iv) the direction of the reach sets (we select oct, but similar results
are achieved when the direction of reach set was set to box). Once the configuration file is
loaded, the reachability analysis algorithm is run over the defined time horizon, which in our
case corresponds to two hours and the reachable sets are generated (cf. Figure 8).
18
| 3cs.SY
|
Beta-gamma algebra identities and Lie-theoretic
exponential functionals of Brownian motion
arXiv:1404.7029v3 [math.PR] 27 Apr 2016
Reda Chhaibi
∗
In memoriam Marc Yor
Abstract
We explicitly compute the exit law of a certain hypoelliptic Brownian motion on a solvable Lie group. The underlying random variable
can be seen as a multidimensional exponential functional of Brownian
motion. As a consequence, we obtain hidden identities in law between
gamma random variables as the probabilistic manifestation of braid
relations. The classical beta-gamma algebra identity corresponds to
the only braid move in a root system of type A2 . The other ones seem
new.
A key ingredient is a conditional representation theorem. It relates
our hypoelliptic Brownian motion conditioned on exiting at a fixed
point to a certain deterministic transform of Brownian motion.
The identities in law between gamma variables tropicalize to identities between exponential random variables. These are continuous
versions of identities between geometric random variables related to
changes of parametrizations in Lusztig’s canonical basis. Hence, we
see that the exit law of our hypoelliptic Brownian motion is the geometric analogue of a simple natural measure on Lusztig’s canonical
basis.
MSC 2010 subject classifications: 60B15, 60B20, 60J65
Keywords: Beta-gamma algebra identities, Exponential functionals of Brownian motion, Braid relations, Total positivity, Brownian motion.
∗
Universität Zürich. Email: [email protected]
1
Contents
1 Introduction
2 Main results
2.1 Preliminaries . . . . . . .
2.2 Conditional representation
2.3 Exit law . . . . . . . . . .
2.4 Beta-gamma algebra . . .
2
. . . . .
theorem
. . . . .
. . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
4
10
12
13
3 Identities in law
17
3.1 Gamma identities . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Exponential identities . . . . . . . . . . . . . . . . . . . . . . 17
3.3 Geometric identities . . . . . . . . . . . . . . . . . . . . . . . 20
4 Proof of the conditional representation theorem
21
5 Review of Dufresne’s identity and the relationship proven
23
by Matsumoto and Yor
5.1 Exponential functionals of Brownian motion . . . . . . . . . . 23
5.2 Proofs of the relationship between Brownian motions of opposing drifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6 Some open questions
1
29
Introduction
Let G be a complex semi-simple group of rank r. We fix a Borel subgroup
B. B = N H where H ≈ (C∗ )r is a maximal complex torus and N is a
lower unipotent subgroup N . We denote by h ≈ Cr the Lie algebra of H. If
a ≈ Rr is the subspace where roots are real, we have h = a + ia. Moreover,
we write A := exp (a).
Since a is an Euclidean space thanks to the Killing form h·, ·i, there is a
natural notion of Brownian motion on a. Then the Brownian motion with
drift µ is denoted by:
(µ)
Xt := Xt + µ t
In the context of geometric crystals, the Robinson-Schensted correspondence with random input X (µ) is performed by solving, on the Borel subgroup B, a left-invariant stochastic differential equation driven by this Brownian motion ([Chh14] section 10, [Chh13]). One then obtains a Lie group
valued stochastic process Bt X (µ) , which was first introduced by [BBO05].
More is said in the preliminary section. We refer to this process as our hypoelliptic Brownian motion because its infinitesimal generator satisfies the
2
(parabolic) Hörmander condition. Although we will not make use of this
fact, it is reassuring to know that it has a smooth transition kernel.
The stochastic process Bt X (µ) has an N A decomposition:
Bt X (µ) = Nt X (µ) At X (µ)
The A part is a multiplicative Brownian motion and does not converge. We
focus on the N part which plays the role of a multidimensional exponential
functional of Brownian motion. Let ∆ be set the set of simple roots and C
be the open Weyl chamber:
C := {x ∈ a | ∀α ∈ ∆, hα, xi > 0}
When µ ∈ C the N part converges to N∞ X (µ) . We refer to the law of
N∞ X (µ) as the exit of law of the hypoelliptic Brownian motion.
The first result is the conditional representation theorem 2.12 that characterizes the law of a certain integral transform of Brownian motion as the
Brownian motion X (µ) conditioned to N∞ X (µ) being fixed.
As a consequence, we are able to give an explicit formula for the exit law
in theorem 2.16. This is our second main result. The expression involves
independent gamma variables. In the case of the group SL2 , one recovers
Dufresne’s identity in law on the exponential functional of a Brownian motion with drift [Duf90]. For W a standard Brownian motion and µ > 0,
R∞
(µ)
this identity states that the random variable 2 0 e−2Ws ds has the same
distribution as γ1µ , where γµ is a gamma random variable with parameter µ.
For groups with higher rank, the presented construction gives the explicit
law of multiple exponential functionals
of Brownian motion.
Almost surely, N∞ X (µ) belongs to the set of totally positive matrices
w0
N>0
⊂ N . The study of totally positivity in reductive groups has been
initiated by George Lusztig (see [Lus08] for a survey), motivated by the
w0
theory of canonical bases. We will only need the fact that N>0
possesses
equivalent charts indexed by reduced words i of the longest element
w0
in the Weyl group. Simply by noticing that the law of N∞ X (µ) uses
gamma variables and charts that depend on a choice of reduced word, we
find hidden identities in law between these gamma variables. Primitive
identities are associated to braid moves. It follows that we have as many
primitive identities as there are rank 2 root systems. That is the content of
our third main result, stated as theorem 2.21.
In fact, the identities between gamma variables tropicalize to identities
between exponential random variables. We also prove the discrete version
involving geometric random variables, using Lusztig’s parametrization of
canonical bases.
3
Structure of the paper
We begin by stating the three main theorems 2.12, 2.16 and 2.21 in section
2, after the necessary preliminaries on Lie theory and total positivity. We
will illustrate our claims thanks to examples from SL2 and SL3 .
It is more convenient to postpone the proof of the conditional representation theorem 2.12 and explain right away how it implies the two others.
In section 3, we show how rational identities between gamma variables tropicalize to min-plus identities between exponential variables. It implies the
discrete version involving geometric random variables.
In section 4, we prove the conditional representation theorem. Thanks
to results from [Chh14] and [Chh13], we reduce the problem to an induction
whose base case is a result by Matsumoto and Yor on a relationship between
Brownian motions with opposite drifts [MY01]. Just before diving into the
proof, we explain how the Matsumoto-Yor theorem is the SL2 case of ours.
We review this result for the sake of completeness in section 5, along with
Dufresne’s identity. This section is absent in the published version.
Finally, we conclude with some open questions.
Acknowledgments
The author is grateful to Marc Yor for fruitful discussions on exponential
functionals of Brownian motion and beta-gamma identities. This paper is
dedicated to his memory. I am also thankful to Philippe Bougerol for his
guidance during my PhD thesis [Chh13], on which this article is based.
2
2.1
Main results
Preliminaries
Lie theory: We will need some (mostly standard) notations and terminology for semi-simple groups and algebras (see for example [Spr09], [Hum72]).
Let g be a complex semi-simple Lie algebra of rank r, and h ≈ Cr is a
maximal abelian subalgebra, the Cartan subalgebra. It has a Cartan decomposition g = n ⊕ h ⊕ u. ∆ = {αi , 1 ≤ i ≤ r} ⊂ h∗ (resp. ∆∨ ⊂ h)
denote the simple roots (resp. coroots) of g. The real part of the Cartan subalgebra a ( h = a + ia) is the subspace where simple roots are
real valued.
The
of g is entirely encoded by the Cartan matrix
structure
∨
A = ai,j = αi αj
. The coefficients in the Cartan matrix de1≤i,j≤r
termine the relations between Chevalley generators (fα , hα = α∨ , eα ), for
α ∈ ∆. Simple roots form a basis of h∗ dual to the fundamental coweights
(ωα∨ )α∈∆ .
Let G be a simply-connected complex Lie group with Lie algebra g. N ,
H and U are the subgroups with Lie algebras n ⊕ h ⊕ u. H ≈ (C∗ )r is a
4
maximal torus. B = N H and B + = HU form a pair of opposite Borel
subgroups. We have a one-parameter subgroup for every α ∈ ∆:
∀t ∈ C, yα (t) := exp(tfα ), xα := exp (teα )
When convenient, they will sometimes be written yi and xi , 1 ≤ i ≤ r.
We identify h and h∗ thanks to the Killing form h·, ·i. In this identifica2β
2α
tion, α∨ = hα,αi
. In general, we will write for any β ∈ h∗ , β ∨ = hβ,βi
∈ h.
The reflection on h with respect to the hyperplane ker β is:
∀h ∈ h, sβ (h) = h − β (h) β ∨
The Weyl group of G is defined as W = Norm(H)/H. It acts on the torus
H by conjugation and hence on h. As a Coxeter group, it is generated by
the reflections (sα )α∈∆ . Every w ∈ W can be written as a product si1 . . . sik
for a sequence i = (i1 , . . . , ik ). A reduced word for w ∈ W is a sequence i
of shortest possible length ℓ(w). The set of all possible reduced words for w
is denoted by R(w). The Weyl group has a unique longest element denoted
by w0 and we set m = ℓ(w0 ).
The Bruhat decomposition states that G is the disjoint union of cells:
G
G
G=
B + ωB + =
Bτ B +
τ ∈W
ω∈W
In the largest opposite Bruhat cell BB + = N HU , every element g admits
a unique Gauss decomposition in the form g = nau with n ∈ N , a ∈ H,
u ∈ U . In the sequel, we will write g = [g]− [g]0 [g]+ , [g]− ∈ N , [g]0 ∈ H and
[g]+ ∈ U for the Gauss decomposition. Also [g]−0 := [g]− [g]0 .
Remark 2.1. The reader unfamiliar with Lie groups can have in mind
the example of SLn (C), of rank r = n − 1. The following matrices can
be chosen as Chevalley generators. If Ei,j = (δi,r δj,s )1≤r,s≤n are the usual
elementary matrices, then hi = Ei,i − Ei+1,i+1 , ei = Ei,i+1 , and fi = Ei+1,i .
h is the set of complex
diagonal matrices with zero trace, which we identify
P
n
with {x ∈ C |
xi = 0}. Then H is the set of diagonal matrices with
determinant 1. N (resp. U ) is the set of lower (resp. upper) triangular
unipotent matrices. We have:
∀t ∈ C, yi (t) = Id + tEi+1,i , xi (t) = Id + tEi,i+1
The Weyl group W is the group of permutation matrices and acts on h by
permuting coordinates. In the identification with the symmetric group acting
on n elements, the reflections si are identified with transpositions (i i + 1).
The longest word w0 reorders the elements 1, 2, . . . , n in decreasing order.
In the case of GLn , the second Bruhat decomposition is known in linear
algebra as the LPU decomposition which states that every invertible matrix
5
can be decomposed into the product of a lower triangular matrix L, a permutation matrix P and an upper triangular matrix U . P is unique. The
largest opposite Bruhat cell corresponds to P = id. It is dense as it is the
locus where all principal minors are non-zero.
Remark 2.2 (Cartan-Killing classification). Complex simple Lie algebras
are known, and are classified by types:
• Type Ar−1 : g = slr G = SLr (C)
• Type Br : g = so2r+1 G = SO2r+1 (C)
• Type Cr : g = spr G = Spr (C)
• Type Dr : g = so2r G = SO2r (C)
• Exceptional types: E6 , E7 , E8 , F4 , G2 .
The number in subscript indicates the rank.
The set of roots is denoted by
Φ := {β ∈ h∗ | sβ ∈ W } .
F
Roots split into positive and negative roots Φ = Φ+ Φ− , with Φ+ being the
set of roots that can be written as a positive sum of simple roots. Reduced
expressions of Weyl group elements give convex orderings of positive roots
(see [Hum90]):
Lemma 2.3. Let (i1 , . . . , ik ) be a reduced expression of w ∈ W . Then for
j = 1, . . . , k:
βi,j := si1 . . . sij−1 αij
produces all the positive roots in the set of inversions of w:
Inv (w) := β ∈ Φ+ , wβ ∈ Φ− .
For w = w0 , it produces all positive roots i.e Inv (w0 ) = Φ+ .
When i ∈ R(w0 ), we call the sequence (βi,j )1≤j≤m an enumeration of
positive roots. When the chosen reduced expression is obvious from context,
we will drop the subscript i. In appendix, we list positive root enumerations
for rank 2 systems.
6
Total positivity: In the classical case of GLn , a totally non-negative matrix is a matrix such that all its minors are non-negative. Thanks to the
classical Cauchy-Binet formula, totally non-negative matrices (GLn )≥0 form
a semi-group. In [Lus94], Lusztig generalized total positivity to reductive
groups motivated by the theory of canonical bases. At that level of generality, taking the semi-group property as a definition is simpler. The totally
non-negative part of G is denoted G≥0 and is defined as the semi-group
generated by the following sets:
• The semi-group H>0 := A = exp (a) ≈ R∗+
r
• The semi-group generated by {xα (t), t > 0, α ∈ ∆}: U≥0
• The semi-group generated by {yα (t), t > 0, α ∈ ∆}: N≥0
We will be interested in parametrizations of G≥0 . Lusztig proved that
totally non-negative elements admit a Gauss decomposition made of totally
non-negative elements.
Theorem 2.4 ([Lus94] Lemma 2.3). Any element g ∈ G≥0 has a unique
Gauss decomposition g = nau with n ∈ N≥0 , a ∈ A and u ∈ U≥0 .
Therefore, one can focus on the parametrizations of U≥0 . The following
theorem generalizes a result of Whitney [Whi52] from GLn to G. It says that
a totally positive element in U can be written as a product of elementary
matrices, depending on where it falls in the Bruhat decomposition.
Theorem 2.5 ([Lus94] Proposition 2.7, [BZ97] Proposition 1.1). For any
w ∈ W with k = ℓ(w), every reduced word i = (i1 , . . . , ik ) in R(w) gives rise
w := U
to a parametrization of U>0
≥0 ∩ BwB by:
xi :
w
→
U>0
Rk>0
(t1 , . . . , tk ) 7→ xi1 (t1 ) . . . xik (tk )
We also define the transition maps:
Ri,i′ := x−1
i ′ ◦ xi
which play a particularly important role.
For the purposes of this paper, we will only need the case of w = w0 and
w0
we will call such a parametrization the Lusztig parametrization of U>0
. The
w0
positive reals (t1 , . . . , tm ) such that u = xi (t1 , . . . , tm ) ∈ U>0 will be called
Lusztig parameters of u, the dependence in i being implicit.
7
On Lusztig’s canonical basis: We will only be interested in the parametrizations of the canonical basis. The combinatorics of total posivity in the group
G are related to the combinatorics of Lusztig’s canonical basis for the dual
group, after a tropicalization procedure [.]trop . More is said on tropical
ization in section 3 where Ri,i′ trop is defined. For now, let us record the
following:
Theorem 2.6 (see [Lus90a], [Lus90b]). There is a canonical basis L of
Uq (n∨ ), the lower unipotent part of the dual quantum group Uq (g∨ ). It has
a natural parametrization associated to every reduced word i ∈ R (w0 ):
Li : Nℓ(w0 ) −→ L
and the changes of parametrizations are given by:
Ri,i′ trop = L−1
i′ ◦ Li .
Brownian motion(s): Since a is made into an Euclidean space thanks
to the Killing form h·, ·i, there is a natural notion of Brownian motion on a.
Indeed, h·, ·i is a scalar product once restricted to a and the induced norm
is denoted by || · ||. Let (Ω, A, P) be the canonical probability space of a
Brownian motion X on a. For µ ∈ a, the Brownian motion with drift µ is
denoted by:
(µ)
Xt := Xt + µ t .
Brownian motion with drift µ and starting position x0 ∈ a is written:
(x0 ,µ)
Xt
:= x0 + Xt + µ t .
One obtains a left-invariant Brownian motion Bt (X (µ) ) on B by solving
the following stochastic differential equation driven by any path X. The
symbol ◦ indicates that the SDE has to be understood in the Stratonovitch
sense:
P
(
(µ)
1
hα,
αif
dt
+
dX
dBt (X (µ) ) = Bt (X (µ) ) ◦
α
t
α∈∆ 2
(2.1)
B0 (X (µ) ) = Id
The SDE can be explicitly solved ([BBO05]):
Z
k
Y
X
||αij ||2 −αij (Xt(µ) ) X (µ)
e t
j
e
Bt (X (µ) ) =
dt
f
.
.
.
f
j
i1
ik
2
t≥tk ≥···≥t1 ≥0
j=1
k≥0
i1 ,...,ik
(2.2)
8
Clearly, Bt (X (µ) ) has an N A decomposition is given by:
2
X Z
(µ)
(µ)
−α (X )···−αik (Xt ) ||αi1 ||
(µ)
k
e i1 t1
Nt (X ) =
...
2
t≥tk ≥···≥t1 ≥0
(2.3)
k≥0
i1 ,...,ik
...
||αik ||2
fi1 . . . fik dt1 . . . dtk
2
(µ)
At X (µ) = eXt
(2.4)
(2.5)
2
Remark 2.7. The presence of half squared norms ||α||
is here to account
2
for the fact that time does not flow in the same fashion for all roots. In the
simply-laced groups (ADE-types), one can choose all roots to be the same
2
length, hence choosing ||α||
2 = 1 for all α.
Remark 2.8. In the A1 case only, we opt out of the normalization made in
the previous remark. Indeed, the classical choice for the only root is α = 2.
2
Hence the factor ||α||
2 = 2 in the following example of SL2 .
Example 2.9 (SL2 - A1 type). Let X (µ) be a Brownian motion with drift
µ on R. The SDE is:
!
(µ)
0
dXt
(µ)
(µ)
dBt (X ) = Bt (X ) ◦
(µ)
2dt
−dXt
and its solution is:
(µ)
Bt (X (µ) ) =
eXt
(µ)
eXt
Rt
0
0
(µ)
2e−2Xs
(µ)
ds e−Xt
!
Example 2.10 (SLn - An−1 type).P
As in remark 2.1, let X be a Brownian
n
motion with drift µ on {x ∈ R |
xi = 0}. For notational reasons, we
drop the superscript (µ) and put indices as exponents. The SDE becomes:
1
0
0
···
0
dXt
..
..
dt dX 2 0
.
.
t
..
..
dBt (X (µ) ) = Bt (X (µ) ) ◦ 0
.
.
0
dt
..
..
..
.
.
. dXtn−1
0
0
···
0
dt
dXtn
and its solution Bt (X (µ) ) is given by:
1
eXt
0
0
R
2
1
2 −X 1
t
X
X
X
e t
0
e t 0 e s s ds
1 Rt 2
R
R
2
1
1
3
2
3
2
s
t
Xs1 −Xs1
eXt
ds1 0 1 eXs2 −Xs2 ds2 eXt 0 eXs −Xs ds eXt
0 e
..
..
..
.
.
.
9
···
· · ·
..
.
..
.
r
The process At (X (µ) ) is a multiplicative Brownian motion on A ≈ R∗+
and does not converge. The main theorems concern the N part, which
converges when the drift is inside the Weyl chamber, and we are able to give
an explicit formula for the exit law. On a few occasions, we will need to
consider the path π driving the flow B. (π) to be any path. The previous
construction carries verbatim if π is a semi-martingale.
2.2
Conditional representation theorem
L
The symbol = stands for equality in law between random variables. Moreover, a generic gamma random variable with parameter a > 0 is denoted
γa .
1 a−1 −x
P (γa ∈ dx) =
x e 1R+ (x) dx
Γ(a)
where Γ is the gamma Euler function.
Let C (R+ , a) be the space of continuous functions with values in a. The
following theorem defines our path transform:
Theorem 2.11 ( [BBO09] Proposition 6.4 , [Chh14] section 5 ). For any
totally non-negative g ∈ U≥0 , there is a path transform Tg : C (R+ , a) −→
C (R+ , a) such that for any path X:
∀t ≥ 0, Bt (Tg X) = [gBt (X)]−0
This transform is well-defined in the sense that the Gauss decomposition
always exists for g ∈ U≥0 .
From now on, fix a reduced word i ∈ R(w0 ) of length m = ℓ(w0 ) and
call (β1 , . . . , βm ) the associated positive roots enumeration. Choose w̄0 to
be any representative in G on the longest element w0 in W = Norm(H)/H.
Also, we introduce the shift vector:
X
hα, αi
ωα∨
θ=
log
2
α∈∆
where ωα∨ are the fundamental coweights. As in remark 2.7, for ADE groups,
θ = 0 because all roots can be chosen and are chosen to be of the same
squared norm 2. We can state the following theorem whose proof is postponed to section 4:
w0
Theorem 2.12 (Conditional representation theorem). Let g ∈ U>0
,µ∈C
and W a standard Brownian motion on
a.
Then Λx0 := x0 + Teθ ge−θ W (w0 µ) is distributed as a Brownian motion
X x0 ,(µ)
• with drift µ
10
• with initial position x0
• conditioned on N∞ (X 0,(µ) ) = Θ(g) where
w0
w0
Θ : U>0
−→ N>0
is the bijective function Θ(g) := [gw̄0 ]−
Moreover, if we pick g = xi1 (t1 ) . . . xim (tm ) , m = ℓ(w0 ) being random with
L
independent Lusztig parameters such that tj = γhβj∨ ,µi , then Λx0 is a standard
Brownian motion with drift µ starting at x0 .
Remark 2.13. The conditioning approach owes a lot to Baudoin and O’Connell
[BO11] in spirit, but ends up being quite different. Indeed, that paper considered
conditioning of Brownian motion X (µ) with respect to simple integrals
R ∞ a−α(X
(µ) )
, for α a simple root. Nevertheless, the geometric path model
0 e
developed in [Chh14] and [Chh13] makes it more natural to condition with
respect to N∞ (X (µ) ). The random variable N∞ (X (µ) ) not only contains the
simple integrals, but also iterated ones.
Example 2.14 (Θ in A1 -type). For G = SL2 , one can choose:
0 −1
w̄0 =
1 0
Then:
Θ
1 0
1 t
1 t
0 −1
= 1
=
0 1
0 1
1 0
t 0
−
Example 2.15 (Θ in A2 -type). For G = SL3 , one can choose:
0 0 1
w̄0 = 0 −1 0
1 0 0
If:
Then:
1 t1 0
1 0 0
1 t3 0
g = 0 1 0 0 1 t2 0 1 0
0 0 1
0 0 1
0 0 1
1
1 t1 + t3 t1 t2
0 0 1
1
1
t2 0 −1 0 = t1
Θ (g) = 0
1
0
0
1
1 0 0
−
t1 t2
11
0
1
t
1+ t1
3
t2
0
0
1
2.3
Exit law
Thanks to the conditional representation theorem, we obtain the exit law
with little effort.
Theorem 2.16 (Law of N∞ (X (µ) )). If X (µ) is a Brownian motion with
drift µ the Weyl chamber, then Nt (X (µ) ) converges almost surely inside the
w0
open cell N>0
and N∞ (X (µ) ) = Θ (xi1 (t1 ) . . . xim (tm )) where the Lusztig
parameters tj are independent random variables with:
L
tj = γhβj∨ ,µi
Proof. The condition µ ∈ C entails the convergence of iterated integrals in
the explicit expression of Nt (X), equation (2.3). The fact that N∞ (X (µ) ) ∈
w0
N>0
follows from a total posivity criterion (Theorem 1.5 in [BZ97] or Theorem 1.11 in [FZ99]) formulated in terms of generalized minors. Basically,
one has only to adapt the proofs of theorem 5.2 in [Chh14], or lemma 4.3 in
[BBO05] to the case of infinite time horizon.
The law of N∞ (X (µ) ) comes directly from theorem 2.12.
It is worth noting that the probability measure has a smooth density and
w0
charges the entire space N>0
, which is an open dense cell of N≥0 . Other
cells are of smaller dimension and therefore of zero measure. Hence, it can
be viewed as an N≥0 -valued random variable.
Let us examine the example of SL2 , the group with smallest rank.
Example 2.17 (A1 -type). For G = SL2 , a = R, α = 2, α∨ = 1 and:
!
1
0
(µ)
N∞ (X ) = R ∞ −2X (µ)
s ds
1
0 2e
We know that Θ
that:
1 0
= 1
, and as such theorem 2.16 tells us
y 1
!
1
0 L 1 0
R ∞ −2X (µ)
= 1
s ds
1
γµ 1
0 2e
1 y
0 1
This is exactly Dufresne’s identity in law [Duf90].
Here, the law of N∞ (X (µ) ) can be seen as a meaningful generalization of
Dufresne’s identity which uses an inverse gamma. In the group setting, we
see that the map Θ plays the role of the inverse map. Higher dimensional
examples give new identities about exponential functionals of Brownian motion.
12
Example 2.18 (A2 -type). For G = SL3 , a = x ∈ R3 | x1 + x2 + x3 = 0 .
(µ)
Consider a Brownian motion
with
drift µ in the Weyl chamber.
X on a
1
0
The simple roots are α1 = −1 , α2 = 1 . We obtain that:
0
−1
0
R ∞ −α (X (µ) )
1
s
ds
1
0
N∞ (X (µ) ) =
0 e
R
R
R ∞ −α (X
(µ)
(µ)
(µ)
s −α2 (Xu )
∞
1
s ) ds
du 0 e−α2 (Xs ) ds 1
0 e
0 e
1
0
We choose the reduced word i = (1, 2, 1) ∈ R(w0 ). If:
L
(t1 , t2 , t3 ) = γhα∨1 ,µi , γhα∨1 +α∨2 ,µi , γhα∨2 ,µi
are independent gamma random variables with corresponding parameters,
then example 2.15 and theorem 2.16 tell us that:
1
0
0
L 1
1
0
(2.6)
N∞ (X (µ) ) = t1
t
1+ t1
1
t1 t2
t2
1
3
This result in itself seems new. It can be restated as follows in terms of
a = hα∨
the two positive parameters a = hα∨
2 , µi > 0 and the two
2 , µi >, e
correlated real Brownian motions:
1
Bt := √ Xt1 − Xt2 ;
2
et := √1 Xt2 − Xt3
B
2
The equality (2.6) is equivalent to the fact that the triple
Z ∞ √
Z ∞ √
Z ∞ √
Z
e
e− 2Bs −as ds,
e− 2Bs −eas ds,
e− 2Bs −as ds
0
0
has the same distribution as
0
1
γa ,
1+ γγa
a
e
γa+ea
,
0
1
γa γae
s
e−
√
eu −e
2B
au
du
.
Notice that the two first marginals are consistent with Dufresne’s identity. In order to see that, one has to perform a Brownian rescaling of time
by a factor 2 and invoke the classical identity
2.4
1+ γγa L
1
a
e
γa+ea = γae .
Beta-gamma algebra
We have then a natural notion of gamma law and inverse gamma law. Here
γ. denote independent gamma random variables.
13
Definition 2.19 (gamma law on U and inverse gamma on N ). For µ ∈ C
and (β1 , . . . , βm ) being the positive roots enumeration associated to a reduced
expression of w0 = si1 . . . sim , define Γµ to be the law of the positive (in the
w0
sense of total positivity) U>0
-valued random variable
L
∨ ,µi
Γµ = xi1 γhβ1∨ ,µi . . . xim γhβm
w0
Define the inverse gamma law on N>0
as:
L
Dµ = Θ (Γµ ) .
Those laws are well defined, in the sense that the above expressions do
not depend on the choice of a reduced expression:
Proposition 2.20. If i and i′ are reduced words of w0 , then for every µ ∈
C, the following equality in law holds between independent gamma random
variables:
L
∨ ,µi , . . . , γhβ ∨ ,µi
γhβ ∨′ ,µi , . . . , γhβ ∨′ ,µi = Ri,i′ γhβi,1
i,m
i ,1
i ,m
Proof. Theorem 2.16 relates the above laws to the exit law of our hypoelliptic
Brownian motion, which is intrinsically defined.
The formula defining Γµ has more to it than meets the eye. Indeed,
in order to have the law being the same for all reduced expressions of w0 ,
there must be hidden non-trivial equalities in law. These can be qualified
as identities from the beta-gamma algebra, as some authors call them (see
e.g [CPY98] and references therein). More is said in the next section. It
is remarkable to think of them as a probabilistic manifestation of a group
structure, more precisely of braid relationships.
If s, s′ ∈ W are reflections associated to simple roots and d is the order
of ss′ , a braid relationship in W is the equality between d terms:
ss′ s · · · = s′ ss′ . . .
A braid move or a d-move occurs when substituting ss′ s . . . for s′ ss′ . . .
within a reduced word. Considering two reduced expressions of the same
Weyl group element, a well known result due to Hideya Matsumoto [Mat64]
and Tits [Tit69] states that one is obtained from the other by successive
braid moves (see also [Hum90]).
si sj si · · · = sj si sj . . .
Using this fact, saying that Γµ is defined unambiguously is equivalent to
saying that proposition 2.20 holds for reduced words which differ by a braid
move. Hence for any µ ∈ a, such that αi (µ) > 0 and αj (µ) > 0, one has:
xi γhα∨i ,µi xj γhsi α∨j ,µi xi γhsi sj α∨i ,µi . . .
14
L
=xj γhα∨j ,µi xi γhsj α∨i ,µi xj γhsj si α∨j ,µi . . .
It turns out that the rank 2 cases (A2 , B2 , C2 and G2 ) contain all the possible
hidden identities. This is the classical rank 2 reduction. Fix one of these
root systems, and consider two reduced words i, i′ of the longest element
w0 in a rank 2 system. The maps Ri,i′ can be computed for every group
explicitly and have been tabulated by Berenstein and Zelevinsky as theorem
5.3 in [BZ97]. Now write thanks to fundamental coweights µ = a1 ωi∨ + a2 ωj∨
and (p1 , p2 , . . . ) = Ri,i′ (t1 , t2 , . . . ). By using the explicit formulas for the
change of Lusztig parameters and the root enumerations given in tables 6,
6, 6 and 6, we obtain the following theorem:
Theorem 2.21 (Beta-gamma algebra identities). Associated to every rank
2 root system, the following identities in law hold:
• Type A2 (table 6): w0 = s1 s2 s1 = s2 s1 s2 . i = (1, 2, 1) and i′ = (2, 1, 2).
t2
t3
, p2 = t1 + t3 , p3 = t1t1+t
. Then:
Let p1 = t1t2+t
3
3
L
(t1 , t2 , t3 ) = (γa1 , γa1 +a2 , γa2 )
if and only if
L
(p1 , p2 , p3 ) = (γa2 , γa1 +a2 , γa1 )
• Type B2 (table 6): w0 = s1 s2 s1 s2 = s2 s1 s2 s1 . i = (1, 2, 1, 2) and
i′ = (2, 1, 2, 1). Let:
p1 =
t2 t23 t4
π2
π2
t1 t2 t3
, p2 =
, p3 = 1 , p4 =
π2
π1
π2
π1
where π1 = t1 t2 + (t1 + t3 )t4 , π2 = t21 t2 + (t1 + t3 )2 t4 . Then:
L
(t1 , t2 , t3 , t4 ) = (γa1 , γa1 +a2 , γa1 +2a2 , γa2 )
if and only if
L
(p1 , p2 , p3 , p4 ) = (γa2 , γa1 +2a2 , γa1 +a2 , γa1 )
• Type C2 (table 6): w0 = s1 s2 s1 s2 = s2 s1 s2 s1 . i = (1, 2, 1, 2) and
i′ = (2, 1, 2, 1). Let:
p1 =
t2 t3 t4
π2
π2
t1 t32 t3
, p2 = 1 , p3 =
, p4 =
π1
π2
π1
π1
where π1 = t1 t2 + (t1 + t3 )t4 , π2 = t3 t24 + (t2 + t4 )2 t1 . Then:
L
(t1 , t2 , t3 , t4 ) = (γa1 , γ2a1 +a2 , γa1 +a2 , γa2 )
15
if and only if
L
(p1 , p2 , p3 , p4 ) = (γa2 , γa1 +a2 , γ2a1 +a2 , γa1 )
• Type G2 (table 6): w0 = s1 s2 s1 s2 s1 s2 = s2 s1 s2 s1 s2 s1 . i = (1, 2, 1, 2, 1, 2)
and i′ = (2, 1, 2, 1, 2, 1). Let:
p1 =
π3
π3
t2 t33 t24 t35 t6
, p2 =
, p3 = 2
π3
π2
π3 π4
p4 =
π3
t1 t2 t23 t4 t5
π3
, p5 = 1 , p6 =
π1 π2
π4
π1
where:
π1 =t1 t2 t23 t4 + t1 t2 (t3 + t5 )2 t6 + (t1 + t3 )t4 t25 t6
π2 =t21 t22 t33 t4 + t21 t22 (t3 + t5 )3 t6 + (t1 + t3 )2 t24 t35 t6
+ t1 t2 t4 t25 t6 (3t1 t3 + 2t2 + 2t3 t5 + 2t1 t5 )
π3 =t31 t22 t33 t4 + t31 t22 (t3 + t5 )3 t6 + (t1 + t3)3 t24 t35 t6
+ t21 t2 t4 t25 t6 (3t1 t3 + 3t2 + 3t3 t5 + 2t1 t5 )
π4 =t21 t22 t33 t4 (t1 t2 t33 t4 + 2t1 t2 (t3 + t5 )3 t6 + (3t1 t3 + 3t23 + 3t3 t5 + 2t1 t5 )t4 t25 t6 )
+ t26 (t1 t2 (t3 + t5 )2 + (t1 + t3 )t4 t25 )3
Then:
L
(t1 , t2 , t3 , t4 , t5 , t6 ) = (γa1 , γ3a1 +a2 , γ2a1 +a2 , γ3a1 +2a2 , γa1 +a2 , γa2 )
if and only if
L
(p1 , p2 , p3 , p4 , p5 , p6 ) = (γa2 , γa1 +a2 , γ3a1 +2a2 , γ2a1 +a2 , γ3a1 +a2 , γa1 )
16
3
Identities in law
3.1
Gamma identities
Let us make some remarks on the obtained gamma identities. The classical
beta-gamma algebra identity states that if γa and γb are two independent
gamma random variables with parameters a > 0 and b > 0, then:
γa
, γa + γb
(3.1)
γa + γb
form a pair of independent random variables.
This fact is easy to retrieve from the A2 case. Indeed, by considering
L
(t1 , t2 , t3 ) = (γa1 , γa1 +a2 , γa2 ) independent variables with the designated laws
and (p1 , p2 , p3 ) algebraically defined as above, we know that p2 and p3 are
independent. And since p2 independent of t2 , we get that p2 = γa1 + γa2 is
1
.
independent of pt23 = γa γa
1 +γa2
Moreover, Lukacs [Luk55] proved that the independence for the pair in
(3.1) characterizes the gamma law. This leads to the first open question in
section 6.
3.2
Exponential identities
A generic exponential random variable with parameter µ > 0 is denoted by
eµ :
P (eµ ∈ dx) = µe−µx 1R+ (x) dx
Consider a rational expression in k variables a ∈ Q (x1 , . . . , xk ) that
has no minus sign. Tropicalizing a to [a]trop tantamounts to replacing the
algebraic operations (+, ×, /) by (min, +, −). A rational expression in the
operations (min, +, −) is now commonly referred to as a tropical expression.
Formally, if a and b are rational subtraction-free functions, then:
[a + b]trop = min(a, b)
[ab]trop = a + b
[a/b]trop = a − b
[a ◦ b]trop = [a]trop ◦ [a ◦ b]trop
For example:
t2 t3
t1 + t3
trop
= t2 + t3 − min(t1 , t3 )
A less algebraic definition could be used, using a limit that always exists:
17
Proposition 3.1 (Analytic tropicalization). For a rational and subtractionfree expression in k variables a ∈ Q (x1 , . . . , xk ), we have that for h > 0:
x1
xk
−h log a e− h , . . . , e− h = [a]trop (x1 , . . . , xk ) + O(h)
(3.2)
where O(h) is a quantity such that
the variables (x1 , . . . , xk ).
O(h)
h
is bounded as h → 0, uniformly in
Proof. Let us prove the statement by induction on the size of the expression
a, meaning the number of operations it uses (addition, multiplication and
division). For the base case, notice that if a is a monomial or a constant, then
the statement is trivially true. Now, for the inductive step, if a is a product
or ratio of two rational subtraction-free expressions, for which the statement
is true, the statement follows using the properties of the logarithm. If a is
a sum whose terms satisfy the induction hypothesis:
a = a1 + a2 ,
then for x = (x1 , . . . , xk ) ∈ Rk :
x1
xk
− h log a e− h , . . . , e− h
xk
x1
xk
xk
= − h log a1 e− h , . . . , e− h + a2 e− h , . . . , e− h
−1
−1
= − h log eO(1)−h [a1 ]trop (x) + eO(1)−h [a2 ]trop(x)
= min ([a1 ]trop (x), [a2 ]trop (x)) + O(h)
−1
− h log 1 + eO(1)−h |[a1 ]trop−[a2 ]trop|(x)
=[a]trop (x) + O(h)
It is well known that Ri,i′ tropicalize to changes of parametrizations in
Lusztig’s canonical basis (Theorem 2.6). Therefore, we consider the following crystallizing procedure for the rational subtraction-free expressions Ri,i′ .
The use of the logarithm function has to be understood component-wise:
−u1 −u2
Ri,i′ trop (u1 , u2 , . . . ) := lim −h log Ri,i′ e h , e h , . . .
h→0
Now, on the probabilistic side, one can recover a tropical version of
gamma identities involving exponential variables. The following lemma
shows how gamma variables degenerate to exponential variables.
Lemma 3.2 (Yor [Yor12]). As h > 0 goes to zero, −h log γhµ converges in
law to eµ , an exponential variable with parameter µ.
18
Proof. A direct proof using densities is possible and straightforward. Let us
rather present an aesthetically more pleasing derivation suggested by Marc
Yor during a private communication. We can reduce the problem to µ = 1
L
because eµ = µ1 e1 and:
lim −h log γhµ =
h→0
1
lim −h log γh
µ h→0
Now, if βa,b is a beta random variable with parameters a > 0 and b > 0, we
can write γh as a product of two independent random variables:
L
γh = βh,1 γ1+h
h is
Knowing that (βh,1 )h is in fact a uniform random variable, − log βh,1
distributed as e1 . Hence:
L
−h log γh = e1 − h log γ1+h
The proof is finished upon noticing that h log γ1+h converges to zero in
probability.
Hence the tropical version of proposition 2.20:
Proposition 3.3. If i and i′ are reduced words of w0 , then for every µ ∈ C,
the following equality in law holds between independent exponential random
variables:
L
∨ ,µi , . . . , ehβ ∨ ,µi
ehβ ∨′ ,µi , . . . , ehβ ∨′ ,µi = Ri,i′ trop ehβi,1
i,m
i ,1
i ,m
Proof. Before tropicalization, for the parameter hµ, we have:
L
∨ ,hµi , . . . , γhβ ∨ ,hµi
γhβ ∨′ ,hµi , . . . , γhβ ∨′ ,hµi = Ri,i′ γhβi,1
i,m
i ,1
i ,m
Then, using the previous lemma:
ehβ ∨′ ,µi , ehβ ∨′ ,µi , . . .
i ,1
i ,2
L
= lim −h log γhβ ∨′ ,hµi , γhβ ∨′ ,hµi , . . .
i ,1
i ,2
h→0
L
∨ ,hµi , γhβ ∨ ,hµi , . . .
= lim −h log Ri,i′ γhβi,1
i,2
h→0
∨ ,µi
∨ ,µi
−h log γhhβi,2
−h log γhhβi,1
, exp
,...
= lim −h log Ri,i′ exp
h→0
h
h
∨ ,µi , ehβ ∨ ,µi , . . .
= Ri,i′ trop ehβi,1
i,2
19
Again, as all the information is contained in the rank two case, there
is a finite list of identities between exponential variables that sums up the
results so far.
Theorem 3.4 (Exponential identities). Associated to every rank 2 root system, the following identities in law hold:
• Type A2 (table 6): w0 = s1 s2 s1 = s2 s1 s2 . Let p1 = t2 + t3 −
min(t1 , t3 ), p2 = min(t1 , t3 ), p3 = t1 + t2 − min(t1 , t3 ).Then:
L
(t1 , t2 , t3 ) = (ea1 , ea1 +a2 , ea2 )
if and only if
L
(p1 , p2 , p3 ) = (ea2 , ea1 +a2 , ea1 )
• Type B2 (table 6): w0 = s1 s2 s1 s2 = s2 s1 s2 s1 . Let:
p1 = t2 + 2t3 + t4 − π2 , p2 = π2 − π1 , p3 = 2π1 − π2 , p4 = t1 + t2 + t3 − π1
where π1 = min(t1 +t2 , min(t1 , t3 )+t4 ), π2 = min(2t1 +t2 , 2 min(t1 , t3 )+
t4 ). Then:
L
(t1 , t2 , t3 , t4 ) = (ea1 , ea1 +a2 , ea1 +2a2 , ea2 )
if and only if
L
(p1 , p2 , p3 , p4 ) = (ea2 , ea1 +2a2 , ea1 +a2 , ea1 )
• In the same fashion one can deduce the exponential identities for types
C2 and G2 .
At this point, it seems very important to mention that in [BBO09], a path
model for Coxeter groups was developed, and exponential laws play a key
role as infinima of a Brownian motion, with appropriate drift. There, once
again, hidden identities in law can be found, involving general Coxeter braid
relations. This goes beyond the crystallographic case we just considered.
This leads to the second open question in section 6.
3.3
Geometric identities
Let a generic geometric random variable with parameter 0 < z < 1 be
denoted by Gz :
∀k ∈ N, P (Gz = k) = z k (1 − z)
∨
∨
Let z ∈ A such that − log z ∈ C. We write zβ = ehβ ,log zi . The
condition − log z ∈ C entails that for every positive coroot β ∨ ∈ Φ+∨ ,
∨
0 < zβ < 1. This allows us to formulate identities in law between geometric
random variables.
20
Proposition 3.5. If i and i′ are reduced words of w0 , then for every µ ∈ C,
the following equality in law holds between independent geometric random
variables:
L
∨ , . . . , G β∨
= Ri,i′ trop G βi,1
G β ∨′ , . . . , G β ∨′
i,m
z
i ,1
z
z
i ,m
z
Proof. The claim is true because Ri,i′ trop is one-to-one from Nm to Nm
and thanks to the elementary fact that:
L
⌊eµ ⌋ = Gz
with z = e−µ .
It is immediate that the same identities as in theorem 3.4 hold between
geometric random variables. As a corollary,
this allows us to prove that
a natural analogue of Γµ (or N∞ W (µ) ) exists at the level of Lusztig’s
canonical basis.
Corollary 3.6. For any
i ∈ R (w0 ) with (β
1 , . . . , βm ) the associated posi∨
independent geometric random
tive root enumeration and Gzβ1∨ , . . . , Gzβm
variables, the random variable defined by
∨
Cz (∞) := Li Gzβ1∨ , . . . , Gzβm
has a distribution on Lusztig’s canonical basis that is independent of the
choice of i.
4
Proof of the conditional representation theorem
Let us state a version of the Matsumoto and Yor relationship between
Brownian motions with opposite drifts, which itself is based on many previous works related to exponential functionals of Brownian motion, including
Dufresne’s identity.
Theorem 4.1 ( Matsumoto-Yor [MY01], theorem 2.2 ). Let B (µ) be a Brownian motion on a Euclidean vector space V ≈ Rn with drift µ and β a linear
form on V such that β(µ) > 0. Denote by sβ the hyperplane reflection with
respect to ker β, by Qy the measure of Brownian motion conditionally on its
exponential functional being equal to y > 0:
Z ∞
(µ)
−β(Bs )
y
e
ds = y
Q := P ·
0
and:
β∨ =
2β
hβ, βi
21
Then:
(sβ µ)
B̂t
(µ)
= Bt
Z
1 t −β(Bs(µ) )
+ log 1 −
e
ds β ∨
y 0
is a Qy -Brownian motion with drift sβ µ = µ − β(µ)β ∨ .
This theorem has a dual version that characterises the reciprocal transform of Brownian motion as a Brownian motion conditioned with respect to
its exponential functional:
Theorem 4.2 ( Matsumoto-Yor [MY01], theorem 2.1). Let W (sβ µ) be a
Brownian motion on V ≈ Rn with drift sβ µ, β(µ) > 0 and:
Z
1 t −β(Ws(sβ µ) )
(sβ µ)
e
ds β ∨
+ log 1 +
Xt = Wt
y 0
Then X is a Brownian motion with drift µ, B (µ) , conditioned to:
Z ∞
(µ)
e−β(Bs ) ds = y.
0
L
If moreover, we pick y as random with y = hβ,βiγ2 ∨ independent from W ,
hβ ,µi
then X is a Brownian motion with drift µ.
Notice that compared to the original formulation, we used a multidimensional setting. The change of sign is simply replaced by a hyperplane
reflection. Let us now focus on proving the conditional representation theorem. In the case of g = xα (ξ) with ξ > 0, the group-theoretic path transform
Tg has a simple expression ([Chh14] Properties 5.19):
Z t
−α(Xs )
e
ds α∨
(Tg X)t = Xt + log 1 + ξ
0
for any continuous path X. The result of Hiroyuki Matsumoto and Yor can
be reformulated as:
Theorem 4.3 (SL2 conditional representation). If g = xα (ξ), ξ > 0,
W (sα µ) a Brownian motion on a with drift sα µ such that α(µ)R > 0 then X :=
∞
Tg (W (sα µ) ) is a Brownian motion with drift µ conditioned to 0 e−α(Xs ) ds =
1
ξ.
L
If moreover we pick ξ as random with ξ = hα,αi
2 γhα∨ ,µi independent from
W then X is a Brownian motion with drift µ.
1 0
1 ξ
.
Remark 4.4. Recall that in the case of SL2 , Θ
= 1
0 1
ξ 1
Hence, theorem 2.12 becomes exactly theorem 4.3 in the particular case of
the group SL2 .
22
It is very surprising and impressive that Matsumoto and Yor fully worked
out the SL2 case without starting from any group-theoretic considerations.
We are now ready to prove the conditional representation theorem.
Proof of theorem 2.12 . We can of course take x0 = 0. Let X = Teθ ge−θ (W (w0 µ) ),
and thanks to the composition property 5.19 in [Chh14] we have:
X = Teθ ge−θ W (w0 µ)
= T
hαi ,αi i
1
1 t
1
2
= T
hαi ,αi i
1
1 t
1
2
xi1
xi1
...xim
hα
im ,αim i
tm
2
◦ ··· ◦ T
xim
hα
W (w0 µ)
im ,αim i
tm
2
W (w0 µ)
We apply inductively theorem 4.3 with Lusztig parameters taken to follow
the right laws, in order to get successive Brownian motions.
The end of proof follows from the deterministic inversion lemma for
Lusztig parameters in [Chh14] theorem 8.16, which we know to be also valid
for an infinite time horizon ([Chh14] subsection 8.5):
= Θ (g)
N∞ Teθ ge−θ W (w0 µ)
Concerning notations, a little precision needs to be made at this point. The
left-invariant flow (Bt (.); t ≥ 0) considered in this paper is a conjugation by
eθ on the one in [Chh14] equation (5.7). Hence the little correction in the
above formula. In the end:
N∞ X 0,(µ) = Θ (g)
concluding the proof.
5
Review of Dufresne’s identity and the relationship proven by Matsumoto and Yor
This section is mainly expository in nature and reviews the probabilistic
results we needed, namely Dufresne’s identity and the proofs of theorems
4.1 4.2.
5.1
Exponential functionals of Brownian motion
An important fact is Dufresne’s identity in law:
Proposition 5.1 ( Dufresne [Duf90] ). If W (µ) is a one dimensional Brownian motion with drift µ > 0, then:
Z ∞
(µ)
L 1
e−2Ws ds =
2γµ
0
23
Quick proof. By time inversion, for any fixed t > 0, the random variable
R t −2W (µ)
(µ) R t
(µ)
2Ws ds. Let (Z ; t > 0) be
s
ds has the same law as e−2Wt
t
0 e
0 e
given by:
Z
t
(µ)
e−Zt := e−2Wt
(µ)
e2Ws ds
0
And, by Ito’s lemma, Zt can be easily checked to be a diffusion process since
it satisfies for t > 0 the SDE:
(µ)
dZt = 2dWt
− eZt dt
Hence it has as infinitesimal generator:
L = 2∂z2 + (2µ − ez )∂z
The sequence Zt converges in law to a unique invariant measure because e−Zt
Rt
(µ)
has the same distribution as 0 e−2Ws ds, which converges almost surely.
R∞
(µ)
This invariant measure will be the law of 0 e−2Ws ds. Therefore, all we
need to do is to prove that the distribution of log 2γµ is an invariant measure
for Zt . This is done easily by checking that the adjoint of L annihilates the
density p(z) of the random variable log 2γµ . We have:
1
1 z
p(z) =
exp µz − e
Γ(µ)2µ
2
Applying L∗ , the adjoint of L:
L∗ = 2∂z2 − ∂z (2µ − ez )
We get:
∗
L p(z) =
2∂z2 p(z)
− 2∂z
1 z
(µ − e )p(z)
2
= 2∂z2 p(z) − 2∂z2 p(z)
=0
5.2
Proofs of the relationship between Brownian motions of
opposing drifts
Let us state a version of the Matsumoto and Yor relationship between Brownian motions with opposite drifts [MY01], which itself is based on many
previous works related to exponential functionals of Brownian motion, including Dufresne’s identity.
First let us start by proving theorem 4.1 using known results on expo(µ)
nential functionals of Brownian motion. Now, let Bt be an n-dimensional
24
Brownian motion with drift µ, FtB its natural filtration, β a linear form such
that β(µ) > 0 and:
Z
t
Nt :=
0
exp(−β(Bs(µ) ))ds
N∞ = lim Nt
t→∞
The law of N∞ comes as a simple corollary of Dufresne’s identity:
Corollary 5.2. One has the identity in law:
Z ∞
L
exp −β(Bs(µ) ) ds =
N∞ :=
0
2
||β||2 γhβ ∨ ,µi
Therefore, the density is:
P (N∞
hβ ∨ ,µi
dn
2
1
2
−hβ ∨ ,µi
∈ dn) =
n
exp −
.
∨
2
2
Γ(hβ , µi)
||β|| n
||β||
n
Proof. Define the real Brownian motion W by β(Bt ) = ||β||Wt for t ≥ 0.
Then:
Z ∞
(µ)
N∞ =
e−β(Bt ) dt
Z0 ∞
e−||β||Wt−β(µ)t dt
=
Z0 ∞ ||β||
L
− √ W − β(µ)tc
c
dt using Brownian scaling for c > 0
e c tc
=
0
Z ∞ ||β||
β(µ)u
1
− √ W −
=
e c u c du using change of variable u = tc
c 0
Z ∞
||β||2
4
−2Wu −4 β(µ)u
||β||2 du by choosing c =
e
=
||β||2 0
4
Z ∞
4
∨
=
e−2(Wu +hβ ,µiu) du .
2
||β|| 0
The result holds using Dufresne’s identity in law. As for the density, for
all f ≥ 0 bounded measurable function, and while writing ν = hβ ∨ , µi, we
have:
E (f (N∞ ))
Z ∞
1
e−t tν dt
=
f(
)
||β||2 t/2 Γ(ν) t
0
ν
Z ∞
2
2
2
2 dn
1
by letting n =
f (n) exp −
=
2
2
2
Γ(ν) 0
||β|| n
||β|| n
||β|| n
||β||2 t
ν
Z ∞
1
dn
2
2
=
n−ν
.
f (n) exp −
2
2
Γ(ν) 0
||β|| n
||β||
n
25
Initial enlargement of the filtration F B using the random variable
N∞
Now, we give a proof of theorem 4.1 following Matsumoto and Yor’s original
presentation. In order to compute the law of N∞ conditionally on FtB , the
following decomposition is essential:
Z t
(µ)
(µ)
e−β(Bs ) ds + e−β(Bt ) Ñ∞ ,
(5.1)
N∞ =
0
with Ñ∞ is a copy of N∞ independent from FtB . Indeed:
Z ∞
(µ)
e−β(Bs ) ds
N∞ =
0
Z t
Z ∞
(µ)
(µ)
(µ)
(µ)
−β(Bs )
−β(Bt )
e
ds + e
e−β(Bs −Bt ) ds
=
0
t
Z t
(µ)
(µ)
e−β(Bs ) ds + e−β(Bt ) Ñ∞ .
=
0
For readability purposes, and because it is not necessary to invoke general filtration enlargement theorems, we will give a complete proof using the
usual tools. Indeed, as proved before the law of N∞ has a (smooth) den∞ ≤y)
with respect to the Lebesgue measure, making possible the
sity dP(Ndy
following computations.
Let Qy = P (·|N∞ = y) be a regular version of the conditional probability,
f ≥ 0 bounded measurable function, and A ∈ FtB . We have:
Z
dP
(N∞ ≤ y) =E QN∞ (A)f (N∞ )
dyf (y)Qy (A)
dy
=E (1A f (N∞ ))
=E 1A E f (N∞ ) |FtB
Z
B
=E 1A f (y)dP(N∞ ∈ dy|Ft )
Z
dP(N∞ ≤ y|FtB )
= dyf (y)E 1A
(Fubini) .
dy
Then:
Qy (A) = E 1A
dP
dy
N∞ ≤ y|FtB
dP
dy
(N∞ ≤ y)
!
.
We conclude that Qy is absolutely continuous with respect to P and that
the Radon-Nikodym derivative on FtB is given by the P-martingale:
(µ)
q(Bt , Nt , y) :=
dQy
dP |FtB
26
=
dP
dy
N∞ ≤ y|FtB
dP
dy
(N∞ ≤ y)
Using the expression for the density of N∞ from corollary 5.2, we get:
(µ)
=
q(Bt , Nt , y)
(µ)
dP
β(Bt ) |F B
Ñ
≤
(y
−
N
)e
∞
t
t
dy
dP
dy
(µ)
=eβ(Bt
Hence:
)
(N∞ ≤ dy)
2
exp −
(µ)
(n − Nt )eβ(Bt
(µ)
β(Bt )
2
||β|| (n−Nt )e
exp − ||β||2 2n n−(1+hβ ∨ ,µi)
)
−(1+hβ ∨ ,µi)
.
(µ)
(µ)
log q(Bt , Nt , n)
= At − hβ
∨
(µ)
, µiβ(Bt )
2e−β(Bt )
−
||β||2 (n − Nt )
where At has a zero quadratic variation. Therefore, the semimartingale
bracket between β(B (µ) ) and log q is:
+
*
(µ)
D
E
2e−β(B. )
(µ)
(µ)
∨
(µ)
β(B ), log q = β(B ), −hβ , µiβ(B. ) −
||β||2 (n − N. )
t
*
+ t
(µ) )
−β(B
.
2
e
= − hβ ∨ , µi||β||2 −
β(B),
2
||β||
n − N.
t
(µ)
Z t −β(Bs )
e
= − 2β(µ) + 2
ds
0 n − Ns
Z t
dNs
= − 2β(µ) + 2
0 n − Ns
Nt
= − 2β(µ) − 2 log 1 −
n
In the end, using the Girsanov theorem ([?] Chapter VIII, theorem 1.4):
B̂t = Bt −
β
hβ(B (µ) ), log qit
||β||2
is a Qy Brownian motion, which completes the proof of theorem 4.1.
Inversion
A natural question is whether we can recover B̂ or B from the other. The
answer is yes and the argument is again due to Matsumoto and Yor [MY01].
The proof of theorem 4.2 follows from 4.1 and the following inversion lemma.
27
Lemma 5.3 (Inversion lemma). Let x and y be V valued paths i.e functions
on R+ . Then
Z
1 t −β(y) ∨
(1) x(t) = y(t) + log(1 +
e
)β
n 0
if and only if
(2)
(
Rt
∀t > 0, 0 e−β(x) < n
y(t) = x(t) + log(1 −
Moreover, in any case:
1
(3) (1 +
n
and if
R∞
0
e−β(y) = ∞ then
Z
t
−β(y)
e
0
R∞
0
1
)(1 −
n
1
n
Rt
0
Z
t
e−β(x) )β ∨
e−β(x) ) = 1
0
e−β(x) = n
Proof. It is immediate to see that (1) and (2) are simultaneously true if and
only if (3) is true. Then all we need to prove is (1) ⇒ (3) and (2) ⇒ (3)
e−β(y)
Rt
(1 + n1 0 e−β(y) )2
"
#t
Z
−1
1 t −β(x)
e
=
Rt
n 0
1 + n1 0 e−β(y) 0
Z
1
1 t −β(x)
=1−
e
Rt
n 0
1 + 1 e−β(y)
(1) ⇒ e−β(x) =
⇒
⇒
n
0
⇒ (3)
and
e−β(x)
Rt
(1 − n1 0 e−β(x) )2
#t
"
Z
1 t −β(y)
1
e
=
Rt
n 0
1 − n1 0 e−β(x) 0
Z
1 t −β(y)
1
=1+
e
Rt
n 0
1 − 1 e−β(x)
(2) ⇒ e−β(y) =
⇒
⇒
n
0
⇒ (3)
Then (3) gives the convergence of
Rt
0
e−β(x) to n, right away.
Now we are ready to prove theorem 4.2:
28
Proof of theorem 4.2: Consider a Brownian motion B (µ) with drift µ conR∞
(µ)
ditioned to 0 e−β(Bs ) ds = n. By the previous filtration enlargement
argument, there is B̂ a Brownian motion in the enlarged filtration, such
that:
Z
1 t −β(Bs(µ) )
(s µ)
(µ)
e
ds β ∨ .
B̂t β = Bt + log 1 −
n 0
Using the inversion lemma:
(µ)
Bt
=
(s µ)
B̂t β
1
+ log 1 +
n
Z
t
(sβ µ)
−β(B̂s
e
0
)
ds β ∨ .
Then the following equalities in law between processes follow:
Z ∞
(µ)
(µ)
e−β(Bs ) ds = n
Bt ; t ≥ 0|
0
Z
1 t −β(B̂s(sβ µ) )
(sβ µ)
∨
+ log 1 +
= B̂t
e
ds β ; t ≥ 0
n 0
Z
1 t −β(Ws(sβ µ) )
(sβ µ)
L
∨
e
ds β ; t ≥ 0
+ log 1 +
= Wt
n 0
= (Xt ; t ≥ 0) .
This ends the proof of the first fact.
R∞
(µ)
The second fact is just a consequence of knowing the law of 0 e−β(Bs ) ds
and the usual disintegration formula. For F continuous functional on the
sample space, we have:
Z ∞
(µ)
−β(Bs )
(µ)
(µ)
ds
e
E F (B. ) = E E F (B. )|
0
6
Some open questions
We end the paper by listing some open questions.
Question 6.1. We noticed in section 3 that the beta-gamma algebra identities in theorem 2.21 for type A2 characterize the gamma random variable
thanks to [Luk55]. Is it true in other types?
We would like to say that if the group theoretic transforms related to
total positivity give independent random variables, then the input is made of
gamma variables.
Question 6.2. In [BBO09], for example in proposition 5.9, a sequence of
mutually independent exponential random variables naturally appears. These
29
depend on a choice of reduced expression, and one can deduce hidden identities in law identical to ours in the crystalligraphic case. However, the general
Coxeter setting of [BBO09] goes beyond our framework, simply because for
a non-crystalligraphic Coxeter group W there is no Lie group for which W
is the Weyl group.
Moreover, the tropical relations that appear there have irrational coefficients, and therefore cannot be the tropicalization of rational subtraction-free
expressions. Indeed, in the proof of theorem 3.12, for the dihedral root system
I(m), transition maps make use of λ = cos 2π
m and Tchebicheff polynomials in λ. λ is indeed rational for the crystallographic values m = 2, 3, 4, 6,
but this is not true in general.
It would be very interesting to gain further insight in the Coxeter case
and being able to explicit such relations. Is it possible to obtain a geometric
lifting to identities between gamma variables?
References
[BBO05] Philippe Biane, Philippe Bougerol, and Neil O’Connell. Littelmann paths and Brownian paths. Duke Math. J., 130(1):127–167,
2005.
[BBO09] Philippe Biane, Philippe Bougerol, and Neil O’Connell. Continuous crystal and Duistermaat-Heckman measure for Coxeter
groups. Adv. Math., 221(5):1522–1583, 2009.
[BO11]
Fabrice Baudoin and Neil O’Connell. Exponential functionals of
Brownian motion and class-one Whittaker functions. Ann. Inst.
Henri Poincaré Probab. Stat., 47(4):1096–1120, 2011.
[BZ97]
Arkady Berenstein and Andrei Zelevinsky. Total positivity in
Schubert varieties. Comment. Math. Helv., 72(1):128–166, 1997.
[Chh13] Reda Chhaibi. Littelmann path model for geometric crystals,
Whittaker functions on Lie groups and Brownian motion. phD
thesis in Université Paris VI, pages 1–226, 2013, arXiv:1302.0902.
[Chh14] Reda Chhaibi. Littelmann path model for geometric crystals.
Preprint v2, pages 1–78, 2014, arXiv:1405.6437.
[CPY98] Philippe Carmona, Frédérique Petit, and Marc Yor. Betagamma random variables and intertwining relations between certain Markov processes. Rev. Mat. Iberoamericana, 14(2):311–367,
1998.
[Duf90]
Daniel Dufresne. The distribution of a perpetuity, with applications to risk theory and pension funding. Scand. Actuar. J.,
(1-2):39–79, 1990.
30
[FZ99]
Sergey Fomin and Andrei Zelevinsky. Double Bruhat cells and
total positivity. J. Amer. Math. Soc., 12(2):335–380, 1999.
[Hum72] James E. Humphreys. Introduction to Lie algebras and representation theory. Springer-Verlag, New York-Berlin, 1972. Graduate
Texts in Mathematics, Vol. 9.
[Hum90] James E. Humphreys. Reflection groups and Coxeter groups, volume 29 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1990.
[Luk55] Eugene Lukacs. A characterization of the gamma distribution.
Ann. Math. Statist., 26:319–324, 1955.
[Lus90a] G. Lusztig. Canonical bases arising from quantized enveloping
algebras. J. Amer. Math. Soc., 3(2):447–498, 1990.
[Lus90b] G. Lusztig. Canonical bases arising from quantized enveloping
algebras. II. Progr. Theoret. Phys. Suppl., (102):175–201 (1991),
1990. Common trends in mathematics and quantum field theories
(Kyoto, 1990).
[Lus94]
G. Lusztig. Total positivity in reductive groups. In Lie theory and
geometry, volume 123 of Progr. Math., pages 531–568. Birkhäuser
Boston, Boston, MA, 1994.
[Lus08]
G. Lusztig. A survey of total positivity. Milan J. Math., 76:125–
134, 2008.
[Mat64] Hideya Matsumoto. Générateurs et relations des groupes de Weyl
généralisés. C. R. Acad. Sci. Paris, 258:3419–3422, 1964.
[MY01]
Hiroyuki Matsumoto and Marc Yor. A relationship between Brownian motions with opposite drifts via certain enlargements of the
Brownian filtration. Osaka J. Math., 38(2):383–398, 2001.
[Spr09]
T. A. Springer. Linear algebraic groups. Modern Birkhäuser Classics. Birkhäuser Boston, Inc., Boston, MA, second edition, 2009.
[Tit69]
Jacques Tits. Le problème des mots dans les groupes de Coxeter. In Symposia Mathematica (INDAM, Rome, 1967/68), Vol.
1, pages 175–185. Academic Press, London, 1969.
[Whi52] A. M. Whitney. A reduction theorem for totally positive matrices.
J. Analyse Math., 2:88–92, 1952.
[Yor12]
Marc Yor. Private communication, 2012.
31
Appendix: Positive root enumerations for rank 2
systems
1
0
β
α1 = −1 , α2 = 1 1
β2
0
−1
β3
121
α1
α1 + α2
α2
212
α2
α1 + α2
α1
Table 1: Positive roots enumerations for type A2
α1 =
β1
1
0
, α2 =
β2
−1
1
β3
β4
1212
α1
α1 + α2
α1 + 2α2
α2
2121
α2
α1 + 2α2
α1 + α2
α1
Table 2: Positive roots enumerations for type B2
α1 =
β1
0
1
, α2 =
β2
2
−1
β3
β4
1212
α1
2α1 + α2
α1 + α2
α2
2121
α2
α1 + α2
2α1 + α2
α1
Table 3: Positive roots enumerations for type C2
0
1
α1 = 1 , α2 = −2 ,
−1
1
β1
β2
β3
β4
β5
β6
121212
α1
3α1 + α2
2α1 + α2
3α1 + 2α2
α1 + α2
α2
212121
α2
α1 + α2
3α1 + 2α2
2α1 + α2
3α1 + α2
α1
Table 4: Positive roots enumerations for type G2
32
| 4math.GR
|
Guarded Dependent Type Theory with
Coinductive Types
arXiv:1601.01586v1 [cs.LO] 7 Jan 2016
Aleš Bizjak1 , Hans Bugge Grathwohl1 , Ranald Clouston1 ,
Rasmus E. Møgelberg2, and Lars Birkedal1
1
Aarhus University
{abizjak,hbugge,ranald.clouston,birkedal}@cs.au.dk
2
IT University of Copenhagen
[email protected]
Abstract. We present guarded dependent type theory, gDTT, an extensional dependent type theory with a ‘later’ modality and clock quantifiers
for programming and proving with guarded recursive and coinductive
types. The later modality is used to ensure the productivity of recursive
definitions in a modular, type based, way. Clock quantifiers are used for
controlled elimination of the later modality and for encoding coinductive
types using guarded recursive types. Key to the development of gDTT
are novel type and term formers involving what we call ‘delayed substitutions’. These generalise the applicative functor rules for the later
modality considered in earlier work, and are crucial for programming
and proving with dependent types. We show soundness of the type theory with respect to a denotational model.
This is the technical report version of a paper to appear in the proceedings
of FoSSaCS 2016.
1
Introduction
Dependent type theory is useful both for programming, and for proving properties of elements of types. Modern implementations of dependent type theories
such as Coq [16], Nuprl [10], Agda [20], and Idris [7], have been used successfully in many projects. However, they offer limited support for programming and
proving with coinductive types.
One of the key challenges is to ensure that functions on coinductive types
are well-defined; that is, productive with unique solutions. Syntactic guarded
recursion [11], as used for example in Coq [12], ensures productivity by requiring
that recursive calls be nested directly under a constructor, but it is well known
that such syntactic checks exclude many valid definitions, particularly in the
presence of higher-order functions.
To address this challenge, a type-based approach to guarded recursion, more
flexible than syntactic checks, was first suggested by Nakano [19]. A new modality, written ⊲ and called ‘later’ [2], allows us to distinguish between data we
have access to now, and data which we will get later. This modality must be
2
used to guard self-reference in type definitions, so for example guarded streams
of natural numbers are described by the guarded recursive equation
StrgN ≃ N × ⊲ StrgN
asserting that stream heads are available now, but tails only later.
Types defined via guarded recursion with ⊲ are not standard coinductive
types, as their denotation is defined via models based on the topos of trees [5].
More pragmatically, the bare addition of ⊲ disallows productive but acausal [15]
functions such as the ‘every other’ function that returns every second element
of a stream. Atkey and McBride proposed clock quantifiers [3] for such functions; these have been extended to dependent types [18,6], and Møgelberg [18,
Thm. 2] has shown that they allow the definition of types whose denotation is
precisely that of standard coinductive types interpreted in set-based semantics.
As such, they allow us to program with real coinductive types, while retaining
productivity guarantees.
In this paper we introduce the extensional guarded dependent type theory
gDTT, which provides a framework where guarded recursion can be used not
just for programming with coinductive types but also for coinductive reasoning.
As types depend on terms, one of the key challenges in designing gDTT is
coping with elements that are only available later, i.e., elements of types of the
form ⊲A. We do this by generalising the applicative functor structure of ⊲ to the
dependent setting. Recall the rules for applicative functors [17]:
Γ ⊢t:A
Γ ⊢ next t : ⊲A
Γ ⊢ f : ⊲(A → B)
Γ ⊢ t : ⊲A
Γ ⊢ f ⊛ t : ⊲B
(1)
The first rule allows us to make later use of data that we have now. The second
allows, for example, functions to be applied recursively to the tails of streams.
Suppose now that f has type ⊲(Πx : A.B), and t has type ⊲A. What should
the type of f ⊛t be? Intuitively, t will eventually reduce to some value next u, and
so the resulting type should be ⊲(B[u/x]), but if t is an open term we may not
be able to perform this reduction. This problem occurs in coinductive reasoning:
if, e.g., A is StrgN , and B a property of streams, in our applications f will be a
(guarded) coinduction assumption that we will want to apply to the tail of a
stream, which has type ⊲ StrgN .
We hence must introduce a new notion, of delayed substitution, similar to
let-binding, allowing us to give f ⊛ t the type
⊲ [x
t] .B
binding x in B. Definitional equality rules then allow us to simplify this type
when t has form next u, i.e., ⊲ [x next u] .B ≡ ⊲(B[u/x]). This construction
generalises to bind a list of variables. Delayed substitution is essential to many
examples, as shown in Sec. 3, and surprisingly the applicative functor termformer ⊛, so central to the standard presentation of applicative functors, turns
out to be definable via delayed substitutions, as shown in Sec. 2.
3
Contributions. The contributions of this paper are:
– We introduce the extensional guarded dependent type theory gDTT, and
show that it gives a framework for programming and proving with guarded
recursive and coinductive types. The key novel feature is the generalisation
of the ‘later’ type-former and ‘next’ term-former via delayed substitutions;
– We prove the soundness of gDTT via a model similar to that used in earlier
work on guarded recursive types and clock quantifiers [18,6].
We focus on the design and soundness of the type theory and restrict attention
to an extensional type theory. We postpone a treatment of an intensional version
of the theory to future work (see Secs. 7 and 8).
In addition to the examples included in this paper, we are pleased to note
that a preliminary version of gDTT has already proved crucial for formalizing a
logical relations adequacy proof of a semantics for PCF using guarded recursive
types by Paviotti et. al. [21].
2
Guarded Dependent Type Theory
gDTT is a type theory with base types unit 1, booleans B, and natural numbers N, along with Π-types, Σ-types, identity types, and universes. For space
reasons we omit all definitions that are standard to such a type theory; see e.g.
Jacobs [14]. Our universes are à la Tarski, so we distinguish between types and
terms, and have terms that represent types; they are called codes of types and
b is the code of the type N.
they can be recognised by their circumflex, e.g., N
We have a map El sending codes of types to their corresponding type. We follow
standard practice and often omit El in examples, except where it is important
to avoid confusion.
We fix a countable set of clock variables CV = {κ1 , κ2 , · · · } and a single clock
constant κ0 , which will be necessary to define, for example, the function hd in
Sec. 5. A clock is either a clock variable or the clock constant; they are intuitively
temporal dimensions on which types may depend. A clock context ∆, ∆′ , · · · is
a finite set of clock variables. We use the judgement ⊢∆ κ to express that either
κ is a clock variable in the set ∆ or κ is the clock constant κ0 . All judgements,
summarised in Fig. 1, are parametrised by clock contexts. Codes of types inhabit
universes U∆ parametrised by clock contexts similarly. The universe U∆ is only
well-formed in clock contexts ∆′ where ∆ ⊆ ∆′ . Intuitively, U∆ contains codes
⊢∆ κ
valid clock
Γ ⊢∆
well-formed context
Γ ⊢∆ A type
well-formed type
Γ ⊢∆ t : A
typing judgment
Γ ⊢∆ A ≡ B
type equality
Γ ⊢∆ t ≡ u : A
term equality
κ
⊢∆ ξ : Γ _ Γ ′
Fig. 1. Judgements in gDTT.
delayed substitution
4
of types that can vary only along dimensions in ∆. We have universe inclusions
from U∆ to U∆′ whenever ∆ ⊆ ∆′ ; in the examples we will not write these
explicitly. Note that we do not have Uc
∆ : U∆′ , i.e., these universes do not form a
hierarchy. We could additionally have an orthogonal hierarchy of universes, i.e.
1
2
for each clock context ∆ a hierarchy of universes U∆
: U∆
: ···.
All judgements are closed under clock weakening and clock substitution. The
former means that if, e.g., Γ ⊢∆ t : A is derivable then, for any clock variable
κ 6∈ ∆, the judgement Γ ⊢∆,κ t : A is also derivable. The latter means that if,
e.g., Γ ⊢∆,κ t : A is derivable and ⊢∆ κ′ then the judgement Γ [κ′ /κ] ⊢∆ t[κ′ /κ] :
A[κ′ /κ] is also derivable, where clock substitution [κ′ /κ] is defined as obvious.
The rules for guarded recursion can be found in Figs. 2 and 3; rules for
coinductive types are postponed until Sec. 4. Recall the ‘later’ type former ⊲,
which expresses that something will be available at a later time. In gDTT we
κ
have ⊲ for each clock κ, so we can delay a type along different dimensions. As
discussed in the introduction, we generalise the applicative functor structure of
κ
each ⊲ via delayed substitutions, which allow a substitution to be delayed until
its substituent is available. We showed in the introduction how a type with a
κ
single delayed substitution ⊲ [x t] .A should work. However if we have a term
κ
f with more than one argument, for example of type ⊲ (Π(x : A).Π(y : B).C),
and wish to type an application f κ t κ u (where κ is the applicative functor
operation ⊛ for clock κ) we may have neither t nor u available now, and so
κ
we need sequences of delayed substitutions to define the type ⊲ [x t, y u] .C.
Our concrete examples of Sec. 3 will show that this issue arises in practice. We
therefore define sequences of delayed substitutions ξ. The new raw types, terms,
and delayed substitutions of gDTT are given by the grammar
κ
A, B ::= · · · | ⊲ξ.A
κ
κ
t, u ::= · · · | nextκ ξ.t | b
⊲ t
ξ ::= · | ξ [x
t] .
Note that we just write ⊲A where its delayed substitution is the empty ·, and
κ
that ⊲ξ.A binds the variables substituted for by ξ in A, and similarly for next.
The three rules DS-Emp, DS-Cons, and Tf-⊲ are used to construct the type
κ
⊲ξ.A. These rules formulate how to generalise these types to arbitrarily long delayed substitutions. Once the type formation rule is established, the introduction
rule Ty-Next is the natural one.
With delayed substitutions we can define κ as
g f
.g x.
f κ t , nextκ
x t
Using the rules in Fig. 2 we can derive the following typing judgement for
κ
Γ ⊢∆ f : ⊲ξ.Π(x : A).B
Γ ⊢∆ f
κ
κ
t : ⊲ξ[x
κ
κ
Γ ⊢∆ t : ⊲ξ.A
t].B
Ty-⊛
When a term has the form nextκ ξ [x nextκ ξ.u] .t, then we have enough
information to perform the substitution in both the term and its type. The rule
5
TmEq-Force applies the substitution by equating the term with the result of
an actual substitution, nextκ ξ.t[u/x]. The rule TyEq-Force does the same for
its type. Using TmEq-Force we can derive the basic term equality
(nextκ ξ.f )
κ
(nextκ ξ.t) ≡ nextκ ξ.(f t).
typical of applicative functors [17].
It will often be the case that a delayed substitution is unnecessary, because
the variable to be substituted for does not occur free in the type/term. This is
what TyEq-⊲-Weak and TmEq-Next-Weak express, and with these we can
justify the simpler typing rule
κ
κ
Γ ⊢∆ f : ⊲ξ.(A → B)
Γ ⊢∆ f
κ
Γ ⊢∆ t : ⊲ξ.A
κ
t : ⊲ξ.B
In other words, delayed substitutions on the type are not necessary when we
apply a non-dependent function.
Further, we have the applicative functor identity law
(nextκ ξ.λx.x)
κ
t ≡ t.
This follows from the rule TmEq-Next-Var, which allows us to simplify a term
nextκ ξ [y t] .y to t.
Sometimes it is necessary to switch the order in the delayed substitution.
Two substitutions can switch places, as long as they do not depend on each
other; this is what TyEq-⊲-Exch and TmEq-Next-Exch express.
Rule TmEq-Next-Comm is not used in the examples of this paper, but it
implies the rule nextκ ξ [x t] . nextκ x ≡ nextκ t, which is needed in Paviotti’s
PhD work.
2.1
Fixed points and guarded recursive types
In gDTT we have for each clock κ valid in the current clock context a fixed-point
combinator fixκ . This differs from a traditional fixed-point combinator in that
the type of the recursion variable is not the same as the result type; instead its
κ
type is guarded with ⊲. When we define a term using the fixed-point, we say that
it is defined by guarded recursion. When the term is intuitively a proof, we say
we are proving by Löb induction [2].
Guarded recursive types are defined as fixed-points of suitably guarded functions on universes. This is the approach of Birkedal and Møgelberg [4], but the
generality of the rules of gDTT allows us to define more interesting dependent
guarded recursive types, for example the predicates of Sec. 3.
We first illustrate the technique by defining the (non-dependent) type of
guarded streams. Recall from the introduction that we want the type of guarded
κ
streams, for clock κ, to satisfy the equation StrκA ≡ A × ⊲ StrκA .
The type A will be equal to El(B) for some code B in some universe U∆
κ
where the clock variable κ is not in ∆. We then define the code SA
of StrκA in the
6
Universes
∆′ ⊆ ∆
Γ ⊢∆
Γ ⊢∆ U∆′ type
Γ ⊢∆ A : U∆′
Univ
Γ ⊢∆ El(A) type
El
Delayed substitutions:
Γ ⊢∆
⊢∆ κ
κ
⊢∆ · : Γ _ ·
κ
⊢∆ ξ : Γ _ Γ ′
DS-Emp
⊢∆ ξ [x
κ
Γ ⊢∆ t : ⊲ξ.A
κ
t] : Γ _ Γ ′ , x : A
DS-Cons
Typing rules:
κ
Γ, Γ ′ ⊢∆ A type
⊢∆ ξ : Γ _ Γ ′
κ
Γ ⊢∆ ⊲ξ.A type
Γ, Γ ′ ⊢∆ t : A
Tf-⊲
κ
⊢∆ ξ : Γ _ Γ ′
κ
Γ ⊢∆ nextκ ξ.t : ⊲ξ.A
Ty-Next
κ
⊢∆′ κ
Γ ⊢∆ A : ⊲U∆′
Ty-b
⊲
Γ ⊢∆ b
⊲κ A : U∆′
⊢∆ κ
κ
Γ, x : ⊲A ⊢∆ t : A
Γ ⊢∆ fixκ x.t : A
Ty-Fix
Fig. 2. Overview of the new typing rules involving ⊲ and delayed substitutions.
κ
κ
bb
b is the code of the (simple)
universe U∆,κ to be SA
, fixκ X.B ×
⊲ X, where ×
κ
product type. Via the rules of gDTT we can show StrκA ≃ A × ⊲ StrκA as desired.
κ
The head and tail operations, hdκ : StrκA → A and tlκ : StrκA → ⊲ StrκA are
simply the first and the second projections. Conversely, we construct streams by
pairing. We use the suggestive consκ notation which we define as
κ
κ
consκ , λ (a : A) as : ⊲ StrκA . ha, asi
consκ : A → ⊲ StrκA → StrκA
Defining guarded streams is also done via guarded recursion, for example the
stream consisting only of ones is defined as ones , fixκ x. consκ 1 x.
The rule TyEq-El-⊲ is essential for defining guarded recursive types as fixedpoints on universes, and it can also be used for defining more advanced guarded
recursive dependent types such as covectors; see Sec. 3.
2.2
Identity types
gDTT has standard extensional identity types IdA(t, u) (see, e.g., Jacobs [14])
but with two additional type equivalences necessary for working with guarded
dependent types. We write rA t for the reflexivity proof IdA(t, t). The first type
equivalence is the rule TyEq-⊲. This rule, which is validated by the model of
Sec. 6, may be thought of by analogy to type equivalences often considered in
homotopy type theory [24], such as
IdA×B(hs1 , s2 i , ht1 , t2 i) ≡ IdA(s1 , t1 ) × IdB(s2 , t2 ).
(2)
7
Definitional type equalities:
κ
κ
⊲ξ [x
κ
⊲ξ [x
t, y
κ
t] .A ≡ ⊲ξ.A
κ
u] ξ ′ .A ≡ ⊲ξ [y
u, x
t] ξ ′ .A
κ
κ
⊲ξ [x
(TyEq-⊲-Weak)
next ξ.t] .A ≡ ⊲ξ.A[t/x]
κ
κ
(TyEq-Force)
κ
El(b
⊲ (next ξ.t)) ≡ ⊲ξ. El(t)
κ
κ
(TyEq-⊲-Exch)
(TyEq-El-⊲)
κ
Idκ⊲ξ.A(next ξ.t, next ξ.s) ≡ ⊲ξ.IdA(t, s)
(TyEq-⊲)
Definitional term equalities:
nextκ ξ [x
κ
next ξ [x
nextκ ξ [x
t, y
t] .u ≡ nextκ ξ.u
(TmEq-Next-Weak)
t] .x ≡ t
u] ξ ′ .v ≡ nextκ ξ [y
(TmEq-Next-Var)
u, x
t] ξ ′ .v
nextκ ξ. nextκ ξ ′ .u ≡ nextκ ξ ′ . nextκ ξ.u
κ
next ξ [x
nextκ ξ.t] .u ≡ nextκ ξ.u[t/x]
fixκ x.t ≡ t[nextκ (fixκ x.t) /x]
(TmEq-Next-Exch)
(TmEq-Next-Comm)
(TmEq-Force)
(TmEq-Fix)
Fig. 3. New type and term equalities in gDTT. Rules TyEq-⊲-Weak and TmEqNext-Weak require that A and u are well-formed in a context without x. Rules
TyEq-⊲-Exch and TmEq-Next-Exch assume that exchanging x and y is allowed,
i.e., that the type of x does not depend on y and vice versa. Likewise, rule TmEqNext-Comm assumes that exchanging the codomains of ξ and ξ ′ is allowed and that
none of the variables in the codomains of ξ and ξ ′ appear in the type of u.
There are two important differences. The first is that (2) is (using univalence) a
propositional type equality, whereas TyEq-⊲ specifices a definitional type equality. This is natural in an extensional type theory. The second difference is that
there are terms going in both directions in (2), whereas we would have a term
κ
of type Idκ⊲ξ.A(nextκ ξ.t, nextκ ξ.u) → ⊲ξ.IdA(t, u) without the rule TyEq-⊲.
The second novel type equality rule, which involves clock quantification, will
be presented in Sec. 4.
3
Examples
In this section we present some example terms typable in gDTT. Our examples will use a term, which we call pη, of type Π(s, t : A × B).IdA(π1 t, π1 s) →
IdB(π2 t, π2 s) → IdA×B(t, s). This term is definable in any type theory with a
strong (dependent) elimination rule for dependent sums. The second property
κ
we will use is that StrκA ≡ A × ⊲ StrκA . Because hdκ and tlκ are simply first
and second projections, pη also has type Π (xs, ys : StrκA ) .IdA(hdκ xs, hdκ ys) →
Idκ⊲ Strκ (tlκ xs, tlκ ys) → IdStrκA(xs, ys).
A
8
zipWithκ preserves commutativity. In gDTT we define the zipWithκ function
which has the type (A → B → C) → StrκA → StrκB → StrκC by
zipWithκ f , fixκ φ.λxs, ys. consκ (f (hdκ xs) (hdκ ys)) (φ
κ
tlκ xs
κ
tlκ ys) .
We show that commutativity of f implies commutativity of zipWithκ f , i.e., that
Π(f : A → A → B). (Π(x, y : A).IdB(f x y, f y x)) →
Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs)
is inhabited. The term that inhabits this type is
λf.λc. fixκ φ.λxs, ys. pη (c (hdκ xs) (hdκ ys)) (φ
κ
tlκ xs
κ
tlκ ys) .
κ
Here, φ has type ⊲(Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs)) so
to type the term above, we crucially need delayed substitutions.
An example with covectors. The next example is more sophisticated, as it involves programming and proving with a data type that, unlike streams, is dependently typed. Indeed the generalised later, carrying a delayed substitution,
is necessary to type even elementary programs. Covectors are the potentially
infinite version of vectors (lists with length). To define guarded covectors we
first
co-natural numbers. The definition in gDTT is CoNκ ,
need guarded
κ
κ
κ
b+
bb
El fix X.(1
⊲ X) ; this type satisfies CoNκ ≡ 1 + ⊲ CoNκ . Using CoNκ we
\κ n), where
can define the type family of covectors CoVecκA n , El(CoVec
A
κ
\κ , fixκ φ : ⊲(CoNκ → U∆,κ ) .λ(n : CoNκ ). case n of
CoVec
A
b
inl u ⇒ 1
κ
bb
inr m ⇒ A ×
⊲ (φ
κ
m).
\κ . As an example of covectors,
We will not distinguish between CoVecκA and CoVec
A
κ
κ
we define ones of type Π(n : CoN ). CoVecN n which produces a covector of any
length consisting only of ones:
ones , fixκ φ.λ(n : CoNκ ). case n of {inl u ⇒ inl hi; inr m ⇒ h1, φ
κ
mi} .
Although this is one of the simplest covector programs one can imagine, it does
not type-check without the generalised later with delayed substitutions.
The map function on covectors is defined as
map : (A → B) → Π(n : CoNκ ). CoVecκA n → CoVecκB n
map f , fixκ φ.λ(n : CoNκ ). case n of
inl u ⇒ λ(x : 1).x
κ
inr m ⇒ λ p : A × ⊲ [n
m] .(CoVecκA n) . hf (π1 p) , φ
κ
m
κ
(π2 p)i .
9
It preserves composition: the following type is inhabited
Π(f : A → B)(g : B → C)(n : CoNκ )(xs : CoVecκA n).
IdCoVecκC n(map g n (map f n xs), map (g ◦ f ) n xs)
by the term
λ(f : A → B)(g : B → C). fixκ φ.λ(n : CoNκ ). case n of
inl u ⇒ λ(xs : 1).r1 xs
inr m ⇒ λ(xs : CoVecκA (inr m)). pη (rC g(f (π1 xs))) (φ
4
κ
m
κ
π2 xs) .
Coinductive types
As discussed in the introduction, guarded recursive types on their own disallow
productive but acausal function definitions. To capture such functions we need
κ
to be able to remove ⊲. However such eliminations must be controlled to avoid
κ
κ
trivialising ⊲. If we had an unrestricted elimination term elim : ⊲A → A every
type would be inhabited via fixκ , making the type theory inconsistent.
κ
However, we may eliminate ⊲ provided that the term does not depend on the
clock κ, i.e., the term is typeable in a context where κ does not appear. Intuitively, such contexts have no temporal properties along the κ dimension, so we
may progress the computation without violating guardedness. Fig. 4 extends the
system of Fig. 2 to allow the removal of clocks in such a setting, by introducing
clock quantifiers ∀κ [3,18,6]. This is a binding construct with associated term
constructor Λκ, which also binds κ. The elimination term is clock application.
Application of the term t of type ∀κ.A to a clock κ is written as t[κ]. One may
think of ∀κ.A as analogous to the type ∀α.A in polymorphic lambda calculus;
indeed the basic rules are precisely the same, but we have an additional construct
κ
prev κ.t, called ‘previous’, to allow removal of the later modality ⊲.
Typing this new construct prev κ.t is somewhat complicated, as it requires
‘advancing’ a delayed substitution, which turns it into a context morphism (an
actual substitution); see Fig. 5 for the definition. The judgement ρ :∆ Γ → Γ ′
expresses that ρ is a context morphism from context Γ ⊢∆ to the context Γ ′ ⊢∆ .
We use the notation ρ[t/x] for extending the context morphism by mapping the
variable x to the term t. We illustrate this with two concrete examples.
First, we can indeed remove later under a clock quantier:
κ
force : ∀κ.⊲A → ∀κ.A
force , λx. prev κ.x[κ] .
κ
The type is correct because advancing the empty delayed substitution in ⊲ turns
it into the identity substitution ι, and Aι ≡ A. The β and η rules ensure that
κ
force is the inverse to the canonical term λx.Λκ. nextκ x[κ] of type ∀κ.A → ∀κ.⊲A.
Second, we may see an example with a non-empty delayed substitution in the
term prev κ. nextκ λn. succ n κ nextκ 0 of type ∀κ.N. Recall that κ is syntactic
10
Γ ⊢∆
Γ ⊢∆,κ A type
Γ ⊢∆ ∀κ.A type
Γ ⊢∆
Tf-∀
Γ ⊢∆,κ t : A
Γ ⊢∆ Λκ.t : ∀κ.A
Γ ⊢∆
∆′ ⊆ ∆
Ty-Λ
Γ ⊢∆ t : ∀κ.U∆′ ,κ
Ty-∀-code
b t : U∆′
Γ ⊢∆ ∀
⊢∆ κ′
Γ ⊢∆ t : ∀κ.A
Ty-app
Γ ⊢∆ t κ′ : A[κ′ /κ]
κ
Γ ⊢∆,κ t : ⊲ξ.A
Γ ⊢∆ prev κ.t : ∀κ.(A(advκ∆ (ξ)))
Ty-prev
Fig. 4. Overview of the new typing rules for coinductive types.
sugar and so more precisely the term is
nextκ λn. succ n
κ f
.f x.
prev κ. next
x nextκ 0
(3)
Advancing the delayed substitution turns it into the substitution mapping the
variable f to the term (prev κ. nextκ λn. succ n)[κ] and the variable x to the term
(prev κ. nextκ 0)[κ]. Using the β rule for prev, then the β rule for ∀κ, this simplifies
to the substitution mapping f to λn. succ n and x to 0. With this we have that
the term (3) is equal to Λκ. ((λn. succ n) 0) which is in turn equal to Λκ.1.
An important property of the term prev κ.t is that κ is bound in t; hence
prev κ.t has type ∀κ.A instead of just A. This ensures that substitution of terms
in types and terms is well-behaved and we do not need the explicit substitutions
used, for example, by Clouston et al. [8] where the unary type-former was
used in place of clocks. This binding structure ensures, for instance, that the
introduction rule Ty-Λ closed under substitution in Γ .
The rule TmEq-∀-fresh states that if t has type ∀κ.A and the clock κ does
not appear in the type A, then it does not matter to which clock t is applied,
as the resulting term will be the same. In the polymorphic lambda calculus, the
corresponding rule for universal quantification over types would be a consequence
of relational parametricity.
We further have the construct b
∀ and the rule Ty-∀-code which witness that
the universes are closed under ∀κ.
To summarise, the new raw types and terms, extending those of Sec. 2, are
A, B ::= · · · | ∀κ.A
t, u ::= · · · | Λκ.t | t[κ] | b
∀ t | prev κ.t
Finally, we have the equality rule TyEq-∀-Id analogous to the rule TyEq⊲. Note that, as in Sec. 2.2, there is a canonical term of type Id∀κ.A(t, s) →
∀κ.IdA(t[κ] , s[κ]) but, without this rule, no term in the reverse direction.
4.1
Derivable type isomorphisms
The encoding of coinductive types using guarded recursive types crucially uses
a family of type isomorphisms commuting ∀κ over other type formers [3,18]. By
11
κ
⊢∆,κ · : Γ _ ·
advκ∆ (·)
⊢∆,κ ξ[x
advκ∆ (ξ[x
Γ ⊢∆
, ι :∆,κ Γ → Γ
κ
t] : Γ _ Γ ′ , x : A
Γ ⊢∆
t]) , advκ∆ (ξ)[(prev κ.t)[κ] /x] :∆,κ Γ → Γ, Γ ′ , x : A
Fig. 5. Advancing a delayed substitution.
a type isomorphism A ∼
= B we mean two well-typed terms f and g of types
f : A → B and g : B → A such that f (g x) ≡ x and g(f x) ≡ x. The first
type isomorphism is ∀κ.A ∼
= A whenever κ is not free in A. The terms g =
λx.Λκ.x of type A → ∀κ.A and f = λx.x[κ0 ] of type A → ∀κ.A witness the
isomorphism. Note that we used the clock constant κ0 in an essential way. The
equality f (g x) ≡ x follows using only the β rule for clock application. The
equality g(f x) ≡ x follows using by the rule TmEq-∀-fresh.
The following type isomorphisms follow by using β and η laws for the constructs involved.
- If κ 6∈ A then ∀κ.Π(x : A).B ∼
= Π(x : A).∀κ.B.
Σ
(y
:
∀κ.A)
(∀κ.B [y[κ]/ x]).
- ∀κ.Σ (x : A) B ∼
=
κ
∼
- ∀κ.A = ∀κ.⊲A.
There is an important additional type isomorphism witnessing that ∀κ commutes with binary sums; however unlike the isomorphisms above we require
equality reflection to show that the two functions are inverse to each other up to
definitional equality. There is a canonical term of type ∀κ.A+∀κ.B → ∀κ.(A+B)
using just ordinary elimination of coproducts. Using the fact that we encode binary coproducts using Σ-types and universes we can define a term com+ of type
∀κ.(A+B) → ∀κ.A+∀κ.B which is a inverse to the canonical term. In particular
com+ satisfies the following two equalities which will be used below.
com+ (Λκ. inl t) ≡ inl Λκ.t
5
com+ (Λκ. inr t) ≡ inr Λκ.t.
(4)
Example programs with coinductive types
b in clock context ∆ and κ a fresh clock variable.
Let A be a type with code A
κ
Let StrA = ∀κ. StrA . We can define head, tail and cons functions
hd : StrA → A
tl : StrA → StrA
cons : A → StrA → StrA
hd , λxs. hdκ0 (xs[κ0 ])
tl , λxs. prev κ. tlκ (xs[κ])
cons , λx.λxs.Λκ. consκ x (nextκ (xs[κ])) .
12
Definitional type equalities:
Γ ⊢∆
Γ ⊢∆
∆′ ⊆ ∆
Γ ⊢∆,κ t : U∆′ ,κ
TyEq-∀-el
b
Γ ⊢∆ El(∀ Λκ.t) ≡ ∀κ. El(t)
Γ ⊢∆,κ A type
Γ ⊢∆ t : ∀κ.A
Γ ⊢∆ s : ∀κ.A
Γ ⊢∆ ∀κ.IdA(t[κ] , s[κ]) ≡ Id∀κ.A(t, s)
TyEq-∀-Id
Definitional term equalities:
⊢∆ κ′
Γ ⊢∆,κ t : A
′
TmEq-∀-β
(Λκ.t) κ ≡ t[κ′ /κ] : A[κ′ /κ]
Γ ⊢∆
Γ ⊢∆
κ 6∈ ∆
Γ ⊢∆ A type
Γ ⊢∆
Γ ⊢∆
Γ ⊢∆ t : ∀κ.A
t κ′ ≡ t κ′′ : A
κ
⊢∆,κ ξ : Γ _ Γ ′
κ 6∈ ∆
Γ ⊢∆ t : ∀κ.A
Γ ⊢∆ Λκ.t[κ] ≡ t : ∀κ.A
⊢∆ κ′
⊢∆ κ′′
Γ, Γ ′ ⊢∆,κ t : A
Γ ⊢∆ prev κ. nextκ ξ.t ≡ Λκ.t(advκ∆ (ξ)) : ∀κ.(A(advκ∆ (ξ)))
Γ ⊢∆
TmEq-∀-η
TmEq-∀-fresh
TmEq-prev-β
κ
Γ ⊢∆,κ t : ⊲A
κ
Γ ⊢∆,κ nextκ ((prev κ.t)[κ]) ≡ t : ⊲A
TmEq-prev-η
Fig. 6. Type and term equalities involving clock quantification.
With these we can define the acausal ‘every other’ function eoκ that removes
every second element of the input stream. It is acausal because the second element of the output stream is the third element of the input. Therefore to type
the function we need to have the input stream always available, so clock quantification must be used. The function eoκ of type StrA → StrκA is defined as
eoκ , fixκ φ.λ (xs : StrA ) . consκ (hd xs) (φ
κ
nextκ ((tl (tl xs)))) .
The result is a guarded stream, but we can easily strengthen it and define eo of
type StrA → StrA as eo , λxs.Λκ. eoκ xs.
We can also work with covectors (not just guarded covectors as in Sec. 3).
This is a dependent coinductive type indexed by conatural numbers which is
the type CoN = ∀κ. CoNκ . It is easy to define 0 and succ as 0 , Λκ. inl hi and
succ , λn.Λκ. inr (nextκ (n[κ])). Next, we can define a transport function comCoN
of type comCoN : CoN → 1 + CoN satisfying
comCoN 0 ≡ inl hi
comCoN (succ n) ≡ inr n.
(5)
This function is used to define the type family of covectors as CoVecA n ,
∀κ. CoVecκA n where CoVecκA : CoN → U∆,κ is the term
n
o
b ⊲κ (φ κ (nextκ n)) .
fixκ φ.λ (n : CoN) . case comCoN n of inl ⇒ b
1; inr n ⇒ A×b
13
Using term equalities (4) and (5) we can derive the type isomorphisms
CoVecA 0 ≡ ∀κ.1 ∼
=1
κ
CoVecA (succ n) ≡ ∀κ. A × ⊲ (CoVecκA n) ∼
= A × CoVecA n
(6)
which are the expected properties of the type of covectors.
A simple function we can define is the tail function
tl : CoVecA (succ n) → CoVecA
tl , λv. prev κ.π2 (v[κ]) .
Note that (6) is needed to type tl. The map function of type
map : (A → B) → Π(n : CoN). CoVecA n → CoVecB n
is defined as map f , λn.λxs.Λκ. mapκ f n (xs[κ]) where mapκ is
mapκ : (A → B) → Π(n : CoN). CoVecκA n → CoVecκB n
mapκ = λf. fixκ φ.λn. case comCoN n of
inl ⇒ λv.v
inr n ⇒ λv. hf (π1 v), φ
5.1
κ
(nextκ n)
κ
π2 (v)i .
Lifting guarded functions
In this section we show how in general we may lift a function on guarded recursive
types, such as addition of guarded streams, to a function on coinductive streams.
Moreover, we show how to lift proofs of properties, such as the commutativity
of addition, from guarded recursive types to coinductive types.
Let Γ be a context in clock context ∆ and κ a fresh clock. Suppose A and
B are types such that Γ ⊢∆,κ A type and Γ, x : A ⊢∆,κ B type. Finally let f be
a function of type Γ ⊢∆,κ f : Π(x : A).B. We define L(f ) satisfying the typing
judgement Γ ⊢∆ L(f ) : Π(y : ∀κ.A).∀κ. (B [y[κ] /x]) as L(f ) , λy.Λκ.f (y[κ]).
Now assume that f ′ is another term of type Π(x : A).B (in the same context)
and that we have proved Γ ⊢∆,κ p : Π(x : A).IdB(f x, f ′ x). As above we can
give the term L(p) the type Π(y : ∀κ.A).∀κ.IdB[y[κ]/x](f (y[κ]), f ′ (y[κ])). which
by using the type equality TyEq-∀-Id and the η rule for ∀ is equal to the type
Π(y : ∀κ.A).Id∀κ.B[y[κ]/x](L(f ) y, L(f ′ ) y). So we have derived a property of lifted
functions L(f ) and L(f ′ ) from the properties of the guarded versions f and
f ′ . This is a standard pattern. Using Löb induction we prove a property of a
function whose result is a “guarded” type and derive the property for the lifted
function.
For example we can lift the zipWith function from guarded streams to coinductive streams and prove that it preserves commutativity, using the result on
guarded streams of Sec. 3.
14
6
Soundness
gDTT can be shown to be sound with respect to a denotational model interpreting the type theory. The model is a refinement of Bizjak and Møgelberg’s [6] but
for reasons of space we leave the description of a full model of gDTT for future
work. Instead, to provide some intuition for the semantics of delayed substitutions, we just describe how to interpret the rule
x : A ⊢ B type
⊢ t : ⊲A
⊢ ⊲ [x t] .B type
(7)
in the case where we only have one clock available.
The subsystem of gDTT with only one clock can be modelled in the category
S, known as the topos of trees [5], the presheaf category over the first infinite
ordinal ω. The objects X of S are families of sets X1 , X2 , . . . indexed by the
positive integers, together with families of restriction functions riX : Xi+1 → Xi
indexed similarly. There is a functor ◮ : S → S which maps an object X to the
object
1
!
X1
r1X
X2
r2X
X3
···
where ! is the unique map into the terminal object.
In this model, a closed type A is interpreted as an object of S and the
type x : A ⊢ B type is interpreted as an indexed family of sets Bi (a), for a
in Ai together with maps riB (a) : Bi+1 (a) → Bi (riA (a)). The term t in (7) is
interpreted as a morphism t : 1 → ⊲A so ti (∗) is an element of Ai (here we write
∗ for the element of 1).
The type ⊢ ⊲ [x t] .B type is then interpreted as the object X, defined by
X1 = 1
Xi+1 = Bi (ti+1 (∗)).
Notice that the delayed substitution is interpreted by substitution (reindexing)
in the model; the change of the index in the model (Bi is reindexed along ti+1 (∗))
corresponds to the delayed substitution in the type theory. Further notice that
if B does not depend on x, then the interpretation of ⊢ ⊲ [x t] .B type reduces
to the interpretation ⊲B, which is defined to be ◮ applied to the interpretation
of B.
The above can be generalised to work for general contexts and sequences of
delayed substitutions, and one can then validate that the definitional equality
rules do indeed hold in this model.
7
Related Work
Birkedal et al. [5] introduced dependent type theory with the ⊲ modality, with
semantics in the topos of trees. The guardedness requirement was expressed
using the syntactic check that every occurrence of a type variable lies beneath
15
a ⊲. This requirement was subsequently refined by Birkedal and Møgelberg [4],
who showed that guarded recursive types could be constructed via fixed-points
of functions on universes. However, the rules considered in these papers do not
allow one to apply terms of type ⊲(Π(x : A).B), as the applicative functor
construction ⊛ was defined only for simple function spaces. They are therefore
less expressive for both programming (consider the covector ones, and function
map, of Sec. 3) and proving, noting the extensive use of delayed substitutions in
our example proofs. They further do not consider coinductive types, and so are
restricted to causal functions.
The extension to coinductive types, and hence acausal functions, is due to
Atkey and McBride [3], who introduced clock quantifiers into a simply typed
setting with guarded recursion. Møgelberg [18] extended this work to dependent
types and Bizjak and Møgelberg [6] refined the model further to allow clock
synchronisation.
Clouston et al. [8] introduced the logic Lgλ to prove properties of terms of
the (simply typed) guarded λ-calculus, gλ. This allowed proofs about coinductive
types, but not in the integrated fashion supported by dependent type theories.
Moreover it relied on types being “total”, a property that in a dependently
typed setting would entail a strong elimination rule for ⊲, which would lead to
inconsistency.
Sized types [13] have been combined with copatterns [1] as an alternative
type-based approach for modular programming with coinductive types. This
work is more mature than ours with respect to implementation and the demonstration of syntactic properties such as normalisation, and so further development of gDTT is essential to enable proper comparison. One advantage of gDTT
is that the later modality is useful for examples beyond coinduction, and beyond
the utility of sized types, such as the guarded recursive domain equations used
to model program logics [23].
8
Conclusion and Future Work
We have described the dependent type theory gDTT. The examples we have
detailed show that gDTT provides a setting for programming and proving with
guarded recursive and coinductive types.
In future work we plan to investigate an intensional version of the type theory
and construct a prototype implementation to allow us to experiment with larger
examples. Preliminary work has suggested that the path type of cubical type
theory [9] interacts better with the new constructs of gDTT than the ordinary
Martin-Löf identity type.
Finally, we are investigating whether the generalisation of applicative functors [17] to apply over dependent function spaces, via delayed substitutions,
might also apply to examples quite unconnected to the later modality.
Acknowledgements. This research was supported in part by the ModuRes Sapere
Aude Advanced Grant and DFF-Research Project 1 Grant no. 4002-00442, both
16
from The Danish Council for Independent Research for the Natural Sciences
(FNU). Aleš Bizjak was supported in part by a Microsoft Research PhD grant.
References
1. Abel, A., Pientka, B.: Wellfounded recursion with copatterns: A unified approach
to termination and productivity. In: ICFP. pp. 185–196 (2013)
2. Appel, A.W., Melliès, P.A., Richards, C.D., Vouillon, J.: A very modal model of a
modern, major, general type system. In: POPL. pp. 109–122 (2007)
3. Atkey, R., McBride, C.: Productive coprogramming with guarded recursion. In:
ICFP. pp. 197–208 (2013)
4. Birkedal, L., Møgelberg, R.E.: Intensional type theory with guarded recursive types
qua fixed points on universes. In: LICS. pp. 213–222 (2013)
5. Birkedal, L., Møgelberg, R.E., Schwinghammer, J., Støvring, K.: First steps in
synthetic guarded domain theory: step-indexing in the topos of trees. LMCS 8(4)
(2012)
6. Bizjak, A., Møgelberg, R.E.: A model of guarded recursion with clock synchronisation. In: MFPS (2015)
7. Brady, E.: Idris, a general-purpose dependently typed programming language: Design and implementation. J. Funct. Programming 23(5), 552–593 (2013)
8. Clouston, R., Bizjak, A., Grathwohl, H.B., Birkedal, L.: Programming and reasoning with guarded recursion for coinductive types. In: FoSSaCS (2015)
9. Cohen, C., Coquand, T., Huber, S., Mörtberg, A.: Cubical type theory: a constructive interpretation of the univalence axiom (2015), unpublished
10. Constable, R.L., Allen, S.F., Bromley, H.M., Cleaveland, W.R., Cremer, J.F.,
Harper, R.W., Howe, D.J., Knoblock, T.B., Mendler, N.P., Panangaden, P., Sasaki,
J.T., Smith, S.F.: Implementing Mathematics with the Nuprl Proof Development
System. Prentice-Hall, Inc., Upper Saddle River, NJ, USA (1986)
11. Coquand, T.: Infinite objects in type theory. In: TYPES. pp. 62–78 (1993)
12. Giménez, E.: Codifying guarded definitions with recursive schemes. In: TYPES.
pp. 39–59 (1995)
13. Hughes, J., Pareto, L., Sabry, A.: Proving the correctness of reactive systems using
sized types. In: POPL. pp. 410–423 (1996)
14. Jacobs, B.: Categorical Logic and Type Theory. No. 141 in Studies in Logic and
the Foundations of Mathematics, North Holland, Amsterdam (1999)
15. Krishnaswami, N.R., Benton, N.: Ultrametric semantics of reactive programs. In:
LICS. pp. 257–266 (2011)
16. The Coq development team: The Coq proof assistant reference manual. LogiCal
Project (2004), http://coq.inria.fr, version 8.0
17. McBride, C., Paterson, R.: Applicative programming with effects. J. Funct. Programming 18(1), 1–13 (2008)
18. Møgelberg, R.E.: A type theory for productive coprogramming via guarded recursion. In: CSL-LICS (2014)
19. Nakano, H.: A modality for recursion. In: LICS. pp. 255–266 (2000)
20. Norell, U.: Towards a practical programming language based on dependent type
theory. Ph.D. thesis, Chalmers University of Technology (2007)
21. Paviotti, M., Møgelberg, R.E., Birkedal, L.: A model of PCF in guarded type
theory. In: MFPS (2015)
17
22. Shulman, M.: Univalence for inverse diagrams and homotopy canonicity. Mathematical Structures in Computer Science 25, 1203–1277 (6 2015)
23. Svendsen, K., Birkedal, L.: Impredicative concurrent abstract predicates. In: ESOP.
pp. 149–168 (2014)
24. The Univalent Foundations Program: Homotopy Type Theory: Univalent Foundations of Mathematics. http://homotopytypetheory.org/book, Institute for Advanced Study (2013)
18
Appendix
A
Overview of the appendix
Sec. B contains type and term equalities of Fig. 3 in full detail. Sec. C starting
on page 19 contains detailed explanations of examples from Sec. 3 explaining
how the rules of gDTT are used. Sec. D starting on page 27 contains detailed
explanations of examples with coinductive types. Sec. E starting on page 29
contains a detailed derivation of the type isomorphism ∀κ.A + B ∼
= ∀κ.A + ∀κ.B
used in Sec. 4.
B
Typing rules
Definitional type equalities:
Γ, Γ ′ ⊢∆ A type
⊢∆ ξ[x
κ
κ
Γ ⊢∆ ⊲ξ [x
⊢∆ ξ [x
t] .A ≡ ⊲ξ.A
TyEq-⊲-Weak
Γ, Γ ′ , x : B, y : C, Γ ′′ ⊢∆ A type
κ
u] ξ ′ : Γ _ Γ ′ , x : B, y : C, Γ ′′
x not free in C
t, y
κ
Γ ⊢∆ ⊲ξ [x
t, y
κ
Γ ⊢∆ ⊲ξ [x
κ
⊢∆′ κ
κ
u] ξ ′ .A ≡ ⊲ξ [y
nextκ ξ.t] .A type
κ
nextκ ξ.t] .A ≡ ⊲ξ.A[t/x]
Γ ⊢∆ ⊲ξ [x
∆′ ⊆ ∆
κ
t] : Γ _ Γ ′ , x : B
u, x
t] ξ ′ .A
TyEq-Force
κ
⊢∆ ξ : Γ _ Γ ′
Γ, Γ ′ ⊢∆ A : U∆′
κ
κ
Γ ⊢∆ El(b
⊲ (nextκ ξ.A)) ≡ ⊲ξ. El(t)
κ
⊢∆ ξ : Γ _ Γ ′
Γ, Γ ′ ⊢∆ t : A
TyEq-⊲-Exch
Γ, Γ ′ ⊢∆ s : A
κ
Γ ⊢∆ Idκ⊲ξ.A(nextκ ξ.t, nextκ ξ.s) ≡ ⊲ξ.IdA(t, s)
TyEq-El-⊲
TyEq-⊲
19
Definitional term equalities:
Γ, Γ ′ ⊢∆ u : A
κ
t] : Γ _ Γ ′ , x : B
⊢∆ ξ [x
Γ ⊢∆ nextκ ξ [x
κ
t] .u ≡ nextκ ξ.u : ⊲ξ.A
TmEq-Next-Weak
κ
Γ ⊢∆ t : ⊲ξ.A
Γ ⊢∆ nextκ ξ [x
κ
t] .x ≡ t : ⊲ξ.A
TmEq-Next-Var
Γ, Γ ′ , x : B, y : C, Γ ′′ ⊢∆ t : A
κ
⊢∆ ξ [x t, y u] ξ ′ : Γ _ Γ ′ , x : B, y : C, Γ ′′
x not free in C
Γ ⊢∆ nextκ ξ [x
u] ξ ′ .v ≡ nextκ ξ [y
t, y
Γ ⊢∆ nextκ ξ [x
Γ ⊢∆ nextκ ξ [x
κ
nextκ ξ.t] .u : ⊲ξ [x
u, x
κ
t] ξ ′ .v : ⊲ξ [y
nextκ ξ.t] .A
κ
nextκ ξ.t] .u ≡ nextκ ξ.u[t/x] : ⊲ξ.A[t/x]
u, x
t] ξ ′ .A
TmEq-Force
Γ ⊢∆ fixκ x.t : A
TmEq-Fix
Γ ⊢∆ fixκ x.t ≡ t[nextκ (fixκ x.t) /x] : A
C
Examples
In this section we provide detailed explanations of typing derivations of examples
described in Sec. 3.
C.1
zipWithκ preserves commutativity
The first proof is the simplest. We will define the standard zipWithκ (zipWith)
function on streams and show that if a binary function f is commutative, then
so is zipWithκ f .
The zipWithκ : (A → B → C) → StrκA → StrκB → StrκC is defined by guarded
recursion as
zipWithκ f , fixκ φ.λ(xs, ys : StrκA ).
consκ (f (hdκ xs) (hdκ ys)) (φ
κ
tlκ xs
κ
tlκ ys)
Note that none of the new generalised ⊲ rules of gDTT are needed to type this
function; this is a function on simple types.
Where we need dependent types is, of course, to state and prove properties. To prove our example, that commutativity of f implies commutativity of
zipWithκ f , means we must show that the type
Π(f : A → A → B). (Π(x, y : A).IdB(f x y, f y x)) →
Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs).
TmEq-Next-Exch
20
is inhabited. We will explain how to construct such a term, and why it is typeable
in gDTT. Although this construction might appear complicated at first, the
actual proof term that we construct will be as simple as possible.
Let f : A → A → B be a function and say we have a term
c : Π(x, y : A).IdB(f x y, f y x)
witnessing commutativity of f . We now wish to construct a term of type
Π(xs, ys : StrκA ).IdStrκC(zipWithκ f xs ys, zipWithκ f ys xs)
We do this by guarded recursion. To this end we assume
κ
φ : ⊲ Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs)
and take xs, ys : StrκA . Using c (the proof that f is commutative) we first have
c (hdκ xs) (hdκ ys) of type
IdB(f (hdκ xs) (hdκ ys), f (hdκ ys) (hdκ xs))
and because we have by definition of zipWithκ
hdκ (zipWithκ f xs ys) ≡ f (hdκ xs) (hdκ ys)
hdκ (zipWithκ f ys xs) ≡ f (hdκ ys) (hdκ xs)
we see that c (hdκ xs) (hdκ ys) has type
IdB(hdκ (zipWithκ f xs ys), hdκ (zipWithκ f ys xs)).
To show that the tails are equal we use the induction hypothesis φ. The terms
κ
tlκ xs and tlκ ys are of type ⊲ StrκA , so we first have φ κ tlκ xs of type
κ
zipWithκ f xs ys,
⊲ [xs tlκ xs] . Π (ys : StrκA ) .IdStrκC
zipWithκ f ys xs
Note the appearance of the generalised ⊲, carrying a delayed substitution. Beκ
cause the variable xs does not appear in ⊲ StrκA we may apply the weakening rule
TmEq-Next-Weak to derive
κ
tlκ ys : ⊲ [xs
tlκ xs] . StrκA
Hence we may use the derived applicative rule to have φ κ tlκ xs κ tlκ ys of type
κ xs
tlκ xs
⊲
.IdStrκC(zipWithκ f xs ys, zipWithκ f ys xs)
ys tlκ ys
and which is definitionally equal to
κ xs
next ys
Idκ⊲ Strκ
C
κ xs
next
ys
the type
tlκ xs
κ
. zipWith f xs ys,
tlκ ys
.
κ
tl xs
κ
. zipWith f ys xs
tlκ ys
21
We also compute
tlκ (zipWithκ f xs ys) ≡ nextκ (zipWithκ f ) κ tlκ xs κ tlκ ys
tlκ xs
κ xs
.(zipWithκ f xs ys)
≡ next
ys tlκ ys
and
tlκ (zipWithκ f ys xs) ≡ nextκ
tlκ ys
.(zipWithκ f ys xs).
κ
tl xs
ys
zs
Using the exchange rule TmEq-Next-Exch we have the equality
xs tlκ xs
ys tlκ ys
.(zipWithκ f xs ys).
.(zipWithκ f xs ys) ≡ nextκ
nextκ
κ
κ
ys tl ys
xs tl xs
Putting it all together we have shown that the term φ
κ
tlκ xs
tlκ ys has type
κ
Idκ⊲ Strκ (tlκ (zipWithκ f xs ys), tlκ (zipWithκ f ys xs))
B
which means that the term
fixκ φ.λ (xs, ys : StrκA ) . pη (c (hdκ xs) (hdκ ys)) (φ
κ
tlκ xs
κ
tlκ ys)
has type Π(xs, ys : StrκA ).IdStrκB(zipWithκ f xs ys, zipWithκ f ys xs).
Notice that the resulting proof term could not be simpler than it is. In particular, we do not have to write delayed substitutions in terms, but only in the
intermediate types.
C.2
An example with covectors
The next example is more sophisticated, as it will involve programming and
proving with a data type that, unlike streams, is dependently typed. In particular, we will see that the generalised later, carrying a delayed substitution, is
necessary to type even the most elementary programs.
Covectors are to colists (potentially infinite lists) as vectors are to lists. To
define guarded covectors we first need guarded co-natural numbers. This is the
type satisfying
κ
CoNκ ≡ 1 + ⊲ CoNκ .
where binary sums are encoded
in a standard way. The defi in the type theory
κ
κ
κ
b
bb
nition in gDTT is CoN , El fix φ.(1 +
⊲ φ) .
κ
Using CoN we define the type of covectors of type A, written CoVecκA , as a
CoNκ -indexed type satisfying
CoVecκA (inl hi) ≡ 1
κ
CoVecκA (inr(nextκ m)) ≡ A × ⊲(CoVecκA m)
22
\κ
In gDTT we first define CoVec
A
\κ , fixκ φ.λ(n : CoNκ ). case n of
CoVec
A
b
inl u ⇒ 1
κ
bb
inr m ⇒ A ×
⊲ (φ
κ
m).
\κ n). In the examples we will not distinguish
and then CoVecκA n , El(CoVec
A
κ
κ
\κ . In the above φ has type ⊲(CoN
between CoVecκA and CoVec
→ U∆,κ ) and
A
κ
κ
inside the branches, u has type 1 and m has type ⊲ CoN , which is evident
from the definition of CoNκ . As an example of covectors, we define ones of type
Π(n : CoNκ ). CoVecκN n which produces a covector of any length consisting only
of ones:
ones , fixκ φ.λ(n : CoNκ ). case n of
inl u ⇒ inl hi
inr m ⇒ h1, φ
κ
mi .
When checking the type of this program, we need the generalised later. The type
κ
κ
of the recursive call is ⊲(Π(n : CoNκ ). CoVecκN n), the type of m is ⊲ CoNκ , and
therefore the type of the subterm φ κ m must be
κ
⊲ [n
m] .Π(n : CoNκ ). CoVecκN n.x
We now aim to define the function map on covectors and show that it preserves composition. Given two types A and B the map function has type
map : (A → B) → Π(n : CoNκ ). CoVecκA n → CoVecκB n.
and is defined by guarded recursion as
map f , fixκ φ.λ(n : CoNκ ).
case n of
inl u ⇒ λ(x : 1).x
κ
λ p : A × ⊲ [n m] .(CoVecκA n) .
inr m ⇒
hf (π1 p) , φ κ m κ (π2 p)i
Let us see why the definition has the correct type. First, the types of subterms
are
κ
φ : ⊲(Π(n : CoNκ ). CoVecκA n → CoVecκB n)
u:1
κ
m : ⊲ CoNκ
23
Let C = CoVecκA n → CoVecκB n, and write C(t) for C[t/n]. By the definition
of CoVecκA and CoVecκB we have C(inl u) ≡ 1 → 1, and so λ(x : 1).x has type
C(inl u).
By the definition of CoVecκA we have
κ
CoVecκA (inr m) ≡ A × El b
⊲ (nextκ (CoVecκA ) κ m)
κ
≡ A × ⊲ [n
m] . (CoVecκA n)
and analogously for CoVecκB (inr m). Hence the type C(inr m) is convertible to
κ
κ
A × ⊲ [n m] . (CoVecκA n) → B × ⊲ [n m] . (CoVecκB n) .
Further, using the derived applicative rule we have
φ
κ
κ
m : ⊲ [n
m] .C(n)
and because π2 p in the second branch has type
κ
⊲ [n
m] .(CoVecκA n)
we may use the (simple) applicative rule again to get
φ
κ
which allows us to type
κ
λ p : A × ⊲ [n
m
κ
κ
(π2 p) : ⊲ [n
m] .(CoVecκB n)
m] .(CoVecκA n) . hf (π1 p) , φ
κ
m
κ
π2 (p)i
with type C(inr m). Notice that we have made essential use of the more general
applicative rule to apply φ κ m to π2 p. Using the strong (dependent) elimination
rule for binary sums we can type the whole case construct with type C(n), which
is what we need to give map the desired type.
Now we will show that map so defined satisfies a basic property, namely that
it preserves composition in the sense that the type (in the context where we have
types A, B and C)
Π(f : A → B)(g : B → C)(n : CoNκ )(xs : CoVecκA n).
IdCoVecκC n(map g n(map f n xs), map(g ◦ f ) n xs)
(8)
is inhabited. The proof is, of course, by Löb induction.
First we record some definitional equalities which follow directly by unfolding
the definitions
map f (inl u) x ≡ x
map f (inr m) xs ≡ hf (π1 xs) , nextκ (map f ) κ m κ π2 (xs)i
m
κ n
.(map f n ys)i
≡ hf (π1 xs), next
ys π2 xs
24
and so iterating these two equalities we get
map g (inl u) (map f (inl u) x) ≡ x
map g (inr m) (map f (inr m) xs) ≡ hg(f (π1 xs)), si
where s is the term
n m
.(map g n zs)
n m
nextκ
zs nextκ
.(map f n ys)
ys π2 xs
which is convertible, by the rule TmEq-Force, to the term
n m
.(map g n (map f n ys)).
nextκ
ys π2 xs
Similarly we have
map(g ◦ f ) (inl u) x ≡ x
and map(g ◦ f ) (inr m) xs convertible to
m
κ n
.(map(g ◦ f ) n ys) .
g(f (π1 xs)), next
ys π2 xs
Now let us get back to proving property (8). Take f : A → B, g : B → C and
assume
κ
φ : ⊲Π(n : CoNκ )(xs : CoVecκA n).IdCoVecκC n(map g n(map f n xs), map(g ◦ f ) n xs)
We take n : CoNκ and write
P (n) = Π(xs : CoVecκA n).IdCoVecκC n(map g n(map f n xs), map(g ◦ f ) n xs).
Then similarly as in the definition of map and the definitional equalities for map
above we compute
P (inl u) ≡ Π(xs : 1).Id1(xs, xs)
and so we have λ(xs : 1).r1 xs of type P (inl u).
The other branch (when n = inr m) is of course a bit more complicated. As
before we have
κ
CoVecκA (inr m) ≡ A × ⊲ [n
m] . CoVecκA n
So take xs of type CoVecκA (inr m). We need to construct a term of type
IdCoVecκC n(map g n(map f n xs), map(g ◦ f ) n xs).
(9)
25
First we have rC g(f (π1 xs)) of type IdC(g(f (π1 xs)), g(f (π1 xs))). Then because
κ
m is of type ⊲ CoNκ we can use the induction hypothesis φ to get φ κ m of type
κ
⊲ [n
m] .Π(xs : CoVecκA n).IdCoVecκC n(map g n(map f n xs), map(g ◦ f ) n xs).
κ
Using (9) we have π2 xs of type ⊲ [n m] . CoVecκA n and so we can use the
applicative rule again to give φ κ m κ π2 xs the type
κ n
map g n(map f n xs),
m
⊲
.IdCoVecκC n
map(g ◦ f ) n xs
xs π2 xs
which by the rule TyEq-⊲ is the same as
m
κ n
next
.
(map
g
n(map
f
n
xs))
,
xs π2 xs
IdD
n m
nextκ
. (map(g ◦ f ) n xs)
xs π2 xs
κ
where D is the type ⊲ [n
m] . CoVecκC n. Thus we can give to the term
λ(xs : CoVecκA (inr m)). pη (rC g(f (π1 xs))) (φ
κ
m
κ
π2 xs)
the type P (inr m). Using the dependent elimination rule for binary sums we get
the final proof of property (8) as the term
λ(f : A → B)(g : B → C). fixκ φ.λ(n : CoNκ ).
case n of
inl u ⇒ λ(xs : 1).r1 xs
inr m ⇒ λ(xs : CoVecκA (inr m)). pη (rC g(f (π1 xs))) (φ
κ
m
κ
π2 xs)
which is as simple as could be expected.
C.3
Lifting predicates to streams
Let P : A → U∆ be a predicate on type A and κ a clock variable not in ∆. We
can define a lifting of this predicate to a predicate P κ on streams of elements of
type A. The idea is that P κ xs will hold precisely when P holds for all elements
of the stream. However we do not have access to all the element of the stream at
the same time. As such we will have P κ xs if P holds for the first element of the
stream xs now, and P holds for the second element of the stream xs one time
step later, and so on. The precise definition uses guarded recursion:
P κ : StrκA → U∆,κ
κ
bb
⊲ (φ
P κ , fixκ φ.λ (xs : StrκA ) .P (hdκ xs) ×
κ
κ
tlκ xs) .
In the above term the subterm φ has type ⊲ (StrκA → U∆,κ ) and so because tlκ xs
κ
κ
κ
has type ⊲ StrκA we may form φ κ tlκ xs of type ⊲U∆,κ and so finally b
⊲ (φ κ tlκ xs)
has type U∆,κ as needed.
26
To see that this makes sense, we have for a stream xs : StrκA
κ
⊲ (nextκ P κ
El (P κ xs) ≡ El (P (hdκ xs)) × El b
Using delayed substitution rules we have
nextκ P κ
κ
tlκ xs ≡ nextκ [xs
κ
κ
tlκ xs) .
tlκ xs] . (P κ xs)
which gives rise to the type equality
El(b
⊲ nextκ P κ
κ
tlκ xs] . (P κ xs) .
κ
tlκ xs) ≡ El b
⊲ nextκ [xs
Finally, the type equality rule TyEq-El-⊲ gives us
κ
κ
El b
⊲ nextκ [xs tlκ xs] . (P κ xs) ≡ ⊲ [xs
tlκ xs] . El(P κ xs).
All of these together then give us the type equality
κ
El (P κ xs) ≡ El(P (hdκ xs)) × ⊲ [xs
tlκ xs] . El(P κ xs).
And so if xs = consκ x (nextκ ys) we can further simplify, using rule TyEqForce, to get
κ
⊲ [xs
κ
κ
nextκ ys] . El(P κ xs) ≡ ⊲ (El(P κ xs)[ys/xs]) ≡ ⊲ El(P κ ys)
κ
which then gives El(P κ xs) ≡ El (P x) × ⊲ El (P κ ys) which is in accordance with
the motivation given above.
Because P κ is defined by guarded recursion, we prove its properties by Löb
induction. In particular, we may prove that if P holds on A then P κ holds on
StrκA , i.e., that the type
(Π(x : A). El (P x)) → (Π(xs : StrκA ). El (P κ xs))
is inhabited (in a context where we have a type A and a predicate P ). Take
p : Π(x : A). El (P x), and since we are proving by Löb induction we assume the
induction hypothesis later
κ
φ : ⊲ (Π(xs : StrκA ). El (P κ xs)) .
Let xs : StrκA be a stream. By definition of P κ we have the type equality
κ
El(P κ xs) ≡ El (P hdκ xs) × ⊲ [xs
tlκ xs] . El (P κ xs)
Applying p to hdκ xs gives us the first component
p(hdκ xs) : El (P (hdκ xs))
and applying the induction hypothesis φ we have
φ
κ
κ
tlκ xs : ⊲ [xs
tlκ xs] . El(P κ xs)
Thus combining this with the previous term we have the proof of the lifting
property as the term
λ (p : Π(x : A). El (P x)) .
fixκ φ.λ (xs : StrκA ) hp (hdκ xs) , φ
κ
tlκ xsi .
27
D
Example programs with coinductive types
Let A be some small type in clock context ∆ and κ, a fresh clock variable. Let
StrA = ∀κ. StrκA . We can define head, tail and cons functions
hd : StrA → A
hd , λxs. hd
κ0
tl : StrA → StrA
tl , λxs. prev κ. tlκ (xs[κ])
(xs[κ0 ])
cons : A → StrA → StrA
cons , λx.λxs.Λκ. consκ x (nextκ (xs[κ])) .
With these we can define the acausal ‘every other’ function eoκ that removes
every second element of the input stream. This is acausal because the second
element of the output stream is the third element of the input. Therefore to type
the function we need to have the input stream always available, necessitating
the use clock quantification. The function eoκ is
eoκ : StrA → StrκA
eoκ , fixκ φ.λ (xs : StrA ) .
consκ (hd xs) (φ
κ
nextκ ((tl (tl xs)))) .
i.e., we return the head immediately and then recursively call the function on
the stream with the first two elements removed. Note that the result is a guarded
stream, but we can easily strengthen it and define eo of type StrA → StrA as
eo , λxs.Λκ. eoκ xs.
A more interesting type is the type of covectors, which is a refinement of the
guarded type of covectors defined in Sec. 3. First we define the type of co-natural
numbers CoN as
CoN = ∀κ. CoNκ .
It is easy to define 0 and succ as
0 : CoN
succ : CoN → CoN
0 , Λκ. inl hi
succ , λn.Λκ. inr (nextκ (n[κ]))
.
Next, we will use type isomorphisms to define a transport function comCoN of
type comCoN : CoN → 1 + CoN as
comCoN , λn. case com+ n of
inl u ⇒ inl u[κ0 ]
inr n ⇒ inr prev κ.n[κ]
This function satisfies term equalities
comCoN 0 ≡ inl hi
comCoN (succ n) ≡ inr n.
(10)
28
Using this we can define type of covectors CoVecA as
CoVecA n , ∀κ. CoVecκA n
where CoVecκA : CoN → U∆,κ is the term
fixκ φ.λ (n : CoN) . case comCoN n of
1
inl ⇒ b
b ⊲κ (φ
inr n ⇒ A×b
κ
(nextκ n)) .
Notice the use of comCoN to transport n of type CoN to a term of type 1 +
CoN which we can case analyse. To see that this type satisfies the correct type
equalities we need some auxiliary term equalities which follow from the way we
have defined the terms.
Using term equalities (4) and (5) we can derive the (almost) expected type
equalities
CoVecA 0 ≡ ∀κ.1
κ
CoVecA (succ n) ≡ ∀κ. A × ⊲ (CoVecκ n)
(11)
and using the type isomorphisms we can extend these type equalities to type
isomorphisms
CoVecA 0 ∼
=1
CoVecA (succ n) ∼
= A × CoVecA n
which are the expected type properties of the covector type.
A simple function we can define is the tail function
tl : CoVecA (succ n) → CoVecA
tl , λv. prev κ.π2 (v[κ]) .
Note that we have used (11) to ensure that tl is type correct.
Next, we define the map function on covectors.
map : (A → B) → Π(n : CoN). CoVecA n → CoVecB n
map f = λn.λxs.Λκ. mapκ f n (xs[κ])
where mapκ is the function of type
mapκ : (A → B) → Π(n : CoN). CoVecκA n → CoVecκB n
defined as
λf. fixκ φ.λn. case comCoN n of
inl ⇒ λv.v
inr n ⇒ λv. hf (π1 v), φ
κ
(nextκ n)
κ
π2 (v)i .
29
Let us see that this has the correct type. Let DA (x) (and analogously DB (x))
be the type
DA (x)
case x of
inl ⇒ b
1
,
b ⊲κ ((nextκ CoVecκA )
inr n ⇒ A×b
κ
(nextκ n)) .
where x is of type 1 + CoN. Using this abbreviation we can write the type of
mapκ as
(A → B) → Π(n : CoN).DA (comCoN n) → DB (comCoN n).
Using this it is straightforward to show, using the dependent elimination rule
for sums, as we did in Sec. 3, that mapκ has the correct type. Indeed we have
κ
DA (inl z) ≡ 1 and DA (inr n) ≡ A × ⊲ (CoVecA n).
E
Type isomorphisms in detail
∼ A. The terms are λx.x [κ0 ] and λx.Λκ.x. The rule
– If κ 6∈ A then ∀κ.A =
TmEq-∀-fresh is crucially needed to show that they constitute a type isomorphism.
– If κ 6∈ A then ∀κ.Π(x : A).B ∼
= Π(x : A).∀κ.B. The terms are
λz.λx.Λκ.z[κ] x
of type ∀κ.Π(x : A).B → Π(x : A).∀κ.B and
λz.Λκ.λx.(z x)[κ]
of type Π(x : A).∀κ.B → ∀κ.Π(x : A).B.
– ∀κ.Σ (x : A) B ∼
= Σ (y : ∀κ.A) (∀κ.B [y[κ]/ x]). The terms are
λz. hΛκ.π1 (z[κ]) , Λκ.π2 (z[κ])i
of type
∀κ.Σ (x : A) B → Σ (y : ∀κ.A) (∀κ.B [y[κ]/ x])
and
λz.Λκ. h(π1 z)[κ] , (π2 z)[κ]i
of the converse type.
κ
– ∀κ.A ∼
= ∀κ.⊲A. The terms are
λz.Λκ. nextκ (z[κ])
κ
of type ∀κ.A → ∀κ.⊲A and
λz. prev κ. (z[κ])
of the converse type. The β and η rules for prev κ. ensure that this pair of
functions constitutes an isomorphism.
30
Using these isomorphisms we can construct an additional type isomorphism
witnessing that ∀κ commutes with binary sums. Recall that we encode binary
coproducts using Σ-types and universes in the standard way. Given two codes
b and B
b in some universe U∆ we define
A
b+
b : U∆
bB
A
b+
b , Σ (b : B) if b then A
b else B
b
bB
A
b+
b . Suppose that ∆′ ⊆ ∆ and κ is a clock variable
bB
and we write A+B for El A
b B
b satisfying
not in ∆. Suppose that Γ ⊢∆ and that we have two codes A,
b : U∆′ ,κ
Γ ⊢∆,κ A
b : U∆′ ,κ
Γ ⊢∆,κ B
We start with an auxiliary function comif . Let b be some term of type B. We
then define
b else B
b
comifb : ∀κ. El if b then A
b else b
b
→ El if b then b
∀ Λκ.A
∀ Λκ.B
comifb , if b then λx.x else λx.x
which is typeable due to the strong elimination rule for B.
We now define the function com+
com+ : ∀κ.(A + B) → ∀κ.A + ∀κ.B
D
E
com+ , λz. π1 (z[κ0 ]) , comifπ1 (z[κ0 ]) (Λκ.π2 (z[κ])) .
We need to check that the types are well-formed and the function well-typed.
The side condition Γ ⊢∆ ensures that the types are well-formed. To see that the
function com+ is well-typed we consider the types of subterms.
- The term z has type ∀κ.(A + B).
- The term π1 (z[κ0 ]) has type B.
- The term Λκ.π2 (z[κ]) has type
b else B
b
∀κ. El if π1 (z[κ]) then A
- From TmEq-∀-fresh we get π1 (z[κ0 ]) ≡ π1 (z[κ]). Indeed, the term
Λκ.π1 (z[κ])
has type B, which does not contain κ, and the required equality follows from
TmEq-∀-fresh and the β rule for clock quantification.
- Thus the term Λκ.π2 (z[κ]) has type
b else B
b
∀κ. El if π1 (z[κ0 ]) then A
31
- And so the term
comifπ1 (z[κ0 ]) Λκ.π2 (z[κ])
has type
b else b
b
∀ Λκ.A
∀ Λκ.B
El if π1 (z[κ0 ]) then b
which is exactly the type needed to typecheck the whole term.
For the term com+ we can derive the following definitional term equalities.
com+ (Λκ. inl t) ≡ inl Λκ.t
com+ (Λκ. inr t) ≡ inr Λκ.t
(12)
There is also a canonical term of type
∀κ.A + ∀κ.B → ∀κ.(A + B)
defined as
λz.Λκ. case z of
inl a ⇒ inl (a[κ])
inl b ⇒ inl (b[κ]).
This term is inverse to com+ , although we require equality reflection to show
that the two functions are inverses to each other. Without equality reflection we
can only prove they are inverses up to propositional equality. The isomorphisms
defined previously do not require equality reflection.
| 6cs.PL
|
Static and Dynamic Semantics of NoSQL Languages
Véronique Benzaken1
arXiv:1303.1716v1 [cs.PL] 7 Mar 2013
1
Giuseppe Castagna2
Kim Nguyễn1
Jérôme Siméon3
LRI, Université Paris-Sud, Orsay, France, 2 CNRS, PPS, Univ Paris Diderot, Sorbonne Paris Cité, Paris, France
3
IBM Watson Research, Hawthorne, NY, USA
Abstract
that is deduced directly from the structure of the program even in
the absence of any explicit type declaration or annotation.
We present a calculus for processing semistructured data that spans
differences of application area among several novel query languages, broadly categorized as “NoSQL”. This calculus lets users
define their own operators, capturing a wider range of data processing capabilities, whilst providing a typing precision so far typical
only of primitive hard-coded operators. The type inference algorithm is based on semantic type checking, resulting in type information that is both precise, and flexible enough to handle structured
and semistructured data. We illustrate the use of this calculus by
encoding a large fragment of Jaql, including operations and iterators over JSON, embedded SQL expressions, and co-grouping, and
show how the encoding directly yields a typing discipline for Jaql
as it is, namely without the addition of any type definition or type
annotation in the code.
Example. We use Jaql [5, 18], a language over JSON [19] developed for BigData analytics, to illustrate how our proposed calculus
works. Our reason for using Jaql is that it encompasses all the features found in the previously cited query languages and includes a
number of original ones, as well. Like Pig [28] it supports sequence
iteration, filtering, and grouping operations on non-nested queries.
Like AQL [3] and XQuery [6], it features nested queries. Furthermore, Jaql uses a rich data model that allows arbitrary nesting of
data (it works on generic sequences of JSON records whose fields
can contain other sequences or records) while other languages are
limited to flat data models, such as AQL whose data-model is similar to the standard relational model used by SQL databases (tuples
of scalars and of lists of scalars). Lastly, Jaql includes SQL as an
embedded sub-language for relational data. For these reasons, although in the present work we focus almost exclusively on Jaql, we
believe that our work can be adapted without effort to a wide array
of sequence processing languages.
The following Jaql program illustrates some of those features.
It performs co-grouping [28] between one JSON input, containing
information about departments, and one relational input containing information about employees. The query returns for each department its name and id, from the first input, and the number of
high-income employees from the second input. A SQL expression
is used to select the employees with income above a given value,
while a Jaql filter is used to access the set of departments and the
elements of these two collections are processed by the group expression (in Jaql “$” denotes the current element).
1. Introduction
The emergence of Cloud computing, and the ever growing importance of data in applications, has given birth to a whirlwind of new
data models [19, 24] and languages. Whether they are developed
under the banner of “NoSQL” [12, 35], for BigData Analytics [5,
18, 28], for Cloud computing [3], or as domain specific languages
(DSL) embedded in a host language [21, 27, 32], most of them
share a common subset of SQL and the ability to handle semistructured data. While there is no consensus yet on the precise boundaries of this class of languages, they all share two common traits:
(i) an emphasis on sequence operations (eg, through the popular
MapReduce paradigm) and (ii) a lack of types for both data and programs (contrary to, say, XML programming or relational databases
where data schemas are pervasive). In [21, 22], Meijer argues that
such languages can greatly benefit from formal foundations, and
suggests comprehensions [7, 33, 34] as a unifying model. Although
we agree with Meijer for the need to provide unified, formal foundations to those new languages, we argue that such foundations
should account for novel features critical to various application domains that are not captured by comprehensions. Also, most of those
languages provide limited type checking, or ignore it altogether. We
believe type checking is essential for many applications, with usage
ranging from error detection to optimization. But we understand the
designers and programmers of those languages who are averse to
any kind of type definition or annotation. In this paper, we propose
a calculus which is expressive enough to capture languages that go
beyond SQL or comprehensions. We show how the calculus adapts
to various data models while retaining a precise type checking that
can exploit in a flexible way limited type information, information
group
(depts -> filter each x (x.size > 50))
by g = $.depid as ds,
(SELECT * FROM employees WHERE income > 100)
by g = $.dept as es
into { dept: g,
deptName: ds[0].name,
numEmps: count(es) };
The query blends Jaql expressions (eg, filter which selects, in
the collection depts, departments with a size of more than 50
employees, and the grouping itself) with a SQL statement (selecting employees in a relational table for which the salary is more
than 100). Relations are naturally rendered in JSON as collections
of records. In our example, one of the key difference is that field
access in SQL requires the field to be present in the record, while
the same operation in Jaql does not. Actually, field selection in Jaql
is very expressive since it can be applied also to collections with the
effect that the selection is recursively applied to the components of
the collection and the collection of the results returned, and similarly for filter and other iterators. In other words, the expression
An extended abstract of this work is included in the proceeding of
POPL 13, 40th ACM Symposium on Principles of Programming Languages,
ACM Press, 2013.
1
2013/3/8
filter each x (x.size > 50) above will work as much when
x is bound to a record (with or without a size field: in the latter
case the selection returns null), as when x is bound to a collection
of records or of arbitrary nested collections thereof. This accounts
for the semistructured nature of JSON compared to the relational
model. Our calculus can express both, in a way that illustrates the
difference in both the dynamic semantics and static typing.
In our calculus, the selection of all records whose mandatory
field income is greater than 100 is defined as:
let Sel =
‘nil => ‘nil
| ({income: x, .. } as y , tail) =>
if x > 100 then (y,Sel(tail)) else Sel(tail)
(collections are encoded as lists à la Lisp) while the filtering among
records or arbitrary nested collections of records of those where the
(optional) size field is present and larger than 50 is:
let Fil =
‘nil => ‘nil
| ({size: x, .. } as y,tail) =>
if x > 50 then (y,Fil(tail)) else Fil(tail)
| ((x,xs),tail) => (Fil(x,xs),Fil(tail))
| (_,tail) => Fil(tail)
The terms above show nearly all the basic building blocks of our
calculus (only composition is missing), building blocks that we dub
filters. Filters can be defined recursively (eg, Sel(tail) is a recursive call); they can perform pattern matching as found in functional
languages (the filter p ⇒ f executes f in the environment resulting
from the matching of pattern p); they can be composed in alternation (f1| f2 tries to apply f1 and if it fails it applies f2 ), they can
spread over the structure of their argument (eg, ( f1 ,f2) —of which
(x,Sel(tail)) is an instance— requires an argument of a product type and applies the corresponding fi component-wise).
For instance, the filter Fil scans collections encoded as lists à
la Lisp (ie, by right associative pairs with ‘nil denoting the empty
list). If its argument is the empty list, then it returns the empty list;
if it is a list whose head is a record with a size field (and possibly
other fields matched by “. .”), then it captures the whole record in
y, the content of the field in x, the tail of the list in tail, and keeps
or discards y (ie, the record) according to whether x (ie, the field)
is larger than 50; if the head is also a list, then it recursively applies
both on the head and on the tail; if the head of the list is neither a
list, nor a record with a size field, then the head is discarded. The
encoding of the whole grouping query is given in Section 5.3.
Our aim is not to propose yet another “NoSQL/cloud computing/bigdata analytics” query language, but rather to show how to
express and type such languages via an encoding into our core calculus. Each such language can in this way preserve its execution
model but obtain for free a formal semantics, a type inference system and, as it happens, a prototype implementation. The type information is deduced via the encoding (without the need of any type
annotation) and can be used for early error detection and debugging
purposes. The encoding also yields an executable system that can
be used for rapid prototyping. Both possibilities are critical in most
typical usage scenarios of these languages, where deployment is
very expensive both in time and in resources. As observed by Meijer [21] the advent of big data makes it more important than ever
for programmers (and, we add, for language and system designers)
to have a single abstraction that allows them to process, transform,
query, analyze, and compute across data presenting utter variability
both in volume and in structure, yielding a “mind-blowing number
of new data models, query languages, and execution fabrics” [21] .
The framework we present here, we claim, encompasses them all.
A long-term goal is that the compilers of these languages could use
the type information inferred from the encoding and the encoding
itself to devise further optimizations.
Types. Pig [28], Jaql [18, 29], AQL [3] have all been conceived
by considering just the map-reduce execution model. The type (or,
schema) of the manipulated data did not play any role in their design. As a consequence these languages are untyped and, when
present, types are optional and clearly added as an afterthought.
Differences in data model or type discipline are particularly important when embedded in a host language (since they yield the
so-called impedance mismatch). The reason why types were/are
disregarded in such languages may originate in an alleged tension
between type inference and heterogeneous/semistructured data: on
the one hand these languages are conceived to work with collections of data that are weakly or partially structured, on the other
hand current languages with type inference (such as Haskell or
ML) can work only on homogeneous collections (typically, lists of
elements of the same type).
In this work we show that the two visions can coexist: we type
data by semantic subtyping [16], a type system conceived for semistructured data, and describe computations by our filters which are
untyped combinators that, thanks to a technique of weak typing introduced in [9], can polymorphically type the results of data query
and processing with a high degree of precision. The conception of
filters is driven by the schema of the data rather than the execution
model and we use them (i) to capture and give a uniform semantics
to a wide range of semi structured data processing capabilities, (ii)
to give a type system that encompasses the types defined for such
languages, if any, notably Pig, Jaql and AQL (but also XML query
and processing languages: see Section 5.3), (iii) to infer the precise result types of queries written in these languages as they are
(so without the addition of any explicit type annotation/definition or
new construct), and (iv) to show how minimal extensions/modifications of the current syntax of these languages can bring dramatic
improvements in the precision of the inferred types.
The types we propose here are extensible record types and heterogeneous lists whose content is described by regular expressions
on types as defined by the following grammar:
Types
t
::=
|
|
|
|
|
|
|
v
{ ℓ:t, . . . , ℓ:t}}
{ ℓ:t, . . . , ℓ:t , ..
..}}
[r]
int | char
any | empty | null
t||t
\t
t\
Regexp
r
::=
ε | t | r* | r+ | r? | r r | r||r
(singleton)
(closed record)
(open record)
(sequences)
(base)
(special)
(union)
(difference)
where ε denotes the empty word. The semantics of types can be
expressed in terms of sets of values (values are either constants
—such as 1, 2, true, false, null, ’1’, the latter denoting the
character 1—, records of values, or lists of values). So the singleton type v is the type that contains just the value v (in particular
null is the singleton type containing the value null). The closed
record type { a:int, b:int}} contains all record values with exactly
two fields a and b with integer values, while the open record
type { a:int, b:int , ..
..}} contains all record values with at least two
fields a and b with integer values. The sequence type [r] is the set
of all sequences whose content is described by the regular expression r; so, for example [char*] contains all sequences of characters (we will use string to denote this type and the standard double
quote notation to denote its values) while [({{a:int}} { a:int}} )+]
For the sake of precision, to comply with Jaql’s semantics the last pattern should rather be ({..},tail) => Fil(tail): since field selection
e.size fails whenever e is not a record or a list, this definition would detect
the possibility of this failure by a static type error.
2
2013/3/8
denotes nonempty lists of even length containing record values of
type { a:int}}. The union type s||t contains all the values of s and
\t contains all the values of s that
of t, while the difference type s\
are not in t. We shall use bool as an abbreviation of the union of
the two singleton types containing true and false: ‘true|| ‘false.
any and empty respectively contain all and no values. Recursive
type definitions are also used (see Section 2.2 for formal details).
These types can express all the types of Pig, Jaql and AQL,
all XML types, and much more. So for instance, AQL includes
only homogeneous lists of type t, that can be expressed by our
types as [ t* ]. In Jaql’s documentation one can find the type
[ long(value=1), string(value="a"), boolean* ] which
is the type of arrays whose first element is 1, the second is the string
"a" and all the other are booleans. This can be easily expressed in
our types as [1 "a" bool*]. But while Jaql only allows a limited use of regular expressions (Kleene star can only appear in tail
position) our types do not have such restrictions. So for example [char* ’@’ char* ’.’ ((’f’ ’r’)|(’i’ ’t’))] is the
type of all strings (ie, sequences of chars) that denote email addresses ending by either .fr or .it. We use some syntactic sugar
to make terms as the previous one more readable (eg, [ .* ’@’
.* (’.fr’|’.it’)]). Likewise, henceforth we use { a?:t}} to denote that the field a of type t is optional; this is just syntactic sugar
for stating that either the field is undefined or it contains a value of
type t (for the formal definition see Appendix G).
Coming back to our initial example, the filter Fil defined before
expects as argument a collection of the following type:
former and limited forms of the latter); finally, it includes first-class
arbitrary deep composition (while most languages offer this operator only at top level), whose power is nevertheless restrained by the
type system.
An important contribution of this work is that it directly compares a programming language approach with the tree transducer
one. Our calculus implements transformations typical of top-down
tree transducers but has several advantages over the transducer approach: (1) the transformations are expressed in a formalism immediately intelligible to any functional programmer; (2) our calculus, in its untyped version, is Turing complete; (3) its transformations can be statically typed (at the expenses of Turing completeness) without any annotation yielding precise result types (4)
even if we restrict the calculus only to well-typed terms (thus losing
Turing completeness), it still is strictly more expressive than wellknown and widely studied deterministic top-down tree transducer
formalisms.
The technical contributions are (i) the proof of Turing completeness for our formalism, (ii) the definition of a type system
that copes with records with computable labels (iii) the definition
of a static type system for filters and its correctness, (iv) the definition of a static analysis that ensures the termination (and the proof
thereof) of the type inference algorithm with complexity bounds
expressed in the size of types and filters and (iv) the proof that
the terms that pass the static analysis form a language strictly more
expressive than top-down tree transducers.
Outline. In Section 2 we present the syntax of the three components of our system. Namely, a minimal set of expressions, the
calculus of filters used to program user-defined operators or to encode the operators of other languages, and the core types in which
the types we just presented are to be encoded. Section 3 defines
the operational semantics of filters and a declarative semantics for
operators. The type system as well as the type inference algorithm
are described in Section 4. In Section 5 we present how to handle a large subset of Jaql. Section 8 reports on some subtler design choices of our system. We compare related works in Section 9
and conclude in Section 10. In order to avoid blurring the presentation, proofs, secondary results, further encodings, and extensions
are moved into a separate appendix.
type Depts = [ ( {size?: int, ..} | Depts )* ]
that is a, possibly empty, arbitrary nested list of records with an
optional size field of type int: notice that it is important to specify
the optional field and its type since a size field of a different type
would make the expression x > 50 raise a run-time error. This
information is deduced just from the structure of the filter (since
Fil does not contain any type definition or annotation).
We define a type inference system that rejects any argument of
Fil that has not type Depts, and deduces for arguments of type
[({size: int, addr: string}| {sec: int} | Depts)+]
(which is a subtype of Depts) the result type [({size: int,
addr: string}|Depts)*] (so it does not forget the field addr
but discards the field sec, and by replacing * for + recognizes that
the test may fail).
By encoding primitive Jaql operations into a formal core calculus we shall provide them a formal and clean semantics as
well as precise typing. So for instance it will be clear that applying the following dot selection [ [{a:3}] {a:5, b:true} ].a
the result will be [ [3] 5 ] and we shall be able to deduce
that _.a applied to arbitrary nested lists of records with an optional integer a field (ie, of type t = { a?:int}} | [ t * ] )
yields arbitrary nested lists of int or null values (ie, of type
u = int | null | [ u * ]).
Finally we shall show that if we accept to extend the current
syntax of Jaql (or of some other language) by some minimal filter
syntax (eg, the pattern filter) we can obtain a huge improvement in
the precision of type inference.
2. Syntax
In this section we present the syntax of the three components of our
system: a minimal set of expressions, the calculus of filters used to
program user-defined operators or to encode the operators of other
languages, and the core types in which the types presented in the
introduction are to be encoded.
The core of our work is the definition of filters and types. The
key property of our development is that filters can be grafted to
any host language that satisfies minimal requirements, by simply
adding filter application to the expressions of the host language.
The minimal requirements of the host language for this to be possible are quite simple: it must have constants (typically for types int,
char, string, and bool), variables, and either pairs or record values (not necessarily both). On the static side the host language must
have at least basic and products types and be able to assign a type to
expressions in a given type environment (ie, under some typing assumptions for variables). By the addition of filter applications, the
host language can acquire or increase the capability to define polymorphic user-defined iterators, query and processing expressions,
and be enriched with a powerful and precise type system.
Contributions. The main contribution of this work is the definition of a calculus that encompasses structural operators scattered
over NoSQL languages and that possesses some characteristics
that make it unique in the swarm of current semi-structured data
processing languages. In particular it is parametric (though fully
embeddable) in a host language; it uniformly handles both width
and deep nested data recursion (while most languages offer just the
2.1 Expressions
The only exception are the “bags” types we did not include in order to focus
on essential features.
In this work we consider the following set of expressions
3
2013/3/8
Definition 1 (expressions).
Exprs e ::= c
(constants)
|
x
(variables)
|
(e, e)
(pairs)
|
{e:e, ..., e:e}
(records)
|
e+e
(record concatenation)
|
e\ℓ
(field deletion)
|
op(e, ..., e)
(built-in operators)
|
fe
(filter application)
where f ranges over filters (defined later on), c over generic constants, and ℓ over string constants.
Record types come in two flavors: closed record types whose values are records with exactly the fields specified by the type, and
open record types whose values are records with at least the fields
specified by the type. Product types are standard and we have a
complete set of type connectives, that is, finite unions, intersections
and negations. We use empty, to denote the type that has no values
and any for the type of all values (sometimes denoted by “_” when
used in patterns). We added a term for recursive types, which allows us to encode both the regular expression types defined in the
introduction and, more generally, the recursive type definitions we
used there. Finally, we use Op (capitalized to distinguish it from
expression operators) to denote the host language’s type operators
(if any). Thus, when filter applications return values whose type
belongs just to the foreign language (eg, a list of functions) we suppose the typing of these functions be given by some type operators.
For instance, if succ is a user defined successor function, we will
suppose to be given its type in the form Arrow(int,int) and, similarly, for its application, say apply(succ,3) we will be given the type
of this expression (presumably int). Here Arrow is a type operator
and apply an expression operator.
The denotational semantics of types as sets of values, that we
informally described in the introduction, is at the basis of the definition of the subtyping relation for these types. We say that a type
t1 is a subtype of a type t2 , noted t1 ≤ t2 , if and only if the set
of values denoted by t1 is contained (in the set-theoretic sense) in
the set of values denoted by t2 . For the formal definition and the
decision procedure of this subtyping relation the reader can refer to
the work on semantic subtyping [16].
Intuitively, these expressions represent the syntax supplied by
the host language —though only the first two and one of the next
two are really needed— that we extend with (the missing expressions and) the expression of filter application. Expressions are
formed by constants, variables, pairs, records, and operation on
records: record concatenation gives priority to the expression on
the right. So if in r1 + r2 both records contains a field with the
same label, it is the one in r2 that will be taken, while field deletion
does not require the record to contain a field with the given label
(though this point is not important). The metavariable op ranges
over operators as well as functions and other constructions belonging to or defined by the host language. Among expressions we single out a set of values, intuitively the results of computations, that
are formally defined as follows:
v
::=
c | (v, v) | {ℓ:v; . . . ; ℓ:v}
We use "foo" for character string constants, ’c’ for characters,
1 2 3 4 5 and so on for integers, and backquoted words, such as
‘foo, for atoms (ie, user-defined constants). We use three distinguished atoms ‘nil, ‘true, and ‘false. Double quotes can be
omitted for strings that are labels of record fields: thus we write
{name:"John"} rather than {"name":"John"}. Sequences (aka,
heterogeneous lists, ordered collections, arrays) are encoded à la
LISP, as nested pairs where the atom ‘nil denotes the empty list.
We use [e1 . . . en ] as syntactic sugar for (e1 , . . . , (en , ‘nil)...).
2.3 Patterns
Filters are our core untyped operators. All they can do are three
different things: (1) they can structurally decompose and transform
the values they are applied to, or (2) they can be sequentially
composed, or (3) they can do pattern matching. In order to define
filters, thus, we first need to define patterns.
Definition 3 (patterns).
Patterns
2.2 Types
Expressions, in particular filter applications, are typed by the following set of types (typically only basic, product, recursive and
—some form of— record types will be provided by the host language):
Definition 2 (types).
Types
t
::=
|
|
|
|
|
|
|
|
|
|
|
|
p
::=
|
|
|
|
|
|
t
(type)
x
(variable)
( p,,p))
(pair)
{ ℓ:p, . . . , ℓ:p}}
(closed rec)
{ ℓ:p, . . . , ℓ:p , ..
..}}
(open rec)
p||p
(or/union)
&p
p&
(and/intersection)
where the subpatterns forming pairs, records, and intersections
have distinct capture variables, and those forming unions have the
same capture variables.
b
(basic types)
v
(singleton types)
( t,,t))
(products)
{ ℓ:t, . . . , ℓ:t}}
(closed records)
{ ℓ:t, . . . , ℓ:t , ..
..}}
(open records)
t||t
(union types)
&t
t&
(intersection types)
¬t
(negation type)
empty
(empty type)
any
(any type)
µT.t
(recursive types)
T
(recursion variable)
Op(t, ..., t)
(foreign type calls)
Patterns are essentially types in which capture variables (ranged
over by x, y, . . . ) may occur in every position that is not under a
negation or a recursion. A pattern is used to match a value. The
matching of a value v against a pattern p, noted v/p, either fails
(noted Ω) or it returns a substitution from the variables occurring
in the pattern, into values. The substitution is then used as an
environment in which some expression is evaluated. If the pattern is
a type, then the matching fails if and only if the pattern is matched
against a value that does not have that type, otherwise it returns
the empty substitution. If it is a variable, then the matching always
succeeds and returns the substitution that assigns the matched value
to the variable. The pair pattern ( p1, p2) succeeds if and only if it
is matched against a pair of values and each sub-pattern succeeds
on the corresponding projection of the value (the union of the two
substitutions is then returned). Both record patterns are similar to
the product pattern with the specificity that in the open record
pattern “..” matches all the fields that are not specified in the
where every recursion is guarded, that is, every type variable is
separated from its binder by at least one application of a type
constructor (ie, products, records, or Op).
Most of these types were already explained in the introduction.
We have basic types (int, bool, ....) ranged over by b and singleton types v denoting the type that contains only the value v.
4
2013/3/8
pattern. An intersection pattern p1& p2 succeeds if and only if
both patterns succeed (the union of the two substitutions is then
returned). The union pattern p1| p2 first tries to match the pattern p1
and if it fails it tries the pattern p2 .
& x,,y)) succeeds only if the
For instance, the pattern ( int&
matched value is a pair of values (v1 , v2 ) in which v1 is an integer —in which case it returns the substitution {x/v1 , y/v2 }—
and fails otherwise. Finally notice that the notation “p as x” we
&x.
used in the examples of the introduction, is syntactic sugar for p&
This informal semantics of matching (see [16] for the formal
definition) explains the reasons of the restrictions on capture variables in Definition 3: in intersections, pairs, and records all patterns
must be matched and, thus, they have to assign distinct variables,
while in union patterns just one pattern will be matched, hence the
same set of variables must be assigned, whichever alternative is selected.
The strength of patterns is their connections with types and the
fact that the pattern matching operator can be typed exactly. This is
entailed by the following theorems (both proved in [16]):
The expression filter e always returns the value corresponding
to the evaluation of e (and discards its argument). The filter p ⇒ f
applies the filter f to its argument in the environment obtained by
matching the argument against p (provided that the matching does
not fail). This rather powerful feature allows a filter to perform two
critical actions: (i) inspect an input with regular pattern-matching
before exploring it and (ii) capture part of the input that can be
reused during the evaluation of the subfilter f . If the argument application of fi to vi returns vi′ then the application of the product
filter ( f1 ,f2) to an argument (v1 , v2 ) returns (v1′ , v2′ ); otherwise, if
any application fails or if the argument is not a pair, it fails. The
record filter is similar: it applies to each specified field the corresponding filter and, as stressed by the “. .”, leaves the other fields
unchanged; it fails if any of the applications does, or if any of the
specified fields is absent, or if the argument is not a record. The filter f1| f2 returns the application of f1 to its argument or, if this fails,
the application of f2 . The semantics of a recursive filter is given by
standard unfolding of its definition in recursive calls. The only real
restriction that we introduce for filters is that recursive calls can be
done only on arguments of a given form (ie, on arguments that have
the form of values where variables may occur). This restriction in
practice amounts to forbid recursive calls on the result of another
recursively defined filter (all other cases can be easily encoded).
The reason of this restriction is technical, since it greatly simplifies the analysis of Section 4.5 (which ensures the termination of
type inference) without hampering expressiveness: filters are Turing complete even with this restriction (see Theorem 7). Filters can
be composed: the filter f1 ;f2 applies f2 to the result of applying
f1 to the argument and fails if any of the two does. The condition
that in every subterm of the form f ;g, f does not contain free recursion variables is not strictly necessary. Indeed, we could allow
such terms. The point is that the analysis for the termination of the
typing would then reject all such terms (apart from trivial ones in
which the result of the recursive call is not used in the composition).
But since this restriction does not restrict the expressiveness of the
calculus (Theorem 7 proves Turing completeness with this restriction), then the addition of this restriction is just a design (rather
than a technical) choice: we prefer to forbid the programmer to
write recursive calls on the left-hand side of a composition, than
systematically reject all the programs that use them in a non-trivial
way.
Finally, we singled out some specific filters (specifically, we
chose groupby and orderby ) whose semantics is generally
specified in a declarative rather than operational way. These do not
bring any expressive power to the calculus (the proof of Turing
completeness, Theorem 7, does not use these declarative operators)
and actually they can be encoded by the remaining filters, but it
is interesting to single them out because they yield either simpler
encodings or more precise typing.
Theorem 4 (Accepted type [16]). For every pattern p, the set of all
values v such that v/p 6= Ω is a type. We call this set the accepted
type of p and note it by *p+.
The fact that the exact set of values for which a matching succeeds
is a type is not obvious. It states that for every pattern p there exists
a syntactic type produced by the grammar in Definition 2 whose
semantics is exactly the set of all and only values that are matched
by p. The existence of this syntactic type, which we note *p+, is
of utmost importance for a precise typing of pattern matching. In
particular, given a pattern p and a type t contained in (ie, subtype of)
*p+, it allows us to compute the exact type of the capture variables
of p when it is matched against a value in t:
Theorem 5 (Type environment [16]). There exists an algorithm
that for every pattern p, and t ≤ *p+ returns a type environment
t/p ∈ Vars(p) → Types such that (t/p)(x) = {(v/p)(x) | v : t}.
2.4 Filters
Definition 6 (filters). A filter is a term generated by:
Filters
f
::=
|
|
|
|
|
|
|
|
e
(expression)
p⇒ f
(pattern)
( f ,f))
(product)
{ ℓ:f, . . . , ℓ:f , ..}}
(record)
f|| f
(union)
µ X..f
(recursion)
Xa
(recursive call)
f ;f
(composition)
o
(declarative operators)
Operators
o
::=
|
groupby f
orderby f
Arguments
a
::=
|
|
|
x
c
( a,a))
{ ℓ:a, ..., ℓ:a}}
(filter grouping)
(filter ordering)
3. Semantics
The operational semantics of our calculus is given by the reduction
semantics for filter application and for the record operations. Since
the former is the only novelty of our work, we save space and omit
the latter, which are standard anyhow.
(variables)
(constants)
(pairs)
(record)
3.1 Big step semantics
such that for every subterm of the form f ;g, no recursion variable
is free inf .
We define a big step operational semantics for filters. The definition
is given by the inference rules in Figure 1 for judgements of the
form δ;γ ⊢eval f (a)
r and describes how the evaluation of
the application of filter f to an argument a in an environment γ
yields an object r where r is either a value or Ω. The latter is
a special value which represents a runtime error: it is raised by
the rule (error) either because a filter did not match the form of
its argument (eg, the argument of a filter product was not a pair)
Filters are like transducers, that when applied to a value return another value. However, unlike transducers they possess more
“programming-oriented” constructs, like the ability to test an input and capture subterms, recompose an intermediary result from
captured values and a composition operator. We first describe informally the semantics of each construct.
5
2013/3/8
δ;γ ⊢eval f1 (v)
r1
δ;γ ⊢eval (f1| f2 )(v)
r1
(union1)
(expr)
δ;γ ⊢eval e(v)
r = eval(γ, e)
r
(prod)
δ;γ ⊢eval f1 (v1 )
r1 δ;γ ⊢eval f2 (v2 )
δ;γ ⊢eval ( f1 ,f2) (v1 , v2 )
(r1 , r2 )
(patt)
δ;γ , v/p ⊢eval f (v)
δ;γ ⊢eval (p ⇒ f )(v)
(comp)
r2
δ;γ ⊢eval f1 (v)
Ω δ;γ ⊢eval f2 (v)
δ;γ ⊢eval (f1| f2 )(v)
r2
(rec)
r
r
δ;γ ⊢eval f1 (v)
r1 δ;γ ⊢eval f2 (r1 )
δ;γ ⊢eval (f1 ;f2 )(v)
r2
(union2)
if r1 =
6 Ω
and r2 =
6 Ω
if v/p 6= Ω
r2
r2
δ, (X →
7 f );γ ⊢eval f (v)
r
µ X..f )(v)
δ;γ ⊢eval (µ
r
(rec-call)
if r1 6= Ω
δ;γ ⊢eval (δ(X))(a)
δ;γ ⊢eval (Xa)(v)
(error)
(recd)
if r1 6= Ω
δ;γ ⊢eval f (a)
r
r
Ω
δ;γ ⊢eval f1 (v1 )
r1
···
δ;γ ⊢eval fn (vn )
rn
}({
{ ℓ1 :v1 , ..., ℓn :vn , ..., ℓn+k :vn+k } )
δ;γ ⊢eval { ℓ1 :f1 , ..., ℓn :fn , ..}
{ ℓ1 :r1 , ..., ℓn :rn , ..., ℓn+k :vn+k }
if no other rule applies
if ∀i, ri 6= Ω
Figure 1. Dynamic semantics of filters
Proof (sketch). We can encode untyped call-by-value λ-calculus
by first applying continuation passing style (CPS) transformations
and encoding CPS term reduction rules and substitutions via filters.
Thanks to CPS we eschew the restrictions on composition.
or because some pattern matching failed (ie, the side condition
of (patt) did not hold). Notice that the argument a of a filter is
always a value v unless the filter is the unfolding of a recursive
call, in which case variables may occurr in it (cf. rule rec-call).
Environment δ is used to store the body of recursive definitions.
The semantics of filters is quite straightforward and inspired
by the semantics of patterns. The expression filter discards its
input and evaluates (rather, asks the host language to evaluate) the
expression e in the current environment (expr). It can be thought
of as the right-hand side of a branch in a match_with construct.
The product filter expects a pair as input, applies its sub-filters
component-wise and returns the pair of the results (prod). This
filter is used in particular to express sequence mapping, as the first
component f1 transforms the element of the list and f2 is applied
to the tail. In practice it is often the case that f2 is a recursive call
that iterates on arbitrary lists and stops when the input is ‘nil. If
the input is not a pair, then the filter fails (rule (error) applies).
The record filter expects as input a record value with at least the
same fields as those specified by the filter. It applies each sub-filter
to the value in the corresponding field leaving the contents of other
fields unchanged (recd). If the argument is not a record value or it
does not contain all the fields specified by the record filter, or if the
application of any subfilter fails, then the whole application of the
record filter fails.
The pattern filter matches its input value v against the pattern p.
If the matching fails so the filter does, otherwise it evaluates its subfilter in the environment augmented by the substitution v/p (patt).
The alternative filter follows a standard first-match policy: If the
filter f1 succeeds, then its result is returned (union-1). If f1 fails,
then f2 is evaluated against the input value (union-2). This filter is
particularly useful to write the alternative of two (or more) pattern
filters, making it possible to conditionally continue a computation
based on the shape of the input.
The composition allows us to pass the result of f1 as input to f2 .
The composition filter is of paramount importance. Indeed, without
it, our only way to iterate (deconstruct) an input value is to use a
product filter, which always rebuilds a pair as result.
Finally, a recursive filter is evaluated by recording its body in
δ and evaluating it (rec), while for a recursive call we replace the
recursion variable by its definition (rec-call).
This concludes the presentation of the semantics of nondeclarative filters (ie, without groupby and orderby). These form a
Turing complete formalism (full proof in Appendix B):
3.2 Semantics of declarative filters
To conclude the presentation of the semantics we have to define the
semantics of groupby and orderby. We prefer to give the semantics
in a declarative form rather than operationally in order not to tie it
to a particular order (of keys or of the execution):
Groupby: groupby f applied to a sequence [v1 . . . vm ] reduces
to a sequence [ (k1 , l1 ) . . . (kn , ln ) ] such that:
1. ∀i, 1 ≤ i ≤ m, ∃j, 1 ≤ j ≤ n, s.t. kj = f (vi )
2. ∀j, 1 ≤ j ≤ n, ∃i, 1 ≤ i ≤ m, s.t. kj = f (vi )
3. ∀j, 1 ≤ j ≤ n, lj is a sequence: [ v1j . . . vnj j ]
4. ∀j, 1 ≤ j ≤ n, ∀k, 1 ≤ k ≤ nj , f (vkj ) = kj
5. ki = kj ⇒ i = j
6. l1 , . . . , ln is a partition of [v1 . . . vm ]
Orderby: orderby f applied to [v1 . . . vn ] reduces to [v1′ . . . vn′ ]
such that:
1. [v1′ . . . vn′ ] is a permutation of [v1 . . . vn ],
2. ∀i, j s.t. 1 ≤ i ≤ j ≤ n, f (vi ) ≤ f (vj )
Since the semantics of both operators is deeply connected to a
notion of equality and order on values of the host language, we
give them as “built-in” operations. However we will illustrate how
our type algebra allows us to provide very precise typing rules,
specialized for their particular semantics. It is also possible to
encode co-grouping (or groupby on several input sequences) with
a combination of groupby and filters (cf. Appendix H).
3.3 Syntactic sugar
Now that we have formally defined the semantics of filters we can
use them to introduce some syntactic sugar.
Expressions. The reader may have noticed that the productions
for expressions (Definition 1) do not define any destructor (eg,
projections, label selection, . . . ), just constructors. The reason is
that destructors, as well as other common expressions, can be
encoded by filter applications:
e.ℓ
fst(e)
snd(e)
let p = e1 in e2
if e then e1 else e2
Theorem 7 (Turing completeness). The language formed by
constants, variables, pairs, equality, and applications of nondeclarative filters is Turing complete.
6
def
=
def
=
def
=
def
=
def
=
({{ℓ:x , ..
..}} ⇒ x)e
(((x,,any)) ⇒ x)e
(((any,,x)) ⇒ x)e
(p ⇒ e2 )e1
(‘true ⇒ e1| ‘false ⇒ e2 )e
2013/3/8
match e with p1 ⇒ e1 |...|pn ⇒ en
def
= (p1 ⇒ e1| . . . | pn ⇒ en )e
whether the type t1 of e1 is finite (ie, it contains only finitely many
values, such as, say, Bool) or not. If a type is finite, (finiteness of
regular types seen as tree automata can be decided in polynomial
time [10]), then it is possible to write it as a finite union of values
(actually, of singleton types). So consider again { e1 :e2} and let t1
be the type of e1 and t2 the type of e2 . First, t1 must be a subtype of string (since record labels are strings). So if t1 is finite
it can be expressed as ℓ1| · · · | ℓn which means that e1 will return
the string ℓi for some i ∈ [1..n]. Therefore { e1 :e2} will have type
{ ℓi : t2} for some i ∈ [1..n] and, thus, the union of all these types,
as expressed by the rule [R CD -F IN ] below. If t1 is infinite instead,
then all we can say is that it will be a record with some (unknown)
labels, as expressed by rule [R CD -I NF ].
These are just a possible choice, but others are possible. For instance in Jaql dot selection is overloaded: when _.ℓ is applied to
a record, Jaql returns the content of its ℓ field; if the field is absent or the argument is null, then Jaql returns null and fails if
the argument is not a record; when applied to a list (‘array’ in Jaql
terminology) it recursively applies to all the elements of the list. So
Jaql’s “_.ℓ” is precisely defined as
µX..({{ ℓ:x , ..
..}} ⇒ x | ({{..}}|null) ⇒ null | (h,,t)) ⇒ (Xh,Xt ))
Besides the syntactic sugar above, in the next section we will use
t1 + t2 to denote the record type formed by all field types in t2
and all the field types in t1 whose label is not already present in t2 .
Similarly t \ ℓ will denote the record types formed by all field types
in t apart from the one labelled by ℓ, if present. Finally, we will also
use for expressions, types, and patterns the syntactic sugar for lists
used in the introduction. So, for instance, [p1 p2 ... pn ] is matched
by lists of n elements provided that their i-th element matches pi .
[R CD -F IN ]
Γ ⊢ e : ℓ1| · · · | ℓn
Γ ⊢ {e:e } : { ℓ1 :t}}| · · · |{ ℓn :t}}
[R CD -I NF ]
Γ⊢e:t
Γ ⊢ e′ : t′
′
Γ ⊢ {e:e } : { ..}}
4. Type inference
Γ ⊢ {e1 :e′1 } : t1
Γ⊢c:c
Γ ⊢ {en :e′n } : tn
[R CD -D EL ]
[R CD -C ONC ]
Variables, constants, and pairs are straightforwardly typed by
Γ ⊢ x : Γ(x)
···
Γ ⊢ {e1 :e′1 , . . . , en :e′n } : t1 + · · · + tn
4.1 Typing of simple and foreign expressions
[C ONSTANT ]
t ≤ string
t is infinite
[R CD -M UL ]
In this section we describe a type inference algorithm for our
expressions.
[VARS ]
Γ ⊢ e′ : t
′
Γ ⊢ e1 : t1 Γ ⊢ e2 : t2 t1 ≤ { ..}}
Γ ⊢ e1 + e2 : t1 + t2 t2 ≤ { ..}}
[P ROD ]
Γ ⊢ e1 : t1 Γ ⊢ e2 : t2
Γ ⊢ (e1 , e2 ) : ( t1, t2)
Γ⊢e:t
}
t ≤ { ..}
Γ ⊢e\ℓ: t\ℓ
Records with multiple fields are handled by the rule [R CD -M UL ]
which “merges” the result of typing single fields by using the type
operator + as defined in CDuce [4, 15], which is a right-priority
record concatenation defined to take into account undefined and
unknown fields: for instance, { a:int, b:int}} + { a?:bool}} =
{ a:int||bool, b:int}} ; unknown fields in the right-hand side may
override known fields of the left-hand side, which is why, for instance, we have { a:int, b:bool}} + { b:int , ..
..}} = { b:int , ..
..}} ;
likewise, for every record type t (ie, for every t subtype of {..}})
we have t + { ..}} = { ..}}. Finally, [R CD -C ONC ] and [R CD -D EL ]
deal with record concatenation and field deletion, respectively, in
a straightforward way: the only constraint is that all expressions
must have a record type (ie, the constraints of the form ... ≤ { ..}}).
See Appendix G for formal definitions of all these type operators.
Notice that these rules do not ensure that a record will not have
two fields with the same label, which is a run-time error. Detecting such an error needs sophisticated type systems (eg, dependent
types) beyond the scope of this work. This is why in the rule [R CD M UL ] we used type operator “+” which, in case of multiple occurring labels, since records are unordered, corresponds to randomly
choosing one of the types bound to these labels: if such a field is
selected, it would yield a run-time error, so its typing can be ambiguous. We can fine tune the rule [R CD -M UL ] so that when all the
ti are finite unions of record types, then we require to have pairwise
disjoint sets of labels; but since the problem would still persist for
infinite types we prefer to retain the current, simpler formulation.
where Γ denotes a typing environment that is a function from expression variables to types and c denotes both a constant and the
singleton type containing the constant. Expressions of the host language are typed by the type function which given a type environment and a foreign expression returns the type of the expression,
and that we suppose to be given for each host language.
[F OREIGN ]
Γ ⊢ e1 : t1 · · · Γ ⊢ en : tn
Γ ⊢ op(e1 ,..., en ) : type((Γ, x1 :t1 , ..., xn :tn ), op(x1 ,..., xn ))
Since the various ei can contain filter applications, thus unknown
to the host language’s type system, the rule [F OREIGN ] swaps them
with variables having the same type.
Notice that our expressions, whereas they include filter applications, do not include applications of expressions to expressions.
Therefore if the host language provides function definitions, then
the applications of the host language must be dealt as foreign expressions, as well (cf. the expression operator apply in Section 2.2).
4.2 Typing of records
The typing of records is novel and challenging because record expressions may contain string expressions in label position, such as
in { e1 :e2} , while in all type systems for record we are aware of,
labels are never computed. It is difficult to give a type to { e1 :e2}
since, in general, we do not statically know the value that e1 will
return, and which is required to form a record type. All we can
(and must) ask is that this value will be a string. To type a record
expression { e1 :e2} , thus, we distinguish two cases according to
4.3 Typing of filter application
Filters are not first-class: they can be applied but not passed around
or computed. Therefore we do not assign types to filters but, as for
any other expression, we assign types to filter applications. The
typing rule for filter application
The function type must be able to handle type environments with types
of our system. It can do it either by subsuming variable with specific types
to the types of the host language (eg, if the host language does not support
singleton types then the singleton type 3 will be subsumed to int) or by
typing foreign expressions by using our types.
[F ILTER -A PP ]
Γ⊢e:t
7
Γ ;∅ ;∅ ⊢fil f (t) : s
Γ ⊢ fe : s
2013/3/8
relies on an auxiliary deduction system for judgments of the form
Γ ;∆ ;M ⊢fil f (t) : s that states that if in the environments
Γ, ∆, M (explained later on) we apply the filter f to a value of
type t, then it will return a result of type s.
To define this auxiliary deduction system, which is the core of
our type analysis, we first need to define *f +, the type accepted by
a filter f . Intuitively, this type gives a necessary condition on the
input for the filter not to fail:
to define recursive types. In order to use them we need to define
three different environments: Γ : Vars → Types denoting type
environments that associate term variables with their types; ∆ :
RVars → Filters denoting definition environments that associate
each filter recursion variable with the body of its definition; M :
RVars × Types → TVars denoting memoization environments
which record that the call of a given recursive filter on a given
type yielded the introduction of a fresh recursion type variable.
Our typing rules, thus work on judgments of the form Γ ;∆ ;M ⊢
f (t) : t′ stating that applying f to an expression of type t in the
environments Γ, ∆, M yields a result of type t′ . This judgment
can be derived with the set of rules given in Figure 2.
These rules are straightforward, when put side by side with the
dynamic semantics of filters, given in Section 3. It is clear that this
type system simulates at the level of types the computations that are
carried out by filters on values at runtime. For instance, rule [F IL E XPR ] calls the typing function of the host language to determine
the type of an expression e. Rule [F IL -P ROD ] applies a product filter
recursively on the first and second projection for each member of
the product decomposition of the input type and returns the union
of all result types. Rule [F IL -R EC ] for records is similar, recursively
applying sub-filters label-wise for each member of the record decomposition and returning the union of the resulting record types.
As for the pattern filter (rule [F IL -PAT ]), its subfilter f is typed in
the environment augmented by the mapping t/p of the input type
against the pattern (cf. Theorem 5). The typing rule for the union
filter, [F IL -U NION ] reflects the first match policy: when typing the
second branch, we know that the first was not taken, hence that at
runtime the filtered value will have a type that is in t but not in *f1 +.
Notice that this is not ensured by the definition of accepted type —
which is a rough approximation that discards grosser errors but,
as we stressed right after its definition, is not sufficient to ensure
that evaluation of f1 will not fail— but by the type system itself:
the premises check that f1 (t1 ) is well-typed which, by induction,
implies that f1 will never fail on values of type t1 and, ergo, that
these values will never reach f2 . Also, we discard from the output
type the contribution of the branches that cannot be taken, that is,
branches whose accepted type have an empty intersection with the
input type t. Composition (rule [F IL -C OMP ]) is straightforward. In
this rule, the restriction that f1 is a filter with no open recursion
variable ensures that its output type s is also a type without free
recursion variables and, therefore, that we can use it as input type
for f2 . The next three rules work together. The first, [F IL -F IX ] introduces for a recursive filter a fresh recursion variable for its output
type. It also memoize in ∆ that the recursive filter X is associated
with a body f and in M that for an input filter X and an input type
t, the output type is the newly introduced recursive type variable.
When dealing with a recursive call X two situations may arise.
One possibility is that it is the first time the filter X is applied to
the input type t. We therefore introduce a fresh type variable T
and recurse, replacing X by its definition f . Otherwise, if the input
type has already been encountered while typing the filter variable
X, we can return its memoized type, a type variable T . Finally,
Rule [F IL -O RD B Y ] and Rule [F IL -G RP B Y ] handle the special cases
of groupby and orderby filters. Their typing is explained in the
following section.
Definition 8 (Accepted type). Given a filter f , the accepted type
of f , written *f + is the set of values defined by:
*e+
= any
*Xa+
= any
*p ⇒ f + = *p + & * f +
µX.. f +
*µ
= *f +
*f1| f2 + = *f1 + | * f2 +
*groupby f + = [any*]
*((f1 ,f2) + = ( * f1 + , * f2 + )
*orderby f + = [any*]
*f1 ;f2 + = *f1 +
*{{ ℓ1 :f1 ,.., ℓn :fn , ..}}+ = { ℓ1 : * f1 + ,.., ℓn : * f2 + , ..
..}}
It is easy to show that an argument included in the accepted type is
a necessary (but not sufficient, because of the cases for composition
and recursion) condition for the evaluation of a filter not to fail:
Lemma 9. Let f be a filter and v be a value such that v ∈
/ *f +.
For every γ, δ, if δ;γ ⊢eval f (v)
r, then r ≡ Ω.
The proof is a straightforward induction on the structure of the
derivation, and is detailed in Appendix C. The last two auxiliary
definitions we need are related to product and record types. In the
presence of unions, the most general form for a product type is a
finite union of products (since intersections distribute on products).
For instance consider the type
( int,,int))|( string,,string))
This type denotes the set of pairs for which either both projections
are int or both projections are string. A type such as
(int||string,,int||string))
is less precise, since it also allows pairs whose first projection is an
int and second projection is a string and vice versa. We see that
it is necessary to manipulate finite unions of products (and similarly
for records), and therefore, we introduce the following notations:
Lemma 10 (Product decomposition). Let t ∈ Types such that
t ≤ ( any,,any)). A product decomposition of t, denoted by π (t) is a
set of types:
n
π (t) = {((t11, t12) , . . . , ( tn
1 , t2 ) }
W
such that t = ti ∈ππ (t) ti . For a given product decomposition, we
say that n is the rank of t, noted rank(t), and use the notation
π ji (t) for the type tji .
There exist several suitable decompositions whose details are
out of the scope of this paper. We refer the interested reader to [15]
and [23] for practical algorithms that compute such decompositions
for any subtype of ( any,,any)) or of {..
..}}. These notions of decomposition, rank and projection can be generalized to records:
Lemma 11 (Record decomposition). Let t ∈ Types such that
t ≤ {..}} . A record decomposition of t, denoted by ρ(t) is a finite
set of types ρ (t)={r1 , . . . , rn } where each ri is either of the form
{ ℓi1 :ti1 , . . . , ℓiniW
:tini } or of the form { ℓi1 :ti1 , . . . , ℓini :tini , ..
..}} and
such that t = ri ∈ρρ(t) ri . For a given record decomposition, we
say that n is the rank of t, noted rank(t), and use the notation ρ jℓ (t)
for the type of label ℓ in the j th component of ρ (t).
4.4 Typing of orderby and groupby
While the “structural” filters enjoy simple, compositional typing
rules, the ad-hoc operations orderby and groupby need specially
crafted rules. Indeed it is well known that when transformation
languages have the ability to compare data values type-checking
(and also type inference) becomes undecidable (eg, see [1, 2]).
We therefore provide two typing approximations that yield a good
In our calculus we have three different sets of variables. The
set Vars of term variables, ranged over by x, y, ..., introduced
in patterns and used in expressions and in arguments of calls of
recursive filters. The set RVars of term recursion variables, ranged
over by X, Y, ... and that are used to define recursive filters. The
set TVars of type recursion variables, ranged over by T, U, ... used
8
2013/3/8
[F IL -E XPR ]
[F IL -PAT ]
Γ ∪ t/p ;∆ ;M ⊢fil f (t) : s
Γ ;∆ ;M ⊢fil p ⇒ f (t) : s
Γ ;∆ ;M ⊢fil e(t) : type(Γ, e,)
[F IL -P ROD ]
i=1..rank(t), j=1, 2
t ≤ *p + & * f +
π ij (t)) : sij
Γ ;∆ ;M ⊢fil fj (π
_
Γ ;∆ ;M ⊢fil ( f1 ,f2) (t) :
( si1, si2)
i=1..rank(t)
[F IL -R EC ]
i=1..rank(t), j=1..m
ρ iℓ (t)) : sij
Γ ;∆ ;M ⊢fil fj (ρ
j
_
}(t) :
}
Γ ;∆ ;M ⊢fil { ℓ1:f1 , . . . , ℓm:fm , ..}
{ ℓ1:si1 , . . . , ℓm:sim , ..
..}
[F IL -U NION ]
i = 1, 2 Γ ;∆ ;M ⊢fil fi (ti ) : si
_
Γ ;∆ ;M ⊢fil f1| f2 (t) :
si
{i|si 6=empty}
i=1..rank(t)
[F IL -C OMP ]
Γ ;∆ ;M ⊢fil f1 (t) : s
Γ ;∆ ;M ⊢fil f2 (s) : s′
[F IL -F IX ]
Γ ;∆, (X 7→ f ) ;M, ((X, t) 7→ T ) ⊢fil f (t) : s
Γ ;∆ ;M ⊢fil f1 ;f2 (t) : s′
µ X..f )(t) : µ T . s
Γ ;∆ ;M ⊢fil (µ
[F IL -C ALL -N EW ]
Γ ;∆ ;M, ((X, t) 7→ T ) ⊢fil ∆(X)(t) : t′
[F IL -O RD B Y ]
∀ti ∈ item(t)
Γ ;∆ ;M ⊢fil f (ti ) : si
Γ ;∆ ;M ⊢fil (orderby f )(t) : OrderBy(t)
T fresh
[F IL -C ALL -M EM ]
t = type(Γ, a)
(X, t) 6∈ dom(M)
T fresh
Γ ;∆ ;M ⊢fil (Xa)(s) : µ T . t′
t ≤ *f1 + | * f2 +
& * f1 +
t1 = t&
& ¬*f1 +
t2 = t&
Γ ;∆ ;M ⊢fil (Xa)(s) : M (X, t)
t = type(Γ, a)
(X, t) ∈ dom(M)
[F IL -G RP B Y ]
∀ti ∈ item(t)
Γ ;∆ ;M ⊢fil f (ti ) : si
t ≤ [any*]
W
Γ ;∆ ;M ⊢fil (groupby f )(t) : [((( i si ),OrderBy(t))) *]
t ≤ [any*]
W
isi is ordered
Figure 2. Type inference algorithm for filter application
compromise between precision and decidability. First we define an
auxiliary function over sequence types:
whose full proof is given in Appendix C. It is easy to write a filter for which the type inference algorithm, that is the deduction
of ⊢fil , does not terminate: µX.. x ⇒ X((x,x)). The deduction of
Γ ;∆ ;M ⊢fil f (t) : s simulates an (abstract) execution of the
filter f on the type t. Since filters are Turing complete, then in
general it is not possible to decide whether the deduction of ⊢fil
for a given filter f will terminate for every input type t. For this
reason we define a static analysis Check(f ) for filters that ensures
that if f passes the analysis, then for every input type t the deduction of Γ ;∆ ;M ⊢fil f (t) : s terminates. For space reasons the
formal definition of Check(f ) is relegated to Appendix A, but its
behavior can be easily explained. Imagine that a recursive filter f
is applied to some input type t. The algorithm tracks all the recursive calls occurring in f ; next it performs one step of reduction of
each recursive call by unfolding the body; finally it checks in this
unfolding that if a variable occurs in the argument of a recursive
call, then it is bound to a type that is a subtree of the original type
t. In other words, the analysis verifies that in the execution of the
derivation for f (t) every call to s/p for some type s and pattern
p always yields a type environment where variables used in recursive calls are bound to subtrees of t. This implies that the rule
[F IL -C ALL - NEW ] will always memoize for a given X, types that are
obtained from the arguments of the recursive calls of X by replacing their variables with a subtree of the original type t memoized
by the rule [F IL -F IX ]. Since t is regular, then it has finitely many
distinct subtrees, thus [F IL -C ALL -N EW ] can memoize only finitely
many distinct types, and therefore the algorithm terminates.
More precisely, the analysis proceeds in two passes. In the first
pass the algorithm tracks all recursive filters and for each of them it
(i) marks the variables that occur in the arguments of its recursive
calls, (ii) assigns to each variable an abstract identifier representing the subtree of the input type to which the variable will be bound
at the initial call of the filter, and (iii) it returns the set of all types
obtained by replacing variables by the associated abstract identifier
in each argument of a recursive call. The last set intuitively represents all the possible ways in which recursive calls can shuffle and
recompose the subtrees forming the initial input type. The second
phase of the analysis first abstractly reduces by one step each recursive filter by applying it on the set of types collected in the first
phase of the analysis and then checks whether, after this reduction,
all the variables marked in the first phase (ie, those that occur in arguments of recursive calls) are still bound to subtrees of the initial
input type: if this checks fails, then the filter is rejected.
Definition 12 (Item set). Let t ∈ Types such that t ≤ [any*].
The item set of t denoted by item(t) is defined by:
item(empty)
=∅
&( any,any)))
item(t)
item(t&
_ 1 2 =[
item( ( ti , ti ) ) = ({t1i } ∪ item(t2i ))
1≤i≤rank(t)
if t 6≤ ( any,any))
1≤i≤rank(t)
The first and second line in the definition ensure that item() returns
the empty set for sequence types that are not products, namely for
the empty sequence. The third line handles the case of non-empty
sequence type. In this case t is a finite union of products, whose
first components are the types of the “head” of the sequence and
second components are recursively the types of the tails. Note also
that this definition is well-founded. Since types are regular trees the
number of distinct types accumulated by item() is finite. We can
now defined typing rules for the orderby and groupby operators.
orderby f : The orderby filter uses its argument filter f to
compute a key from each element of the input sequence and then
returns the same sequence of elements, sorted with respect to their
key. Therefore, while the types of the elements in the result are still
known, their order is lost. We use item() to compute the output
type of an orderby application:
_
OrderBy(t) = [( ti ) ∗ ]
ti ∈item(t)
groupby f : The typing of orderby can be used to give a rough
approximation of the typing of groupby as stated by rule [F IL G RP B Y ]. In words, we obtain a list of pairs where the key component is the result type of f applied to the items of the sequence,
and use OrderBy to shuffle the order of the list. A far more precise typing of groupby that keeps track of the relation between list
elements and their images via f is given in Appendix E.
4.5 Soundness, termination, and complexity
The soundness of the type inference system is given by the property
of subject reduction for filter application
Theorem 13 (subject reduction). If ∅ ;∅ ;∅ ⊢fil f (t) : s, then
for all v : t, ∅;∅ ⊢eval f (v)
r implies r : s.
9
2013/3/8
5. Jaql
It is not difficult to see that the type inference algorithm converges if and only if for every input type there exists a integer n
such that after n recursive calls the marked variables are bound
only to subtrees of the initial input type (or to something that does
not depend on it, of course). Since deciding whether such an n
exists is not possible, our analysis checks whether for all possible
input types a filter satisfies it for n=1, that is to say, that at every
recursive call its marked variables satisfy the property; otherwise it
rejects the filter.
In this Section, we show how filters can be used to capture some
popular languages for processing data on the Cloud. We consider
Jaql [18], a query language for JSON developed by IBM. We give
translation rules from a subset of Jaql into filters.
Definition 15 (Jaql expressions). We use the following simplified
grammar for Jaql (where we distinguish simple expressions, ranged
over by e, from “core expressions” ranged over by k).
Theorem 14 (Termination). If Check(f ), then for every type t the
deduction of Γ ;∅ ;∅ ⊢fil f (t) : s is in 2-EXPTIME. Furthermore,
if t is given as a non-deterministic tree automaton (NTA) then
Γ ;∅ ;∅ ⊢fil f (t) : s is in EXPTIME, where the size of the problem
is |f | × |t|.
e ::=
|
|
|
|
|
|
|
k ::=
|
|
|
(for proofs see Appendix A for termination and Appendix D for
complexity). This complexity result is in line with those of similar
formalisms. For instance in [20], it is shown that type-checking non
deterministic top-down tree transducers is in EXPTIME when the
input and output types are given by a NTA.
All filters defined in this paper (excepted those in Appendix B)
pass the analysis. As an example consider the filter rotate that applied to a list returns the same list with the first element moved to
the last position (and the empty list if applied to the empty list):
µ X.. ( ( x,((y,z))) ⇒ ( y,X(( x,z))) | w ⇒ w )
The analysis succeeds on this filter. If we denote by ιx the abstract
subtree bound to the variable x, then the recursive call will be executed on the abstract argument ( ιx ,ιz ) . So in the unfolding of the
recursive call x is bound to ιx , whereas y and z are bound to two
distinct subtrees of ιz . The variables in the recursive call, x and z,
are thus bound to subtrees of the original tree (even though the argument of the recursive call is not a subtree of the original tree),
therefore the filter is accepted . In order to appreciate the precision
of the inference algorithm consider the type [int+ bool+], that is,
the type of lists formed by some integers (at least one) followed by
some booleans (at least one). For the application of rotate to an
argument of this type our algorithm statically infers the most precise type, that is, [int* bool+ int]. If we apply it once more the
inferred type is [int* bool+ int int]|[bool* int bool].
Generic filters are Turing complete. However, requiring that
Check() holds —meaning that the filter is typeable by our system—
restricts the expressive power of our filters by preventing them
from recomposing a new value before doing a recursive call. For
instance, it is not possible to typecheck a filter which reverses the
elements of a sequence. Determining the exact class of transformations that typeable filters can express is challenging. However it is
possible to show (cf. Appendix F) that typeable filters are strictly
more expressive than top-down tree transducers with regular lookahead, a formalism for tree transformations introduced in [14]. The
intuition about this result can be conveyed by and example. Consider the tree:
a(u1 (. . . (un ()))v1 (. . . (vm ())))
that is, a tree whose root is labeled a with two children, each being
a monadic tree of height n and m, respectively. Then it is not possible to write a top-down tree transducer with regular look-ahead
that creates the tree
a(u1 (. . . (un (v1 (. . . vm ())))))
which is just the concatenation of the two children of the root, seen
as sequences, a transformation that can be easily programmed by
typeable filters. The key difference in expressive power comes from
the fact that filters are evaluated with an environment that binds
capture variables to sub-trees of the input. This feature is essential
to encode sequence concatenation and sequence flattening —two
pervasive operations when dealing with sequences— that cannot be
expressed by top-down tree transducers with regular look-ahead.
c
(constants)
x
(variables)
$
(current value)
[e,... , e]
(arrays)
{ e:e,... , e:e }
(records)
e.l
(field access)
op(e,... ,e)
(function call)
e -> k
(pipe)
filter (each x )? e
(filter)
transform (each x)? e
(transform)
expand ((each x)? e)?
(expand)
group ((each x)? by x = e (as x)?)? into e (grouping)
5.1 Built-in filters
In order to ease the presentation we extend our syntax by adding
“filter definitions” (already informally used in the introduction) to
filters and “filter calls” to expressions:
e ::= let filter F [F1 , . . . , Fn ] = f in e
f ::= F [f, . . . , f ]
(filter defn.)
(call)
where F ranges over filter names. The mapping for most of the
language we consider rely on the following built-in filters.
let filter Filter [F ] = µ X..
‘nil ⇒ ‘nil
| (( x,, xs)), tl)) ⇒ ( X(x, xs),X(tl)))
| ( x,,tl)) ⇒ F x ;(‘true ⇒ ( x, X(tl)))| ‘false ⇒ X(tl))
let filter Transform [F ] = µ X..
‘nil ⇒ ‘nil
| (( x,, xs)), tl)) ⇒ ( X(x, xs),X(tl)))
| ( x,,tl)) ⇒ ( F x, X(tl)))
let filter Expand = µ X..
‘nil ⇒ ‘nil
| ( ‘nil,,tl)) ⇒ X(tl)
| (( x,, xs)), tl)) ⇒ ( x,X(xs, tl)))
5.2 Mapping
Jaql expressions are mapped to our expressions as follows (where
$ is a distinguished expression variable interpreting Jaql’s $ ):
= c
JcK
JxK
= x
= $
J$ K
J{e1 :e′1 ,...,en :e′n }K = { Je1 K :Je′1 K , ...,Jen K :Je′n K }
Je.lK
= JeK .l
= op(Je1 K , ...,Jen K)
Jop(e1 , ..., en )K
J[e1 ,...,en ]K
= (Je1 K , ...(Jen K , ‘nil)...)
= JeK ;JkKF
Je -> kK
Jaql core expressions are mapped to filters as follows:
10
2013/3/8
Jfilter eKF
= Jfilter each $ eKF
Jfilter each x eKF
= Filter [x ⇒ JeK ]
= Jtransform each $ eKF
Jtransform eKF
Jtransform each x eKF = Transform [x ⇒ JeK ]
= JexpandKF ;Jtransform each x eKF
Jexpand each x eKF
= Expand
JexpandKF
Jgroup into eKF
= Jgroup by y=true into eKF
Jgroup by y=e1 into e2 KF = Jgroup each $ by y=e1 into e2 KF
Jgroup each x by y=e1 into e2 KF =
Jgroup each x by y = e1 as $ into e2 KF
Jgroup each x by y = e1 as g into e2 KF =
groupby x ⇒ Je1 K ; Transform [((y,g)) ⇒ Je2 K ]
This translation defines the (first, in our knowledge) formal semantics of Jaql. Such a translation is all that is needed to define the
semantics of a NoSQL language and, as a bonus, endow it with the
type inference system we described without requiring any modification of the original language. No further action is demanded
since the machinery to exploit it is all developed in this work.
As for typing, every Jaql expression is encoded into a filter for
which type-checking is ensured to terminate: Check() holds for
Filter[], Transform[], and Expand (provided it holds also for
their arguments) since they only perform recursive calls on recombinations of subtrees of their input; by its definition, the encoding
does not introduce any new recursion and, hence, it always yields a
composition and application of filters for which Check() holds.
Let us now illustrate how the above composition of filters is
typed. Consider an instance where:
• employees has type [ Remp* ], where
Remp ≡ { dept: int, income:int, ..}
• depts has type [ (Rdep | Rbranch)* ], where
Rdep ≡ {depid:int, name: string, size: int}
Rbranch ≡ {brid:int, name: string}
(this type is a subtype of Dept as defined in the introduction)
The global input type is therefore (line 1)
[ [ Remp* ] [ (Rdep | Rbranch)* ] ]
which becomes, after selection and filtering (line 2)
[ [ Remp* ] [ Rdep* ] ]
(note how all occurrences of Rbranch are ignored by Fil). Tagging
with an integer (line 3) and flattening (line 4) yields
[ (1,Remp)* (2,Rdep)* ]
which illustrates the precise typing of products coupled with singleton types (ie, 1 instead of int). While the groupby (line 5) introduces an approximation the dependency between the tag and the
corresponding type is kept
[ (int, [ ((1,Remp) | (2,Rdep) )+ ]) * ]
Lastly the transform is typed exactly, yielding the final type
[ {dept:int, deptName:string|null, numEmps:int }* ]
Note how null is retained in the output type (since there may be
employees without a department, then Head may be applied to an
empty list returning null, and the selection of name of null returns null). For instance suppose to pipe the Jaql grouping defined
in the introduction into the following Jaql expression, in order to
produce a printable representation of the records of the result
5.3 Examples
To show how we use the encoding, let us encode the example of
the introduction. For the sake of the concision we will use filter
definitions (rather than expanding them in details). We use Fil
and Sel defined in the introduction, Expand and Transform[]
defined at the beginning of the section, the encoding of Jaql’s field
selection as defined in Section 3.3, and finally Head that returns the
first element of a sequence and a family of recursive filters Rgrpi
with i ∈ N+ both defined below:
transform each x (
(x.deptName)@":"@(to_string x.dep)@":"@(x.numEmps))
where @ denotes string concatenation and to_string is a conversion operator (from any type to string). The composition is ill-typed
for three reasons: the field dept is misspelled as dep, x.numEmps
is of type int (so it must be applied to to_string before concatenation), and the programmer did not account for the fact that
the value stored in the field deptName may be null. The encoding
produces the following lines to be appended to the previous code:
let filter Head = ‘nil => null | (x,xs) => x
let filter Rgrpi = ‘nil => ‘nil
| ((i,x),tail) => (x , Rgrpi tail)
| _ => Rgrpi tail
Then, the query in the introduction is encoded as follows
1
2
3
4
5
6
7
8
9
10
11
[employees depts];
[Sel Fil];
[Transform[x =>(1,x)] Transform[x =>(2,x)]];
Expand;
groupby ( (1,$)=>$.dept | (2,$)=>$.depid );
Transform[(g,l)=>(
[(l; Rgrp1) (l; Rgrp2)];
[es ds] =>
{ dept: g,
deptName: (ds ; Head).name),
numEmps: count(es) } )]
12
13
Transform[ x =>
(x.deptName)@":"@(to_string x.dep)@":"@(x.numEmps)]
in which all the three errors are detected by our type system. A
subtler example of error is given by the following alternative code
Transform[
{ dept : d, deptName: n&String, numEmps: e } =>
14
n @ ":" @ (to_string d) @ ":" @ (to_string e)
15 | { deptName: null, .. } => ""
16 | _ => "Invalid department" ]
12
13
which corrects all the previous errors but adds a new one since, as
detected by our type system, the last branch can be never selected.
As we can see, our type-system ensures soundness, forcing the programmer to handle exceptional situations (as in the null example
above) but is also precise enough to detect that some code paths
can never be reached.
In order to focus on our contributions we kept the language of
types and filters simple. However there already exists several contributions on the types and expressions used here. Two in particular
are worth mentioning in this context: recursive patterns and XML.
Definition 3 defines patterns inductively but, alternatively, we
can consider the (possibly infinite) regular trees coinductively generated by these productions and, on the lines of what is done in
CDuce, use the recursive patterns so obtained to encode regular
expressions patterns (see [4]). Although this does not enhance ex-
In words, we perform the selection on employees and filter the
departments (lines 1-2); we tag each element by 1 if it comes from
employees, and by 2 if it comes from departments (line 3); we
merge the two collections (line 4); we group the heterogeneous list
according to the corresponding key (line 5); then for each element
of the result of grouping we capture in g the key (line 6), split the
group into employees and depts (line 7), capture each subgroup into
the corresponding variable (ie, es and ds) (line 8) and return the
expression specified in the query after the “into” (lines 8-10). The
general definition of the encoding for the co-grouping is given in
Appendix H.
11
2013/3/8
and this latter, with the syntactic sugar defined for CDuce, can be
written as:
pressiveness, it greatly improves the writing of programs since it
makes it possible to capture distinct subsequences of a sequence by
a single match. For instance, when a sequence is matched against
a pattern such as [ (int as x | bool as y | _)* ], then x
captures (the list of) all integer elements (capture variables in regular expression patterns are bound to lists), y captures all Boolean
elements, while the remaining elements are ignored. By such patterns, co-grouping can be encoded without the Rgrp. For instance,
the transform in lines 6-11 can be more compactly rendered as:
6
7
8
9
<product system="US-size">[
<number>[ 557 ]
<name>[ Blouse ]
]
Clearly in our system there are no restrictions in merging and
nesting JSON and XML and no further extension is required to our
system to define XML query and processing expressions. Just, the
introduction of syntactic sugar to make the expressions readable,
seems helpful:
e ::= <e e>e
f ::= <f f >f
Transform[(g,[ ((1,es)|(2,ds))* ]) =>
{ dept: g,
deptName: (ds;Head).name,
numEmps: count(es) }]
For what concerns XML, the types used here were originally defined for XML, so it comes as a no surprise that they can seamlessly
express XML types and values. For example CDuce uses the very
same types used here to encode both XML types and elements as
triples, the first element being the tag, the second a record representing attributes, and the third a heterogeneous sequence for the
content of the element. Furthermore, we can adapt the results of [9]
to encode forward XPath queries in filters. Therefore, it requires
little effort to use the filters presented here to encode languages
such as JSONiq [12] designed to integrate JSON and XML, or to
precisely type regular expressions, the import/export of XML data,
or XPath queries embedded in Jaql programs. This is shown in the
section that follows.
The system we introduced here is already able to reproduce (and
type) the same transformations as in JSONiq, but without the restrictions and drawbacks of the latter (this is why we argue that it
is better to extend NoSQL languages with XML primitives directly
derived from our system rather than to use our system to encode
JSONiq). For instance, the example given in the JSONiq draft to
show how to render in Xhtml the following JSON data:
{
"col labels" : ["singular", "plural"],
"row labels" : ["1p", "2p", "3p"],
"data" :
[
["spinne", "spinnen"],
["spinnst", "spinnt"],
["spinnt", "spinnen"]
]
6. JSON, XML, Regex
There exist various attempts to integrate JSON and XML. For
instance JSONiq [12] is a query language designed to allow XML
and JSON to be used in the same query. The motivation is that
JSON and XML are both widely used for data interchange on the
Internet. In many applications, JSON is replacing XML in Web
Service APIs and data feeds, while more and more applications
support both formats. More precisely, JSONiq embeds JSON into
an XML query language (XQuery), but it does it in a stratified way:
JSONiq does not allow XML nodes to contain JSON objects and
arrays. The result is thus similar to OCamlDuce, the embedding of
CDuce’s XML types and expressions into OCaml, with the same
drawbacks.
Our type system is derived from the type system of CDuce,
whereas the theory of filters was originally designed to use CDuce
as an host language. As a consequence XML types and expressions
can be seamlessly integrated in the work presented here, without
any particular restriction. To that end it suffices to use for XML
elements and types the same encoding used in the implementation
of CDuce, where an XML element is just a triple formed by a tag
(here, an expression), a record (whose labels are the attributes of
the element), and a sequence (of characters and or other XML elements) denoting its content. So for instance the following element
}
can be encoded in the filters presented in this work (with the new
syntactic sugar) as:
{ "col labels" : cl ,
"row labels" : rl ,
"data" : dl
} =>
<table border="1" cellpadding="1" cellspacing="2">[
<tr>[ <th>[ ] !(cl; Transform[ x -> <th>x ]) ]
!(rl; Transform[ h ->
<tr>[ <th>h !(dl; Transform[ x -> <td>x ]) ]
]
)
]
(where ! expands a subsequence in the containing sequence). The
resulting Xhtml document is rendered in a web browser as:
<product system="US-size">
<number>557</number>
<name>Blouse</name>
</product>
Similarly, Jaql built-in libraries include functions to convert and
manipulate XML data. So for example as it is possible in Jaql to
embed SQL queries, so it is possible to evaluate XPath expressions, by the function xpath() which takes two arguments, an
XML document and a string containing an xpath expression—eg,
xpath(read(seq(("conf/addrs.xml")) , "content/city")
—. Filters can encode forward XPath expressions (see [23]) and
precisely type them. So while in the current implementation there
is no check of the type of the result of an external query (nor for
is encoded by the following triple:
("product" , { system : "US-size" } ,
[
("number" , {} , [ 557 ])
("name", {}, "Blouse")
]
)
12
2013/3/8
tegers. An even greater precision can be obtained for grouping expressions when the generation of the key is performed by a filter
that discriminates on types: the result type can keep a precise correspondence between keys and the corresponding groups. As an
example consider the following (extended) Jaql grouping expression:
XPath or for SQL) and the XML document is produced independently from Jaql, by encoding (forward) XPath into filters we can
not only precisely type calls to Jaql’s xpath() function but also
feed them with documents produced by Jaql expressions.
Finally, the very same regular expressions types that are used
to describe heterogeneous sequences and, in particular, the content of XML elements, can be used to type regular expressions.
Functions working on regular expressions (regex for short) form,
in practice, yet another domain specific language that is embedded
in general purpose languages in an untyped or weakly typed (typically, every result is of type string) way. Recursive patterns can
straightforwardly encode regexp matching. Therefore, by combining the pattern filter with other filters it is possible to encode any
regexp library, with the important advantage that, as stated by Theorem 4, the set of values (respectively, strings) accepted by a pattern
(respectively, by a regular expression), can be precisely computed
and can be expressed by a type. So a function such as the built-in
Jaql’s function regex_extract(), which extracts the substrings
that match a given regex, can be easily implemented by filters and
precisely typed by our typing rules. Typing will then amount to
intersect the type of the string (which can be more precise than
just string) with the type accepted by the pattern that encodes the
regex at issue.
group e by({town: ("Roma"|"Pisa"), ..} => "Italia"
|{town: "Paris", ..} => "France"
| _ => "?")
if e has type [{town:string, addr:string}*], then the
type inferred by our system for this groupby expression is
[( ("Italia", [{town:"Roma"|"Pisa", addr:string}+])
|("France", [{town:"Paris", addr:string}+])
|("?", [{town:string/("Roma"|"Pisa"|"Paris"),
addr:string}+])
)*]
which precisely associates every key to the type of the elements it
groups.
Finally, in order to allow a modular usage of filters, adding just
filter application to the expression of the foreign language does not
suffice: parametric filter definitions are also needed.
e ::= f e | let filter F [F1 , . . . , Fn ] = f in e
7. Programming with filters
However, on the line of what we already said about the disruption of
the execution model, recursive parametric filter definitions should
be probably disallowed, since a compilation according to a mapreduce model would require to disentangle recursive calls.
Up to now we have used the filters to encode operators hardcoded in some languages in order to type them. Of course, it is
possible to embed the typing technology we introduced directly
into the compilers of these language so as to obtain the flexible
typing that characterizes our system. However, an important aspect
of filters we have ignored so far is that they can be used directly by
the programmer to define user-defined operators that are typed as
precisely as the hard-coded ones. Therefore a possibility is extend
existing NoSQL languages by adding to their expressions the filter
application f e expression.
The next problem is to decide how far to go in the definition
of filters. A complete integration, that is taking for f all the definitions given so far, is conceivable but might disrupt the execution
model of the host language, since the user could then define complex iterators that do not fit map-reduce or the chosen distributed
compilation policy. A good compromise could be to add to the host
language only filters which have “local” effects, thus avoiding to
affect the map-reduce or distributed compilation execution model.
The minimal solution consists in choosing just the filters for patterns, unions, and expressions:
8. Commentaries
Finally, let us explain some subtler design choices for our system.
Filter design: The reader may wonder whether products and
record filters are really necessary since, at first sight, the filter
( f1 ,f2) could be encoded as (x, y) ⇒ (f1 x, f2 y) and similarly
for records. The point is that f1 x and f2 y are expressions —and
thus their pair is a filter— only if the fi ’s are closed (ie, wihtout
free term recursion variables). Without an explicit product filter it
would not be possible to program a filter as simple as the identity
map, µ X.. ‘nil ⇒ ‘nil||(h, t) ⇒ ( h,Xt)) since Xt is not an expression (X is a free term recursion variable). Similarly, we need
an explicit record filter to process recursively defined record types
such as µ X..({{head:int, tail:X}}| ‘nil).
Likewise, one can wonder why we put in filters only the “open”
record variant that copy extra fields and not the closed one. The
reason is that if we want a filter to be applied only to records with
exactly the fields specified in the filter, then this can be simply
obtained by a pattern matching. So the filter { ℓ1 :f1 , . . . , ℓn :fn}
(ie, without the trailing “. .”) can be simply introduced as syntactic
sugar for { ℓ1 :any, . . . , ℓn :any}} ⇒ { ℓ1 :f1 , . . . , ℓn :fn , ..}}
f ::= e | p => f | f |f
Adding such filters to Jaql (we use the “=>” arrow for patterns in
order to avoid confusion with Jaql’s “->” pipe operator) would not
allow the user to define powerful operators, but their use would already dramatically improve type precision. For instance we could
define the following Jaql expression
transform ( { a:x,..}} as y => { y.*, sum:x+x}} | y => y )
(with the convention that a filter occurring as an expression denotes its application to the current argument $ ). With this syntax,
our inference system is able to deduce that feeding this expression
} *] returns a result
with an argument of type [{{a?:int, c:bool}
} | { c:bool}} )*]. This
of type [({{ a:int, c:bool, sum:int}
precision comes from the capacity of our inference system to discriminate between the two branches of the filter and deduce that
a sum field will be added only if the a field is present. Similarly
by using pattern matching in a Jaql “filter” expression, we can
deduce that filter ( int => true | _ => false ) fed with any
sequence of elements always returns a (possibly empty) list of in-
Constructors: The syntax for constructing records and pairs is
exactly the same in patterns, types, expressions, and filters. The
reader may wonder why we did not distinguish them by using, say,
× for product types or = instead of : in record values. This, combined with the fact that values and singletons have the same syntax,
is a critical design choice that greatly reduces the confusion in these
languages, since it makes it possible to have a unique representation
Syntactically we could write µ X..‘nil ⇒ ‘nil||(h, t) ⇒ ( h,(Xt)v))
where v is any value, but then this would not pass type-checking since the
expression (Xt)v must be typeable without knowing the ∆ environment
(cf. rule [F ILTER -A PP ] at the beginning of Section 4.3). We purposedly
stratified the system in order to avoid mutual recursion between filters and
expressions.
13
2013/3/8
9. Related Work
for constructions that are semantically equivalent. Consider for instance the pattern ( x,,( 3,, ‘nil))) . With our syntax ( 3,, ‘nil)) denotes
both the product type of two singletons 3 and ‘nil, or the value
( 3,, ‘nil)), or the singleton that contains this value. According to
the interpretation we choose, the pattern can then be interpreted as
a pattern that matches a product or a pattern that matches a value.
If we had differentiated the syntax of singletons from that of values
(eg, {v}) and that of pairs from products, then the pattern above
could have been written in five different ways. The point is that
they all would match exactly the same sets of values, which is why
we chose to have the same syntax for all of them.
In the (nested) relational (and SQL) context, many works have
studied the integration of (nested)-relational algebra or SQL into
general purpose programming languages. Among the first attempts
was the integration of the relational model in Pascal [31] or in
Smalltalk [11]. Also, monads or comprehensions [7, 33, 34] have
been successfully used to design and implement query languages
including a way to embed queries within host languages. Significant efforts have been done to equip those languages with type
systems and type checking disciplines [8, 25, 26] and more recently [27] for integration and typing aspects. However, these approaches only support homogeneous sequences of records in the
context of specific classes of queries (practically equivalent to
a nested relational algebra or calculus), they do not account for
records with computable labels, and therefore they are not easily
transposable to a setting where sequences are heterogeneous, data
are semi-structured, and queries are much more expressive.
While the present work is inspired and stems from previous
works on the XML iterators, targeting NoSQL languages made the
filter calculus presented here substantially different from the one
of [9, 23] (dubbed XML filters in what follows), as well in syntax
as in dynamic and static semantics. In [9] XML filters behave as
some kind of top-down tree transducers, termination is enforced by
heavy syntactic restrictions, and a less constrained use of the composition makes type inference challenging and requires sometimes
cumbersome type annotations. While XML filters are allowed to
operate by composition on the result of a recursive call (and, thus,
simulate bottom-up tree transformations), the absence of explicit
arguments in recursive calls makes programs understandable only
to well-trained programmers. In contrast, the main focus of the
current work was to make programs immediately intelligible to any
functional programmer and make filters effective for the typing
of sequence transformations: sequence iteration, element filtering,
one-level flattening. The last two are especially difficult to write
with XML filters (and require type annotations). Also, the integration of filters with record types (absent in [9] and just sketched in
[23]) is novel and much needed to encode JSON transformations.
Record types: The definition of records is redundant (both for
types and patterns). Instead of the current definition we could
have used just { ℓ:t}}, {} , and { ..}}, since the rest can be encoded by intersections. For instance, { ℓ1 :t1 , . . . , ℓn :tn , ..
..}} =
&{ ℓn :tn}&{ ..}} . We opted to use the redundant defini{ ℓ1 :t}}& ...&
tion for the sake of clarity. In order to type records with computed
labels we distinguished two cases according to whether the type of
a record label is finite or not. Although such a distinction is simple,
it is not unrealistic. Labels with singleton types cover the (most
common) case of records with statically fixed labels. The dynamic
choice of a label from a statically known list of labels is a usage
pattern seen in JavaScript when building an object which must conform to some interface based on a run-time value. Labels with infinite types cover the fairly common usage scenario in which records
are used as dictionaries: we deduce for the expression computing
the label the type string, thus forcing the programmer to insert
some code that checks that the label is present before accessing it.
The rationale behind the typing of records was twofold. First
and foremost, in this work we wanted to avoid type annotations
at all costs (since there is not even a notion of schema for JSON
records and collections —only the notion of basic type is defined—
we cannot expect the Jaql programmer to put any kind of type
information in the code). More sophisticated type systems, such
as dependent types, would probably preclude type reconstruction:
dependent types need a lot of annotations and this does not fit our
requirements. Second, we wanted the type-system to be simple
yet precise. Making the finite/infinite distinction increases typing
precision at no cost (we do not need any extra machinery since
we already have singleton types). Adding heuristics or complex
analysis just to gain some precision on records would have blurred
the main focus of our paper, which is not on typing records but
on typing transformations on records. We leave such additions for
future work.
10. Conclusion
Our work addresses two very practical problems, namely the typing of NoSQL languages and a comprehensive definition of their
semantics. These languages add to list comprehension and SQL
operators the ability to work on heterogeneous data sets and are
based on JSON (instead of tuples). Typing precisely each of these
features using the best techniques of the literature would probably
yield quite a complex type-system (mixing row polymorphism for
records, parametric polymorphism, some form of dependent typing,...) and we are skeptical that this could be achieved without using any explicit type annotation. Therefore we explored the formalization of these languages from scratch, by defining a calculus and
a type system. The thesis we defended is that all operations typical
of current NoSQL languages, as long as they operate structurally
(ie, without resorting on term equality or relations), amount to a
combination of more basic bricks: our filters. On the structural side,
the claim is that combining recursive records and pairs by unions,
intersections, and negations suffices to capture all possible structuring of data, covering a palette ranging from comprehensions, to
heterogeneous lists mixing typed and untyped data, through regular
expressions types and XML schemas. Therefore, our calculus not
only provides a simple way to give a formal semantics to, reciprocally compare, and combine operators of different NoSQL languages, but also offers a means to equip these languages, in they
current definition (ie, without any type definition or annotation),
with precise type inference.
Record polymorphism: The type-oriented reader will have noticed that we do not use row variables to type records, and nevertheless we have a high degree of polymorphism. Row variables
are useful to type functions or transformations since they can keep
track of record fields that are not modified by the transformation. In
this setting we do not need them since we do not type transformations (ie, filters) but just the application of transformations (filters
are not first-class terms). We have polymorphic typing via filters
(see how the first example given in Section 7 keeps track of the c
field) and therefore open records suffice.
Record selection: Some languages —typically, the dynamic ones
such as Javascript, Ruby, Python— allow the label of a field selection to be computed by an expression. We considered the definition
of a fine-grained rule to type expressions of the form e1. e2 : whenever e2 is typed by a finite unions of strings, the rule would give
a finite approximation of the type of the selection. However, such
an extension would complex the definition of the type system, just
to handle few interesting cases in which a finite union type can be
deduced. Therefore, we preferred to omit its study and leave it for
future work.
14
2013/3/8
As such we accounted for both components that, according to
Landin, constitute the design of a language: operators and data
structures. But while Landin considers the design of terms and
types as independent activities we, on the contrary, advocate an
approach in which the design of former is driven by the form of
latter. Although other approaches are possible, we tried to convey
the idea that this approach is nevertheless the only one that yields
a type system whose precision, that we demonstrated all the work
long, is comparable only to the precision obtained with hard-coded
(as opposed to user-defined) operators. As such, our type inference
yields and surpasses in precision systems using parametric polymorphism and row variables. The price to pay is that transformations are not first class: we do not type filters but just their applications. However, this seems an advantageous deal in the world
of NoSQL languages where “selects” are never passed around (at
least, not explicitly), but early error detection is critical, especially
in the view of the cost of code deployment.
The result are filters, a set of untyped terms that can be easily
included in a host language to complement in a typeful framework
existing operators with user-defined ones. The requirements to include filters into a host language are so minimal that every modern
typed programming language satisfies them. The interest resides
not in the fact that we can add filter applications to any language,
rather that filters can be used to define a smooth integration of calls
to domain specific languages (eg, SQL, XPath, Pig, Regex) into
general purpose ones (eg, Java, C#, Python, OCaml) so as both can
share the same set of values and the same typing discipline. Likewise, even though filters provide an early prototyping platform for
queries, they cannot currently be used as a final compilation stage
for NoSQL languages: their operations rely on a Lisp-like encoding of sequences and this makes the correspondence with optimized
bulk operations on lists awkward. Whether we can derive an efficient compilation from filters to map-reduce (recovering the bulk
semantics of the high-level language) is a challenging question.
Future plans include practical experimentation of our technique:
we intend to benchmark our type analysis against existing collections of Jaql programs, gauge the amount of code that is ill typed
and verify on this how frequently the programmer adopted defensive programming to cope with the potential type errors.
[6] S. Boag, D. Chamberlain, M. F. Fernández, D. Florescu, J. Robie, and
J. Siméon. XQuery 1.0: An XML query language, W3C recommendation, 2007.
[7] P. Buneman, L. Libkin, D. Suciu, V. Tannen, and L. Wong. Comprehension syntax. SIGMOD Record, 23(1):87–96, 1994.
[8] P. Buneman, R. Nikhil, and R. Frankel. A Practical Functional Programming System for Databases. In Proc. Conference on Functional
Programming and Architecture. ACM, 1981. 1981.
[9] G. Castagna and K. Nguyễn. Typed iterators for XML. In Proceeding
of the 13th ACM SIGPLAN international conference on Functional
programming, ICFP ’08, pages 15–26, New York, NY, USA, 2008.
ACM.
[10] H. Comon, M. Dauchet, R. Gilleron, F. Jacquemard, C. Löding,
D. Lugiez, S. Tison, and M. Tommasi. Tree automata techniques
and applications. http://www.grappa.univ-lille3.fr/tata, 2007.
[11] G. Copeland and D. Maier. Making smalltalk a database system. In
In Proceedings of the ACM SIGMOD International Conference on
Management of Data, pages 316–325, 1984.
[12] J. R. (editor). Jsoniq. http://jsoniq.org.
[13] J. Engelfriet. Bottom-up and Top-down Tree Transformations – A
comparison. Theory of Computing Systems, 9(2):198–231, 1975.
[14] J. Engelfriet. Top-down tree transducers with regular look-ahead.
Mathematical Systems Theory, 10(1):289–303, Dec. 1976.
[15] A. Frisch. Théorie, conception et réalisation d’un langage de programmation fonctionnel adapté à XML. PhD thesis, Université Paris
7 Denis Diderot, 2004.
[16] A. Frisch, G. Castagna, and V. Benzaken. Semantic subtyping: Dealing set-theoretically with function, union, intersection, and negation
types. Journal of the ACM, 55(4):1–64, 2008.
[17] H. Hosoya. Foundations of XML Processing: The Tree Automata
Approach. Cambridge University Press, 2010.
[18] Jaql. http://code.google.com/p/jaql.
[19] Javascript object notation (json). http://json.org/.
[20] W. Martens and F. Neven. Typechecking top-down uniform unranked
tree transducers. In Proceedings of the 9th International Conference
on Database Theory, ICDT ’03, pages 64–78, London, UK, 2002.
Springer-Verlag.
[21] E. Meijer. The world according to LINQ. ACM Queue, 9(8):60, 2011.
References
[22] E. Meijer and G. Bierman. A co-relational model of data for large
shared data banks. Communications of the ACM, 54(4):49–58, 2011.
[1] N. Alon, T. Milo, F. Neven, D. Suciu, and V. Vianu. XML with
data values: typechecking revisited. In Proceedings of the twentieth ACM SIGMOD-SIGACT-SIGART symposium on Principles of
database systems, PODS ’01, pages 138–149, New York, NY, USA,
2001. ACM.
[23] K. Nguyễn. Language of Combinators for XML: Conception, Typing,
Implementation. PhD thesis, Université Paris-Sud 11, http://www.lri.fr/~kn/file
2008.
[24] Odata. http://www.odata.org/.
[2] N. Alon, T. Milo, F. Neven, D. Suciu, and V. Vianu. Typechecking
XML views of relational databases. ACM Trans. Comput. Logic,
4:315–354, July 2003.
[25] A. Ohori and P. Buneman. Type Inference in a Database Programming Language. In ACM Conference on LISP and Functional Programming, pages 174–183, 1988.
[3] A. Behm, V. R. Borkar, M. J. Carey, R. Grover, C. Li, N. Onose,
R. Vernica, A. Deutsch, Y. Papakonstantinou, and V. J. Tsotras. Asterix: towards a scalable, semistructured data platform for evolvingworld models. Distributed and Parallel Databases, 29(3):185–216,
2011.
[26] A. Ohori, P. Buneman, and V. Tannen. Database Programming in
Machiavelli -a Polymorphic Language with Static Type Inference.
Proc. ACM SIGMOD Conference, pages 46–57, 1989.
[27] A. Ohori and K. Ueno. Making standard ML a practical database programming language. In M. M. T. Chakravarty, Z. Hu, and O. Danvy,
editors, ICFP, pages 307–319. ACM, 2011.
[4] V. Benzaken, G. Castagna, and A. Frisch. CDuce: an XML-friendly
general purpose language. In ICFP ’03, 8th ACM International Conference on Functional Programming, pages 51–63, Uppsala, Sweden,
2003. ACM Press.
[28] C. Olston, B. Reed, U. Srivastava, R. Kumar, and A. Tomkins. Pig
latin: a not-so-foreign language for data processing. In SIGMOD
Conference, pages 1099–1110, 2008.
[5] K. S. Beyer, V. Ercegovac, R. Gemulla, A. Balmin, M. Y. Eltabakh,
C.-C. Kanne, F. Özcan, and E. J. Shekita. Jaql: A scripting language
for large scale semistructured data analysis. PVLDB, 4(12):1272–
1283, 2011.
[29] F. Özcan, D. Hoa, K. S. Beyer, A. Balmin, C. J. Liu, and Y. Li. Emerging trends in the enterprise data analytics: connecting Hadoop and
DB2 warehouse. In Proceedings of the 2011 international conference
on Management of data, SIGMOD ’11, pages 1161–1164, 2011.
Only filter-encoded operators are not first class: if the host language provides, say, higher-order functions, then they stay higher-order and are typed
by embedding the host type system, if any, via “foreign type calls”.
[30] A. Sabry and P. Wadler. A reflection on call-by-value. In Proceedings
of the first ACM SIGPLAN international conference on Functional
15
2013/3/8
programming, ICFP ’96, pages 13–24. ACM, 1996.
[31] J. Schmidt and M. Mall. Pascal/R Report. Technical Report 66,
Fachbereich Informatik, université de Hamburg, 1980.
[32] Squeryl: A Scala ORM and DSL for talking with Databases with
minimum verbosity and maximum type safety
.
http://squeryl.org/.
[33] V. Tannen, P. Buneman, and L. Wong. Naturally embedded query
languages. In ICDT, pages 140–154, 1992.
[34] P. Trinder and P. Wadler. Improving list comprehension database
queries. In Fourth IEEE Region 10 Conference (TENCON), pages
186–192, Nov. 1989.
[35] Unql. http://www.unqlspec.org/.
16
2013/3/8
Appendix
17
2013/3/8
A. Termination analysis algorithm
1. It builds a substitution σ : Vars → Ids from expression
variables to variables identifiers, thus associating each capture
variable occurring in a pattern of the filter to a fresh identifier
for the abstract subtree of the input type it will be bound to.
In order to deduce the result type of the application of a filter to
an expression, the type inference algorithm abstractly executes the
filter on the type of the expression. As explained in Section 4, the
algorithm essentially analyzes what may happen to the original
input type (and to its subtrees) after a recursive call and checks
that, in all possible cases, every subsequent recursive call will
only be applied to subtrees of the original input type. In order to
track the subtrees of the original input type, we use an infinite set
Ids = {ι1 , ι2 , . . .} of variable identifiers (ranged over by, possible
indexed, ι). These identifiers are used to identify the subtrees of the
original input type that are bound to some variable after a recursive
call. Consider for instance the recursive filter:
2. It uses the substitution σ to compute the set of symbolic arguments of recursive calls of X.
In other words, if n recursive calls of X occur in the filter f ,
then TreesX (σ, f ) returns the set {A1 , ..., An } of the n symbolic
arguments of these calls, obtained under the hypothesis that the free
variables of f are associated to subtrees as specified by σ. The
formal definition is as follows:
TreesX (σ, e)
= ∅
S
TreesX (σ, p ⇒ f ) = TreesX (σ ∪
{xi 7→ ιi }, f ) (ιi fresh)
µ X..( x1,( x2, x3)) ⇒ X((x1 ,x3) | _ ⇒ ‘nil.
xi ∈vars(p)
TreesX (σ, ( f1 ,f2) ) = TreesX (σ, f1 ) ∪ TreesX (σ, f2 )
TreesX (σ, f1| f2 ) = TreesX (σ, f1 ) ∪ TreesX (σ, f2 )
TreesX (σ, µ X.. f ) = ∅
TreesX (σ, µ Y . f ) = TreesX (σ, f )
(X 6= Y )
TreesX (σ, Xa)
= {aσ}
TreesX (σ, f1 ;f2 ) = TreesX (σ, f2 )
TreesX (σ, of )
= TreesXS
(σ, f ) (o = groupby , orderby )
TreesX (σ, fi )
TreesX (σ, { ℓi :fi , ..}} i∈I ) =
The algorithm records that at the first call of this filter each variable
xi is bound to a subtree ιi of the input type. The recursive call
X((x1 ,x3) is thus applied to the “abstract argument” ( ι1 ,ι3) . If we
perform the substitutions for this call, then we see that x1 will be
bound to ι1 and that x2 and x3 will be bound to two subtrees of
the ι3 subtree. We do not care about x2 since it is not used in
any recursive call. What is important is that both x1 and x3 are
bound to subtrees of the original input type (respectively, to ι1 and
to a subtree of ι3 ) and therefore, for this filter, the type inference
algorithm will terminate for every possible input type.
The previous example introduces the two important concepts
that are still missing for the definition of our analysis:
i∈I
where aσ denotes the application of the substitution σ to a.
The definition above is mostly straightforward. The two important cases are those for the pattern filter where the substitution σ
is updated by associating every capture variable of the pattern with
a fresh variable identifier, and the one for the recursive call where
the symbolic argument is computed by applying σ to the actual
argument.
1. Recursive calls are applied to symbolic arguments, such as
( ι1 ,ι3) , that are obtained from arguments by replacing variables
by variable identifiers. Symbolic arguments are ranged over by
A and are formally defined as follows:
Symb Args
A
::=
|
|
|
|
Second phase. The second phase is implemented by the function
Check(). The intuition is that Check(f ) must “compute” the application of f to all the symbolic arguments collected in the first phase
and then check whether the variables occurring in the arguments
of recursive calls (ie, the “marked” variables) are actually bound
to subtrees of the original type (ie, they are bound either to variable identifiers or to subparts of variable identifiers). If we did not
have filter composition, then this would be a relatively easy task:
it would amount to compute substitutions (by matching symbolic
arguments against patterns), apply them, and finally verify that the
marked variables satisfy the sought property. Unfortunately, in the
case of a composition filter f1 ;f2 the analysis is more complicated
than that. Imagine that we want to check the property for the application of f1 ;f2 to a symbolic argument A. Then to check the
property for the recursive calls occurring in f2 we must compute
(or at least, approximate) the set of the symbolic arguments that
will be produced by the application of f1 to A and that will be thus
fed into f2 to compute the composition. Therefore Check(f ) will
be a function that rather than returning just true or false, it will return a set of symbolic arguments that are the result of the execution
of f , or it will fail if any recursive call does not satisfy the property
for marked variables.
More precisely, the function Check() will have the form
ι
(variable identifiers)
c
(constants)
( A,A))
(pairs)
{ ℓ:A, . . . , ℓ:A}}
(records)
⊥
(indeterminate)
2. We said that we “care” for some variables and disregard others.
When we analyze a recursive filter µ X..f the variables we care
about are those that occur in arguments of recursive calls of X.
Given a filter f and a filter recursion variable X the set of these
variables is formally defined as follows
MarkX (f ) = {x | Xa ⊑ f and x ∈ vars(a)}
where ⊑ denotes the subtree containment relation and vars(e)
is the set of expression variables that occur free in e. With an
abuse of notation we will use vars(p) to denote the capture
variables occurring in p (thus, vars(f e) = vars(f ) ∪ vars(e)
and vars(p ⇒ f ) = vars(f ) \ vars(p), the rest of the definition
being standard).
As an aside notice that the fact that for a given filter f the type
inference algorithm terminates on all possible input types does
not imply that the execution of f terminates on all possible input
values. For instance, our analysis correctly detects that for the filter
µ X..( x1, x2) ⇒ X(( x1 ,x2)| _ ⇒ _ type inference terminates on
all possible input types (by returning the very same input type)
although the application of this same filter never terminates on
arguments of the form ( v1 ,v2) .
We can now formally define the two phases of our analysis
algorithm.
CheckV (σ, f, {A1 , ..., An })
where V ⊆ Vars stores the set of marked variables, f is a filter, σ
is a substitution for (at least) the free variables of f into Ids, and
Ai are symbolic arguments. It will either fail if the marked variables do not satisfy the property of being bound to (subcomponents
of) variable identifiers or return an over-approximation of the result
of applying f to all the Ai under the hypothesis σ. This approximation is either a set of new symbolic arguments, or ⊥. The latter
simply indicates that Check() is not able to compute the result of
the application, typically because it is the result of some expression
First phase. The first phase is implemented by the function
Trees_ (_, _). For every filter recursion variable X, this function
explores a filter and does the following two things:
18
2013/3/8
• CheckV (σ, f, {A1 , ..., An }) =
n
S
CheckV (σ, f, Ai )
i=1
• If (⌉⌉A⌈⌈& * f++ = ∅) then CheckV (σ, f, A) = ∅ otherwise:
CheckV (σ, e, A)
=
(
CheckV (σ, f1 ;f2 , A)
=
CheckV (σ, f2 , CheckV (σ, f1 , A))
CheckV (σ, Xa, A)
=
CheckV (σ, µ X.. f, A)
=
⊥
(
CheckV (σ, f1| f2 , A)
CheckV (σ, ( f1 ,f2) , A)
CheckV (σ, { ℓi :fi , ..}}1≤i≤n , A)
CheckV (σ, p ⇒ f, A)
=
=
=
=
{aσ} if e ≡ a
⊥
otherwise
fail
⊥
if CheckMarkX (f ) (σ, f, A) = fail
otherwise
if * f1 + &⌉ A⌈⌈ = ∅
CheckV (σ, f2 , A)
CheckV (σ, f1 , A)
if ⌉ A⌈⌈ \ * f1+ = ∅
Check (σ, f , A) ∪ Check (σ, f , A) otherwise
V
1
V
2
CheckV (σ, f1 , A1 ) × CheckV (σ, f2 , A2 ) if A ≡ (A1 , A2 )
CheckV (σ, f1 , ι1 ) × CheckV (σ, f2 , ι2 ) if A ≡ ι
(where ι1 and ι2 are fresh)
fail
if A ≡ ⊥
S
{ ℓ1 :B1 , ..., ℓn :Bn , ..., ℓn+k :An+k , ..}} if A ≡ { ℓi :fi , ..}}1≤i≤n+k
Bi ∈CheckS
V (σ,fi ,Ai )
{ ℓ1 :B1 , ..., ℓn :Bn , ..., ℓn+k :An+k , ..}} if A ≡ ι
(where all ιi are fresh)
Bi ∈CheckV (σ,fi ,ιi )
fail
if A ≡ ⊥
(
fail
CheckV (σ ∪ (A V/ p), f, A)
if A V/ p = fail
otherwise
Figure 3. Definition of the second phase of the analysis
belonging to the host language (eg, the application of a function) or
because it is the result of a recursive filter or of a recursive call of
some filter.
The full definition of Check_ (_, _, _) is given in Figure 3. Let us
comment the different cases in detail.
CheckV (σ, f, {A1 , ..., An }) =
n
[
syntactic equivalence), in which case we return it after having
applied the substitution σ to it; or it is some other expression, in
which case the function says that it is not able to compute a result
and returns ⊥.
When the filter is a composition of two filters, f1 ;f2 , we first
call CheckV (σ, f1 , A) to analyze f1 and compute the result of
applying f1 to A, and then we feed this result to the analysis of
f2 :
CheckV (σ, f, Ai )
i=1
simply states that to compute the filter f on a set of symbolic arguments we have to compute f on each argument and return the union
of the results. Of course, if any CheckV (σ, f, Ai ) fails, so does
CheckV (σ, f, {A1 , ..., An }). The next case is straightforward: if
we know that an argument of a given form makes the filter f fail,
then we do not perform any check (the recursive call of f will never
be called) and no result will be returned. For instance, if we apply
the filter { ℓ:f , ..}} to the symbolic argument ( ι1 ,ι2) , then this will
always fail so it is useless to continue checking this case. Formally,
what we do is to consider ⌉ A⌈⌈, the set of all values that have the
form of A (this is quite simply obtained from A by replacing any
for every occurrence of a variable identifier or of ⊥) and check
whether in it there is any value accepted by f , namely, whether the
intersection ⌉ A⌈⌈& * f++ is empty. If it is so, we directly return the
empty set of symbolic arguments, otherwise we have to perform a
fine grained analysis according to the form of the filter.
If the filter is an expression, then there are two possible cases:
either the expression has the form of an argument (≡ denotes
CheckV (σ, f1 ;f2 , A) = CheckV (σ, f2 , CheckV (σ, f1 , A))
When the filter is recursive or a recursive call, then we are not able
to compute the symbolic result. However, in the case of a recursive
filter we must check whether the type inference terminates on it. So
we mark the variables of its recursive calls, check whether its definition passes our static analysis, and return ⊥ only if this analysis
—ie, CheckMarkX (f ) (σ, f, A)— did not fail. Notice that ⊥ is quite
different from failure since it allows us to compute the approximation as long as the filter after the composition does not bind (parts
of) the result that are ⊥ to variables that occur in arguments of recursive calls. A key example is the case of the filter Filter[F ]
defined in Section 5.1. If the parameter filter F is recursive, then
without ⊥ Filter[F ] would be rejected by the static analysis.
Precisely, this is due to the composition F x;(‘false ⇒ Xl|| · · · ).
When F is recursive the analysis supposes that F x produces ⊥,
which is then passed over to the pattern. The recursive call Xl is
executed in the environment ⊥/‘true, which is empty. Therefore
19
2013/3/8
A/x
=
A/((p1, p2)
=
A/p1& p2
=
A/p1| p2
=
A/{{ℓ1 :p1 , ..., ℓn :pn}
A/{{ℓ1 :p1 , ..., ℓn :pn , ..
..}}
A/t
=
=
=
(
{x 7→ A} if A ≡ v or A ≡ ι or x ∈
/V
fail
otherwise
(A1 /p1 ) ∪ (A2 /p2 )
(ι /p ) ∪ (ι /p )
1
1
2
2
fail
Ω
if A ≡ ( A1, A2)
if A ≡ ι
(where ι1 , ι2 are fresh)
if A ≡ ⊥
otherwise
A/p1 ∪ A/p2
(
A/p1 if A/p1 6= Ω
A/p2 otherwise
S
Ai /pi
i≤i≤n
S ι /p
i
i
i≤i≤n
fail
Ω
S
Ai /pi
i≤i≤n
S
ιi /pi
i≤i≤n
fail
Ω
if A ≡ { ℓ1 :A1 , ..., ℓn :An}
if A ≡ ι
(where ιi are fresh)
if A ≡ ⊥
otherwise
if A ≡ { ℓ1 :A1 , ..., ℓn :An , ..., ℓn+k :An+k}
if A ≡ ι
(where ιi are fresh)
if A ≡ ⊥
otherwise
∅
Figure 4. Pattern matching with respect to a set V of marked variables
variables in p it also checks that all capture variables that are also
marked are actually bound to a subtree of the initial input type.
A V/ p is defined in Figure 4 (to enhance readability we omitted the
V index since it matters only in the first case and it does not change
along the definition). The check for marked variables is performed
in the first case of the definition: when a symbolic argument A is
matched against a variable x, if the variable is not marked (x 6∈ V),
then the substitution {x 7→ A} is returned; if it is marked, then the
matching (and thus the analysis) does not fail only if the symbolic
argument is either a value (so it does not depend on the input type)
or it is a variable identifier (so it will be exactly a subtree of the
original input type). The other cases of the definition of A V/ p
are standard. Just notice that in the product and record patterns
when A is a variable identifier ι, the algorithm creates fresh new
variable identifiers to denote the new subtrees in which ι will be
decomposed by the pattern. Finally, the reader should not confound
Ω and fail. The former indicates that the argument does not match
the value, while the latter indicates that a marked variable is not
bound to a value or to exactly a subtree of the initial input type
(notice that the case Ω cannot happen in the definition of Check()
since A V/ p is called only when ⌉ A⌈⌈ is contained in * p++).
the result of the F x call, whatever it is, cannot affect the termination of the subsequent recursive calls. Even though the composition
uses the result of F x, the form of this result cannot be such as to
make typechecking of Filter[F ] diverge.
For the union filter f1| f2 there are three possible cases: f1 will
always fail on A, so we return the result of f2 ; or f2 will never
be executed on A since f1 will never fail on A, so we return just
the result of f1 ; or we cannot tell which one of f1 or f2 will be
executed, and so we return the union of the two results.
If the filter is a product filter ( f1 ,f2) , then its accepted type
*( f1 ,f2)+ is a subtype of ( any,any)). Since we are in the case where
⌉ A⌈⌈&*( f1 ,f2)+ is not empty, then A can have just two forms: either
it is a pair or it is a variable identifier ι. In the former case we check
the filters component-wise and return as result the set-theoretic
product of the results. In the latter case, recall that ι represents a
subtree of the original input type. So if the application does not fail
it means that each subfilter will be applied to a subtree of ι. We
introduce two fresh variable identifiers to denote these subtrees of
ι (which, of course are subtrees of the original input type, too) and,
as in the previous case, apply the check component-wise and return
the set-theoretic product of the results. The case for the record filter
is similar to the case of the product filter.
Finally, the case of pattern filter is where the algorithm checks
that the marked variables are indeed bound to subtrees of the initial
input type. What Check() does is to match the symbolic argument
against the pattern and update the substitution σ so as to check the
subfilter f in an environment with the corresponding assignments
for the capture variables in p. Notice however that the pattern
matching A V/ p receives as extra argument the set V of marked
variables, so that while computing the substitutions for the capture
The two phases together. Finally we put the two phases together. Given a recursive filter µ X..f we run the analysis by marking the variables in the recursive calls of X in f , running the
first phase and then and feeding the result to the second phase:
CheckMarkX (f ) (∅, f, TreesX (∅, f )). If this function does not fail
then we say that µ X.. f passed the check:
µX..f ) ⇐⇒ CheckMarkX (f ) (∅, f, TreesX (∅, f )) 6= fail
Check(µ
20
2013/3/8
We are now able to prove Theorem 16. Let t be a type and f a
filter define
For filters that are not recursive at toplevel it suffices to add a
def
µX.. f ) with X fresh.
dummy recursion: Check(f ) = Check(µ
D(f, t) = {(X, t′ ) | X ⊑ f, t′ ∈ Support(f, t)}
Theorem 16 (Termination). If Check(f ) then the type inference
algorithm terminates for f on every input type t.
Notice that since Support(f, t) is finite and there are finitely many
different recursion variables occuring in a filter, then D(f, t) is
finite, too. Now let f and t be the filter and type mentioned in
the statement of Theorem 16 and consider the (possibly infinite)
derivation of Γ ;∆ ;M ⊢fil f (t) : s. Assign to every judgment
Γ′ ;∆′ ;M ′ ⊢fil f ′ (t′ ) : s′ the following weight
In order to prove the theorem above we need some auxiliary
definitions:
Definition 17 (Plinth [15]). A plinth i ⊂ Types is a set of types
with the following properties:
Wgt(Γ′ ;∆′ ;M ′ ⊢fil f ′ (t′ ) : s′ ) = ( |D(f, t)\dom(M ′ )| , size(f ′ )))
• i is finite
• i contains any, empty and is closed under Boolean connectives
where |S| denotes the cardinality of the set S (notice that in the
definition S is finite), dom(M ′ ) is the domain of M ′ , that is, the set
of pairs (X, t) for which M ′ is defined, and size(f ′ ) is the depth of
the syntax tree of f ′ .
Notice that the set of all weights lexicographically ordered form
a well-founded order. It is then not difficult to prove that every
application of a rule in the derivation of Γ ;∆ ;M ⊢fil f (t) : s
strictly decreases Wgt, and therefore that this derivation must be
finite. This can be proved by case analysis on the applied rule, and
we must distinguish only three cases:
& ,||,¬
¬)
(&
• for all types t = ( t1, t2) in i, t1 ∈ i and t2 ∈ i
• for all types t = { ℓ1 :t1 , . . . , ℓn :tn , (..
..
..)}} in i, ti ∈ i for all
i ∈ [1..n].
We define the plinth of t, noted i(t), as the smallest plinth containing t.
Intuitively, the plinth of t is the set of types that can be obtained
by all possible boolean combination of the subtrees of t. Notice
that i(t) is always defined since our types are regular: they have
finitely many distinct subtrees, which (modulo type equivalence)
can thus be combined in finitely many different ways.
[F IL -F IX ] Notice that in this case the first component of Wgt
either decreases or remains constant, and the second component
strictly decreases.
[F IL -C ALL -N EW ] In this case Lemma 19 ensures that the first
component of the Wgt of the premise strictly decreases. Since
this is the core of the proof let us expand the rule. Somewhere in
the derivation of Γ ;∆ ;M ⊢fil f (t) : s we have the following
rule:
Definition 18 (Extended plinth and support). Let t be a type and
f a filter.WThe extended plinth of t and f , noted î(f, t) is defined
as i(t ∨ v⊑f v) (where v ranges over values).
The support of t and f , noted as Support(f, t), is defined as
[
def
{Aσ | σ : Ids(A) → î(f, t)}
Support(f, t) = î(f, t) ∪
[F IL -C ALL -N EW ]
Γ′ ;∆′ ;M ′ , ((X, t′′ ) 7→ T ) ⊢fil ∆′ (X)(t′′ ) : t′
A∈Trees∅ (∅,f )
Γ′ ;∆′ ;M ′ ⊢fil (Xa)(s) : µ T..t′
The extended plinth of t and f is the set of types that can be
obtained by all possible boolean combination of the subtrees of t
and of values that occur in the filter f . The intuition underlying the
definition of support of t and f is that it includes all the possible
types of arguments of recursive calls occurring in f , when f is
applied to an argument of type t.
Lastly, let us prove the following technical lemma:
and we want to prove that (D(f, t) \ dom(M ′ )) ) (D(f, t) \
(dom(M ′ ) ∪ (X, t′′ )) and the containment must be strict to
ensure that the measure decreases. First of all notice that
(X, t′′ ) 6∈ dom(M ′ ), since it is a side condition for the application of the rule. So in order to prove that containment is strict
it suffices to prove that (X, t′′ ) ∈ D(t, f ). But this is a consequence of Lemma 19 which ensures that t′′ ∈ Support(f, t),
whence the result.
Lemma 19. Let f be a filter such that Check(f ) holds. Let t be a
type. For every derivation D (finite or infinite) of
Γ ;∆ ;M ⊢fil f (t) : s
,
for every occurrence of the rule [F IL -C ALL -N EW ]
Γ′ ;∆′ ;M ′ , ((X, t′′ ) 7→ T ) ⊢fil ∆′ (X)(t′′ ) : t′
Γ′ ;∆′ ;M ′ ⊢fil (X a)(s) : µ T..t′
t′′ = type(Γ′ , a)
(X, t′′ ) 6∈ dom(M ′ )
T fresh
[F IL -*] With all other rules the first component of Wgt remains
constant, and the second component strictly decreases.
t′′ = type(Γ′ , a)
(X, t′′ ) 6∈ dom(M ′ )
T fresh
in D, for every x ∈ vars(a), Γ′ (x) ∈ î(f, t) (or equivalently,
type(Γ′ , a) ∈ Support(f, t))
Proof. By contradiction. Suppose that there exists an instance of
the rule for which x ∈ vars(a) ∧ x ∈
/ î(f, t). This means that
Γ′ (x) is neither a singleton type occurring in f nor a type obtained
from t by applying left projection, right projection or label selection (⋆). But since Check(f ) holds, x must be bound to either an
identifier or a value v (see Figure 4) during the computation of
CheckMarkX (f ) (∅, f, TreesX (∅, f )). Since identifiers in Check(f )
are only introduced for the input parameter or when performing a
left projection, right projection or label selection of another identifier, this ensure that x is never bound to the result of an expression
whose type is not in î(f, t), which contradicts (⋆).
A.1 Improvements
Although the analysis performed by our algorithm is already fine
grained, it can be further refined in two simple ways. As explained
in Section 4, the algorithm checks that after one step of reduction
the capture variables occurring in recursive calls are bound to subtrees of the initial input type. So a possible improvement consists
to try to check this property for a higher number of steps, too. For
instance consider the filter:
µ X..( ‘nil,,x)) ⇒ X({{ ℓ:x}})
| ( y,,_)) ⇒ X(((‘nil,,( y,,y))) )
| _ ⇒ ‘nil
This filter does not pass our analysis since if ιy is the identifier
bound to the capture variable y, then when unfolding the recursive
call in the second branch the x in the first recursive call will be
bound to ( ιy , ιy ) . But if we had pursued our abstract execution
one step further we would have seen that ( ιy , ιy ) is not used in a
recursive call and, thus, that type inference terminates. Therefore,
a first improvements is to modify Check() so that it does not stop
21
2013/3/8
type M = (K,V) | (V,V,K)
type V = { var : string } | { lambda2 : (string, string, M) }
type K = { var : string } | { lambda1 : (string, M) }
let filter Eval =
| ( { lambda2 : (x, k, m) }, v , h ) -> m ; Subst[x,v] ; Subst[k,h]; Eval
| ( { lambda1 : (x, m) }, v ) -> m ; Subst[x,v] ; Eval
| x -> x
let filter Subst[c,F] =
| ( Subst[c,F] , Subst[c,F] , Subst[c,F] )
| ( Subst[c,F] , Subst[c,F] )
| { var : c } -> F
| { lambda2 : (x&¬c, k&¬c, m) } -> { lambda2 : (x, k, m;Subst[c,F]) }
| { lambda1 : (x&¬c, m) } -> { lambda1 : (x, m;Subst[c,F]) }
| x -> x
Figure 5. Filter encoding of λcps
just after one step of reduction but tries to go down n steps, with
n determined by some heuristics based on sizes of the filter and of
the input type.
A second and, in our opinion, more promising improvement
is to enhance the precision of the test ⌉ A⌈⌈& * f++ = ∅ in the
definition of Check, that verifies whether the filter f fails on the
given symbolic argument A. In the current definition the only
information we collect about the type of symbolic arguments is
their structure. But further type information, currently unexploited,
is provided by patterns. For instance, in the following (admittedly
stupid) filter
with the following reduction rules (performed at top-level, without
any reduction under context).
(λx.λk.M )V K
(λx.M )V
−→
−→
M [x := V ][k := K]
M [x := V ]
In order to prove Turing completeness it suffices to show how to
encode λcps terms and reductions in our calculus. For the sake
of readability, we use mutually recursive types (rather than their
encoding in µ-types), we use records (though pairs would have
sufficed), we write just the recursion variable X for the filter x ->
Xx, and use ¬t to denote the type any\t. Term productions are
encoded by the recursive types given at the beginning of Figure 5.
Next we define the evaluation filter Eval. In its body it calls the
filter Subst[c,F] which implements the capture free substitution
and, when c denotes a constant, is defined as right below.
It is straightforward to see that the definitions in Figure 5 implement the reduction semantics of CPS terms. Of course the definition above would not pass our termination condition. Indeed while
Subst would be accepted by our algorithm, Eval would fail since
it is not possible to ensure that the recursive calls of Eval will receive from Subst subtrees of the original input type. This is expected: while substitution always terminate they may return trees
that are not subtrees of the original term.
µ X..( int,,x)) ⇒ x
& int ⇒ X(((y,,y)))
| y&
| z⇒z
if in the first pass we associate y with ιy , then we know that ( ιy , ιy )
has type ( int,, int)). If we record this information, then we know
that in the second pass ( ιy , ιy ) will always match the first pattern,
and so it will never be the argument of a further recursive call. In
other words, there is at most one nested recursive call. The solution
is conceptually simple (but yields a cumbersome formalization,
which is why we chose the current formulation) and amounts to
modify Trees(,) so that when it introduces fresh variables it records
their type information with them. It then just suffices to modify the
definition of ⌉ A⌈⌈ so as it is obtained from A by replacing every
occurrence of a variable identifier by its type information (rather
than by any) and the current definition of Check() will do the rest.
C. Proof of subject reduction (Theorem 13)
We first give the proof of Lemma 9, which we restate:
Let f be a filter and v be a value such that v ∈
/ *f +. Then for
every γ, δ, if δ;γ ⊢eval f (v)
r, r ≡ Ω.
B. Proof of Turing completeness (Theorem 7)
Proof. By induction on the derivation of δ;γ ⊢eval f (v)
case analysis on the last rule of the derivation:
In order to prove Turing completeness we show how to define by
filters an evaluator for untyped (call-by-value) λ-calculus. If we allowed recursive calls to occur on the left-hand side of composition,
then the encoding would be straightforward: just implement β and
context reductions. The goal however is to show that the restriction on compositions does not affect expressiveness. To that end
we have to avoid context reductions, since they require recursive
calls before composition. To do so, we first translate λ-terms via
Plotkin’s call-by-value CPS translation and apply Steele and Rabbit’s administrative reductions to them obtaining terms in λcps . The
latter is isomorphic to cbv λ-calculus (see for instance [30]) and
defined as follows.
M
V
K
::=
::=
::=
r, and
(expr): here, *e+ = any for any expression e, therefore this rule
cannot be applied (since ∀v, v ∈ any).
(prod): assume v ∈
/ *((f1 ,f2) + ≡ ( * f1 + , * f2 + ) . then either
v ∈
/ ( any,,any)), in which case only rule (error) applies and
therefore r ≡ Ω. Or v ≡ (v1 , v2 ) and v1 ∈
/ *f1 +. Then by
induction hypothesis on the first premise, δ;γ ⊢eval f1 (v1 )
r1
and r1 ≡ Ω, which contradicts the side condition r1 6= Ω (and
We were a little bit sloppy in the notation and used a filter parameter
as a pattern. Strictly speaking this is not allowed in filters and, for instance, the branch { var : c } -> F in Subst should rather be written as { var : y } -> ( (y=c); (true -> F | false -> { var
: y }) ). Similarly, for the other two cases.
KV | V V K
x | λx.λk.M
k | λx.M
22
2013/3/8
(error): this rule can never occur. Indeed, if v : t and Γ ;∆ ;M ⊢fil
f (t) : s, that means in particular that t ≤ *f + and therefore
that all of the side conditions for the rules prod, pat and recd
hold.
similarly for the second premise). Therefore this rule cannot be
applied to evaluate ( f1 ,f2) .
& *f +
(patt): Similarly to the previous case. If v ∈
/ *p ⇒ f + = *p+&
then either v ∈
/ *p+ (which contradicts v/p 6= Ω) or v ∈
/ *f +
and by induction hypotheis, δ;γ, v/p ⊢eval f (v)
Ω.
(comp): If v ∈
/ *f1 ;f2 + ≡ *f1 +, then by induction hypothesis on
the first premise, r1 ≡ Ω, which contradicts the side condition
r1 6= Ω. Therefore only rule (error) can be applied here.
(recd): Similar to product type: either v is not a record and therefore only the rule (errro) can be applied, or one of the ri = Ω
(by induction hypothesis) which contradicts the side condition.
(union1) and (union2): since v ∈
/ *f1| f2 +, this means that v ∈
/
*f1 + and v ∈
/ *f2 +. Therefore if rule (union1) is chosen, by
induction hypothesis, r1 ≡ Ω which contradicts the side condition. If rule (union2) is chosen, then by induction hypothesis
r2 ≡ Ω which gives r ≡ Ω.
(rec-call) trivially true, since it cannot be that v ∈
/ any
(rec) we can apply straightfowardly the induction hypothesis on
the premise and have that r ≡ Ω.
(groupby) and (orderby): If v ∈
/ f then the only rule that applies
is (error) which gives r ≡ Ω.
D. Complexity of the typing algorithm
Proof. For clarity we restrict ourselves to types without record
constructors but the proof can straihgtforwardly be extended to
them. First remark that our types are isomorphic to alternating tree
automata (ATA) with intersection, union and complement (such
as those defined in [10, 17]). From such an ATA t it is possible
to compute a non-deterministic tree automaton t′ of size O(2|t| ).
When seeing t′ in our type formalism, it is a type generated by the
following grammar:
τ
const
atom
::=
::=
::=
µX.τ | τ||const | const
( τ,,τ)) | atom
X | b
where b ranges over negation and intersections of basic types. Intuitively each recursion variable X is a state in the NTA, a finite
union of products is a set of non-deterministict transitions whose
right-hand-side are each of the products and atom productions are
leaf-states. Then it is clear than if Check(f ) holds, the algorithm
considers at most |f | × |t′ | distinct cases (thanks to the memoization set M ). Furthermore for each rule, we may test a subtyping
problem (eg, t ≤ *f +), which itself is EXPTIME thus giving the
bound
We are now equipped to prove the subject reduction theorem
which we restate in a more general manner:
For every Γ, ∆, M , γ, δ such that ∀x ∈ dom(γ), x ∈ dom(Γ) ∧
γ(x) : Γ(x), if Γ ;∆ ;M ⊢fil f (t) : s, then for all v : t,
δ;γ ⊢eval f (v)
r implies r : s.
The proof is by induction on the derivation of δ;γ ⊢eval f (v)
r, and by case analysis on the rule. Beside the basic case, the only
rules which are not a straightforward application of induction are
the rules union1 and union2 that must be proved simulatneously.
Other cases being proved by a direct application of the induction
hypothesis, we only detail the case of the product constructor.
E. Precise typing of groupby
The process of inferring a precise type for groupby f (t) is decomposed in several steps. First we have to compute the set D(f ) of
discriminant domains for the filter f . The idea is that these domains
are the types on which we know that f may give results of different types. Typically this corresponds to all possible branchings that
the filter f can do. For instance, if {t1 , ..., tn } are pairwise disjoint
types and f is the filter t1 ⇒ e1| · · · | tn ⇒ en , then the set of
discriminant domains for the filter f is {t1 , ..., tn }, since the various si obtained from ∅ ;∅ ;∅ ⊢fil f (ti ) : si may all be different,
and we want to keep track of the relation between a result type si
and the input type ti that produced it. Formally D(f ) is defined as
follows:
D(e)
= {any}
D(((f1 ,f2) )
= D(f1 ) × D(f2 )
D(p ⇒ f )
= {**p + & t | t ∈ D(f )}
D(f1| f2 )
= D(f1 ) ∪ {t\**f1+ | t ∈ D(f2 )}
µX.. f )
D(µ
= D(f )
D(Xa)
= {any}
D(f1 ;f2 )
= D(f1 )
D(of )
= {[any*]}
(o = groupby , orderby )
S
D({{ℓi :fi , ..}}i∈I ) =
{ ℓi :ti , ..}}i∈I
(expr): we suppose that the host languages enjoys subject reduction, hence e(v) = r : s.
W
(prod): Here, we know that t ≡ i≤rank(t) (ti1 , ti2 ). Since v is a
value, and v : t, we obtain that v ≡ (v1 , v2 ). Since (v1 , v2 )
is a value, ∃i ≤ rank(t) such that (v1 , v2 ) : (ti1 , ti2 ). The
observation that v is a value is W
crucial here, since in general
j
j
given a type t′ ≤ t with t′ ≡ j≤rank(t′ ) (t′1 , t′2 ) it is not
j
j
true that ∃i, j s.t. (t′1 , t′2 ) ≤ (t1 i , t2 i ). However this property
holds for singleton types and therefore for values. We have
therefore that v1 : ti1 and v2 : ti1 . Furthermore, since we
suppose that a typing derivation exists and the typing rules are
syntax directed, then Γ ;∆ ;M ⊢fil f (ti1 ) : si1 must be a used
to prove our typing judgement. We can apply the induction
hypothesis and deduce that δ;γ ⊢eval f (v1 )
r1 and similarly
for v2 . We have therefore that
the filter evaluates to (r1 , r2 )
W
which has type (si1 , si2 ) ≤ j≤rank(s) (s1 i , s2 i ) which proves
this case.
ti ∈D(fi )
Now in order to have a high precision in typing we want to
compute the type of the intermediate list of the groupby on a set of
disjoint types. For that we define a normal form of a set of types that
given a set of types computes a new set formed of pairwise disjoint
types whose union covers the union of all the original types.
(union1) and (union2): both rules must be proved together. Indeed, given a filter f1| f2 for which we have a typing derivation
&*
for the judgement Γ ;∆ ;M ⊢fil f1| f2 (t) : s, either v : t&
f1 + and we can apply the induction hypothesis and therefore,
& * f1 +) : s1 and if δ;γ ⊢eval f (v1 )
Γ ;∆ ;M ⊢fil f1| (t&
r1
& * f1 +.
(case union1) r1 : s1 . However it might be that v ∈
/ t&
Then by Lemma 9 we have that δ;γ ⊢eval f (v1 )
Ω (case
union2) and we can apply the induction hypothesis on the second premise, which gives us δ;γ ⊢eval f (v1 )
r2 which allows
us to conclude.
N ({ti | i ∈ S}) =
[
(
\
∅⊂I⊆S i∈I
ti \
[
tj )
j∈S\I
Recall that we are trying to deduce a type for groupby f (t), that
is for the expression groupby f when applied to an argument of
type t. Which types shall we use as input for our filter f to compute
the type of the intermediate result? Surely we want to have all the
23
2013/3/8
where every rule rj ∈ R
types that are in the item(t). This however is not enough since we
would not use the information of the descriminant domains for the
filter. For instance if the filter gives two different result types for
positive and negative numbers and the input is a list of integers,
we want to compute two different result types for positive and
negatives and not just to compute the result of the filter application
on generic integers (indeed item([int*]) = {int}). So the idea
is to add to the set that must be normalized all the types of the set
of discriminant types. This however is not precise enough, since
these domains may be much larger than the item types of the input
list. For this reason we just take the part of the domain types that
intersects at least some possible value of the list. In other terms we
consider the normal form of the following set
_
ti | σi ∈ item(t)}
item(t) ∪ {σi ∧
rj ≡ qi (aj (x1 , . . . , xn )), Dj → bj (qj1 (xjk1 ), . . . , qjm (xjkm ))
is translated into:
(‘aj , [ x1& JDj (x1 )K ... xn& JDj (xn )K ]) ⇒ ‘bj ,[]
if bj ∈ ∆0
(‘aj , [ x1 &JDj (x1 )K ... xn& JDj (xn )K ]) ⇒
( ‘bj ,((xjk1 ;TR(S ′ , qj1 ),(( . . . ,((xjkm ;TR(S ′ , qjm ),‘nil)))))
whereS ′ = S ∪ {qi }
otherwise
The fact that ∀t ∈ dom(T ), JT (t)K ≡ TR(∅, q0 )JtK is proved by a
straightforward induction on t. The only important point to remark
is that since T is deterministic, there is exactly one branch of the
alternate filter f1| . . . | fn that can be selected by pattern matching
for a given input v. For as to why Check(fT ) holds, it is sufficient
to remark that each recursive call is made on a strict subtree of the
input which guarantees that Check(fT ) returns true.
ti ∈D(f )
The idea is then that if {ti }i∈I is the normalized
set of the set
W
above, then the type of grouby is the type [ i∈I ( f (ti ),,[ti +]))*]
with the optimization that we can replace the * by a + if we know
that the input list is not empty.
Lemma 23 (TDTTR 6← filters). Filters are strictly more expressive than TDTTRs.
F. Comparison with top-down tree transducers
We first show that every Top-down tree transducers with regular
look-ahead can be encoded by a filter. We use the definition of [14]
for top-down tree transducers:
Definition 20 (Top-down Tree transducer with regular look-ahead). A top-down tree transducer with regular look-ahead (TDTTR)
is a 5-tuple < Σ, ∆, Q, Qd , R > where Σ is the input alphabet,
∆ is the output alphabet, Q is a set of states, Qd the set of initial
states and R a set of rules of the form:
q(a(x1 , . . . , xn )), D → b(q1 (xi1 ), . . . , qm (xim ))
where a ∈ Σn , b ∈ ∆m , q, q1 ,. . . , qm ∈ Q, ∀j ∈ 1..m, ij ∈ 1..n
and D is a mapping from {x1 , . . . , xn } to 2TΣ (TΣ being the set of
regular tree languages over alphabet Σ). A TDTTR is deterministic
if Qd is a singleton and the left-hand sides of the rules in R are
pairwise disjoint.
Since our filters encode programs, it only make sense to compare filters and deterministic TDTTR. We first show that given such
a deterministic TDTTR we can write an equivalent filter f and, furthermore, that Check(f ) does not fail. First we recall the encoding
of ranked labeled trees into the filter data-model:
Definition 21 (Tree encoding). Let t ∈ TΣ . We define the treeencoding of t and we write JtK the value defined inductively by:
JaK
Ja(t1 , . . . , tn )K
=
=
(‘a,[])
(‘a,[ Jt1 K ... Jtn K ])
∀a ∈ Σ0
∀a ∈ Σn
In particular it is clear than when S is regular, JSK is a type.
∀t ∈ dom(T ), JT (t)K ≡ fT JtK
Proof. The encoding is as follows. For every state qi ∈ Q, we
will introduce a recursion variable Xi . Formally, the translation is
performed by the function TR : 2Q × Q → Filters defined as:
x ⇒ Xi x
µXi .(f1| . . . | fn )
→
→
→
→
→
→
a(q1 (x))
a(a)
a(b)
b(q2 (x))
b(a)
b(b)
where Σ = ∆ = {a, b}, Q = {q0, q1, q2}, Qd = {q0 } and
D = {x 7→ TΣ }. It is however impossible to write a TDTTR that
replaces a leaf with a copy of the whole input, for inputs of arbitrary
size. A similar example is used in [13] to show that TDTTR and
bottom-up tree transducers are not comparable.
Lemma 22 (TDTTR → filters). Let T =< Σ, ∆, Q, {q0 }, R >
be a deterministic TDTTR. There exists a filter fT such that:
=
=
This filter works on monadic trees of the form u1 (. . . un−1 (un ) . . .)
where ui ∈ {a, b}, and essentially replaces the leaf un of a tree
t by a copy of t itself. This cannot be done by a TDTTR. Indeed,
TDTTR have only two ways to “remember” a subtree and copy it.
One is of course by using variables; but their scope is restricted to a
rule and therefore an arbitrary subtree can only be copied at a fixed
distance of its original position. For instance in a rule of the form
q(a(x)), D → a(q1 (x), b(b(q1 (x)))), assuming than q1 copies its
input, the copy of the original subtree is two-levels down from its
next sibling but it cannot be arbitrary far. A second way to copy
a subtree is to remember it using the states. Indeed, states can encode the knowledge that the TDTTR has accumulated along a path.
However, since the number of states is finite, the only thing that a
TDTTR can do is copy a fixed path. For instance for any given n,
there exists a TDTTR that performs the transformation above, for
trees of height n (it has essentially 2n − 1 states which remember
every possible path taken). For instance for n = 2, the TDTTR is:
q0 (a(x)), D
q1 (a()), D
q1 (b()), D
q0 (b(x)), D
q2 (a()), D
q2 (b()), D
where the list notation [ v1 ... vn ] is a short-hand for (v1 ,
..., (vn , ‘nil)). We generalize this encoding to tree languages and types. Let S ⊆ Tσ , we write JSK the set of values
such that ∀t ∈ S, JtK ∈ JSK.
TR(S, qi )
TR(S, qi )
Proof. Even if we restrict filters to have the same domain as TDTTRs (meaning, we used a fixed input alphabet Σ for atomic types)
we can define a filter that cannot be expressed by a TDTTR. For
instance consider the filter:
y ⇒ (µX.((‘a,,‘nil)) ⇒ y
|( ‘b,,‘nil)) ⇒ y
|( ‘a,,[x])) ⇒ (‘a, [ X x ]))
|( ‘b,,[x])) ⇒ (‘b, [ X x ]))
G. Operators on record types.
We use the theory of records defined for CDuce. We summarize
here the main definitions. These are adapted from those given in
Chapter 9 of Alain Frisch’s PhD thesis [15] where the interested
if qi ∈ S
if qi ∈
/S
24
2013/3/8
3. if r2 (ℓ) may be undefined (ie, r2 (ℓ) = t||⊥ for some type t),
then we take the union of the two corresponding fields since
it can results either in r1 (ℓ) or r2 (ℓ) according to whether the
record typed by r2 is undefined in ℓ or not: (r1 + r2 )(ℓ) =
r1 (ℓ)||(r2 (ℓ) \ ⊥).
reader can also find detailed definitions of the semantic interpretation of record types and of the subtyping relation it induces, the
modifications that must be done to the algorithm to decide it, finer
details on pattern matching definition and compilation techniques
for record types and expressions.
Let Z denote some set, a function r : L → Z is quasi-constant
if there exists z ∈ Z such that the set {ℓ∈L | r(ℓ) 6= z} is finite; in
this case we denote this set by dom(r) and the element z by def(r).
We use L _ Z to denote the set of quasi-constant functions from
L to Z and the notation {ℓ1 = z1 , . . . , ℓn = zn , _ = z} to denote
the quasi-constant function r : L _ Z defined by r(ℓi ) = zi for
i = 1..n and r(ℓ) = z for ℓ ∈ L \ {ℓ1 , . . . , ℓn }. Although this
notation is not univocal (unless we require zi 6= z), this is largely
sufficient for the purposes of this section.
Let ⊥ be a distinguished constant, then the sets string _
Types ∪ {⊥} and string _ Values ∪ {⊥} denote the set of
all record types expressions and of all record values, respectively.
The constant ⊥ represents the value of the fields of a record that
are “undefined”. To ease the presentation we use the same notation
both for a constant and the singleton type that contains it: so when
⊥ occurs in L _ Values ∪ {⊥} it denotes a value, while in
string _ Types ∪ {⊥} it denotes the singleton type that contains
only the value ⊥.
Given the definitions above, it is clear that the record types
in Definition 2 are nothing but specific notations for some quasiconstant functions in string _ Types ∪ {⊥}. More precisely,
the open record type expression { ℓ1 :t1 , . . . , ℓn :tn , ..
..}} denotes the
quasi-constant function {ℓ1 = t1 , . . . , ℓn = tn , _ = any} while
the closed record type expression { ℓ1 :t1 , . . . , ℓn :tn} denotes the
quasi-constant function {ℓ1 = t1 , . . . , ℓn = tn , _ = ⊥}. Similarly, the optional field notation { ..., ℓ?:t, ...}} denotes the record
type expressions in which ℓ is mapped either to ⊥ or to the type t,
that is, {..., ℓ = t | ⊥, ...}.
Let t be a type and r1 , r2 two record type expressions, that is
r1 , r2 : string _ Types ∪ {⊥}. The merge of r1 , and r2 with
respect to t, noted ⊕t and used infix, is the record type expression
defined as follows:
&t ≤ empty
def
r1 (ℓ)
if r1 (ℓ)&
(r1 ⊕t r2 )(ℓ) =
(r1 (ℓ) \ t)||r2 (ℓ) otherwise
This explains all the examples we gave in the main text. In particular, { a:int, b:int}} + { a?:bool}} = { a:int||bool, b:int}} since
“a” may be undefined in the right hand-side record while “b” is
undefined in it, and { a:int}} + { ..}} = { ..}}, since “a” in the right
hand-side record is defined (with a 7→ any) and therefore has priority over the corresponding definition in the left hand-side record.
H. Encoding of Co-grouping
As shown in Section 5.2, our groupby operator can encode JaQL’s
group each x = e as y by e′ , where e computes the grouping key,
and for each distinct key, e′ is evaluated in the environment where
x is bound to the key and y is bound to the sequence of elements in
the input having that key. Co-grouping is expressed by:
group
l1 by x = e1 as y1
.
.
.
ln by x = en as yn
into e
Co-grouping is encoded by the following composition of filters:
[ Jl1 K ... Jln K ];
[ Transform[x ⇒ (1, x)] ... Transform[x ⇒ (n, x)] ];
Expand;
groupby (((1,, $)) ⇒ Je1 K | . . . |( n,, $)) ⇒ Jen K);
Transform[ ( x,,l)) ⇒ ([(l;Rgrp1)...(l; Rgrpn)];
[y1 , . . . , yn ] ⇒ JeK )]
where
let filter Rgrpi = ‘nil => ‘nil
| ((i,x),tail) => (x , Rgrpi tail)
| _ => Rgrpi tail
Essentially, the co-grouping encoding takes as argument a sequence
of sequences of values (the n sequences to co-group). Each of
these n sequence, is tagged with an integer i. Then we flatten this
sequences of tagged values. We can on this single sequence apply
our groupby operator and modify each key selector ei so that it is
only applied to a value tagged with integer i. Once this is done, we
obtain a sequence of pairs (k, l) where k is the commen grouping
key and l a sequence of tagged values. We only have to apply
the auxiliary Regrp filter which extracts the n subsequences from
l (tagged with 1..n) and removes the integer tag for each value.
Lastly we can call the recombining expression e which has in scope
x bound to the current grouping key and y1 ,. . . ,yn bound to each of
the sequences of values from input l1 , . . . ,ln whose grouping key
is x.
Recall that by Lemma 11 a record type (ie, a subtype of {..}) is
equivalent to a finite union of record type expressions (ie, quasiconstant functions in string _ Types ∪ {⊥}). So the definition
of merge can be easily extended to all record types as follows
_
_ ′ def _
( r i ) ⊕t (
rj ) =
(ri ⊕t rj′ )
i∈I
j∈J
i∈I,j∈J
Finally, all the operators we used for the typing of records in the
rules of Section 4.2 are defined in terms of the merge operator:
t1 + t2
=
def
t2 ⊕⊥ t1
(1)
t\ℓ
def
{ℓ = ⊥, _ = c0 } ⊕c0 t
(2)
=
where c0 is any constant different from ⊥ (the semantics of the
operator does not depend on the choice of c0 as long as it is different
from ⊥).
Notice in particular that the result of the concatenation of two
record type expressions r1 + r2 may result for each field ℓ in three
different outcomes:
1. if r2 (ℓ) does not contain ⊥ (ie, the field ℓ is surely defined), then
we take the corresponding field of r2 : (r1 + r2 )(ℓ) = r2 (ℓ)
2. if r2 (ℓ) is undefined (ie, r2 (ℓ) = ⊥), then we take the corresponding field of r1 : (r1 + r2 )(ℓ) = r1 (ℓ)
25
2013/3/8
This figure "xhtml.png" is available in "png" format from:
http://arxiv.org/ps/1303.1716v1
| 6cs.PL
|
1
Joint Source-Channel Secrecy Using Uncoded
Schemes: Towards Secure Source Broadcast
Lei Yu, Houqiang Li, Senior Member, IEEE, and Weiping Li, Fellow, IEEE
arXiv:1607.07040v3 [cs.IT] 22 Jul 2017
Abstract
This paper investigates a joint source-channel secrecy problem for the Shannon cipher broadcast system. We
suppose list secrecy is applied, i.e., a wiretapper is allowed to produce a list of reconstruction sequences and the
secrecy is measured by the minimum distortion over the entire list. For discrete communication cases, we propose a
permutation-based uncoded scheme, which cascades a random permutation with a symbol-by-symbol mapping. Using
this scheme, we derive an inner bound for the admissible region of secret key rate, list rate, wiretapper distortion,
and distortions of legitimate users. For the converse part, we easily obtain an outer bound for the admissible region
from an existing result. Comparing the outer bound with the inner bound shows that the proposed scheme is optimal
under certain conditions. Besides, we extend the proposed scheme to the scalar and vector Gaussian communication
scenarios, and characterize the corresponding performance as well. For these two cases, we also propose another
uncoded scheme, orthogonal-transform-based scheme, which achieves the same performance as the permutationbased scheme. Interestingly, by introducing the random permutation or the random orthogonal transform into the
traditional uncoded scheme, the proposed uncoded schemes, on one hand, provide a certain level of secrecy, and on
the other hand, do not lose any performance in terms of the distortions for legitimate users.
Index Terms
Uncoded scheme, secrecy, permutation, orthogonal transform, Shannon cipher system.
I. I NTRODUCTION
Investigations on joint source-channel coding (JSCC) could trace back to Shannon’s pioneering work [1], where
a geometric method was developed to design a communication system. For the JSCC of transmitting a Gaussian
source over a Gaussian broadcast channel, Goblick observed [2] that when the source and channel bandwidths are
matched (i.e., one channel use per source sample), directly sending a scaled version of the source samples on the
channel (i.e., linear scheme) is in fact optimal; while for this case the separation scheme that cascades source coding
with channel coding indeed suffers a performance loss [3]. For vector Gaussian communication cases, the optimal
linear coding was studied in [4]. In general, the schemes that consist of symbol-by-symbol mappings (not limited to
the linear one) are named uncoded schemes. The optimality of uncoded schemes for the general source-channel pair
L. Yu is with the Department of Electrical and Computer Engineering, National University of Singapore, Singapore (e-mail: [email protected]).
This work was done when he was at University of Science and Technology of China. H. Li and W. Li are with the Department of Electronic
Engineering and Information Science, University of Science and Technology of China, Hefei, China (e-mail: {lihq,wpli}@ustc.edu.cn).
2
has been investigated in [3], which showed that the Shannon limit can be achieved by uncoded schemes only when
the source and channel satisfy a certain probabilistic matching condition. To further improve the performance for
mismatched source-channel pairs, the hybrid coding (or hybrid digital-analog coding) has been studied in [5]-[10],
which combines the traditional digital coding and symbol-by-symbol mapping together. As for the converse part of
JSCC problem, Reznic et al. [11] and Tian et al. [12] derived some nontrivial converse results for Gaussian source
broadcast problem. Besides, Yu et al. [9], [10] generalized the achievability and converse results for the Gaussian
communication to the general source-channel case.
On information-theoretic security, the Shannon cipher system (the noisy broadcast version depicted in Fig. 1) was
first investigated in Shannon’s pioneering work [13], where a sender A communicates with a legitimate receiver B
secretly by exploiting a shared secret key. For lossy source communication, wiretapper might only want to decrypt
a lossy version of the source. Schieler et al. [14] studied a distortion-based secrecy measure in the Shannon cipher
system around the assumption that the wiretapper has ability to conduct list decoding with fixed list size, and the
secrecy is measured by the minimum distortion over the entire list. Yu et al. [15] showed that the systems with
this secrecy measure are equivalent to those with secrecy measured by a new quantity lossy-equivocation, which
could be considered as a lossy extension of the traditional equivocation. Hence the list secrecy is closely related
to the traditional equivocation as well. Furthermore, Yu et al. used this secrecy measure to study the problem
of source-channel secrecy for the Shannon cipher system, and showed that for the source-channel pair satisfying
certain conditions, an uncoded scheme could outperform the separate one.
JSCC improves the robustness of communication or the performance of broadcast, while secrecy coding improves
the security of communication by exploiting the secret key and/or the wiretap channel. Therefore, intuitively the
robustness and the security could be obtained simultaneously if we combine JSCC and secrecy coding together.
This joint source-channel secrecy (JSCS) problem has been considered in several works already. Yamamoto in [16]
studied the secure lossy transmission over the noisy wiretap channel with secrecy measured by the wiretapper’s best
reconstruction distortion. However, it is shown in [14] this secrecy measure is cheap and fragile, since only one bit
of secret key suffices to achieve the optimality of secrecy, and meanwhile, only one bit of additional information
for the wiretapper suffices to decrypt this optimal encryption scheme. A different formulation of the problem was
considered in [17], where the authors assumed there is a fixed information leakage to the wiretapper and wish to
minimize the distortion at the legitimate receiver, while at the same time providing a graceful distortion degradation
when there is an SNR (Signal Noise Ratio) mismatch. They showed that, for a positive leakage, this can be achieved
by a hybrid digital-analog coding. This scenario was extended to consider side information at the receiver in [18]
or side information at the sender in [19].
Analog encryption (or analog scrambling) technologies, e.g., sign-change based scheme [20], permutation based
scheme [20] and bandwidth-keeping scheme [21], can be seen as uncoded JSCS schemes as well, although they are
not designed for a specified source-channel pair. Sign-change based scheme improves secrecy by changing the sign
of each sample according to the secret key. But owing to at most one bit secret key used per sample, this scheme
could not provide higher secrecy even with a higher key rate available. The permutation based scheme improves
secrecy by shuffling the positions of samples. Unlike the sign-change based scheme, it supports any arbitrarily high
3
key rate. Furthermore, Kang and Liu [23] recently applied the permutation operation in a digital encryption scheme,
and showed that the permutation is another powerful encryption technique (besides the one-time pad) to achieve
the optimality of secrecy.
A. Contributions
In this paper, we consider the joint source-channel secrecy problem of secure source broadcast in the bandwidthmatched Shannon cipher system (see Fig. 1). The list secrecy [14] is used to measure secrecy, that is, the wiretapper
is allowed to conduct list decoding with fixed list size, and the secrecy is measured by the minimum distortion over
the entire list. We study an achievable region of secret key rate, list rate, wiretapper distortion, and distortions of
all legitimate users and show optimality under certain conditions. Our contributions are as follows:
1) For the discrete source case, we propose a permutation-based uncoded scheme, which cascades a random
permutation with a symbol-by-symbol mapping. Our scheme differs from the permutation based scheme
proposed in [23] in two main aspects: 1) our scheme, coupling a permutation operation with a traditional
uncoded scheme, is designed for the source-channel secrecy problem, however, the scheme in [23] couples
a permutation operation with a digital scheme, and is designed for the source-secrecy coding problem; 2)
in addition to the finite alphabet case, we also extend the scheme to source-channel pairs with countably
infinite alphabets and Gaussian source-channel pairs, which require us to use some more powerful techniques,
including unified typicality, d´tilted information, and geometric analysis. By analyzing the proposed scheme,
we provide an inner bound for the admissible region. For the converse part, we give an outer bound by using
our recent result [15]. Comparing the outer bound with the inner bound shows that the proposed scheme is
optimal under certain conditions.
2) We extend the proposed scheme to scalar and vector bandwidth-matched Gaussian communication scenarios.
For these two cases, we also propose another uncoded scheme, orthogonal-transform-based scheme, which
achieves the same inner bounds as the one achieved by the permutation-based scheme. Interestingly, by
introducing the random permutation or the random orthogonal transform into the traditional uncoded scheme,
the proposed uncoded schemes, no matter for the discrete source case or the Gaussian source-channel case,
on one hand, provide a certain level of secrecy, and on the other hand, do not lose any performance in terms
of the distortions for legitimate users.
Schieler and Cuff [14] studied the list secrecy problem for the noiseless point-to-point1 version of Shannon cipher
system, and showed a digital scheme, in which the secret key is used to choose a source codebook to code the
source sequence, is optimal. For this problem, a separate coding, cascading source coding and one-time pad,
has been proven optimal as well [15]. Yu et al. [15] extended this problem to the noisy channel case, and
showed that the separate strategy (cascading source coding, one-time pad, and channel coding) is suboptimal
in general and a single-letter uncoded scheme could outperform the separate scheme. In this paper we extend
1 Here
system.
the word noiseless means the wiretap channel is noiseless, and the word point-to-point means there is only one legitimate user in the
4
Key
K 2nRK
Y1n
Sn
A
Xn
B1
Sˆ1n
n
PY1Y2 Z | X Y2
Zn
B2
Sˆ2n
Eve
Fig. 1. The Shannon cipher broadcast system.
the problem to noisy broadcast scenarios, and propose two kind of uncoded schemes that adopt two different
encryption strategies—random permutation and random orthogonal-transform (instead of the traditional one-time
pad encryption). We show the proposed uncoded schemes could achieve the optimality under certain cases.
The rest of the paper is organized as follows. Section II formulates the joint source-channel secrecy problem.
Section III proposes a permutation-based uncoded scheme for the discrete communication, and analyzed the
corresponding performance. Sections IV and V extend the proposed scheme to the scalar and vector Gaussian
communications respectively, and another scheme, orthogonal-transform based scheme, is also proposed in these
two sections. Finally, Section VI concludes the paper.
II. P ROBLEM F ORMULATION
A. Problem setup
Consider a bandwidth-matched2 Shannon cipher broadcast system with two legitimate users3 shown in Fig. 1,
where a sender A and two legitimate receivers B1 and B2 share a secret key K that is uniformly distributed over
“ nR ‰4
2 K and independent of a source S n . The sender A observes the discrete memoryless (DM) source sequence
S n with each element i.i.d. (independent and identically distributed) according to PS , and then transmits it to the
legitimate users B1 and B2 over a DM wiretap broadcast channel PY1 Y2 Z|X confidentially by utilizing the secret
key and the wiretap channel. Finally, the legitimate users B1 and B2 produce source reconstructions Sp1n and Sp2n ,
respectively.
Definition 1. An pn, RK q block code consists of
“
‰
1) Encoder: ϕ : S n ˆ 2nRK ÞÑ X n ;
2 Although
here we consider a bandwidth-matched communication system, our results in this section are easy to be extended to any bandwidth-
mismatched system since any system with source-channel bandwidth ratio
ns
nc
can be converted into a bandwidth-matched system, by considering
ns source symbols and nc channel symbols as a source supersymbol and a channel supersymbol, respectively.
3 Although
we only consider the system with two legitimate users, our results derived in this paper can be easily extended to the cases with
more legitimate users.
4 In
this paper, the set t1, ..., mu is sometimes denoted by rms.
5
“
‰
2) Decoders: ψi : Yin ˆ 2nRK ÞÑ Spin , i “ 1, 2.
The encoder and decoders can be stochastic.
Another output Z n of the channel is accessed by a wiretapper Eve. Based on Z n , the wiretapper produces a list
LpZ n q Ď Sqn and the induced distortion is set to the minimum one over the entire list, i.e., minsqn PLpZ n q dE pS n , sqn q,
řn
where dE psn , sqn q fi n1 t“1 dE pst , sqt q is a distortion measure for the wiretapper. For given distortion levels
D0 , D1 , D2 , Nodes A and B1, B2 want to communicate the source within distortions D1 , D2 (for B1 and B2
respectively) by exploiting the secret key and the wiretap channel, while ensuring that the wiretapper’s strategy
always suffers distortion above D0 with high probability.
Definition 2. The tuple pRK , RL , D0 , D1 , D2 q is achievable if there exists a sequence of pn, RK q codes such that
@ ą 0,
1) Distortion constraint:
”
ı
nÑ8
P dB pS n , Spin q ď Di ` ÝÝÝÑ 1, i “ 1, 2;
where dB psn , spn q fi
řn
1
n
t“1
(1)
dB pst , spt q5 is a distortion measure for the legitimate users;
2) Secrecy constraint:
minn
Ln pz q:
1
log |Ln |ďRL ´
lim supnÑ8 n
”
P
ı
nÑ8
n n
q
d
pS
,
s
q
ě
D
´
ÝÝÝÑ 1.
min
E
0
n
n
sq PLpZ q
(2)
Definition 3. The admissible region R fi tAchievable pRK , RL , D0 , D1 , D2 qu.
We assume that the wiretapper knows the pn, RK q code and the distributions PS and PY1 Y2 Z|X .
B. Henchman problem
The problem above is equivalent to the henchman problem [14], in which wiretapper reconstructs a single sequence
with the help of a rate-limited henchman who can access to the source S n and the wiretapper’s observation Z n .
As depicted in Fig. 2, the wiretapper receives the best possible nRn bits from the henchman to assist in producing
a reconstruction sequence Sqn .
Definition 4. The Rn henchman code (Hcode) of a pn, RK q block code consists of
“
‰
1) Encoder: ϕH : S n ˆ Z n ÞÑ 2nRn ;
“
‰
2) Decoder: ψH : 2nRn ˆ Z n ÞÑ Sqn .
We assume that the wiretapper and henchman are aware of the pn, RK q block code adopted by Nodes A and B,
and they cooperate to design a henchman code based on the pn, RK q block code.
Definition 5. The tuple pRK , RL , D0 , D1 , D2 q is achievable in the henchman problem if there exists a sequence of
pn, RK q codes such that @ ą 0,
5 For
simplicity, we only consider the legitimate users have the same distortion measure. Note that our results derived in this paper still hold
for the case with different distortion measures.
6
Key
K 2nRK
Y1n
Sn
A
B1
n
PY1Y2 Z | X Y2
Xn
B2
Zn
S
n
Sˆ1n
Eve
,Zn
Henchman
S
Sˆ2n
n
M 2nRL
Fig. 2. Henchman problem.
1) Distortion constraint: (1);
2) Secrecy constraint:
min
Rn Hcodes:
lim supnÑ8 Rn ďRL ´
”
ı
nÑ8
P dE pS n , Sqn q ě D0 ´ ÝÝÝÑ 1.
(3)
The equivalence between the list secrecy problem and the henchman problem, shown in the following proposition,
has been proven by Schieler and Cuff [14, Prop. 1].
Proposition 1. [14] The tuple pRK , RL , D0 , D1 , D2 q is achievable in the list reconstruction problem if and only if
it is achievable in the henchman problem.
Furthermore, the list secrecy problem and henchman problem are also equivalent to a lossy-equivocation secrecy
problem; see [15].
In addition to the DM system, we also consider the Shannon cipher system with a Gaussian source S „ N p0, λq
transmitted over a power-constrained Gaussian wiretap broadcast channel
Yi “ X ` Vi , i “ 1, 2,
(4)
Z “ X ` V0 ,
(5)
where Vi , i “ 0, 1, 2 are zero-mean additive Gaussian noises with variances Ni , i “ 0, 1, 2, independent of X. For
this case, the constraint on channel input power
“
‰ nÑ8
P ρ pX n q ď P ` ÝÝÝÑ 1, @ ą 0,
should be added to Definitions 2 and 5, where ρ pxn q “
1
n
řn
i“1
(6)
x2i . For the system involving the channel power
constraint, Proposition 1 still holds.
III. D ISCRETE C OMMUNICATION
A. Permutation based Scheme (Finite Alphabets)
In this section, we propose a secure uncoded scheme by coupling the permutation operation with the traditional
uncoded JSCC scheme. The uncoded scheme for JSCC system (with two receivers) consists of three symbol-by-
7
symbol mappings: x psq , sp1 py1 q, sp2 py2 q. The induced distortions are Di “ EdB pS, spi pYi qq, i “ 1, 2. It is easy to
show that we can benefit from replacing the encoder x psq with a stochastic one PX|S when secrecy is considered
řn
for the system. On the other hand, observe that dB psn , spni q “ n1 t“1 dB pst , spi,t q “ ETsn ,spn dB ps, spi q, where Tsn ,psni
i
denotes the joint type (empirical distribution) of psn , spni q. That is, the induced distortions only depend on the joint
type of source and reconstruction sequences. Therefore, if we want to improve the secrecy of a scheme and at the
same time retain the induced distortions unchanged, we only need to require the encryption and decryption operations
does not change the joint type of source and reconstruction sequences. That is, for the encryption s1n psn , kq and
s1n
“ Tsn ,psni . To that end, here we consider a random permutation as the
the decryption spni pp
s1n
i , kq, we require Ts1n ,p
i
encryption operation, and the inverse permutation as the decryption operation. Obviously, the permutation and its
inverse operation do not change the joint type of the source sequence and its reconstructions.
“
‰(
Codebook (Public Key) Generation: Generate a permutation set C “ Ψk , k P 2nRK
with each element
uniformly at random and independently selected from the set of permutations of rns (which is denoted as Sn ). As
a public key, the codebook C is revealed to the sender and all the receivers (including the wiretapper).
Encoding: Upon observing a source sequence sn and a key k , the encoder first generates
s1n “ Ψk psn q ,
and then generates xn according to
śn
t“1
(7)
PX|S pxt |s1t q. Here for a permutation sequence Ψk “ σ n , Ψk psn q fi
psσ1 , sσ2 , ..., sσn q denotes the permutation operation on sn (more precisely, on the indices of sn ) respect to the
permutation sequence Ψk .6
Decoding (for Legitimate Users): For the legitimate user Bi, i “ 1, 2, upon the received sequence yin and the
key k, the decoder first reconstructs s1n as
sp1i,t “ spi pyi,t q, t P rns,
(8)
by using the symbol-by-symbol mapping spi p¨q, and then produces
spni “ Ψ´1
s1n
i q,
k pp
(9)
by using the inverse permutation operation Ψ´1
k p¨q of Ψk p¨q.
The proposed scheme above cascades a random permutation operation with the traditional uncoded JSCC scheme.
The uncoded JSCC part provide a graceful degradation of the source for legitimate users with different channel
qualities. The random permutation operation part that shifts the sequence in the same type provides a certain level
of secrecy. Next we will analyze the asymptotic performance of the proposed scheme as blocklength n goes to
infinity. At first, we need introduce some basic properties of the random codebook C.
Observe that for any permutation sequence Ψ, the mapping between Ψ p¨q and Ψ´1 p¨q is bijective, hence we
have the following lemma.
6 In
this paper, the permutation sequence is termed as permutation sequence, and to distinguish from it, the permutation mapping from one
sequence to another sequence is termed as permutation operation. When there is no disambiguation, we call both of them permutation.
8
Lemma 1. Suppose Ψ is a permutation sequence uniformly at random selected from Sn , the set of permutations
of rns. Then Ψ´1 is also uniformly distributed on Sn , and moreover for any permutation sequence ψ P Sn , both
Ψ pψq and Ψ´1 pψq also have the uniform distribution on Sn .
Utilizing Lemma 1, we can establish the following lemma.
Lemma 2. Suppose Ψ is a permutation sequence uniformly at random selected from Sn . Then Ψ psn q transforms
an arbitrary sequence sn P S n into a random sequence that is uniformly distributed on the set of sequences of type
Tsn (the type class of Tsn ). Moreover, for finite S, the set of sequences of type Tsn has cardinality 2npHpTsn q´op1qq ,
and hence
“
‰
P Ψ psn q “ s1n “ 2´npHpTsn q´op1qq 1 tTs1n “ Tsn u , s1n P S n ,
(10)
where op1q denotes a term tending zero as n Ñ 8.
Proof:
ÿ
“
‰
P Ψ psn q “ s1n “
P pΨ “ ψq
(11)
1
n!
(12)
ψPSn :ψpsn q“s1n
ÿ
“
ψPSn :ψpsn q“s1n
ś
pnTsn psqq!
1 tTs1n “ Tsn u
n!
1 tTsn “ Ts1n u
,
“
|ts1n P S n : Ts1n “ Tsn u|
“
sPS
where (14) follows from |ts1n P S n : Ts1n “ Tsn u| “
n!
ś
sPS pnTs
n psqq!
(13)
(14)
. This implies Ψ psn q transforms an arbitrary
sequence sn P S n into a random sequence uniformly distributed on the set of sequences of type Tsn .
From the type counting lemma [29, Lem. 2.3], we have that for finite S,
pn ` 1q
´|S|
ˇ
(ˇ
2nHpTsn q ď ˇ s1n P S n : Ts1n “ Tsn ˇ ď 2nHpTsn q .
(15)
Hence |ts1n P S n : Ts1n “ Tsn u| “ 2npHpTsn q´op1qq . Combining it with (14) gives us
“
‰
P Ψ psn q “ s1n “ 2´npHpTsn q´op1qq 1 tTs1n “ Tsn u .
(16)
Lemma 2 shows a nice property of the random permutation operation: The resulting sequence will be uniformly
distributed on the set of sequences of type Tsn for the input sequence sn , if the permutation is randomly and
uniformly chosen from the set of permutations of rns. Utilizing this property, we can characterize the performance
of the proposed scheme, as shown in the following theorem. The proof of this theorem is given in Appendix A.
9
Theorem 1 (Permutation based Scheme for Finite Alphabets). For DM communication with finite alphabets
q X , Z, Yi , Spi , i “ 1, 2 are all finite), the permutation based scheme above achieves the region Rpiq Ď R,
(S, S,
where
Rpiq
$
’
’
’ pRK , RL , D0 , D1 , D2 q :
ď &
fi
D ě minspi EdB pS, spi pYi qq, i “ 1, 2,
’ i
PX|S ’
’
% R ď min R ` R pD q, R pD q(
L
K
S|Z
0
S
0
,
/
/
/
.
,
(17)
/
/
/
-
with
pS, Y1 , Y2 , Zq „
ÿ
PS PX|S PY1 Y2 Z|X ,
(18)
x
RS pDq “
min
q
IpS; Sq
(19)
q
IpS; S|Zq
(20)
q
PS|S
q :EdE pS,SqďD
denoting the rate-distortion function of S, and
RS|Z pDq “
min
q
PS|SZ
:EdE pS,SqďD
q
denoting the conditional rate-distortion function of S given two-sided information Z.
Note that for the Rpiq above, the components pD1 , D2 q and the components pRK , RL , D0 q depend on each other
through PX|S . Observe that for a given PX|S , minspi EdB pS, spi pYi qq, i “ 1, 2 are the minimal distortions that the legitimate users can achieve even for the non-secrecy communication case. On the other hand, min RK ` RS|Z pD0 q, RS pD0 q
is larger than RS|Z pD0 q, the optimal RL can be achieved by uncoded schemes when there is no key. Hence compared
with traditional uncoded schemes, the proposed scheme, on one hand, improves improve the performance of secrecy
to a certain extent, and on the other hand, does not lose any performance in terms of the distortions of legitimate
users.
The first constraint of Rpiq is consistent with the performance of traditional uncoded schemes. The second
constraint of Rpiq , roughly speaking, follows from the following argument. On one hand, the henchman and
the wiretapper can ignore the signal Z n altogether and use a RS pD0 q-rate point-to-point source code to describe
S n within distortion D0 . On the other hand, the proposed scheme forces the wiretapper’s optimal strategy to
be an indirect guessing strategy: First, the wiretapper decrypts the secret key by using RK rate; then upon the
observation Z n , the wiretapper reconstructs the sequence S 1n within distortion D0 by using rate RS|Z pD0 q (denote
the reconstruction as Sq1n ); finally, upon the secret key and Sq1n , the wiretapper reconstructs the source S n as
n
1n
q1n
qn
Sqn “ Ψ´1
and Sq1n , since the
k pS q. Obviously the distortion between S and S is the same as that between S
average distortion only depends the joint type of the sequences. Hence the wiretapper needs rate RK ` RS|Z pD0 q
to achieve the distortion D0 .
Now we consider a special case: sending a binary source over a binary wiretap broadcast channel. For the binary
` ˘
communication, the source is a Bernoulli source S „ Bern 21 with the Hamming distortion measure dB ps, spq “
dE ps, spq fi 0, if s “ sp; 1, otherwise. The binary wiretap broadcast channel is Yi “ X ‘ Vi , i “ 1, 2, Z “ X ‘ V0
with Vi „ Bern ppi q , V0 „ Bern pp0 q , 0 ď p0 , p1 , p2 ď 12 . Set X “ S ‘ E with E „ Bernpp1 q. Then from Theorem
1, we get the following corollary.
(
10
Corollary 1 (Binary Communication). For the binary communication above, we have Rpiq Ď R, where
$
,
’
/
/
’
pR
,
R
,
D
,
D
,
D
q
:
K
L
0
1
2
’
/
.
ď &
piq
1
R fi
,
Di ě p ‹ pi , i “ 1, 2,
’
!
) /
/
’
0ďp1 ď 21 ’
% RL ď min RK ` rH2 pp1 ‹ p0 q ´ H2 pD0 qs` , r1 ´ H2 pD0 qs` /
`
with rxs fi max t0, xu, ‹ denoting the binary convolution, i.e.,
x ‹ y “ p1 ´ xqy ` xp1 ´ yq,
(21)
and H2 denoting the binary entropy function, i.e.,
H2 ppq “ ´p log p ´ p1 ´ pq logp1 ´ pq.
(22)
B. Permutation based Scheme (More General Alphabets)
Theorem 1 can be extended to more general alphabets cases, as shown in the following theorem. The proof of
this theorem is given in Appendix C.
Theorem 2 (Permutation based Scheme for More General Alphabets). Assume S is countable, Sq is finite, and
X , Z, Yi , Spi , i “ 1, 2 are general7 . Assume H pSq is finite, and PS satisfies
ˆ
˙
ˆ ˙
n
1
“o
,
NPS
n
log n
ˆ
˙
ˆ ˙
1
1
“o
,
ΦPS
n
log n
ˆ
˙
ˆ
˙
δ
n
rP
“o
, @0 ă δ ď 1,
N
S
log n
log2 n
(23)
(24)
(25)
where NPS pαq fi |ts : PS psq ě αu| denotes the number of probability values that are not smaller than α,
ř
rP pβq fi minα:Φ pαqďβ NP pαq
ΦPS pαq fi s:PS psqăα PS psq denotes the sum of probability values that are smaller than α, N
S
S
PS
denotes the minimum number N such that the sum of the probability values except N largest ones is not larger
than β. Then Theorem 1 still holds.
Remark 1. The conditions (23)-(25) is equivalent to as x Ó 0,8
ˆ
˙
1
,
NPS pxq “ o
x log x1
ˆ
˙
1
ΦPS pxq “ o
,
log x1
¯
´
rP pxq “ o x2 e xδ , @0 ă δ ď 1.
N
S
(26)
(27)
(28)
Remark 2. The conditions (23)-(25) require that the sequence PS psq , s P S should vanish as fast as possible.
Obviously, (23)-(25) hold for any finite S. Besides, for any countably infinite S, it is easy to verify that any
distribution PS such that PS psq “ o ps´α q , s “ 1, 2, ...9 for some α ą 1 satisfies (23)-(25) as well. However, if
7 An
alphabet is countable means that it is either finite or countably infinite. An alphabet is general means that it is either countable or
uncountable (e.g., continuous).
8 This
claim holds when we ignore n is an integer in (23)-(25).
9 Without
loss of generality, any countably infinite S can be converted into t1, 2, 3, ...u by some bijective mapping.
11
PS psq converges slower than or as slow as
1
s,
then
ř
sě1
PS psq does not converge, and hence PS cannot be a
probability distribution. This implies Theorem 2 holds for almost all probability distributions.
Note that for a countably infinite alphabet S, we need the conditions (23)-(25) to guarantee the existence of a
high-probability set (unified typicality set), for each sequence of which, Lemma 2 still holds. This further makes
Theorem 2 hold, just as done for the finite alphabets case.
C. Outer Bound
For the system with a single legitimate user (remove the legitimate user B2 from the system in Fig. 1), the
following outer bound for the admissible region of pRK , RL , D0 , D1 q has been proven by us recently [15].
Lemma 3. [15] For the DM communication with only one legitimate user,
$
’
’ pRK , RL , D0 , D1 q : C1 ě IpS; Sp1 q,
’
’
’
’
ď &
D1 ě EdB pS, Sp1 q,
poq
!
´
¯
R Ď R fi
’ RL ď min RK ` Γ IpS; Sp1 q, PY |X , PZ|X ` R p pD0 q,
PSp |S ’
1
S|S1
’
1
’
)
’
’
%
R pD q
S
0
,
/
/
/
/
/
/
.
,
/
/
/
/
/
/
-
where C1 denotes the channel capacity for the legitimate user, and
˘
`
Γ R, PY |X , PZ|X fi
min
max
QY Z|X :QY |X “PY |X , QX :IQ pX;Y qěR
QZ|X “PZ|X
IQ pX; Y |Zq
(29)
with IQ p¨q denoting the mutual information under distribution QX QY Z|X , is a function specified by the wiretap
channel.
The first two constraints of Rpoq follow from the source-channel coding theorem [30], and the last constraint
follows from an indirect decryption strategy for the wiretapper: Roughly speaking, the wiretapper first reconstructs
Sp1n using rate ΓpIpS; Sp1 q, PY1 |X , PZ|X q, next decrypts the secret key using rate RK , then upon Y1n and secret key,
produces the legitimate user’s reconstruction Sp1n , and finally upon Sp1n produces a final reconstruction Sqn using rate
RS|Sp1 pD0 q. The details can be seen in [15].
By applying this lemma to the system with two legitimate users (the system considered in this paper), the
following outer bound is immediate.
Theorem 3 (Outer Bound). For the DM communication (with two legitimate users),
$
’
’
’ pRK , RL , D0 , D1 , D2 q : Ci ě IpS; Spi q,
ď &
R Ď Rpoq fi
Di ě EdB pS, Spi q, i “ 1, 2,
’
’
PSp Sp |S ’
1 2
% R ď min tR , R , R pD qu
L
1
2
S
0
,
/
/
/
.
,
/
/
/
-
where Ci denotes the channel capacity for the legitimate user i, and
´
¯
Ri “ RK ` Γ IpS; Spi q, PYi |X , PZ|X ` RS|Spi pD0 q, i “ 1, 2.
When specialized to the binary communication, we have the following corollary.
(30)
12
Corollary 2 (Binary Communication). For binary communication,
$
’
’
pRK , RL , D0 , D1 , D2 q :
’
&
poq
R Ď R fi
Di ě pi , i “ 1, 2,
’
!
)
’
’
% RL ď min R1 , R2 , r1 ´ H2 pD0 qs`
,
/
/
/
.
.
/
/
/
-
where
`
`
Ri “ RK ` rH2 pp0 q ´ H2 ppi qs ` rH2 pDi q ´ H2 pD0 qs , i “ 1, 2.
(31)
Comparing Theorem 1 and Corollary 2, we can identify the optimality of the proposed scheme for binary
communication.
Theorem 4 (Optimality of the Proposed Scheme). For the binary communication (with 2 legitimate users), the
proposed uncoded scheme is optimal if p0 ď pi ď Di ď D0 or p0 ě pi “ Di ě D0 holds for i “ 1 or 2.
Remark 3. Theorem 4 implies under conditions that compared with one of legitimate users, the wiretapper has a
better channel and wants to produce a worse reconstruction, or the legitimate user’s distortion is restricted to be the
Shannon limit and meanwhile the wiretapper has a worse channel and wants to produce a better reconstruction, the
proposed uncoded scheme is optimal. It is worth noting that these optimality conditions do not include the practical
scenario in which the wiretapper has a worse channel than the legitimate users and a higher distortion requirement.
But it does not mean our scheme is not optimal for the practical scenario. We believe that for the binary broadcast
communication without secrecy requirement, the proposed uncoded scheme with p1 “ 0 and with no permutation
operation is the unique scheme to achieve the Shannon limits for both the legitimate users. If so, when the secrecy
requirement is involved, the proposed scheme is optimal as well, no matter what the wiretapper’s channel condition
is and what his desired distortion level is. This is because RK rate of secret key could increase RL at most by RK ,
and our scheme satisfies this point. Of course, we need a rigorous proof about this claim, but unfortunately, now
we have no idea how to prove it.
We know that when there is no secrecy constraint, the traditional uncoded scheme could outperform the separate
scheme for broadcast communication scenarios. It is not surprising that when secrecy constraint is involved, the
proposed uncoded scheme still could outperform the separate scheme. However, surprisingly, the example given
in [15] shows the proposed uncoded scheme may strictly outperform the separate coding even for the secure
point-to-point communication (with only one legitimate user).
IV. S CALAR G AUSSIAN C OMMUNICATION
In this section, we consider a Gaussian source S „ N p0, λq transmitted over a bandwidth-matched10 and powerconstrained Gaussian wiretap broadcast channel (the average input power is constrained by P ). The distortion
2
measures are set to dB ps, spq “ dE ps, spq “ d ps, spq fi ps ´ spq .
10 Although
we can also convert a bandwidth-mismatched Gaussian system into a bandwidth-matched system, just as done in Remark 2, our
results in this section cannot be easily extended to the bandwidth-mismatched system since the linear coding used in our schemes is specified
for the bandwidth-matched one.
13
For this communication system, we provide two uncoded schemes. The first one is just the scheme proposed in
previous section. Next we will show that the permutation based scheme also works in the Gaussian communication
case. The other one is an orthogonal-transform based scheme, which cascades a random orthogonal transform
(instead of random permutation operation) with a symbol-by-symbol mapping.
A. Permutation based Scheme
It has been shown that linear coding is optimal for the bandwidth-matched Gaussian broadcast communication
when there is no secrecy requirement [2]. Hence we set PX|S and spi pyi q, i “ 1, 2 to the linear functions x “
b
1
αs, spi “ βi yi , i “ 1, 2 in the proposed scheme for DM communications, where α “ Pλ with 0 ď P 1 ď P and
?
βi “
λP 1
P 1 `Ni .
Then we apply this permutation based scheme to the Gaussian communication. The performance of
this scheme is provided in the following theorem, the proof of which is given in Appendix E.
Theorem 5 (Permutation based Scheme). For the Gaussian communication, the proposed permutation based scheme
achieves the region Rpiq Ď R, where
$
’
’
’ pRK , RL , P, D0 , D1 , D2 q :
ď &
piq
i
R fi
Di ě PλN
1 `N , i “ 1, 2,
’
!i
´
0ďP 1 ďP ’
’
% RL ď min RK ` 1 log`
2
,
/
/
/
.
,
λN0
D0 pP 1 `N0 q
¯
, 12 log`
´
λ
D0
(32)
¯) /
/
/
-
with log` x fi max t0, log xu.
Remark 4. The Rpiq here is just the one given in Theorem 1 with PX|S and spi pyi q, i “ 1, 2 set to x “ αs and
spi “ βi yi , i “ 1, 2, respectively. This is because they are achieved by the same scheme.
Remark 5. The first constraint of Rpiq is consistent with the performance of linear coding [2]. The second constraint
of Rpiq follows from the similar argument to the DM case.
Note that for Rpiq , P 1 is a variable. Moreover, the region of pD1 , D2 q and the region of pRK , RL , D0 q depend on
each other through P 1 which satisfies 0 ď P 1 ď P . This finding is similar to the discrete communication case. Given
pRK , D0 q, the minimum of achievable D1 (or D2 ) and the maximum of achievable RL are both decreasing in P 1 ,
which implies for the proposed scheme, transmitting the source using a larger power results in smaller distortions
for legitimate users, but also leads to decrypting the source more easily for the wiretapper. The proposed scheme
with P 1 “ P , on one hand, provides a certain level of secrecy, and on the other hand, it achieves the Shannon’s
distortion limits for both legitimate users. The region in Theorem 5 with λ “ 1 and P 1 “ 1 is illustrated in Fig. 3.
Given P 1 , pD1 , D2 q has no effect on the pRK , RL , D0 q tradeoff.
B. Orthogonal-Transform based Scheme
The proposed scheme above uses a random permutation operation (which shuffles the sequence within the same
type class) to improve the level of secrecy. It works not only for the discrete communication but also for the
continuous communication, such as the Gaussian communication. In this subsection we propose another secure
uncoded scheme for the Gaussian communication case which is designed from a geometric point of view.
14
4
3.5
3
RL
2.5
2
1.5
1
0.5
0
0
1
0.2
0.5
0.4
0.6
0.8
1
0
N0
D0
Fig. 3. The region in Theorem 5 with λ “ 1 and P 1 “ 1.
To give an interpretation for the motivation of our proposed scheme, we consider a special case where the
wiretapper has a noiseless channel. Apply linear coding to the Gaussian communication, then we know that given
the Euclidean norm, the sequence of a Gaussian source uniformly distributed on some sphere, and so are the
sequences of channel input, outputs, and source reconstructions. Assume we generate a set of bijective transforms
(as a codebook), and randomly choose one of them (according to the key) to transform the source sequence before
applying linear coding on it. To keep the power unchanged, these transforms are required to map a sphere into
itself. On the other hand, by using the secret key the legitimate users could transform it back. Hence the induced
distortions of legitimate users do not change as well. Furthermore, without knowing the secret key but with knowing
the norm of the source sequence and the codebook, in the view of the wiretapper, the source sequence is uniformly
distributed over the vectors that are possible to generate the channel output (wiretapper’s observation) through some
key values. To make the wiretapper guess the source as difficultly as possible, these vectors should be “uniformly”
(at equal distance) located on the sphere. This is because if so, the wiretapper has to cover either all these vectors
or the whole sphere to meet the decryption requirement. It can be shown the orthogonal transform is one of such
transforms. Hence it is adopted in our second scheme.
“
‰
Codebook (Public Key) Generation: Generate random n ˆ n matrices Qk , k P 2nRK independently whose
elements are generated i.i.d. according to N p0, 1q. Then apply Gram-Schmidt orthonormalization process to the
columns of each matrix, hence all the resulting matrices are orthogonal and they constitute a subset of orthogonal
“
‰(
matrices C “ Ψk , k P 2nRK . As a public key, the codebook C is revealed to the sender and all the receivers
(including the wiretapper).
Encoding: Upon observing a source sequence sn and a key k, the encoder generates xn as follows.
xn “ αΨk sn ,
(33)
15
b
where α “
P1
λ
with 0 ď P 1 ď P .
Decoding (for Legitimate Users): For legitimate user Bi, i “ 1, 2, upon the received sequence yin and the key k,
the decoder reconstructs the source as follows.
spni “ βi ΨTk yin ,
?
where βi “
λP 1
P 1 `Ni ,
(34)
and ΨTk denotes the transpose of the matrix Ψk .
Next we will analyze the asymptotic performance of this scheme. Similar to the case of permutation based
scheme, we need first introduce some basic properties of the random codebook C.
Lemma 4. [31] Suppose Q is a random n ˆ n matrix with each element independently distributed according to
Gaussian distribution N p0, 1q. Let Q1 , Q2 , ¨ ¨ ¨ , Qn be the columns of Q and let Ψ be the random matrix whose
columns are obtained by applying the Gram-Schmidt orthonormalization procedure to Q1 , Q2 , ¨ ¨ ¨ , Qn . Then both
Ψ and ΨT have the uniform distribution (Haar measure under orthogonal transform) on the set of nˆn orthogonal
matrices F pnq, and moreover for any orthogonal matrix A, both AΨ and ΨA also have the uniform distribution
on F pnq.
Utilizing Lemma 4, we can establish the following lemma.
Lemma 5. Random orthogonal transform xn “ Ψsn with Ψ uniformly distributed on orthogonal matrices set F pnq,
transforms an arbitrary vector sn P Rn into a random vector that is uniformly distributed on the pn ´ 1q-sphere
with radius }sn }.
Proof: From Lemma 4, without loss of generality we can assume Ψ is obtained in the manner described
in Lemma 4. Let Ψ1 , Ψ2 , ¨ ¨ ¨ , Ψn be the columns of Ψ. From Gram-Schmidt orthonormalization, we know that
Ψ1 “
Q1
}Q1 } ,
and for any rotation matrix (or more generally, orthogonal matrix) A, AΨ1 “
AQ1
}Q1 }
“
AQ1
}AQ1 } .
On
the other hand, Q1 is a random vector with each element i.i.d. „ N p0, 1q, and it is easy to verify that for any
rotation matrix A, AQ1 has the same distribution as Q1 , i.e., a normally distributed random vector is invariant to
rotation. Therefore, AΨ1 has the same distribution as Ψ1 , i.e., Ψ1 is also invariant to rotation. This implies Ψ1 is
T
uniformly distributed on the unit pn ´ 1q-sphere. In addition, observe Ψ p1, 0, ¨ ¨ ¨ , 0q “ Ψ1 . Hence the random
matrix Ψ transforms vector p1, 0, ¨ ¨ ¨ , 0q
T
to a random vector uniformly distributed on the pn ´ 1q-sphere. For
arbitrary vector sn P Rn , we can easily find an orthogonal matrix B with the first column
T
sn
}sn } .
Hence sn can
T
be expressed as sn “ }sn } B p1, 0, ¨ ¨ ¨ , 0q . Then we have Ψsn “ }sn } ΨB p1, 0, ¨ ¨ ¨ , 0q . From Lemma 4, ΨB
T
has the same distribution as Ψ. Hence ΨB p1, 0, ¨ ¨ ¨ , 0q is also uniformly distributed on the unit pn ´ 1q-sphere,
which implies Ψsn is uniformly distributed on the pn ´ 1q-sphere with radius }sn }.
Lemma 5 implies the resulting vector will be uniformly distributed on the sphere where the input vector is, if
the transform matrix is randomly and uniformly chosen from the set of orthogonal matrices. This is a nice property
of the random orthogonal transform, similar to the property of the random permutation operation. Utilizing the
properties, we can establish the following theorem, the proof of which is given in Appendix F.
16
Theorem 6 (Orthogonal-Transform based Scheme). For the Gaussian communication, the inner bound Rpiq given
in Theorem 5 can be achieved by the scheme above as well.
The inner bound Rpiq can be understood from a geometric point of view. The random orthogonal transform in
the proposed scheme guarantees that given Z n , S n has a uniform distribution on 2nRK small pn ´ 2q ´spheres with
b
0
whose centers are uniformly distributed on the pn ´ 1q ´sphere with center O (the origin) and
radius r2 “ PnλN
1
b `N0
nλP 1
radius r1 “ P 1 `N0 . However, owing to the uniform conditional distribution of the source given Z n and the lack
?
r2
qn balls with radius nD0 to cover these pn ´ 2q ´spheres.
of secret key, the wiretapper needs at least 2nRK p ?nD
0
On the other hand, under the unconditional case, the source has a uniform distribution on the pn ´ 1q ´sphere with
?
r0
center O and radius r0 “ nλ. Hence if ignoring Z n , the wiretapper needs at least p ?nD
qn balls with radius
0
?
nD0 to cover the sphere. This results in the inner bound Rpiq .
It seems somewhat counterintuitive that the permutation based scheme achieves the same performance as the
orthogonal-transform based scheme, as shown by Theorems 5 and 6; it is easy to observe that for low-dimension
cases, e.g., 2-dimension case (see Fig. 4), permutations cannot always transform a source sequence into vectors
“uniformly” (at equal distance) distributed over a sphere, so why does this property hold (with high probability) when
the dimension goes to infinity? Actually, it indeed does. This is because as the dimension increases, such “bad”11
source sequences will occur with vanishing probability. This can be seen from that12 P prSsn P Uδn prSsqq Ñ 1 as
n Ñ 8 (i.e., besides on the sphere, the source sequence should also with high probability appear the neighborhoods
of the vectors in Uδn prSsq), and moreover, Uδn prSsq consists of a set of “good” source sequences. Hence the “good”
source sequences will occur with high probability as the dimension increases, that is, permutations will transform
an arbitrary source sequence from a high probability set into vectors “uniformly” distributed over a sphere.
C. Comparison with Sign-Change Based Scheme
In previous two subsections, we give an analysis of the asymptotic performance of permutation based scheme or
orthogonal-transform based scheme. However, is it necessary to let the blocklength n go to infinity? What if n is
set to be a finite value? In this subsection, we study the simplest finite blocklength case: n “ 1 (single-letter codes).
For this case, the permutation based scheme is obviously inferior to the asymptotic case, since for 1 dimension
case no permutation exists except for the source sequence itself. Hence in the following, we mainly consider the
orthogonal-transform based scheme.
For n “ 1, the orthogonal-transform based scheme reduces to a sign-change based scheme [20], [15]. Next we
compare the proposed schemes with this sign-change based scheme [20], [15]. Assume RK “ 1, and the secret key
is uniformly distributed on t0, 1u.
Encoding: Upon observing a source sequence s and a key k , the encoder generates x as follows.
x “ αΨk s,
11 Here
(35)
a source sequence is said to be “good” if its permutations are “uniformly” distributed over a sphere; otherwise it is “bad”. Obviously,
the permutations of a “good” source sequence are also “good”.
´ ¯
12 Here rSs “ ∆ ¨ Round S
and Uδn prSsq is the δ-unified typical set for PrSs ; see the proof in E.
∆
17
O
Fig. 4. Illustration of permutations of a source sequence for n “ 2 case.
b
where α “
P1
λ
with 0 ď P 1 ď P , and
Ψk “
$
’
&´1, if k “ 0;
’
%1,
(36)
if k “ 1.
Decoding (for Legitimate Receivers): For legitimate receiver Bi, i “ 1, 2, upon the received sequence yi and the
key k, the decoder reconstructs the source as follows.
spi “ βi Ψk yi ,
(37)
?
where βi “
λP 1
P 1 `Ni .
It is easy to verify that pSt , Kt , Xt , Y1,t , Y2,t , Zt , Sp1,t , Sp2,t q8
t“1 are i.i.d. and
1
rfV0 pz ´ αsq ` fV0 pz ` αsqs
2
‰
1“
fV01 ps ´ β0 zq ` fV01 ps ` β0 zq ,
“ fZ pzq ¨
2
fS,Z ps, zq “ fS psq ¨
(38)
(39)
?
where β0 “
fV01
λP 1
P 1 `N0 ,
fV0 denotes the probability distribution function (pdf) of the wiretapper’s channel noise V0 , and
´
¯
0
denotes the pdf of V01 „ N 0, PλN
. Given Z, S can be regarded as a Gaussian mixture with two components
1 `N
0
of equal weight and variance. For such single-letter scheme, in [15] we have shown the maximum achievable RL
(or equivalently the minimum rate needed to code S within distortion D0 with two-sided information Z) equals
the conditional rate-distortion function RS|Z pD0 q. The performance of the sign-change based scheme is given by
the following theorem.
Theorem 7 (Sign-Change based Scheme). [15] For the Gaussian communication with RK “ 1, the sign-change
piq
based scheme above achieves the region Rsign Ď R, where
$
’
’
pRK , RL , P, D0 , D1 , D2 q :
’
&
ď
piq
i
Rsign fi
Di ě PλN
1 `N , i “ 1, 2,
i
’
0ďP 1 ďP ’
’
% R ď R pD q
L
S|Z
0
,
/
/
/
.
,
/
/
/
-
18
with RS|Z pD0 q denoting the conditional rate-distortion function of S given two-sided information Z, defined in
(20).
Since it is hard (even if possible) to express RS|Z pD0 q in closed form, for ease of comparison, we will derive
a closed-form upper bound for RS|Z pD0 q. The result is shown in the following lemma, and the proof is given in
Appendix I.
Lemma 6. If pS, Zq follows the distribution (38) or (39), then
"
ˆ
˙*
1
λ
pUBq
RS|Z pD0 q ď min RS|Z pD0 q, log`
,
2
D0
where
(40)
$
1
’
0
& pλ´D0 qpP1 `N0 q ,
ă D0 ď λ;
if PλN
1 `N
λP
pUBq
0
RS|Z pD0 q fi
¯
´
’
λN0
λN0
%1 ` 1 log
2
D0 pP 1 `N0 q , if 0 ď D0 ď P 1 `N0 .
(41)
Since RS|Z pD0 q denotes the minimum rate needed to code S within distortion D0 when Z is available at both
encoder and decoder, we can give an interpretation for the upper bound from the perspective of source coding.
´ ¯
´ ¯
First, by ignoring the side information, we have RS|Z pD0 q ď 12 log` Dλ0 , where 21 log` Dλ0 is the minimum
rate needed to code S without any side information. Second, if
λN0
P 1 `N0
ď D0 ď λ, then consider the following
timesharing coding strategy.13 If we code the secret key K (1 bit per symbol), then using a linear decoder (similar
to those of legitimate users), we can reconstruct the source within distortion
λN0
P 1 `N0 .
On the other hand, if we do not
code anything, then it results in rate 0 and distortion λ. By using a timesharing strategy between these two schemes,
pλ´D0 qpP 1 `N0 q
0
we need
rate to reconstruct the source within distortion D0 . Finally, if 0 ď D0 ď PλN
1 `N , then we
λP 1
0
λN0
P 1 `N0
by using rate 1 to code the secret key, and upon the reconstruction,
¯
´
λN0
we further code the residual error within distortion D0 by using rate 12 log D0 pP
1 `N q .
0
reconstruct the source within distortion
Combining Theorem 7 and Lemma 6 gives us the following result.
piq
Theorem 8 (Outer Bound of Rsign ). For the Gaussian communication with RK “ 1, the region achieved by the
piq
poq
sign-change based scheme satisfies Rsign Ď Rsign , where
$
’
’ pRK , RL , P, D0 , D1 , D2 q :
&
ď ’
poq
i
Rsign fi
Di ě PλN
1 `N , i “ 1, 2,
’
!i
´
1
’
0ďP ďP ’
% RL ď min RpUBq pD0 q, 1 log`
S|Z
2
,
/
/
/
.
.
λ
D0
(42)
¯) /
/
/
-
Remark 6. Observe that only 1 bit/symbol of key can be exploited by the sign-change based scheme even when
RK ą 1. Hence for that case, its performance is still that given by Theorem 7 and outer bounded by 42.
´
¯
pUBq
λN0
From Lemma 6, it can be observed that when RK “ 1, RS|Z pD0 q “ 1 ` 21 log` D0 pP
for 0 ď D0 ď
1 `N q
0
´
¯
pUBq
poq
`
λN0
λN0
1
0
for PλN
ă D0 ď λ. Hence Rsign Ř Rpiq , where Rpiq
1 `N
P 1 `N0 , and RS|Z pD0 q ă 1 “ 1 ` 2 log
D0 pP 1 `N0 q
0
13 Note
that the argument here is only available for the inequality (40), and does not apply to the secrecy problem considered in this paper.
For the secrecy problem the wiretapper and henchman cannot benefit from adopting a timesharing strategy since the constraint (2) or (3) is to
restrict the excess-distortion probability, instead of the average distortion.
19
1
Proposed JSCS
Sign change (upper bound)
0.9
0.8
0.7
RL
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.2
0.4
0.6
0.8
1
D0
Fig. 5. Comparison of the achievable RL by the proposed (infinite blocklength) schemes and that by the sign-change based scheme. λ “ 1,
N0 “ 0 (noiseless wiretap channel) and RK “ 1.
given in Theorem 5 denotes the achievable region by the permutation based scheme or orthogonal-transform based
scheme. This implies for the same P 1 , the sign-change based scheme is strictly inferior to the proposed schemes
under the condition
λN0
P 1 `N0
ă D0 ď λ. That is, the single-letter version of orthogonal-transform based scheme is
inferior to the corresponding infinite blocklength version. To see it clearer, the RL achieved by the proposed (infinite
blocklength) schemes (given in Theorem 5) and the upper bound of RL achieved by the sign-change based scheme
(given in Theorem 8) are illustrated in Fig. 5.
D. Outer Bound
For the Gaussian communication, the following outer bound has been proven for the system with only one
legitimate user [15].
Lemma 7. [15] For the Gaussian communication with only one legitimate user,
$
,
’
/
’ pRK , RL , P, D0 , D1 q :
/
’
/
&
.
poq
λN1
R Ď R fi
,
D1 ě P `N1 ,
’
!
´ ¯) /
’
/
’
/
% RL ď min R1 , 1 log` λ
2
(43)
D0
where
R1 “ RK `
1
log`
2
ˆ
1 ` P{N1
1 ` P{N0
˙
`
1
log`
2
ˆ
D1
D0
˙
.
(44)
Using this result, we have the following outer bound for the system with 2 legitimate users (the system considered
in this paper).
20
Theorem 9 (Outer Bound). For the Gaussian communication (with 2 legitimate users),
,
$
/
’
/
’ pRK , RL , P, D0 , D1 , D2 q :
/
’
.
&
poq
λNi
R Ď R fi
,
Di ě P `Ni , i “ 1, 2,
’
!
´ ¯) /
/
’
/
’
% RL ď min R1 , R2 , 1 log` λ
2
(45)
D0
where
Ri “ RK `
1
log`
2
ˆ
1 ` P{Ni
1 ` P{N0
˙
`
1
log`
2
ˆ
Di
D0
˙
, i “ 1, 2.
(46)
Comparing Theorem 6 and Corollary 9, we can identify the optimality of the proposed schemes for the Gaussian
communication. This result is similar to Theorem 4 for the binary communication.
Theorem 10 (Optimality of the Proposed Schemes). For the Gaussian communication (with 2 legitimate users),
the proposed scheme is optimal if N0 ď Ni , D0 ě Di or N0 ě Ni , D0 ď Di “
λNi
P `Ni
holds for i “ 1 or 2.
A similar remark to Remark 3 applies to this theorem.
V. V ECTOR G AUSSIAN C OMMUNICATION
The proposed schemes are easily extended to vector Gaussian communication scenarios. Consider an m-vector
Gaussian source S „ N p0, diag pλ1 , λ2 , ¨ ¨ ¨ , λm qq14 transmitted over an m-vector Gaussian broadcast channel
Yi “ X ` Vi , i “ 1, 2,
(47)
where Yi is the channel output vector observed by the i-th legitimate user, and Vi „ N p0, diagpNi,1 , Ni,2 , ¨ ¨ ¨ , Ni,m qq
is an additive Gaussian noise vector. A wiretapper Eve accesses to another channel output Z through a channel
Z “ X ` V0 ,
(48)
where V0 „ N p0, diag pN0,1 , N0,2 , ¨ ¨ ¨ , N0,m qq is an additive Gaussian noise vector as well. The distortion
řm
řm
measures are set to dB ps, spq “ dE ps, spq “ j“1 psj ´p
sj q2 , and the channel cost function is set to ρ pxq “ j“1 x2j .
pi , S
qi as super-symbols, then the proposed permutation based scheme can be
Consider the vectors S, X, Yi , Z, S
applied to the vector Gaussian case directly. The performance of this scheme can be proven by following similar
steps to the proof for the scalar Gaussian case.
Furthermore, we can apply the proposed orthogonal-transform based scheme to each subsource-subchannel pair,
as shown in the following.
‰
“
Codebook (Public Key) Generation: Generate m¨2nRK random nˆn matrices Qj,k , j P rms , k P 2nRK independently whose elements are generated i.i.d. according to N p0, 1q. Then we apply Gram-Schmidt orthonormalization
process on every matrix, hence all the resulting matrices are orthogonal, and constitute a subset of orthogonal
“
‰(
matrices C “ Ψj,k , j P rms , k P 2nRK . As a public key, the codebook C is revealed to the sender and all the
receivers (including the wiretapper).
14 In
this paper, we use bold font to denote vector or matrix, e.g., pS1 , ¨ ¨ ¨ , Sm q and ps1 , ¨ ¨ ¨ , sm q are denoted by S and s, respectively.
21
Encoding: Upon observing a source sequence sn “ psn1 , sn2 , ¨ ¨ ¨ , snm q and a key k , the encoder generates
xn “ pxn1 , xn2 , ¨ ¨ ¨ , xnm q as follows.
xnj “ αj Ψj,k snj , j P rms ,
b
where αj “
Pj
λj
with transmitting power P1 , P2 , ¨ ¨ ¨ , Pm such that 0 ď
(49)
řm
j“1
Pj ď P .
Decoding (for Legitimate Users): For the legitimate user Bi, i “ 1, 2, upon the received sequence yin and the
key k, the decoder reconstructs the source as follows.
n
spni,j “ βi,j ΨTj,k yi,j
, j P rms ,
(50)
?
where βi,j “
λj Pj
Pj `Ni .
The achievable regions by the proposed schemes (permutation based scheme and orthogonal-transform based
scheme) are given in the following theorem, the proof of which is given in Appendix J.
Theorem 11 (Performance of the Proposed Schemes). For the vector Gaussian communication, the permutation
based scheme or the orthogonal-transform based scheme above achieves the same region Rpiq Ď R, where
,
$
/
’
/
’
pR
,
R
,
P,
D
,
D
,
D
q
:
K
L
0
1
2
/
’
.
&
ď
řm λj Ni,j
piq
,
R fi
Di ě j“1 Pj `Ni,j , i “ 1, 2,
’
/
’
( /
/
’
%
RL ď min RK ` RS|Z pD0 q, RS pD0 q
P1 , P2 , ¨ ¨ ¨ , Pm ě 0,
řm
0 ď j“1 Pj ď P
with
ˆ ˙
m
ÿ
1
λj
`
log
RS pD0 q “
2
µ
j“1
ˆ
˙
m
ÿ 1
λj N0,j
log`
RS|Z pD0 q “
2
θ pPj ` N0,j q
j“1
(51)
(52)
and with µ and θ such that
D0
m
ÿ
min tµ, λj u ,
(53)
"
*
λj N0,j
“
min θ,
.
Pj ` N0,j
j“1
(54)
“
j“1
D0
m
ÿ
Remark 7. Actually, in Theorem 11, RS pD0 q denotes the rate-distortion function of the source S, and RS|Z pD0 q
denotes the rate-distortion function of the source S with the side information Z available at both the encoder and
b
P
decoder, where Zj “ λjj Sj ` V j , j P rms with V „ N p0, diag pN0,1 , N0,2 , ¨ ¨ ¨ , N0,m qq independent of S.
VI. C ONCLUDING R EMARKS
In this paper, we studied the joint source-channel secrecy problem for secure source broadcast in the Shannon
cipher system, in which the list secrecy is used to measure the secrecy of communication. We proposed two secure
uncoded schemes: a permutation based scheme for discrete, scalar Gaussian, and vector Gaussian communications,
and an orthogonal-transform based scheme for the latter two communications. In these two uncoded schemes, a
22
random permutation or a random orthogonal transform is cascaded with the traditional uncoded JSCC scheme.
The analysis showed that the proposed schemes outperform the sign-change based scheme. Interestingly, by adding
the random permutation operation or the random orthogonal transform into the traditional uncoded scheme, the
proposed uncoded schemes, on one hand, provide a certain level of secrecy, and on the other hand, do not lose any
performance in terms of the distortions for legitimate users.
Although the proposed schemes adopt two different random transforms, permutation operation and orthogonal
transform, they are consistent in two aspects: First, actually the permutation operation is one kind of orthogonal
transform; second, for the Gaussian communication, the orthogonal transform can be also considered as a shift
operation that shifts a sequence to another in the same “type”, if we treat the Euclidean norm of the source
sequence as its “type”15 . Furthermore, it is worth noting that different from the common construction of codebook
in information theory (including spherical codes such as the one used in [27]), the codebooks in the proposed
schemes are constructed by generating a sequence of i.i.d. random permutations or random matrices, instead of
a sequence of i.i.d. random samples. In other words, the codebooks used here specify a sequence of bijective
operations or transforms and hence they apply to uncoded schemes; while the common codebooks in information
theory only specify a sequence of samples and hence can only be used in quantization operation (or digital schemes).
Furthermore, such random-permutation or random-matrix based codebook construction can be also found in [22],
[23], [24], [25], [26], where they were used to design digital schemes for communication, secrecy communication,
and antijamming communication problems. But different from those works, in our case they were used to design
uncoded schemes, instead of digital schemes.
It is worth noting that the proofs used in this paper follow basic outline of the proofs in [14]. But different
from [14], besides the finite alphabet case, we also considered the countably infinite alphabet and continuous
(Gaussian) alphabet cases. Hence some powerful techniques, including unified typicality, d´tilted information,
geometric analysis, and discretization, are used in our proofs. Furthermore, the unified typicality used in our proofs
is different from the existing one defined in [34]. The unified typical set defined by us has a good property that the
sequences in it only have (nearly) sub-exponential number of types. This property coincides with the finite alphabet
case, and is of crucial importance to our proofs. We believe our definition of unified typicality could be used to
further extend the method of types to countably infinite alphabet cases (besides the extension in [34]).
A PPENDIX A
P ROOF OF T HEOREM 1
Denote
15 This
n
Z 1n fi Ψ´1
K pZ q ,
(55)
n
X 1n fi Ψ´1
K pX q ,
(56)
n
Yi1n fi Ψ´1
K pYi q .
(57)
kind of type can be called “weak type”, since the relationship of it and the weak typicality is similar to that of the traditional type
(empirical distribution) and strong typicality.
23
Then from the fact that the permutation operation is bijective, we have that
PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn
i
i
i
i
“ PK PC PS n PS 1n |S n ΨK PX n |S 1n PYin Z n |X n PSp1n |Y n ΨK PSpn |Sp1n PX 1n |X n ΨK PYi1n |Yin ΨK PZ 1n |Z n ΨK
(58)
i
i
i
i
“ PK PC PS n PS 1n |S n ΨK PX 1n |S 1n ΨK PYi1n Z 1n |X 1n ΨK PSp1n |Y 1n ΨK PSpn |Y 1n ΨK PX n |X 1n ΨK PYin |Yi1n ΨK PZ n |Z 1n ΨK (59)
i
i
i
i
“ PK PC PS n PX 1n |S n PYi1n Z 1n |X 1n PSpn |Y 1n PS 1n |S n ΨK PSp1n |Y 1n ΨK PX n |X 1n ΨK PYin |Yi1n ΨK PZ n |Z 1n ΨK ,
i
i
i
(60)
i
and similarly, PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn can be also expressed as
i
i
i
i
PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn
i
i
i
i
“ PK PC PS 1n PX n |S 1n PYin Z n |X n PSp1n |Y n PS n |S 1n ΨK PSpn |Sp1n PX 1n |X n ΨK PYi1n |Yin ΨK PZ 1n |Z n ΨK .
i
i
i
(61)
i
Hence pΨK , Z n q Ñ Z 1n Ñ S n forms a Markov chain. Furthermore, since the permutation operation does not change
the joint distribution of the sequences, we have PS n PX 1n |S n PYi1n Z 1n |X 1n PSpn |Y 1n “ PS 1n PX n |S 1n PYin Z n |X n PSp1n |Y n “
i
i
i
i
ś
PS PX|S PYi Z|X PSpi |Yi , where PSpi |Yi pp
s|yq fi 1 tp
s “ spi pyqu denotes the conditional distribution induced by the
decoder i, and PS PX|S PYi Z|X PSpi |Yi is the distribution given in (18).
´
¯
Since S n , Spin is an i.i.d. sequence, by the law of large numbers,
”
ı
nÑ8
P dB pS n , Spin q ď EdB pS, Spi q ` ÝÝÝÑ 1,
(62)
for any ą 0. Hence the distortion constraints for legitimate users are satisfied.
Next we prove the secrecy constraint is also satisfied, i.e., if
(
lim sup Rn ă min RK ` RS|Z pD0 q, RS pD0 q ,
(63)
nÑ8
”
“
‰ı
then lim ECZ n maxRn Hcodes P dE pS n , Sqn q ď D0 “ 0. To that end, we need the following lemma.
nÑ8
Lemma 8. [15] For a sequence of random variables tXn u, and a sequence of events tAn u, limnÑ8 P pAn q “ 0,
if and only if limnÑ8 P rP pAn |Xn q ą τn s “ 0 for some sequence tτn u with τn ą 0 and limnÑ8 τn “ 0.
From Lemma 8, to prove the secrecy constraint we only need to show that if Rn satisfies (63), then
”
ı
“
‰
lim PCZ n max P dE pS n , Sqn q ď D0 ą τn “ 0,
nÑ8
Rn Hcodes
(64)
for some sequence tτn u with τn ą 0 and limnÑ8 τn “ 0. Next we prove this.
Define event
A fi
` n 1n ˘
`
˘(
S ,Z
P Tδn S, Z 1 ,
for δ ą 0. The δ-typical set is defined according to the notion of strong typicality, see [30]:
ÿ
Tδn pSq fi tsn P S n :
|Tsn psq ´ PS psq| ď δu,
(65)
(66)
sPS
where Tsn denotes the type (or empirical distribution) of sn . For simplicity, Tδn pSq is also shortly denoted as Tδn .
Since pS n , Z 1n q is an i.i.d. sequence, from the fact that the typical set has total probability close to one [30], we
have the following lemma.
24
Lemma 9. [30] For any δ ą 0, P rAs Ñ 1, as n Ñ 8.
“
‰
Consider that for each n, the optimal Rn -rate henchman code that maximizes P dE pS n , Sqn q ď D0 |CZ n is
” “
ı
‰
adopted, then we only need to show lim PCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn “ 0 for these codes. By utilizing
nÑ8
Lemmas 8 and 9, we have
ı
” “
‰
PCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn
ı
” “
‰
ďPCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn , P rAc |CZ n s ď n ` P rP rAc |CZ n s ą n s
ı
” “
‰
ďPCZ n P dE pS n , Sqn q ď D0 , A|CZ n ` P rAc |CZ n s ą τn , P rAc |CZ n s ď n ` 1n
ı
” “
‰
ďPCZ n P dE pS n , Sqn q ď D0 , A|CZ n ą τn1 ` 1n ,
(67)
(68)
(69)
for some n and 1n that both vanish as n Ñ 8, where τn1 “ τn ´ n . By choosing proper τn , τn1 can be set to
some sequence that converges to zero sub-exponentially fast (i.e., τn1 “ 2´opnq ). Since n vanishes as n Ñ 8, this
guarantees that τn also vanishes as n Ñ 8.
Owing to the rate constraint, given pC, Z n q, the reconstruction Sqn cannot take more than Rn values. Denote the
set of possible values as cpC, Z n q, then
”
“
‰
P dE pS n , Sqn q ď D0 , A|CZ n “ P
min
sqn PcpC,Z n q
ı
dE pS n , sqn q ď D0 , A|CZ n .
(70)
Now we apply a union bound to the right-hand side of (70) and write
ı
”
P n min n dE pS n , sqn q ď D0 , A|CZ n
sq PcpC,Z q
”
ı
P dE pS n , sqn q ď D0 , A|CZ n
(71)
”
ı
n n
n
s
q
P
d
pS
,
q
ď
D
,
A|CZ
max
E
0
n
n
(72)
ÿ
ď
sqn PcpC,Z n q
ď2nRn
sq PcpC,Z q
”
ı
ď2nRn max P dE pS n , sqn q ď D0 , A|CZ n
(73)
sqn PSqn
nRn
max
“2
nRK
2ÿ
P rK “ k|CZ n s P rdE pS n , sqn q ď D0 , A|CZ n , K “ ks
sqn PSqn k“1
“2npRn ´RK q max
nRK
2ÿ
sqn PSqn k“1
P rdE pS n , sqn q ď D0 , A|Ψk , Z n s ,
(74)
(75)
where (75) follows from the Markov chain CKZ n Ñ ΨK Z n Ñ S n Z n A and P rK “ k|C “ c, Z n “ z n s “ 2´nRK
(see (61)).
Combine (69), (70), and (75), then we have
ı
” “
‰
PCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn
”
ďP
CZ n
max
nRK
2ÿ
sqn PSqn k“1
ı
P rdE pS n , sqn q ď D0 , A|Ψk , Z n s ą τn1 2´npRn ´RK q ` 1n
nRK
ˇ ˇ
” 2ÿ
ı
ˇ qn ˇ
ď ˇS ˇ max PCZ n
ξk,zn pq
sn q ą τn1 2´npRn ´RK q ` 1n ,
sqn PSqn
k“1
(76)
(77)
25
where
sn q fi P rdE pS n , sqn q ď D0 , A|Ψk , Z n s ,
ξk,zn pq
(78)
Therefore, if we can show that the probability in (77) decays doubly exponentially fast with n, then the proof will
be complete.
“
‰
sn q , k P 2nRK are i.i.d. random variables, with mean
Consider that given sqn and z n , ξk,zn pq
EC ξk,zn pq
sn q “ EC P rdE pS n , sqn q ď D0 , A|Ψk , z n s
“ EΨk P rdE pS n , sqn q ď D0 , A|Ψk , z n s
(79)
(80)
To complete the proof, we need introduce the following lemmas. The proof of Lemma 10 is given in Appendix B.
Lemma 10. Assume S n is i.i.d. according to PS , then for any type t of sequences in S n and any sqn ,
PrdE pS n , sqn q ď D, S n P Tδn |TS n “ ts ď 2´npRS pDq´op1qq ,
(81)
where TS n denotes the type of S n , and op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
Lemma 11. [14] Fix PS|Z and z n . If S n is distributed according to
śn
i“1
PS|Z“zi , then for any sqn ,
PrdE pS n , sqn q ď D, pS n , z n q P Tδn |z n s ď 2´npRS|Z pDq´op1qq ,
(82)
where op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
Lemma 12. [14] If X m is a sequence of i.i.d. random variables on the interval r0, as with ErXi s “ p, then
m
”ÿ
ı ´ e¨m¨p ¯k{a
.
(83)
P
Xi ą k ď
k
i“1
From (60), we have
P rS n “ sn |Ψk , z n s “
ź
`
˘
PS|Z si |zi1 .
(84)
Hence Lemma 11 implies
“
‰
n n
1n
ξk,zn pq
sn q “ P dE pS , sq q ď D0 , A|z
ď 2´npRS|Z pD0 q´op1qq .
(85)
(86)
On the other hand,
n n
n
n
n
sn q ď EΨk PrdE pS , sq q ď D0 , S P Tδ |Ψk , z s
EC ξk,zn pq
ÿ
“
PrS 1n “ s1n |z n sEΨk PrdE pS n , sqn q ď D0 , S n P Tδn |S 1n “ s1n , Ψk s
(87)
(88)
s1n
ÿ
“
PrS 1n “ s1n |z n sPrdE pS n , sqn q ď D0 , S n P Tδn |S 1n “ s1n s
(89)
PrS 1n “ s1n |z n sPrdE pS n , sqn q ď D0 , S n P Tδn |TS n “ Ts1n s
(90)
s1n
ÿ
“
s1n
ď 2´npRS pD0 q´op1qq .
(91)
26
Using these bounds, we apply Lemma 12 to the probability in (77) by identifying
m “ 2nRK ,
Then we have
(92)
a “ 2´npRS|Z pD0 q´op1qq ,
(93)
p ď 2´npRS pD0 q´op1qq ,
(94)
k “ τn1 2´npRn ´RK q .
(95)
nRK
” 2ÿ
ı
nβ
sn q ą τn1 2´npRn ´RK q ď 2´nα2 ,
P
ξk,zn pq
(96)
k“1
where
α “ RS pD0 q ´ Rn ´ op1q
(97)
β “ RK ` RS|Z pD0 q ´ Rn ´ op1q.
For small enough δ and large enough n, both α and β are positive and bounded away from zero, and (96) vanishes
doubly exponentially fast. Therefore, the expression in (77) vanishes. This completes the proof of Theorem 1.
A PPENDIX B
P ROOF OF L EMMA 10
|t psq ´ PS psq| ą δ, then PrdE pS n , sqn q ď D, S n P Tδn |TS n “ ts “ 0. Hence we only need to consider
ř
the t’s such that sPS |t psq ´ PS psq| ď δ.
If
ř
sPS
Consider
ˇ
(ˇ
PrTS n “ ts “ ˇ s1n P S n : Ts1n “ t ˇ 2´npDpt||PS q`Hptqq
“ 2´npDpt||PS q`op1qq
(98)
(99)
for any type t of sequences in S n , where D pt||PS q denotes the relative entropy between t and PS , and (99) follows
from (15). Moreover, from [32, Thm. 25] we have
¸
˜
ř
˙
ˆ
2
p s |t psq ´ PS psq|q
δ2
ď log 1 `
Ñ 0,
D pt||PS q ď log 1 `
2PS,min
2PS,min
(100)
as δ Ñ 0, where PS,min “ minsPS PS psq. Therefore,
PrTS n “ ts ě 2´nop1q .
(101)
Utilizing (101), we get
PrdE pS n , sqn q ď D, S n P Tδn |TS n “ ts
PrdE pS n , sqn q ď D, S n P Tδn , TS n “ ts
PrTS n “ ts
n n
PrdE pS , sq q ď D, S n P Tδn s
ď
.
2´nop1q
“
To complete the proof, we need the following lemma.
(102)
(103)
27
Lemma 13. [14] Assume S n is i.i.d. according to PS , then for any sqn ,
PrdE pS n , sqn q ď D, S n P Tδn s ď 2´npRS pDq´op1qq .
(104)
By the lemma above, (103) implies that
PrdE pS n , sqn q ď D, S n P Tδn |TS n “ ts ď 2´npRS pDq´op1qq .
(105)
A PPENDIX C
P ROOF OF T HEOREM 2
Define X 1n , Yi1n , Z 1n same as (55)-(57), then the distribution PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn also satisfies (60)
i
i
i
i
and (61). Similar to the finite alphabet case, it is easy to show the distortion constraints for legitimate users are
satisfied.
Next following similar steps to the proof for the finite alphabet case, we prove the secrecy constraint is also
satisfied for this case. Before proving that, we need introduce d´tilted information and conditional d´tilted
information first.
Let PSq‹ |S be a distribution that achieves the rate-distortion function RS pDq (which is not necessarily unique).
Then d´tilted information is defined as follows.
Definition 6 (d´tilted information [15]). For D ą Dmin fi inf tD : RS pDq ă 8u, the d´tilted information in s
is defined as
1
S ps, Dq “ log “
`
˘‰ ,
‹
E exp λ D ´ λ‹ dps, Sq‹ q
(106)
where the expectation is with respect to PSq‹ , i.e. the unconditional distribution of the reproduction random variable
that achieves RS pDq, and
λ‹ “ ´RS1 pDq.
(107)
For pS, Zq that follow the distribution in (18), we define
RS|Z“z pβq “
“min
‰
q “ zq.
IpS; S|Z
(108)
q
PS|S,Z“z
:E dE pS,Sq|Z“z
ďβ
q
Let PSq‹ |S,Z“z be a distribution that achieves RS|Z“z pβq. Define b‹ pzq fi ES,Sq‹ |Z“z dpS, Sq‹ q with the expectation
taken with respect to PS|Z“z PSq‹ |S,Z“z .
(
Definition 7 (Conditional d´tilted information [15]). For b‹ pzq ą βmin pzq fi inf β : RS|Z“z pβq ă 8 , the
conditional d´tilted information in s under condition Z “ z is defined as
S|Z“z ps, b‹ pzqq “ log
1
”
´
¯ı ,
‹
‹
ESq‹ |Z“z exp λ pzq b pzq ´ λ‹ pzq dps, Sq‹ q
(109)
where the expectation is with respect to PSq‹ |Z“z , i.e. the margin distribution of PS|Z“z PSq‹ |S,Z“z , and
1
λ‹ pzq “ ´RS|Z“z
pb‹ pzqq.
Next we prove the secrecy constraint. To that end, we need re-define
(110)
28
n
!
1 ÿ
A fi S n P Uδn ,
S pS i , D0 q ě RS pD0 q ´ δ,
n i“1
n
n
)
1 ÿ ‹ 1
1 ÿ
S|Z“Zi1 pS i , b‹ pZi1 qq ě RS|Z pD0 q ´ δ,
b pZi q ě D0 ´ δ
n i“1
n i“1
(111)
for δ ą 0. The δ-unified typical set is defined as16
Uδn pSq fi T nδ pSq X Wδn pSq,
(112)
log n
where T nδ pSq defined in (66), denotes the
log n
δ
log n -strongly
!
Wδn pSq fi sn P S n
typical set, and
ˇ 1
ˇ
)
ˇ
ˇ
: ˇ´ log PS n psn q ´ H pSqˇ ď δ ,
n
(113)
denotes the δ-weakly typical set [28]. For simplicity, Uδn pSq is also shortly denoted as Uδn .
Since pS n , Z 1n q is an i.i.d. sequence, we have the following lemma.
´
rP
Lemma 14. [15, Lem. 18] [34, Lem. 2] Assume PS satisfies N
S
δ1
log n
¯
´
“o
n
log2 n
¯
, @0 ă δ 1 ď 1. Then for
any δ ą 0, P rAs Ñ 1 as n Ñ 8.
Then the derivation up to (77) still holds, i.e.,
” “
ı
‰
PCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn
nRK
ˇ ˇ
” 2ÿ
ı
ˇ qn ˇ
ξk,zn pq
sn q ą τn1 2´npRn ´RK q ` 1n ,
ď ˇS ˇ max PCZ n
sqn PSqn
(114)
k“1
Therefore, if we can show that the probability in (114) decays doubly exponentially fast with n, then the proof
will be complete. To that end, we need introduce the following lemmas. The proof of Lemma 15 is given in
Appendix D.
Lemma 15. Assume PS satisfies NPS
`1˘
n
´
“o
n
log n
¯
, ΦP S
`1˘
n
´
“o
1
log n
¯
, and S n is i.i.d. according to PS ,
then for any type t of sequences in S n and any sqn P Sqn ,
n
‰
“
1 ÿ
P dE pS n , sqn q ď D, S n P Uδn ,
S pSi , Dq ě RS pDq ´ δ|TS n “ t ď 2´npRS pDq´op1qq ,
n i“1
(115)
where op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
Lemma 16. [15] Fix PSZ and z n P Z n . Assume given Z n “ z n , S n is distributed according to
śn
i“1
PS|Z“zi ,
then for any sqn P Sqn ,
n
n
”
ı
1 ÿ
1 ÿ ‹
P dE pS n , sqn q ď D,
S|Z“zi pS i , b‹ pzi qq ě RS|Z pDq ´ δ,
b pzi q ě D ´ δ|Z n “ z n
n i“1
n i“1
ď 2´npRS|Z pDq´op1qq ,
(116)
16 Here
the δ-unified typical set is different from the one defined in [34]. Our definition has the benefit that it makes the following property
ˇ
(ˇ
hold: For each sequence sn P Uδn pSq, ˇ s1n P Uδn pSq : Ts1n “ Tsn ˇ “ 2npHpSq´op1qq , or equivalently, PrTS n “ Tsn s “ 2´nop1q , where
op1q is a term that vanishes as δ Ñ 0 and n Ñ 8. This property coincides with (101) for the finite alphabet case, and it is of crucial importance
to our proof here (see (126)).
29
where op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
Apply Lemmas 12, 15 and 16, then we have that the probability in (114) decays doubly exponentially fast with
n. This completes the proof of Theorem 2.
A PPENDIX D
P ROOF OF L EMMA 15
δ
log n
does not hold, then PrdE pS n , sqn q ď D, S n P Uδn |TS n “ ts “ 0. Hence we only
ř
need to consider the t’s satisfying sPS |t psq ´ PS psq| ď logδ n .
If
ř
sPS
|t psq ´ PS psq| ď
The Lemma 2.6 of [29] says that for any type t of sequences in S n ,
PrTS n “ ts ě pn ` 1q
´|suppptq|
2´nDpt||PS q ,
(117)
where supp ptq fi ts P S : t psq ą 0u denotes the suppose of t.
n
log n
pδ ` n q holds, where n is a term that vanishes as n Ñ 8.
(
(
To that end, we divide S into two parts: s : PS psq ě n1 and s : PS psq ă n1 . Then
ˇ"
*ˇ ˇ"
*ˇ
ˇ
1 ˇˇ ˇˇ
1 ˇˇ
ˇ
|supp ptq| “ ˇ s : t psq ą 0, PS psq ě
` s : t psq ą 0, PS psq ă
(118)
n ˇ ˇ
n ˇ
ˇ"
*ˇ ˇ"
*ˇ
ˇ
1 ˇˇ ˇˇ
1 ˇˇ
ď ˇˇ s : PS psq ě
` s : t psq ą 0, PS psq ă
(119)
n ˇ ˇ
n ˇ
ˆ ˙
ÿ
1
“ NPS
`
1 tt psq ą 0u
(120)
n
1
s:PS psqă n
ˆ ˙
ÿ
1
`n
t psq ,
(121)
ď NPS
n
1
Now we prove that for any δ ą 0, |supp ptq| ď
s:PS psqă n
where (120) follows from the definition of NPS
`1˘
n
, and (121) follows from the fact t psq ě
1
n
for any s such that
t psq ą 0.
ř
ř
ř
ř
Since sPS |t psq ´ PS psq| ě s:PS psqă 1 |t psq ´ PS psq| ě s:PS psqă 1 pt psq ´ PS psqq and sPS |t psq ´ PS psq| ď
n
δ
log n ,
n
we have
ÿ
1
s:PS psqă n
t psq ď
ÿ
1
s:PS psqă n
PS psq `
δ
log n
ˆ ˙
1
δ
“ ΦPS
`
.
n
log n
(122)
(123)
Therefore,
ˆ ˙
ˆ ˙
1
1
δn
|supp ptq| ď NPS
` nΦPS
`
.
(124)
n
n
log n
´
¯
´
¯
` ˘
` ˘
Since NPS n1 “ o logn n , ΦPS n1 “ o log1 n , we have |supp ptq| ď logn n pδ ` n q. Therefore, (117) implies
PrTS n “ ts ě 2´npDpt||PS q`δ`n q .
(125)
30
Furthermore, S n P Uδn implies D pt||PS q ď 2δ, which is obtained by following part of proof steps of [34, Thm. 3]
(but with and δ replaced with δ and
n
log n ,
respectively). Hence it holds that
PrTS n “ ts ě 2´nop1q ,
(126)
where op1q is a term that vanishes as δ Ñ 0 and n Ñ 8. Utilizing (126), we can get
n
1 ÿ
S pSi , Dq ě RS pDq ´ δ|TS n “ ts
n i“1
řn
PrdE pS n , sqn q ď D, S n P Uδn , n1 i“1 S pSi , Dq ě RS pDq ´ δ, TS n “ ts
“
PrTS n “ ts
řn
1
n n
PrdE pS , sq q ď D, n i“1 S pSi , Dq ě RS pDq ´ δs
.
ď
2´nop1q
PrdE pS n , sqn q ď D, S n P Uδn ,
(127)
(128)
To complete the proof, we need the following lemma.
Lemma 17. [15] Assume S and Sq are general (not necessarily countable) alphabets, and S n is i.i.d. drawn from
S n according to PS . Then for any D ą Dmin (Dmin is defined in Definition 6) and any sqn P Sqn ,
PrdE pS n , sqn q ď D,
n
1 ÿ
S pSi , Dq ě RS pDq ´ δs ď 2´npRS pDq´op1qq .
n i“1
(129)
Hence by the lemma above, (128) implies that
PrdE pS n , sqn q ď D, S n P Uδn ,
n
1 ÿ
S pSi , Dq ě RS pDq ´ δ|TS n “ ts ď 2´npRS pDq´op1qq .
n i“1
(130)
A PPENDIX E
P ROOF OF T HEOREM 5
Similar to the DM cases, it is easy to show the distortion constraints for legitimate users are satisfied.
Next by following similar steps to the proof for the DM cases, we prove the secrecy constraint is also satisfied.
q Let
To that end, we first need to discretize the source S and the reconstruction S.
q P N fi t¨ ¨ ¨ , ´2∆, ´∆, 0, ∆, 2∆, ¨ ¨ ¨ u ,
rSs, rSs
(131)
q obtained by mapping S and Sq to the closest quantization point, i.e., rSs “
be quantized versions of S and S,
´ ¯
`S˘
q “ ∆ ¨ Round Sq . Then we have for any sn P Rn ,
∆ ¨ Round ∆ , rSs
∆
0ď
n
1 ÿ
∆2
2
psi ´ rsi sq ď
.
n i“1
4
(132)
Furthermore, it holds that
a
n
n
ndpsn , sqn q “ ||sn ´ rq
ss ` rq
ss ´ sqn ||
n
(133)
n
ě ||sn ´ rq
ss || ´ || rq
ss ´ sqn ||
n
n
n
(134)
n
ss || ´ ||sn ´ rss || ´ || rq
ss ´ sqn ||
ě || rss ´ rq
b
n
n
ě ndprss , rq
ss q ´ ∆
(135)
(136)
31
where (134) follows from triangle inequality. Utilizing this inequality, we have
„
¯2
´a
n
n
n n
n
n
D0 ` ∆ , A|Ψk , Z
P rdpS , sq q ď D0 , A|Ψk , Z s ď P d prSs , rq
ss q ď
‰
“
n
ss q ď D01 , A|Ψk , Z n ,
“ P d prSsn , rq
where D01 fi
`?
(137)
(138)
˘2
D0 ` ∆ .
Reorder the probabilities PrSs prssq, rss P N in decreasing order, and denote the result as Pi , i “ 1, 2, .... Then
P1 “ PrSs p0q, P2j “ P2j`1 “ PrSs pj∆q, j ě 1. Obviously,
∆fS ppj ` 1q∆q ď P2j “ P2j`1 ď ∆fS pj∆q ,
(139)
2
and hence for Gaussian sources, P2j “ P2j`1 “ ope´j q. From Remark 2, we have that PrSs satisfies the conditions
(23)-(25).
Define event
n
!
1 ÿ
A fi S n P Wδn , rSsn P Uδn prSsq,
S prSsi , D0 q ě RrSs pD0 q ´ δ,
n i“1
n
n
)
1 ÿ
1 ÿ ‹ 1
rSs|Z“Zi1 prSsi , b‹ pZi1 qq ě RrSs|Z pD0 q ´ δ,
b pZi q ě D0 ´ δ
n i“1
n i“1
(140)
for δ ą 0. Observe that the distribution PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn also satisfies (60) and (61), which implies
i
i
i
i
pS n , Z 1n q is an i.i.d. sequence. Hence Lemma 14 still holds for this case. Following similar steps to the proof of
Theorem 1, we can get
” “
ı
‰
PCZ n P dE pS n , Sqn q ď D0 |CZ n ą τn
”
ď PCZ n
max
n
n
ss PN
rq
”
“ PCZ n
max
n
n
ss PB
rq
nRK
2ÿ
(141)
ı
“
‰
n
ss q ď D01 , A|Ψk , Z n ą τn1 2´npRn ´RK q ` 1n
P dprSsn , rq
(142)
k“1
nRK
2ÿ
k“1
ď |B n | max
PCZ n
n
n
ss PB
rq
ı
“
‰
n
ss q ď D01 , A|Ψk , Z n ą τn1 2´npRn ´RK q ` 1n
P dprSsn , rq
nRK
” 2ÿ
ı
n
ss q ą τn1 2´npRn ´RK q ` 1n ,
ξk,zn prq
(143)
k“1
where
!
? )
ssn P N n : }rq
ssn } ď nΓ
B n fi rq
with
?
Γ fi
(144)
a
?
a
λ p1 ` δq ` ∆ ` D01 ,
‰
“
n
n
ss q fi P d prSsn , rq
ss q ď D01 , A|Ψk , Z n ,
ξk,zn prq
(145)
a
and (142) follows that S n only appears in the ball with radius nλ p1 ` δq which implies rSsn only appears in
a
?
n
ss P B n instead of the whole set
the ball with radius λ p1 ` δq ` ∆ , hence it is sufficient to only consider rq
N n.
32
Furthermore, observe that
Volume of n´ball with radius
|B | ď
∆n
´?
¯n
?
n{2
π
nΓ ` n∆2
˘
`
“
∆n Γ n2 ` 1
?
n
n
ď 22
n
“ 22
nΓ `
log
q
p
∆2
´ 12 log πn`op1q
n∆2
(146)
(147)
?
?
log π`n logp nΓ` n∆2 q´n log ∆´ n
2 log
?
?
2
Γ` ∆2
2πe
?
n
1
2e ´ 2
log πn`op1q
(148)
.
(149)
Therefore, if we can show that the probability in (143) decays doubly exponentially fast with n, then the proof
will be complete.
Apply Lemmas 12, 15 and 16, then we have the probability in (143) decays doubly exponentially fast with m.
”
“
‰ı
Hence lim ECZ n maxRn Hcodes P dE pS n , Sqn q ď D0 “ 0 if
nÑ8
(
lim sup Rn ă min RK ` RrSs|Z pD01 q, RrSs pD01 q .
(150)
nÑ8
To complete the proof, we need to show RrSs|Z pD01 q ě RS|Z pD0 q and RrSs pD01 q ě RS pD0 q as ∆ Ñ 0.
1
1
1
q˚
q˚
Suppose PrSs
q ˚ |rSsZ achieves RrSs|Z pD0 q, then RrSs|Z pD0 q “ IprSs; rSs |Zq and EdprSs, rSs q ď D0 . Since for
PrSsrSs
q ˚ |SZ “ PrSs|S PrSs
q ˚ |rSsZ ,
´
¯2
q ˚ q “ E S ´ rSs ` rSs ´ rSs
q˚
EdpS, rSs
´
¯2
´
¯
q ˚ ` E pS ´ rSsq2 ` 2E pS ´ rSsq rSs ´ rSs
q˚
“ E rSs ´ rSs
b
q ˚ q ` ∆ EdprSs, rSs
q ˚q
ď EdprSs, rSs
a
ď D01 ` ∆ D01 ,
(151)
(152)
(153)
(154)
a ˘
`
where (153) follows from the Cauchy–Schwarz inequality. On the other hand, RS|Z D01 ` ∆ D01 is defined
a ˘
a
`
q ˚ |Zq “
q
q ď D1 ` ∆ D1 . Hence RS|Z D1 ` ∆ D1 ď IpS; rSs
as the minimum IpS; S|Zq
such that EdpS, Sq
0
0
0
0
q ˚ |Zq “ RrSs|Z pD1 q. Let ∆ Ñ 0, then we have RS|Z pD0 q ď lim∆Ñ0 RrSs|Z pD1 q. Similarly, we can
IprSs; rSs
0
0
prove RS pD0 q ď lim∆Ñ0 RrSs pD01 q. This completes the proof of Theorem 5.
A PPENDIX F
P ROOF OF T HEOREM 6
Denote X 1n , Yi1n , Z 1n as (55)-(57), where ΨK p¨q denotes the orthogonal transform, instead of the permutation operation. Then it can be verified that for Gaussian source-channel pair, the distribution PCS n KS 1n X n Y n Z n X 1n Y 1n Z 1n Sp1n Spn
i
i
i
also satisfies (60) and (61). Hence pΨK , Z n q Ñ Z 1n Ñ S n forms a Markov chain.
Similar to the permutation-based scheme, it is easy to show the power constraint and the distortion constraints
for legitimate users are satisfied. Next by following similar steps to the proof for the permutation-based scheme, we
prove the secrecy constraint is also satisfied. But a slight difference is that here we use an argument from a geometric
point of view, instead of the one from the view of rate-distortion theory (or method of types) used for Theorems
i
33
1, 2, and 5. Here we do not need to discretize the source. But we still need to discretize the reconstruction Sq as
(131), since it will enable us to take the maximizing operation out of the probability, just as done in (141)-(143).
Define event
A fi
(
` n 1n ˘
P Wδn pS, Zq ,
S ,Z
(155)
for δ ą 0. For jointly Gaussian variables X and Z, where Z “ X ` U and U is independent of X, the δ-weakly
typical set and the δ-weakly jointly typical set become
ˇ }xn }2
ˇ
!
)
ˇ
ˇ
Wδn pXq fi xn P X n : ˇ
´ 1ˇ ď δ ,
nNX
(156)
and
Wδn pX, Zq fi
ˇ }xn }2
ˇ
ˇ
ˇ
pxn , z n q P R2n : ˇ
´ 1ˇ ď δ,
nNX
ˇ }z n }2
ˇ
ˇ
ˇ
´ 1ˇ ď δ,
ˇ
nNZ
2
ˇ }xn }2
ˇ
)
}z n ´ xn }
ˇ
ˇ
`
´ 2ˇ ď δ ,
ˇ
nNX
nNU
!
(157)
(158)
(159)
respectively, where NZ , NX and NU denote the variances of Z, X and U , respectively.
Since pS n , Z 1n q is an i.i.d. sequence, from the fact that (weakly) typical set has total probability close to one
[28], we have the following lemma.
Lemma 18. [28] For any δ ą 0, P rAc s Ñ 0, as n Ñ 8.
Following similar steps to the proof of Theorem 5, we can get
” “
ı
‰
PCZ n P dpS n , Sqn q ď D0 |CZ n ą τn
ď |B n | max
PCZ n
n
n
ss PB
rq
nRK
” 2ÿ
ı
n
ss q ą τn1 2´npRn ´RK q ` 1n ,
ξk,zn prq
(160)
k“1
where B n is given by (144), and
“
‰
n
n
ss q fi P d pS n , rq
ss q ď D01 , A|Ψk , Z n .
ξk,zn prq
(161)
Since as shown in (149), |B n | is upper bounded by an exponential function, we only need to show that the
probability in (160) decays doubly exponentially fast with n. To that end, we need introduce the following lemmas.
The proofs of Lemmas 20 and 19 are given in Appendixes G and H, respectively.
Lemma 19. Assume S n “ Z n ` U n , where Z n „ N p0, NZ Iq and U n „ N p0, NU Iq17 are independent, then for
any z n , s̄n P Rn ,
Prd pS n , s̄n q ď D, pS n , z n q P Wδn |z n s ď 2´npRS|Z pDq´op1qq ,
where RS|Z pDq “
17 Here
1
2
log`
`N ˘
U
D
, and op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
0 denotes an all-zero vector and I denotes an identity matrix
(162)
34
Lemma 20. Assume S n „ N p0, NS Iq and S 1n “ ΨS n , with Ψ uniformly distributed on orthogonal matrices set
and independent of S n , then for any s1n , s̄n P Rn ,
Prd pS n , s̄n q ď D, S n P Wδn |s1n s ď 2´npRS pDq´op1qq ,
where RS pDq “
1
2
log`
`N ˘
S
D
(163)
, and op1q is a term that vanishes as δ Ñ 0 and n Ñ 8.
Then we have
n
n
n
ss q ď D01 , S n P Wδn |Ψk , z n s
ss q ď EΨk Prd pS , rq
EC ξk,zn prq
ż
˘
`
n
“ EΨk Prd pS n , rq
ss q ď D01 , S n P Wδn |Ψk , z n , s1n sfS 1n |Z n s1n |Ψk , z n ds1n
ż
˘
`
n
ss q ď D01 , S n P Wδn |s1n sfS 1n |Z n s1n |z n ds1n
“ PΨk rd pS n , rq
1
ď 2´npRS pD0 q´op1qq ,
(164)
(165)
(166)
(167)
where (167) follows from Lemma 20. Furthermore, Lemma 19 implies
“
‰
n
n
n
ss q ď D01 , A|z 1n
ξk,zn prq
ss q “ P d pS , rq
1
ď 2´npRS|Z pD0 q´op1qq ,
(168)
(169)
where (167) follows from ΨK Z n Ñ Z 1n Ñ S n .
Applying Lemmas 12, we have that the probability in (160) decays doubly exponentially fast with n. This
completes the proof of Theorem 6.
A PPENDIX G
P ROOF OF L EMMA 19
Consider that
Prd pS n , s̄n q ď D, pS n , z n q P Wδn |z n s
2
2
2
ď Prd pS n , s̄n q ď D,
}z n }
}S n ´ z n }
}z n }
P 1 ˘ δ,
`
P 2 ˘ δ|z n s
nNZ
nNZ
nNU
ď Prd pS n , s̄n q ď D,
}S n ´ z n }
P 1 ˘ 2δ|z n s.
nNU
(170)
2
Denote R “
1
n
(171)
2
}S n ´ z n } , then
Prd pS n , s̄n q ď D, pS n , z n q P Wδn |z n s
ż NU p1`2δq
ď
fR|Z n pr|z n q Prd pS n , s̄n q ď D|R “ r, Z n “ z n sdr.
(172)
NU p1´2δq
Given Z n “ z n , S n ´ z n „ N p0, NU Iq, and on the other hand, Gaussian distribution is isotropic (or invariant
to rotation). Hence under condition of Z n “ z n and R “ r, S n is uniformly distributed over the sphere with center
?
z n and radius nr.
Prd pS n , s̄n q ď D|R “ r, Z n “ z n s ď
Ω pθq
,
Ω pπq
(173)
35
1
sin
O
Fig. 6. Cap cut out by the cone on the unit sphere.
where
c
θ “ arcsin
D
,
r
(174)
and Ω pθq be solid angle in n space of a cone with half-angle θ, i.e., the area of a spherical cap on a unit sphere
Ωpθq
Ωpπq ,
(see Fig. 6). To approximate
we need the following lemma.
Lemma 21. [33] Let Ω pθq be solid angle in n space of a cone with half-angle θ, then it holds that
ˆ
ˆ ˙˙
Ω pθq
sinn´1 θ
1
“?
1`O
.
Ω pπq
n
2πn cos θ
(175)
Lemma 21 implies
?
1
1
1
Ω pθq
“ 2nplog sin θ´ n logp 2πn sin θ cos θq` n logp1`Op n qqq
Ω pπq
“ 2nplog sin θ`op1qq .
(176)
(177)
Combine (173), (174) and (177), then we have for any r P NU p1 ˘ 2δq,
´
¯
?D
n log
n n
n
n
r `op1q
Prd pS , s̄ q ď D|R “ r, Z “ z s ď 2
´
b
n log N
ď2
U
´
´n
“2
1
2
log
(178)
¯
D
p1´2δq `op1q
NU
D
(179)
¯
´op1q
,
(180)
where op1q is a term that vanishes
as δ ѯ 0 and n Ñ 8. Combining (172) and (180) gives us Prd pS n , s̄n q ď
´
´n
D, pS n , z n q P Wδn |z n s ď 2
1
2
log
NU
D
´op1q
. This completes the proof of Lemma 19.
A PPENDIX H
P ROOF OF L EMMA 20
Observe that
Prd pS n , s̄n q ď D, S n P Wδn |s1n s
2
}S n }
P 1 ˘ δ|s1n s
nNS
#
+
1n 2
` T 1n n ˘
}s
}
“ PΨ rd Ψ s , s̄ ď D|s1n s1
P1˘δ .
nNS
“ Prd pS n , s̄n q ď D,
(181)
(182)
36
Since Ψ is uniformly distributed on orthogonal matrices set (so is ΨT as stated in Lemma (4)), Lemma 5 implies
that for any s1n , ΨT s1n is uniformly distributed over the sphere with center at the origin O and radius }s1n }. Hence
˘
`
Ω pθq
PΨ rd ΨT s1n , s̄n ď D|s1n s ď
,
Ω pπq
(183)
where as described in the previous section, Ω pθq denotes solid angle in n space of a cone with half-angle θ, and
d
D
.
(184)
θ “ arcsin 1
1n 2
n }s }
2
}s1n }
From Lemma 21, we have for any s1n such that
nNS
P 1 ˘ δ,
˜
¸
c
n
˘
PΨ rd ΨT s1n , s̄n ď D|s1n s ď 2
`
log
D
1 s1n }2
n}
`op1q
(185)
´
b
n log N
D
`op1q
S p1´δq
ď2
´
´n
ď2
1
2
log
NS
D
¯
(186)
¯
`op1q
,
(187)
where op1q is a term that vanishes as δ Ñ 0 and n Ñ 8. Combining (182) and (187) gives us
´
n
n
n
Prd pS , s̄ q ď D, S P
Wδn |s1n s
ď2
´n
1
2
log
NS
D
¯
`op1q
.
(188)
This completes the proof of Lemma 20.
A PPENDIX I
P ROOF OF L EMMA 6
q
q “
By choosing Sq to be independent of Z, we have RS|Z pD0 q “
min
IpS; S|Zq
ď
min
IpS; Sq
q
q
PS|SZ
:EdpS,SqďD
PS|S
0
q
q :EdpS,SqďD0
´ ¯
pUBq
RS pD0 q “ 12 log` Dλ0 . Then we only need to prove RS|Z pD0 q ď RS|Z pD0 q.
First consider the case of
λN0
P 1 `N0
ď D0 ď λ. Assume
$
’
&1, with probability p;
Q“
’
%0, with probability 1 ´ p,
independent of pS, Zq, denotes a timesharing random variable, and also assume
$
’
&β0 ΨK Z, if Q “ 1;
SqQ “
’
%0,
if Q “ 0,
where β0 “
?
λP 1
P 1 `N0
(189)
(190)
and ΨK is defined in (36). Then
”
ı
EdpS, SqQ q “ EQ E dpS, SqQ q|Q
“p
λN0
` p1 ´ pq λ.
` N0
P1
(191)
(192)
37
Therefore, to satisfy distortion constraint EdpS, SqQ q ď D0 , it is sufficient to set p “
pλ´D0 qpP 1 `N0 q
.
λP 1
Substituting
SqQ into RS|Z pD0 q, we have
RS|Z pD0 q ď IpS; SqQ |Zq
(193)
ď IpS; SqQ Q|Zq
(194)
“ IpS; SqQ |QZq
(195)
“ pIpS; SqQ |Z, Q “ 1q ` p1 ´ pq IpS; SqQ |Z, Q “ 0q
(196)
“ pIpS; SqQ |Z, Q “ 1q
(197)
ď pIpS; K|Z, Q “ 1q
(198)
ď pHpKq
(199)
“p
(200)
1
“
pλ ´ D0 q pP ` N0 q
,
λP 1
(201)
where (195) follows from Q is independent of pS, Zq.
Next consider the case of 0 ď D0 ď
λN0
P 1 `N0 .
Observe that
RS|Z pD0 q “
q
IpS; S|Zq
min
(202)
q
PS|SZ
:EdpS,SqďD
0
q
min
“
q
IpS; S|Zq
(203)
q
IpS; SK|Zq
(204)
q
PS|SZK
:EdpS,SqďD
0
q
min
ď
q
PS|SZK
:EdpS,SqďD
0
q
“ IpS; K|Zq `
min
q
IpS; S|ZKq
(205)
q
PS|SZK
:EdpS,SqďD
0
q
ď HpKq `
min
q
IpS; S|ZKq,
(206)
q
PS|SZK
:EdpS,SqďD
0
q
where (203) follows since, on one hand, by setting PS|SZK
in (203) as PS|SZ
we have (202)ě(203); on the other
q
q
ř
hand, given PSZK , both the constraint and the optimization objective only depend on PS|SZ
“ k PS|SZK
PK|SZ ,
q
q
hence it suffices to optimize (203) over PS|SZ
.
q
The first term of (206) satisfies
HpKq “ 1.
(207)
By (38) and (39), we have S and ΨK Z are jointly Gaussian, i.e.,
?
where β0 “
λP 1
P 1 `N0 ,
ΨK
S “ β0 ΨK Z ` V01
(208)
´
¯
0
is defined in (36), and V01 „ N 0, PλN
is independent of ΨK Z. Hence we can also
1 `N
0
write
Sq˚ “ β0 ΨK Z ` V02
(209)
38
and
S “ Sq˚ ` ∆V02 ,
where V02
(210)
¯
´
0
and ∆V02 „ N p0, D0 q are independent of each other and also independent of
„ N 0, PλN
1 `N ´ D0
0
ΨK Z. Therefore, we can bound the second term in (206) as
q
IpS; S|ZKq
ď IpS; Sq˚ |ZKq
min
(211)
q
PS|SZK
:EdpS,SqďD
0
q
“ hpS|ZKq ´ hpS|ZK Sq˚ q
(212)
“ hpS|Kq ` hpZ|SKq ´ hpZ|Kq ´ hpS ´ Sq˚ |ZK Sq˚ q
(213)
1
1
1
log 2πeλ ` log 2πeN0 ´ log 2πepP 1 ` N0 q ´ hpS ´ Sq˚ q
2
2
2
ˆ
˙
λN0
1
,
ď log
2
D0 pP 1 ` N0 q
ď
(214)
(215)
where (211) follows since PSq˚ |SZK satisfies the constraint EdpS, Sq˚ q ď D0 , and (214) follows since ΨK Z Ñ
Sq˚ Ñ S forms a Markov chain and S ´ Sq˚ is independent of Sq˚ .
pUBq
Combining (206), (207) and (215) gives us RS|Z pD0 q ď RS|Z pD0 q. This completes the proof of Lemma 6.
A PPENDIX J
P ROOF OF T HEOREM 11
Following similar steps to the proof for the scalar Gaussian case, it is easy to prove Rpiq is achievable by the
permutation based scheme. However, for the orthogonal-transform based scheme, the proof for the scalar Gaussian
case cannot be applied to the vector Gaussian case directly, and some details need to be treated specially. Next we
give a proof for this case.
Following similar steps to the proof of Theorem 6, it can be shown that for vector Gaussian case, the distortion
constraints and power constraint are satisfied for the tuples given in Theorem 11. Next we prove the secrecy
constraint is also satisfied.
Define events
(
` n 1n ˘
Sj , Z j P Wδn pSj , Zj q ,
ź
A fi
Aj ,
Aj fi
(216)
(217)
jPrms
“ ‰
for δ ą 0. Similar to Lemma 18, it can be shown that for any δ ą 0, P Acj Ñ 0, as n Ñ 8.
The derivation up to (160) still holds for vector Gaussian case. Hence
”
ı
“
‰
qn q ď D0 |CZ n ą τn
PCZ n max P dpS n , S
Rn Hcodes
”
ď PCZ n
where D01 fi
`?
max
qn PRmn
s
nRK
2ÿ
k“1
˘2
D0 ` m∆ .
“
n
n
ss q ď
P dpS , rq
D01 , A|Ψj,k , j
P rms, Z
n
‰
ą
τn1 2´npRn ´RK q
ff
ı
` 1n ,
(218)
39
˘
`
ş
śm
n
Observe that P rdpS n , rq
ss q ď D0 , A|Ψj,k , j P rms, Z n s “ řm dpsn ,rqsj sn qďD1 ,A j“1 f snj |Ψj,k , zjn dsnj . One
j“1
j
‰
ş
ś
śm 0 “
n
sj s q ď dj , Aj |Ψj,k , Zjn
may expect to exchange with , in order to write the expression as j“1 P dpSjn , rq
řm
for some dj , j P rms such that j“1 dj ď D01 . However, obviously this is not feasible. To address this problem, we
ş
ř
need to discretize the source, and then eliminate the operation since after discretization it becomes a
operation
with the number of summands polynomial in n.
`S˘
Discretize S by rSs “ ∆ ¨ Round ∆
. Then we have
‰
“
n
P dpS n , rq
ss q ď D01 , A|Ψj,k , j P rms, Z n
‰
“
n
n
ď P d prSs , rq
ss q ď D02 , A|Ψj,k , j P rms, Z n ,
where D02 fi
(219)
`a 1
˘2
D0 ` m∆ . In addition, observe that
‰
“
n
n
P d prSs , rq
ss q ď D01 , A|Ψj,k , j P rms, Z n
m
”ÿ
ı
n
n
sj s q ď D01 , A|Ψj,k , j P rms, Z n
“P
dprSj s , rq
(220)
j“1
“
m
ÿ ź
“
‰
n
n
sj s q “ dj , Aj |Ψj,k , Zjn
P dprSj s , rq
(221)
dPD mn j“1
ď
m
ÿ ź
“
‰
n
P dpSjn , rq
sj s q ď d1j , Aj |Ψj,k , Zjn ,
(222)
dPD mn j“1
where d1j fi
`a
˘2
dj ` ∆ , and
n
n
n
n
n
n
Dmn fi dprss , rq
ss q : rss , rq
ss P B mn , d prss , rq
ss q ď D01
!
)
a
n
n
B mn fi rq
ss P N mn : }rq
sj s } ď nΓj , 1 ď j ď m
(
a
a
a
n
λj p1 ` δq ` ∆ ` mD01 . (221) follows from that Aj implies }rSj s } ď nλj p1 ` δq ` ∆, hence
a
n
it is sufficient to only consider the case of }rq
sj s } ď nΓj . (222) is obtained by using triangle inequality again.
with
a
Γj fi
Combining (218), (219) and (222) gives us
”
ı
“
‰
qn q ď D0 |CZ n ą τn
PCZ n max P dpS n , S
Rn Hcodes
ď |B mn |
max
PCZ n
n
mn
ss PB
rq
nRK
” 2ÿ
ı
n
ss q ą τn1 2´npRn ´RK q ` 1n ,
ξk,zn prq
(223)
k“1
where
m
ÿ ź
“
‰
n
P dpSjn , rq
sj s q ď d1j , Aj |Ψj,k , Zjn .
n
ξk,zn prq
ss q fi
(224)
dPD mn j“1
“
‰
n
n
Given rq
ss and z n , ξk,zn prq
ss q , k P 2nRK are i.i.d. with mean
n
EC ξk,zn prq
ss q “ EC
m
ÿ ź
‰
“
n
sj s q ď d1j , Aj |Ψj,k , Zjn
P dpSjn , rq
(225)
dPD mn j“1
m
ÿ ź
“
dPD mn j“1
‰
“
n
sj s q ď d1j , Aj |Ψj,k , Zjn .
EΨj,k P dpSjn , rq
(226)
40
To bound (223), we need to bound |B mn | and |Dmn | first. Similar to (149), it can be shown
|B
n
mn
|ď2
n
n
2
j“1 log
?
Γj `
∆2
2
q
∆2
`Oplog nq
.
(227)
a
a
n
sj s } ď nΓj , using triangle inequality we have
nΓj , }rq
n
sj s such that }rsj s } ď
In addition, for rsj s , rq
n
?
p
2πe
řm
n
n 2
n
dprsj s , rq
sj s q “ }rsj s ´ rq
sj s } ď 4nΓj .
(228)
Combine it with
n
n
¯2
∆2 ÿ ´
1 ÿ
2
prsi,j s ´ rq
si,j sq “
li,j ´ q
li,j ,
n i“1
n i“1
´ ¯
` si,j ˘
sqi,j
fi Round ∆
and q
li,j fi Round ∆
are both integers, then we have
n
n
dprsj s , rq
sj s q “
where li,j
n ´
ÿ
¯2
li,j ´ q
li,j
ď
i“1
In addition,
řn
i“1
4n2 Γj
.
∆2
(229)
(230)
¯2
´
li,j ´ q
li,j
P N Y t0u, hence
|D
mn
m ˆ
ź
4n2 Γj
|ď
j“1
∆2
˙
`1 .
(231)
That is, |Dmn | is bounded by a polynomial term of n.
If we can show that the probability in (223) decays doubly exponentially fast with n, then the proof will be
complete. To that end, by using Lemma 19 we have
řm
n
ξk,zn prq
ss q ď |Dmn | 2´np
j“1
RSj |Zj pd1j q´op1qq
ř
1
´np m
j“1 RSj |Zj pdj q´op1qq
(232)
“2
(233)
ď 2´nmpRS|Z pD0 `∆ q´op1qq ,
(234)
where RS|Z pDq given in (52) denotes the conditional rate-distortion function for source S with side information
Z at both encoder and decoder, op1q is a term that vanishes as δ Ñ 0 and n Ñ 8, and ∆ is a term that vanishes
as ∆ Ñ 0. (233) follows from that |Dmn | grows only polynomially fast with n.
From Lemma 20, we have for any z n ,
n
EC ξk,zn prq
ss q ď 2´nmpRS pD0 `∆ q´op1qq ,
(235)
where RS pDq given in (51) denotes the point-to-point rate-distortion function for S.
Using these bounds and applying Lemmas 12, we have that the probability in (223) decays doubly exponentially
fast with n. This completes the proof of Theorem 11.
R EFERENCES
[1] C. E. Shannon, “Communication in the presence of noise,” Proc. IRE, vol. 37, no. 1, pp. 10-21, Jan. 1949.
[2] T. J. Goblick, “Theoretical limitations on the transmission of data from analog sources,” IEEE Trans. Inf. Theory, vol. IT-11, no. 4, pp.
558-567, Oct. 1965.
41
[3] M. Gastpar, B. Rimoldi, and M. Vetterli, “To code, or not to code: lossy source-channel communication revisited,” IEEE Trans. Inf. Theory,
vol. 49, no. 5, pp. 1147-1158, May 2003.
[4] K. H. Lee, and D. P. Petersen, “Optimal linear coding for vector channels,” IEEE Trans. Commun., vol. 24, no. 12, pp. 1283-1290, Dec.
1976.
[5] S. Shamai, S. Verdú, and R. Zamir, “Systematic lossy source/channel coding,” IEEE Trans. Inf. Theory, vol. 44, no. 3, pp. 564-579, Mar.
1998.
[6] V. Prabhakaran, R. Puri, and K. Ramchandran, “Hybrid digital-analog codes for source-channel broadcast of Gaussian sources over Gaussian
channels,” IEEE Trans. Inf. Theory, vol. 57, no. 7, pp.4573-4588, Aug. 2011.
[7] L. Yu, H. Li, and W. Li, “Wireless scalable video coding using a hybrid digital-analog scheme,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 24, no. 2, pp. 331-345, Feb. 2014.
[8] L. Yu, H. Li, and W. Li, “Wireless cooperative video coding using a hybrid digital-analog scheme,” IEEE Trans. Circuits Syst. Video
Technol., vol. 25, no. 3, pp. 436-450, Mar. 2015.
[9] L. Yu, H. Li, and W. Li, “Distortion bounds for source broadcast over degraded channel,” IEEE Int. Symp. Inf. Theory (ISIT), 2016.
[10] L. Yu, H. Li, and W. Li, “Distortion bounds for source broadcast problem,” Submitted to IEEE Trans. Inf. Theory, 2016.
[11] Z. Reznic, M. Feder, and R. Zamir, “Distortion bounds for broadcasting with bandwidth expansion,” IEEE Trans. Inf. Theory, vol. 52, no.
8, pp. 3778-3788, Aug. 2006.
[12] C. Tian, S. Diggavi, and S. Shamai, “Approximate characterizations for the Gaussian source broadcasting distortion region,” IEEE Trans.
Inf. Theory, vol. 57, no. 1, pp. 124-136, Jan. 2011.
[13] C. E. Shannon, “Communication theory of secrecy systems,” Bell Syst. Tech. J., vol. 28, no. 4, pp. 656–715, Oct. 1949.
[14] C. Schieler, and P. Cuff, “The henchman problem: Measuring secrecy by the minimum distortion in a list,” IEEE Trans. Inf. Theory, vol.
62, no. 6, pp. 3436–3450, Jun. 2016.
[15] L. Yu, H. Li, and W. Li, “Source-channel secrecy for Shannon cipher system,” IEEE Trans. Inf. Theory, vol. 63, no. 4, pp. 2596-2622,
Apr. 2017.
[16] H. Yamamoto, “Rate-distortion theory for the Shannon cipher system,” IEEE Trans. Inf. Theory, vol. 43, no. 3, pp. 827–835, May 1997.
[17] M. P. Wilson, and K. Narayanan, “Transmitting an analog Gaussian source over a Gaussian wiretap channel under SNR mismatch,” IEEE
International Conference on Telecommunications, pp.44-47, April 2010.
[18] G. Bagherikaram and K. Plataniotis, “Secure hybrid digital-analog Wyner-Ziv coding,” in IEEE 22nd International Symposium on Personal
Indoor and Mobile Radio Communications, pp. 1161 –1166, Sep. 2011.
[19] G. Bagherikaram and K. Plataniotis, “Secure joint source-channel coding with interference known at the transmitter,” IET Communications,
vol. 6, no. 17, pp. 2796 - 2808, Jan. 2013.
[20] S. C. Kak and N. S. Jayant, “On speech encryption using waveform scrambling,” Bell Syst. Tech. J., vol. 56, pp. 781-808, May-Jun. 1977.
[21] A. D. Wyner, “An analog scrambling scheme which does not expand bandwidth, Part I: Discrete time,” IEEE Trans. Inf. Theory, vol.
IT-25, pp. 261-274, May 1979.
[22] R. Ahlswede and G. Dueck, “Good codes can be produced by a few permutations,” IEEE Trans. Inf. Theory, vol. 28, no. 3, pp. 430–443,
May 1982.
[23] W. Kang and N. Liu “Compressing encrypted data: Achieving optimality and strong secrecy via permutations,” IEEE Trans. on Inf. Theory,
vol. 62, no. 12, pp. 7153-7163, Dec. 2016.
[24] D. Slepian, “Permutation modulation,” Proc. IEEE, vol. 53, pp. 228–236, Mar. 1965.
[25] D. Slepian, “Group codes for the Gaussian channel,” Bell Syst. Tech. J., vol. 47, pp. 575–602, Apr. 1968.
[26] T. Ericson, “A min-max theorem for antijamming group codes,” IEEE Trans. Inf. Theory, vol. IT-30, pp. 792–799, Nov. 1984.
[27] A. Lapidoth and S. Tinguely, “Sending a bivariate Gaussian over a Gaussian MAC,” IEEE Trans. Inf. Theory, vol. 56, no. 6, pp. 2714–2752,
Jun. 2010.
[28] T. M. Cover and J. A. Thomas, Elements of Information Theory, Wiley, New York, 1991.
[29] I. Csiszár, and J. Körner, Information theory: coding theorems for discrete memoryless systems. Cambridge University Press, 2011.
[30] A. El Gamal and Y.-H. Kim, Network Information Theory. Cambridge University Press, 2011.
[31] M. L. Eaton, Multivariate Statistics: A Vector Space Approach, Wiley and Sons, New York, NY, 1983.
[32] I. Sason and S. Verdú “f -divergence inequalities,” IEEE Trans. on Inf. Theory, vol. 62, no. 11, pp. 5973–6006, Nov. 2016.
[33] C. E. Shannon, “Probability of error for optimal codes in a Gaussian channel,” Bell Syst. Tech. J., vol. 38, pp. 611–656, 1959.
42
[34] S.-W. Ho and R. W. Yeung, “On information divergence measures and a unified typicality,” IEEE Trans. Inf. Theory, vol. 56, no. 12, pp.
5893–5905, Dec. 2010.
| 7cs.IT
|
Size-Independent Sample Complexity of Neural Networks
arXiv:1712.06541v2 [cs.LG] 10 Jan 2018
Noah Golowich
Harvard University
Alexander Rakhlin
University of Pennsylvania
Ohad Shamir
Weizmann Institute of Science
and Microsoft Research
Abstract
We study the sample complexity of learning neural networks, by providing new bounds on their
Rademacher complexity assuming norm constraints on the parameter matrix of each layer. Compared
to previous work, these complexity bounds have improved dependence on the network depth, and under
some additional assumptions, are fully independent of the network size (both depth and width). These
results are derived using some novel techniques, which may be of independent interest.
1 Introduction
One of the major challenges involving neural networks is explaining their ability to generalize well, even if
they are very large and have the potential to overfit the training data [Neyshabur et al., 2014, Zhang et al.,
2016]. Learning theory teaches us that this must be due to some inductive bias, which constrains one to learn
networks of specific configurations (either explicitly, e.g., via regularization, or implicitly, via the algorithm
used to train them). However, understanding the nature of this inductive bias is still largely an open problem.
A useful starting point is to consider the much more restricted class of linear predictors (x 7→ w⊤ x).
For this class, we have a very good understanding of how its generalization behavior is dictated by the norm
of w . In particular, assuming that kwk ≤ M (where k · k signifies Euclidean norm), and the distribution
is such that kxk ≤ B almost surely, it is well-known that the generalization error (w.r.t. Lipschitz losses)
√
given m training examples scales as O(M B/ m), completely independent of the dimension of w. Thus,
it is very natural to ask whether in the more general case of neural networks, one can obtain similar “sizeindependent” results (independent of the networks’ depth and width), under appropriate norm constraints on
the parameters. This is also a natural question, considering that the size of modern neural networks is often
much larger than the number of training examples.
Classical results on the sample complexity of neural networks (see Anthony and Bartlett [2009]) do not
satisfy this desideratum, and have a strong explicit dependence on the network size. In particular, they
can be trivial once the number of parameters exceeds the number of training examples. More recently,
there have been several works aiming at improving the sample complexity bounds, assuming various norm
constraints on the parameter matrices. For example, Neyshabur et al. [2015] use Rademacher complexity
tools to show that if the parameter matrices W1 , . . . , Wd in each of the d layers have Frobenius norms upperbounded by MF (1), . . . , MF (d) respectively, and under suitable assumptions on the activation functions,
the generalization error scales (with high probability) as
!
Q
B2d dj=1 MF (j)
√
.
(1)
O
m
1
Although this bound has no explicit dependence on the network width (that is, the dimensions of W1 , . . . ,
Wd ), it has a very strong, exponential dependence on the network depth d, even if MF (j) ≤ 1 for all
j. Neyshabur et al. [2015] also showed that this dependence can sometimes be avoided for anti-symmetric
activations, but unfortunately this is a non-trivial assumption, which is not satisfied for common activations
such as the ReLU. Bartlett et al. [2017] use a covering numbers argument to show a bound scaling as
2/3 !3/2
Q
P
T
kWj k2,1
d
d
B
j=1
j=1 kWj k
kWj k
√
Õ
(2)
,
m
P pP
2
where kW k denotes the spectral norm of W , kW T k2,1 := l
k Wj (l, k) denotes the 1-norm of the
2-norms of the rows of W , and where we ignore factors logarithmic in m and the network width. Unlike
Eq. (1), here there is no explicit exponential dependence on the depth. However,
there is still a strong and
√
unavoidable polynomial dependence: To see this, note that for any Wj ,
the bound above can never be smaller than
r
d
Y
d3
kWj k
.
Õ B
m
P
l
P
2
k Wj (l,k)
kWj k
≥
kWj kF
kWj k
≥ 1, so
j=1
In particular, even if we assume that B
Ω(m1/3 ).
Q
d
j=1 kWj k
is a constant, the bound becomes trivial once d ≥
Finally, and using the same notation, Neyshabur et al. [2017] utilize a PAC-Bayesian analysis to
prove a bound scaling as
v
u
2
P
kW
k
d
j
u
F
2
d
Y
t d h i=1 kWj k2
,
kW
k
B
Õ
(3)
j
m
j=1
kW k
where h denotes the network width.1 Again, since kWjj kF ≥ 1 for any parameter matrix Wj , this bound
q
√
Qd
d3 h
can never be smaller than Õ B
, and becomes trivial once d 3 h ≥ Ω(m1/3 ). To
j=1 kWj k
m
summarize, although some of the bounds above have logarithmic or no dependence on the network width,
we are not aware of a bound in the literature which avoids a strong dependence on the depth, even if various
norms are controlled.
Can this depth dependency be avoided, assuming the norms are sufficiently constrained? We argue that
in some cases, it must be true. To see this, let us return to the well-understood case of linear predictors, and
consider generalized linear predictors of the form
n
o
x 7→ σ w⊤ x : kwk ≤ M ,
where σ(z) = max{0, z} is the ReLU function. Like plain-vanilla linear predictors, the generalization
√
error of this class is well-known to be O(M B/ m), assuming the inputs satisfy kxk ≤ B almost surely.
1
Bartlett et al. [2017] note that this PAC-Bayesian bound is never better than the bound in Eq. (2) derived from a covering
numbers argument.
2
However, it is not difficult to show that this class can be equivalently written as a class of “ultra-thin” ReLU
networks of the form
d
Y
|wj | ≤ M ,
(4)
x 7→ σ(wd · σ(. . . w2 · σ(w1⊤ x))) : kw1 k ·
j=2
(where w1 is a vector and w2 , . . . , wd are scalars), where the depth d is arbitrary. Therefore, the sample
√
complexity of this class must also scale as O(M B/ m): This depends on the norm product M , but is completely independent of the network depth d as well as the dimension of w1 . We argue that a “satisfactory”
sample complexity analysis should have similar independence properties when applied on this class.
In more general neural networks, the vector w1 and scalars w2 , . . . , wd become matrices W1 , . . . , Wd ,
and the simple observation above no longer applies.
Qd However, using the same intuition, it is natural to try
and derive generalization bounds by controlling j=1 kWj k, where k · k is a suitable matrix norm. Perhaps
the simplest choice is the spectral norm (and indeed, a product of spectral norms was utilized in some of the
previous results mentioned earlier). However, as we formally show in Sec. 5, the spectral norm alone is too
weak to get size-independent bounds, even if the network depth is small. Instead, we show that controlling
other suitable norms can indeed lead to better depth dependence, or even fully size-independent bounds,
improving on earlier works. Specifically, we make the following contributions:
• In Sec. 3, we show that the exponential depth dependence in Rademacher complexity-based analysis
(e.g. Neyshabur et al. [2015]) can be avoided by applying contraction to a slightly different object than
what has become standard since the work of Bartlett and Mendelson [2002]. For example, for networks with parameter matrices of Frobenius norm at most MF (1), . . . , MF (d), the bound in Eq. (1)
can be improved to
!
√ Q
B d dj=1 MF (j)
√
O
.
(5)
m
The technique can also be applied to other types of norm constraints. For example, if we consider an
ℓ1 /ℓ∞ setup, corresponding to the class of depth-d networks, where the 1-norm of each row of Wj is
at most M (j), we attain a bound of
!
p
Q
B d + log(n) · dj=1 M (j)
√
,
O
m
where n is the input dimension. Again, the dependence on d is polynomial and quite mild. In contrast,
Neyshabur et al. [2015] studied a similar setup and only managed to obtain an exponential dependence
on d.
• In Sec. 4, we develop a generic technique to convert depth-dependent bounds to depth-independent
bounds, assuming some control over any Schatten norm of the parameter matrices (which includes,
for instance, the Frobenius norm and the trace norm as special cases). The key observation we utilize
is that the prediction function computed by such networks can be approximated by the composition of
a shallow network and univariate Lipschitz functions. For example, again assuming that the Frobenius
norms of the layers are bounded by MF (1), . . . , MF (d), we can further improve Eq. (5) to
(
r )
d
Y
d
1
MF (j) · min
Õ B
,
.
(6)
1/4
m
m
j=1
3
Q
ignoring logarithmic factors. Note that this can be upper bounded by Õ(B dj=1 MF (j)/m1/4 ),
which to the best of our knowledge, is the first explicit bound for standard neural networks which
is fully size-independent, assuming only suitable norm constraints. Moreover, it captures the depthindependent sample complexity behavior of the network class in Eq. (4) discussed earlier. We also
apply this technique to get a depth-independent version of the bound in [Bartlett et al., 2017]: SpecifkW T k2,1
j
ically, if we assume that kWj k ≤ M (j) for all j, and maxj kW
jk
provided by Bartlett et al. [2017] becomes
r
d
Y
d3
.
M (j) ·
Õ BL
m
≤ L, then the bound in Eq. (2)
j=1
In contrast, we show the following bound for any p ≥ 1 (ignoring some lower-order logarithmic
factors):
Q
21
d
+p r
B
d
3
log
M
(j)
Y
p
j=1
Γ
d3
M (j) · min
Õ BL
,
,
1
m
m 2+3p
j=1
where Mp (j) is a bound on the Schatten p-norm of Wj , and Γ is the largest possible Euclidean norm
of the network’s output, over all input vectors of norm at most B. Again, by upper bounding the min
by its first argument, we get a bound independent of the depth d, assuming the norms are suitably
constrained.
• In Sec. 5, we provide a lower bound, showing that for any p, the class of depth-d, width-h neural
networks, where each parameter matrix Wj has Schatten p-norm at most Mp (j), can have Rademacher
complexity of at least
o
n
Qd
max 0, 12 − p1
B j=1 Mp (j) · h
√
Ω
.
m
This somewhat improves on Bartlett et al. [2017, Theorem 3.6], which only showed such a result for
p = ∞ (i.e. with spectral norm control), and without the h term. For p = 2, it matches the upper
bound in Eq. (6) in terms of the norm dependencies and B. Moreover, it establishes that controlling
the spectral norm alone (and indeed, any Schatten p-norm control with p > 2) cannot lead to bounds
independent of the size of the network. Finally, the bound shows (similar to Bartlett et al. [2017]) that
a dependence on products of norms across layers is generally inevitable.
Besides the above, we provide some additional remarks and observations in Sec. 6. Most of our technical
proofs are presented in Sec. 7.
2 Preliminaries
Notation. We use bold-faced letters to denote vectors, and capital letters to denote matrices or fixed parameters (which should be clear from context). Given a vector w ∈ Rh , kwk will refer to the Euclidean norm,
1/p
P
h
p
and for p ≥ 1, kwkp =
will refer to the ℓp norm. For a matrix W , we use kW kp , where
|w
|
i
i=1
p ∈ [1, ∞], to denote the Schatten p-norm (that is, the p-norm of the spectrum of W , written as a vector).
4
For example, p = ∞ refers to the spectral norm, p = 2 refers to the Frobenius norm, and p = 1 refers to
the trace norm. For the case of the spectral norm, we will drop the ∞ subscript, and use just kW k. Also, in
order to follow standard convention, we use kW kF to denote the Frobenius norm. Finally, given a matrix
q/p 1/q
P P
p
denote the q-norm of the p-norms of
W and reals p, q ≥ 1, we let kW kp,q :=
k
j |Wj,k |
the columns of W .
Neural Networks. Given the domain X = {x : kxk ≤ B} in Euclidean space, we consider (scalar or
vector-valued) standard neural networks, of the form
x 7→ Wd σd−1 (Wd−1 σd−2 (. . . σ1 (W1 x))),
where each Wj is a parameter matrix (with some fixed dimensions), and each σj is some fixed Lipschitz
continuous function between Euclidean spaces, satisfying σj (0) = 0. In the above, we denote d as the depth
of the network, and its width h is defined as the maximal row or column dimensions of W1 , . . . , Wd . Without
loss of generality, we will assume that σj has a Lipschitz constant of at most 1 (otherwise, the Lipschitz
constant can be absorbed into the norm constraint of the neighboring parameter matrix). We say that σ is
element-wise if it can be written as an application of the same univariate function over each coordinate of
its input (in which case, somewhat abusing notation, we will also use σ to denote that univariate function).
We say that σ is positive-homogeneous if it is element-wise and satisfies σ(αz) = ασ(z) for all α ≥ 0 and
z ∈ R. An important example of the above are ReLU networks, where every σj corresponds to applying the
(positive-homogeneous) ReLU function σ(z) = max{0, z} on each element. To simplify notation, we let
Wbr be shorthand for the matrix tuple {Wb , Wb+1 , . . . , Wr }, and NWbr denote the function computed by the
sub-network composed of layers b through r, that is
x 7→ Wr σr−1 (Wr−1 σr−2 (. . . σb (Wb x))) .
Rademacher Complexity. The results in this paper focus on Rademacher complexity, which is a standard tool to control the uniform convergence (and hence the sample complexity) of given classes of predictors (see Bartlett and Mendelson [2002], Shalev-Shwartz and Ben-David [2014] for more details). Formally,
given a real-valued function class H and some set of data points x1 , . . . , xm ∈ X , we define the (empirical)
Rademacher complexity R̂m (H) as
#
"
m
1 X
εi h(xi ) ,
(7)
R̂m (H) = Eǫ sup
h∈H m
i=1
where ε = (ε1 , . . . , εm ) is a vector uniformly distributed in {−1, +1}m . Our main results provide bounds
on the Rademacher complexity (sometimes independent of x1 , . . . , xm , as long as they are assumed to have
norm at most B), with respect to classes of neural networks with various norm constraints. Using standard
arguments, such bounds can be converted to bounds on the generalization error, assuming access to a sample
of m i.i.d. training examples.
3 From Exponential to Polynomial Depth Dependence
To get bounds on the Rademacher complexity of deep neural networks, a reasonable approach (employed in
Neyshabur et al. [2015]) is to use a “peeling” argument, where the complexity bound for depth d networks
is reduced to a complexity bound for depth d − 1 networks, and then applying the reduction d times. For
5
example, consider the class Hd of depth-d ReLU real-valued neural networks, with each layer’s parameter
matrix with Frobenius norm at most MF (j). Using some straightforward manipulations, it is possible to
show that R̂m (Hd ), which by definition equals
Eǫ sup
h∈Hd
m
m
i=1
i=1
1 X
1 X
ǫi h(xi ) = Eǫ sup
ǫi Wd σ(h(xi )),
sup
m
h∈Hd−1 Wd :kWd kF ≤MF (d) m
can be upper bounded by
m
MF (d) · Eǫ sup
h∈Hd−1
1 X
ǫi σ(h(xi ))
m
i=1
m
≤ 2MF (d) · Eǫ sup
h∈Hd−1
1 X
ǫi h(xi )) .
m
(8)
i=1
Q
Iterating this inequality d times, one ends up with a bound scaling as 2d dj=1 MF (j) (as in Neyshabur et al.
[2015], see also Eq. (1)). The exponential 2d factor follows from the 2 factor in Eq. (8), which in turn
follows from applying the Rademacher contraction principle to get rid of the σ function. Unfortunately, this
2 factor is generally unavoidable (see the discussion in Ledoux and Talagrand [1991] following Theorem
4.12).
In this section, we point out a simple trick, which can be used to reduce such exponential depth dependencies to polynomial ones. In a nutshell,
P using Jensen’s inequality, we can rewrite the (scaled) Rademacher
complexity m · R̂m (H) = Eǫ suph∈H m
i=1 ǫi h(xi ) as
!
!!
m
m
X
X
1
1
ǫi h(xi ) ≤ log Eǫ sup exp λ
ǫi h(xi )
,
log exp λ · Eǫ sup
λ
λ
h∈H
h∈H
i=1
i=1
where λ > 0 is an arbitrary parameter. We then perform a “peeling” argument similar to before, resulting
in a multiplicative 2 factor after every peeling step. Crucially, these factors accumulate inside the log factor,
so that the √
end result contains only a log(2d ) = d factor, which by appropriate tuning of λ, can be further
reduced to d.
The formalization of this argument depends on the matrix norm we are using, and we will begin with
the case of the Frobenius norm. A key technical condition for the argument to work is that we can perform
the “peeling” inside the exp function. This is captured by the following lemma:
Lemma 1. Let σ be a 1-Lipschitz, positive-homogeneous activation function which is applied element-wise
(such as the ReLU). Then for any vector-valued class F, and any convex and monotonically increasing
function g : R → [0, ∞),
!
!
m
m
X
X
.
≤ 2 · Eǫ sup g R ·
Eǫ
sup
g
ǫi f (xi )
ǫi σ(W f (xi ))
f ∈F ,W :kW kF ≤R
f ∈F
i=1
i=1
Proof. Letting w1 , w2 , . . . , wh be the rows of the matrix W , we have
m
X
i=1
2
ǫi σ(W f (xi ))
=
h
X
j=1
kwj k2
6
m
X
i=1
ǫi σ
wj⊤
kwj k
!!2
f (xi )
.
P
The supremum of this over all w1 , . . . , wh such that kW k2F = hj=1 kwj k2 ≤ R2 must be attained when
kwj k = R for some j, and kwi k = 0 for all i 6= j. Therefore,
!
!
m
m
X
X
⊤
ǫi σ(W f (xi ))
Eǫ
sup
g
= Eǫ
sup
g
ǫi σ(w f (xi ))
.
f ∈F ,W :kW kF ≤R
f ∈F ,w:kwk=R
i=1
i=1
Since g(|z|) ≤ g(z) + g(−z), this can be upper bounded by
!
!
!
m
m
m
X
X
X
⊤
⊤
⊤
ǫi σ(w f (xi )) ,
ǫi σ(w f (xi )) = 2·Eǫ sup g
ǫi σ(w f (xi )) +Eǫ sup g −
Eǫ sup g
i=1
i=1
i=1
where the equality follows from the symmetry in the distribution of the ǫi random variables. The right hand
side in turn can be upper bounded by
!
!
m
m
X
X
⊤
ǫi f (xi )
ǫi w f (xi ) ≤ 2 · Eǫ
sup
g kwk
2 · Eǫ
sup
g
f ∈F ,w:kwk=R
f ∈F ,w:kwk=R
i=1
= 2 · Eǫ sup g R ·
f ∈F
i=1
m
X
ǫi f (xi )
i=1
!
.
(see equation 4.20 in Ledoux and Talagrand [1991]).
With this lemma in hand, we can provide a bound on the Rademacher complexity of Frobnius-norm√
bounded neural networks, which is as clean as Eq. (1), but where the 2d factor is replaced by d:
Theorem 1. Let Hd be the class of real-valued networks of depth d over the domain X , where each parameter matrix Wj has Frobenius norm at most MF (j), and with activation functions satisfying Lemma 1.
Then
Q
p
v
d
um
d
B
2
log(2)d
+
1
X
Y
p
u
j=1 MF (j)
1
2
t
√
kxi k ≤
MF (j) ·
.
2 log(2)d + 1
R̂m (Hd ) ≤
m
m
i=1
j=1
Proof. Fix λ > 0, to be chosen later. The Rademacher complexity can be upper bounded as
mR̂m (Hd ) = Eǫ
sup
N
W1d−1
m
X
,Wd i=1
ǫi Wd σd−1 (NW d−1 (xi ))
1
!
m
X
1
≤ log Eǫ sup exp λ
ǫi Wd σd−1 (NW d−1 (xi ))
1
λ
i=1
m
X
1
≤ log Eǫ sup exp MF (d) · λ
ǫi σd−1 (NW d−1 (xi ))
1
λ
i=1
!
We write this last expression as
m
X
1
log Eǫ
sup
exp MF (d) · λ
ǫi σd−1 (Wd−1 f (xi ))
λ
f,kWd−1 kF ≤MF (d−1)
i=1
!!
m
X
1
ǫi f (xi )
≤ log 2 · Eǫ sup exp MF (d) · MF (d − 1) · λ
λ
f
i=1
7
!
where f ranges over all possible functions σd−2 ◦ NW d−2 (x). Here we applied Lemma 1 with g(α) =
1
exp{MF (d)λ · α}. Repeating the process, we arrive at
!!
m
X
1
(9)
ǫi xi
mR̂m (Hd ) ≤ log 2d · Eǫ exp M λ
λ
i=1
where M =
Qd
j=1 MF (j).
Define a random variable
Z=M·
m
X
ǫi xi ,
i=1
(random as a function of the random variables ǫ1 , . . . , ǫm ). Then
n
o d log(2) 1
1
log 2d · E exp λZ =
+ log {E exp λ(Z − EZ)} + EZ.
λ
λ
λ
(10)
By Jensen’s inequality, E[Z] can be upper bounded by
v
v
v
u
u
um
2
m
m
u
u
X
X
uX
u
u
⊤
M tEǫ
ǫi xi = M tEǫ
kxi k2 .
ǫi ǫi′ xi xi′ = M t
i,i′ =1
i=1
i=1
To handle the log {E exp λ(Z − EZ)} term in Eq. (10), note that Z is a deterministic function of the i.i.d.
random variables ǫ1 , . . . , ǫm , and satisfies
Z(ǫ1 , . . . , ǫi , . . . , ǫm ) − Z(ǫ1 , . . . , −ǫi , . . . , ǫm ) ≤ 2M kxi k .
This means that Z satisfies a bounded-difference condition, which by the proof of Theorem 6.2 in [Boucheron et al.,
2013], implies that Z is sub-Gaussian, with variance factor
m
m
i=1
i=1
X
1X
kxi k2 ,
(2M kxi k)2 = M 2
v=
4
and satisfies
Choosing λ =
P
P
2
2
1
1 λ2 M 2 m
λM 2 m
i=1 kxi k
i=1 kxi k
log {E exp λ(Z − EZ)} ≤
=
.
λ
λ
2
2
√
2 log(2)d
√P
and using the above, we get that Eq. (9) can be upper bounded as follows:
m
2
M
i=1
kxi k
v
v
um
um
n
o
X
p
p
u
uX
1
d
2
t
2 log(2)d + 1 t
kxi k ≤ M
kxi k2 ,
log 2 · E exp λZ ≤ EZ + 2 log(2)d · M
λ
i=1
i=1
from which the result follows.
Remark 1. We note that for simplicity, the bound in Thm. 1 is stated for real-valued networks, but the
argument easily carries over to vector-valued networks, composed with some real-valued Lipschitz loss
function. In that case, one uses a variant of Lemma 1 to peel off the losses, and then proceed in the same
manner as in the proof of Thm. 1. We omit the precise details for brevity.
8
A result similar to the above can also be derived for other entry-based matrix norms. For example,
given a matrix W , let kW k1,∞ denote the maximal 1-norm of its rows, and consider the class Hd of depthd networks, where each parameter matrix Wj satisfies kWj k1,∞ ≤ M (j) for all j (this corresponds to
a setting, also studied in Neyshabur et al. [2015], where the 1-norm of the weights of each neuron in the
network is bounded). In this case, we can derive a variant of Lemma 1, which in fact does not require
positive-homogeneity of the activation function:
Lemma 2. Let σ be a 1-Lipschitz activation function with σ(0) = 0, applied element-wise. Then for any
vector-valued class F, and any convex and monotonically increasing function g : R → [0, ∞),
!
!
m
m
X
X
,
(11)
≤ 2 · Eǫ sup g R ·
ǫi f (xi )
Eǫ
sup
g
ǫi σ(W f (xi ))
f ∈F ,W :kW k1,∞ ≤R
i=1
f ∈F
∞
i=1
∞
where k · k∞ denotes the vector infinity norm.
Using the same technique as before, we can use this lemma to get a bound on the Rademacher complexity
for Hd :
Theorem 2. Let Hd be the class of real-valued networks of depth d over the domain X , where kWj k1,∞ ≤
M (j) for all j ∈ {1, . . . , d}, and with activation functions satisfying the condition of Lemma 2. Then
v
p
Q
u
m
d
X
Y
p
u
2B d + 1 + log(n) · dj=1 M (j)
2
2
√
xi,j ≤
M (j) · d + 1 + log(n) · t max
,
R̂m (Hd ) ≤
m
m
j∈{1,...,n}
i=1
j=1
where xi,j is the j-th coordinate of the vector xi .
The proofs of the theorem, as well as Lemma 2, appear in Sec. 7.
The constructions used in the results of this section use the function g(z) = exp(λz) together with
its inverse g−1 (z) = λ1 log(z), to get depth dependencies scaling as g−1 (2d ). Thus, it might be tempting
to try and further improve the depth dependence, by using other functions gP
for which g−1 increases sublogarithmically. Unfortunately, the argument still requires us to control Eǫ g (k m
i=1 ǫi xi k), which is difficult
if g increases more than exponentially. In the next section, we introduce a different idea, which under
suitable assumptions, allows us to get rid of the depth dependence altogether.
4 From Depth Dependence to Depth Independence
In this section, we develop a general result, which allows one to convert any depth-dependent bound on
the Rademacher complexity of neural networks, to a depth-independent one, assuming that the Schatten
p-norms of the parameter matrices (for any p ∈ [1, ∞)) is controlled. We develop and formalize the main
result in Subsection 4.1, and provide applications in Subsection 4.2. The proofs the results in this section
appear in Sec. 7.
4.1 A General Result
To motivate our approach, let us consider a special case of depth-d networks, where
• Each parameter matrix W1 , . . . , Wd−1 is constrained to be diagonal and of size h × h.
9
• The Frobenius norm of every W1 , . . . , Wd is at most 1.
• All activation functions are the identity (so the network computes a linear function).
Letting wi be the diagonal of Wi , such networks are equivalent to
x 7→ (wd ◦ wd−1 ◦ . . . ◦ w1 )⊤ x ,
where ◦ denotes element-wise product. Therefore, if we would like the network to compute a non-trivial
function, we clearly need that wd ◦ . . . ◦ w1 be bounded away from zero (e.g., not exponentially small in
d), while still satisfying the constraint kwj k ≤ 1 for all j. In fact, the only way to satisfy both requirements
simultaneously is if w1 , . . . , wd are all close to some 1-sparse unit vector, which implies that the matrices
W1 , . . . , Wd must be close to being rank-1.
It turns out that this intuition holds much more generally, even if we do not restrict ourselves to identity
activations and diagonal parameter matrices as above. Essentially, what we can show is that if some network
computes a non-trivial function, and the product of its Schatten p-norms (for any p < ∞) is bounded, then
there must be at least one parameter matrix, which is not far from being rank-1. Therefore, if we replace that
parameter matrix by an appropriate rank-1 matrix, the function computed by the network does not change
much. This is captured by the following theorem:
Q
Theorem 3. For any p ∈ [1, ∞), any network NW d such that supx∈X kNW d (x)k ≥ Γ and dj=1 kWj kp ≤
1
1
M , and for any r ∈ {1, . . . , d}, there exists another network NW̃ d (of the same depth and layer dimensions)
1
with the following properties:
• W̃1d = {W̃1 , . . . , W̃d } is identical to W1d , except for the parameter matrix W̃r in the r-th layer, which
is of rank at most 1, and equals suv⊤ where s, u, v are some leading singular value and singular
vectors pairs of Wr .
1/p
Q
2p log(M B/Γ)
d
.
• supx∈X kNW d (x) − NW̃ d k ≤ B
kW
k
j
j=1
r
1
1
We now make the following crucial observation: A real-valued network with a rank-1 parameter matrix
Wr = suv⊤ computes the function
x 7→ Wd σd−1 (. . . σr (suv⊤ σr−1 (. . . σ1 (W1 x) . . .))) .
This can be seen as the composition of the depth-r network
x 7→ v⊤ σr−1 (. . . σ1 (W1 x) . . .),
and the univariate function
x 7→ Wd σd−1 (. . . σr (sux)) .
Moreover, the norm constraints imply that the latter function is Lipschitz. Therefore, the class of networks
we are considering is a subset of the class of depth-r networks composed with univariate Lipschitz functions. Fortunately, given any class with bounded complexity, one can effectively bound the Rademacher
complexity of its composition with univariate Lipschitz functions, as formalized in the following theorem.
10
Theorem 4. Let H be a class of functions from Euclidean space to [−R, R]. Let FL,a be the class of of LLipschitz functions from [−R, R] to R, such that f (0) = a for some fixed a. Letting FL,a ◦ H := {f (h(·)) :
f ∈ FL,a , h ∈ H}, its Rademacher complexity satisfies
R
3/2
R̂m (FL,a ◦ H) ≤ cL √ + log (m) · R̂m (H) ,
m
where c > 0 is a universal constant.
Remark 2. The log3/2 (m) · R̂m (H) can be replaced by log(m) · Ĝm (H), where Ĝm (H) is the empirical
Gaussian complexity of H – see the proof in Sec. 7 for details.
Combining these ideas, our plan of attack is the following: Given some class of depth-d networks, and an
arbitrary parameter r ∈ {1, . . . , d}, we use Thm. 3 to relate their Rademacher complexity to the complexity
of similar networks, but where one of the first r parameter matrices is of rank-1. We then use Thm. 4
to bound that complexity in turn using the Rademacher complexity of depth-r networks. Crucially, the
resulting bound has no explicit dependence on the original depth d, only on the new parameter r. Formally,
we have the following theorem, which is the main result of this section:
Theorem 5. Consider the following hypothesis class of networks on X = {x : kxk ≤ B}:
(
)
maxx∈X kNW d (x)k ≥ Γ
1
n
o
,
H = NW d :
kWj k kWj kp
1
∀j ∈ {1 . . . d}, Wj ∈ Wj , max M (j)
, Mp (j) ≤ 1
for some parameters p, Γ ≥ 1, {M (j), Mp (j), Wj }dj=1 . Also, for any r ∈ {1, . . . , d}, define
NW1r maps to R
∀j
∈
{1
.
.
.
r
−
1},
W
∈
W
j
j
r
H r = NW 1 :
.
o
n
kWj k kWj kp
, Mp (j) ≤ 1
∀j ∈ {1 . . . r}, max M (j)
Finally, for m > 1, let ℓ ◦ H = {(ℓ1 (h(x1 )), . . . , ℓm (h(xm ))) : h ∈ H}, where ℓ1 , . . . , ℓm are real-valued
loss functions which are γ1 -Lipschitz and satisfy ℓ1 (0) = ℓ2 (0) = . . . = ℓm (0).
Then the Rademacher complexity R̂m (ℓ ◦ H) is upper bounded by
1/p
Q
Qd
d
B
log3/2 (m) · R̂ (H )
log Γ j=1 Mp (j)
cB j=1 M (j)
1
m
r
Qr
+
+√ ,
min
B j=1 M (j)
γ
r
m
r∈{1,...,d}
where c > 0 is a universal constant.
In particular, one can upper bound this result by any choice of r in {1, . . . , d}. By tuning r appropriately,
we can get bounds independent of the depth d. In the next subsection, we provide some concrete applications
for specific choices of H.
Remark 3. The parameters Γ and γ, which divide the norm terms in Thm. 5, are both closely related to the
notion of a margin. Indeed, if we consider binary or multi-class classification, then bounds as above w.r.t.
1
γ -Lipschitz losses can be converted to a bound on the misclassification error rate in terms of the average
γ-margin error on the training data (see Bartlett et al. [2017, Section 3.1] for a discussion). Also, Γ can be
viewed as the “maximal” margin attainable over the input domain.
11
4.2 Applications of Thm. 5
In this section we exemplify how Thm. 5 can be used to obtain depth-independent bounds on the sample
complexity of various classes of neural networks. The general technique is as follows: First, we prove a
√
bound on R̂m (Hr ), which generally depends on the depth r, and scales as r α / m for some α > 0. Then,
we plug it into Thm. 5, and utilize the following lemma to tune r appropriately:
Lemma 3. For any α > 0, β ∈ (0, 1] and b, c, n ≥ 1, it holds that
)
(
α
cr α
dα
b dα
b α+β
min
min
,
.
+ β ,
≤ min 3 ·
β
r
n
n
r∈{1,...,d} n
(n/c) α+β
We begin with proving a depth-independent version of Thm. 1. That theorem implies that for the class
Hr of depth-r neural networks with Frobenius norm bounds MF (1), . . . , MF (r) (up to and including r =
d),
r
d
Y
r
(12)
R̂m (Hr ) ≤ O B
MF (j)
m
j=1
Plugging this into Thm. 5, and using Lemma 3, it is straightforward to derive the following corollary (see
Sec. 7 for a formal derivation):
Corollary 1. Let H be the class of depth-d neural networks, where each parameter matrix Wj satisfies
kWj kF ≤ MF (j), and with 1-Lipschitz, positive-homogeneous, element-wise activation functions. Assuming the loss function ℓ and H satisfy the conditions of Thm. 5 (with the sets Wj being unconstrained), it
holds that
r
Qd
3/4
B
r
Qd
log (m) log Γ j=1 MF (j)
B j=1 MF (j)
d
.
,
·
min
R̂m (ℓ ◦ H) ≤ O
γ
m
m1/4
Ignoring logarithmic factors and replacing the min by its first argument, the bound in the corollary is at
most
!
Q
B dj=1 MF (j)
Õ
.
γm1/4
The bound is completely independent of either the width or the depth of the networks. In other words, it is
possible to make this
Qd bound smaller than any fixed ǫ, with a sample size m independent of the network’s
size, as long as B
j=1 MF (j) is bounded. On the other hand, the bound in Corollary 1 is also bounded by
Γ
O
B
Qd
√ !
d
j=1 MF (j)
√
γ m
,
which is the bound one would get from an immediate application of Thm. 1, and implies that the asymptotic
rate (as a function of m) is still maintained.
Next, we apply Thm. 5 to the results in Bartlett et al. [2017], which as discussed in the introduction,
provide a depth-dependent bound using a different set of norms. Specifically, they obtain the following
intermediary result in deriving their generalization bound:
12
Theorem 6 (Bartlett et al. [2017]). Let H be the hypothesis class of depth-d, width-h real-valued networks
on X = {x : kxk ≤ B}, using 1-Lipschitz activation functions, given by
)
(
kWjT k2,1
≤ L(j) ,
H = NW d : ∀j ∈ {1, . . . , d}, kWj k ≤ M (j),
1
kWj k
for some fixed parameters {L(j), M (j)}dj=1 . Then the Rademacher complexity R̂m (H) is upper-bounded
by
3/2
Qd
d
B log(h) log(m) j=1 M (j) X
√
L(j)2/3 .
R̂m (H) ≤ O
·
m
j=1
p As discussed in the introduction, L(j) can never be smaller than 1, hence the bound scales at least as
d3 /m. However, using the bound above together with Thm. 5 and Lemma 3, we can get the following
corollary (where for simplicity, we assume that L(j) for all j are uniformly bounded by some L):
Corollary 2. Let H be the class of depth-d, width-h networks with 1-Lipschitz, positive-homogeneous,
element-wise
activation functions.
Assuming the loss function ℓ and H satisfy the conditions of Thm. 5 (with
kWjT k2,1
Wj = Wj : kWj k ≤ L for all j), it holds that the Rademacher complexity R̂m (ℓ ◦ H) is at most
BL log(h) log(m)
O
γ
21
Q
13
d
1+ 2 p
3/2
B
3 +p
M
(j)
log
log
(m)
3/2
p
j=1
Γ
d
j=1 M (j)
· min
,√
.
1
m
m 2+3p
Qd
As before, by replacing the min by its first argument, we get a bound which is fully independent of the
network size, assuming the norms are suitably bounded. To give a concrete example, if we take p = 2 (so
that the Mp (j) constraints correspond to the Frobenius norm), and ignore lower-order logarithmic factors,
we get a bound scaling as
v
Q
u
r
Qd
d
3/2 B
u
4 log
3
BL j=1 M (j)
t
j=1 Mp (j)
Γ
d
.
· min
R̂m (ℓ ◦ H) ≤ Õ
,
γ
m
m1/2
In contrast, a direct application of Thm. 6 in the same setting leads to a bound of
r !
Q
BL dj=1 M (j)
d3
R̂m (ℓ ◦ H) ≤ Õ
.
·
γ
m
Finally, we note that since the Eq. (3), based on a PAC-Bayes analysis, is always weaker than Eq. (2) (as
noted in Bartlett et al. [2017]), Corollary 2 also gives a size-independent version of Eq. (3).
5 A Lower Bound for Schatten Norms
In this section, we present a lower bound on the Rademacher complexity, for the class of neural networks
with parameter matrices of bounded Schatten norms. The formal result is the following:
13
Theorem 7. Let H be the class of depth-d, width-h neural networks, where each parameter matrix Wj
satisfies kWj kp ≤ Mp (j) for some Schatten p-norm k · kp (and where we use the convention that p = ∞
refers to the spectral norm). Then there exists a choice of γ1 -Lipschitz loss ℓ and data points x1 , . . . , xm ∈ X ,
with respect to which
n
o
Qd
max 0, 21 − p1
B j=1 Mp (j) · h
√
R̂m (ℓ ◦ H) ≥ Ω
.
γ m
This theorem strengthens Theorem 3.6 in Bartlett et al. [2017], which only considered the case p = ∞,
and did not have a dependence on h (on the other hand, they consider bounds which hold for any choice
of x1 , . . . , xm , while we consider bounds uniform over x1 , . . . , xm for simplicity). Furthermore, it has the
following implications:
• Like Bartlett et al. [2017], the theorem implies that by controlling just the norms of each parameter
matrix, a dependence on the product of the norms is generally inevitable.
• For p = ∞, we see that there is an inevitable h1/2 factor in the bound, which implies that controlling
the spectral norm is insufficient to get size-independent bounds (at least, independent of the width
h). More generally, any Schatten p-norm control with p > 2 will be insufficient to get such size
independence.
• For p = 2 (i.e. Frobenius
norm bounds MF (1), . . . , MF (d)), the lower bound becomes size-independent,
Q
B
d
MF (j)
j=1
√
. This matches our upper bound in Corollary 1 up to logarithmic facand on the order of
m
tors, except for the worse polynomial dependence on m.
6 Additional Remarks
6.1 Post-hoc Guarantees
So far we have proved upper bounds on the empirical Rademacher complexity of a fixed class of neural
networks, of the form
HL = {h : compl(h) ≤ L},
for some complexity measure compl and a parameter L. These imply high-probability learning guarantees
for algorithms which return predictors in HL . However, in the context of norm-based constraints, practical
algorithms for neural networks usually perform unconstrained optimization, and therefore are not guaranteed a-priori to return a predictor in some fixed HL . Fortunately, it is straightforward to convert these bounds
into probabilistic guarantees for any neural network h, with the bound scaling appropriately with the complexity compl(h) of the particular network. We note that such post-hoc guarantees have also been stated
in the context of some previous sample complexity bounds for neural networks (e.g., Bartlett et al. [2017],
Neyshabur et al. [2017]). It is achieved, for instance, by a union bound over, say, a doubling scale of the
complexity. We refer to the proof of the margin bound of Koltchinskii and Panchenko [2002, Theorem 2]
for an example of this technique.
14
6.2 Complexity of Lipschitz Networks
In proving the results of Sec. 4, the key element has been the observation that under appropriate norm
constraints, a neural network must have a layer with parameter matrix close to being rank-1, and therefore
the network can be viewed as a composition of a shallower network and a univariate Lipschitz function. In
fact, this can be generalized: Whenever we have a network with parameter matrix close to being rank-k,
we can view it as a composition of a shallow network and a Lipschitz function on Rk . Although we do not
develop this idea further in this paper, this observation might be useful in analyzing other types of neural
network classes.
Taking this to the extreme, we can also bound the complexity of neural networks computing Lipschitz
functions, by studying the complexity of all Lipschitz functions over the domain X . It is easily verified that
in our setting, if we consider the class H of depth-d
networks, where each parameter matrix has spectral
Q
norm at most M (j), then the network must be dj=1 M (j)-Lipschitz. Using well-known estimates of the
covering numbers of Lipschitz functions over X , we get that
!
Q
B dj=1 M (j)
· ,
R̂m (ℓ ◦ H) ≤ O
γm1/dim(X )
where the loss ℓ is assumed γ1 -Lipschitz, and where dim(X ) is the dimensionality of X . Of course, the
bound has a very bad dependence on the input dimension (or equivalently, the width of the first layer in the
network), but on the other hand, has no dependence on the network depth, nor on any matrix norm other
than the spectral norm. Again, as discussed in the previous subsection, it is also possible to use this bound
to get post-hoc guarantees, without constraining the Lipschitz parameter of the learned network in advance.
7 Proofs
7.1 Proofs of Lemma 2 and Thm. 2
We first prove Lemma 2. Letting wj denote the j-th row of a matrix W , we have
!
!
m
m
X
X
⊤
sup
g
ǫi σ(W f (xi ))
=
sup
max g
ǫi σ(wk f (xi ))
f ∈F ,W :kW k1,∞ ≤R
i=1
f ∈F ,W :kwj k1 ≤R
∞
=
sup
g
f ∈F ,kwk1 ≤R
k
m
X
i=1
ǫi σ(w⊤ f (xi ))
i=1
!
.
Since g(|z|) ≤ g(z) + g(−z), the left-hand side of Eq. (11) is upper bounded by
!
!
m
m
X
X
ǫi σ(w⊤ f (xi ))
ǫi σ(w⊤ f (xi )) + Eǫ
sup
g −
Eǫ
sup
g
f ∈F ,kwk1 ≤R
f ∈F ,kwk1 ≤R
i=1
i=1
and the proof is concluded exactly as in Lemma 1 by appealing to Ledoux and Talagrand [1991].
We now turn to Thm. 2, whose proof is rather similar to that of Thm. 1. Fixing λ > 0 to be chosen later,
15
the Rademacher complexity can be upper bounded as
mR̂m (Hd ) = Eǫ
m
X
sup
N
,Wd
W d−1
1
ǫi Wd σd−1 (NW d−1 (xi ))
1
i=1
m
X
1
ǫi Wd σd−1 (NW d−1 (xi ))
log Eǫ sup exp λ
1
λ
i=1
(
m
X
1
≤ log Eǫ sup exp MF (d) · λ
ǫi σd−1 (NW d−1 (xi ))
1
λ
≤
i=1
∞
)
Applying the same argument as in the proof of Thm. 1, and using Lemma 2, we can upper bound the above
by
!!
m
X
1
mR̂m (Hd ) ≤ log 2d · Eǫ exp M λ
,
(13)
ǫi xi
λ
i=1
∞
Qd
where M = j=1 M (j). Letting xi,j denote the j-th coordinate of xi , and using symmetry, the expectation
inside the log can be re-written as
!
!
m
n
m
X
X
X
ǫi xi,j
Eǫ exp M λ ·
≤
ǫi xi,j
Eǫ exp M λ · max
j
≤
n
X
Eǫ exp M λ
j=1
= 2
n
X
j=1 i=1
ǫi xi,j
m
X
ǫi xi,j
i=1
j=1
= 2
m
X
i=1
Eǫ exp M λ
m
n Y
X
i=1
j=1
i=1
"
!
!
+ exp −M λ
= 2
n Y
m
X
m
X
ǫi xi,j
i=1
!#
Eǫ exp (M λǫi xi,j )
j=1 i=1
n
m
j=1
i=1
X
X
exp (M λxi,j ) + exp (−M λxi,j )
x2i,j
exp M 2 λ2
≤ 2
2
!
,
≤ exp(z 2 /2). Further upper bounding this by
where in the
last step we used
the fact that exp(z)+exp(−z)
2
P
2n maxj exp M 2 λ2 di=1 x2i,j and plugging back to Eq. (13), we get
m
X
1
x2i,j
log 2d+1 n · max exp M 2 λ2
j
λ
i=1
Choosing λ =
r
d+1+log(n)
P
2 ,
M 2 maxj m
i=1 xi,j
!!
m
X
d + 1 + log(n)
=
x2i,j .
+ M 2 λ max
j
λ
i=1
we can upper bound the above by
v
u
m
X
u
t
x2i,j ,
2M (d + 1 + log(n)) max
j
from which the result follows.
16
i=1
7.2 Proof of Thm. 3
The proof will build on the following few technical lemmas.
Lemma 4. For any matrix W , and any Schatten p-norm k · kp such that p < ∞, there exists a rank-1 matrix
W̃ of the same size such that
kW̃ k ≤ kW k , kW̃ kp ≤ kW kp , kW − W̃ k ≤ kW kpp − kW kp
1/p
.
Proof. Let U SV ⊤ denote the SVD decomposition of W , where S = diag(s1 , s2 , . . . , sr ), and choose
W̃ = u1 s1 v1⊤ , where u1 , v1 , s1 are top singular vectors and values of W . The first two inequalities in the
lemma are easy to verify. As to the third inequality, using the unitarial invariance of the spectral norm, we
have
1/p
1/p
r
r
X
X
spj
spj − sp1 ,
kW −W̃ k = kU SV ⊤ −u1 s1 v1⊤ k = kU diag(0, s2 , . . . , sr )V ⊤ k = s2 ≤
=
j=2
1/p
which equals (kW kpp − kW kp )
j=1
.
Lemma 5. Given network parameters W1d = {W1 , . . . , Wd }, let W̃1d = {W1 , . . . , Wr−1 , W̃r , Wr+1 , . . . , Wd }
be the same parameters, where the parameter matrix Wr of the r-th layer (for some fixed r ∈ {1, . . . , d}) is
changed to some other matrix W̃r . Then
d
Y
kWr − W̃r k
.
kWj k
sup kNW d (x) − NW̃ d (x)k ≤ B
1
1
kWr k
x∈X
j=1
Proof. By a simple calculation, we have that the Lipschitz constant of the function NWbr is at most
Assume for now that 2 ≤ r ≤ d − 1. By definition, we have
Qr
j=b kWj k.
NW d (x) = NW d (σr (Wr σr−1 (NW r−1 (x))))
r+1
1
1
and
NW̃ d (x) = NW d (σr (W̃r σr−1 (NW r−1 (x)))) .
r+1
1
1
Qd
r (x) is at most
The Lipschitz constant of the function NW d is at most j=r+1 kWj k, and the norm of NW
1
r+1
17
kxk
Qr
j=1 kWj k.
Therefore, for any x ∈ X ,
kNW d (x) − NW̃ d (x)k = kNW d (σr (Wr σr−1 (NW r−1 (x)))) − NW d (σr (W̃r σr−1 (NW r−1 (x)))) k
1
r+1
1
1
1
r+1
d
Y
kWj k · kσr (Wr σr−1 (NW r−1 (x))) − σr (W̃r σr−1 (NW r−1 (x)))k
≤
1
1
j=r+1
≤
d
Y
j=r+1
≤
d
Y
j=r+1
≤
d
Y
j=r+1
kWj k · kWr σr−1 (NW r−1 (x)) − W̃r σr−1 (NW r−1 (x)))k
1
1
kWj k · kWr − W̃r k · kσr−1 (NW r−1 (x)k
1
kWj k · kWr − W̃r k ·
r−1
Y
j=1
kWj k · kxk,
from which the result follows after a simplification. The cases r = 1 and r = d are handled in exactly the
same manner.
Q
Q
Lemma 6. Suppose that NW d is such that dj=1 kWj k ≥ Γ and dj=1 kWj kp ≤ M . Then for any r ∈
1
{1, . . . , d},
1/r
kWj kp
M
≤
.
min
Γ
j∈{1,...,r} kWj k
Proof. Fixing some r, and using the stated assumptions as well as the fact that kW kp ≥ kW k for any p, we
have
Qd
d
r
Y
Y
kWj kp
kWj kp
kWj kp r
M
j=1 kWj kp
=
≥ Qd
≥
≥
min
.
Γ
kWj k
kWj k
j∈{1,...,r} kWj k
j=1 kWj k
j=1
j=1
Taking the r-th root from both sides, the result follows.
With these lemmas in hand, we can now turn to prove Thm. 3. Combining Lemma 4 and Lemma 5, we
have that there indeed exists a network NW̃ d with a rank-1 matrix in layer r, such that
1
d
1/p
Y
(kWr kpp − kWr kp )
sup kNW d (x) − NW̃ d (x)k ≤ B
kWj k ·
1
1
kWr k
x∈X
j=1
1/p
d
Y
kWr kpp
kWj k
−1
.
(14)
= B
kWr kp
j=1
By the definition of NW d , we have that kNW d (x)k ≤ kxk
1
1
Qd
j=1 kWj k. Since
Q
that dj=1 kWj k ≥
we assume there exists x
(of norm at most B) such that kNW d (x)k ≥ Γ, it follows
Γ/B. Using this and the
1
Qd
assumption that j=1 kWj kp ≤ M , and plugging into Lemma 6, it follows that for any r ∈ {1, . . . , d},
kWj kp
≤
min
j∈{1,...,r} kWj k
18
MB
Γ
1/r
.
Substituting into Eq. (14), we get that
sup kNW d (x) − NW̃ d (x)k ≤ B
x∈X
1
1
= B
d
Y
j=1
d
Y
j=1
kWj k
MB
Γ
kWj k exp
p/r
!1/p
−1
p log(M B/Γ)
r
1/p
.
−1
Suppose for now that r is such that r ≥ p log(M B/Γ). Using the fact that exp(z) ≤ 1 + 2z for any
z ∈ [0, 1], it follows that the above is at most
1/p
d
Y
2p
log(M
B/Γ)
.
kWj k
B
r
j=1
It remains to consider the case where r < p log(M B/Γ). However, in this regime the theorem trivially
holds: Let W̃r in the network NW̃ r be the all-zeros matrix (which is rank zero and ensures that NW̃ r (x) = 0
1
1
for all x), and we have by definition that
d
d
Y
Y
2p log(M B/Γ) 1/p
kWj k
kWj k < B
sup kNW d (x)−NW̃ d (x)k = sup kNW d (x)k ≤ B
.
1
1
1
r
x∈X
x:kxk≤B
j=1
j=1
7.3 Proof of Thm. 4
To prove the theorem, we use a straightforward covering number argument, beginning with a few definitions.
Given any function class F, a metric d on the elements of F, and ǫ > 0, we let the covering number N (F, d, ǫ) denote the minimal number n of functions f1 , f2 , . . . , fn in F, such that for all f ∈ F,
mini=1,...,n d(fi , f ) ≤ ǫ. In particular, fix some set of data points x1 , . . . , xm , and define the empirical L2
distance
v
u
m
u1 X
′
ˆ
(f (xi ) − f ′ (xi ))2 .
dm (f, f ) = t
m
i=1
Also, given a function class F and x1 , . . . , xm , we let
"
Ĝm (F) := Eη
#
m
1 X
ηi f (xi )
sup
f ∈F m
i=1
denote the (empirical) Gaussian complexity of F, where η1 , . . . , ηn arep
i.i.d. standard Gaussian random variables. It is well known that R̂m (H) and Ĝm (H) are equivalent up to a c log(m) factor [Ledoux and Talagrand,
1991, pg. 97]. By Sudakov’s minoration theorem (see theorem 3.18 in Ledoux and Talagrand [1991]), we
have that for all α > 0
!2
√
m
·
Ĝ
(H)
m
(15)
log(N (H, dˆm , α)) ≤ c
α
for a universal constant c > 0.
19
With these definitions in hand, we turn to prove the theorem. We first note that FL ◦ H is equivalent to
the class {Lg(·) + a : g ∈ F1,0 ◦ H}, and therefore
"
#
#
"
m
m
1 X
1 X
sup
ǫi g(xi ) = Eǫ
ǫi (Lg(xi ) + a)
sup
R̂m (FL,a ◦ H) = Eǫ
g∈F1,0 ◦H m i=1
g∈FL,a ◦H m i=1
"
#
!
m
m
1 X
a X
= Eǫ L · sup
ǫi g(xi ) +
ǫi
m
g∈F1,0 ◦H m i=1
i=1
"
#
m
1 X
sup
= L · Eǫ
ǫi g(xi ) = L · R̂m (F1,0 ◦ H).
(16)
g∈F1,0 ◦H m
i=1
Therefore, it is enough to consider R̂m (F1,0 ◦ H). To simplify notation in what follows, we will drop the
1, 0 subscript from F.
We first argue that
R
log(N (F, dˆm , ǫ)) ≤ c′ 1 +
,
(17)
ǫ
again for some numerical constant c′ > 0. To prove this, we note that for any functions f, f ′ , it holds that
dˆm (f, f ′ ) ≥ dˆ∞ (f, f ′ ) := supx |f (x) − f ′ (x′ )|, so it is enough to upper bound N (H, dˆ∞ , ǫ). To do so, we
first notice that the range of any f ∈ F is in [−R, R]. Discretize [−R, R] × [−R, R] into a two-dimensional
grid Ux × Uy , where
Ux := {−R, −R + ǫ, −R + 2ǫ, . . . , −R + ⌊2R/ǫ⌋ ǫ, R} , Uy := {0, ±ǫ, ±2ǫ, . . . , ±⌊R/ǫ⌋ǫ, ±R}.
Given any f ∈ F, construct the piecewise-linear function f ′ as follows: For any input x ∈ Ux , let f ′ (x)
be the point in Uy nearest to f (x) (breaking ties arbitrarily), and let the rest of f ′ be constructed as a linear
interpolation of these points on Ux . It is easily verified that supx∈[−R,R] |f (x) − f ′ (x)| ≤ ǫ. Moreover,
note that for two neighboring points x, x′ in Ux , the points f ′ (x), f ′ (x′ ) on Uy must be neighboring or
the same. Therefore, each such function f ′ can be parameterized by a vector of the form {−, 0, +}|Ux |−1 ,
which specifies whether (starting from the origin) f ′ goes up, down, or remains the same on each of its
linear segments. The number of such functions is at most 3|Ux |−1 ≤ 32R/ǫ+1 , and therefore N (F, dˆ∞ , ǫ) ≤
32R/ǫ+1 . Recalling that dˆm (f, f ′ ) majorizes dˆ∞ (f, f ′ ), we get Eq. (17).
Next, we argue that
ǫ
ǫ
N (F ◦ H, dˆm , ǫ) ≤ N F, dˆm ,
· N H, dˆm ,
.
(18)
2
2
To see this, pick any f ∈ F and h ∈ H, and let f ′ , h′ be the respective closest functions in the cover of F
and H (at scale ǫ/2). By the triangle inequality and the easily verified fact that f ′ is 1-Lipschitz, we have
dˆm (f h, f ′ h′ ) ≤ dˆm (f h, f ′ h) + dˆm (f ′ h, f ′ h′ ) ≤ dˆ∞ (f h, f ′ h) + dˆm (f ′ h, f ′ h′ )
v
u
m
u1 X
′
(f ′ h(xi ) − f ′ h′ (xi ))2
≤ sup |f (x) − f (x)| + t
m
x∈[−R,R]
i=1
v
u
m
ǫ u
1 X
≤ +t
(h(xi ) − h′ (xi ))2
2
m
i=1
ǫ
ǫ
ǫ
= + dˆm (h, h′ ) ≤ + = ǫ .
2
2 2
20
Therefore, we can cover F ◦ H (at scale ǫ) by taking f ′ (h′ (·)) for all possible choices of f ′ , h′ from the
covers of F, H (at scale ǫ/2), leading to Eq. (18).
Combining Eq. (15),Eq. (17) and Eq. (18), we get that
!2
√
m
·
Ĝ
(H)
R
m
,
log(N (F ◦ H, dˆm , ǫ)) ≤ c′′ 1 + +
ǫ
ǫ
for some numerical constant c′′ > 0. Finally, we use Dudley’s entropy integral, which together with the
equation above, implies the following for some numerical constant c > 0 (possibly changing from row to
row):
(
)
Z supg∈F ◦H dˆm (g,0) q
1
log(N (F ◦ H, dˆm , ǫ))dǫ
R̂m (F ◦ H) ≤ c inf α + √
α≥0
m α
Z Rq
1
≤ c inf α + √
log(N (F ◦ H, dˆm , ǫ))dǫ
α≥0
m α
v
!2
√
Z Ru
u
1
m · Ĝm (H)
t1 + R +
≤ c inf α + √
dǫ
α≥0
ǫ
ǫ
m α
(
!
)
r
√
Z R
1
R
m · Ĝm (H)
≤ c inf α + √
1+
+
dǫ
α≥0
m α
ǫ
ǫ
)
(
r !
Z R
Z R
1
R
1
dǫ + Ĝm (H)
1+
dǫ
≤ c inf α + √
α≥0
m 0
ǫ
α ǫ
R
R
≤ c inf α + √ + Ĝm (H) log
.
α≥0
α
m
√
Choosing in particular α = R/ m, we get the upper bound
R
R̂m (F ◦ H) ≤ c √ + log(m) · Ĝm (H)
m
p
for some c > 0. Plugging this into Eq. (16), and upper bounding Ĝm (H) by c log(m)R̂m (H) (see
[Ledoux and Talagrand, 1991]), the result follows.
7.4 Proof of Thm. 5
It is enough to prove the bound for any fixed r ∈ {1, . . . , d}, and then take the infimum over any such r.
Given H and r, construct a new hypothesis class H̃ by replacing each network h ∈ H by the network h̃
as defined in Thm. 3 (namely, where the parameter matrix in the r-th layer is replaced by a rank-1 matrix).
We will use the notation h̃h to clarify the dependence of h̃ on h. According to that theorem, as well as the
21
definition of Rademacher complexity, we have
#
"
m
1 X
ǫi ℓ(h(xi ))
R̂m (ℓ ◦ H) = Eǫ sup
h∈H m i=1
"
!#
m
m
1 X
1 X
= Eǫ sup
ǫi ℓ(h̃h (xi )) +
ǫi ℓ(h(xi )) − ℓ(h̃h (xi ))
m
h∈H m i=1
i=1
"
!#
m
X
1
≤ Eǫ sup
ǫi ℓ(h̃h (xi )) + sup |ℓ(h(x)) − ℓ(h̃h (x))|
x∈X
h∈H m i=1
"
!#
m
1 X
1
≤ Eǫ sup
ǫi ℓ(h̃(xi )) + · sup |h(x) − h̃h (x)|
γ x∈X
h̃∈H̃ m i=1
1/p
Q
Q
d
d
B
M
(j)
2p
log
B
M
(j)
p
j=1
j=1
Γ
.
≤ R̂m (ℓ ◦ H̃) +
γ
r
(19)
We now reach the crucial observation which lies at the heart of the proof. Consider any network NW̃ d in
1
H̃, and let suv⊤ be the SVD decomposition of its (rank-1) parameter matrix W̃r in layer r (where s, u, v are
also leading singular value and vectors of Wr , by construction). By definition, we have that the composition
of NW̃ d with any loss ℓj equals
1
x 7→ ℓj (Wd σd−1 (Wd−1 σd−2 (. . . σr (suv⊤ σr−1 (. . . σ1 (W1 x)))).
This function is equivalent to the composition of the function
x 7→ sv⊤ σr−1 (. . . σ1 (W1 x))
with the univariate function
x 7→ ℓj (Wd σd−1 (Wd−1 σd−2 (. . . σr (ux)))) .
Note that since ksv⊤ k = s = kWr k and ksv⊤ kp = s ≤ kWr kp , the former function is contained in
Q
Hr as defined in the theorem; whereas the latter function has Lipschitz constant at most γ1 dj=r+1 kWj k ≤
1 Qd
j=r+1 M (j), and maps the input 0 to the same fixed output (which we’ll denote as a) for all j. Therefore,
γ
we obtain
Qr that ℓ ◦ H̃ is contained in the composition of Hr (with
Qdfunctions whose output is bounded in
Q
±B j=1 M (j)), and the class F 1 d
j=r+1 M (j)-Lipschitz functions. As a
M (j),a consisting of
γ
j=r+1
result, we can apply Thm. 4 and obtain that
Q
R̂m (ℓ ◦ Ĥ) ≤ R̂m Hr ◦ F 1 d
j=r+1 M (j),a
γ
d
r
Y
B
c Y
M (j) √
M (j) + log3/2 (m) · R̂m (Hr )
≤
γ
m
j=r+1
j=1
!
d
3/2
c Y
B
log
(m)
·
R̂
(H
)
m
r
Qr
.
=
M (j) √ +
γ
m
j=1 M (j)
j=1
22
Plugging this back into Eq. (19) and simplifying a bit (also noting that p1/p can be upper bounded by a
universal constant), we get that R̂m (ℓ ◦ H) is upper bounded by
Q
1/p
Q
d
B
d
log Γ j=1 Mp (j)
c′ B j=1 M (j) log3/2 (m) · R̂m (Hr )
+ √1
Qr
+
γ
B j=1 M (j)
r
m
for an appropriate constant c′ . As mentioned at the beginning of the proof, this upper bound holds for any
fixed r ∈ {1, . . . , d}, from which the result follows.
7.5 Proof of Lemma 3
We will show that for any α, β, b, c, n as stated in the lemma, there always exists a choice of r ∈ {1, . . . , d}
such that
α
α
cr
b dα
b α+β
.
min
+ β ,
≤ 3·
β
n
r
n
(n/c) α+β
α
Since the left hand side is also trivially at most dn , the result follows. We prove this inequality by a case
analysis:
k
j
1
1
• If (bn/c) α+β ∈ [1, d], pick r = (bn/c) α+β ∈ {1, 2, . . . , d}, in which case
α
1
α
α
α
α+β
α
c
(bn/c)
cr
(b) α+β
b α+β
b α+β
b
b
β
+2
≤ 3
.
+ β ≤
+
β =
β
β
β
1
n
r
n
1
α+β
α+β
α+β
α+β
(n/c)
(n/c)
(n/c)
(bn/c)
2
1
• If (bn/c) α+β > d, it follows that
α
α
α
cr
b dα
(bn/c) α+β
b α+β
dα
.
min
+ β ,
< c
=
≤
β
n
r
n
n
n
(n/c) α+β
7.6 Proof of Corollary 1
A direct application of Thm. 1, as well as the fact that the loss ℓ is 1/γ Lipschitz, implies that
r !
Q
B dj=1 MF (j) d
.
R̂m (ℓ ◦ H) ≤ O
γ
m
(20)
On the other hand, plugging Eq. (12) into Thm. 1, (using p = 2, letting each Wj be the space of all matrices,
kWj k
kW k
and choosing M (j) = MF (j) for all j, noting that M (j)
≤ MFj(j)F ≤ 1), we get that
v
Q
u
Qd
d
u
B
√
log3/2 (m) r t log Γ j=1 MF (j)
B j=1 MF (j)
.
min
√
+
R̂m (ℓ ◦ H) ≤ O
r∈{1,...,d}
γ
m
r
(21)
Upper bounding R̂m (ℓ ◦ H) by the minimum of Eq. (20) and Eq. (21), and using Lemma 3 with α = β = 21 ,
r
Qd
√
M
(j)
, c = log3/2 (m), and n = m, the result follows.
b = log B
F
j=1
Γ
23
7.7 Proof of Corollary 2
Consider the class
Hr =
NW1r maps to R
NW1r :
kW T k2,1
j
≤L
∀j ∈ {1 . . . r − 1}, kW
.
jk
n
o
kWj k kWj kp
∀j ∈ {1 . . . r}, max M (j) , Mp (j) ≤ 1
Since NW1r in the definition of Hr maps to R, Wr is a vector, meaning that kWrT k2,1 = kWr k2 = kWr k.
Therefore, we can use Thm. 6 to bound R̂m (Hr ); in particular, for r ≥ 1,
!
Q
B log(h) log(m)L rj=1 M (j)r 3/2
√
· .
R̂m (Hr ) ≤ O
m
It therefore follows from Thm. 5 that R̂m (ℓ ◦ H) is at most
O
BL log(h) log(m)
γ
where c = log3/2 (m), n =
Qd
j=1 M (j)
·
min
r∈{1,...,d}
(
b
cr 3/2
+ 1/p
n
r
)!
,
Q
1/p
√
d
m, and b = log B
M
(j)
.
p
j=1
Γ
On the other hand, a direct application of Thm. 6 also implies that R̂m (ℓ ◦ H) is at most
!
Q
B dj=1 M (j) log(h) log(m)Ld3/2
√
.
·
O
γ
m
Combining the above two bounds, and applying Lemma 3, we get that R̂m (ℓ ◦ H) is at most
BL log(h) log(m)
O
γ
7.8 Proof of Thm. 7
21
Q
13
d
1+ 2 p
3/2
B
3 +p
M
(j)
log
log
(m)
3/2
p
j=1
Γ
d
j=1 M (j)
· min
,√
.
1
m
m 2+3p
Qd
By definition of the Rademacher complexity, it is enough to lower bound the complexity of ℓ ◦ H′ where H′
is some subset of H. In particular, consider the class H′ of neural networks over Rh of the form
x 7→ Mp (d) · Mp (d − 1) · · · Mp (2) · W x,
where W = diag(kwk) is a diagonal matrix satisfying kwkp ≤ Mp (1) (here, k · kp refers to the vector
p-norm). Furthermore, suppose that ℓ(z) = γ1 max{z1 , . . . , zh }. Finally, we will choose x1 , . . . , xm in Rh
as xi = Be(i mod h) for all i, where et is the t-th standard basis vector.. Letting Ak = {i ∈ {1, . . . , m} :
24
i mod h = k}, it holds that
d
m
Y
X
1
R̂m (ℓ ◦ H′ ) = Eǫ
sup
Mp (j)W xi
ǫi ℓ
m
W :kW kp ≤Mp (1)
j=2
i=1
=
=
B
B
Qd
j=2 Mp (j)
Qd
γm
j=1 Mp (j)
γm
In particular, by choosing wk = h−1/p · sign
B
Qd
j=1 Mp (j)
γm · h1/p
· Eǫ
· Eǫ
P
sup
w:kwkp ≤Mp (1) k=1
h
X
sup
w:kwkp ≤1 k=1
max{0, wk } ·
max{0, wk } ·
X
X
ǫi
i∈Ak
ǫi .
i∈Ak
for all k, we can lower bound the above by
ǫ
i
i∈Ak
h
X
X
· Eǫ
ǫi = Ω
max 0,
k=1
h
X
i∈Ak
B
Qd
j=1 Mp (j)
γm · h1/p
·
h p
X
k=1
!
|Ak |
and since |Ak | ≥ ⌊m/h⌋ by its definition, we get a lower bound of
Q
1
Q
r !
− p1
d
2
B
M
(j)
·
h
B dj=1 Mp (j)
p
m
j=1
.
√
·h
= Ω
Ω
1/p
h
γ m
γm · h
,
(22)
An alternative bound (which is better when p < 2) can be obtained by considering the class H′ of real-valued
neural networks over R, of the form
x 7→ Mp (d) · Mp (d − 1) · · · Mp (2) · wx,
where |w| ≤ Mp (1). Furthermore, supposing that ℓ is the identity function times γ1 , and xi = B for all i, it
holds that
m
d
1 X Y
Mp (j)Bw
ǫi
w:|w|≤Mp(1) γm i=1
j=2
!
Q
Q
m
X
B dj=1 Mp (j)
B dj=1 Mp (j)
√
ǫi = Ω
· Eǫ
=
.
γm
m
R̂m (ℓ ◦ H′ ) = Eǫ
sup
i=1
Taking the best of this lower bound and the lower bound in Eq. (22), the result follows.
References
Martin Anthony and Peter L Bartlett. Neural network learning: Theoretical foundations. cambridge university press, 2009.
Peter Bartlett, Dylan J Foster, and Matus Telgarsky. Spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1706.08498, 2017.
25
Peter L Bartlett and Shahar Mendelson. Rademacher and gaussian complexities: Risk bounds and structural
results. Journal of Machine Learning Research, 3(Nov):463–482, 2002.
Stéphane Boucheron, Gábor Lugosi, and Pascal Massart. Concentration inequalities: A nonasymptotic
theory of independence. Oxford university press, 2013.
Vladimir Koltchinskii and Dmitry Panchenko. Empirical margin distributions and bounding the generalization error of combined classifiers. Annals of Statistics, pages 1–50, 2002.
Michel Ledoux and Michel Talagrand. Probability in Banach Spaces. Springer, 1991.
Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. In search of the real inductive bias: On the role of
implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014.
Behnam Neyshabur, Ryota Tomioka, and Nathan Srebro. Norm-based capacity control in neural networks.
In Conference on Learning Theory, pages 1376–1401, 2015.
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. A pac-bayesian approach
to spectrally-normalized margin bounds for neural networks. arXiv preprint arXiv:1707.09564, 2017.
Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms.
Cambridge university press, 2014.
Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep
learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016.
26
| 9cs.NE
|
RoboCupSimData: A RoboCup soccer research
dataset
Olivia Michael1 , Oliver Obst∗1 ,
Falk Schmidsberger2 , Frieder Stolzenburg2
1
arXiv:1711.01703v1 [cs.AI] 6 Nov 2017
2
: Western Sydney University, Centre for Research in Mathematics
: Harz University of Applied Sciences, Automation and Computer Sciences Department
∗
: Corresponding author: [email protected]
Abstract
RoboCup is an international scientific robot competition in which teams of multiple robots
compete against each other. Its different leagues provide many sources of robotics data, that can
be used for further analysis and application of machine learning. This paper describes a large
dataset from games of some of the top teams (from 2016 and 2017) in RoboCup Soccer Simulation
League (2D), where teams of 11 robots (agents) compete against each other. Overall, we used
10 different teams to play each other, resulting in 45 unique pairings. For each pairing, we ran
25 matches (of 10 mins), leading to 1125 matches or more than 180 hours of game play. The
generated CSV files are 17 GB of data (zipped), or 229 GB (unzipped). The dataset is unique in
the sense that it contains both the ground truth data (global, complete, noise-free information of
all objects on the field), as well as the noisy, local and incomplete percepts of each robot. These
data are made available as CSV files, as well as in the original soccer simulator formats.
lator as an input, and, asynchronously to sensor
input, also decides on a next action that is sent
back to the simulator, several times a second.
This complexity of the environment, with continuous state and action spaces, together with
the opportunity to compete against each other,
makes RoboCup soccer an interesting testbed for
multi-robot learning among many other applications.
1 Introduction
RoboCup is an international scientific robot
competition in which teams of multiple robots
compete against each other. The RoboCup soccer leagues provide platforms for a number of
challenges in robotics research, including locomotion, vision, real-time decision making, dealing with partial information, multi-robot coordination and teamwork. In RoboCup, several different leagues exist to emphasize specific research problems by using different kinds
of robots and rules. There are different soccer
leagues in the RoboCup with different types and
sizes of hardware and software: small size, middle size, standard platform league, humanoid, 2D
and 3D simulation (Kitano et al., 1997). In the
soccer simulation leagues (Akiyama et al., 2015),
the emphasis is on multi-robot team work with
partial and noisy information, in real-time. Each
of the robots is controlled by a separate program
that receives sensor information from the simu-
To assist automated learning of team behavior, we provide a large dataset generated using 10
of the top participants in RoboCup 2016 or 2017.
While it is possible to use the simulator for robot
learning, we also generate additional data that is
not normally available from playing other teams
directly: We modified the simulator to record
data from each robots local perspective, that is,
with the restricted views that depend on each
robots situation and actions, and also include
the sensor noise. In addition, for every step of
the game (100 ms), we recorded ground truth information (such as positions and velocities) of
1
receives additional status information, including
energy levels, referee decisions, and the state of
the game, every 100 ms. Each robot can issue
parameterized actions every 100 ms, to control
its locomotion, the direction of its view, and its
field of view. A more detailed description of
the information transmitted can be found in the
simulator manual (Chen et al., 2003).
all objects on the soccer field, as well as basic
actions of each robot. This ground truth information is usually recorded in a logfile, but not
available to teams during a match.
To create our dataset, we ran all pairings of the
selected teams with 25 repetitions of each game,
that is, 1125 games in total. With 11 robots
in each team, a single game dataset consists
of 22 local views plus a global (ground-truth)
view. These views are made available as CSV
files (comma-separated values). We also provide
the original logfiles that include additional sensors, actions of each robot recorded as text files,
and the software we created to patch the simulator and to convert recordings into CSV.
The provided data are useful for various different tasks including imitation learning (e.g.,
Ben Amor et al., 2013), learning or testing of
self-localization (e.g., Olson, 2000), predictive
modeling of behavior, transfer learning and reinforcement learning (e.g., Taylor and Stone,
2009), and representation learning for time series
data (Michael et al., 2018). The next sections describe the environment, robots, and data in more
detail.
3 Overview on the provided data
In robotics, data collections often comprise lidar data recorded by laser scans. This is very
useful in many applications, e.g., field robotics,
simultaneous localization and mapping (SLAM)
(Tong et al., 2013), or specific purposes such as
agriculture (Chebrolu et al., 2017). However, in
many contexts, there is not only one but several
robots which may be observed. The data from
RoboCup that we consider here includes information about other robots in the environment
and hence about the whole multi-robot system.
Data from multi-agent systems like the
RoboCup or the multi-player real-time strategy
video game StarCraft (Lin et al., 2017) provide
information on (simulated) environments as in
robotics. However, in addition, they contain
data on other agents and thus lay the basis for
machine learning research to analyze and predict agent behavior and strategies, which is important in applications such as service robotics
and multi-robot systems in general. To provide
a diverse dataset, we include several teams from
the last two RoboCup competitions, allowing for
different behaviors and strategies.
Perception and behavior of each robot during a game depends on the behavior of all other
robots on the field. Game logfiles, that is, files
containing ground truth information obtained
from recording games, can be produced from the
simulator, and are recorded in a binary format.
Access to individual player percepts, however,
is only possible from within the player code. To
learn from behavior of other teams, it is useful to
use the exact information that individual players receive, rather than the global (and noisefree) information in recorded logfiles. We therefore modified the simulator to additionally also
record all local and noisy information as received
by the robots on the field, in individual files for
2 Description of the environment
The RoboCup Soccer Simulation Server
rcssserver (Noda et al., 1998) is the software
used for the annual RoboCup competitions
that have been held since 1997. It is hosted
at github.com/rcsoccersim/. We used the
rcssserver version 15.3.0 to create the data for
this paper. The simulator implements the (2D)
physics and rules of the game, and also handles
the interface to the programs controlling each
player. By default, players use a 90 degree field
of view, and receive visual information every
150 ms. Throughout the game, this frequency
can actively be changed by each player individually to 75 ms, 150 ms, or 300 ms, by changing
the field of view to 45 degrees, 90 degrees, or
180 degrees, respectively. Visual information
is transmitted in the form of (Lisp-like) lists
of identified objects, with the level of detail
of information depending on object distances.
Potential objects include all other players on
the field, the ball, and landmarks like goal
posts, flags, and side lines. Each player also
2
2016
CSU Yunlu
Gliders
HELIOS
Ri-one
2017
CYRUS
Fra-UNIted
HELIOS
HfutEngine
MT
Oxsy
tion of the robot players, the pitch is filled with
several landmark objects as depicted in Fig. 2:
flags (f), which are punctual objects, lines (l),
the goal (g), and the penalty area (p). The origin
of the coordinate system is the center point (c).
The pitch is divided horizontally (x-direction)
into a left (l) and right (r) half and vertically (ydirection) into a top (t) and a bottom (b) half.
Additional numbers (10, 20, 30, 40, or 50) indiFigure 1: Teams from the RoboCup 2D soc- cate the distance to the origin in meters. Since
cer simulation world championships every soccer game takes place on the same pitch,
2016 in Leipzig (Germany) and 2017 there is only one file with infomation about the
in Nagoya (Japan) selected for the landmarks for all games that lists all these coordataset.
dinates, given as a table in CSV format, with
name landmarks.csv. For example, the row
each player. This information is stored in the f r t,52.5,34 says that the right top flag of
same format as it is sent to players. We also the pitch has the (x, y)-coordinates (52.5, 34).
Further table files provide information about
provide code to translate these individual logs
the
respective game. The names of all these
into CSV files that contain relative positions and
files – all naming conventions are summarized
velocities.
in Fig. 3 – contain the names of the competing
We chose the top-ten teams from the RoboCup
teams, the final scores for each team, possibly
2D soccer simulation world championships 2016
extended by the result of a penalty shootout, a
in Leipzig (Germany) and 2017 in Nagoya
time stamp (when the game was recorded), and
(Japan) (see Figure 1). Team binaries includsome further identifier.
ing further descriptions can be downloaded from:
The central SoccerServer (Chen et al., 2003)
archive.robocup.info/Soccer/Simulation/2D.
controls every virtual game with built-in physics
We played each team against each other team
rules. When a game starts, the server may be
for 25 times, resulting in about 17 GB of data
configured by several parameters which are col(zipped CSV files). For each game, we record the
lected in one file with the identifier parameters.
original logfiles including message logs. We also
For example, the row ball_decay,0.94 denotes
generate files with ground truth data as well as
that the ball speed decreases by the specified faclocal player data in human-readable format. Fitor (cf. Stolzenburg et al., 2002). However, from
nally, we made our generating scripts available,
a robotics point of view, most of the information
so that they can be used to reproduce our results
in this file is not very relevant, like the stamina of
or to produce additional datasets using other
the robots, the noise model, or the format of the
robotic soccer teams. There is also a smaller
coach instructions. We thus skip further details
subset of 10 games where the top-five teams play
here.
against each other once (163 MB CSV files plus
A soccer simulation game in the RoboCup 2D
217 MB original logfiles). Our data is available at
simulation league lasts 10 mins in total, divided
bitbucket.org/oliverobst/robocupsimdata/.
into 6000 cycles where the length of each cycle
is 100 ms. Logfiles comprise information about
the game, in particular about the current posi4 Description of the ground truth data
tions of all players and the ball including their
According to rules of the world soccer association velocity and orientation for each cycle. This inFIFA, a soccer pitch has the size of 105 m×68 m. formation is collected for the whole game in a
This is adopted for the RoboCup soccer simula- table with the identifier groundtruth. For each
tion league. Nevertheless, the physical boundary time point, the play mode (e.g. kickoff), the curof the area that may be sensed by the robots has rent ball position coordinates and its velocity is
an overall size of 120 m × 80 m. For the localiza- listed. Furthermore, the positions and velocities
3
Figure 2: Flags and lines in the robotic soccer simulation (cf. Chen et al., 2003).
1. landmarks: landmarks.csv (static information for all games, 1 file)
2. game data: <time>-<team left>_<score left>-vs-<team right>_<score right>-<id>.csv
where <id> =
• parameters (server configuration parameters, 1 file)
• groundtruth (logfile information for each game, 1 file)
• <team name>_<player number>-<suffix>
where <suffix> =
– landmarks (relative distances and angles to landmarks, 22 files)
– moving (relative distances and angles to ball and other players, 22 files)
Figure 3: Name conventions for the data files.
4
As described above, the frequency of these messages is dependent on the width of view that the
player selects, one message every 150 ms by default. Every 100 ms, players also receive status
information, so-called body messages. These individual messages are not recorded in the simulator logfiles. Developers of teams can implement
recording of messages that their own agents receive, but in order to record messages sent to any
team, the simulator software had to be modified
instead. Our code contains patches to the simulator that allow recording of visual and body
messages in individual files for each player. The
messages are stored in their original format to
keep the amount of additional processing during the game minimal. Playing a game with
the modified simulator will result in a number
of recorded files:
of each player of the left (L) and the right (R)
team including the goalkeeper (G) is stated. For
example, the column with head LG1 vx contains
the velocity of the left goalkeeper in x-direction.
Finally, information about the robots body and
head orientation and their view angle and quality is included. The absolute direction a player is
facing (with respect to the pitch coordinate system) is the sum of the body and head direction
of that player.
5 Description of local player data
The visual sensor of the players reports the objects currently seen. The information is automatically sent to players every sense step with
a frequency depending on the player view width
and quality; by default it is set to 150 ms. Thus,
in addition to the three files mentioned above,
44 more files are available for each game. For
each of the altogether 22 robots (ten field players and one goalkeeper per team), two files with
local player data is provided, hosting information about where the respective player sees the
landmarks and moving objects, respectively.
The file with final identifier landmarks provides the distances (in meters) and angles (in
degrees) to the respective landmarks relative to
the robot head orientation for each step. Analogously, the file with final identifier moving provides the actual relative distances and angles to
the ball and all other players. Sometimes the
player number or even the team name is not visible and hence unknown (u) to the robot. In
this case, the respective piece of information is
left out. If data is not available at all, then this
is marked by NAN in the respective table element.
The server also provides information about the
velocity, stamina, (yellow and red) cards, and
the commands (e.g. dash, turn, or kick) of the
robots. In some cases there is also information
about the observed state of other robots available, in particular, whether they are kicking (k),
tackling (t), or are a goalkeeper (g).
• the visual and body messages (two files for
each player). The file names follow the same
naming convention as the game data CSV
files (cf. Fig. 3), but use the .rcv suffix
for visual messages and .rcb for body messages.
• a recording of the game (the ground truth
in a binary format), using the suffix .rcg
• the commands from players as received by
the simulator (in plain text), using .rcl as
a suffix
To convert visual messages into CSV files,
we provide a see2csv.py Python program that
translates player visual messages into two files:
a CSV file for moving objects (other players and
the ball), and a CSV file with perceived landmarks.
To convert the simulator logfile (ground
truth) into a CSV file,
we provide
rcg2csv, a C++ program that is built
using the open source librcsc library
(see osdn.net/projects/rctools/releases/p3777).
Logfiles are recorded at regular intervals of
100 ms. rcg2csv can optionally also store all
simulation parameters in an additional CSV file.
6 Code
7 Conclusions
The soccer simulator communicates with players We released a large and unique dataset
using text messages in form of lists, via UDP. (RoboCupSimData), creating using a tourna-
5
ment of top RoboCup simulation league teams.
7.07 and Later. The RoboCup Federation.
Creating this dataset required modification of
URL https://sourceforge.net/projects/
the simulation software. We ran 25 repetitions of
sserver/files/rcssmanual/manual-7.08.
45 matches that last 10 mins, each. The dataset
1/manual.pdf.
will allow a number of problems to be investigated, from learning and testing approaches to Kitano H, Asada M, Kuniyoshi Y, Noda I, Osawa E and Matsubara H (1997) RoboCup: A
self-localization, predictive world-models, reinchallenge problem for AI. AI Magazine 18(1):
forcement learning. In this respect, the dataset
73–85.
shall be useful for the whole robotics community.
Lin Z, Gehring J, Khalidov V and Synnaeve G
(2017) STARDATA: A StarCraft AI research
Funding
dataset. CoRR – Computing Research ReposThe research reported in this paper has
itory abs/1708.02139, Cornell University Libeen supported by the German Academic Exbrary. URL http://arxiv.org/abs/1708.
change Service (DAAD) by funds of the Ger02139.
man Federal Ministry of Education and Research (BMBF) in the Programmes for Project- Michael O, Obst O, Schmidsberger F and
Stolzenburg F (2018) Analysing soccer games
Related Personal Exchange (PPP) under grant
with clustering and conceptors. In: Akyama
no. 57319564 and Universities Australia (UA) in
H, Obst O, Sammut C and Tonidandel F (eds.)
the Australia-Germany Joint Research CooperaRoboCup 2017: Robot Soccer World Cup XXI.
tion Scheme within the project Deep Conceptors
RoboCup International Symposium. Nagoya,
for Temporal Data Mining (Decorating).
Japan: Springer. To appear.
References
Noda I, Matsubara H, Hiraki K and Frank I
(1998) Soccer server: A tool for research on
Akiyama H, Dorer K and Lau N (2015) On
multiagent systems. Applied Artificial Intellithe progress of soccer simulation leagues. In:
gence 12(2–3): 233–250.
Bianchi RAC, Akin HL, Ramamoorthy S and
Sugiura K (eds.) RoboCup 2014: Robot World Olson CF (2000) Probabilistic self-localization
for mobile robots. IEEE Transactions on
Cup XVIII. RoboCup International SympoRobotics and Automation 16(1): 55–66.
sium, Lecture Notes in Computer Science, volume 8992. Springer, pp. 599–610.
Stolzenburg F, Obst O and Murray J (2002)
Qualitative velocity and ball interception. In:
Ben Amor H, Vogt D, Ewerton M, Berger E,
Jarke M, Köhler J and Lakemeyer G (eds.)
Jung B and Peters J (2013) Learning responKI-2002: Advances in Artificial Intelligence
sive robot behavior by imitation. In: Inter– Proceedings of 25th Annual German Connational Conference on Intelligent Robots and
ference on Artificial Intelligence, LNAI 2479.
Systems (IROS-2013). pp. 3257–3264.
Aachen: Springer, pp. 283–298.
Chebrolu N, Lottes P, Schaefer A, Winterhalter W, Burgard W and Stachniss C (2017) Taylor ME and Stone P (2009) Transfer learning for reinforcement learning domains: A surAgricultural robot dataset for plant classificavey. Journal of Machine Learning Research
tion, localization and mapping on sugar beet
10: 1633–1685.
fields. The International Journal of Robotics
Research 36(10): 1045–1052.
Tong CH, Gingras D, Larose K, Barfoot
TD and Dupuis É (2013) The Canadian
Chen M, Dorer K, Foroughi E, Heintz F, Huang
planetary emulation terrain 3D mapping
Z, Kapetanakis S, Kostiadis K, Kummeneje J,
dataset. International Journal of Robotics ReMurray J, Noda I, Obst O, Riley P, Steffens
search URL http://asrl.utias.utoronto.
T, Wang Y and Yin X (2003) RoboCup
ca/datasets/3dmap/.
Soccer Server — for Soccer Server Version
6
| 2cs.AI
|
Local Linearizability∗
Andreas Haas1 , Thomas A. Henzinger2 , Andreas Holzer3 ,
Christoph M. Kirsch4 , Michael Lippautz1 , Hannes Payer1 ,
Ali Sezgin5 , Ana Sokolova4 , and Helmut Veith6,7
arXiv:1502.07118v4 [cs.PL] 24 Jun 2016
1
2
3
4
5
6
7
Google Inc.
IST Austria, Austria
University of Toronto, Canada
University of Salzburg, Austria
University of Cambridge, UK
Vienna University of Technology, Austria
Forever in our hearts
Abstract
The semantics of concurrent data structures is usually given by a sequential specification and a
consistency condition. Linearizability is the most popular consistency condition due to its simplicity and general applicability. Nevertheless, for applications that do not require all guarantees
offered by linearizability, recent research has focused on improving performance and scalability
of concurrent data structures by relaxing their semantics.
In this paper, we present local linearizability, a relaxed consistency condition that is applicable
to container-type concurrent data structures like pools, queues, and stacks. While linearizability
requires that the effect of each operation is observed by all threads at the same time, local
linearizability only requires that for each thread T, the effects of its local insertion operations and
the effects of those removal operations that remove values inserted by T are observed by all threads
at the same time. We investigate theoretical and practical properties of local linearizability and
its relationship to many existing consistency conditions. We present a generic implementation
method for locally linearizable data structures that uses existing linearizable data structures as
building blocks. Our implementations show performance and scalability improvements over the
original building blocks and outperform the fastest existing container-type implementations.
1998 ACM Subject Classification D.3.1 [Programming Languages]: Formal Definitions and
Theory—Semantics; E.1 [Data Structures]: Lists, stacks, and queues; D.1.3 [Software]: Programming Techniques—Concurrent Programming
Keywords and phrases (concurrent) data structures, relaxed semantics, linearizability
1
Introduction
Concurrent data structures are pervasive all along the software stack, from operating system
code to application software and beyond. Both correctness and performance are imperative
for concurrent data structure implementations. Correctness is usually specified by relating concurrent executions, admitted by the implementation, with sequential executions,
admitted by the sequential version of the data structure. The latter form the sequential
specification of the data structure. This relationship is formally captured by consistency
conditions, such as linearizability, sequential consistency, or quiescent consistency [25].
Linearizability [26] is the most accepted consistency condition for concurrent data structures due to its simplicity and general applicability. It guarantees that the effects of all
∗
This paper is an extended version of [17]
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Local Linearizability
operations by all threads are observed consistently. This global visibility requirement imposes the need of extensive synchronization among threads which may in turn jeopardize
performance and scalability. In order to enhance performance and scalability of implementations, recent research has explored relaxed sequential specifications [23, 40, 2], resulting in
well-performing implementations of concurrent data structures [2, 18, 23, 28, 38, 6]. Except
for [27], the space of alternative consistency conditions that relax linearizability has been
left unexplored to a large extent. In this paper, we explore (part of) this gap by investigating local linearizability, a novel consistency condition that is applicable to a large class
of concurrent data structures that we call container-type data structures, or containers for
short. Containers include pools, queues, and stacks. A fine-grained spectrum of consistency conditions enables us to describe the semantics of concurrent implementations more
precisely, e.g., we show in our appendix that work stealing queues [35] which could only be
proven to be linearizable wrt pool are actually locally linearizable wrt double-ended queue.
Local linearizability is a (thread-)local consistency
condition that guarantees that insertions per thread are
enq(1)
deq(2)
T1
observed consistently. While linearizability requires
enq(2)
deq(1)
T2
a consistent view over all insertions, we only require
that projections of the global history—so called thread- The thread-induced history of thread T1
induced histories—are linearizable. The induced his- is enclosed by a dashed line while the
thread-induced history of thread T2 is
tory of a thread T is a projection of a program exe- enclosed by a solid line.
cution to the insert-operations in T combined with all
remove-operations that remove values inserted by T irFigure 1 Local Linearizability
respective of whether they happen in T or not. Then,
the program execution is locally linearizable iff each thread-induced history is linearizable.
Consider the example (sequential) history depicted in Figure 1. It is not linearizable wrt a
queue since the values are not dequeued in the same order as they were enqueued. However,
each thread-induced history is linearizable wrt a queue and, therefore, the overall execution
is locally linearizable wrt a queue. In contrast to semantic relaxations based on relaxing
sequential semantics such as [23, 2], local linearizability coincides with sequential correctness for single-threaded histories, i.e., a single-threaded and, therefore, sequential history is
locally linearizable wrt a given sequential specification if and only if it is admitted by the
sequential specification.
Local linearizability is to linearizability what coherence is to sequential consistency. Coherence [22], which is almost universally accepted as the absolute minimum that a shared
memory system should satisfy, is the requirement that there exists a unique global order per
shared memory location. Thus, while all accesses by all threads to a given memory location
have to conform to a unique order, consistent with program order, the relative ordering of
accesses to multiple memory locations do not have to be the same. In other words, coherence
is sequential consistency per memory location. Similarly, local linearizability is linearizability per local history. In our view, local linearizability offers enough consistency for the
correctness of many applications as it is the local view of the client that often matters. For
example, in a locally linearizable queue each client (thread) has the impression of using a
perfect queue—no reordering will ever be observed among the values inserted by a single
thread. Such guarantees suffice for many e-commerce and cloud applications. Implementations of locally linearizable data structures have been successfully applied for managing
free lists in the design of the fast and scalable memory allocator scalloc [5]. Moreover, except for fairness, locally linearizable queues guarantee all properties required from Dispatch
Queues [1], a common concurrency programming mechanism on mobile devices.
A. Haas et al.
In this paper, we study theoretical and practical properties of local linearizability. Local
linearizability is compositional—a history over multiple concurrent objects is locally linearizable iff all per-object histories are locally linearizable (see Thm. 12) and locally linearizable
container-type data structures, including queues and stacks, admit only “sane” behaviours—
no duplicated values, no values returned from thin air, and no values lost (see Prop. 4). Local
linearizability is a weakening of linearizability for a natural class of data structures including
pools, queues, and stacks (see Sec. 4). We compare local linearizability to linearizability,
sequential, and quiescent consistency, and to many shared-memory consistency conditions.
Finally, local linearizability leads to new efficient implementations. We present a generic
implementation scheme that, given a linearizable implementation of a sequential specification S, produces an implementation that is locally linearizable wrt S (see Sec. 6). Our
implementations show dramatic improvements in performance and scalability. In most cases
the locally linearizable implementations scale almost linearly and even outperform state-ofthe-art pool implementations. We produced locally linearizable variants of state-of-the-art
concurrent queues and stacks, as well as of the relaxed data structures from [23, 28]. The
latter are relaxed in two dimensions: they are locally linearizable (the consistency condition is relaxed) and are out-of-order-relaxed (the sequential specification is relaxed). The
speedup of the locally linearizable implementation to the fastest linearizable queue (LCRQ)
and stack (TS Stack) implementation at 80 threads is 2.77 and 2.64, respectively. Verification of local linearizability, i.e. proving correctness, for each of our new locally linearizable
implementations is immediate, given that the starting implementations are linearizable.
2
Semantics of Concurrent Objects
The common approach to define the semantics of an implementation of a concurrent data
structure is (1) to specify a set of valid sequential behaviors—the sequential specification, and
(2) to relate the admissible concurrent executions to sequential executions specified by the
sequential specification—via the consistency condition. That means that an implementation
of a concurrent data structure actually corresponds to several sequential data structures, and
vice versa, depending on the consistency condition used. A (sequential) data structure D is
an object with a set of method calls Σ. We assume that method calls include parameters,
i.e., input and output values from a given set of values. The sequential specification S of D
is a prefix-closed subset of Σ∗ . The elements of S are called D-valid sequences. For ease of
presentation, we assume that each value in a data structure can be inserted and removed at
most once. This is without loss of generality, as we may see the set of values as consisting
of pairs of elements (core values) and version numbers, i.e. V = E × N. Note that this
is a technical assumption that only makes the presentation and the proofs simpler, it is
not needed and not done in locally linearizable implementations. While elements may be
inserted and removed multiple times, the version numbers provide uniqueness of values. Our
assumption ensures that whenever a sequence s is part of a sequential specification S, then,
each method call in s appears exactly once. An additional core value, that is not an element,
is empty. It is returned by remove method calls that do not find an element to return. We
denote by Emp the set of values that are versions of empty, i.e., Emp = {empty} × N.
I Definition 1 (Appears-before Order, Appears-in Relation). Given a sequence s ∈ Σ∗ in which
each method call appears exactly once, we denote by ≺s the total appears-before order over
method calls in s. Given a method call m ∈ Σ, we write m ∈ s for m appears in s.
Throughout the paper, we will use pool, queue, and stack as typical examples of containers. We specify their sequential specifications in an axiomatic way [24], i.e., as sets of
XX:3
XX:4
Local Linearizability
(1)
(2)
(3)
(4)
(5)
∀i, j ∈ {1, . . . , n}. s = m1 . . . mn ∧ mi = mj ⇒ i = j
∀x ∈ V. r(x) ∈ s ⇒ i(x) ∈ s ∧ i(x) ≺s r(x)
∀e ∈ Emp. ∀x ∈ V. i(x) ≺s r(e) ⇒ r(x) ≺s r(e)
∀x, y ∈ V. i(x) ≺s i(y) ∧ r(y) ∈ s ⇒ r(x) ∈ s ∧ r(x) ≺s r(y)
∀x, y ∈ V. i(x) ≺s i(y) ≺s r(x) ⇒ r(y) ∈ s ∧ r(y) ≺s r(x)
Table 1 The pool axioms (1), (2), (3); the queue order axiom (4); the stack order axiom (5)
axioms that exactly define the valid sequences.
I Definition 2 (Pool, Queue, & Stack). A pool, queue, and stack with values in a set V
have the sets of methods ΣP = {ins(x), rem(x) | x ∈ V } ∪ {rem(e) | e ∈ Emp}, ΣQ =
{enq(x), deq(x) | x ∈ V } ∪ {deq(e) | e ∈ Emp}, and ΣS = {push(x), pop(x) | x ∈ V } ∪
{pop(e) | e ∈ Emp}, respectively. We denote the sequential specification of a pool by SP ,
the sequential specification of a queue by SQ , and the sequential specification of a stack
by SS . A sequence s ∈ Σ∗P belongs to SP iff it satisfies axioms (1) - (3) in Table 1—the
pool axioms—when instantiating i() with ins() and r() with rem(). We keep axiom (1)
for completeness, although it is subsumed by our assumption that each value is inserted
and removed at most once. Specification SQ contains all sequences s that satisfy the pool
axioms and axiom (4)—the queue order axiom—after instantiating i() with enq() and r()
with deq(). Finally, SS contains all sequences s that satisfy the pool axioms and axiom
(5)—the stack order axiom—after instantiating i() with push() and r() with pop().
We represent concurrent executions via concurrent histories. An example history is
shown in Figure 1. Each thread executes a sequence of method calls from Σ; method
calls executed by different threads may overlap (which does not happen in Figure 1). The
real-time duration of method calls is irrelevant for the semantics of concurrent objects;
all that matters is whether method calls overlap. Given this abstraction, a concurrent
history is fully determined by a sequence of invocation and response events of method calls.
We distinguish method invocation and response events by augmenting the alphabet. Let
Σi = {mi | m ∈ Σ} and Σr = {mr | m ∈ Σ} denote the sets of method-invocation events
and method-response events, respectively, for the method calls in Σ. Moreover, let I be the
set of thread identifiers. Let ΣIi = {mki | m ∈ Σ, k ∈ I} and ΣIr = {mkr | m ∈ Σ, k ∈ I}
denote the sets of method-invocation and -response events augmented with identifiers of
executing threads. For example, mki is the invocation of method call m by thread k. Before
we proceed, we mention a standard notion that we will need in several occasions.
I Definition 3 (Projection). Let s be a sequence over alphabet Σ and M ⊆ Σ. By s|M
we denote the projection of s on the symbols in M , i.e., the sequence obtained from s by
removing all symbols that are not in M .
I Definition 4 (History). A (concurrent) history h is a sequence in (ΣIi ∪ ΣIr )∗ where (1) no
invocation or response event appears more than once, i.e., if h = m1 . . . mn and mh = mk∗ (x)
and mj = ml∗ (x), for ∗ ∈ {i, r}, then h = j and k = l, and (2) if a response event mkr appears
in h, then the corresponding invocation event mki also appears in h and mi ≺h mr .
I Example 5. A queue history (left) and its formal representation as a sequence (right):
T1
T2
enq(2)
deq(1)
enq(1)
enq(2)1i enq(1)2i enq(2)1r deq(1)1i enq(1)2r deq(1)1r
A. Haas et al.
A history is sequential if every response event is immediately preceded by its matching invocation event and vice versa. Hence, we may ignore thread identifiers and identify
a sequential history with a sequence in Σ∗ , e.g., enq(1)enq(2)deq(2)deq(1) identifies the
sequential history in Figure 1.
A history h is well-formed if h|k is sequential for every thread identifier k ∈ I where h|k
denotes the projection of h on the set {mki | m ∈ Σ} ∪ {mkr | m ∈ Σ} of events that are local
to thread k. From now on we will use the term history for well-formed history. Also, we
may omit thread identifiers if they are not essential in a discussion.
A history h determines a partial order on its set of method calls, the precedence order:
I Definition 6 (Appears-in Relation, Precedence Order). The set of method calls of a history h
is M (h) = {m | mi ∈ h}. A method call m appears in h, notation m ∈ h, if m ∈ M (h).
The precedence order for h is the partial order <h such that, for m, n ∈ h, we have that
m <h n iff mr ≺h ni . By <kh we denote <h|k , the subset of the precedence order that relates
pairs of method calls of thread k, i.e., the program order of thread k.
We can characterize a sequential history as a history whose precedence order is total. In
particular, the precedence order <s of a sequential history s coincides with its appears-before
order ≺s . The total order for history s in Fig. 1 is enq(1) <s enq(2) <s deq(2) <s deq(1).
I Definition 7 (Projection to a set of method calls). Let h be a history, M ⊆ Σ, MiI = {mki |
m ∈ M, k ∈ I}, and MrI = {mkr | m ∈ M, k ∈ I}. Then, we write h|M for h|(MiI ∪ MrI ).
Note that h|M inherits h’s precedence order: m <h|M n ⇔ m ∈ M ∧ n ∈ M ∧ m <h n
A history h is complete if the response of every invocation event in h appears in h. Given
a history h, Complete(h) denotes the set of all completions of h, i.e., the set of all complete
histories that are obtained from h by appending missing response events and/or removing
pending invocation events. Note that Complete(h) = {h} iff h is a complete history.
A concurrent data structure D over a set of methods Σ is a (prefix-closed) set of concurrent histories over Σ. A history may involve several concurrent objects. Let O be a set of
concurrent objects with individual sets of method calls Σq and sequential specifications Sq
for each object q ∈ O. A history h over O is a history over the (disjoint) union of method
S
calls of all objects in O, i.e., it has a set of method calls q∈O {q.m | m ∈ Σq }. The added
prefix q. ensures that the union is disjoint. The projection of h to an object q ∈ O, denoted
by h|q, is the history with a set of method calls Σq obtained by removing the prefix q. in
every method call in h|{q.m | m ∈ Σq }.
I Definition 8 (Linearizability [26]). A history h is linearizable wrt the sequential specification S if there is a sequential history s ∈ S and a completion hc ∈ Complete(h) such
that (1) s is a permutation of hc , and (2) s preserves the precedence order of hc , i.e., if
m <hc n, then m <s n. We refer to s as a linearization of h. A concurrent data structure
D is linearizable wrt S if every history h of D is linearizable wrt S. A history h over a set
of concurrent objects O is linearizable wrt the sequential specifications Sq for q ∈ O if there
exists a linearization s of h such that s|q ∈ Sq for each object q ∈ O.
3
Local Linearizability
Local linearizability is applicable to containers whose set of method calls is a disjoint union
Σ = Ins ∪ Rem ∪ DOb ∪ SOb of insertion method calls Ins, removal method calls Rem, dataobservation method calls DOb, and (global) shape-observation method calls SOb. Insertions
(removals) insert (remove) a single value in the data set V or empty; data observations return
XX:5
XX:6
Local Linearizability
a single value in V ; shape observations return a value (not necessarily in V ) that provides
information on the shape of the state, for example, the size of a data structure. Examples
of data observations are head(x) (queue), top(x) (stack), and peek(x) (pool). Examples of
shape observations are empty(b) that returns true if the data structure is empty and false
otherwise, and size(n) that returns the number of elements in the data structure.
Even though we refrain from formal definitions, we want to stress that a valid sequence
of a container remains valid after deleting observer method calls:
S | (Ins ∪ Rem) ⊆ S.
(1)
There are also containers with multiple insert/remove methods, e.g., a double-ended queue
(deque) is a container with insert-left, insert-right, remove-left, and remove-right methods,
to which local linearizability is also applicable. However, local linearizability requires that
each method call is either an insertion, or a removal, or an observation. As a consequence,
set is not a container according to our definition, as in a set ins(x) acts as a global observer
first, checking whether (some version of) x is already in the set, and if not inserts x. Also
hash tables are not containers for a similar reason.
Note that the arity of each method call in a container being one excludes data structures
like snapshot objects. It is possible to deal with higher arities in a fairly natural way,
however, at the cost of complicated presentation. We chose to present local linearizability
on simple containers only. We present the definition of local linearizability without shape
observations here and discuss shape observations in Appendix A.
I Definition 9 (In- and out-methods). Let h be a container history. For each thread T
we define two subsets of the methods in h, called in-methods IT and out-methods OT of
thread T , respectively:
IT = {m | m ∈ M (h|T ) ∩ Ins}
OT =
{m(a) ∈ M (h) ∩ Rem | ins(a) ∈ IT } ∪ {m(e) ∈ M (h) ∩ Rem | e ∈ Emp}
∪ {m(a) ∈ M (h) ∩ DOb | ins(a) ∈ IT }.
Hence, the in-methods for thread T are all insertions performed by T . The out-methods
are all removals and data observers that return values inserted by T . Removals that remove
the value empty are also automatically added to the out-methods of T as any thread (and
hence also T ) could be the cause of “inserting” empty. This way, removals of empty serve as
means for global synchronization. Without them each thread could perform all its operations
locally without ever communicating with the other threads. Note that the out-methods OT
of thread T need not be performed by T , but they return values that are inserted by T .
I Definition 10 (Thread-induced History). Let h be a history. The thread-induced history
hT is the projection of h to the in- and out-methods of thread T , i.e., hT = h| (IT ∪ OT ).
I Definition 11 (Local Linearizability). A history h is locally linearizable wrt a sequential
specification S if (1) each thread-induced history hT is linearizable wrt S, and (2) the
thread-induced histories hT form a decomposition of h, i.e., m ∈ h ⇒ m ∈ hT for some
thread T . A data structure D is locally linearizable wrt S if every history h of D is locally
linearizable wrt S. A history h over a set of concurrent objects O is locally linearizable wrt
the sequential specifications Sq for q ∈ O if each thread-induced history is linearizable over O
and the thread-induced histories form a decomposition of h, i.e., q.m ∈ h ⇒ q.m ∈ hT for
some thread T .
Local linearizability is sequentially correct, i.e., a single-threaded (necessarily sequential)
history h is locally linearizable wrt a sequential specification S iff h ∈ S. Like linearizabil-
A. Haas et al.
XX:7
ity [25], local linearizability is compositional. The complete proof of the following theorem
and missing or extended proofs of all following properties can be found in Appendix B.
I Theorem 12 (Compositionality). A history h over a set of objects O with sequential specifications Sq for q ∈ O is locally linearizable iff h|q is locally linearizable wrt Sq for every q ∈ O.
Proof (Sketch). The property follows from the compositionality of linearizability and the
fact that (h|q)T = hT |q for every thread T and object q.
J
The Choices Made. Splitting a global history into subhistories and requiring consistency
for each of them is central to local linearizability. While this is common in shared-memory
consistency conditions [22, 31, 32, 3, 16, 4, 20], our study of local linearizability is a first
step in exploring subhistory-based consistency conditions for concurrent objects.
We chose thread-induced subhistories since thread-locality reduces contention in concurrent objects and is known to lead to high performance as confirmed by our experiments. To
assign method calls to thread-induced histories, we took a data-centric point of view by (1)
associating data values to threads, and (2) gathering all method calls that insert/return a
data value into the subhistory of the associated thread (Def. 9). We associate data values to
the thread that inserts them. One can think of alternative approaches, for example, associate with a thread the values that it removed. In our view, the advantages of our choice are
clear: First, by assigning inserted values to threads, every value in the history is assigned
to some thread. In contrast, in the alternative approach, it is not clear where to assign the
values that are inserted but not removed. Second, assigning inserted values to the inserting
thread enables eager removals and ensures progress in locally linearizable data structures.
In the alternative approach, it seems like the semantics of removing empty should be local.
An orthogonal issue is to assign values from shape observations to threads. In Appendix A, we discuss two meaningful approaches and show how local linearizability can be
extended towards shape and data observations that appear in insertion operations of sets.
Finally, we have to choose a consistency condition required for each of the subhistories.
We chose linearizability as it is the best (strong) consistency condition for concurrent objects.
4
Local Linearizability vs. Linearizability
We now investigate the connection between local linearizability and linearizability.
I Proposition 1 (Lin 1). In general, linearizability does not imply local linearizability.
Proof. We provide an example of a data structure that is linearizable but not locally
linearizable. Consider a sequential specification SNearlyQ which behaves like a queue except when the first two insertions were performed without a removal in between—then
the first two elements are removed out of order. Formally, s ∈ SNearlyQ iff (1) s =
s1 enq(a)enq(b)s2 deq(b)s3 deq(a)s4 where s1 enq(a)enq(b)s2 deq(a)s3 deq(b)s4 ∈ SQ and s1 ∈
{deq(e) | e ∈ Emp}∗ for some a, b ∈ V , or (2) s ∈ SQ and s 6= s1 enq(a)enq(b)s2 for
s1 ∈ {deq(e) | e ∈ Emp}∗ and a, b ∈ V . The example below is linearizable wrt SNearlyQ .
However, T1 ’s induced history enq(1)enq(2)deq(1)deq(2) is not.
T1
T2
enq(1)
enq(2) deq(3)
enq(3)
deq(2)
deq(1)
J
The following condition on a data structure specification is sufficient for linearizability
to imply local linearizability and is satisfied, e.g., by pool, queue, and stack.
XX:8
Local Linearizability
T1
T2
i(1)
r(empty)
i(2)
r(2)
r(1)
Figure 2 LL, not SC (Pool, Queue, Stack)
T1
T2
i(1)
r(1)
r(empty)
Figure 3 SC, not LL (Pool, Queue, Stack)
I Definition 13 (Closure under Data-Projection). A seq. specification S over Σ is closed
under data-projection 1 iff for all s ∈ S and all V 0 ⊆ V , s|{m(x) ∈ Σ | x ∈ V 0 ∪ Emp} ∈ S.
For s = enq(1)enq(3)enq(2)deq(3)deq(1)deq(2) we have s ∈ SNearlyQ , but s|{enq(x), deq(x) |
x ∈ {1, 2} ∪ Emp} ∈
/ SNearlyQ , i.e., SNearlyQ is not closed under data-projection.
I Proposition 2 (Lin 2). Linearizability implies local linearizability for sequential specifications that are closed under data-projection.
Proof (Sketch). The property follows from Definition 13 and Equation (1).
J
There exist corner cases where local linearizability coincides with linearizability, e.g., for
S = ∅ or S = Σ∗ , or for single-producer/multiple-consumer histories.
We now turn our attention to pool, queue, and stack.
I Proposition 3. The seq. specifications SP , SQ , and SS are closed under data-projection.
Proof (Sketch). Let s ∈ SP , V 0 ⊆ V , and let s0 = s| ({ins(x), rem(x) | x ∈ V 0 ∪ Emp}).
Then, it suffices to check that all axioms for pool (Definition 2 and Table 1) hold for s0 . J
I Theorem 14 (Pool & Queue & Stack, Lin). For pool, queue, and stack, local linearizability
is (strictly) weaker than linearizability.
Proof. Linearizability implies local linearizability for pool, queue, and stack as a consequence of Proposition 2 and Proposition 3. The history in Figure 2 is locally linearizable
but not linearizable wrt pool, queue and stack (after suitable renaming of method calls). J
Although local linearizability wrt a pool does not imply linearizability wrt a pool (Theorem 14), it still guarantees several properties that ensure sane behavior as stated next.
I Proposition 4 (LocLin Pool). Let h be a locally linearizable history wrt a pool. Then:
1. No value is duplicated, i.e., every remove method appears in h at most once.
2. No out-of-thin-air values, i.e., ∀x ∈ V. rem(x) ∈ h ⇒ ins(x) ∈ h ∧ rem(x)6<h ins(x).
3. No value is lost, i.e., ∀x ∈ V. ∀e ∈ Emp. rem(e) <h rem(x) ⇒ ins(x) 6<h rem(e) and
∀x ∈ V. ∀e ∈ Emp. ins(x) <h rem(e) ⇒ rem(x) ∈ h ∧ rem(e)6<h rem(x).
Proof. By direct unfolding of the definitions.
J
Note that if a history h is linearizable wrt a pool, then all of the three stated properties
hold, as a consequence of linearizability and the definition of SP .
5
Local Linearizability vs. Other Relaxed Consistency Conditions
We compare local linearizability with other classical consistency conditions to better understand its guarantees and implications.
1
The same notion has been used in [7] under the name closure under projection.
A. Haas et al.
XX:9
Sequential Consistency (SC). A history h is sequentially consistent [25, 30] wrt a sequential specification S, if there exists a sequential history s ∈ S and a completion hc ∈
Complete(h) such that (1) s is a permutation of hc , and (2) s preserves each thread’s program order, i.e., if m <Th n, for some thread T , then m <s n. We refer to s as a sequential
witness of h. A data structure D is sequentially consistent wrt S if every history h of D is
sequentially consistent wrt S.
Sequential consistency is a useful consistency condition for shared memory but it is not
really suitable for data structures as it allows for behavior that excludes any coordination
between threads [39]: an implementation of a data structure in which every thread uses a
dedicated copy of a sequential data structure without any synchronization is sequentially
consistent. A sequentially consistent queue might always return empty in one (consumer)
thread as the point in time of the operation can be moved, e.g., see Figure 3. In a producerconsumer scenario such a queue might end up with some threads not doing any work.
I Theorem 15 (Pool, Queue & Stack, SC). For pool, queue, and stack, local linearizability
is incomparable to sequential consistency.
J
Figures 2 and 3 give example histories that show the statement of Theorem 15. In contrast
to local linearizability, sequential consistency is not compositional [25].
(Quantitative) Quiescent Consistency (QC & QQC). Like linearizability and sequential
consistency, quiescent consistency [13, 25] also requires the existence of a sequential history,
a quiescent witness, that satisfies the sequential specification. All three consistency conditions impose an order on the method calls of a concurrent history that a witness has to
preserve. Quiescent consistency uses the concept of quiescent states to relax the requirement
of preserving the precedence order imposed by linearizability. A quiescent state is a point
in a history at which there are no pending invocation events (all invoked method calls have
already responded). In a quiescent witness, a method call m has to appear before a method
call n if and only if there is a quiescent state between m and n. Method calls between
two consecutive quiescent states can be ordered arbitrarily. Quantitative quiescent consistency [27] refines quiescent consistency by bounding the number of reorderings of operations
between two quiescent states based on the concurrent behavior between these two states.
The next result about quiescent consistency for pool is needed to establish the connection
between quiescent consistency and local linearizability.
I Proposition 5. A pool history h satisfying 1.-3. of Prop. 4 is quiescently consistent.
J
From Prop. 4 and 5 follows that local linearizability implies quiescent consistency for pool.
I Theorem 16 (Pool, Queue & Stack, QC). For pool, local linearizability is (strictly) stronger
than quiescent consistency. For queue and stack, local linearizability is incomparable to
quiescent consistency.
J
Local linearizability also does not imply the stronger condition of quantitative quiescent consistency. Like local linearizability, quiescent consistency and quantitative quiescent
consistency are compositional [25, 27]. For details, please see Appendix D.
Consistency Conditions for Distributed Shared Memory. There is extensive research on
consistency conditions for distributed shared memory [3, 4, 8, 16, 20, 22, 30, 31, 32]. In
Appendix E, we compare local linearizability against coherence, PRAM consistency, processor consistency, causal consistency, and local consistency. All these conditions split a
history into subhistories and require consistency of the subhistories. For our comparison,
XX:10
Local Linearizability
T1
T2
ins(1)
ins(2) head(2) head(1) head(2) head(1)
Figure 4 Problematic shared-memory history.
we first define a sequential specification SM for a single memory location. We assume that
each memory location is preinitialized with a value vinit ∈ V . A read-operation returns
the value of the last write-operation that was performed on the memory location or vinit if
there was no write-operation. We denote write-operations by ins and read-operations by
head. Formally, we define SM as SM = {head(vinit )}? · {ins(v)head(v)i | i ≥ 0, v ∈ V }? .
Note that read-operations are data observations and the same value can be read multiple
times. For brevity, we only consider histories that involve a single memory location. In the
following, we summarize our comparison. For details, please see Appendix E.
While local linearizability is well-suited for concurrent data structures, this is not necessarily true for the mentioned shared-memory consistency conditions. On the other hand,
local linearizability appears to be problematic for shared memory. Consider the locally
linearizable history in Figure 4. There, the read values oscillate between different values
that were written by different threads. Therefore, local linearizability does not imply any
of the shared-memory consistency conditions. In Appendix E, we further show that local
linearizability is incomparable to all considered shared-memory conditions.
6
Locally Linearizable Implementations
In this section, we focus on locally linearizable data structure implementations that are generic as follows: Choose a linearizable implementation of a data structure Φ wrt a sequential
specification SΦ , and we turn it into a (distributed) data structure called LLD Φ that is
locally linearizable wrt SΦ . An LLD implementation takes several copies of Φ (that we call
backends) and assigns to each thread T a backend ΦT . Then, when thread T inserts an
element into LLD Φ, the element is inserted into ΦT , and when an arbitrary thread removes
an element from LLD Φ, the element is removed from some ΦT eagerly, i.e., if no element is
found in the attempted backend ΦT the search for an element continues through all other
backends. If no element is found in one round through the backends, then we return empty.
I Proposition 6 (LLD correctness). Let Φ be a data structure implementation that is linearizable wrt a sequential specification SΦ . Then LLD Φ is locally linearizable wrt SΦ .
Proof. Let h be a history of LLD Φ. The crucial observation is that each thread-induced
history hT is a backend history of ΦT and hence linearizable wrt SΦ .
J
Any number of copies (backends) is allowed in this generic implementation of LLD Φ.
If we take just one copy, we end up with a linearizable implementation. Also, any way of
choosing a backend for removals is fine. However, both the number of backends and the
backend selection strategy upon removals affect the performance significantly. In our LLD
Φ implementations we use one backend per thread, resulting in no contention on insertions,
and always attempt a local remove first. If this does not return an element, then we continue
a search through all other backends starting from a randomly chosen backend.
LLD Φ is an implementation closely related to Distributed Queues (DQs) [18]. A DQ
is a (linearizable) pool that is organized as a single segment of length ` holding ` backends.
DQs come in different flavours depending on how insert and remove methods are distributed
A. Haas et al.
across the segment when accessing backends. No DQ variant in [18] follows the LLD approach described above. Moreover, while DQ algorithms are implemented for a fixed number
of backends, LLD Φ implementations manage a segment of variable size, one backend per
(active) thread. Note that the strategy of selecting backends in the LLD Φ implementations is similar to other work in work stealing [35]. However, in contrast to this work
our data structures neither duplicate nor lose elements. LLD (stack) implementations have
been successfully applied for managing free lists in the fast and scalable memory allocator
scalloc [5]. The guarantees provided by local linearizability are not needed for the correctness of scalloc, i.e., the free lists could also use a weak pool (pool without a linearizable
emptiness check). However, the LLD stack implementations provide good caching behavior
when threads operate on their local stacks whereas a weak pool would potentially negatively
impact performance.
We have implemented LLD variants of strict and relaxed queue and stack implementations. None of our implementations involves observation methods, but the LLD algorithm
can easily be extended to support observation methods. For details, please see App. F.4.
Finally, let us note that we have also experimented with other locally linearizable implementations that lacked the genericity of the LLD implementations, and whose performance
evaluation did not show promising results (see App. G). As shown in Sec. 4, a locally linearizable pool is not a linearizable pool, i.e., it lacks a linearizable emptiness check. Indeed,
LLD implementations do not provide a linearizable emptiness check, despite of eager removes. We provide LL+ D Φ, a variant of LLD Φ, that provides a linearizable emptiness
check under mild conditions on the starting implementation Φ (see App. F.1 for details).
Experimental Evaluation. All experiments ran on a uniform memory architecture (UMA)
machine with four 10-core 2GHz Intel Xeon E7-4850 processors supporting two hardware
threads (hyperthreads) per core, 128GB of main memory, and Linux kernel version 3.8.0.
We also ran the experiments without hyper-threading resulting in no noticeable difference.
The CPU governor has been disabled. All measurements were obtained from the artifactevaluated Scal benchmarking framework [12, 19, 11], where you can also find the code of
all involved data structures. Scal uses preallocated memory (without freeing it) to avoid
memory management artifacts. For all measurements we report the arithmetic mean and
the 95% confidence interval (sample size=10, corrected sample standard deviation).
In our experiments, we consider the linearizable queues Michael-Scott queue (MS) [34]
and LCRQ [36] (improved version [37]), the linearizable stacks Treiber stack (Treiber) [42]
and TS stack [14], the k-out-of-order relaxed k-FIFO queue [28] and k-Stack [23] and linearizable well-performing pools based on distributed queues using random balancing [18] (1-RA
DQ for queue, and 1-RA DS for stack). For each of these implementations (but the pools)
we provide LLD variants (LLD LCRQ, LLD TS stack, LLD k-FIFO, and LLD k-Stack) and,
when possible, LL+ D variants (LL+ D MS queue and LL+ D Treiber stack). Making the
pools locally linearizable is not promising as they are already distributed. Whenever LL+ D
is achievable for a data structure implementation Φ we present only results for LL+ D Φ as,
in our workloads, LLD Φ and LL+ D Φ implementations perform with no visible difference.
We evaluate the data structures on a Scal producer-consumer benchmark where each
producer and consumer is configured to execute 106 operations. To control contention, we
add a busy wait of 5µs between operations. This is important as too high contention results in measuring hardware or operating system (e.g., scheduling) artifacts. The number of
threads ranges between 2 and 80 (number of hardware threads) half of which are producers
and half consumers. To relate performance and scalability we report the number of data
XX:11
million operations per sec (more is better)
Local Linearizability
26
24
22
20
18
16
14
12
10
8
6
4
2
0
MS
LCRQ
k-FIFO
LL+D MS
LLD LCRQ
LLD k-FIFO
1-RA DQ
2
10
20
30
40
50
60
70
80
number of threads
“queue-like” data structures
million operations per sec (more is better)
XX:12
26
24
22
20
18
16
14
12
10
8
6
4
2
0
Treiber
TS Stack
k-Stack
LL+D Treiber
LLD TS Stack
LLD k-Stack
1-RA DS
2
10
20
30
40
50
60
70
80
number of threads
“stack-like” data structures
Figure 5 Performance and scalability of producer-consumer microbenchmarks with an increasing
number of threads on a 40-core (2 hyperthreads per core) machine
structure operations per second. Data structures that require parameters to be set are configured to allow maximum parallelism for the producer-consumer workload with 80 threads.
This results in k = 80 for all k-FIFO and k-Stack variants (40 producers and 40 consumers
in parallel on a single segment), p = 80 for 1-RA-DQ and 1-RA-DS (40 producers and
40 consumers in parallel on different backends). The TS Stack algorithm also needs to be
configured with a delay parameter. We use optimal delay (7µs) for the TS Stack and zero
delay for the LLD TS Stack, as delays degrade the performance of the LLD implementation.
Figure 5 shows the results of the producer-consumer benchmarks. Similar to experiments performed elsewhere [14, 23, 28, 36] the well-known algorithms MS and Treiber do
not scale for 10 or more threads. The state-of-the-art linearizable queue and stack algorithms
LCRQ and TS-interval Stack either perform competitively with their k-out-of-order relaxed
counter parts k-FIFO and k-Stack or even outperform and outscale them. For any implementation Φ, LLD Φ and LL+ D Φ (when available) perform and scale significantly better
than Φ does, even slightly better than the state-of-the-art pool that we compare to. The best
improvement show LLD variants of MS queue and Treiber stack. The speedup of the locally
linearizable implementation to the fastest linearizable queue (LCRQ) and stack (TS Stack)
A. Haas et al.
implementation at 80 threads is 2.77 and 2.64, respectively. The performance degradation
for LCRQ between 30 and 70 threads aligns with the performance of fetch-and-inc—the
CPU instruction that atomically retrieves and modifies the contents of a memory location—
on the benchmarking machine, which is different on the original benchmarking machine [36].
LCRQ uses fetch-and-inc as its key atomic instruction.
7
Conclusion & Future Work
Local linearizability splits a history into a set of thread-induced histories and requires consistency of all such. This yields an intuitive consistency condition for concurrent objects
that enables new data structure implementations with superior performance and scalability.
Local linearizability has desirable properties like compositionality and well-behavedness for
container-type data structures. As future work, it is interesting to investigate the guarantees
that local linearizability provides to client programs along the line of [15].
Acknowledgments
This work has been supported by the National Research Network RiSE on Rigorous Systems
Engineering (Austrian Science Fund (FWF): S11402-N23, S11403-N23, S11404-N23, S11411N23), a Google PhD Fellowship, an Erwin Schrödinger Fellowship (Austrian Science Fund
(FWF): J3696-N26), EPSRC grants EP/H005633/1 and EP/K008528/1, the Vienna Science
and Technology Fund (WWTF) trough grant PROSEED, the European Research Council
(ERC) under grant 267989 (QUAREM) and by the Austrian Science Fund (FWF) under
grant Z211-N23 (Wittgenstein Award).
References
1
2
3
4
5
6
7
8
9
10
URL:
https://developer.apple.com/library/ios/documentation/General/
Conceptual/ConcurrencyProgrammingGuide/OperationQueues/OperationQueues.
html.
Y. Afek, G. Korland, and E. Yanovsky. Quasi-Linearizability: Relaxed Consistency for
Improved Concurrency. In OPODIS, pages 395–410, 2010.
M. Ahamad, R.A. Bazzi, R. John, P. Kohli, and G. Neiger. The Power of Processor
Consistency. In SPAA, pages 251–260, 1993.
M. Ahamad, G. Neiger, J.E. Burns, P. Kohli, and P.W. Hutto. Causal memory: definitions,
implementation, and programming. Distributed Computing, 9(1):37–49, 1995.
M. Aigner, C. M. Kirsch, M. Lippautz, and A. Sokolova. Fast, multicore-scalable, lowfragmentation memory allocation through large virtual memory and global data structures.
In OOPSLA, pages 451–469, 2015.
D. Alistarh, J. Kopinsky, J. Li, and N. Shavit. The SprayList: A Scalable Relaxed Priority
Queue. In PPoPP, pages 11–20, 2015.
A. Bouajjani, M. Emmi, C. Enea, and J. Hamza. On Reducing Linearizability to State
Reachability. In ICALP, pages 95–107, 2015.
S. Burckhardt, A. Gotsman, H. Yang, and M. Zawirski. Replicated Data Types: Specification, Verification, Optimality. In POPL, pages 271–284, 2014.
A. Cerone, A. Gotsman, and H. Yang. Parameterised Linearisability. In ICALP, pages
98–109, 2014.
S. Chakraborty, T. A. Henzinger, A. Sezgin, and V. Vafeiadis. Aspect-Oriented Linearizability Proofs. Logical Methods in Computer Science, 11(1:20):1–33, 2015.
XX:13
XX:14
Local Linearizability
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
POPL 2015 Artifact Evaluation Committee. POPL 2015 Artifact Evaluation. Accessed on
01/14/2015. URL: http://popl15-aec.cs.umass.edu/home/.
Computational Systems Group, University of Salzburg. Scal: High-Performance MulticoreScalable Computing. URL: http://scal.cs.uni-salzburg.at.
J. Derrick, B. Dongol, G. Schellhorn, B. Tofan, O. Travkin, and H. Wehrheim. Quiescent
Consistency: Defining and Verifying Relaxed Linearizability. In FM, pages 200–214, 2014.
M. Dodds, A. Haas, and C.M. Kirsch. A Scalable, Correct Time-Stamped Stack. In POPL,
pages 233–246, 2015.
I. Filipovic, P.W. O’Hearn, N. Rinetzky, and H. Yang. Abstraction for concurrent objects.
Theor. Comput. Sci., 411(51-52):4379–4398, 2010.
J.R. Goodman. Cache consistency and sequential consistency. University of WisconsinMadison, Computer Sciences Department, 1991.
A. Haas, T.A. Henzinger, A. Holzer, C.M. Kirsch, M. Lippautz, H. Payer, A. Sezgin,
A. Sokolova, and H. Veith. Local Linearizability for Concurrent Container-Type Data
Structures. In CONCUR, 2016.
A. Haas, T.A. Henzinger, C.M. Kirsch, M. Lippautz, H. Payer, A. Sezgin, and A. Sokolova.
Distributed Queues in Shared Memory: Multicore Performance and Scalability through
Quantitative Relaxation. In CF, 2013.
A. Haas, T. Hütter, C.M. Kirsch, M. Lippautz, M. Preishuber, and A. Sokolova. Scal: A
Benchmarking Suite for Concurrent Data Structures. In NETYS, pages 1–14, 2015.
A. Heddaya and H. Sinha. Coherence, Non-coherence and Local Consistency in Distributed
Shared Memory for Parallel Computing. Technical report, Computer Science Department,
Boston University, 1992.
S. Heller, M. Herlihy, V. Luchangco, M. Moir, W.N. Scherer, and N. Shavit. A Lazy
Concurrent List-based Set Algorithm. In OPODIS, 2005.
J.L. Hennessy and D.A. Patterson. Computer Architecture, Fifth Edition: A Quantitative
Approach. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 5th edition, 2011.
T.A. Henzinger, C.M. Kirsch, H. Payer, A. Sezgin, and A. Sokolova. Quantitative relaxation
of concurrent data structures. In POPL, pages 317–328, 2013.
T.A. Henzinger, A. Sezgin, and V. Vafeiadis. Aspect-Oriented Linearizability Proofs. In
CONCUR, pages 242–256, 2013.
M. Herlihy and N. Shavit. The Art of Multiprocessor Programming. Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA, 2008.
M. Herlihy and J.M. Wing. Linearizability: A Correctness Condition for Concurrent Objects. ACM Trans. Program. Lang. Syst., 12(3):463–492, 1990.
R. Jagadeesan and J. Riely. Between Linearizability and Quiescent Consistency - Quantitative Quiescent Consistency. In ICALP, pages 220–231, 2014.
C.M. Kirsch, M. Lippautz, and H. Payer. Fast and Scalable, Lock-free k-FIFO Queues. In
PaCT, pages 208–223, 2013.
A. Kogan and E. Petrank. Wait-free queues with multiple enqueuers and dequeuers. In
PPoPP, pages 223–234, 2011.
L. Lamport. How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs. IEEE Trans. Comput., 28(9):690–691, September 1979.
R.J. Lipton and J.S. Sandberg. PRAM: A Scalable Shared Memory. Technical Report Nr.
180, Princeton University, Department of Computer Science, 1988.
R.J. Lipton and J.S. Sandberg. Oblivious memory computer networking, September 28
1993. CA Patent 1,322,609.
M.M. Michael. Hazard Pointers: Safe Memory Reclamation for Lock-Free Objects. IEEE
Trans. Parallel Distrib. Syst., 15(6):491–504, 2004.
A. Haas et al.
34
35
36
37
38
39
40
41
42
M.M. Michael and M.L. Scott. Simple, fast, and practical non-blocking and blocking
concurrent queue algorithms. In PODC, pages 267–275, 1996.
M.M. Michael, M.T. Vechev, and V.A. Saraswat. Idempotent Work Stealing. In PPoPP,
pages 45–54, 2009.
A. Morrison and Y. Afek. Fast Concurrent Queues for x86 Processors. In PPoPP, pages
103–112, 2013.
Multicore Computing Group, Tel Aviv University. Fast Concurrent Queues for x86 Processors. Accessed on 01/28/2015. URL: http://mcg.cs.tau.ac.il/projects/lcrq/.
H. Rihani, P. Sanders, and R. Dementiev. MultiQueues: Simpler, Faster, and Better
Relaxed Concurrent Priority Queues. CoRR, 2014. arXiv:1411.1209.
A. Sezgin. Sequential Consistency and Concurrent Data Structures. CoRR, abs/1506.04910,
2015.
N. Shavit. Data Structures in the Multicore Age. CACM, 54(3):76–84, March 2011.
R.C. Steinke and J.G. Nutt. A Unified Theory of Shared Memory Consistency. J. ACM,
51(5):800–849, September 2004.
R.K. Treiber. Systems Programming: Coping with Parallelism. Technical Report RJ-5118,
IBM Research Center, 1986.
XX:15
XX:16
Local Linearizability
A
Local Linearizability with Shape Observers
There are two possible ways to deal with shape observers: treat them locally, in the threadinduced history of the performing thread, or treat them globally. While a local treatment
is immediate and natural to a local consistency condition, a global treatment requires care.
We present both solutions next.
I Definition 17 (Local Linearizability LSO). A history h is locally linearizable with local
shape observers (LSO) wrt a sequential specification S if it is locally linearizable according
to Definition 11 with the difference that the in-methods (Definition 9) also contain all shape
observers performed by thread T , i.e., IT = {m | m ∈ M (h|T ) ∩ (Ins ∪ SOb)}.
Global observations require more notation and auxiliary notions. Let sj for j ∈ J be
a collection of sequences over alphabet Σ with pairwise disjoint sets of symbols M (sj ). A
S
sequence s is an interleaving of sj for j ∈ J if M (s) = j M (sj ) and s|M (sj ) = sj for all
Q
j ∈ J. We write j sj for the set of all interleavings of sj with j ∈ J.
Given a history h and a method call m ∈ h, we write h≤m for the (incomplete) history
that is the prefix of h up to and without mr , the response event of m. Hence, h≤m contains
all invocation and response events of h that appear before mr .
I Definition 18. Let S denote the sequential specification of a container D. A shape
observer m in a history h has a witness if there exists a sequence s ∈ Σ∗ such that sm ∈ S
Q
and s ∈ T sT for some sT that is a linearization of the thread-induced history (h≤m )T .
Informally, the above definition states that a global shape observer m must be justified
by a (global) witness. Such a global witness is a sequence that (1) when extended by m
belongs to the sequential specification, and (2) is an interleaving of linearizations of the
thread-induced histories up to m.
I Definition 19 (Local Linearizability GSO). A history h is locally linearizable with global
shape observers (GSO) wrt a sequential specification S if it is locally linearizable and each
shape observer m ∈ SOb has a witness.
We illustrate the difference in the local vs. the global approach for shape observers with
the following example.
I Example 20. Consider the following queue history with global observer size()
T1
T2
enq(1)
enq(2)
deq(1)
size(n)
where n is just a placeholder for a concrete natural number. For n = 0, the history h is
locally linearizable LSO, but not locally linearizable GSO. For n = 1, the history h is locally
linearizable GSO, but not locally linearizable LSO.
Global observers and non-disjoint operations are expected to have negative impact on
performance. If one cares for global consistency, local linearizability is not the consistency
condition to be used. The restriction to containers and disjoint operations specifies, in an
informal way, the minimal requirements for local consistency to be acceptable.
Neither sets nor maps are containers according to our definition. However, it is possible to
extend our treatment to sets and maps similar to our treatment of global observers. Locally
A. Haas et al.
XX:17
linearizable sets and maps will be weaker than their linearizable counterparts, but, due to
the tight coupling between mutator and observer effects, the gain in performance is unlikely
to be as substantial as the one observed in other data structures. The technicalities needed
to extend local linearizability to sets and maps would complicate the theoretical development
without considerable benefits and we, therefore, excluded such data structures.
B
Additional Results and Proofs
Theorem 12 (Compositionality). A history h over a set of objects O with sequential
specifications Sq for q ∈ O is locally linearizable if and only if h|q is locally linearizable with
respect to Sq for every q ∈ O.
Proof. The property follows from the compositionality of linearizability and the fact that
(h|q)T = hT |q for every thread T and object q. Assume that h over O is locally linearizable. This means that all thread-induced histories hT over O are linearizable. Hence, since
linearizability is compositional, for each object q ∈ O the history hT |q is linearizable with
respect to Sq . Now from (h|q)T = hT |q we have that for every object q the history (h|q)T
is linearizable for every thread T .
Similarly, assume that for every object q ∈ O the history h|q is locally linearizable. Then,
for every q, (h|q)T = hT |q is linearizable for every thread T . From the compositionality of
linearizability, hT is linearizable for every thread T . This proves that h is locally linearizable.
J
Proposition 2 (Lin vs. LocLin 2).
Linearizability implies local linearizability for
sequential specifications that are closed under data-projection.
Proof. Assume we are given a history h which is linearizable with respect to a sequential
specification S that is closed under data-projection. Further assume that, without loss of
generality, h is complete. Then there exists a sequential history s ∈ S such that (1) s is a
permutation of h, and (2) if m <h n, then also m <s n. Given a thread T , consider the
thread-induced history hT and let sT = s| (IT ∪ OT ). Then, sT is a permutation of hT since
hT and sT consist of the same events. Furthermore, sT ∈ S since S is closed under dataprojection and since Equation (1) holds for containers. Finally, we have for each m ∈ hT
and n ∈ hT that, if m <hT n, then also m <sT n since m <h n and therefore m <s n which
implies m <sT n. Thereby, we have shown that hT is linearizable with respect to S, for an
arbitrary thread T . Hence h is locally linearizable with respect to S.
J
Proposition 3 (Data-Projection Closedness).
queue, and stack are closed under data-projection.
The sequential specifications of pool,
Proof. Let s ∈ SP , V 0 ⊆ V , and let
s0 = s| ({ins(x), rem(x) | x ∈ V 0 ∪ Emp}) .
Then, it suffices to check that all axioms for pool (Definition 2 and Table 1) hold for s0 .
Clearly, all methods in s0 appear at most once, as they do so in s. If rem(x) ∈ s0 , then
rem(x) ∈ s and, since s ∈ SP , ins(x) ≺s rem(x). But then also rem(x) ∈ s0 and hence
ins(x) ≺s0 rem(x). Finally, if ins(x) ≺s0 rem(e) for e ∈ Emp, then ins(x) ≺s rem(e) implying
that rem(x) ∈ s and rem(x) ≺s rem(e). But then rem(x) ∈ s0 as well and rem(x) ≺s0 rem(e).
This shows that SP is closed under data-projection.
XX:18
Local Linearizability
Assume now that s ∈ SQ and s0 is as before (with enq() and deq() for ins() and rem(),
respectively). Then, as SP is closed under data-projection, s0 satisfies the pool axioms.
Moreover, the queue-order axiom (Definition 2 and Table 1) also holds: Assume enq(x) ≺s0
enq(y) and deq(y) ∈ s0 . Then enq(x) ≺s enq(y) and deq(y) ∈ s. Since s ∈ SQ we get
deq(x) ∈ s and deq(x) ≺s deq(y). But this means deq(x) ∈ s0 and deq(x) ≺s0 deq(y).
Hence, SQ is closed under data-projection.
Finally, if s ∈ SS and s0 is as before (with push() and pop() for ins() and rem(),
respectively), we need to check that the stack-order axiom (Definition 2 and Table 1) holds.
Assume push(x) ≺s0 push(y) ≺s0 pop(x). This implies push(x) ≺s push(y) ≺s pop(x)
and since s ∈ SS we get pop(y) ∈ s and pop(y) ≺s pop(x). But then pop(y) ∈ s0 and
pop(y) ≺s0 pop(x). So, SS is closed under data-projection.
J
Proposition 4 (LocLin Pool). Let h be a locally linearizable history wrt a pool. Then:
1. No value is duplicated, i.e., every remove method appears in h at most once.
2. There are no out-of-thin-air values, i.e.,
∀x ∈ V. rem(x) ∈ h ⇒ ins(x) ∈ h ∧ rem(x)6<h ins(x).
3. No value is lost, i.e., ∀x ∈ V. ∀e ∈ Emp. ins(x) <h rem(e) ⇒ rem(x) ∈ h ∧ rem(e)6<h rem(x)
and ∀x ∈ V. ∀e ∈ Emp. rem(e) <h rem(x) ⇒ ins(x) 6<h rem(e).
Proof. Note that if a history h is linearizable wrt a pool, then all of the three stated
properties hold, as a consequence of linearizability and the definition of SP . Now assume
that h is locally linearizable wrt a pool.
If rem(x) appears twice in h, then it also appears twice in some thread-induced history
hT contradicting that hT is linearizable with respect to a pool. This shows that no value is
duplicated.
If rem(x) ∈ h, then rem(x) ∈ hT for some T and, since hT is linearizable with respect to
a pool, ins(x) ∈ hT and rem(x)6<hT ins(x). This yields ins(x) ∈ h and rem(x)6<h ins(x).
Hence, there are no thin-air values.
Finally, if rem(e) ∈ h for e ∈ Emp then rem(e) ∈ hT for all T . Let ins(x) <h rem(e) and
let T 0 be such that ins(x) ∈ hT 0 . Then ins(x) <hT 0 rem(e) and since hT 0 is linearizable
with respect to a pool, rem(x) ∈ hT 0 and rem(e)6<hT 0 rem(x). This yields rem(x) ∈ h and
rem(e)6<h rem(x). Similarly, the other condition holds. Hence, no value is lost.
J
Theorem 25 (Queue Local Linearizability). A queue concurrent history h is locally
linearizable with respect to the queue sequential specification SQ if and only if
1. h is locally linearizable with respect to the pool sequential specification SP , and
2. ∀x, y ∈ V. ∀i. enq(x) <ih enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x).
Proof. Assume h is locally linearizable with respect to SQ . Since SQ ⊆ SP (with suitably
renamed method calls), h is locally linearizable with respect to SP . Moreover, since all hi
are linearizable with respect to SQ , by Theorem 24, for all i we have ∀x, y ∈ V. enq(x) <hi
enq(y) ∧ deq(y) ∈ hi ⇒ deq(x) ∈ hi ∧ deq(y) 6<hi deq(x).
Assume x, y ∈ V are such that enq(x) <ih enq(y) and deq(y) ∈ h. Then enq(x) <hi
enq(y) and deq(y) ∈ hi so deq(x) ∈ hi and deq(y) 6<hi deq(x). This implies deq(x) ∈ h
and deq(y) 6<h deq(x).
For the opposite, assume that conditions 1. and 2. hold for a history h. We need to
show that (1) hi form a decomposition of h, which is clear for a queue, and (2) each hi is
linearizable with respect to SQ .
A. Haas et al.
XX:19
By 1., each hi is linearizable with respect to a pool. Assume enq(x) <hi enq(y) and
deq(y) ∈ hi . Then enq(x) <ih enq(y) ∧ deq(y) ∈ h and hence by 2., deq(x) ∈ h ∧
deq(y) 6<h deq(x). Again, as enq(x), deq(x) ∈ hi we get deq(x) ∈ hi ∧ deq(y) 6<hi deq(x).
According to Theorem 24 this is enough to conclude that each hi is linearizable with respect
to SQ .
J
Theorem 15 (Pool, Queue, & Stack, SC). For pool, queue, and stack, local linearizability is incomparable to sequential consistency.
Proof. The following histories, when instantiating i() with ins(), enq(), and push(), respectively, and instantiating r() with rem(), deq(), and pop(), respectively, are sequentially
consistent but not locally linearizable wrt pool, queue and stack:
(a) Pool:
T1
i(1)
r(1)
r(empty)
T2
(b) Queue:
r(2)
T1
T2
i(1)
i(2)
i(1)
i(2)
r(1)
(c) Stack:
T1
r(1)
T2
r(2)
History (a) is already not locally linearizable wrt pool, queue, and stack, respectively,
histories (b) and (c) provide interesting examples. The history in Figure 2 is locally linearizable but not sequentially consistent wrt a pool. The following histories are locally
linearizable but not sequentially consistent wrt a queue and a stack, respectively:
(d) Queue:
T1
T2
i(1)
i(2)
i(3)
r(1)
r(2)
i(4)
r(4)
r(3)
The two thread-induced histories i(1)i(2)i(3)r(1)r(2)r(3) and i(4)r(4) are both linearizable with respect to a queue. However, the overall history has no sequential witness and is
therefore not sequentially consistent: To maintain the queue behavior, the order of operations r(1) and r(2) cannot be changed. However, this implies that the value 3 instead of
the value 4 would have to be removed directly after i(4).
(e) Stack:
XX:20
Local Linearizability
qi = ins(xi,1 )rem(xi,1 ) . . . ins(xi,p )rem(xi,p )rem(xi,p+1 )rem(xi,q )rem(empty)ri ins(xi,q+1 ) . . . ins(xi,m ).
Figure 6 Sequential history qi .
T1
i(1)
i(2)
r(2)
T2
i(3)
r(1)
r(3)
The two thread-induced histories i(1)i(2)r(2)r(1) and i(3)r(3) are both linearizable with
respect to a stack. The operations i(2) and r(2) prevent the reordering of operations
i(1) and i(3). Therefore, the overall history has no sequential witness and hence it is not
sequentially consistent.
J
Proposition 5 (Pool, QC). Let h be a pool history in which no data is duplicated, no
thin-air values are returned, and no data is lost, i.e., h satisfies 1.-3. of Proposition 4.
Then h is quiescently consistent.
Proof. Assume h is a pool history that satisfies 1.-3. of Proposition 4. Let h1 , . . . , hn
be histories that form a sequential decomposition of h. That is h = h1 · · · hn and the
only quiescent states in any hi are at the beginning and at the end of it. Note that this
decomposition has nothing to do with a thread-local decomposition. Let Mi = Mhi be the
set of methods of hi , for i ∈ {1, . . . , n}. Note that the sanity conditions 1.-3. ensure that
none of the following two situations can happen:
rem(x) ∈ Mi , ins(x) ∈ Mj , j > i,
ins(x) ∈ Mi , rem(empty) ∈ Mj , rem(x) ∈ Mk , k > j > i,
Let Vi = {xi,1 , . . . , xi,m } denote the set of values in Mi ordered in a way that there is a p
and q such that
ins(xi,j ), rem(xi,j ) ∈ Mi for j ≤ p;
rem(xi,j ) ∈ Mi for j > p, j ≤ q; and
ins(xi,j ) ∈ Mi for j > q.
Moreover, let ri be the number of occurrences of rem(empty) in hi .
We now construct a sequential history for h, which has the form q = q1 · · · qn where
each sequential history qi is a permutation of Mi shown in Figure 6. Using the observations
above, it is easy to check that q is indeed a quiescent witness for h.
J
Theorem 16 (Pool, Queue, & Stack, QC). For pool, local linearizability is stronger than
quiescent consistency. For queue and stack, local linearizability is incomparable to quiescent
consistency.
Proof. The following histories are quiescently consistent but not locally linearizable wrt
pool, queue, and stack, respectively:
(a) Pool:
rem(empty)
T1
T2
ins(1)
rem(empty)
rem(1)
A. Haas et al.
XX:21
(b) Queue:
enq(1)
T1
T2
enq(2)
enq(3)
deq(3)
deq(2)
(c) Stack:
push(1)
T1
T2
push(2) push(3)
pop(2)
pop(3)
In all three histories, the only quiescent states are before and after the longest operation.
Therefore, all operations in thread T2 can be reordered arbitrarily, in particular in a way
such that they satisfy the sequential specification of the respective concurrent data structure. However, each of the thread-induced histories for thread T2 are not linearizable with
respect to pool, queue, and stack, respectively. Therefore, none of these histories is locally
linearizable. Also here history (a) suffices.
On the other hand, the following histories are not quiescently consistent but locally
linearizable wrt queue, and stack, respectively:
(d) Queue:
T1
enq(1)
deq(2)
enq(2)
T2
deq(1)
(e) Stack:
T1
T2
push(1)
pop(1)
push(2)
pop(2)
In histories (d) and (e), between each two operations, the concurrent data structure
is in a quiescent state. Therefore, none of the operations can be reordered and, hence, no
sequential witness exists. However, all thread-induced histories are linearable and, therefore,
the overall histories are locally linearizable. In particular, on a history where each pair of
operations is separated by a quiescent state, i.e., there is no overlap of operations, a quiescent
consistent data structure behaves as it would be linearizable with respect to its sequential
specification and we see the same semantic differences to local linearizability as we see
between linearizability and local linearizability.
J
C
Case Study: Work Stealing Queues
Consider a data structure D which admits two operation types: ins(x), which inserts the
element x into the container, and rem(), which returns and removes an element from the
container. Now imagine that the implementation uses a Work Stealing Queue (WSQ) [35].
Every thread T that uses D has its unique designated buffer QT in the WSQ. Whenever
thread T calls ins(x), x is appended to the tail of QT . When T calls rem(), WSQ first
XX:22
Local Linearizability
T1
T2
enq(3)
enq(2) enq(1)
deq(1) deq(2) deq(3)
Figure 7 History that is QQC but not LL.
checks whether QT is non-empty; if it is, then it returns the element at the tail of QT (LIFO
semantics) and removes it. Otherwise, it chooses some other QT 0 and tries to return an
element from that buffer. But any time a different thread’s buffer is checked, the element to
be removed is taken from the head (FIFO semantics). If T and T 0 are both trying to access
the same buffer at the same time, then usual synchronization measures are taken to ensure
that exactly one thread removes one element.
Given this implementation, the developer of D wants to write a specification for the
potential users of D. Since D is essentially a collection of deques, the developer is tempted
to state that D is a deque with a particular consistency condition. However, D is not a
linearizable deque because ins(x) by T followed by ins(y) by T 0 followed by rem() returns
either x or y depending on whether T or T 0 calls it; i.e. rem() has ambiguous semantics.
D can be seen as a sequentially consistent (SC) deque but then D does not allow many
behaviors that an SC deque would allow; i.e. SC does not capture the behaviors of D
tightly. Relaxed sequential specifications will not work either since D does converge to
sequential semantics (of a LIFO stack) when a single thread uses it. In short, the developer
will fail to capture the semantics of D in a satisfactory manner.
D on the other hand is a locally linearizable deque in which rem() by T from QT 0 is
treated as FIFO removal whenever T 6= T 0 and as LIFO removal whenever T = T 0 . In other
words, local linearizability provides a succinct and clean representation of a well-known
implementation framework (WSQ) hiding away implementation details. Compare this with
the fact that even though WSQ has a queue in it, to argue its correctness it is proved to be
a linearizable pool even though it has stronger semantics than a pool; i.e. linearizable pool
semantics is too weak for D. Observe also that since what we have described in the example
is essentially providing the illusion of using a monolithic structure which is implemented
in terms of distributed components (shared memory is typically implemented on message
passing), we expect local linearizability to be widely applicable.
D
Quiescent Consistency & Quantitative Quiescent Consistency
Without going into the details of the definition of quantitative quiescent consistency we give
a history in Figure 7 that is quantitatively quiescently consistent but not locally linearizable
wrt a queue. Quantitative quiescent consistency allows to reorder the two insert-operations
in thread T2 and thereby violates local linearizability.
E
Consistency Conditions for Distributed Shared Memory
Decomposition per
LL
thread
Coherence
PRAM
#SHs
Write-Operations Ih (i)
Read-Operations Oh (i)
CCfSH
LoD
n
{ins(v) ∈ h|Ti | v ∈ V }
{head(vinit ) ∈ h}
∪ {head(v) ∈ h | ins(v) ∈ Ih (i)}
Lin.
no
memory location
k
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h | v ∈ V }
SC
yes
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SC
yes
a
yes
PC
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SC
CC
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SCb
yes
{head(v) ∈ h|Ti | v ∈ V }
c
yes
LC
thread & memory location
n·k
{ins(v) ∈ h|Ti | v ∈ V }
∪ {ins(v) ∈ h | head(v) ∈ h|Ti }
SC
A. Haas et al.
Consistency
Condition
SCa : SC and ins-operations are in the same order for each witness.
SCb : SC and ins-operations are ordered by the transitive closure of the thread program orders and write-read pairs.
SCc : SC and ins-operations from threads other than Ti can be reordered even if they are from the same thread and only logical contradictions
in the local history are considered for consistency.
n: number of threads, k: number of memory locations,
#SHs: number of subhistories, CCfSH: consistency condition for subhistories, LoD: loss of data
Table 2 Comparison of consistency conditions for a single distributed shared memory location, i.e., k = 1
XX:23
XX:24
Local Linearizability
In Table 2 we compare local linearizability (LL) against the consistency conditions coherence [3], pipelined RAM (PRAM) consistency [31, 32, 41, 4], processor consistency (PC) [3,
16], causal consistency (CC) [4], and local consistency (LC) [20]. Local linearizability shares
with all these consistency conditions the idea of decomposing a concurrent history into
several subhistories.
Coherence projects a concurrent history to the operations on a single memory location
and each resulting history has to be sequentially consistent. Since sequential consistency is
not compositional, coherence does not imply sequential consistency for the overall history [3]
whereas local linearizability for each single memory location implies local linearizability for
the overall history.
In contrast to coherence and local consistency, local linearizability, PRAM consistency,
PC, and CC all decompose the history into per-thread subhistories, i.e., if there are n threads
then these conditions consider n subhistories and need n sequential witnesses. Coherence
requires one witness per memory location and local consistency requires one witness per
thread and memory location.
For determining the subhistory for a thread Ti , coherence, PRAM consistency, PC, and
CC consider all write-operations in a given history, i.e., Ih (i) = {ins(v) ∈ h | v ∈ V }. In
contrast, local linearizability only considers the write-operations in thread Ti , i.e., Ih (i) =
{ins(v) ∈ h|Ti | v ∈ V } and local consistency considers all write-operations in thread Ti as
well as all write-operations whose values are read in thread Ti , i.e., Ih (i) = {ins(v) ∈ h|Ti |
v ∈ V } ∪ {ins(v) ∈ h | head(v) ∈ h|Ti }. Regarding read-operations, PRAM consistency,
PC, CC, and LC consider only the read-operations in thread Ti . Coherence considers all
read-operations in a given history and local linearizability only considers read-operations
that read the initial value vinit and read-operations that read values that were written by a
write-operation in thread Ti . Reading the initial value is analogous to returning empty in a
data structure.
Local linearizability requires that each subhistory, i.e., thread-induced history, is linearizable with respect to the sequential specification under consideration. In contrast, coherence,
PRAM consistency, PC, CC, and LC require that each subhistory is sequentially consistent
(or a variant thereof) with respect to the sequential specification. However, the variants
of sequential consistency that are used by these consistency conditions are vulnerable to
a loss of data as discussed in Section 5 and, therefore, make these consistency conditions
unsuitable for concurrent data structures.
When considering PRAM consistency, the sequentialization of the write-operations of different threads might be observed differently by different threads, e.g., a thread T1 might observe all write operations of thread T2 before the write operations of thread T3 but a thread T4
might observe all write operations of T3 before the write operations of T2 . In contrast, threadinduced histories as defined by local linearizability do not involve write-operations from other
threads but involve (some) read-operations performed by other threads. Like PRAM consistency, processor consistency requires for each thread Ti that the read- and write-operations
performed by Ti are seen in Ti ’s program order and that the write-operations performed by
other threads are seen in their respective program order. Furthermore, processor consistency
also requires that two write-operations to the same memory location appear in the same order in each sequential witness of each thread even if they are from different threads [3, 16].
This additional condition makes processor consistency strictly stronger than PRAM consistency [3]. This condition also creates a similar effect as the consideration of read-operations
in different threads when forming the thread-induced history in local linearizability. Causal
consistency considers a causal order instead of the thread program orders alone. Like local
A. Haas et al.
XX:25
` = 4, v = 4
0
`−1
0
backend
` = 5, v = 5
`−1
s:
s:
n:
` = 4, v = 4
max
`−1
0
s:
ni :
ni :
(b) Add new node ni
(c) Adjust `, then v
alive
(a) Initial state
Figure 8 Segment modifications throughout announce_thread().
linearizability, causal consistency matches write-read pairs across different threads. In particular, the causal order is the transitive closure of the thread program orders and write-read
pairs. By considering the causal order, writes from different threads can become ordered
which is not the case for local linearizability.
F
LLD and LL+ D Implementation Details
As already mentioned, each thread inserts elements into a local backend and removes elements either from its local backend (preferred) or from other backends (fall-back) accessed
through a single segment (thread-indexed array), effectively managing single-producer/multipleconsumer backends for a varying number of threads.
The segment is dynamic in length (with a predefined maximum). A slot in this segment
refers to a node that consists of a backend and a flag indicating whether the corresponding thread is alive or has terminated. Similar to other work [2, 21] the flag is used for
logically removing the node from the segment (it stays in the segment until its backend is
empty). Additionally, a (global) version number keeps track of all changes in the segment.
The algorithm is divided into two parts: (1) maintaining the segment, and (2) adding and
removing elements to backends.
In the following we refer to the segment as s, a thread’s Ti local node as ni , the version
number of the segment as v and the current length of the segment as `. The range of indices
r is then defined as 0 ≤ r < `.
For maintaining the segment we provide two methods announce_thread() and
cleanup_thread(node) that are used to add and remove nodes to the segment. Upon
removal of a node the segment is also compacted, i.e., the hole that is created by removing
a node pointer is filled with the last node pointer in the segment. As nodes are added and
removed the length of the segment ` and thus the range of valid indices i of the segment,
0 ≤ i < `, is updated. All changes to the segment involve incrementing the version number.
More detailed, the operations for maintaining the segment and compacting it as nodes
are cleaned up are:
announce_thread(): Allocates a node for the thread as follows: searches for an existing
node of a terminated thread and reuses it if it finds one; otherwise it creates a new node,
adds the node to s, and adjusts `. In both cases it then increments v and returns the
node. The creation of new node is illustrated in Figure 8.
cleanup_thread(Node n): Searches for the node n in s using linear search. If it finds n
at slot j, it copies the pointer of s[` − 1] to s[j], decrements `, increments v, and resets
s[`] to null using the new `. If n is not found, then a concurrent thread has already
performed the cleanup and the operation just returns. Figure 9 illustrates an example
XX:26
Local Linearizability
` = 5, v = 5
0
` = 5, v = 5
`−1
s:
` = 4, v = 6
`−1
0
0
s:
`−1
s:
backend
n:
dead
n:
backend
m:
m:
dead
(a) n at s[0] is empty and
dead
(b) Write s[` − 1] into s[0]
Adjust ` then v
Figure 9 Segment modifications throughout cleanup_thread(Node n).
where initially ` = 5, the thread owning the node at s[0] is dead and the corresponding
backend is empty.
Note that updating the segment state is only needed when threads are joining or when
backends of terminated threads become empty. We consider both scenarios as infrequent and
implement the corresponding operations using locks. Alternatively those operations can be
implemented using helping approaches, similar to wait-free algorithms [29]. Also note that
although operations on segments are protected by locks, partial changes can be observed,
e.g., a remove operation (as defined below) can observe a segment in an intermediate state
with two pointers pointing to a node during cleanup. The invariant is that no change can
destroy the integrity of the segment within the valid range, i.e., all slots within the range
either point to a valid node or nothing (null).
The actual algorithm for adding and removing elements is then defined as follows:
ins(): Upon first insertion, a thread Ti gets assigned a node ni (containing backend bi )
using announce_thread(). The element is then inserted into bi . Subsequent insertions
from this thread will use ni throughout the lifetime of the thread.
rem(): The remove operation consists of two parts: (a) finding and removing an element
and (b) cleaning up nodes of terminated threads. For (a) a thread Ti tries to get an
element from its own backend in ni . If ni does not exist (because the thread has not
yet performed a single ins() operation) or the corresponding backend is empty, then a
different node n is selected randomly within the valid range. If the backend contained
in n is empty, the operation scans all other nodes’ backends in linear fashion. However,
if the version number changed during the round of scanning through all backends, the
operation is restarted immediately. Note that since ` is dynamic a remove operation may
operate on a range that is no longer valid. Checking the version number ensures that the
operation is restarted in such a case. For (b) a thread calls cleanup_thread(n) upon
encountering a node n that has its alive-flag set to false (dead) and contains an empty
backend. A cleanup also triggers a restart of the remove operation.
terminate(): Upon termination a thread Ti changes the alive flag of ni to false (dead).
Dynamic memory used for nodes is susceptible to the ABA problem and requires proper
handling to free memory. Our implementations use 16-bit ABA counters to avoid the ABA
problem and refrain from freeing memory. Hazard pointers [33] can be used for solving the
ABA problem as well as for freeing memory.
F.1
LL+D: LLD with Linearizable Emptiness Check
We call a data structure implementation Φ stateful if the remove methods of Φ can be
modified to return a so-called state that changes upon an insert or a remove of an element,
A. Haas et al.
but does not change between two removes that return empty unless an element has been
inserted in the data structure in the meantime. For stateful implementations Φ we can create
the locally linearizable version with linearizable emptiness check LL+ D Φ. Michael-Scott
queue [34] and Treiber stack [42] are stateful implementations, whereas LCRQ [36] is not.
Also TS stack [14], and k-FIFO [28] and k-Stack [23] are stateful implementations, but the
notion of a state in these data structures is huge making it unsuitable for LL+ D.
For LL+ D implementations, linearizable emptiness checks are achieved via an atomic
snapshot [25], just like for DQs. A detailed description of the LLD and LL+ D implementations, as well as the pseudo code, can be found in the appendix. Here, we only present the
results of the experimental performance evaluation.
F.2
Correctness of LL+ D
I Proposition 7 (LLD and LL+ D). Let Φ be a stateful data structure implementation that is
linearizable with respect to a sequential specification SΦ . Then LL+ D Φ is linearizable with
respect to a pool.
Proof. Proving that LL+ D Φ is linearizable with respect to pool, in particular that it has
a linearizable emptiness check, follows the proof for DQ in general, see [18]: The emptiness
check is performed by creating an atomic snapshot [25] of the states of all backends (stored
in the states array) using the first loop (lines 28-43). If the atomic snapshot is valid
(checked via the second loop, lines 46-52, in particular line 48) and all backends are empty
in this atomic snapshot, then there existed a point in time during the creation of the atomic
snapshot where all backends were indeed empty.
Notice that since the segment is dynamic in length it can happen that some backends
are not contained in the atomic snapshot. To guarantee that no elements are missed in the
emptiness check the atomic snapshot is extended by the version number v of the segment. If
a new backend is added to the segment during the generation of the atomic snapshot, then
the version number is increased and the atomic snapshot becomes invalid (line 45).
The linearization point of the remove operation that returns empty is inbetween the two
loops (the last remove attempt of the first loop) if the version check and second loop go
through.
J
F.3
LLD Pseudo Code
All implementations use the interfaces depicted in Listing 1. For simplicity, the interface
only mentions pool, queue, and stack. The highlighted code refers to linearizable emptiness
check, i.e., it is only part of the LL+ D implementations: Methods retrieving elements (e.g.
rem) are assumed (or modified when possible) to also return a State object that uniquely
identifies the state of the data structure with respect to methods inserting elements (e.g.
ins). The same state can be accessed via the get_state() observer method.
Listing 2 illustrates the pseudo-code for maintaining the segment. The backend on line 2
can either be declared as Stack or Queue as defined in Listing 1 (or any other linearizable
data structure).
Listing 3 shows the pseudo-code for LL+ D. When removing the highlighted code, we
obtain the code for LLD. Each thread maintains its own backend, enclosed in a thread-local
node (line 3), for insertion. The local backend is always accessed through get_local_node
(line 5). This method also makes sure that a thread is announced (line 7) upon first insertion and acquires a node. An ins() operation then always uses a thread’s local backend
XX:27
XX:28
Local Linearizability
1
2
3
4
5
Pool {
< Element , State > rem () ;
v o i d ins ( Element e ) ;
State get_state();
}
6
7
8
9
10
11
12
Queue : Pool {
< Element , State > dequeue () ;
v o i d enqueue ( Element e ) ;
v o i d ins ( Element e ) = > enqueue ( e ) ;
< Element , State > rem () = > dequeue () ;
}
13
14
15
16
17
18
19
Stack : Pool {
< Element , State > pop () ;
v o i d push ( Element e ) ;
v o i d ins ( Element e ) = > push ( e ) ;
< Element , State > rem () = > pop () ;
}
Listing 1 Pool, queue, and stack interfaces
(line 13 and 14) for insertion. For removing an element in rem(), a thread tries to remove an
element from its local backend first (line 19-23). If no element can be found, all backends
in the valid range are searched in a linear fashion, starting from a random index. The
highlighted code (lines 46-52) illustrates checking the atomic snapshot for LL+ D.
F.4
LLD with Observer Methods
We have implemented LLD variants of (strict and relaxed) queue and stack implementations.
None of our LLD implementations involves observer methods, but the LLD algorithm can
easily be extended to support observer methods:
A data observer on LLD Φ (independently of which thread performs it) amounts to a
data observer on any ΦT .
A local shape observer on LLD Φ performed by thread T executes the shape observer
on ΦT .
A global shape observer on LLD Φ executes the shape observer on each backend ΦT and
produces an aggregate value.
G
Additional Implementations
We now present and evaluate additional algorithms that provide locally linearizable variants
of queues and stacks, obtained by modifying relaxed k-out of order queues and stacks [23,
28] in a way that makes them sequentially correct. We have also tried another generic
implementation, related to the construction in [9], that implements a flat-combining wrapper
with sequential (to be precise, single-producer multiple-consumer) backends. In our initial
experiments the performance of such an implementation was not particularly promising.
G.1
Locally Linearizable k-FIFO Queue and k-Stack
k-FIFO queues [28] and k-Stacks [23] are relaxed queues and stacks based on lists of segments
where each segment holds k slots for elements, effectively allowing reorderings of elements of
up to k − 1. The list of segments is implemented by a variant of Michael-Scott queue [34] for
A. Haas et al.
1
2
3
4
Node {
Pool backend ;
Bool alive ;
}
XX:29
// Any linearizable data structure .
5
6
7
8
9
Segment {
Node nodes [ MAX_THREADS ];
Int l = 0;
Int version = 0;
10
// Returns all indexes between 0 and l ( exclusive ) in random order .
[ Int ] range () ;
11
12
13
// Announces a node in the buffer , effectively adding it to nodes_ ,
// adjusting l , and changing the version .
Node a nn ou nc e _t hr ea d () {
segment_lock () ; // Protecting against concurrent announce or cleanup operations .
Node n = find_ dead_nod e () ;
i f n == null {
n = Node ( b : Backend () ) ;
nodes [ l ] = n ;
l ++;
}
n . alive = true ;
version ++;
segme nt_unloc k () ;
r e t u r n n;
}
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
// Removes a node from the buffer , effectively removing it from nodes_ ,
// adjusting l , and changing the version .
v o i d clea nup_thre ad ( Node n , Int old_version ) {
segment_lock () ; // Protecting against concurrent announce or cleanup operations .
<j , error > = f i n d _ n o d e _ i n _ s e g m e n t ( n ) ;
i f error || n . alive || old_version != version {
segm ent_unlo ck () ;
return ;
}
nodes [ j ] = nodes [l -1];
l - -;
version ++;
nodes [ l ] = null ;
segme nt_unloc k () ;
}
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
}
Listing 2 Node and segment structure for LLD and LL+ D (queue or stack)
k-FIFO and a variant of Treiber stack [42] for k-Stack. Insert and remove methods operate
on the segments ignoring any order of elements within the same segment. Segments used
for insertion and removal are identified by insertion and removal pointers, respectively.
For queues, elements are removed from the oldest segment and inserted into the mostrecent not-full segment. Upon trying to remove an element from an empty segment the
segment is removed and the removal pointer advanced to the next segment. Upon trying to
insert an element into a full segment a new segment is appended and the insertion pointer
is advanced to this new segment. Similarly (but different) for a stack, removal and insertion
operate on the most-recent segment, i.e., removal and insertion pointer are synonyms and
identify the same segment at all times. Again, upon trying to remove an element from
an empty segment the segment is removed and the removal pointer advanced to the next
segment. Upon trying to insert an element into a full segment a new segment is prepended
and the insertion pointer is set to this new segment.
k-FIFO queues and k-Stacks are relaxed queues and stacks that are: (1) linearizable with
respect to k-out-of-order queue and stack [23], respectively; (2) linearizable with respect to
a pool [23, 28]; (3) not locally linearizable with respect to queue and stack, respectively, for
k ≥ 1 since reordering elements that are inserted in the same segment (even sequentially by
XX:30
Local Linearizability
1
2
3
DynamicLocallyLinearizableDQ {
Segment s ;
thread_local Node local_node ;
4
Node get_loca l_node ( Bool c r e a t e _ i f_ a b s e n t ) {
i f ( c r e a t e _ i f _ a b s e n t ) && ( local_node == null ) {
local_node = s . an n ou nc e_ t hr ea d () ;
}
r e t u r n local_node ;
}
5
6
7
8
9
10
11
v o i d ins ( Element e ) {
n = get_loca l_node ( cr e a t e _ i f _ a b s e n t : true ) ;
n . backend . ins ( e ) ;
}
12
13
14
15
16
Element rem () {
// Fast path of retrieving an element from the thread - local backend .
n = get_loca l_node ( cr e a t e _ i f _ a b s e n t : false ) ;
i f n != null {
<e , state > = n . backend . rem () ;
i f e != null { r e t u r n e ; }
}
w h i l e true {
retry = false ;
old_version = s . version ;
range = s . range () ;
f o r i in range {
n = s . nodes [ i ];
i f old_version != s . version {
retry = true ; b r e a k ; }
Bool alive = n . alive ;
<e , state > = n . backend . rem () ;
i f e == null {
states[i] = state;
i f ! alive {
s . c leanup_ thread (n , old_version ) ;
retry = true ; b r e a k ;
}
} else {
r e t u r n e;
}
}
i f retry { c o n t i n u e ; }
i f old_version != s . version { c o n t i n u e ; }
f o r i in range {
n = s . nodes [ i ];
i f n == null || n . backend . get_state () != states [ i ] {
retry = true ;
break ;
}
}
i f retry { c o n t i n u e ; }
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
r e t u r n null ; // Empty case .
55
}
56
}
57
58
// Called upon thread termination .
v o i d terminate () {
n = get_loca l_node ( c r e a t e _ i f _ a b s e n t : false ) ;
i f n != null { n . alive = false ; }
}
59
60
61
62
63
64
}
Listing 3 LLD and LL+ D (queue and stack)
a single thread) is allowed, see the histories (b) and (c) in the proof of Theorem 15; and (4)
not sequentially consistent with respect to queue and stack, as shown by the histories (d)
and (e) in the proof of Theorem 15 that are k-FIFO and k-Stack histories, respectively, for
k ≥ 1.
A. Haas et al.
XX:31
We now present LL k-FIFO and LL k-Stack, modifications of k-FIFO and k-Stack, that
enforce local linearizability by ensuring that no thread inserts more than once in a single
segment. Assuming that segments are unique (by tagging pointers), LL k-FIFO remembers
the last used insertion pointer per thread. For LL k-Stack the situation is more subtle
as (due to the stack semantics) segments can be reached multiple times for insertion and
removal. Figure 10 illustrates an example where the top segment of a k-Stack is reached
multiple times by the same thread (T1 ). Since in the general case all segments could be
reached multiple times by a single thread it is required to maintain the full history of each
thread’s insertions. Assuming the maximum number of threads is known in advance, a
bitmap is used to maintain the information in which segment a thread has already pushed a
value. One can similarly implement a locally linearizable version of the Segment Queue [2].
T1
ins(1)
ins(3)
ins(4)
ins(2)
T2
rem(3)
rem(2)
2
⊥
1
1
T1 segment state
insert/remove
⊥
⊥
2
⊥
2
⊥
4
⊥
1
1
3
1
1
⊥
Figure 10 LL k-Stack run (k = 2). T1 can only insert in uncolored segments and needs to
prepend a new segment (for insertion) otherwise.
G.1.1
k-FIFO Queue and LL k-FIFO Queue Pseudo Code.
Listing 4 shows the pseudo code for LL k-FIFO queue. Again we highlight the code we
added to the original pseudo code [28]. Similar to the locally linearizable k-Stack each
thread inserts at most one element into a segment. However, in the k-FIFO queue we do
not need flags in each segment to achieve this property. It is sufficient to remember the last
segment used for insertion for each thread (set_last_tail; line 15). For each enqueue the
algorithm checks whether the executing thread has already used this segment for enqueueing
an element (get_last_tail; line 5). If the segment has already been used, the thread tries
to append a new segment (effectively adding a new tail).
G.1.2
Correctness Proof of LL k-FIFO Queue.
Having Theorem 25, the proof of correctness of LL k-FIFO queue is easy.
I Theorem 21 (Correctness of LL k-FIFO). LL k-FIFO queue presented in Listing 5 is locally
linearizable.
Proof. Using Theorem 25, as a first proof obligation we have to show that any history h of
the LL k-FIFO queue is locally linearizable with respect to the pool sequential specification
SP . This proof is analogous to the proof that any history of the LL k-Stack is locally
linearizable with respect to the pool sequential specification SP , and is therefore postponed
until the corresponding LL k-Stack theorem.
What remains to show is that
∀x, y ∈ V. ∀i. enq(x) <ih enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x)
XX:32
Local Linearizability
1
2
3
4
5
6
7
LocallyLinearizableKFIFOQueue {
enqueue ( item ) :
w h i l e true :
tail_old = get_tail () ;
i f get_last_tail ( thread_id ) == tail_old :
advance_tail ( tail_old , k ) ;
c o n t i n u e ; // Restart while loop .
head_old = get_head () ;
item_old , index = f in d _e mp ty _ sl ot ( tail_old , k ) ;
i f tail_old == get_tail () :
i f item_old . value == EMPTY :
item_new = atomic_value ( item , item_old . version + 1) ;
i f CAS (& tail_old - > segment [ index ] , item_old , item_new ) :
i f committed ( tail_old , item_new , index ) :
8
9
10
11
12
13
14
set_last_tail ( thread_id , tail_old ) ;
15
r e t u r n true ;
else:
advance_tail ( tail_old , k ) ;
16
17
18
19
bool committed ( tail_old , item_new , index ) :
i f tail_old - > segment [ index ] != item_new :
r e t u r n true ;
head_current = get_head () ;
tail_current = get_tail () ;
item_empty = atomic_value ( EMPTY , item_new . version + 1) ;
i f i n _ q u e u e _ a f t e r _ h e a d ( tail_old , tail_current , head_current ) :
r e t u r n true ;
e l s e i f not_in_queue ( tail_old , tail_current , head_current ) :
i f ! CAS (& tail_old - > segment [ index ] , item_new , item_empty ) :
r e t u r n true ;
e l s e : // in queue at head
head_new = atomic_value ( head_current . value , head_current . version + 1) ;
i f CAS (& head , head_current , head_new ) :
r e t u r n true ;
i f ! CAS (& tail_old - > segment [ index ] , item_new , item_empty ) :
r e t u r n true ;
r e t u r n false ;
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
item dequeue () :
w h i l e true :
head_old = get_head () ;
item_old , index = find_item ( head_old , k ) ;
tail_old = get_tail () ;
i f head_old == get_head () :
i f item_old . value != EMPTY :
i f head_old . value == tail_old . value :
advance_tail ( tail_old , k ) ;
item_empty = atomic_value ( EMPTY , item_old . version + 1) ;
i f CAS (& head_old [ index ] , item_old , item_empty ) :
r e t u r n item_old . value ;
else:
i f head_old . value == tail_old . value && tail_old . value == get_tail () :
r e t u r n null ;
advance_head ( head_old , k ) ;
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
}
Listing 4 Locally Linearizable k-FIFO Queue
Assume enq(x) <ih enq(y). This means that x and y were enqueued by the same thread
i and therefore inserted into different segments. Moreover, the segment of x is closer to
the head of the list than the segment of y. A deq(y) method call can remove y only if the
segment of y is the head segment. The segment of y can only become the head segment if
all segments closer to the head of the list get empty. This means that also the segment of x
has to become empty. Therefore there has to exist a deq(x) method call which removes x
from the segment, and deq(x) 6<h deq(y).
J
A. Haas et al.
G.1.3
k-Stack and LL k-Stack Pseudo Code.
Listing 5 shows the pseudo code for LL k-Stack. The highlighted code is the code we
added to the original pseudo code [23] to achieve local linearizability. The difference to
the original algorithm is that a thread inserts at most one element into a segment. To
achieve this property each segment in the k-stack contains a flag per thread which is set
when an element is inserted into the segment (mark_segment_as_used; line 14 and line 60).
If a thread encounters a segment where its flag is already set, the thread does not insert
its element into that segment but tries to prepend a new segment (is_segment_marked;
line 50). Otherwise the element is inserted into the existing segment and the flag of the
thread in that segment is set.
1
2
LocallyLinearizableKStack {
SegmentPtr top ;
3
4
5
6
v o i d init () :
new_ksegment = calloc ( s i z e o f ( ksegment ) ) ;
top = atomic_value ( new_ksegment , 0) ;
7
8
9
10
11
12
13
14
15
16
17
bool t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) :
i f top_old == top :
new_ksegment = calloc ( s i z e o f ( ksegment ) ) ;
new_ksegment - > next = top_old ;
new_ksegment - > s [0] = atomic_value ( item , 0) ; // Use first slot for item .
top_new = atomic_value ( new_ksegment , top_old . ver +1) ;
m a r k _ s e g m e n t _ a s _ u s e d ( top_new ) ;
i f CAS (& top , top_old , top_new ) :
r e t u r n true ;
r e t u r n false ;
18
19
20
21
22
23
24
25
26
27
v o i d t r y _ r e m o v e _ k s e g m e n t ( top_old ) :
i f top_old == top :
i f top_old - > next != null :
a t o m i c _ i n c r e me n t (& top_old - > remove ) ;
i f empty ( top_old ) :
top_new = atomic_value ( top_old - > next , top_old . ver +1) ;
i f CAS (& top , top_old , top_new ) :
return ;
a t o m i c _ d e c r e me n t (& top_old - > remove ) ;
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
bool committed ( top_old , item_new , index ) :
i f top_old - > s [ index ] != item_new :
r e t u r n true ;
e l s e i f top_old - > remove == 0:
r e t u r n true ;
e l s e : // top_old - > remove >= 1
item_empty = atomic_value ( EMPTY , item_new . ver +1) ;
i f top_old != top :
i f ! CAS (& top_old - > s [ index ] , item_new , item_empty ) :
r e t u r n true ;
else:
top_new = atomic_value ( top_old . val , top_old . ver +1) ;
i f CAS (& top , top_old , top_new ) :
r e t u r n true ;
i f ! CAS (& top_old - > s [ index ] , item_new , item_empty ) :
r e t u r n true ;
r e t u r n false ;
46
47
48
49
50
51
52
53
54
55
56
57
58
59
v o i d push ( item ) :
w h i l e true :
top_old = top ;
i f s e g m e n t _ i s _ m a r k e d ( top_old ) :
i f t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) ;
r e t u r n true ;
c o n t i n u e ; // Restart while loop .
item_old , index = f i nd _e mp t y_ sl ot ( top_old ) ;
i f top_old == top :
i f item_old . val == EMPTY :
item_new = atomic_value ( item , item_old . ver +1) ;
i f CAS (& top_old - > s [ index ] , item_old , item_new ) :
i f committed ( top_old , item_new , index ) :
XX:33
XX:34
Local Linearizability
m a r k _ s e g m e n t _ a s _ u s e d ( old_top ) ;
60
r e t u r n true ;
61
else:
i f t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) :
r e t u r n true ;
62
63
64
65
item pop () :
w h i l e true :
top_old = top ;
item_old , index = find_item ( top_old ) ;
i f top_old == top :
i f item_old . val != EMPTY :
item_empty = atomic_value ( EMPTY , item_old . ver +1) ;
i f CAS (& top_old - > s [ index ] , item_old , item_empty ) :
r e t u r n item_old . val ;
else:
i f only_ksegment ( top_old ) :
i f empty ( top_old ) :
i f top_old == top :
r e t u r n null ;
else:
t r y _ r e m o v e _ k s e g m e n t ( top_old ) ;
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
}
Listing 5 Locally Linearizable k-Stack
G.1.4
Correctness Proof of LL k-Stack.
The local linearizability proof of LL k-Stack is more involved, but very interesting. We use
a theorem from the published artifact of [14], which has been mechanically proved in the
Isabelle HOL theorem prover.
I Theorem 22 (Empty Returns for Stack). Let h be a history, and let h0 be the projection of
h to Σ \ pop(empty). If h is linearizable with respect to the sequential specification SP of a
pool (see Definition 2), and h0 is linearizable with respect to the sequential specification SS
of a stack (see Definition 2), then h is linearizable with respect to SS .
Proof. Here we repeat the key insights of the proof and leave out technical details. A
complete and mechanized version of the proof is available in the published artifact of [14].
As h is linearizable with respect to SP , and h0 is linearizable with respect to SS , there
exists a sequential history s ∈ SP such that s is a linearization of h, and there exists
a sequential history s0 ∈ SS such that s0 is a linearization of h0 . We show that we can
construct a sequential history t ∈ SS such that t is a linearization of h.
The linearization t is constructed as follows: the position of pop(empty) in s is preserved in t. This means for any method call m ∈ s that if pop(empty) ≺s m, then also
pop(empty) ≺t m, and if m ≺s pop(empty), then also m ≺t pop(empty). Moreover, if two
method calls m, n ∈ s are ordered as m ≺s pop(empty) ≺s n and therefore by transitivity it
holds that m ≺s n, then also m ≺t n.
For all other method calls the order of s0 is preserved. This means for any two method
calls m, n ∈ s with m ≺s0 n, that if for all pop(empty) it holds that pop(empty) ≺s m if and
only if pop(empty) ≺s n, then m ≺t n.
By construction, the history t is sequential and a permutation of h. Next we show that
t is a linearization of h by showing that t preserves the precedence order of h. Also by
construction, it holds that if m ≺t n for any two method calls m, n ∈ t, then also either
m ≺s n or m ≺s0 n. Both s and s0 are linearizations of h and h0 , respectively. Therefore
it cannot be for any m, n with m <h n that n ≺s m or n ≺s0 m, it can also not be that
n ≺t m. Since t is sequential, this means that t preserves the precedence order of h.
Next we show that t ∈ SP according to Definition 2:
A. Haas et al.
(1) Every method call, but pop(empty), appears in s at most once: This is guaranteed since
t is a permutation of s, and s ∈ SP .
(2) If pop(x) appears in t, then also push(x) does and push(x) ≺t pop(x): again, since
t is a permutation of s and s ∈ SP , if pop(x) ∈ t, then also push(x) ∈ t. Since
push(x) ≺s pop(x) and push(x) ≺s0 pop(x) (because both s and s0 are in SP ) it also
holds that push(x) ≺t pop(x), as we argued already above.
(3) ∀x ∈ V. push(x) ≺t pop(empty) ⇒ pop(x) ≺t pop(empty): this property is satisfied
trivially as all pop(empty) operations are ordered the same in t as in s, and s ∈ SP .
It only remains to check that all elements are removed in a stack fashion. We have to
show the following:
∀x, y ∈ V. push(x) ≺t push(y) ≺t pop(x) ⇒ pop(y) ∈ t ∧ pop(y) ≺t pop(x)
First we show that if push(x) ≺t push(y) ≺t pop(x), then also push(x) ≺s0 push(y) ≺s0
pop(x). We do this by showing that there cannot exist a pop(empty) such that push(x) ≺t
pop(empty) ≺t push(y) or push(y) ≺t pop(empty) ≺t pop(x).
Assume, towards a contradiction, push(x) ≺t pop(empty) ≺t push(y). By the transitivity of ≺t this implies that push(x) ≺t pop(empty) ≺t pop(x), which contradicts our
observation above that t ∈ SP . Therefore push(x) ≺t pop(empty) ≺t push(y) is not possible, and for the same reason also push(y) ≺t pop(empty) ≺t pop(x) is not possible.
Now, as s0 ∈ SS and push(x) ≺s0 push(y) ≺s0 pop(x), there has to exist a pop(y) ∈ s0 with
pop(y) ≺s0 pop(x). For the same reason as above it cannot be that pop(x) ≺s pop(empty) ≺s
pop(y). Therefore pop(y) and pop(x) are ordered in t the same as in s0 , i.e. pop(y) ≺t pop(x),
and therefore t ∈ SS .
J
I Theorem 23 (Correctness of LL k-Stack). The LL k-Stack algorithm presented in Listing 5
is locally linearizable.
Proof. We have to show that every history h of LL k-Stack is locally linearizable with
respect to the sequential specification SS defined in Definition 2. This means that we have
to show that every thread-induced history hi of h is linearizable with respect to SS for any
thread i.
Having Theorem 22 we only have to show that hi is linearizable with respect to the
sequential specification SP of a pool (defined in Definition 2), and that h0 i , the projection
of hi to Σ \ pop(empty), is linearizable with respect to the sequential specification SS of a
stack.
We start with the proof that hi is linearizable with respect to SP . We construct a
sequential history si from hi by identifying the linearization points of the push and pop
method calls of the LL k-Stack. This means that two method calls m, n are ordered in si ,
m ≺si n if the linearization point of m is executed before the linearization point of n in hi .
The linearization point of push method calls is either the successful insertion of a new
segment in line 15, or the last successful CAS which writes the element into a segment slot
in line 58. The linearization point of pop method calls is the successful CAS which removes
an element from its segment slot in line 73.
For the linearization point of pop(empty) we take the linearization point of the call to
empty in line 77. The empty method creates an atomic snapshot [25] of the top segment.
This atomic snapshot is the state of the top segment at some point (i.e. linearization point
of empty) within the execution of empty. If empty returns true, then there exists no element
in the atomic snapshot of the segment.
XX:35
XX:36
Local Linearizability
Next we show that si is in SP as defined in Definition 2.
(1) Since there exists exactly one linearization point per method call, every method call,
but rem(empty), appears in si at most once.
(2) If pop(x) appears in si , then it reads x in a slot of the top segment before its linearization
point. Since only push method calls write their elements into segment slots, there has to
exist a push(x) which wrote x into that slot. Therefore the linearization point of push(x)
is always before the linearization point of pop(x), and therefore push(x) ≺si pop(x).
(3) Segments are only removed from the list of segments when they become empty. The call
to committed guarantees that elements are not inserted into segments which are about
to be removed.
A pop method calls empty only if there is a single segment left in the LL k-Stack and
no element was found in that segment in find_item.
Now assume a push(x) method call inserts an element x which is missed by find_item.
If push(x) wrote x into a segment before the linearization point of pop(empty) and
the segment was not the last segment, then the top segment changed since pop(empty)
searched for an element and therefore the check in line 70 would fail. If push(x) wrote x
into the last segment of the LL k-Stack, then a pop(x) method call removed x from the
segment because otherwise x would be in the atomic snapshot of empty and therefore
empty would return false. Therefore, if push(x) ≺si pop(empty), then also pop(x) ≺si
pop(empty).
Therefore si is in the sequential specification SP of a pool.
Next we show that h0 i is linearizable with respect to SS . We construct again a sequential
history s0 i from h0 i by identifying the linearization points of the push and pop method calls
of LL k-Stack.
The linearization point of the push operations is the successful insertion of a new segment
in line 15 if it is executed, or the reading of the empty slot (line 54) in the last (and therefore
successful) iteration of the main loop. The linearization point of a pop operation is the
reading of a non-empty slot (line 69) in the last (and therefore successful) iteration of the
main loop. There do not exist any pop(empty) method calls in s0 i . Since we assume a
sequentially consistent memory model, these read operations define a total order on the LL
k-Stack method calls in h0 i .
First we show that s0 i is in the sequential specification SP of a pool as defined in Definition 2.
(1) Since there exists exactly one linearization point per method call, every method call
appears in s0 i at most once.
(2) If pop(x) appears in s0 i , then it read x in a slot of the top segment at its linearization
point. Since only push operations write their elements into segment slots, there has
to exist a push(x) which wrote x into that slot. The linearization point of push(x) is
always before x is written into a segment slot. Therefore push(x) ≺s0 i pop(x).
(3) Since there exist no pop(empty) operations in s0 i the third pool condition is trivially
correct.
Next we show that s0 i also provides a stack order, which means that we have to show
that
∀x, y ∈ V. push(x) ≺s0 i push(y) ≺s0 i pop(x) ⇒ pop(y) ∈ s0 i ∧ pop(y) ≺s0 i pop(x).
We start by observing some invariants.
A. Haas et al.
1. A thread never inserts elements into the same segment twice. This is guaranteed by the
call to segment_is_marked.
2. Between the linearization point of a push and the time it writes its element into a segment
the segment the element gets written into is not removed: if the push operation inserts
a new segment this is trivially correct. If the push operation writes the element into an
existing segment, then the call to committed in line 59 guarantees that the segment was
not removed.
3. At the time of the linearization point of the pop, which is the time when the pop reads
the non-empty slot (line 69) in the last (and therefore successful) iteration, the pop reads
the non-empty slot from the top segment. This is guaranteed by the check in line 70.
Now assume there exist the operations push(x), push(y) and pop(x) in s0 i and
push(x) ≺s0 i push(y) ≺s0 i pop(x). Since push(x) and push(y) are both in s0 i , this means
that both operations are executed by the same thread. Therefore, according to Invariant 1.,
x and y get inserted into different segments, with the segment y on top of the segment of x.
The linearization point of pop(x) cannot be before y is written into its segment because
according to Invariant 2. the segment y gets inserted into does not get removed between the
linearization point of push(y) and the time y is written into the segment. With Invariant
3. this means that x is unaccessible for pop(x) before y gets written into a segment. Also
because of the third invariant the top segment changes between the insertion of y and the
linearization point of pop(x).
Next we observe that as long as y is not removed, no segment below the segment of y can
become the top segment. Therefore for the segment of x to become the top segment so that
pop(x) can remove it, y has to be removed first. Only a pop(y) can remove y, and therefore
there exists a pop(y) and the linearization point of pop(y) is before the linearization point
of pop(x).
Hence s0i is in the sequential specification of a stack. Using Theorem 22 this means that
LL k-Stack in listing 5 is locally-linearizable with respect to the sequential specification of
a stack.
J
H
Additional Experiments
We also evaluate the implementations on another Scal workload, the sequential alternating
workload. However, we note that in this workload in the locally linearizable implementations
threads only access their local backends, so no wonder they perform perfectly well.
Mixed Workload. In order to evaluate the performance and scalability of mixed workloads, i.e., workloads where threads produce and consume values, we exercise the so-called
sequential alternating workload in Scal. Each thread is configured to execute 106 pairs of
insert and remove operations, i.e., each insert operation is followed by a remove operation.
As in the producer-consumer workload, the contention is controlled by adding a busy wait
of 5µs. The number of threads is configured to range between 1 and 80. Again we report
the number of data structure operations per second.
Data structures that require parameters to be set are configured like in the producerconsumer benchmark. Figure 11 shows the results of the mixed workload benchmark for all
considered data structures.
The MS queue and Treiber stack do not perform and scale for more than 10 threads.
As in the producer-consumer benchmark, LCRQ and TS Stack either perform competitively
XX:37
32
28
24
20
16
12
8
4
0
1
10
20
30
40
50
60
70
80
million operations per sec (more is better)
Local Linearizability
million operations per sec (more is better)
XX:38
32
28
24
20
16
12
8
4
0
1
10
20
30
40
50
60
70
80
number of threads
number of threads
MS
LCRQ
k-FIFO
LL+D MS
LLD LCRQ
LLD k-FIFO
1-RA DQ
Queues, LL queues, and “queue-like” pools
Treiber
TS Stack
k-Stack
LL+D Treiber
LLD TS Stack
LLD k-Stack
1-RA DS
Stacks, LL stacks, and “stack-like” pools
Figure 11 Performance and scalability of sequential alternating microbenchmarks with an increasing number of threads on a 40-core (2 hyperthreads per core) machine
with their k-out-of-order relaxed counter parts k-FIFO and k-Stack or even outperform and
outscale them (in the case of LCRQ, that even outperforms the pool).
LL+ D MS queue, LLD LCRQ, and LL+ D Treiber stack perform very well and scale
(nearly) linearly in the number of threads. A surprising result is that LLD k-FIFO performs
poorly in this experiment. The reason is that k-FIFO performs poorly when it is almost
empty, and in this experiment each backend instance of LLD k-FIFO contains at most one
element at any point in time. The k-Stack performs better on a nearly-empty state. The
benefit of trying to perform a local operation first in the LLD algorithms is visible when
comparing to 1-RA DQ and DS that do not utilize a local fast path.
I
Verifying Local Linearizability
In general, verifying local linearizability amounts to verifying linearizability for a set of
smaller histories. This might enable verification in a modular/compositional way. Aside
from this, it is important to mention (again) that for our locally linearizable data structures
in Section 6 built from linearizable building blocks, the correctness proofs are straightforward
assuming the building blocks are proven to be linearizable. In addition, for queue we can
state an “axiomatic" verification theorem for local linearizability in the style of [24, 10],
whose main theorem we recall next (with a slight reformulation).
I Theorem 24 (Queue Linearizability). A queue concurrent history h is linearizable wrt the
queue sequential specification SQ if and only if
1. h is linearizable wrt the pool sequential specification SP (with suitable renaming of method
calls), and
2. ∀x, y ∈ V. enq(x) <h enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x). J
We note that an analogous change to the axioms in the sequential specification of a pool
and a stack does not lead to a characterisation of linearizability for pools and stacks, cf. [14].
An axiomatic characterisation of linearizability for pools and stacks would involve an infinite
number of axioms/infinite axioms, due to the need to prohibit infinitely many problematic
shapes, cf. [7].
We are now able to state the queue-local-linearizability-verification result.
A. Haas et al.
I Theorem 25 (Queue Local Linearizability). A queue concurrent history h is locally linearizable wrt the queue sequential specification SQ if and only if
1. h is locally linearizable wrt the pool sequential specification SP (after suitable renaming
of method calls), and
2. ∀x, y ∈ V. ∀T. enq(x) <Th enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h
deq(x).
J
XX:39
| 6cs.PL
|
Simulating the Dynamics of T Cell Subsets Throughout
the Lifetime
Stephanie Foan, Andrew Jackson, Ian Spendlove, Uwe Aickelin
Academic Unit of Clinical Oncology and Intelligent Modelling and Analysis Research Group,
University of Nottingham
Abstract. It is widely accepted that the immune system undergoes age-related
changes correlating with increased disease in the elderly. T cell subsets have been
implicated. The aim of this work is firstly to implement and validate a simulation
of T regulatory cell (Treg) dynamics throughout the lifetime, based on a model by
Baltcheva. We show that our initial simulation produces an inversion between
precursor and mature Treys at around 20 years of age, though the output differs
significantly from the original laboratory dataset. Secondly, this report discusses
development of the model to incorporate new data from a cross-sectional study of
healthy blood donors addressing balance between Treys and Th17 cells with novel
markers for Treg. The potential for simulation to add insight into immune aging is
discussed.
1 Introduction to System Dynamics Modelling of Immunity
Simulation has been defined as methods and applications mimicking the behaviour of
a real system [1]. The benefits of simulation to immunology include time- and costeffectiveness as well as less labour- and resource-intensiveness resulting from
removal from the biological environment. In vitro experimentation is useful for
investigating individual interactions but is far removed from the whole picture, and in
vivo experimentation is useful for the whole picture but is unlikely to answer specific
questions [2]. Using simulation, flexibility is available for systematically generating
hypotheses and conducting experiments impossible to do practically, yet informed by
robust data and literature.
System dynamics simulations are useful for looking at complex systems over time.
They are characterised by stocks of an entity and flows between stocks [3]. Immune
system examples of stocks include precursor and mature T cell pools and flows might
represent transition of cells from precursor to mature. This technique is useful for
modelling relationships defined by differential equations. An example of a differential
equation describing T cell dynamics is the change in number of precursor cells
equated to proliferation of precursors minus death and maturation rates. This ongoing
work will apply system dynamics simulation technique to complement in vitro studies
of Th 17s and Treg s throughout the lifetime.
2 The Need for Balance: Th 17s and Treys Throughout the Lifetime
The immune system maintains a balance between mounting an adequate immune response to protect from infection and restricting the size of the immune response to
2
prevent damage to self. There is evidence to suggest an age-related tendency to a proinflammatory environment [4] contributing to more collateral damage and autoimmune
diseases. This work addresses the hypothesis that important contributors to this state of
imbalance are Th17 cells (amplifying immune responses) and Tregs (dampening down
immune responses). Although some studies have shown an age-related increase in the
number of Tregs in human peripheral blood [5][6], it has also been shown that
homeostasis is maintained [7]. One study concluded upon the oscillatory nature of Treg
numbers through life, with peaks in adolescence and in over 60 year olds [8]. A recent
study in Th17 cells with age showed a small decrease in the frequency of Th17 cells in
the CD4+ memory population in elderly donors relative to young [9]. However, for the
balance between Tregs and Th17 cells, there is currently no published literature.
There is evidence that the balance between Tregs and Th 17s is altered in agerelated diseases such as acute coronary syndrome [10]. Thus it is intuitive that this
balance should be examined with a cross sectional study in healthy donors of different
ages. Laboratory experimentation will begin by using flow cytometry to enumerate
peripheral blood cells expressing CD4, CD25 and signature transcription factors of
these subsets: Foxp3 and Helios for Tregs and RORC for Th17 cells.
3 Method: Simulation of Treg Dynamics
Ultimately we wish to build a model of the dynamics of Tregs and Th17 cells throughout
life from data currently being collected. Preliminary work for this has involved building
a system dynamics simulation in AnyLogic 6.5.0 University Edition, based on a
mathematical model by Baltcheva [11]. This model was selected as it comprehensively
incorporates the functional dynamics of Tregs in terms of homeostasis and during an
acute immune response. It characterises the changing precursor and mature Treg
populations throughout the human lifetime. Key assumptions include that there is no
change in function or responsiveness throughout the lifetime, nor a change to other
influential factors on their dynamics such as dendritic cell number and function [11].
Also, the immune response considered includes an expansion and contraction phase, and
only one response can occur at a given timepoint [11]. The original model was based on
numbers of CD4+CD25±CD45R0— (precursor) and CD4+CD25±CD45R0+ (mature)
populations in 119 peripheral blood samples of donors aged 19 to 81 [11]. Although
total numbers of CD4+CD25+ cells remained constant, the ratio of precursor to mature
was inverted in early adulthood. This represents an important dimension to the observed
homeostasis in Treg numbers throughout the lifespan, especially when considering
thymic involution from adolescence, reducing the number of new cells entering the
system.
Ordinary differential equations describe the dynamics of the above mentioned cells,
and stochastic processes control the frequency, duration and antigen-specific nature of
primary and secondary immune responses on the different cell compartments [11]. In
this work, a simple scenario was chosen in order to test the hypothesis that the model
could be implemented in AnyLogic. The scenario assumes a lack of antigen-induced
proliferation and death of both precursor and mature Tregs, no density-dependent proliferation and death and thymic output as the only external input into the various Treg
3
subsets [11]. The parameter values used correspond to means of the distributions for
scenario 2iiia given in Baltcheva's work [11]. The simulation is shown below:
0 immuneResponse [..]
Day
OF
Year
gPt
Oh o„ ,
YdoneP
)
Pp
[=]
gQt
.gp,,
sigm aQ
a
[01. _________
[o]
hQ
gQt1
sigmaP
3 piN
hP
QuiescentReactivation
PreoursorActivaton
C. alpha
YdoneR
.Ez1- -- •
YdoneQ
dP PreoursorDeath
[0]
[Yr
Rh
PreoursorProlf ation
ActlyeProlf eration
aP
[
aR
[_]
[2:]
ActiveDeath
ActivetoQuescent
QuiescentDeath
e,
dQ
g dR
Figure 1: Main View. Events are shown with for time in days (t) and years (y). Flow
variables given by [z] involve a rate of conversion multiplied by the number of cells in the
stocks named Y cloneP, —R and —Q. P corresponds to precursors, R to active matures
and Q to quiescent matures. These stocks are prefixed by Y clone as they represent total
and antigen-specific Tregs determined by an array. Homeostatic parameters apply to the
total Treg population, whereas immune response parameters are applied to a proportion
given by piN. For example, flow of specific precursors into the active mature stock is
given by specific precursor stock multiplied by the maturation rate. Each time point,
AnyLogic recalculates each stock using the flows defined.
An additional class (immuneResponse) controls the immune system functional status
using IRstatechart:
qt
IRstatechart
_ NoPesponse
Day
K2
Year
e,th °t
0Y
gr K1
Figure 2: immuneResponse View. At the beginning of a run the immune system is in the
N °Response state. Every 100.95 time steps, the immune system mounts a primary
immune response with probability qt or defaults to a secondary response. Primaryl and
Secondaryl, represent the expansion phase. PrimaryI I and SecondaryI I, represent the
contraction phase which continues until a new response is instigated. During Primaryl,
parameter b is applied. During Primaryll, b is set to zero and parame-
4
ters c, dR and dQ are applied. In Secondary) parameters b and f are applied and in
SecondaryI I these are set to zero and c, dR and dQ are applied.
25
2 5
Total Precursors
Total Quiescent Matures
fl h l d 1 1 1 1 0 1 1 1 1 1 - 0 1 1 1 0 1 1 I 1 1 0
20
GO
BO
100
120
GO
Ag e ( Ye a rs)
Age (Years)
80
100
Figure 3 (generated using MATLAB 122010a): Output data from AnyLogic over a
complete run. a) Total precursor and quiescent mature Trey stocks. b) Total active
Tregs. Each peak corresponds to antigen-specific clones experiencing either primary
or secondary immune responses.
Figure 4 shows simulation output compared to the original dataset. Data has been collected for each stock over 3 complete replications. The maximum standard deviation
between three runs for the total precursors was 2.753x10 -7. The maximum standard
deviation for total quiescent matures was 9411.
08
09
08
06
0 5
01
20
40
GO
Age (Years)
80
100
120
0
4
0
3
0
2
20
40
GO
80
100
120
Age (Years)
Figure 4: Output data compared to Baltcheva's dataset. a) The proportion of total
Tregs in the precursor stock. b) The proportion of total Tregs in the quiescent stock.
In order to quantify how similar the simulation output was to the laboratory data, both
datasets were split into 10-year age groups. The median was calculated for these
groups, and the difference between medians for simulation output data and laboratory
data have been documented below. A Mann Whitney test was then performed for the
null hypothesis of no difference between laboratory and output data and the p values
are given below:
5
Age (Years)
10-19
20-29
30-39
40-49
50-59
60-69
70-79
80-89
Median Difference
Mann Whitney Test
Proportion of Proportion of Proportion of Proportion of
Precursors
Matures
Precursors
Matures
0.3101
0.0207
p=0.083
p=0.658
0.0525
0.1092
p<0.001
p<0.001
0.1227
0.0290
p=0.034
p<0.001
0.1182
0.0690
p=0.001
p<0.001
0.0587
0.1007
p<0.001
p<0.001
0.1230
p=0.127
0.0072
p<0.001
0.0011
0.1204
p=0.808
p<0.001
p=0.014
0.0805
0.0775
p=0.014
Table 1: Comparison of median output and laboratory data for each 10-year age
group.
4 Discussion and Concluding Remarks
The implementation of Baltcheva's model as a system dynamics simulation has been
documented here and compared to experimental evidence. It has been shown that the
simulation mimics the key feature of inversion of precursor and memory cells in early
adulthood. The lack of statistical similarity between simulation output and laboratory data
indicates that further validation of this model is necessary and will involve a comparison
of other scenarios proposed in Baltcheva's work. Ultimately we will develop and validate
a simulation of our novel dataset of Treg and Th17 cells using this sort of approach.
CD4+CD25±Foxp3+ and CD4±Foxp3±Helios+ cell numbers instead of CD4+CD25+ s
for Treg s will also be collected as they are arguably more specific markers [12][13]. In
terms of improving the simulation, alternatives to continually reactivating a single Treg
clone are required, as is simulation of more than one immune stimulus at a time.
Baltcheva discloses various assumptions including no difference in Treg function [11]. It
may be possible to improve this model by considering functional as well as numerical
changes to Treg subsets with age.
A more abstract research question is whether a simplistic model of immunosenescence can lend useful insight into the biological problem. It can be argued that the process
of simulation alone might allow researchers to address assumptions and allow for
systematic generation of hypotheses. Also, hypotheses which are difficult to test in the
laboratory might be testable with a simulation. For example, we might introduce an
intervention to mimic ablative chemotherapy by depleting each stock at a single time
point. Total values of each stock might then be compared for simulation runs with or
without intervention to make hypotheses about Treg recovery. Simulating the dynamics of
Th17 cells in parallel to Treg s may also allow us to make predictions about the maintenance of their balance throughout life, would allow for extreme parameter values to be
tested and may indicate a maximum length of time for homeostasis to be maintained.
Our primary hypothesis is that age alters Treg and Th17 cells with consequences for
health in older age and we aim to conduct a cross sectional study to obtain the distribution
of particular changes. We anticipate that a strategy of both laboratory investigation
6
and system dynamics simulation as exemplified by Baltcheva's work will be useful to
address relationships between T cell subsets over time. The model might also be developed
to consider new questions about response to interventions and the length of time the
immune system might be able to maintain Treg and Th17 cell homeostasis.
With thanks to Irina Baltcheva for providing her raw data.
References
1. Kelton, W.D., Sadowski, R.P., Swets, N.B.: Simulation with Arena: 5th Edition.
McGraw-Hill International Edition (2010)
2. Kim, P.S., Levy, D., Lee, P.P.: Modeling and Simulation of the Immune System as a
Self-Regulating Network. Volume 467 of Methods in Enzymology. Academic Press
(2009) 79 —109
3. Figueredo, G.P., Aickelin, U.: Investigating immune system aging: System dynamics
and agent-based modelling. In: Proceedings of the Summer Computer Simulation
Conference 2010. (2010)
4. Boren, E., Gershwin, M.E.: Inflamm-aging: Autoimmunity, and the Immune-Risk
Phenotype. Autoimmunity Reviews 3 (2004) 401-406
5. Rosenkranz, D., Weyer, S., Tolosa, E., Gaenslen, A., Berg, D., Leyhe, T., Gasser, T.,
Stoltze, L.: Higher Frequency of Regulatory T Cells in the Elderly and Increased Suppressive
Activity in Neurodegeneration. Journal of Neuroimmunology 188 (2007) 117-127
6. Gregg, R., Smith, C.M., Clark, EJ., Dunnion, D., Khan, N., Chakraverty, R., Nayak, L.,
Moss, P.A.: The Number of Human Peripheral Blood CD4+ CD251"gh Regulatory T Cells
Increases with Age. Clinical and Experimental Immunology 140 (2005) 540-546
7. Hwang, K.A., Kim, H.R., Kang, I.: Aging and Human CD4+ Regulatory T Cells.
Clinical Immunology 130 (2009) 509-517
8. Faria, A.M., de Moraes, S.M., de Freitas, L.H., Speziali, E., Soares, T.F., FigueiredoNeves, S.P., Vitelli-Avelar, D.M., Martins, M.A., Barbosa, K.V., Soares, E.B., SathlerAvelar, R., Peruhype-Magalhaes, V., Cardoso, G.M., Comin, E, Teixeira, R., Eloi-Santos,
S.M., Queiroz, D.M., Correa-Oliveira, R., Bauer, M.E., Teixeira-Carvalho, A., MartinsFilho, 0.A.: Variation Rhythms of Lymphocyte Subsets During Healthy Aging. Neuroimmunomodulation 15 (2008) 365-379
9. Lee, J.S., Lee, W.W., Kim, S.H., Kang, Y., Lee, N., Shin, M.S., Kang, S.W., Kang, I.: Ageassociated alteration in naive and memory th17 cell response in humans. Clinical Immunology
In Press, Corrected Proof (2011) —
10. Li, Q., Wang, Y., Chen, K., Zhou, Q., Wei, W., Wang, Y., Wang, Y.: The Role of Oxidized
Low-Density Lipoprotein in Breaking Peripheral Th17/Treg Balance in Patients with Acute
Coronary Syndrome. Biochemical and Biophysical Research Communications 394 (2010)
836-842
11. Baltcheva, I., Codarri, L., Pantaleo, G., Boudec, J.Y.L.: Lifelong Dynamics of
Human CD4+CD25+ Regulatory T Cells: Insights from in vivo Data and Mathematical
Modeling. Journal of Theoretical Biology 266 (2010) 307-322
12. Thornton, A.M., Korty, P.E., Tran, D.Q., Wohlfert, E.A., Murray, P.E., Belkaid, Y.,
Shevach, E.M.: Expression of Helios, an Ikaros Transcription Factor Family Member,
Differentiates Thymic-Derived from Peripherally Induced Foxp3+ T Regulatory Cells.
The Journal of Immunology 184 (2010) 3433-3441
13. Ziegler, S.F., Buckner, J.H.: Foxp3 and the Regulation of Treg/Th17 Differentiation.
Microbes and Infection 11 (2009) 594-598
| 5cs.CE
|
UNIFORM NULL CONTROLLABILITY FOR A DEGENERATING
REACTION-DIFFUSION SYSTEM APPROXIMATING A SIMPLIFIED CARDIAC
MODEL
arXiv:1106.1788v6 [math.OC] 25 Sep 2015
FELIPE WALLISON CHAVES-SILVA∗ AND MOSTAFA BENDAHMANE
Abstract. This paper is devoted to the analysis of the uniform null controllability for a family of nonlinear reaction-diffusion systems approximating a parabolic-elliptic system which models the electrical
activity of the heart. The uniform, with respect to the degenerating parameter, null controllability of
the approximating system by means of a single control is shown. The proof is based on the combination
of Carleman estimates and weighted energy inequalities.
1. Introduction
Let Ω ⊂ RN (N = 2, 3) be a bounded connected open set whose boundary, ∂Ω, is sufficiently regular.
Let T > 0, and let ω and O be two (small) nonempty subsets of Ω, which we will refer to as control
domains. We will use the notation Q = Ω × (0, T ) and Σ = ∂Ω × (0, T ).
The main objective of this paper is to study the properties of controllability and observability for a
family of nonlinear reaction-diffusion systems which degenerates into a nonlinear parabolic-elliptic system
which models the electrical activity in the cardiac tissue.
To state the model, we let ui = ui (t, x) and ue = ue (t, x) represent the intracellular and extracellular
electric potentials, respectively. Their difference, v = ui − ue , is called the transmembrane potential. The
anisotropic properties of the media are modeled by intracellular and extracellular conductivity tensors
Mi (x) and Me (x).
The widely accepted model (see [11, 22, 35]) describing the electrical activity in the cardiac tissue
reads as follows:
cm ∂t v − div (Mi (x)∇ui ) + h(v) = f 1ω
in Q,
(1.1)
cm ∂t v + div (Me (x)∇ue ) + h(v) = g1O
in Q,
where cm > 0 is the surface capacitance of the membrane, the nonlinear function h : R → R is the
transmembrane ionic current (the most interesting case being when h is a cubic polynomial), and f and
g are stimulation currents applied, respectively, to ω and O.
System (1.1) is known as the bidomain model and is completed with Dirichlet boundary conditions for
the intra- and extracellular electric potentials
ui = ue = 0 on Σ
(1.2)
2010 Mathematics Subject Classification. 35K57, 93B05, 93B07, 93C10.
Key words and phrases. reaction-diffusion system, monodomain model, Carleman estimates, uniform null controllability,
observability.
∗ F. W. Chaves-Silva has been supported by the ERC project Semi Classical Analysis of Partial Differential Equations,
ERC-2012-ADG, project number 320845; the Grant BFI-2011-424 of the Basque Government and partially supported by
the Grant MTM2011-29306-C02-00 of the MICINN, Spain, the ERC Advanced Grant FP7-246775 NUMERIWAVES, ESF
Research Networking Programme OPTPDE and the Grant PI2010-04 of the Basque Government.
1
2
F. W. CHAVES-SILVA AND M. BENDAHMANE
and initial data for the transmembrane potential
v(0, x) = v0 (x),
x ∈ Ω.
(1.3)
We point out that realistic models describing electrical activities in the heart also include a system
of ODE’s for computing the ionic current as a function of the transmembrane potential and a series
of additional “gating variables” which aim to model the ionic transfer across the cell membrane (see
[23, 25, 31, 32]).
In the case where f 1ω = g1O and Mi = µMe , for some constant µ ∈ R, the bidomain model is
simplified into the following parabolic-elliptic system:
µ
in Q,
cm ∂t v − µ+1 div Me (x)∇v + h(v) = f 1ω
−div M(x)∇u = div M (x)∇v
in Q,
e
i
(1.4)
v = ue = 0
on Σ,
v(0) = v0
in Ω,
where M = Mi + Me .
System (1.4) is known as monodomain model and is a very interesting model from the implementation
point of view, since it conserves some of the essential features of the bidomain model as excitability
phenomena (see [11, 27, 36]).
The main difference between the bidomain model (1.1) and the monodomain model (1.4) is the fact
that the first model is a system of two coupled parabolic equations, while the second one is a system of
parabolic-elliptic type. Therefore, from the control point of view, one could expect these two systems to
have, at least a priori, different control properties. In this work we show that the properties of controllability and observability for the monodomain model can be seen as a limit process of the controllability
properties of a family of coupled parabolic systems. Indeed, given ε ∈ R such that 0 < ε ≤ 1, we
approximate the monodomain model by the following family of parabolic systems:
µ
ε
ε
+ h(v ε ) = f ε 1ω
in Q,
cm ∂t v − µ+1 div Me (x)∇v
ε∂ uε − div M(x)∇uε = div M (x)∇v ε
in
Q,
t e
i
e
(1.5)
ε
ε
on Σ,
v = ue = 0
ε
in Ω.
v (0) = v0 , uεe (0) = ue,0
In this paper we give a positive answer to the following question:
Question 1.1. If, for each ε > 0, there exists a control f ε that drives the solution (v ε , uεe ) of (1.5) to
zero at time t = T , i.e.,
v ε (T ) = uεe (T ) = 0,
is it true that, when ε → 0+ , the control sequence {f ε }ε>0 converges to a function f which drives the
associated solution (v, ue ) of (1.4) to zero at time t = T ?
This question of approximating an equation by another having different physical properties has been
used several times in the case of parabolic equations degenerating into hyperbolic ones (see, for example,
[12, 16, 21]) and hyperbolic equations degenerating into parabolic ones (see, for example, [29, 30]).
However, as far as we know, this is the first time that controllability of parabolic systems degenerating
into parabolic-elliptic systems is studied. It is also important to mention that families of parabolic
systems which degenerate into parabolic-elliptic ones arise in many areas, such as biology, chemistry and
astrophysics (see [9, 10, 26]).
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
3
As usual, in control theory, when dealing with the controllability of a nonlinear problem, we first
consider the linearized version of (1.5):
µ
div Me (x)∇v ε + a(t, x)v ε = f ε 1ω
in Q,
cm ∂t v ε − µ+1
ε∂ uε − div M(x)∇uε = div M (x)∇v ε
in Q,
t e
i
e
(1.6)
ε
ε
v
=
u
=
0
on Σ,
e
ε
v (0) = v0 , uεe (0) = ue,0
in Ω,
where a is a bounded function.
Given ε > 0, the first obstacle to answering, positively, Question 1.1, will be to drive (v ε , uεe ), solution
of (1.6), to zero at time T by means of a control f ε in such a way that the sequence of controls {f ε }ε>0
converges when ε → 0+ . Once it is shown that such a convergent sequence of control, {f ε }ε>0 , for the
linear system (1.6), exists, we employ a fixed point argument and conclude that the same is true for the
nonlinear system (1.5).
Thus, we introduce the adjoint system of (1.6):
µ
div (Me (x)∇ϕε ) + a(t, x)ϕε = div (Mi (x)∇ϕεe ) in Q,
−cm ∂t ϕε − µ+1
−ε∂ ϕε − div (M (x)∇ϕε ) = 0
in Q,
t e
e
(1.7)
ε
ε
ϕ
=
ϕ
=
0
on
Σ,
e
ε
in Ω.
ϕ (T ) = ϕT , ϕεe (T ) = ϕe,T
Using duality arguments, it is very easy to prove that the task of building such a convergent sequence
of controls, {f ε }ε>0 , for (1.6) is equivalent to prove the following (uniform) observability inequality for
the solutions of (1.7):
ZZ
||ϕε (0)||2L2 (Ω) + ε||ϕεe (0)||2L2 (Ω) ≤ C
|ϕε |2 dxdt,
Qω := ω × (0, T ),
(1.8)
Qω
where (ϕT , ϕe,T ) ∈ L2 (Ω)2 and the constant C = C(ε, Ω, ω, ||a||L∞ , T ) remains bounded when ε → 0+ .
We prove inequality (1.8) as a consequence of an appropriate Carleman inequality for the solution
(ϕε , ϕεe ) of (1.7) (see section 3). We notice that, due to the fact the control is acting on the first equation
of (1.6), in our Carleman inequality, we need to bound global integrals of ϕε and ϕεe in terms of a local
integral of ϕε , uniformly with respect to ε. Two main difficulties appear: first, the coupling in the first
equation of (1.7) is in div (Mi (x)∇ϕεe ) and not in ϕεe ; second, we must show that the constant we get in
our Carleman inequality does not blow up when ε → 0+ .
The first difficulty is not so hard to overcome. Indeed, for each ε > 0 fixed, inequality (1.8) is known
to be true for system (1.7) (see [20]). However, the main novelty here is the fact that we obtain the
boundedness of the observability constant C with respect to ε. As we will see, Carleman inequalities
alone are not enough for this task, and we need to combine sharp Carleman estimates, with respect to ε,
and weighted energy inequalities.
As far as the controllability of non degenerate coupled parabolic systems is concerned, the situation is,
by now, fairly well understood. For instance, in [20], the controllability of a quite general linear coupled
parabolic system is studied and a null controllability result is obtained by means of Carleman inequalities.
In [2], using a different strategy, the controllability of a nonlinear reaction-diffusion system of two coupled
parabolic equations is analyzed, and the authors prove the null controllability for the linear system and
the local null controllability of the nonlinear one. Another relevant work concerning the controllability
of coupled systems is [15], in which the authors analyze the null controllability of a cascade system of
m (m > 1) coupled parabolic equations and the authors are able to obtain null controllability for the
4
F. W. CHAVES-SILVA AND M. BENDAHMANE
cascade system, whenever they have a good coupling structure. It is also worth mentioning the works [3],
[4] and [18], where local and global controllability results for phase field systems were studied.
For a general discussion about the controllability of coupled parabolic systems, see the survey paper
[1].
Concerning controllability results for the bidomain model, since in both equations the couplings are
given by the time derivatives of the electrical potentials, it seems very difficult to study controllability
properties for such a model. To the best of our knowledge, for the bidomain model (1.1), the problems of
null and approximate controllability are still open (even with two controls). Regarding the null controllability of the monodomain model (1.4), since the solution of the parabolic equation enters as a source
term in the elliptic one, the following controllability result holds.
Theorem 1.2.
(1) If h is C 1 (R), globally Lipschitz and h(0) = 0, then, for every v0 ∈ L2 (Ω), there
exists a control f ∈ L2 (ω × (0, T )) such that the solution (v, ue ) of (1.4) satisfies:
v(T ) = ue (T ) = 0.
Moreover, the control f satisfies the following estimate:
2
2
kf 1ω kL2 (Q) ≤ C kv0 kL2 (Ω) ,
(1.9)
for a constant C = C(Ω, ω, T ) > 0.
(2) If h is C 1 (R) and h(0) = 0, there exists γ > 0 such that, for every v0 ∈ W 2/3,6 (Ω) ∩ H01 (Ω) with
||v0 ||W 2/3,6 (Ω)∩H01 (Ω) ≤ γ, there exists a control f ∈ L6 (ω × (0, T )) such that the solution (v, ue )
of (1.4) satisfies:
v(T ) = ue (T ) = 0.
Moreover, the control f satisfies the following estimate:
kf 1ω kL6 (Q) ≤ C kv0 kL2 (Ω) ,
(1.10)
for a constant C = C(Ω, ω, T ) > 0.
Theorem 1.2, case 1, follows from [14, Theorem 3.1] and case 2 follows from [19, Theorem 3.5] (see
also [14, Theorem 4.2]).
This paper is organized as follows. In section 2, we state our main results. In section 3, we prove a
uniform Carleman inequality for the adjoint system (1.7). Next, we show, in section 4, the uniform null
controllability of (1.6). In section 5, we deal with the uniform null controllability of the nonlinear system
(1.5).
2. Main results
Throughout this paper we will assume that the matrices Mj , j = i, e are C ∞ , bounded, symmetric
and positive semidefinite.
Our first main result is a uniform Carleman estimate for the adjoint system (1.7).
Theorem 2.1. Given any 0 < ≤ 1, there exist positive constants C = C(Ω, ω), λ0 = λ0 (Ω, ω) ≥ 1 and
s0 = s0 (Ω, ω) ≥ 1 such that, for any (ϕT , ϕe,T ) ∈ L2 (Ω)2 and any a ∈ L∞ (Q), the solution (ϕε , ϕεe ) of
(1.7) satisfies:
ZZ
ZZ
e3sα |ρε |2 dxdt + s3 λ4
φ3 e3sα |ϕε |2 dxdt
Q
Q
ZZ
≤ Ce6λ||ψ|| s8 λ4
φ8 e2sα |ϕε |2 dxdt,
(2.1)
Qω
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
5
2/3
for every s ≥ (T + (1 + ||a||L∞ )T 2 )s0 and λ ≥ λ0 , where ρε (x, t) = div (M (x)∇ϕεe (x, t)) and the weight
functions φ and α are defined in (3.3) and (3.4), respectively.
The proof of Theorem 2.1 follows from a combination of Carleman inequalities, for the heat equation,
with a precise dependence on the degenerating parameter, and an energy inequality for the adjoint system
(1.7). We prove Theorem 2.1 in section 3.
Remark 2.2. As a direct consequence of the Carleman inequality (2.1), we have the unique continuation
property for the solutions (ϕε , ϕεe ) of (1.7):
“Given ε > 0, if ϕε = 0 in ω × (0, T ), then (ϕε , ϕεe ) ≡ (0, 0) in Q”.
This unique continuation property for the adjoint system (1.7) implies, for each ε > 0, the approximate
controllability at time T of system (1.6), with a control acting only on the first equation.
The second main result of this paper gives the global null controllability of the linear system (1.6).
Theorem 2.3. For any 0 < ε ≤ 1 and any (v0 , ue,0 ) ∈ L2 (Ω)2 , there exists a control f ε ∈ L2 (ω × (0, T ))
such that the associated solution, (v ε , uεe ), to (1.6) is driven to zero at time T . That is to say, the
associated solution satisfies:
v ε (T ) = 0, uεe (T ) = 0.
Moreover, the control f ε satisfies the estimate:
2
2
2
kf ε 1ω kL2 (Q) ≤ C kv0 kL2 (Ω) + ε kue,0 kL2 (Ω) ,
(2.2)
for a constant C = C(Ω, ω, ||a||L∞ , T ) > 0.
From Theorem 2.1, the proof of Theorem 2.3 is standard. However, for the sake of completeness, we
prove Theorem 2.3 in section 4.
The third main result of this paper is concerned with the uniform null controllability of the nonlinear
parabolic system (1.5).
Theorem 2.4. Given any 0 < ε ≤ 1, we have:
(1) If h is C 1 (R), globally Lipschitz and h(0) = 0, then, for every (v0 , ue,0 ) ∈ L2 (Ω)2 , there exists a
control f ε ∈ L2 (ω × (0, T )) such that the solution (v ε , uεe ) of (1.5) satisfies:
v ε (T ) = uεe (T ) = 0.
Moreover, the control f ε satisfies the estimate:
2
2
2
kf ε 1ω kL2 (Q) ≤ C kv0 kL2 (Ω) + ε kue,0 kL2 (Ω) ,
(2.3)
for a constant C = C(Ω, ω, T ) > 0.
(2) If h is C 1 (R) and h(0) = 0, there exists γ > 0, does not depending on ε, such that, for every
2
(v0 , ue,0 ) ∈ W 2/3,6 (Ω) ∩ H01 (Ω) with ||(v0 , ue,0 )||W 2/3,6 (Ω) ≤ γ, there exists a control f ε ∈
L6 (ω × (0, T )) such that the solution (v ε , uεe ) of (1.5) satisfies:
v ε (T ) = uεe (T ) = 0.
Moreover, the control f ε satisfies the estimate:
2
2
2
kf ε 1ω kL6 (Q) ≤ C kv0 kL2 (Ω) + ε kue,0 kL2 (Ω) ,
(2.4)
for a constant C = C(Ω, ω, T ) > 0.
The proof of Theorem 2.4 is achieved through fixed point arguments, and it will be done in section 5.
6
F. W. CHAVES-SILVA AND M. BENDAHMANE
Remark 2.5. In this paper we restrict the dimension to N = 2, 3, because the bidomain model makes
sense only in such dimensions. Nevertheless, from the mathematical point of view, systems (1.4), (1.5)
and (1.6) make sense for any N ∈ N (the 1-d case corresponding to the cable equation) and, taking the
initial data in the appropriate space, all the results of this paper can be extended to higher dimensions.
3. Carleman inequality
In this section we prove Theorem 2.1.
To simplify the notation, we neglect the index ε and, since the only constant which matters in the
analysis is ε, we assume that all the other constants are normalized to be the unity. In this case, the
adjoint system (1.7) reads:
−∂t ϕ − div (Me (x)∇ϕ) + a(x, t)ϕ = div (Mi (x)∇ϕe )
−ε∂ ϕ − div (M (x)∇ϕ ) = 0
t e
e
ϕ
=
ϕ
=
0
e
ϕ(T ) = ϕT , ϕe (T ) = ϕe,T
in Q,
in Q,
on Σ,
in Ω.
(3.1)
We notice that, if ϕT and ϕe,T are regular enough, taking ρ(x, t) = div (Mi (x)∇ϕe (x, t)), the pair
(ϕ, ρ) satisfies:
−∂t ϕ − div (Me (x)∇ϕ) + a(x, t)ϕ = ρ
in Q,
−ε∂ ρ − div (M (x)∇ρ) = 0
in
Q,
t
(3.2)
ϕ = ρ = 0
on Σ,
ϕ(T ) = ϕT , ρ(T ) = ρT
in Ω.
We prove the Carleman inequality (2.1) using system (3.2).
Before starting the proof of the Carleman inequality, let us first define several weight functions which
will be usefull in the sequel.
Lemma 3.1. Let ω0 be an arbitrary nonempty open set such that ω0 ⊂ ω ⊂ Ω. There exists a function
ψ ∈ C 2 (Ω) such that
ψ(x) > 0, ∀x ∈ Ω, ψ ≡ 0 on ∂Ω, |∇ψ(x)| > 0 ∀x ∈ Ω\ω0 .
Proof. See [14].
Using Lemma 3.1, we introduce the weight functions
φ(x, t) =
eλ(ψ(x)+m||ψ||)
eλm||ψ||
; φ∗ (t) = min φ(x, t) =
;
t(T − t)
t(T − t)
x∈Ω
eλ(ψ(x)+m||ψ||) − e2λmkψk
eλ(m+1)kψk − e2λmkψk
; α∗ (t) = max α(x, t) =
,
t(T − t)
t(T − t)
x∈Ω
for a parameter λ > 0 and a constant m > 1. Here,
α(x, t) =
(3.3)
(3.4)
kψ(x)k = max |ψ(x)| .
x∈Ω
∗
Remark 3.2. From the definition of α and α it follows that, for λ large enough, 3α∗ ≤ 2α. Moreover,
φ∗ (t) ≤ φ(x, t) ≤ eλkψk φ∗ (x, t)
and
|∂t α∗ | ≤ e2λkψk T φ2 .
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
7
Proof of Theorem 2.1 . For a better comprehension, we divide the proof into several steps.
Step 1. First estimate for the parabolic system.
In this step we obtain a first Carleman estimate for the adjoint system (1.7). For that, we will apply
sharp Carleman inequalities, with respect to ε, to the system and get a global estimate of ϕ and ρ in
terms of a local integral of ϕ and another in ρ.
We consider a set ω1 such that ω0 ⊂⊂ ω1 ⊂⊂ ω and apply the sharp Carleman inequality (6.2), with
ε = 1, and (6.15) to ϕ and ρ, respectively. We get
ZZ
ZZ
N
X
2
2
∂x2i xj ϕ dxdt
s−1 φ−1 e2sα |ϕt | dxdt + s−1
φ−1 e2sα
Q
Q
+ s3 λ4
ZZ
i,j=1
ZZ
2
2
φ3 e2sα |ϕ| dxdt + sλ2
φe2sα |∇ϕ| dxdt
Q
Q
Z Z
ZZ
2
2
2
2sα
3 4
3 2sα
≤C
e (|ρ| + |ϕ| )dxdt + s λ
φ e
|ϕ| dxdt
Q
(3.5)
Qω1
and
ZZ
2
e2sα |∂t ρ| dxdt + ε−2
ZZ
Q
N
X
e2sα
Q
2
∂x2i xj ρ dxdt
i,j=1
ZZ
2
2
φ4 e2sα |ρ| dxdt + s2 λ2 ε−2
φ2 e2sα |∇ρ| dxdt
Q
Q
ZZ
2
λkψk 4 4 −2
4 2sα
≤ Ce
s λ ε
φ e
|ρ| dxdt,
+ s4 λ4 ε−2
ZZ
(3.6)
Qω1
2/3
kakL∞ )T 2 )s0
for s ≥ (T + (1 +
and λ ≥ λ0 .
Adding (3.5) and (3.6), and absorbing the lower order terms in the right-hand side, we get
ZZ
ZZ
N
X
2
2
−1 2sα
−1 2sα
φ e
|ϕt | dxdt +
φ e
∂x2i xj ϕ dxdt
Q
Q
4 4
ZZ
+s λ
3 2sα
φ e
2
i,j=1
2 2
ZZ
Q
+ ε2
ZZ
Q
2
e2sα |∂t ρ| dxdt +
ZZ
Q
+ s4 λ4
e2sα
Q
ZZ
2
φ4 e2sα |ρ| dxdt + s2 λ2
Q
≤C e
2
φe2sα |∇ϕ| dxdt
|ϕ| dxdt + s λ
N
X
2
(3.7)
∂x2i xj ρ dxdt
i,j=1
ZZ
2
φ2 e2sα |∇ρ| dxdt
Q
λ||ψ|| 4 4
ZZ
s λ
4 2sα
φ e
Qω1
2
4 4
ZZ
3 2sα
|ρ| dxdt + s λ
φ e
|ϕ| dxdt ,
2
Qω1
2/3
for s ≥ (T + (1 + kakL∞ )T 2 )s0 and λ ≥ λ0 .
Remark 3.3. If we were trying to drive the solution of (1.6) to zero by means of controls on both
equations, inequality (3.7) would be sufficient.
Step 2. Estimate of the local integral of ρ.
8
F. W. CHAVES-SILVA AND M. BENDAHMANE
In this step we estimate the local integral on ρ in the right-hand side of (3.7). This will be done using
equation (3.2)1 . Indeed, we consider a function ξ satisfying
ξ ∈ C0∞ (ω), 0 ≤ ξ ≤ 1, ξ(x) = 1 ∀x ∈ ω1
and write
Ceλ||ψ|| s4 λ4
ZZ
ZZ
e2sα φ4 |ρ|2 ξdxdt = Ceλ||ψ|| s4 λ4
Qω
e2sα φ4 ρ(−ϕt − div (Me ∇ϕ) + aϕ)ξdxdt
Qω
(3.8)
:= E + F + G.
In the sequel, we estimate each parcel in the expression above. First, we have
ZZ
E = Ceλ||ψ|| s4 λ4
s∂t αe2sα φ4 ρϕξdxdt
Qω
ZZ
ZZ
+ Ceλ||ψ|| s4 λ4
e2sα φ3 φt ρϕξdxdt + Ceλ||ψ|| s4 λ4
e2sα φ4 ∂t ρϕξdxdt
Qω
Qω
:= E1 + E2 + E3 ,
and it is not difficult to see that
E1 + E2 ≤
1 4 4
s λ
10
ZZ
e2sα φ4 |ρ|2 dxdt + Ce2λ||ψ|| s8 λ4
ZZ
Qω
e2sα φ8 |ϕ|2 dxdt
Qω
and
ε2
E3 ≤
2
ZZ
e
2sα
2λ||ψ|| −2 8 8
2
|∂t ρ| dxdt + Ce
ε
Qω
ZZ
s λ
e2sα φ8 |ϕ|2 dxdt.
Qω
Next, integrating by parts, we get
e−λ||ψ|| s−4 λ−4 F =
N ZZ
X
i,j=1
+
s∂xi α e2sα φ4 ρ(Meij ∂xj ϕ)ξdxdt
Qω
N ZZ
X
i,j=1
+
N ZZ
X
i,j=1
+
e2sα φ3 ∂xi φ ρ(Meij ∂xj ϕ)ξdxdt
Qω
e2sα φ4 ∂xi ρ (Meij ∂xj ϕ)ξdxdt
Qω
N ZZ
X
i,j=1
e2sα φ4 ρ(Meij ∂xj ϕ)∂xi ξdxdt
Qω
and we can show that
F ≤
ZZ
1
e2sα φ4 |ρ|2 dxdt + s2 λ2
e2sα φ2 |∇ρ|2 dxdt
6
Qω
Qω
ZZ
ZZ
N
X
2
1
e2sα
+ Ce2λ||ψ|| s8 λ8
e2sα φ8 |ϕ|2 dxdt +
∂x2i xj ρ dxdt.
2
Qω
Qω
i,j=1
1 4 4
s λ
10
ZZ
(3.9)
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
9
Finally, we have
1
G ≤ s4 λ4
10
ZZ
e2sα φ4 |ρ|2 dxdt
Qω
2λ||ψ|| 4 4
+ Ce
s λ
||a||2L∞
ZZ
e2sα φ4 |ϕ|2 dxdt.
Qω
Putting E, F and G together in (3.7), we obtain
ZZ
ZZ
N
X
2
e2sα |ϕt |2 dxdt +
e2sα
∂x2i xj ϕ dxdt
Q
Q
4 4
ZZ
4 2sα
+s λ
φ e
i,j=1
2
2 2
ZZ
2 2sα
|ϕ| dxdt + s λ
Q
φ e
2
|∇ϕ| dxdtε
Q
N
X
e2sα
+
ZZ
Q
2
∂x2i xj ρ dxdt + s4 λ4
i,j=1
2
ZZ
e2sα |∂t ρ|2 dxdt
Q
ZZ
φ4 e2sα |ρ|2 dxdt
(3.10)
Q
ZZ
+ s2 λ2
φ2 e2sα |∇ρ|2 dxdt
Q
ZZ
e2sα φ8 |ϕ|2 dxdt,
≤ Ce2λ||ψ|| ε−2 s8 λ8
Qω
2/3
kakL∞ )T 2 )s0
for s ≥ (T + (1 +
and λ ≥ λ0 .
Using (3.10), we can prove that, for every ε > 0, system (1.6) is null controllable. However, the
sequence of controls obtained in this way will not be bounded when ε → 0+ . Therefore, we need to go a
step further and improve estimate (3.10). This is the goal of the next step.
Step 3. Weighted energy inequality.
The reason why we do not get a bounded sequence of controls out of step 2 is because of the term ε−2
in the right-hand side of (3.10). In this step we prove a weighted energy inequality for equation (3.2)2 ,
which will be used to compensate this ε−2 term.
Let us introduce the function
∗
3
y = e 2 sα ρ.
This new function satisfies
∗
3
ε∂t y − div (M (x)∇y) = ε 32 s∂t α∗ e 2 sα ρ
in Q,
(3.11)
y=0
on Σ,
y(0) = y(T ) = 0
in Ω.
Multiplying (3.11) by y and integrating over Ω, we get
Z
∗
3
ε d
3
2
2
||y(t)||L2 (Ω) + C||∇y(t)||L2 (Ω) ≤ ε
s∂t α∗ (t)e 2 sα (t) ρ(t)y(t)dx.
2 dt
2 Ω
Integrating this last inequality form 0 to T and using Poincaré’s and Young’s inequalities, it is not difficult
to see that
ZZ
ZZ
3sα∗
2
2 4λ||ψ||
e
|ρ| dxdt ≤ Cε e
s4 φ4 e2sα |ρ|2 dxdt.
(3.12)
Q
Q
Finally, from (3.10) and (3.12), we obtain
ZZ
ZZ
∗
e3sα |ρ|2 dxdt ≤ Ce6λ||ψ|| s8 λ4
Q
Qω
φ8 e2sα |ϕ|2 dxdt.
(3.13)
10
F. W. CHAVES-SILVA AND M. BENDAHMANE
This last estimate gives a global estimate of ρ in terms of a local integral of φ, with a constant C which
is bounded with respect to ε.
Step 4. Last estimates and conclusion.
In order to finish the proof of Theorem 2.1, we combine inequality (3.13) and a slightly different Carleman
inequality to the equation (3.2)1 . Indeed, the following Carleman inequality holds:
ZZ
ZZ
N
X
2
∂x2i xj ϕ dxdt
s−1 φ−1 e3sα |ϕt |2 dxdt + s−1
φ−1 e3sα
Q
Q
i,j=1
ZZ
ZZ
+ s3 λ4
φ3 e3sα |ϕ|2 dxdt + sλ2
φe3sα |∇ϕ|2 dxdt
Q
Q
Z Z
ZZ
3sα
2
3 4
3 3sα
2
≤C
e |ρ| dxdt + s λ
φ e |ϕ| dxdt ,
Q
(3.14)
Qω
2/3
kakL∞ )T 2 )s0
for s ≥ (T + (1 +
and λ ≥ λ0 , where ϕ is, together with ρ, solution of (3.2).
Notice that here we have just changed the weight e2sα by e3sα . The proof of (3.14) is exactly the same
as the proof of Theorem 6.1, just taking the appropriate change of variable in (6.3).
∗
Next, since e3sα ≤ e3sα , we have
ZZ
ZZ
∗
e3sα |ρ|2 dxdt ≤
e3sα |ρ|2 dxdt
and by (3.13), we have that
ZZ
e
3sα
Q
Q
2
6λ||ψ|| 8 4
|ρ| dxdt ≤ Ce
Q
ZZ
φ8 e2sα |ϕ|2 dxdt.
s λ
Qω
From (3.13) and (3.14), it follows that
ZZ
ZZ
ZZ
3sα
2
3 4
3 3sα
2
6λ||ψ|| 8 4
e |ρ| dxdt + s λ
φ e |ϕ| dxdt ≤ Ce
s λ
Q
Q
φ8 e2sα |ϕ|2 dxdt,
(3.15)
Qω
which is exactly (2.1).
By density, we can show that (3.15) remains true when we consider initial data in L2 (Ω). Therefore,
the proof of Theorem 2.1 is finished.
4. Null controllability for the linearized system
This section is devoted to proving the null controllability of linearized equation (1.6). It will be done
by showing the observability inequality (1.8) for the adjoint system (1.7), and solving a minimization
problem. The arguments used here are classical in control theory for linear PDE’s. Hence, we just give
a sketch of the proof.
Proof of Theorem 2.3. Combining the standard energy inequalities for system (3.2) and the Carleman
inequality given by Theorem 2.1, we can show the following observability inequality for the solutions of
(3.2):
ZZ
2/3
||ϕ(0)||2L2 (Ω) + ε||ρ(0)||2L2 (Ω) ≤ eC(1+1/T +||a||L∞ +||a||L∞ T )
|ϕ|2 dxdt,
Qω
where C = C(Ω, ω) is a positive constant.
Next, since ρ(x, t) = div (M (x)∇ϕe (x, t)) and ϕe = 0 on ∂Ω, we have that
||ϕe (t)||H 2 (Ω) ≤ C||ρ(t)||L2 (Ω) ,
(4.1)
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
11
for all t ∈ [0, T ]. Therefore, it follows from (4.1) that
2/3
||ϕ(0)||2L2 (Ω) + ε||ϕe (0)||2L2 (Ω) ≤ eC(1+1/T +||a||L∞ +||a||L∞ T )
ZZ
|ϕ|2 dxdt,
(4.2)
Qω
which is the observability inequality (1.8).
From (4.2) and the density of smooth solutions in the space of solutions of (3.1) with initial data in
L2 (Ω), we see that the above observability inequality is satisfied by all solutions of (1.7) with initial data
in L2 (Ω).
Now, in order to obtain the null controllability for linear system (1.6), we solve, for any δ > 0, the
following minimization problem:
Minimize Jδ (ϕT , ϕe,T ), with
( Z Z
1 T
2
|ϕε | dx dt + ε(ue,0 , ϕεe (0))
Jδ (ϕT , ϕe,T ) =
2 0 ω
(4.3)
)
+ (v0 , ϕε (0)) + δ(kϕT kL2 (Ω) + ε1/2 kϕe,T kL2 (Ω) ) ,
where (ϕ, ϕe ) is the solution of the adjoint problem (1.7) with initital data (ϕT , ϕe,T ) ∈ L2 (Ω)2 .
It is an easy matter to check that Jδ is strictly convex and continuous. So, in order to guarantee the
existence of a minimizer, the only thing remaining to prove is the coercivity of Jδ .
Using the observability inequality (1.8) for the adjoint system (1.7), the coercivity of Jδ is straightfoward. Therefore, for each δ > 0, there exists a unique minimizer (ϕδe,T , ϕδT ) of Jδ . Let us denote by
ϕε,δ the corresponding solution to (1.7) associated to this minimizer. Taking f ε,δ = ϕε,δ 1ω as a control
for (1.6), the duality between (1.6) and (1.7) gives
||v ε,δ (T )||L2 (Ω) + ε1/2 ||uε,δ
e (T )||L2 (Ω) ≤ δ,
ε,δ
where (v ε,δ , uε,δ
. It also gives
e ) is the solution of (1.6) associated to the control f
||f ε,δ 1ω ||2L2 (Q) ≤ C ||v0 ||2L2 (Ω) + ε||ue,0 ||2L2 (Ω) .
(4.4)
(4.5)
From (4.4) and (4.5), we get a control f ε (the weak limit of a subsequence of f ε,δ 1ω in L2 (ω × (0, T )))
that drives the solution of (1.6) to zero at time T . From (4.5), we have the following estimate on the
control f ε ,
||f ε 1ω ||2L2 (Q) ≤ C ||v0ε ||2L2 (Ω) + ε||uεe,0 ||2L2 (Ω) .
(4.6)
This finishes the proof of Theorem 2.3.
5. The nonlinear system
In this section we prove Theorem 2.4. The proof is achieved through fixed point arguments.
Proof of Theorem 2.4 (case 1): We consider the following linearization of system (1.5):
µ
ε
ε
+ g(z)v ε = f ε 1ω
in Q,
cm ∂t v − µ+1 div Me (x)∇v
ε∂ uε − div M(x)∇uε = div M (x)∇v ε
in Q,
t e
i
e
ε
ε
v = ue = 0
on Σ,
ε
ε
v (0) = v0 , ue (0) = ue,0
in Ω,
(5.1)
12
F. W. CHAVES-SILVA AND M. BENDAHMANE
where
h(s)
, if |s| > 0,
g(s) =
s
h0 (0), if s = 0.
(5.2)
It follows from Theorem 2.3 that, for each (v0 , ue,0 ) ∈ L2 (Ω)2 and z ∈ L2 (Q), there exists a control
function f ε ∈ L2 (Q) such that the solution of (5.1) satisfies:
v ε (T ) = uεe (T ) = 0.
As we said before, the idea is to use a fixed point argument. For that, we need the following generalized
version of Kakutani’s fixed point Theorem, due to Glicksberg [17].
Theorem 5.1. Let B be a non-empty convex, compact subset of a locally convex topological vector space
X. If Λ : B −→ B is a convex set-valued mapping with closed graph and Λ(B) is closed, then Λ has a
fixed point.
In order to apply Glicksberg‘s Theorem, we define a mapping Λ : B −→ X as follows
Λ(z) = {v ε ; (v ε , uεe ) is a solution of (5.1), such that v ε (T ) = uεe (T ) = 0,
for a control f ε satisfying (2.2)}.
Here, X = L2 (Q) and B is the ball
B = {z ∈ L2 (0, T, H01 (Ω)), ∂t z ∈ L2 (0, T, H −1 (Ω));
||z||2L2 (0,T ;H 1 (Ω)) + ||∂t z||2L2 (0,T ;H −1 (Ω)) ≤ M }.
0
It is easy to see that Λ is well defined and that B is a convex and compact subset of L2 (Q).
Let us now prove that Λ is convex, compact and has closed graph.
• Λ(B) ⊂ B.
Let z ∈ B and v ε ∈ Λ(z). Since v ε satisfies (5.1)1 , the following inequality holds
||v ε ||2L2 (0,T ;H 1 (Ω)) + ||∂t v ε ||2L2 (0,T ;H −1 (Ω)) ≤ K1 .
0
(5.3)
In this way, if z ∈ B then Λ(z) ⊂ B, if we take M = K1 .
• Λ(z) is closed in L2 (Q).
Let z ∈ B fixed, and vnε ∈ Λ(z), such that vnε → v ε . Let us prove that v ε ∈ Λ(z).
In fact, by definition we have that vnε is, together with a function uεe,n , and a control fnε , the solution of
(5.1), with ||fnε 1ω ||2L2 (Q) ≤ C ||v0 ||2L2 (Ω) + ε||ue,0 ||2L2 (Ω) . Therefore, we can extract a subsequence of fnε ,
denoted by the same index, such that
fnε 1ω → f ε 1ω weakly in L2 (Q).
Since fnε is bounded, we can argue as in the previous section and show that
||vnε ||2L2 (0,T ;H 1 (Ω)) + ||∂t vnε ||2L2 (0,T ;H −1 (Ω)) ≤ M.
0
Hence,
vnε → v ε weakly in L2 (0, T ; H01 (Ω)),
vnε → v ε strongly in L2 (Q),
∂t vnε → ∂t v ε weakly in L2 (0, T ; H −1 (Ω)).
(5.4)
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
13
Using the converges above and (5.1)2 , we see that there exists a function uεe such that
uεe,n → uεe weakly in L2 (0, T ; H01 (Ω)),
uεe,n → uεe strongly in L2 (Q),
∂t uεe,n → ∂t uεe weakly in L2 (0, T ; H −1 (Ω)).
It follows that (v ε , uεe ) is a controlled solution of (5.1) associated to the control f ε . Hence, v ε ∈ Λ(z) and
Λ(z) is closed and compact in L2 (Q).
• Λ has closed graph in L2 (Q) × L2 (Q).
We need to prove that if zn → z, vnε → v ε strongly in L2 (Q) and vnε ∈ Λ(zn ), then v ε ∈ Λ(z). Using
the two previous steps, it is easy to show that v ε ∈ Λ(z).
Therefore, we can apply Glicksberg’s Theorem to conclude that Λ has a fixed point. This proves
Theorem 2.4 in the case where the nonlinearity is a C 1 globally Lipschitz function.
Proof of Theorem 2.4 (case 2): We consider the linear system:
µ
div Me (x)∇v ε + a(z)v ε = f ε 1ω
cm ∂t v ε − µ+1
ε∂ uε − div M(x)∇uε = div M (x)∇v ε
i
t e
e
ε
ε
v
=
u
=
0
e
ε
v (0) = v0 , uεe (0) = ue,0
2
with (v0 , ue,0 ) ∈ W 2/3,6 (Ω) ∩ H01 (Ω) , z ∈ L∞ (Q) and
Z 1
dh
a(z) =
(sz)ds.
0 dz
in Q,
in Q,
on Σ,
in Ω,
(5.5)
Arguing as in the proof Theorem 2.3, we can show the null controllability of (5.5) with controls in
L2 (ω × (0, T )). However, these L2 controls are not sufficient to apply fixed point arguments and obtain
the null controllability of the nonlinear system (1.5). For this reason, we modify a little the functional
(4.3), obtaining controls which will allow us to employ Schauder’s fixed point Theorem. Indeed, for any
δ > 0, we consider the problem:
Minimize Jδ (ϕT , ϕeT ), with
( Z Z
1 T
2
Jδ (ϕT , ϕeT ) =
e2sα φ8 |ϕε | dx dt + ε(ue,0 , ϕεe (0))
2 0 ω
(5.6)
)
ε
1/2
+ (v0 , ϕ (0)) + δ kϕT kL2 (Ω) + kϕe,T kL2 (Ω)
,
where (ϕε , ϕεe ) is the solution of the adjoint system (1.7) with initital data (ϕT , ϕe,T ) ∈ L2 (Ω)2 .
ε,δ
As in section 4, we show that problem (5.6) has a unique minimizer (ϕε,δ , ϕε,δ
=
e ). Defining f
ε,δ
2sα 8 ε,δ
ε,δ
ε,δ
e φ ϕ and using the fact that ϕ is, together with a ϕe , the solution of (1.7), we see that f is a
solution of a parabolic equation, with homogeneous Dirichlet boundary conditions, null initial data and
a right-hand side in L2 (Q). Hence, f ε,δ ∈ L2 (0, T ; H 2 (Ω)) ∩ H 1 (0, T ; L2 (Ω)) and in particular we have
that f ε,δ ∈ L6 (Q).
Arguing as in [2, Lemma 5], we can easily show that
2
2
||f ε,δ 1ω ||2L6 (Q) ≤ C kv0 kL2 (Ω) + ε kue,0 kL2 (Ω) ,
(5.7)
where C > 0 is independent of and δ.
14
F. W. CHAVES-SILVA AND M. BENDAHMANE
ε,δ
Moreover, the solution (v ε,δ , uε,δ
, satisfies
e ) of (5.5), associated to f
||v ε,δ (T )||L2 (Ω) + ε1/2 ||uε,δ
e (T )||L2 (Ω) ≤ δ.
+
ε
(5.8)
6
Taking the limit when δ → 0 , we get a control f ∈ L (Q) (the weak limit of a subsequence of f ε,δ )
such that the associated solution (v ε , uεe ) to (5.5) satisfies
v ε (T ) = uεe (T ) = 0.
Next, we define a map F : L∞ (Q) −→ L∞ (Q) which, to each z ∈ L∞ (Q), associates v ε the solution,
together with uεe , of (5.5) corresponding to z and to the control f ε built above. Note that this application
is well defined since, from the regularity theory for parabolic equations (see, for instance, [28]), we have
that v ε ∈ X := L6 (0, T ; W 2,6 (Ω)) ∩ W 1,6 (0, T ; L6 (Ω)) ∩ C([0, T ]; W 2/3,6 (Ω) ∩ H01 (Ω)).
Let us now consider the set A defined as
A := z ∈ L∞ (Q); ||z||L∞ (Q) ≤ 1 .
It is clear that A is a convex closed subset of L∞ (Q). From (5.7), which still holds for f ε , and the
smallness assumption on the initial data, we can easily show that F is continuous and that F (A) ⊂ A.
Finally, since the space X is compactly embedded in L∞ (Q), we have that F (A) is compact in L∞ (Q).
Therefore, F has a fixed point and the the proof of case 2 of Theorem 2.4 is finished.
6. Appendix: Some technical results
In this section we prove the two sharp Carleman inequalities used in the proof of Theorem 2.1.
We consider the parabolic equation
N
X
∂xi (aij (x)∂xi v(t, x)) = g(x, t)
in Q,
−∂t v(x, t) −
i,j=1
v=0
v(T ) = vT
(6.1)
on Σ,
in Ω,
where vT ∈ L2 (Ω) and g ∈ L2 (Q).
We assume that the matrix aij has the form
aij =
Mij
,
ε
and (Mij )ij is an elliptic matrix, i.e., there exists β > 0 such that
PN
i,j
Mij ξj ξi ≥ β|ξ|2 for all ξ ∈ RN .
6.1. A degenerating Carleman inequality. The first sharp Carleman inequality we prove is the
following.
Theorem 6.1. For any 0 < ≤ 1, there exist λ0 = λ0 (Ω, ω) ≥ 1 and s0 = s0 (Ω, ω) ≥ 1 such that, for
every λ ≥ λ0 and s ≥ s0 (T + T 2 ), the solution v of (6.1) satisfies
ZZ
ZZ
N
X
2
s−1 φ−1 e2sα |∂t v|2 dxdt + s−1 ε−2
φ−1 e2sα
∂x2i xj v dxdt
Q
Q
i,j=1
ZZ
ZZ
+ s3 λ4 ε−2
φ3 e2sα |v|2 dxdt + sλ2 ε−2
φe2sα |∇v|2 dxdt
Q
Q
Z Z
ZZ
2sα
2
3 4 −2
3 2sα
2
≤C
e |g| dxdt + s λ ε
φ e |v| dxdt ,
Q
with C > 0 depending only on Ω, ω0 , ψ and β.
Qω
(6.2)
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
15
Proof. For s > 0 and λ > 0, we consider the change of variable
w(t, w) = esα v(t, w),
(6.3)
which implies
w(T, x) = w(0, x) = 0.
Using the fact that v is the solution of (6.1), we write
L1 w + L2 w = gs ,
(6.4)
where
N
X
L1 w = −∂t w + 2sλ
2
φaij ∂xj ψ ∂xi w + 2sλ
i,j=1
L2 w = −
N
X
N
X
φaij ∂xi ψ ∂xj ψw,
(6.5)
i,j=1
∂xi (aij ∂xj w) − s2 λ2
i,j=1
N
X
φ2 aij ∂xi ψ ∂xj ψ w + s∂t α w
(6.6)
i,j=1
and
gs = esα g + sλ2
N
X
φaij ∂xi ψ ∂xj ψ w − sλ
i,j=1
N
X
φ∂xi (aij ∂xj ψ) w.
(6.7)
i,j=1
From (6.4), we have that
||L1 w||2L2 (Q) + ||L2 w||2L2 (Q) + 2(L1 w, L2 w)L2 (Q) = ||gs ||2L2 (Q) .
(6.8)
The rest of the proof is devoted to analyze the terms appearing in (L1 w, L2 w)L2 (Q) .
First, we write
N
X
(L1 w, L2 w)L2 (Q) =
Iij ,
i,j=1
where Iij is the inner product in L2 (Q) of the ith term in the expression of L1 w and the jth term in
L2 w.
After a long, but straightforward, calculation, we can show that the following estimate holds
ZZ
ZZ
2(L1 w, L2 w)L2 (Q) ≥ 2s3 λ4 β 2 ε−2
φ3 |∇ψ|4 |w|2 dxdt + 2sλ2 β 2 ε−2
φ|∇ψ|2 |∇w|2 dxdt
Q
Q
Z Z
−2
2 2 4
2 2
2
3 3
2
− Cε
T s λ + Ts λ + T s + s λ + Ts λ
φ3 |w|2 dxdt
Q
ZZ
−2
2
2
− Cε (sλ + λ )
φ|∇w| dxdt.
(6.9)
Q
We take λ ≥ λ0 and s ≥ s0 (T + T 2 ), and it follows, from Remark 6.2 below, that
2(L1 w, L2 w)
L2 (Q)
3 4 2 −2
ZZ
φ3 |w|2 dxdt
+ 2s λ β ε
Qω0
2 2 −2
ZZ
φ|∇w|2 dxdt
+ 2sλ β ε
Qω0
≥ 2s3 λ4 β 2 ε−2
ZZ
φ3 |w|2 dxdt + 2sλ2 β 2 ε−2
Q
ZZ
φ|∇w|2 dxdt.
Q
(6.10)
16
F. W. CHAVES-SILVA AND M. BENDAHMANE
Remark 6.2. Since Ω\ω0 is compact and |∇ψ| > 0 on Ω\ω0 , there exists δ > 0 such that
β|∇ψ| ≥ δ on Ω\ω0 .
Putting (6.10) in (6.8), we get
ZZ
φ3 |w|2 dxdt
||L1 w||2L2 (Q) + ||L2 w||2L2 (Q) + 2β −2 s3 λ4 δ 4 ε−2
Q
ZZ
+ 2sλ2 δ 2 ε−2
φ|∇w|2 dxdt
Q
ZZ
≤ ||gs ||2L2 (Q) + 2β −2 s3 λ4 δ 4 ε−2
φ3 |w|2 dxdt
(6.11)
Qω0
+ 2sλ2 δ 2 ε−2
ZZ
φ|∇w|2 dxdt.
Qω0
Now we deal with the local integral involving ∇w on the right-hand side of (6.11). To this end, we
introduce a cutt-off function ξ such that
ξ ∈ C0∞ (ω), 0 ≤ ξ ≤ 1, ξ(x) = 1 ∀x ∈ ω0 .
Using the ellipticity condition on aij , we can prove that
ZZ
βε−1
φξ 2 |∇w|2 dxdt
Qω
Z Z
ZZ
≤C
L2 wφξ 2 wdxdt + (sT + ε−1 s2 λ2 )
φ3 |w|2 dxdt
Q
Qω
ZZ
φ1/2 |∇w|ξφ1/2 wdxdt .
+ λε−1
Qω
Therefore, by Young’s inequality, we have that
ZZ
sλ2 δ 2 ε−2
φξ 2 |∇w|2 dxdt
Q
ZZ ω
ZZ
1
≤
|L2 w|2 dxdt + Cβ −2 s3 λ4 (δ 4 + δ 2 )ε−2
φ3 |w|2 dxdt.
4
Q
Qω
Thus, inequality (6.11) gives
||L1 w||2L2 (Q)
−2 3 4 −2
ZZ
φ3 |w|2 dxdt
ZZ
2 −2
+ sλ ε
φ|∇w|2 dxdt
Q
ZZ
sα
2
−2 3 4 −2
3
2
≤ C ||e g||L2 (Q) + β s λ ε
φ |w| dxdt .
+
||L2 w||2L2 (QT )
+β
s λ ε
Q
(6.12)
Qω
Let us now we use the first two terms in left-hand side of (6.12) in order to add the integrals of |∆w|2
and |wt |2 to the left-hand side of (6.12). This is done using the expressions of L1 w and L2 w. Indeed,
from (6.5) and (6.6), we have
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
ZZ
s−1 φ−1 |∂t w|2 dxdt + ε−2
Q
ZZ
s−1 φ−1
Q
N
X
17
2
∂xi (Mij ∂xj w) dxdt
i,j=1
ZZ
ZZ
(6.13)
+ s3 λ4 ε−2
φ3 |w|2 dxdt + sλ2 ε−2
φ|∇w|2 dxdt
Q
Q
ZZ
≤ C ||esα g||2L2 (Q) + s3 λ4 ε−2
φ3 |w|2 dxdt .
Qω
Using the term in |∂xi (Mij ∂xj w)|2 on the lef-hand side of (6.13) and elliptic regularity, it is easy to show
that
s−1 ε−2
ZZ
φ−1
Q
N
X
2
∂x2i xj w dxdt
ZZ
≤ C ||esα g||2L2 (Q) + s3 λ4 ε−2
φ3 |w|2 dxdt .
Qω
i,j=1
Estimate (6.13) then gives
ZZ
s−1 φ−1 |∂t w|2 dxdt + s−1 ε−2
ZZ
Q
φ−1
Q
ZZ
N
X
2
∂x2i xj w dxdt
i,j=1
ZZ
+s3 λ4 ε−2
φ3 |w|2 dxdt + sλ2 ε−2
φ|∇w|2 dxdt
Q
Q
ZZ
≤ C ||esα g||2L2 (Q) + s3 λ4 ε−2
φ3 |w|2 dxdt .
(6.14)
Qω
From (6.14) and the fact that w = esα v, we finish the proof of Theorem 6.1.
6.2. A Slightly changed Carleman inequality. Our second sharp Carleman inequality is the following.
Theorem 6.3. For any 0 < ≤ 1, there exist λ0 = λ0 (Ω, ω) ≥ 1 and s0 = s0 (Ω, ω) ≥ 1 such that, for
every λ ≥ λ0 and s ≥ s0 (T + T 2 ), the solution v of (6.1) satisfies
ZZ
e2sα |∂t v|2 dxdt + ε−2
Q
ZZ
e2sα
Q
N
X
2
∂x2i xj v dxdt
i,j=1
ZZ
ZZ
+ s4 λ4 ε−2
φ4 e2sα |v|2 dxdt + s2 λ2 ε−2
φ2 e2sα |∇ρ|2 dxdt
Q
Q
ZZ
ZZ
≤ Ceλ||ψ|| (s
φe2sα |g|2 dxdt + s4 λ4 ε−2
φ4 e2sα |v|2 dxdt),
Q
(6.15)
Qω
with C > 0 depending only on Ω, ω0 , ψ and β.
Proof. The starting point is the application of the Carleman inequality given in Theorem 6.1 to the
equation (6.1). Indeed, we have
18
F. W. CHAVES-SILVA AND M. BENDAHMANE
ε
2
ZZ
−1 −1 2sα
s
φ
e
ZZ
2
Q
Q
+ s3 λ4
N
X
s−1 φ−1 e2sα
|vt | dxdt +
2
∂x2i xj v dxdt
i,j=1
ZZ
ZZ
φ3 e2sα |v|2 dxdt + sλ2
φe2sα |∇v|2 dxdt
Q
ZZ
ZZ
2sα
2
3 4
≤ C(
e |g| dxdt + s λ
φ3 e2sα |v|2 dxdt).
(6.16)
Q
Q
Qω
1
Next, we introduce the function y(x, t) = v(x, t)(φ∗ (t)) 2 . This new function satisfies
1
ε∂t y − div (M (x)∇y) = −ε (T −2t)
in Q,
φ∗ y + ε(φ∗ (t)) 2 g
2
y=0
on Σ.
(6.17)
Applying again the Carleman inequality given by Theorem 6.1, this time for y, we obtain, for s large
enough, that
ZZ
s−1 φ−1 e2sα |∂t y|2 dxdt + ε−2
ZZ
Q
s−1 φ−1 e2sα
Q
+ s3 λ4 ε−2
N
X
2
∂x2i xj y dxdt
i,j=1
ZZ
ZZ
φ3 e2sα |y|2 dxdt + sλ2 ε−2
φe2sα |∇y|2 dxdt
Q
ZZ
ZZ
∗ 2sα
2
3 4 −2
≤ C(
φ e |g| dxdt + s λ ε
φ3 e2sα |y|2 dxdt).
(6.18)
Q
Q
Qω
From the definition of y, it is easy to show that
ZZ
ZZ
ZZ
1
s−1 φ−1 e2sα |vt (φ∗ ) 2 |2 dxdt ≤
s−1 φ−1 e2sα |∂t y|2 dxdt +
e2sα φ|y|2 dxdt.
Q
Q
(6.19)
Q
Using (6.19), inequality (6.18) becomes
ZZ
s−1 φ−1 φ∗ e2sα |vt |2 dxdt + ε−2
Q
ZZ
s−1 φ∗ φ−1 e2sα
Q
+ s3 λ4 ε−2
N
X
2
∂x2i xj v dxdt
i,j=1
ZZ
ZZ
φ3 φ∗ e2sα |v|2 dxdt + sλ2 ε−2
φφ∗ e2sα |∇v|2 dxdt
Q
Q
ZZ
ZZ
∗ 2sα
2
3 4 −2
3 ∗ 2sα
2
≤ C(
φ e |g| dxdt + s λ ε
φ φ e |v| dxdt).
Q
(6.20)
Qω
From Remark 3.2, the result follows.
Acknowledgments
This paper has been partially established during the visit of M. Bendahmane to the Basque Center
for Applied Mathematics. The authors thank professor Enrique Zuazua for various fruitful discussions
about this work and Erich Foster for a careful reading of this paper.
UNIFORM CONTROLLABILITY FOR A DEGENERATING SYSTEM
19
References
[1] F. Ammar-Khodja, A. Benabdallah, M. González-Burgos, L. de Teresa, Recents results on the controllability of linear
coupled parabolic problems: a survey, Mathematical Control and Related Fields, 1 (3)(2011), 267–306.
[2] F. Ammar-Khodja, A. Benabdallah, C. Dupaix, Null-controllability of some reaction-diffusion systems with one control
force, J. Math. Anal. Appl., 320 (2006), 928–943.
[3] F. Ammar-Khodja, A. Benabdallah, C. Dupaix, I. Kostine, Controllability to the trajectories of phase-field models by
one control force, SIAM J. Control Optim., 42 (5)(2003), 1661–1680.
[4] V. Barbu, Local controllability of the phase field system, Nonlinear Anal. Ser. A Theory Methods, 50 (2002), 363–372.
[5] M. Bendahmane, R. Bürger, R. Ruiz-Baier, A finite volume scheme for cardiac propagation in media with isotropic
conductivities, Math. Comp. Simul., 80 (2010), 1821–1840.
[6] M. Bendahmane, F. W. Chaves-Silva, Null controllability of a degenerated reaction-diffusion system in cardiac electrophysiology, C. R. Math. Acad. Sci. Paris, 350 (11-12)(2012), 587–590.
[7] M. Bendahmane, K. H. Karlsen, Analysis of a class of degenerate reaction-diffusion systems and the bidomain model
of cardiac tissue, Netw. Heterog. Media, 1 (2006), 185–218.
[8] M. Bendahmane, K. H. Karlsen, Convergence of a finite volume scheme for the bidomain model of cardiac tissue, Appl.
Numer. Math., 59 (2009), 2266–2284.
[9] P. Biler, Existence and asymptotics of solutions for a parabolic-elliptic system nonlinear non-flux boundary conditions,
Nonilinear Analysis T. M. A., 19 (1992), 1121–1136.
[10] P. Biler, W. Hebisch, T. Nadzieja, The Debye system: Existence and long time behavior of solutions, Nonlinear
Analysis T. M. A., 23 (1994), 1189–1209.
[11] P. Colli Franzone, G. Savaré, Degenerate evolution systems modeling the cardiac electric field at micro- and macroscopic level, in Evolution equations, semigroups and functional analysis (Milano, 2000), Progr. Nonlinear Differential
Equations Appl. 50, Birkhäuser, Basel, 2002, 49–78.
[12] J.-M. Coron, S. Guerrero, A singular optimal control: A linear 1-D parabolic hyperbolic example, Asymptot. Analisys,
44 (2005), 237–257.
[13] E. Fernández-Cara, E. Zuazua, Null and approximate controllability for weakly blowing up semilinear heat equations,
Ann. Inst. Henri Poincaré, Analise Non Linéaire, 17 (2000), 583–616.
[14] A. V. Fursikov, O. Yu. Imanuvilov, Controllability of Evolution Equations, Lecture Notes Series 34, Research Institute
of Mathematics, Seoul National University, Seoul, 1996.
[15] M. González-Burgos, L. de Teresa, Controllability results for cascade systems of m coupled parabolic PDEs by one
control force, Port. Math., 67 (2010), 91–113.
[16] O. Glass, A complex-analytic approach to the problem of uniform controllability of a transport equation in the vanishing
viscosity limit, J. Funct. Analysis, 358 (2010), 852–868.
[17] I. L. Glicksberg, A further Generalization of the Kakutani fixed point theorem, with applications to Nash equilibrium
points, Proc. Amer. Math. Soc., 3 (1952), 170–174.
[18] M. González-Burgos, R. Pérez-Garcı́a, Controllability results for some nonlinear coupled parabolic systems by one
control force, Asymptot. Anal., 46 (2)(2006), 123–162.
[19] S. Guerrero, Introduction to the Controllability of Partial Differential Equations, Lectures at the Pan-American Advances Studies Institute, Santiago, Chile, 2012.
[20] S. Guerrero, Null Controllability of some systems of two parabolic equations with one control force, SIAM J. Control
Optim., 46 (2007), 379–394.
[21] S. Guerrero, G. Lebeau, Singular Optimal Control for a Transport-Diffusion equation, Comm. Partial Diff. Equations,
32 (2007), 1813–1836.
[22] C. S. Henriquez, Simulating the electrical behavior of cardiac tissue using the bidomain model, Crit. Rev. Biomed.
Engr., 21 (1993), 1–77.
[23] A. L. Hodgkin, A. F. Huxley, A quantitative description of membrane current and its application to conduction and
excitation in nerve, J. Physiol., 117 (1952), 500–544.
[24] O. Yu. Imanuvilov, M. Yamamoto, Carleman estimate for a parabolic equation in a Sobolev space of negative order
and its applications, in Control of Nonlinear Distributed Parameter Systems (College Station, Tex, 1999), Lectures
Notes in Pure and Appl. Math. 218, Marcel Dekker, New York, 2001, 113-137.
[25] J. Keener, J. Sneyd, Mathematical physiology, Interdisciplinary Applied Mathematics 8, Springer-Verlag, New York,
1998.
[26] E. F. Keller, L. A. Segel, Initiation of slime mold aggregation viewed as an instability, J. Theor. Biol., 26 (1970),
399–415.
[27] K. Kunisch, M. Wagner, Optimal control of the bidomain system (I): The monodomain approximation with the RogersMcCulloch model, Nonlinear Analysis: Real World applications, 13 (2012), 1525–1550.
20
F. W. CHAVES-SILVA AND M. BENDAHMANE
[28] O. A. Ladyzhenskaja, V. A. Solonnikov, N. N. Ural’ceva, Linear and Quasilinear equations of Parabolic type, 23
(Translated from the Russian by S. Smith), Translations of Mathematical Monographs, American Mathematical Society,
Providence), 1967.
[29] A. Lopez, X. Zhang, E. Zuazua, Null Controllability of the heat equation as singular limit of the exact controllability
of dissipative wave equations, J. Math. Pures Appl., 79 (2000), 741–808.
[30] A. Lopez, E. Zuazua, Null controllability of the 1 − d heat equation as a singular limit of the controllability of damped
wave equations, C. R. Acad. Sci. Paris, 327 (1998), 753–758.
[31] C.-H. Luo, Y. Rudy, A model of the ventricular cardiac action potential: Depolarization, repolarization, and their
interaction, Circ Res., 68 (1991), 1501–1526.
[32] D. Noble, A modification of the Hodgkin-Huxley equation applicable to Purkinje fibre action and pacemaker potentials,
J. Physiol., 160 (1962), 317–352.
[33] J.-P. Puel, Controllability of Partial Differential Equations, Lectures in the Universidade Federal do Rio de Janeiro,
Rio de Janeiro, Brazil, 2002.
[34] S. Sanfelici, Convergence of the Galerkin approximation of a degenerate evolution problem in electro-cardiology, Numer.
Methods Partial Differential Equations, 18 (2002), 218–240.
[35] L. Tung, A bidomain model for describing ischemic myocardial D-C potentials, PhD thesis, MIT, 1978.
[36] M.Veneroni, Reaction-diffusion systems for the microscopic cellular model of the cardiac electric field, Math. Methods
Appl.Sci., 29 (2006),1631–1661.
[37] E. Zuazua, Exact boundary controllability of the semilinear wave equation, in Nonlinear Partial Differential Equations
and their applications, Vol. 10, Pitman Res. Notes Math. Ser., 220, Longman, Harlow, UK, 357-391, 1991.
(F. W. Chaves-Silva)
Université de Nice Sophia-Antipolis, Laboratoire Jean Dieudonné, UMR CNRS 6621, Parc Valrose, 06108 Nice
Cedex 02, France.
E-mail address: [email protected]
(M. Bendahmane)
Institut Mathématiques de Bordeaux, Université Victor Segalen Bordeaux 2, 3 ter Place de la Victoire, 33076
Bordeaux, France.
E-mail address: [email protected]
| 3cs.SY
|
State of the Art Review for Applying Computational
Intelligence and Machine Learning Techniques to
Portfolio Optimisation
Evan Hurwitz
School of Electrical and Information Engineering
University of the Johannesburg
Johannesburg, Gauteng, South Africa
[email protected]
Tshilidzi Marwala
Faculty of Engineering
University of Johannesburg
Johannesburg, Gauteng, South Africa
[email protected]
Abstract
Computational techniques have shown much promise in the field of Finance, owing to their ability to extract sense out of
dauntingly complex systems. This paper reviews the most promising of these techniques, from traditional computational
intelligence methods to their machine learning siblings, with particular view to their application in optimising the
management of a portfolio of financial instruments. The current state of the art is assessed, and prospective further work
is assessed and recommended.
Keywords: reinforcement, learning, temporal, difference, neural, network, portfolio optimisation, genetic algorithm,
genetic programming, markowitz portfolio theory, black-scholes, investment theory.
1
Introduction
This paper is to serve as a literature review for a Phd thesis to be carried out at the University of Johannesburg. This paper
is to examine the state of the art of machine learning and computational intelligence techniques to the task of portfolio
optimisation. First the background of what precisely what constitutes computational intelligence and machine learning
techniques will be explored, in addition to the successful implementations and applications of these techniques, and the
facets that make them so promising. Following on that, the problem of portfolio optimisation will be defined, including
current objections to the standard model . Current applications and research will then be evaluated, after which
limitations will be examined, both technical limitations and industry limitations. Finally, further research will be
recommended based on the findings detailed in the paper.
2
Background
The field of portfolio optimisation is a field of immense importance to the economy of a country and indeed the global
economy in general. Like it or not, the majority of stored money in the world, ranging from money in funds, in banks, in
insurance schemes and more, is stored in portfolios of some sort or another, and therefore subject to all of the risks that
financial markets pose. The management of these portfolios has therefore been a critical area of research ever since the
inception of the marketplace, and the importance of this research is only increasing as financial instruments become ever
more complex. The recent sub-prime crisis only further illustrates that our understanding of the systems we have created
need further examination and understanding if we are to navigate these waters without bouncing from reef to reef.
The fields of computational intelligence and machine learning are concerned with the modelling of systems, and in
particular with modelling systems that prove difficult, if not impossible to model by conventional mathematical means. In
the case of computational intelligence, this is done by observing and optimising hueristics based upon gathered data.
Conversely, the field of machine learning relies on learning from observation, oftentimes involving online learning, ie
learning while the process to be modelled is in action / operation. Both of these disciplines require relatively intensive
computing power, and as such are only coming into their prime now, owing to the advances that have been made in the
past two decades in computing power, and in particular said computing power’s ready availability to researchers.
3
Computational Intelligence
A number of computational intelligence techniques will be examined in this section, including their strengths, weaknesses
and typical applications. The broad purpose of these techniques is to solve either classification or regression problems in
order to model a given system. These models can be used for prediction purposes, or further than that for system control
purposes.
3.1
Genetic Algorithms (GA)
Genetic algorithms attempt to imitate the evolutionary process of genetic evolution in order to optimise a given problem
[1]. The problem is formalised by defining a fitness function, which determines a given value for the set of parameters to
be optimised in such a way that maximising the value of this fitness function returns the optimal set of parameters [1].
The Algorithm derives its name from the manner in which it optimises its parameters. This is done by generating an
initial population of possible solutions, and evaluating the fitness function for each possible solution [2]. These solutions
are then used to generate a new population (each new population termed a generation), new individual solutions being
generated by inheriting properties from its specific parents (choosing of parents can be done in many ways, as can the
inheriting of features, but invariably the individuals with better fitness values are favoured by the chosen process) [2]. The
precise choices of parent-selection-criteria and feature inheritance, as well as type of mutation (a random changing of a
given value, used to explore further into the search-space) are all optimisation parameters that can be fine-tuned to
optimise the process itself [2].
Genetic algorithms have found much success in optimising large searching systems that are difficult to quantify, and have
in particular been very successful in the solving of Scheduling and Routing problems, in addition to financial applications
[3].
A relatively recent offshoot of genetic algorithms is genetic programming [4], in which a problem-solving method, or
policy, is evaluated and optimised by means of a genetic algorithm [4]. This in particular shows much promise in the
Portfolio Optimisation realm for being able to combine multiple portfolio optimisation strategies into one coherent,
optimised strategy.
3.2
Particle-Swarm Optimisation (PSO)
Particle-swarm optimisation is another method of solving the same types of problems as genetic algorithms, based on the
paths of swarms rather than on the workings of genetics [5]. A fitness function is defined in the same way as with a GA,
but the optimisation instead follows the following formulae for the velocity and position of each particle (solution) [5]:
vi , j = c 0 vi + c1r1 (GB j − x i , j ) + c 2 r2 (LBi , j − xi , j ) + c3 r3 (LBi , j − x1, j )
(1)
xi , j = x i , j + v i , j
(2)
Where Ci are learning constants, and Ri being random vectors, typically between 0 and 1, and GB and LB represent
Global Best and Local Best respectively [5].
Owing to its similarity to GAs in terms of function, PSOs have typically been applied to similar problems as GAs [6].
Both have also been applied to Portfolio Optimisation problems in various forms [7].
3.3
Neural Networks (NN)
The fundamental building-blocks of neural networks are neurons [8]. These neurons are simply a multiple-input, singleoutput mathematical function [8]. Each neuron has a number of weights connecting inputs from another layer to itself,
which are then added together, possibly with a bias, the result of which is then passed into the neuron’s activation
function [8]. The activation function is a function that represents the way in which the neural network “thinks”. Different
activation functions lend themselves to different problem types, ranging from yes-or-no decisions to linear and nonlinear
mathematical relationships. Each layer of a neural network is comprised of a finite number of neurons. A network may
consist of any number of layers, and each layer may contain any number of neurons [8]. When a neural network is run,
each neuron in each consecutive layer sums its inputs and multiplies each input by its respective weight, and then treats
the weighted sum as an input to its activation function. The output will then be passed on as an input to the next layer,
and so on until the final output layer is reached. Hence the input data is passed through a network of neurons in order to
arrive at an output. Figure 1 illustrates an interconnected network, with 2 input neurons, three hidden layer neurons, and
two output neurons. The hidden layer and output layer neurons can all have any of the possible activation functions. This
type of neural network is referred to as a multi-layer perceptron [8], and while not the only configuration of neural
network, it is the most widely used configuration for regression-type problems [8].
Figure 1. Sample Connectionist network
Neural networks have a number of characteristic properties that mark them as ideal for modelling. Specifically, these
characteristics are [9]:
•
Universal approximators
•
Generalisation
•
Pattern Recognition
Neural Networks have been applied in a multitude of differing topologies, and to a wide variety of problems [10]. Many
attempts have been made to apply Neural Networks to the problem of predicting stock-market prices, although with very
limited success [11]. One problem that is often overlooked when applying neural networks is the implicit assumptions
that data structuring has on the underlying problem – for example, many NN stock-market attempts have assumed a
simple time-series approach, making the implicit assumption that the historical prices of a single equity are sufficient to
predict its future value [12].
4
Machine Learning
Machine learning is a subset of computational intelligence [13]. It is a field specifically focusing on allowing a system to
maximise some reward signal, and thereby learn to operate within its environment [13]. The complexity of the modelling
system can vary greatly, allowing the researcher a high degree of customisable freedom, with appropriate trade-offs
inherent to each specific choice [13].
4.1
Reinforcement Learning
Reinforcement learning is a specific form of machine learning, in which the system can only react to the reward signal it
is given, and is not told by the engineer how to react given a specific result [13]. A typical reinforcement learning
problem will have the following topology:
Figure 2. Reinforcement Learning Topology.
The reward signal R indicates the success / failure of the attempted policy’s action within the environment, as well as the
degree thereof. The learning algorithm L defines how either the policy parameters are updated based on the reward signal
R, or the environment model E is updated with respect to R. In this way, the policy P learns to exploit the environment
with which it interacts, in fact bypassing the act of modelling the system and instead directly modelling its own control
system for working within the system [13].
4.2
On-Policy Versus Off-Policy Learning
Depending on the level of generalisation built into the model, a reinforcement learning system may learn either on-policy
or off-policy – that is, the learning may be specific to the policy tested in order to acquire a given reward signal, or
alternatively the learning may generalise to situations beyond the specific policy that has been used. The advantage of offpolicy learning is that one can achieve far swifter learning, since every single scenario need not be encountered in order to
learn to handle it [14]. The downside of off-policy learning is that learning can have unintended consequences since it
applies to scenarios beyond that which has been observed [14]. In addition, models and learning algorithms both require
greater complexity in order to successfully learn off-policy [14].
4.3
Tabular Learning / Dynamic Programming
This form of reinforcement learning is the simplest form. Each scenario is evaluated, and iteratively updated using an
appropriate equation (such as the Bellman Equation) [14], in order to create a perfect model of the environment [14].
From this, a policy can be constructed to take advantage of the environment, picking the choice with the highest expected
return.
4.4
Temporal Difference Learning
Reinforcement learning involves the training of an artificial intelligence system by means of trial-and-error, reflecting the
same manner of learning that living beings exhibit [14]. Reinforcement learning is very well suited to episodic tasks [14].
In reference to figure 2, Policy P would be represented by some appropriate function approximator, which would then be
updated according to the chosen learning algorithm [14]. This methodology allows for real-time learning, and also
eliminates the need for expert knowledge [14].
Recent work by Sutton, Svepesvari and Maei [15] has proven an algorithm for off-policy learning with a linear function
approximator using temporal difference learning and allows for linear complexity in memory and per-time-step
computation [15]. Current learning algorithms still struggle to remain stable when dealing with nonlinear function
approximators [16] [17].
5
Portfolio Theory
Portfolio theory deals with the problem of allocating funds within a portfolio [18]. This problem is generalised to include
any defined universe of financial instruments [18]. The goal of the portfolio optimisation problem is to achieve maximum
return on investment for the minimum amount of risk [18]. Optimising this risk-reward trade-off forms the heart of
portfolio theory, which is quite appropriate as this is the key question in the portfolio management industry, namely
managing clients’ funds in such a way as to keep an appropriate level or risk while maximising the clients’ return on
investment.
5.1
Modern Portfolio Theory (MPT)
Centered around Markowitz’ theories [18], this forms the central tenet of modern investment theory [18]. Price moves
within the market are assumed to be independent, stochastic events [18] and are then treated as random-walk problems
[18] (in some literature referred to instead as Markov Decision Processes) following typical Brownian Motion. With this
model, portfolios are then examined through a statistical framework, modelled only by two parameters, namely their mean
return and their standard deviation, the latter of which is interpreted as the inherent riskiness of the examined asset [18].
By quantifying assets according to these two parameters, assets and strategies can be compared on a risk-adjusted reward
basis, with efficient strategies being the highest reward for a given amount of risk [18]. Alternatively, fund managers of a
targeted-return fund will attempt to achieve a specific return, striving for the lowest possible risk in a given portfolio.
These optimal (efficient) strategies lie along the efficient horizon [18] as indicated in Figure 3.
Figure 3. The Efficient Horizon, with strategies plotted according to return (R) and risk (σ)
Unfortunately, the assumptions within the Markowitz framework have been seen to be invalid within the investment arena
[19] [20] [21]. In particular, assumptions of independence, both to internal historical moves, and to external moves, have
been seen to be false [21]. Similarly, the assumption of Brownian Motion in price moves does not hold up to scrutiny and
in most cases grossly underestimates the extent of possible price swings [21], with the industry standard approach being to
discard such events as outliers [21], even though they occur with too high a frequency to be validly treated as such [21].
5.2
Post-Modern Portfolio Theory
If one accepts the premise, as do many investors and investment houses, that returns above the expected return are not
risky (or an indication of risk in the underlying instrument) then the use of variance as a measure of risk is in fact
inappropriate [22]. Markowitz in fact recommended the use of semivariance as a measure of risk, but was of the opinion
that the computational difficulties were too formidable [24]. This, unsurprisingly, is certainly no longer the case.
Semivariance is a measure of the downside variance of a dataset, excluding values that are above the mean from the risk
calculation [22]. With a different measure of risk formulated, one can then apply MPT as before, but now designed to only
treat losses as risks, not gains [23].
5.3
Quantifying Risk
The task of quantifying risk is one of extreme importance to the financial industry, yet also extremely difficult [25]. In the
field of portfolio management risk is specific to the amount of money one can expect to lose should the market not act
favourably [25]. The traditional measure of risk, as has been mentioned in previous sections, is the statistically measurable
standard deviation of a portfolio [22]. Application of this risk measure, however, uses the assumption of a normally
distributed price moves [21], which has been shown historically to be an inaccurate and insufficient method of measuring
risk [21]. One suggested method, recommended by Mandelbrot, is to use a fractal index as a measurement of risk, which
is conceivably more reasonable due to it possessing a scaling property more indicative of real price moves than the static
Gaussian distribution [21]. Note that this only applies to volatility risk, which does not encompass the entirety of the risk
involved in investing [21].
6
Portfolio Optimisation
Portfolio optimisation is typically done by choosing a specific strategy (i.e. a mid cap fund, targeted return, etc) and then
optimising the risk-reward ratio within the limitations of the particular strategy [26]. The strategy chosen will constrain
the available instruments for the portfolio, while thereafter the optimisation will select from those available to create a
portfolio [26]. As an example, consider a fund investing money on behalf of a pension fund. The money is required to
grow, but is legally prohibited from selling short, and is mandated by the clients to have a relatively low level of risk, on
the grounds that preserving capital is more important than superior growth. Lastly, if we assume the fund is a large fund,
it will also have a liquidity limitation, on the grounds that if the fund manager should decide an asset needs to be sold,
that buyers can be found for the amounts being held. (Small penny stocks are unlikely to be traded in the quantities
required, and hence would be disqualified from the fund’s universe of stocks). The Manager would then plot the
remaining instruments and strategies on a graph of risk versus reward (as in Figure 3) and then compile a portfolio based
upon his research in order to comply with his given mandate [26]. This process will need to be re-done on a regular basis
as the various instruments’ performance changes (changing their ratings) and as profits and losses change the
composition of the portfolio, affecting its overall risk-reward ratio.
7
Applied Combinations
Owing to the promising properties of machine learning and computational intelligence techniques, numerous attempts
have been made to apply these techniques in the investment arena, with widely varying rates of success. Predictive
methods have been attempted in numerous different formats, using genetic algorithms [27], neural networks [12] [28] and
others. Attempts have also been made to utilise these techniques for option pricing [29], as an alternative to the BlackScholes and Monte-Carlo modelling approaches that are the current industry norm. Recently, Rosen and Saunders [30]
have examined using analytical techniques for pricing collaterolised debt obligations (CDOs), although they also fall into
the old trap assuming a normally distributed data set (in this case the creditworthiness of the constituents of the portfolio).
The need for utilising systems that include the dependant nature of the market that the initial assumptions of Markowitz
denies has been brought up again by Comte [31], looking specifically at long-term memory in volatility models.
Neural network applications within finance have even been attempted as far afield as in marketing campaigns, such as the
work done in [33] where notably a neural network-based approach was used to assign discounts in a marketing campaign
in order to maximise its effectiveness. Notably, the researchers utilised machine learning to optimise the policy, but still
kept the neural network model static once it had been trained [32].
Neely [33] used a genetic programming approach to find optimum trading rules, working with a technical analaysis (TA)
basis, but found that their results were inferior to a simple buy-and-hold strategy. Allen and Karjalienen attempted a
similar approach [34], using a genetic programming to tune technical trading rules, but in this case found that their
approach actually lost money when backtested. It is notable that both approaches utilise only historical data of the actual
to-be-traded security as inputs, and thus adhere to the assumption that individual price moves are independent events (an
assumption quite vigorously challenged, as indicated earlier). Lin et al [37] attempted a similar approach, using genetic
algorithms to fine-tune the parameters in TA-based trading strategies and Filter Rules. This research was evaluated using
simulated trading in the Australian Stock Market [37]. The method was compared with a greedy algorithm providing
optimal returns for efficiency comparison. The GA proved to be far more efficient than the greedy algorithm,
accomplishing in 10 seconds what the greedy algorithm achieved in 10 hours [37]. The testing appears to also have been
on a limited selection of stocks, and thus somewhat suspect. Cura [38] in turn investigated using a particle swarm
optimisation approach, and compared it to GA, tabular search and simulated annealing aproaches, using the Markowitzmean model. While the PSO results were promising, no testing was done on real-world data [38].
More recently, Freitas et al [35] investigated using auto-regressive moving reference neural networks (AR-MRNNs) in
order to predict price moves for short-term trading. They worked within the Markowitz framework, aiming to create an
efficiently diversified portfolio, trading using the predictor based on the AR-MRNN. Not only did they manage to beat the
mean-variance model, but also in fact beat the market (something that traditional portfolio theory states is impossible,
based on the challenged assumptions detailed earlier) [35]. It is worth noting that even this highly promising result still
used a relatively simple input-output structure, which limits the ability of the network.
The necessity of adapting to changing market condition was explored by Nagayama and Yoshii [36] as they explored
forgetting appropriate data in a binary classifier, in order adapt the classifier to changing conditions by forgetting data
that was no longer relevant. The explored both active and passive agent forgetting in their trials, with promising results. It
should be noted, however, that the test data used is very limited, and the results could very well be a feature of the data.
8
Application Limitations
In the field of machine learning, it is still remarkably difficult to implement reinforcement learning when using nonlinear
function approximators, such as neural networks [16]. This limitation makes application to nonstationary systems (as
financial markets have shown themselves to be) a challenging endeavour. Sutton has recently managed to develop an
algorithm for off-policy learning using temporal difference learning with a linear function approximator that allows for
linear complexity in memory and per-time-step computation, almost but not quite enough to handle the nonlinear case of
neural networks. A further limitation that has become apparent is the inadequacy of traditional risk measurement
benchmarks, as their unrealistic assumptions taint the underlying analysis [21] [39]. The assumptions implicit in the
Markowitz model are in fact no longer convenient, but now actually a hindrance to meaningful work. Much of the
research is also out of touch with the realities of the investment world, concentrating on arbitrary measures of reward and
only paying superficial attention to the issue of risk, which in reality dominates the
9
Recommended Research
In this field, there is much scope for meaningful work that, if current tumbling financial markets are any indication, is
desperately needed. The potential for machine-learning and computational intelligence techniques to perform predictionbased and classification-based portfolio optimisation is readily apparent owing to their various properties, and the various
niches are beckoning. The development of an algorithm to perform temporal difference updates that remains stable for
nonlinear function approximators would be an invaluable advance. For the tasks of predicting; modelling and classifying
the markets, the space is open for research with far greater input space, in turn making fewer inherent assumptions about
the market that are unlikely to hold up to scrutiny. In particular, the astonishingly common assumption that historical
prices / price moves of a given financial instrument or index is sufficient to model its behaviour is ludicrous – the
assumption that all information necessary is contained within is a rigid interpretation of the efficient market hypothesis,
and one that does not bear up to scrutiny. While all information may be reflected in a price, it is obfuscated by overlying
effects, and does not account at all for longer-term dynamics in the system, which cannot be ignored if a meaningful /
useful model is to be developed. Further research into risk measurement is required, with fractal indexing being a
particularly promising field currently. Modifying traditional risk measures to make fewer assumptions about the market
also promises to be a fruitful area of research. Lastly, using computational intelligence and machine learning approaches,
an adaptive portfolio optimiser is theoretically feasible, combining multiple trading strategies to create and maintain an
optimised portfolio according to modern portfolio theory measures, while utilising objective measures of risk without
unrealistic assumptions.
References
[1] Beasley, D., Bull, D.R., Martin, R.R., 1993. An overview of genetic algorithms: part I, fundamentals. University
Computing 15, 58-69.
[2] Beasley, D., Bull, D.R., Martin, R.R., 1993. An overview of genetic algorithms: part I, fundamentals. University
Computing 15, 58-69. (GA Func Dets)
[3] Chen, S. H. Genetic Algorithms and Genetic Programming in Computational Finance. Boston, MA: Kluwer. 2002
[4] Koza, J.R., 1992. Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT
Press, Cambridge.
[5] J. Kennedy, R.C. Eberhart, Particle Swarm Optimization, in: Proc. IEEE Int’l Conf. on Neural Networks, vol. IV,
1995, pp. 1942–1948.
[6] Margarita R; C Carlos. Multi-Objective Particle Swarm Optimizers: A survery of the State of the Art. International
Journal of Computational Intelligence Research. 2006. Vol 2. Issue 3.
[7] Tunchan Cura, Particle Swarm Optimization Approach to Portfolio Optimization. Nonlinear Analysis: Real World
Applications. Elsevier. April 2008.
[8] Wesley Hines J. Matlab supplement to Fuzzy and Neural Approaches in Engineering. John Wiley & Sons Inc. New
York. 1997
[9] Miller, Sutton and Werbos. Neural Networks for Control. MIT Press. Cambridge, Massecheussettes. 1990
[10] Guoqiang Zhang, B. Eddy Patuwo, Michael Y. Hu. Forecasting with artificial neural networks: The state of the art.
Elsevier 1998. International Journal of Forecasting. Vol 1.
[11] Fernando Fernández-Rodríguez, Christian González-Martel and Simón Sosvilla-Rivero. On the profitability of
Technical Trading Rules based on Artificial Neural Networks: Evidence from the Madrid Stock Market. Elsevier 2000.
[12] Leke Betechuoh B, Marwala T. Optimization of the stock market input time-window using Bayesian neural
networks. 2005 IEEE International Conference on Service Operations and Logistics, and Informatics , Beijing , Peoples
Republic of China , August 10-12 2005, pp 883 to 888..
[13] Sutton, R. S. Learning to predict by the methods of temporal differences. Mach. Learning 3, (1988), 9-44
[14] Sutton, R. S, Barto A. G. Reinforcement learning: An Introduction, MIT press, 1998.
[15] Sutton, R. S., Szepesvari, Cs., Maei, H. R. (2009). A Convergent O(n) Algorithm for Off-policy Temporal-difference
Learning with Linear Function Approximation. In Advances in Neural Information Processing Systems 21, Vancouver,
BC. December 2008. MIT Press..
[16] Hurwitz, Marwala. Optimising Reinforcement Learning for Neural Networks. Game-On 2005, Eurosis. (pp 13-18)
[17] G. Tessauro. Temporal Difference Learning and TD-Gammon. Communications of the ACM. March 1995. Vol 38.
No. 3.
[18] H. Markowitz. Portfolio Selection: Efficient diversification of investments. B. Blackwell, Cambridge, MA, second
edition edition, 1991.
[19] B. G. Malkiel, The efficient market hypothesis and its critics, The Journal of Economic Perspectives 17 (1) (2003)
59–82.
[20] E. F. Fama, Efficient capital markets II, Journal of Finance 46 (5) (1991) 1575–1617
[21] Benoit B Mandlebrot. The (Mis)behaviour of Markets: A Fractal view of Risk, Ruin and Reward. 2004.
[22] Investment Analysis and Portfolio Management, Eighth Edition, by Frank K. Reilly, CFA and Keith C. Brown, CF.
[23] H. Konno, H. Yamazaki, Mean-absolute deviation portfolio optimization model and its application to Tokyo stock
market, Management Science 37 (1991) 519–531.
[24] Balzer, L. A. (1994). Measuring Investment Risk: A Review. Journal of Investing
[25] M. Fritelli, E. Rosazza Gianin, Putting order in risk measures, Journal of Banking and Finance 26 (2002) 1473–
1486
[26] E. J. Elton, M. J. Gruber, S. J. Brown, W. N. Goetzmann, Modern Portfolio Theory and Investment Analysis, 7th
Edition, JohnWiley & Sons, Inc., 2007.
[27] Bauer, R.J. Jr., 1994. Genetic Algorithms and Investment Strategies. Wiley, New York.
[28] Kimoto, T. Asakawa, K. Yoda, M. Takeoka, M. Stock market prediction system with modular neural networks.
International Joint Conference on Neural Networks. Vol 1. June 1990.
[29] Pires M, Marwala T. American option pricing using Bayesian multi-layer perceptrons and Bayesian support vector
machines. IEEE 3 rd International Conference on Computational Cybernetics , Mauritius , April 13-16 2005, pp 219 to
224.
[30] Dan Rosen. David Saunders. Analytical Methods for Hedging Systematic Credit Risk with Linear Factor Portfolios.
Journal of Economic Dynamics and Control. Elsevier. 2008
[31] Fabienne Comte Long MemoryIn Continuous-Time Stochastic Volatility Models.. Mathematical Finance. 1998
[32] Gabriel Gómez-Pérez a, José D. Martín-Guerrero a,*, Emilio Soria-Olivas a, Emili Balaguer-Ballester b, Alberto
Palomares b, Nicolás Casariego Assigning Discounts in a Marketing Campaign Using Reinforcement Learning and
Neural Nets. c. Expert Systems With Applications. Elsevier. 2009.
[33] Neely, Christopher J.,Using Genetic Algorithms to Find Technical Trading Rules: A Comment on Risk Adjustment
(October 27, 1999). Federal Reserve Bank of St. Louis Working Paper No. 99-015A. Available at SSRN:
http://ssrn.com/abstract=189488 or DOI: 10.2139/ssrn.189488.
[34] Franklin Allen; Risto Karjalainen Using genetic algorithms to Find technical trading rules.. Journal of Financial
Economics 51. 1999. 245-271.
[35] Fabio. D. Freitas; Alberto. F. de Souza; Ailson. R. de Almeida. Prediction-Based Portfolio Optimisation Using
Neural Networks. Neurocomputing. Accepted August 2008. Elsevier.
[36] Hirotaka Nagayama, Kengo Yoshii. Active Forgetting in Machine Learning and its Application to Financial
Problems. 2000. IEEE
[37] Li Lin, Longbing Cao, Jiaqi Wang, Chengqi Zhang. The Applications of Genetic Algorithms in Stock Market Data
Mining Optimisation. Faculty of Information Technology, University of Technology, Sydney,NSW 2007, Australia
Capital Market CRC, Sydney NSW 2000, Australia
[38] Tunchan Cura. Particle Swarm Optimization Approach to Portfolio Optimization. Nonlinear Analysis: Real World
Applications. Elsevier 2009.
[39] Naomi Miller. Andrej Ruszcynski. Risk-adjusted probability measures in portfolio optimization with coherent
measures of risk. European Journal of Operational Research. Elsevier 2008.
| 5cs.CE
|
1
Robust Large Margin Deep Neural Networks
Jure Sokolić, Student Member, IEEE, Raja Giryes, Member, IEEE,
arXiv:1605.08254v3 [stat.ML] 23 May 2017
Guillermo Sapiro, Fellow, IEEE, and Miguel R. D. Rodrigues, Senior Member, IEEE
Abstract
The generalization error of deep neural networks via their classification margin is studied in this
work. Our approach is based on the Jacobian matrix of a deep neural network and can be applied to
networks with arbitrary non-linearities and pooling layers, and to networks with different architectures
such as feed forward networks and residual networks. Our analysis leads to the conclusion that a bounded
spectral norm of the network’s Jacobian matrix in the neighbourhood of the training samples is crucial for
a deep neural network of arbitrary depth and width to generalize well. This is a significant improvement
over the current bounds in the literature, which imply that the generalization error grows with either the
width or the depth of the network. Moreover, it shows that the recently proposed batch normalization
and weight normalization re-parametrizations enjoy good generalization properties, and leads to a novel
network regularizer based on the network’s Jacobian matrix. The analysis is supported with experimental
results on the MNIST, CIFAR-10, LaRED and ImageNet datasets.
Index Terms
Deep learning, deep neural networks, generalization error, robustness.
I. I NTRODUCTION
In recent years, deep neural networks (DNNs) achieved state-of-the-art results in image recognition,
speech recognition and many other applications [1]–[4]. DNNs are constructed as a series of non-linear
J. Sokolić and M. R. D. Rodrigues are with the Department of Electronic and Electrical Engineering, Univeristy College
London, London, UK (e-mail: {jure.sokolic.13, m.rodrigues}@ucl.ac.uk).
R. Giryes is with the School of Electrical Engineering, Faculty of Engineering, Tel-Aviv University, Tel Aviv, Israel (e-mail:
[email protected]).
G. Sapiro is with the Department of Electrical and Computer Engineering, Duke University, NC, USA (e-mail:
[email protected]).
The work of Jure Sokolić and Miguel R. D. Rodrigues was supported in part by EPSRC under grant EP/K033166/1. The work
of Raja Giryes was supported in part by GIF, the German-Israeli Foundation for Scientific Research and Development. The work
of Guillermo Sapiro was supported in part by NSF, ONR, ARO, and NGA.
May 24, 2017
DRAFT
2
signal transformations that are applied sequentially, where the parameters of each layer are estimated
from the data [3]. Typically, each layer applies on its input a linear (or affine) transformation followed by
a point-wise non-linearity such as the sigmoid function, the hyperbolic tangent function or the Rectified
Linear Unit (ReLU) [5]. Many DNNs also include pooling layers, which act as down-sampling operators
and may also provide invariance to various input transformations such as translation [6], [7]. They may
be linear, as in average pooling, or non-linear, as in max-pooling.
There were various attempts to provide a theoretical foundation for the representation power, optimization
and generalization of DNNs. For example, the works in [8], [9] showed that neural networks with a single
hidden layer – shallow networks – can approximate any measurable Borel function. On the other hand, it
was shown in [10] that a deep network can divide the space into an exponential number of sets, which
can not be achieved by shallow networks that use the same number of parameters. Similarly, the authors
in [11] conclude that functions implemented by DNNs are exponentially more expressive than functions
implemented by shallow networks. The work in [12] shows that for a given number of parameters and a
given depth, there always exists a DNN that can be approximated by a shallower network only if the
number of parameters in the shallow network is exponential in the number of layers of the deep network.
Scattering transform - a convolutional DNN like transform, which is based on the wavelet transform
and pointwise non-linearities - provides insights into translation invariance and stability to deformations
of convolutional DNNs [13]–[15].
DNNs with random weights are studied in [16], where it is shown that such networks perform distance
preserving embedding of low-dimensional data manifolds. The authors in [17] model a loss function of
DNN with a spin-glass model and show that for large networks the local optima of the loss function
are close to the global optima. Optimization aspects of DNNs are studied from the perspective of tensor
factorization in [18] where it is shown that if a network is large, then it is possible to find the global minima
from any initialization with a gradient descent algorithm. The role of DNNs in improving convergence
speed of various iterative algorithms is studied in [19].
Optimization dynamics of a deep linear network is studied in [20], where it is shown that the learning
speed of deep networks may be independent of their depth. Reparametrization of DNN for more efficient
learning is studied in depth in [21]. A modified version of stochastic gradient descent for optimization of
DNNs that are invariant to weight rescaling in different layers is proposed in [22], where it is shown
that such an optimization may lead to a smaller generalization error (GE) - the difference between the
empirical error and the expected error, than the one achieved with the classical stochastic gradient descent.
The authors in [23] propose the batch normalization – a technique that normalizes the output of each
layer and leads to faster training and also a smaller GE. A similar technique based on normalization of
May 24, 2017
DRAFT
3
the weight matrix rows is proposed in [24]. It is shown empirically that such reparametrization leads
to a faster training and a smaller GE. Learning of DNN by bounding the spectral norm of the weight
matrices is proposed in [25]. Other methods for DNN regularization include weight decay, dropout [26],
constraining the Jacobian matrix of encoder for regularization of auto-encoders [27], and enforcing a
DNN to be a partial isometry [28].
An important theoretical aspect of DNNs is the effect of their architecture, e.g. depth and width, on their
GE. Various measures such as the VC-dimension [29], [30], the Rademacher or Gaussian complexities
[31] and algorithmic robustness [32] have been used to bound the GE in the context of DNNs. For
example, the VC-dimension of DNN with the hard-threshold non-linearity is equal to the number of
parameters in the network, which implies that the sample complexity is linear in the number of parameters
of the network. The GE can also be bounded independently of the number of parameters, provided that
the norms of the weight matrices (the network’s linear components) are constrained appropriately. Such
constraints are usually enforced by training networks with weight decay regularization, which is simply
the `1 - or `2 -norm of all the weights in the network. For example, the work [33] studies the GE of DNN
with ReLUs and constraints on the norms of the weight matrices. However, it provides GE bounds that
scale exponentially with the network depth. Similar behaviour is also depicted in [34]. The authors in
[32] show that DNNs are robust provided that the `1 -norm of the weights in each layer is bounded. The
bounds are exponential in the `1 -norm of the weights if the norm is greater than 1.
The GE bounds in [30], [32], [33] suggest that the GE of a DNN is bounded only if the number of
training samples grows with the DNN depth or size. However, in practice increasing network’s depth or
size often leads to a lower GE [4], [35]. Moreover, recent work in [36] shows that a 2 layer DNN with
ReLUs may fit any function of n samples in d dimensions provided that it has 2n + d parameters, which
is often the case in practice. They show that the nature of the GE depends more on the nature of the data
than on the architecture of the network as the same network is able to fit both structured data and random
data, where for the first the GE is very low and for the latter it is very large. The authors conclude that
data agnostic measures such as the Rademacher complexity or VC-dimension are not adequate to explain
the good generalization properties of modern DNN.
Our work complements the previous works on the GE of DNNs by bounding the GE in terms of the
DNN classification margin, which is independent of the DNN depth and size, but takes into account the
structure of the data (considering its covering number) and therefore avoids the issues presented above.
The extension of our results to invariant DNN is provided in [37].
May 24, 2017
DRAFT
4
A. Contributions
In this work we focus on the GE of a multi-class DNN classifier with general non-linearities. We
establish new GE bounds of DNN classifiers via their classification margin, i.e. the distance between
the training sample and the non-linear decision boundary induced by the DNN classifier in the sample
space. The work capitalizes on the algorithmic robustness framework in [32] to cast insight onto the
generalization properties of DNNs. In particular, the use of this framework to understand the operation of
DNNs involves various innovations, which include:
•
We derive bounds for the GE of DNNs by lower bounding their classification margin. The lower
bound of the classification margin is expressed as a function of the network’s Jacobian matrix.
•
Our approach includes a large class of DNNs. For example, we consider DNNs with the softmax
layer at the network output; DNNs with various non-linearities such as the Rectified Linear Unit
(ReLU), the sigmoid and the hyperbolic tangent; DNNs with pooling, such as down-sampling, average
pooling and max-pooling; and networks with shortcut connections such as Residual Networks [4].
•
Our analysis shows that the GE of a DNN can be bounded independently of its depth or width
provided that the spectral norm of the Jacobian matrix in the neighbourhood of the training samples is
bounded. We argue that this result gives a justification for a low GE of DNNs in practice. Moreover,
it also provides an explanation for why training with the recently proposed weight normalization or
batch normalization can lead to a small GE. In such networks the `2 -norm of the weight matrices
is fixed and `2 -norm regularization does not apply. The analysis also leads to a novel Jacobian
matrix-based regularizer, which can be applied to weight normalized or batch normalized networks.
•
We provide a series of examples on the MNIST, CIFAR-10, LaRED and ImageNet datasets that
validate our analysis and demonstrate the effectiveness of the Jacobian regularizer.
Our contributions differ from the existing works in many ways. In particular, the GE of DNNs has
been studied via the algorithmic robustness framework in [32]. Their bounds are based on the per-unit
`1 -norm of the weight matrices, and the studied loss is not relevant for classification. Our analysis is
much broader, as it aims at bounding the GE of 0-1 loss directly and also considers DNNs with pooling.
Moreover, our bounds are a function of the network’s Jacobian matrix and are tighter than the bounds
based on the norms of the weight matrices.
The work in [28] shows that learning transformations that are locally isometric is robust and leads
to a small GE. Though they apply the proposed technique to DNNs they do not show how the DNN
architecture affects the GE as our work does.
The authors in [25] have observed that contractive DNNs with ReLUs trained with the hinge loss lead
May 24, 2017
DRAFT
5
to a large classification margin. However, they do not provide any GE bounds. Moreover, their results
are limited to DNNs with ReLUs, whereas our analysis holds for arbitrary non-linearities, DNNs with
pooling and DNNs with the softmax layer.
The work in [27] is related to ours in the sense that it proposes to regularize auto-encoders by
constraining the Frobenious norm of the encoder’s Jacobian matrix. However, their work is more empirical
and is less concerned with the classification margin or GE bounds. They use the Jacobian matrix to
regularize the encoder whereas we use the Jacobian matrix to regularize the entire DNN.
Finally, our DNN analysis, which is based on the network’s Jacobian matrix, is also related to the
concept of sensitivity analysis that has been applied to feature selection for SVM and neural networks
[38], [39], and for the construction of radial basis function networks [40], since the spectral norm of the
Jacobian matrix quantifies the sensitivity of DNN output with respect to the input perturbation.
B. Paper organization
Section II introduces the problem of generalization error, including elements of the algorithmic robustness
framework, and introduces DNN classifiers. Properties of DNNs are described in Section III. The bounds
on the classification margin of DNNs and their implication for the GE of DNNs are discussed in Section
IV. Generalizations of our results are discussed in Section V. Section VI presents experimental results.
The paper is concluded in Section VII. The proofs are deferred to the Appendix.
C. Notation
We use the following notation in the sequel: matrices, column vectors, scalars and sets are denoted
by boldface upper-case letters (X), boldface lower-case letters (x), italic letters (x) and calligraphic
upper-case letters (X ), respectively. The convex hull of X is denoted by conv(X ). IN ∈ RN ×N denotes
the identity matrix, 0M ×N ∈ RM ×N denotes the zero matrix and 1N ∈ RN denotes the vector of ones.
The subscripts are omitted when the dimensions are clear from the context. ek denotes the k -th basis
vector of the standard basis in RN . kxk2 denotes the Euclidean norm of x, kXk2 denotes the spectral
norm of X, and kXkF denotes the Frobenious norm of X. The i-th element of the vector x is denoted by
(x)i , and the element of the i-th row and j -th column of X is denoted by (X)ij . The covering number of
X with d-metric balls of radius ρ is denoted by N (X ; d, ρ).
II. P ROBLEM S TATEMENT
We start by describing the GE in the framework of statistical learning. Then, we dwell on the GE bounds
based on the robustness framework by Xu and Manor [32]. Finally, we present the DNN architectures
studied in this paper.
May 24, 2017
DRAFT
6
A. The Classification Problem and Its GE
We consider a classification problem, where we observe a vector x ∈ X ⊆ RN that has a corresponding
class label y ∈ Y . The set X is called the input space, Y = {1, 2, . . . , NY } is called the label space and
NY denotes the number of classes. The samples space is denoted by S = X × Y and an element of S is
denoted by s = (x, y). We assume that samples from S are drawn according to a probability distribution
m
P defined on S . A training set of m samples drawn from P is denoted by Sm = {si }m
i=1 = {(xi , yi )}i=1 .
The goal of learning is to leverage the training set Sm to find a classifier g(x) that provides a label
estimate ŷ given the input vector x. In this work the classifier is a DNN, which is described in detail in
Section II-C.
The quality of the classifier output is measured by the loss function `(g(x), y), which measures the
discrepancy between the true label y and the estimated label ŷ = g(x) provided by the classifier. Here
we take the loss to be the 0-1 indicator function. Other losses such as the hinge loss or the categorical
cross entropy loss are possible. The empirical loss of the classifier g(x) associated with the training set
and the expected loss of the classifier g(x) are defined as
`emp (g) = 1/m
X
` (g(xi ), yi )
(1)
si ∈Sm
and
`exp (g) = Es∼P [` (g(x), y)] ,
(2)
respectively. An important question, which occupies us throughout this work, is how well lemp (g) predicts
lexp (g). The measure we use for quantifying the prediction quality is the difference between lexp (g) and
lemp (g), which is called the generalization error:
GE(g) = |`exp (g) − `emp (g)| .
(3)
B. The Algorithmic Robustness Framework
In order to provide bounds to the GE for DNN classifiers we leverage the robustness framework [32],
which is described next.
The algorithmic robustness framework provides bounds for the GE based on the robustness of a learning
algorithm that learns a classifier g leveraging the training set Sm :
May 24, 2017
DRAFT
7
Definition 1 ( [32]). Let Sm be a training set and S the sample space. A learning algorithm is (K, (Sm ))robust if the sample space S can be partitioned into K disjoint sets denoted by Kk , k = 1, . . . , K ,
Kk ⊆ S,
k = 1, . . . , K,
(4)
S = ∪K
k=1 Kk ,
Kk ∩ Kk0 = ∅,
(5)
∀k 6= k 0 ,
(6)
such that for all si ∈ Sm and all s ∈ S
si = (xi , yi ) ∈ Kk ∧ s = (x, y) ∈ Kk =⇒
|`(g(xi ), yi ) − `(g(x), y)| ≤ (Sm ) .
(7)
Note that si is an element of the training set and s is an arbitrary element of the sample space S .
Therefore, a robust learning algorithm chooses a classifier g for which the losses of any s and si in the
same partition Kk are close. The following theorem provides the GE bound for robust algorithms.1
Theorem 1 (Theorem 3 in [32]). If a learning algorithm is (K, (Sm ))-robust and `(g(x), y) ≤ M for
all s = (x, y) ∈ S , then for any δ > 0, with probability at least 1 − δ ,
r
2K log(2) + 2 log(1/δ)
.
GE(g) ≤ (Sm ) + M
m
(8)
The first term in the GE bound in (8) is constant and depends on the training set Sm . The second term
√
behaves as O(1/ m) and vanishes as the size of the training set Sm approaches infinity. M = 1 in the
case of 0-1 loss, and K corresponds to the number of partitions of the samples space S .
A bound on the number of partitions K can be found by the covering number of the samples space S .
The covering number is the smallest number of (pseudo-)metric balls of radius ρ needed to cover S , and it
is denoted by N (S; d, ρ), where d denotes the (pseudo-)metric.2 The space S is the Cartesian product of
a continuous input space X and a discrete label space Y , and we can write N (S; d, ρ) ≤ NY · N (X ; d, ρ),
where NY corresponds to the number of classes. The choice of metric d determines how efficiently one
may cover X . A common choice is the Euclidean metric
d(x, x0 ) = kx − x0 k2 , ,
x, x0 ∈ X ,
(9)
which we also use in this paper. The covering number of many structured low-dimensional data models
can be bounded in terms of their “intrinsic” properties, for example:
1
Additional variants of this theorem are provided in [32].
2
Note that we can always obtain a set of disjoint partitions from the set of metric balls used to construct the covering.
May 24, 2017
DRAFT
8
•
a Gaussian mixture model (GMM) with L Gaussians and covariance matrices of rank at most k
leads to a covering number N (X ; d, ρ) = L(1 + 2/ρ)k [41];
•
k -sparse signals in a dictionary with L atoms have a covering number N (X ; d, ρ) =
L
k
(1 + 2/ρ)k
[16];
•
CM regular k -dimensional manifold, where CM is a constant that captures its “intrinsic” properties,
k
[42].
has a covering number N (X ; d, ρ) = CρM
1) Large Margin Classifier: An example of a robust learning algorithm is the large margin classifiers,
which we consider in this work. The classification margin is defined as follows:
Definition 2 (Classification margin). The classification margin of a training sample si = (xi , yi ) measured
by a metric d is defined as
γ d (si ) = sup{a : d(xi , x) ≤ a =⇒ g(x) = yi ∀x} .
(10)
The classification margin of a training sample si is the radius of the largest metric ball (induced by d)
in X centered at xi that is contained in the decision region associated with class label yi . The robustness
of large margin classifiers is given by the following Theorem.
Theorem 2 (Adapted from Example 9 in [32]). If there exists γ such that
γ d (si ) > γ > 0
∀si ∈ Sm ,
(11)
then the classifier g(x) is (NY · N (X ; d, γ/2), 0)-robust.
Theorems 1 and 2 imply that the GE of a classifier with margin γ is upper bounded by (neglecting the
log(1/δ) term in (8))
1 p
2 log(2) · NY · N (X ; d, γ/2) .
GE(g) . √
m
(12)
Note that in case of a large margin classifier the constant (Sm ) in (8) is equal to 0, and the GE approaches
√
zero at a rate m as the number of training samples grows. The GE also increases sub-linearly with
the number of classes NY . Finally, the GE depends on the complexity of the input space X and the
classification margin via the covering number N (X ; d, γ/2).
For example, if we take X to be a CM regular k -dimensional manifold then the upper bound to the
GE behaves as:
May 24, 2017
DRAFT
9
x
φ1 (x, θ1 )
z1
φ2 (z1 , θ2 )
φL (zL−1 , θL )
z
Fig. 1. DNN transforms the input vector x to the feature vector z by a series of (non-linear) transforms.
Corollary 1. Assume that X is a (subset of) CM regular k -dimensional manifold, where N (X ; d, ρ) ≤
k
CM
. Assume also that classifier g(x) achieves a classification margin γ and take `(g(xi ), yi ) to be
ρ
the 0-1 loss. Then for any δ > 0, with probability at least 1 − δ ,
s
r
log(2) · NY · 2k+1 · (CM )k
2 log(1/δ)
+
GE(g) ≤
.
k
m
γ m
(13)
Proof: The proof follows directly from Theorems 1 and 2.
Note that the role of the classifier is captured via the achieved classification margin γ . If we can always
ensure a classification margin γ = 1, then the GE bound only depends on the dimension of the manifold
k and the manifold constant CM . We relate this bound, in the context of DNNs, to other bounds in the
literature in Section IV.
C. Deep Neural Network Classifier
The DNN classifier is defined as
g(x) = arg max(f (x))i ,
(14)
i∈[NY ]
where (f (x))i is the i-th element of the NY dimensional output of a DNN f : RN → RNY . We assume
that f (x) is composed of L layers:
f (x) = φL (φL−1 (· · · φ1 (x, θ1 ), · · · θL−1 ), θL ) ,
(15)
where φl (·, θl ) represents the l-th layer with parameters θl , l = 1, . . . , L. The output of the l-th layer is
denoted by zl , i.e. zl = φl (zl−1 , θl ), zl ∈ RMl ; the input layer corresponds to z0 = x; and the output of
the last layer is denoted by z = f (x). Such a DNN is visualized in Fig. 1. Next, we define various layers
φl (·, θl ) that are used in the modern state-of-the-art DNNs.
1) Linear and Softmax Layers: We start by describing the last layer of a DNN that maps the output of
previous layer into RNY , where NY corresponds to the number of classes.3 This layer can be linear:
z = ẑ,
3
ẑ = WL zL−1 + bL ,
(16)
Assuming that there are NY one-vs.-all classifiers.
May 24, 2017
DRAFT
10
TABLE I
P OINT- WISE NON - LINEARITIES
Name
Function: σ(x)
ReLU
max(x, 0)
Sigmoid
1
1+e−x
Hyperbolic tangent
tanh(x) =
Derivative:
d
dx σ(x)
Derivative bound: supx
{1 if x > 0; 0 if x ≤ 0}
σ(x)(1 − σ(x)) =
ex −e−x
ex +e−x
e−x
(1+e−x )2
1 − σ(x)2
d
dx σ(x)
≤1
≤
1
4
≤1
where WL ∈ RNY ×ML−1 is the weight matrix associated with the last layer and b ∈ RNY is the bias
vector associated with the last layer. Note that according to (14), the i-th row of WL can be interpreted
as a normal to the hyperplane that separates class i from the others. If the last layer is linear the usual
choice of learning objective is the hinge loss.
A more common choice for the last layer is the softmax layer:
z = ζ(ẑ) = eẑ / 1T eẑ , ẑ = WL zL−1 + bL ,
(17)
where ζ(·) is the softmax function and WL and bL are the same as in (16). Note that the exponential is
applied element-wise. The elements of z are in range (0, 1) and are often interpreted as “probabilites”
associated with the corresponding class labels. The decision boundary between class y1 and class y2
corresponds to the hyperplane {z : (z)y1 = (z)y2 }. The softmax layer is usually coupled with categorical
cross-entropy training objective.
For the remainder of this work we will take the softmax layer as the last layer of DNN, but note that
all results still apply if the linear layer is used.
2) Non-linear layers: A non-linear layer is defined as
zl = [ẑl ]σ = [Wl zl−1 + bl ]σ ,
(18)
where [ẑl ]σ represents the element-wise non-linearity applied to each element of ẑl ∈ RMl , and ẑl
represents the linear transformation of the layer input: ẑl = Wl zl−1 + bl . Wl ∈ RMl ×Ml−1 is the weight
matrix and bl ∈ RMl is the bias vector. The typical non-linearities are the ReLU, the sigmoid and the
hyperbolic tangent. They are listed in Table I. The choice of non-linearity σ is usually the same for all
the layers in the network.
Note that the non-linear layer in (18) includes the convolutional layers which are used in the convolutional
neural networks. In that case the weight matrix is block-cyclic.
May 24, 2017
DRAFT
11
Class 1
Class 2
0.6
Class 1
Class 2
0.7
0.4
0.6
0.2
γ d (si )
0.5
(z) 2
(x) 2
0.0
0.2
0.3
0.4
0.2
0.6
0.1
0.8
1.0
0.8
0.4
0.0
0.6
0.4
0.2 0.0
0.2
(x) 1
0.4
0.6
0.8
1.0
0.0
(a) Input space.
0.1
0.2
0.3
(z) 1
0.4
0.5
0.6
(b) Output space.
Fig. 2. Decision boundaries in the input space and in the output space. Plot (a) shows samples of class 1 and 2 and the decision
regions produced by a two-layer network projected into the input space. Plot (b) shows the samples transformed by the network
and the corresponding decision boundary at the network output.
3) Pooling layers: A pooling layer reduces the dimension of intermediate representation and is defined
as
zl = Pl (zl−1 )zl−1 ,
(19)
where Pl (zl−1 ) is the pooling matrix. The usual choices of pooling are down-sampling, max-pooling and
average pooling. We denote by pli (zl−1 ) the i-th row of Pl (zl−1 ) and assume that there are Ml pooling
regions Pi , i = 1, . . . , Ml . In the case of down-sampling pli (zl−1 ) = ePi (1) , where Pi (1) is the first element
of the pooling region Pi ; in the case of max-pooling pli (zl−1 ) = ej ? , where j ? = arg maxj 0 ∈Pi |(zl−1 )j 0 |;
P
and in the case of average pooling pli (zl−1 ) = |P1i | j∈Pi ej .
III. T HE G EOMETRICAL P ROPERTIES OF D EEP N EURAL N ETWORKS
The classification margin introduced in Section II-A is a function of the decision boundary in the
input space. This is visualized in Fig. 2 (a). However, a training algorithm usually optimizes the decision
boundary at the network output (Fig. 2 (b)), which does not necessarily imply a large classification margin.
In this section we introduce a general approach that allows us to bound the expansion of distances between
the network input and its output. In Section IV we use this to establish bounds of the classification margin
and the GE bounds that are independent of the network depth or width.
We start by defining the Jacobian matrix (JM) of the DNN f (x):
L
J(x) =
df (x) Y dφl (zl−1 ) dφ1 (x)
=
·
.
dx
dx
dzl−1
(20)
l=1
May 24, 2017
DRAFT
12
Note that by the properties of the chain rule, the JM is computed as the product of the JMs of the
individual network layers, evaluated at the appropriate values of the layer inputs x, z1 , . . . , zL−1 . We use
the JM to establish a relation between a pair of vectors in the input space and the output space.
Theorem 3. For any x, x0 ∈ X and a DNN f (·), we have
Z 1
0
f (x ) − f (x) =
J(x + t(x0 − x)) dt (x0 − x)
(21)
0
= Jx,x0 (x0 − x),
(22)
where
1
Z
Jx,x0 =
J(x + t(x0 − x)) dt
(23)
0
is the average Jacobian on the line segment between x and x0 .
Proof: The proof appears in Appendix A.
As a direct consequence of Theorem 3 we can bound the distance expansion between x and x0 at the
output of the network f (·):
Corollary 2. For any x, x0 ∈ X and a DNN f (·), we have
kf (x0 ) − f (x)k2 = kJx,x0 (x0 − x)k2
≤
sup
kJ(x00 )k2 kx0 − xk2 .
(24)
x00 ∈conv(X )
Proof: The proof appears in Appendix B.
Note that we have established that Jx,x0 corresponds to a linear operator that maps the vector x0 − x
to the vector f (x0 ) − f (x). This implies that the maximum distance expansion of the network f (x) is
bounded by the maximum spectral norm of the network’s JM. Moreover, the JM of f (x) corresponds to
the product of JMs of all the layers of f (x) as shown in (20). It is possible to calculate the JMs of all
the layers defined in Section II-C:
1) Jacobian Matrix of Linear and Softmax Layers: The JM of the linear layer defined in (16) is equal
to the weight matrix
dz
= WL .
dzL−1
(25)
Similarly, in the case of softmax layer defined in (17) the JM is
dz
dz
dẑ
=
· L−1
L−1
dz
dẑ dz
= −ζ(ẑ)ζ(ẑ)T + diag(ζ(ẑ) · WL .
(26)
Note that −ζ(ẑ)ζ(ẑ)T + diag(ζ(ẑ) corresponds to the JM of the softmax function ζ(ẑ).
May 24, 2017
DRAFT
13
2) Jacobian Matrix of Non-Linear Layers: The JM of the non-linear layer (18) can be derived in the
same way as the JM of the softmax layer. We first define the JM of the point-wise non-linearity, which is
a diagonal matrix4
dzl
dẑl
ii
dσ (ẑl )i
=
,
d(ẑl )i
i = 1, . . . , Ml .
(27)
The derivatives associated with various non-linearities are provided in Table I. The JM of the non-linear
layer can be expressed as
dzl
dzl
=
· Wl .
dẑl−1
dẑl
(28)
3) Jacobian Matrix of Pooling Layers: The pooling operator defined in (19) is linear or a piece-wise
linear operator. The corresponding JM is therefore also linear or piece-wise linear and is equal to:
Pl (zl−1 ) .
(29)
The following Lemma collects the bounds on the spectral norm of the JMs for all the layers defined in
Section II-C.
Lemma 1. The following statements hold:
1) The spectral norm of JMs of the linear layer in (16), the softmax layer in (17) and non-linear layer
in (18) with the ReLU, Sigmoid or Hyperbolic tangent non-linearities is upper bounded by
dzl
dẑl−1
≤ kWl k2 ≤ kWl kF .
(30)
2
2) Assume that the pooling regions of the down-sampling, max-pooling and average pooling operators
are non-overlapping. Then the spectral norm of their JMs can be upper bounded by
dzl
dẑl−1
≤ 1.
(31)
2
Proof: The proof appears in Appendix C.
Lemma 1 shows that the spectral norms of all layers can be bounded in terms of their weight matrices.
As a consequence, the spectral norm of the JM is bounded by the product of the spectral norms of the
weight matrices. We leverage this facts to provide GE bounds in the next section.
4
Note that in case of ReLU the derivative of max(x, 0) is not defined for x = 0, and we need to use subderivatives (or
subgradients) to define the JM. We avoid this technical complication and simply take the derivative of max(x, 0) to be 0 when
x = 0. Note that this does not change the results in any way because the subset of X for which the derivatives are not defined
has zero measure.
May 24, 2017
DRAFT
14
We also briefly explore a relationship between the Jacobian matrix and the Fisher information matrix.
To simplify the derivations we assume M = 1, N = 1, x0 = x + θn and n ∼ N (0, 1), where θ is the
model parameter and x is deterministic. The Fisher information F (θ) measures how much information
about the parameter θ is contained in the random variable y = f (x0 ), where f represents a DNN. In this
particular case the Fisher information is given as
"
#
d log f (x0 ) 2
F (θ) = En
dθ
"
#
d log f (x0 ) df (x0 ) dx0 2
= En
df (x0 )
dx0 θ
"
2 #
d log f (x0 )
0
.
= En
J(x )n
df (x0 )
(32)
In our setup the parameter θ can be interpreted as a magnitude of the input perturbation. It is clear from
(32) that a small norm of the Jacobian matrix leads to a small Fisher information, which indicates that
the distribution of y is not very informative about the parameters θ. By ensuring that the norm of the
Jacobian is small we then naturally endow the network with robustness against perturbations of the input.
IV. G ENERALIZATION ERROR OF A D EEP N EURAL N ETWORK C LASSIFIER
In this section we provide the classification margin bounds for DNN classifiers that allow us to bound
the GE. We follow the common practice and assume that the networks are trained by a loss that promotes
separation of different classes at the network output, e.g. categorical cross entropy loss or the hinge loss.
In other words, the training aims at maximizing the score of each training sample, where the score is
defined as follows.
Definition 3 (Score). Take score of a training sample si = (xi , yi )
√
o(si ) = min 2(δ yi − δ j )T f (xi ) ,
j6=yi
(33)
where δ i ∈ RNY is the Kronecker delta vector with (δ i )i = 1.
Recall the definition of the classifier g(x) in (14) and note that the decision boundary between class i
and class j in the feature space Z is given by the hyperplane {z : (z)i = (zj )}. A positive score indicates
that at the network output, classes are separated by a margin that corresponds to the score. However, a
large score o(si ) does not necessarily imply a large classification margin γ d (si ). Theorem 4 provides
classification margin bounds expressed as a function of the score and the properties of the network.
May 24, 2017
DRAFT
15
Theorem 4. Assume that a DNN classifier g(x), as defined in (14), classifies a training sample xi with
the score o(si ) > 0. Then the classification margin can be bounded as
γ d (si ) ≥
≥
o(si )
supx:kx−xi k2 ≤γ d (si ) kJ(x)k2
, γ1d (si )
(34)
o(si )
, γ2d (si )
supx∈conv(X ) kJ(x)k2
(35)
≥Q
o(si )
, γ3d (si )
kW
k
l 2
Wl ∈W
(36)
≥Q
o(si )
, γ4d (si ) ,
kW
k
l F
Wl ∈W
(37)
where W is the set of all weight matrices of f (x).
Proof: The proof appears in Appendix D.
Given the bounds of the classification margin we can specialize Corollary 1 to DNN classifiers.
Corollary 3. Assume that X is a (subset of) CM regular k -dimensional manifold, where N (X ; d, ρ) ≤
k
CM
. Assume also that DNN classifier g(x) achieves a lower bound to the classification margin
ρ
γbd (si ) > γb for b ∈ {1, 2, 3, 4}, ∀si ∈ Sm and take `(g(xi ), yi ) to be the 0-1 loss. Then for any δ > 0,
with probability at least 1 − δ ,
s
GE(g) ≤
log(2) · NY · 2k+1 · (CM )k
+
γbk m
r
2 log(1/δ)
.
m
(38)
Proof: The proof follows from Theorems 1, 2 and 4.
Corollary 3 suggests that the GE will be bounded by C √1m γ −k/2 , where C =
p
log(2) · NY 2k+1 (CM )k ,
provided that the classification margin bounds satisfy γbd (si ) > γ for some b ∈ {1, 2, 3, 4}, ∀si ∈ Sm .
We now leverage the classification margin bounds in Theorem 4 to construct constraint sets Wb =
{Wl ∈ W : γbd (si ) > γ∀si }, b ∈ 1, 2, 3, 4 such that W ∈ Wb ensures that the GE is bounded by
May 24, 2017
DRAFT
16
C √1m γ −k/2 . Using (34)-(37) we obtain
W1 = Wl ∈ W :
sup
x:kx−xi k2 ≤γ d (si )
kJ(x)k2 < γ · o(si )
∀si = (xi , yi ) ,
W2 = Wl ∈ W :
sup
x∈conv(X )
(39)
kJ(x)k2 < γ · o(si )
∀si = (xi , yi ) ,
(40)
Y
W3 = Wl ∈ W :
kWl k2 < γ · o(si )
Wl ∈W
∀si = (xi , yi ) ,
(41)
Y
kWl kF < γ · o(si )
W4 = Wl ∈ W :
Wl ∈W
∀si = (xi , yi ) .
(42)
Note that while we want to maximize the score o(si ), we also need to constrain the network’s Jacobian
matrix J(x) (following W1 and W2 ) or the weight matrices Wl ∈ W (following W3 and W4 ). This
stands in line with the common training rationale of DNN in which we do not only aim at maximizing
the score of the training samples to ensure a correct classification of the training set, but also have a
regularization that constrains the network parameters, where this combination eventually leads to a lower
GE. The constraint sets in (39)-(42) impose different regularization techniques:
•
The term supx:kx−xi k2 ≤γ d (si ) kJ(x)k2 < γ · o(si ) in (39) considers only the supremum of the spectral
norm of the Jacobian matrix evaluated at the points within Ni = {x : kx − xi k2 ≤ γ d (si )}, where
γ d (si ) is the classification margin of training sample si (see Definition 2). We can not compute the
margin γ d (si ), but can still obtain a rationale for regularization: as long as the spectral norm of the
Jacobian matrix is bounded in the neighbourhood of a training sample xi given by Ni we will have
the GE guarantees.
•
The constraint on the Jacobian matrix supx∈conv(X ) kJ(x)k2 < γ · o(si ) in (40) is more restrictive as
•
it requires bounded spectral norm for all samples x in the convex hull of the input space X .
Q
Q
The constraints in (41) and (42) are of similar form, Wl ∈W kWl k2 < γ·o(si ) and Wl ∈W kWl kF <
γ · o(si ), respectively. Note that the weight decay, which aims at bounding the Frobenious norms
of the weight matrices might be used to satisfy the constrains in (42). However, note also that the
bound based on the spectral norm in (41) is tighter than one based on the Frobenious norm in (42).
For example, take Wl ∈ W to have orthonormal rows and be of dimension M × M . Then the
constraint in (41), which is based on the spectral norm, is of the form 1 < γo(si ) and the constraint
in (42), which is based on the Frobenious norm, is M L/2 < γo(si ). In the former case we have
May 24, 2017
DRAFT
17
a constraint on the score, which is independent of the network width or depth. In the latter the
constraint on the output score is exponential in network depth and polynomial in network width. The
difference is that the Frobenious norm does not take into account the correlation (angles) between
the rows of the weight matrix Wl , while the spectral norm does. Therefore, the bound based on the
Frobenious norm corresponds to the worst case when all the rows of Wl are aligned. In that case
√
√
kWl kF = kWl k2 = M . On the other hand, if the rows of Wl are orthonormal kWl kF = M ,
but kWl k2 = 1.
Remark 1. To put results into perspective we compare our GE bounds to the GE bounds based on the
Rademacher complexity in [33], which hold for DNNs with ReLUs. The work in [33] shows that if
W ∈ WF = {Wi ∈ W :
L
Y
kWi kF < CF }
(43)
i=1
and the energy of training samples is bounded then:
1
GE(g) / √ 2L−1 CF .
m
(44)
Although the bounds (38) and (44) are not directly comparable, since the bounds based on the robustness
framework rely on an underlying assumption on the data (covering number), there is still a remarkable
difference between them. The behaviour in (44) suggests that the GE grows exponentially with the network
depth even if the product of the Frobenious norms of all the weight matrices is fixed, which is due to the
term 2L . The bound in (34) and the constraint sets in (39)-(42), on the other hand, imply that the GE
does not increase with the number of layers provided that the spectral/Frobenious norms of the weight
matrices are bounded. Moreover, if we take the DNN to have weight matrices with orthonormal rows
then the GE behaves as
√1 (CM )k/2
m
(assuming o(si ) ≥ 1, i = 1, . . . , m), and therefore relies only on the
complexity of the underlying data manifold and not on the network depth. This provides a possible answer
to the open question of [33] that depth independent capacity control is possible in DNNs with ReLUs.
Remark 2. An important value of our bounds is that they provide an additional explanation to the success
of state-of-the-art DNN training techniques such as batch normalization [23] and eight normalization
[24].
Weight normalized DNNs have weight matrices with normalized rows, i.e.
Wl = diag(ŴlT Ŵl )−1 Ŵl ,
(45)
where diag(·) denotes the diagonal part of the matrix. While the main motivation for this method is
a faster training, the authors also show empirically that such networks achieve good generalization.
√
Note that for row-normalized weight matrices kWl kF = Ml and therefore the bounds based on the
May 24, 2017
DRAFT
18
Frobenious norm can not explain the good generalization of such networks as adding layers or making
Wl larger will lead to a larger GE bound. However, our bound in (34) and the constraint sets in (39)-(41)
show that a small Frobenious norm of the weight matrices is not crucial for a small GE. A supporting
experiment is presented in Section VI-A2.
We also note that batch normalization also leads to row-normalized weight matrices in DNNs with
ReLUs:5
Theorem 5. Assume that the non-linear layers of a DNN with ReLUs are batch normalized as:
ˆl
ẑl = Wl zl ,
zl+1 = [N {zli }m
,
W
l z ]σ ,
i=1
(46)
where σ denotes the ReLU non-linearity and
N ({zi }m
i=1 , W)
m
X
= diag
!− 1
2
Wzi zTi WT
(47)
i=1
is the normalization matrix. Then all the weight matrices are row normalized. The exception is the weight
matrix of the last layer, which is of the form N({zL−1
}m
i=1 , WL )WL .
i
Proof: The proof appears in Appendix E.
A. Jacobian Regularizer
The constraint set (39) suggests that we can regularize the DNN by bounding the norm of the network’s
JM for the inputs close to xi . Therefore, we propose to penalize the norm of the network’s JM evaluated
at each training sample xi ,
m
1 X
RJ (W) =
kJ(xi )k22 .
m
(48)
i=1
The implementation of such regularizer requires computation of its gradients or subgradients. In this case
the computation of the subgradient of the spectral norm requires the calculation of a SVD decomposition
[44], which makes the proposed regularizer inefficient. To circumvent this, we propose a surrogate
regularizer based on the Frobenious norm of the Jacobian matrix:
m
RF (W) =
1 X
kJ(xi )k2F .
m
(49)
i=1
5
To simplify the derivation we omit the bias vectors and therefore also the centering applied by the batch normalization. This
does not affect the generality of the result. We also follow [43] and omit the batch normalization scaling, as it can be included
into the weight matrix of the layer following the batch normalization. We also omit the regularization term and assume that the
matrices are invertible.
May 24, 2017
DRAFT
19
Note that the Frobenious norm and the spectral norm are related as follows: 1/rank(J(xi ))kJ(xi )k2F ≤
kJ(xi )k22 ≤ kJ(xi )k2F , which justifies using the surrogate regularizer. We will refer to RF (W) as the
Jacobian regularizer.
1) Computation of Gradients and Efficient Implementation: Note that the k -th row of J(xi ) corresponds
to the gradient of (f (x))k with respect to the input x evaluated at xi . It is denoted by gk (xi ) =
d(f (x))k
|x=xi .
dx
Now we can write
m
RF (W) =
M
1 XX
gk (xi )gk (xi )T .
m
(50)
i=1 k=1
As the regularizer will be minimized by a gradient descent algorithm we need to compute its gradient
with respect to the DNN parameters. First, we express gk (xi ) as
gk (xi ) = gkl (xi )Wl Jl−1 (xi )
where gkl (xi ) =
Jl−1 (xi ) =
d(f (x))k
dẑl |x=xi
dzl−1
dx |x=xi
(51)
is the gradient of (f (x))k with respect to ẑl evaluated at the input xi and
is the JM of l − 1-th layer output zl−1 evaluated at the input xi . The gradient of
gk (xi )gk (xi )T with respcet to Wl is then given as [45]
∇Wl gk (xi )gk (xi )T = 2gkl (xi )T gkl (xi )Wl Jl−1 (xi ).
The computation of the gradient of the regularizer at layer l requires the computation of gradients gkl (xi ),
k = 1, . . . , M , i = 1, . . . , m and the computation of the Jacobian matrices Jl−1 (xi ), i = 1, . . . , m. The
computation of the gradient of a typical loss used for training DNN usually involves a computation of m
gradients with computational complexity similar to the computational complexity of gkl (xi ). Therefore,
the computation of gradients required for an implementation of the Jacobian regularizer can be very
expensive.
To avoid excessive computational complexity we propose a simplified version of the regularizer (49),
which we name per-layer Jacobian regularizer. The per-layer Jacobian regularizer at layer l is defined as
m
1 X l−1
l−1
RFl (Wl ) =
g̃π(i) (xi )(g̃π(i)
(xi ))T ,
(52)
m
i=1
l−1
where g̃π(i)
(xi ) =
d(f (x))π(i)
dzl−1 |x=xi ,
and π(i) ∈ {1, . . . , M } is a random index. Compared to (49) we have
made two simplifications. First, we assumed that input of layer l is fixed. This way we do not need to
compute the JM Jl−1 (xi ) between the output of the layer l − 1 and the input. Second, by choosing only
one index π(i) per training sample we have to compute only one additional gradient per training sample.
l−1
l−1
This significantly reduces the computational complexity. The gradient of g̃π(i)
(xi )(g̃π(i)
(xi ))T is simply
l−1
l−1
l
l
∇Wl g̃π(i)
(xi )(g̃π(i)
(xi ))T = 2gπ(i)
(xi )T gπ(i)
(xi )Wl .
We demonstrate the effectiveness of this regularizers in Section VI.
May 24, 2017
DRAFT
20
V. D ISCUSSION
In the preceding sections we analysed the standard feed-forward DNNs and their classification margin
measured in the Euclidean norm. We now briefly discuss how our results extend to other DNN architectures
and different margin metrics.
A. Beyond Feed Forward DNN
There are various DNN architectures such as Residual Networks (ResNets) [4], [46], Recurrent Neural
Networks (RNNs) and Long Short-Term Memory (LSTM) networks [47], and Auto-encoders [48] that
are used frequently in practice. It turns out that our analysis – which is based on the network’s JM – can
also be easily extended to such DNN architectures. In fact, the proposed framework encompasses all
DNN architectures for which the JM can be computed.Below we compute the JM of a ResNet.
The ResNets introduce shortcut connection between layers. In particular, let φ(·, θl ) denote a concatenation of several non-linear layers (see (18)). The l-th block of a Residual Network is then given
as
zl = zl−1 + φ(zl−1 , θl ) .
(53)
We denote by Jl (zl−1 ) the JM of φ(zl−1 , θl ). Then the JM of the l-th block is
dzl
= I + Jl (zl−1 ) ,
dzl−1
(54)
and the JM of a ResNet is of the form
JSM (zL−1 ) ·
I+
L
X
!!
l−1
Y
Jl (zl−1 )
(I + Jl−i (zl−2 ))
,
l=1
i=1
(55)
where JSM (zL−1 ) denotes the JM of the soft-max layer. In particular, the right element of the product in
(55) can be expanded as
I + J1 (x)
+J2 (z1 ) + J2 (z1 )J1 (x)
+J3 (z2 ) + J3 (z2 )J2 (z1 )J1 (x) + J3 (z2 )J2 (x) + J3 (x)J1 (x)
+...
This is a sum of JMs of all the possible sub-networks of a ResNet. In particular, there are L elements of
the sum consiting only of one 1-layer sub-networks and there is only one element of the sum consisting
of L-layer sub-network. This observation is consistent with the claims in [49], which states that ResNets
resemble an ensemble of relatively shallow networks.
May 24, 2017
DRAFT
21
B. Beyond the Euclidean Metric
Moreover, we can also consider the geodesic distance on a manifold as a measure for margin instead
of the Euclidean distance. The geodesic distance can be more appropriate than the Euclidean distance
since it is a natural metric on the manifold. Moreover, the covering number of the manifold X may be
smaller if we use the covering based on the geodesic metric balls, which will lead to tighter GE bounds.
We outline the approach below.
Assume that X is a Riemannian manifold and x, x0 ∈ X . Take a continuous, piecewise continuously
differentiable curve c(t), t = [0, 1] such that c(0) = x, c(1) = x0 and c(t) ∈ X ∀t ∈ [0, 1]. The set of all
such curves c(·) is denoted by C . Then the geodesic distance between x and x0 is defined as
Z 1
dc(t)
0
dG (x, x ) = inf
dt .
dt 2
c(t)∈C 0
(56)
Similarly as in Section III, we can show that the JM of DNN is central to bounding the distance expansion
between the signals at the DNN input and the signals at the DNN output.
Theorem 6. Take x, x0 ∈ X , where X is the Riemmanian manifold and take c(t), t = [0, 1] to be
a continuous, piecewise continuously differentiable curve connecting x and x0 such that dG (x, x0 ) =
R 1 dc(t)
dt. Then
dt
0
2
kf (x0 ) − f (x)k2 ≤ sup kJ(c(t))k2 dG (x0 , x)
(57)
t∈[0,1]
Proof: The proof appears in Appendix F.
Note that we have established a relationship between the Euclidean distance of two points in the output
space and the corresponding geodesic distance in the input space. This is important because it implies that
promoting a large Euclidean distance between points can lead to a large geodesic distance between the
points in the input space. Moreover, the ratio between kf (x0 ) − f (x)k2 and dG (x, x0 ) is upper bounded
by the maximum value of the spectral norm of the network’s JM evaluated on the line c(t). This result is
analogous to the results of Theorem 3 and Corollary 2. It also implies that regularizing the network’s JM
as proposed in Section IV is beneficial also in the case when the classification margin is not measured in
the Euclidean metric.
Finally, note that in practice the training data may not be balanced. The provided GE bounds are still
valid in such cases. However, the classification error may not the best measure of performance in such
cases as it is dominated by the classification error of the class with the highest prior probability. Therefore,
alternative performance measures need to be considered. We leave a detailed study of training DNN with
unbalanced training sets for possible future work.
May 24, 2017
DRAFT
22
99
Accuracy [%]
Accuracy [%]
98
97
96
95
94
93
2
3
Number of Layers
(a) MNIST
4
64
62
60
58
56
54
52
50
48
46
44
2
3
Number of Layers
4
(b) CIFAR-10
Fig. 3. Classification accuracy of DNNs trained with the Jacobian regularization (solid lines) and the weight decay (dashed
lines). Different numbers of training samples are used: 5000 (red), 20000 (blue) and 50000 (black).
VI. E XPERIMENTS
We now validate the theory with a series of experiments on the MNIST [50], CIFAR-10 [51], LaRED
[52] and ImageNet (ILSVRC2012) [53] datasets. The Jacobian regularizer is applied to various DNN
architectures such as DNN with fully connected layers, convolutional DNN and ResNet [4]. We use the
ReLUs in all considered DNNs as this is currently the most popular non-linearity.
A. Fully Connected DNN
In this section we compare the performance of fully connected DNNs regularized with Jacobian
Regularization or with the weight decay. Then we analyse the behaviour of the JM of a fully connected
DNNs of various depth and width.
1) Comparison of Jacobian Regularization and Weight Decay: First, we compare standard DNN with
fully connected layers trained with the weight decay and the Jacobian regularization (49) on the MNIST
and CIFAR-10 datasets. Different number of training samples are used (5000, 20000, 50000). We consider
DNNs with 2, 3 and 4 fully connected layers where all layers, except the last one, have dimension equal
to the input signal dimension, which is 784 in case of MNIST and 3072 in case of CIFAR-10. The last
layer is always the softmax layer and the objective is the CEE loss. The networks were trained using
the stochastic gradient descent (SGD) with momentum, which was set to 0.9. Batch size was set to 128
and learning rate was set to 0.01 and reduced by factor 10 after every 40 epochs. The networks were
trained for 120 epochs in total. The weight decay and the Jacobian regularization factors were chosen on
a separate validation set. The experiments were repeated with the same regularization parameters on 5
random draws of training sets and weight matrix initializations. Classification accuracies averaged over
different experimental runs are shown in Fig. 3. We observe that the proposed Jacobian regularization
always outperforms the weight decay. This validates our theoretical results in Section IV, which predict
May 24, 2017
DRAFT
23
that the Jacobian matrix is crucial for the control of (the bound to) the GE. Interestingly, in the case of
MNIST, a 4 layer DNN trained with 20000 training samples and Jacobian regularization (solid blue line
if Fig. 3 (a)) performs on par with DNN trained with 50000 training samples and weight decay (dashed
black line Fig. 3 (a)), which means that the Jacobian regularization can lead to the same performance
with significantly less training samples.
2) Analysis of Weight Normalized Deep Neural Networks: Next, we explore weight normalized DNNs,
which are described in Section IV. We use the MNIST dataset and train DNNs with a different number of
fully connected layers (L = 2, 3, 4, 5) and different sizes of weight matrices (Ml = 784, 2 · 784, 3 · 784, 4 ·
784, 5 · 784, 6 · 784, l = 1, . . . , L − 1). The last layer is always the softmax layer and the objective is the
CCE loss. The networks were trained using the stochastic gradient descent (SGD) with momentum, which
was set to 0.9. Batch size was set to 128 and learning rate was set to 0.1 and reduced by factor 10 after
every 40 epochs. The networks were trained for 120 epochs in total. All experiments are repeated 5 times
with different random draws of a training set and different random weight initializations. We did not
employ any additional regularization as our goal here is to explore the effects of the weight normalization
on the DNN behaviour. We always use 5000 training samples.
The classification accuracies are shown in Fig. 4 (a) and the smallest classification score obtained
on the training set is shown in Fig. 4 (b). We have observed for all configurations that the training
accuracies were 100% (only exception is the case L = 2, Ml = 784 where the training accuracy was
99.6%). Therefore, the (testing set) classification accuracies increasing with the network depth and the
weight matrix size directly imply that the GE is smaller for deeper and wider DNNs. Note also that the
score increases with the network depth and width. This is most obvious for the 2 and 3 layer DNNs,
√
whereas for the 3 and 4 layer DNNs the score is close to 2 for all network widths.
Since the DNNs are weight normalized, the Frobenious norms of the weight matrices are equal to the
square root of the weight matrix dimension, and the product of Frobenious norms of the weight matrices
grows with the network depth and the weight matrix size. The increase of score with the network depth
and network width does not offset the product of Frobenious norms, and clearly, the bound in (38) based
on the margin bound in (37) and the bound in (44), which leverage the Frobenious norms of the weight
matrices, predict that the GE will increase with the network depth and weight matrix size in this scenario.
Therefore, the experiment indicates that these bounds are too pessimistic.
We have also inspected the spectral norms of the weight matrices of the trained networks. In all cases
the spectral norms were greater than one. We can argue that the bound in (38) based on the margin bound
in (36) predicts that the GE will increase with network depth, as the product of the spectral norms grows
with the network depth in a similar way than in previous paragraph. We note however, that the spectral
May 24, 2017
DRAFT
24
norms of the weight matrices are much smaller than the Frobenious norms of the weight matrices.
Finally, we look for a possible explanation for the success of the weight normalization in the bounds
in (38) based on the margin bounds in (34) and (35), which are a function on the JM. The largest value
of the spectral norm of the network’s JM evaluated on the training set is shown in Fig. 4 (c) and the
largest value of the spectral norm of the network’s JM evaluated on the testing set is shown in Fig. 4 (d).
We can observe an interesting phenomena. The maximum value of the JM’s spectral norm on the
training set decreases with the network depth and width. On the other hand, the maximum value of the
JM’s spectral norm on the testing set increases with network depth (and slightly with network width).
From the perspective of the constraint sets in (39) and (40) we note that in the case of the latter we have
to take into account the worst case spectral norm of the JM for inputs in conv(X ). The maximum value
of the spectral norm on the testing set indicates that this value increases with the network depth and
implies that the bound based on (35) is still loose. On the other hand, the bound in (34) implies that
we have to consider the JM in the neighbourhood of the training samples. As an approximation, we can
take the spectral norms of the JMs evaluated at the training set. As it is shown in Fig. 4 (c) this values
decrease with the network depth and width. We argue that this provides a reasonable explanation for the
good generalization of deeper and wider weight normalized DNNs.
B. Convolutional DNN
In this section we compare the performance of convolutional DNNs regularized with the Jacobian
regularizer or with the weight decay. We also show that Jacobian Regularization can be applied to batch
normalized DNNs. We will use the standard MNIST and CIFAR-10 dataset and the LaRED dateset which
is briefly described below.
The LaRED dataset contains depth images of 81 distinct hand gestures performed by 10 subjects with
approximately 300 images of each gesture per subject. We extracted the depth images of the hands using
the masks provided in [52] and resized the images to 32 × 32. The images of the first 6 subjects were
used to create non-overlapping training and testing sets. In addition we also constructed a testing set
composed from the images of the last 4 subjects in the dataset in order to test generalization across
different subjects. The goal is classification of gestures based on the depth image.
1) Comparison of Jacobian Regularization and Weight Decay: We use a 4 layer convolutional DNN
with the following architecture: (32, 5, 5)-conv, (2, 2)-max-pool, (32, 5, 5)-conv, (2, 2)-max-pool followed
by a softmax layer, where (k, u, v)-conv denotes the convolutional layer with k filters of size u × v and
(p, p)-max-pool denotes max-pooling with pooling regions of size p × p. The training procedure follows
the one described in the previous paragraphs. The results are reported in Table II.
May 24, 2017
DRAFT
L=2
L=3
L=4
L=5
784 1568 2352 3136 3920 4704
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
2.5
max
i |J(xi )|2 (train set)
96.0
95.8
95.6
95.4
95.2
95.0
94.8
94.6
mini o(si )
Accuracy [%]
25
2.0
1.0
0.5
0.0
784 1568 2352 3136 3920 4704
Layer width
(a) Classification accuracy.
L=2
L=3
L=4
L=5
1.5
Layer width
784 1568 2352 3136 3920 4704
Layer width
(b) Smallest o(si ) (training set). (c) Largest kJ(xi )k2 (training
set).
max
i |J(xi )|2 (test set)
10
9
8
7
6
5
4
3
2
784 1568 2352 3136 3920 4704
Layer width
(d) Largest kJ(xi )k2 (test set).
Fig. 4. Weight normalized DNN with L = 2, 3, 4, 5 layers and different sizes of weight matrices (layer width). Plot (a) shows
classification accuracy, plot (b) shows the smallest score of training samples, plot (c) shows the largest spectral norm of the
network’s JM evaluated on the training set and plot (d) shows the largest spectral norm of the network’s JM evaluated on the
testing set.
TABLE II
C LASSIFICATION ACC . [%] OF CONVOLUTIONAL DNN
MNIST
ON
LaRED (same subject)
MINST AND L A RED.
LaRED (different subject)
# train samples
Weight dec.
Jac. reg.
# train samples
Weight dec.
Jac. reg.
# train samples
Weight dec.
Jac. reg.
1000
94.00
96.03
2000
61.40
63.56
2000
31.53
32.62
5000
97.59
98.20
5000
76.59
79.14
5000
38.11
39.62
20000
98.60
99.00
10000
87.01
88.24
10000
41.18
42.85
50000
99.10
99.35
50000
97.18
97.54
50000
45.12
46.78
We observe that training with the Jacobian regularization outperforms the weight decay in all cases.
This is most obvious at smaller training set sizes. For example, on the MNIST dataset, the DNN trained
using 1000 training samples and regularized with the weight decay achieves classification accuracy of
94% and the DNN trained with the Jacobian regularization achieves classification accuracy of 96.3%.
Similarly, on the LaRED dataset the Jacobian regularization outperforms the weight decay with the
difference most obvious at the smallest number of training samples. Note also that the generalization of
May 24, 2017
DRAFT
26
TABLE III
C LASSIFICATION ACC . [%] OF CONVOLUTIONAL DNN
ON
CIFAR-10.
# train samples
Batch norm.
Batch norm. + Jac. reg.
2500
60.86
66.15
10000
76.35
80.57
50000
87.44
88.95
the network to the subjects outside the training set is not very good; i.e., using 50000 training samples
the classification accuracy on the testing set containing the same subjects is higher than 97% whereas
the classification accuracy on the testing set containing different subjects is only 46%. Nevertheless, the
Jacobian regularization outperforms the weight decay also on this testing set by a small margin.
2) Batch Normalization and Jacobian Regularization: Now we show that the Jacobian regularization
(49) can also be applied to a batch normalized DNN. Note that we have shown in Section IV that the
batch normalization has an effect of normalizing the rows of the weight matrices.
We us the CIFAR-10 dataset and use the All-convolutional-DNN proposed in [54] (All-CNN-C) with 9
convolutional layers, an average pooling layer and a softmax layer. All the convolutional layers are batch
normalized and the softmax layer is weight normalized. The networks were trained using the stochastic
gradient descent (SGD) with momentum, which was set to 0.9. Batch size was set to 64 and the learning
rate was set to 0.1 and reduced by a factor 10 after every 25 epochs. The networks were trained for 75
epochs in total. The classification accuracy results are presented in Table III for different sizes of training
sets (2500, 10000, 50000).
We can observe that the Jacobian regularization also leads to a smaller GE in this case.
C. Residual Networks
Now we demonstrate that the Jacobian regularizer is also effective when applied to ResNets. We use
the CIFAR-10 and ImageNet datasets. We use the per-layer Jacobian regularization (52) for experiments
in this section.
1) CIFAR-10: The Wide ResNet architecture proposed in [35], which follows [46], but proposes wider
and shallower networks which leads to the same or better performance than deeper and thinner networks
is used here. In particular, we use the ResNet with 22 layers of width 5.
We follow the data normalization process of [35]. We also follow the training procedure of [35] except
for the learning rate and use the learning rate sequence: (0.01, 5), (0.05, 20), (0.005, 40), (0.0005, 40),
(0.00005, 20), where the first number in parenthesis corresponds to the learning rate and the second
May 24, 2017
DRAFT
27
TABLE IV
C LASSIFICATION ACC . [%] OF R ES N ET CIFAR-10
# train samples
ResNet
ResNet + Jac. reg.
2500
55.69
62.79
10000
71.79
78.70
50000 + aug.
93.34
94.32
number corresponds to the number of epochs. We train ResNet on small training sets (2500 and 10000
training samples) without augmentation and on the full training set with the data augmentation as in [35].
The regularization factor were set to 1 and 0.1 for the smaller training sets (2500 and 10000) and the full
augmented training set, respectively.
The results are presented in Table IV. In all cases the ResNet with Jacobian regularization outperforms
the standard ResNet. The effect of regularization is the strongest with the smaller number of training
samples, as expected.
2) ImageNet: We use the 18 layer ResNet [4] with identity connection [46]. The training procedure
follows [4] with the learning rate sequence: (0.1, 30), (0.01, 30), (0.001, 30). The Jacobian regularization
factor is set to 1.
The images in the dataset are resized to 128 × 128. We run an experiment without data augmentation
and with data augmentation following [1], which includes random cropping of images of size 112 × 112
from the original image and color augmentation. The classification accuracies during training are shown
in Fig. 5 and the final results are reported in Table V.
We first focus on training without data augmentation. The ResNet trained using the Jacobian regularization
has a much smaller GE (23.83%) compared to the baseline ResNet (61.53%). This again demonstrates that
the Jacobian regularization decreases the GE, as our theory predicts. Note that the smaller GE of Jacobian
regularized ResNet partially transfers to a higher classification accuracy on the testing set. However, in
practice DNNs are often trained with data augmentation. In this case the GE of a baseline ResNet is much
lower (13.14%) and is very close to the GE of the ResNet with the Jacobian regularization (12.03%).
It is clear that data augmentation reduces the need for strong regularization. Nevertheless, note that the
ResNet trained with the Jacobian regularization achieves a slightly higher testing set accuracy (47.51%)
compared to the baseline ResNet (46.75%).
May 24, 2017
DRAFT
28
TABLE V
Train
Test
GE
Baseline
89.82
28.29
61.53
Baseline + Jac. reg.
59.52
35.69
23.83
Baseline + aug.
59.89
46.75
13.14
Baseline + aug + Jac. reg.
59.54
47.51
12.03
0 10 20 30 40 50 60 70 80 90
100
90
80
70
60
50
40
30
Epoch
60
Top-1 accuracy [%]
90
80
70
60
50
40
30
20
10
Setup
Top-5 accuracy [%]
Top-1 accuracy [%]
C LASSIFICATION ACC . [%] AND GE [%] OF R ES N ET ON I MAGENET.
0 10 20 30 40 50 60 70 80 90
Epoch
(b) No data augmentation.
Top-5 accuracy [%]
(a) No data augmentation.
90
80
70
60
50
40
30
20
50
40
30
20
10
0 10 20 30 40 50 60 70 80 90
Epoch
(c) Data augmentation.
Train (baseline)
Test (baseline)
Train (Jac. reg.)
Test (Jac. reg.)
0 10 20 30 40 50 60 70 80 90
Epoch
(d) Data augmentation.
Fig. 5. Training set (dashed) and testing set (solid) classification accuracies during training. Blue curves correspond to the
ResNet with Jacobian regularization and red curves correspond to the baseline ResNet. Top-1 and top-5 classification accuracies
are reported for training without data augmentation (a,b) and for training with data augmentation (c,d).
D. Computational Time
Finally, we measure how the use of Jacobian regularization affects training time of DNNs. We have
implemented DNNs in Theano [55], which includes automatic differentiation and computation graph
optimization. The experiments are run on the Titan X GPU. The average computational time per batch
for the convolutional DNN on the MNIST dataset in Section VI-B1 and for the ResNet on the ImageNet
dataset in Section VI-C2 are reported in Table VI. Note that in the case of MNIST the regularizer in
(49) is used and in the case of ImageNet the per-layer regularizer in (52) is used. These results are also
representative of the other datasets and network architectures.
May 24, 2017
DRAFT
29
TABLE VI
AVERAGE COMPUTATION TIME [s/ BATCH ]
Experiment
no reg.
Jac. reg.
Increase factor
MNIST (Sec.VI-B1)
0.003
0.030
10.00
ImageNet (Sec. VI-C2)
0.120
0.190
1.580
We can observe that using the Jacobian regularizer in (49) introduces additional computational time.
This may not be critical if the number of training samples is small and training computational time is not
too critical. On the other hand, the per-layer Jacobian regularizer in (52) has a much smaller cost. As
shown in the experiments this regularizer is still effective and leads to only 58% increase in computation
time on the ImageNet dataset. Due to its efficiency the per-layer Jacobian regularizer might be more
appropriate for large scale experiments where computational time is important.
VII. C ONCLUSION
This paper studies the GE of DNNs based on their classification margin. In particular, our bounds
express the generalization error as a function of the classification margin, which is bounded in terms of
the achieved separation between the training samples at the network output and the network’s JM.
One of the hallmarks of our bounds relates to the fact that the characterization of the behaviour of the
generalization error is tighter than that associated with other bounds in the literature. Our bounds predict
that the generalization error of deep neural networks can be independent of their depth and size whereas
other bounds say that the generalization error is exponential in the network width or size.
Our bounds also suggest new regularization strategies such as the regularization of the network’s
Jacobian matrix, which can be applied on top of other modern DNN training strategies such as the weight
normalization and the batch normalization, where the standard weight decay can not be applied. These
regularization strategies are especially effective in the limited training data regime in comparison to other
approaches, with moderate increase in computational complexity.
A PPENDIX
A. Proof of Theorem 3
We first note that the line between x and x0 is given by x + t(x0 − x), t ∈ [0, 1] . We define the
function F (t) = f (x + t(x0 − x)), and observe that
May 24, 2017
dF (t)
dt
= J(x + t(x0 − x))(x0 − x). By the generalized
DRAFT
30
fundamental theorem of calculus or the Lebesgue differentiation theorem we write
Z 1
dF (t)
0
f (x ) − f (x) = F (1) − F (0) =
dt
dt
0
Z 1
=
J(x + t(x0 − x)) dt (x0 − x) .
(58)
0
This concludes the proof.
B. Proof of Corollary 2
First note that kJx,x0 (x0 − x)k2 ≤ kJx,x0 k2 kx0 − xk2 and that Jx,x0 is an integral of J(x + t(x0 − x)).
In addition, notice that we may always apply the following upper bound:
kJx,x0 k2 ≤
sup
kJ(x + t(x0 − x))k2 .
(59)
x,x0 ∈X ,t∈[0,1]
Since x + t(x0 − x) ∈ conv(X ) ∀t ∈ [0, 1], we get (24).
C. Proof of Lemma 1
In all proofs we leverage the fact that for any two matrices A, B of appropriate dimensions it holds
kABk2 ≤ kAk2 kBk2 . We also leverage the bound kAk2 ≤ kAkF .
We start with the proof of statement 1). For the non-linear layer (18), we note that the JM is a product
of a diagonal matrix (27) and the weight matrix Wl . Note that for all the considered non-linearities the
diagonal elements of (27) are bounded by 1 (see derivatives in Table I), which implies that the spectral
norm of this matrix is bounded by 1. Therefore the spectral norm of the JM is upper bounded by kWl k2 .
The proof for the linear layer is trivial. In the case of the softmax layer (17) we have to show that the
spectral norm of the softmax function −ζ(ẑ)ζ(ẑ)T + diag(ζ(ẑ) is bounded by 1. We use the Gershgorin
disc theorem, which states that the eigenvalues of −ζ(ẑ)ζ(ẑ)T + diag(ζ(ẑ) are bounded by
max(ζ(ẑ))i (1 − (ζ(ẑ))i ) + (ζ(ẑ))i
i
Noticing that
P
j6=i (ζ(ẑ))j
X
(ζ(ẑ))j .
(60)
j6=i
≤ 1 leads to the upper bound
maxi (ζ(ẑ))i (2 − (ζ(ẑ))i ) .
(61)
Since (ζ(ẑ))i ∈ [0, 1] it is trivial to show that (61) is upper bounded by 1.
The proof of statement 2) is straightforward. Because the pooling regions are non-overlapping it is
straightforward to verify that the rows of all the defined pooling operators Pl (zl−1 ) are orthonormal.
Therefore, the spectral norm of the JM is equal to 1.
May 24, 2017
DRAFT
31
D. Proof of Theorem 4
Throughout the proof we will use the notation o(si ) = o(xi , yi ) and vij =
√
2(δ i − δ j ). We start by
proving the inequality in (34). Assume that the classification margin γ d (si ) of training sample (xi , yi )
is given and take j ? = arg minj6=yi min vyTi j f (xi ). We then take a point x? that lies on the decision
boundary between yi and j ? such that o(x? , yi ) = 0. Then
o(xi , yi ) = o(x, yi ) − o(x? , yi ) = vyTi j ? (f (xi ) − f (x? ))
= vyTi j ? Jxi ,x? (xi − x? ) ≤ kJxi ,x? k2 kxi − x? k2 .
Note that by the choice of x? , kxi −x? k2 = γ d (si ) and similarly kJxi ,x? k2 ≤ supx:kx−xi k2 ≤γ d (si ) kJ(x)k2 .
Therefore, we can write
o(si ) ≤
sup
x:kx−xi k2 ≤γ d (si )
kJ(x)k2 γ d (si ),
(62)
which leads to (34).
Next, we prove (35). Recall the definition of the classification margin in (10):
γ d (si ) = sup{a : kxi − xk2 ≤ a =⇒ g(x) = yi ∀x}
= sup{a : kxi − xk2 ≤ a =⇒ o(x, yi ) > 0 ∀x} ,
where we leverage the definition in (33). We observe o(x, yi ) > 0 ⇐⇒ minj6=yi vyTi j f (x) > 0 and
min vyTi j f (x) = min vyTi j f (xi ) + vyTi j (f (x) − f (xi )) .
j6=yi
j6=yi
Note that
min vyTi j f (xi ) + vyTi j (f (x) − f (xi ))
j6=yi
(63)
≥ min vyTi j f (xi ) + min vyTi j (f (x) − f (xi ))
j6=yi
j6=yi
= o(xi , yi ) + min vyTi j (f (x) − f (xi )) .
j6=yi
(64)
Therefore,
o(xi , yi ) + min vyTi j (f (x) − f (xi )) > 0 =⇒ o(x, yi ) > 0 .
j6=yi
This leads to the bound of the classification margin
γ d (si ) ≥ sup{a : kxi − xk2 ≤ a =⇒
o(xi , yi ) + min vyTi j (f (x) − f (xi )) > 0 ∀x} .
j6=yi
May 24, 2017
DRAFT
32
Note now that
o(xi , yi ) + min vyTi j (f (x) − f (xi )) > 0
j6=yi
(65)
⇐⇒
o(xi , yi ) − max vyTi j (f (xi ) − f (x)) > 0
j6=yi
(66)
⇐⇒
o(xi , yi ) > max vyTi j (f (xi ) − f (x)) .
j6=yi
(67)
Moreover,
max vyTi j (f (xi ) − f (x)) ≤
j6=yi
kJ(x)k2 kxi − xk2 ,
sup
x∈conv(X )
where we have leveraged the fact that kvij k2 = 1 and the inequality (24) in Corollary 2. We may write
γ d (si ) ≥ sup{a :kxi − xk2 ≤ a =⇒
o(xi , yi ) >
sup
kJ(x)k2 kxi − xk2 ∀x}.
x∈conv(X )
a that attains the supremum can be obtain easily and we get:
γ d (si ) ≥
o(xi , yi )
,
supx∈conv(X ) kJ(x)k2
(68)
which proves (35). The bounds in (36) and (37) follow from the bounds provided in Lemma 1 and the
fact that the spectral norm of a matrix product is upper bounded by the product of the spectral norms.
This concludes the proof.
E. Proof of Theorem 5
We denote by WlN the row normalized matrix obtained from Wl (in the same way as (45)). By noting
that the ReLU and diagonal non-negative matrices commute, it is straight forward to verify that
l
l m
N [WN zl ] .
[N {zli }m
σ
i=1 , Wl Wl z ]σ = N {zi }i=1 , Wl
l
N
Note now that we can consider N({zli }m
i=1 , Wl ) as the part of the weight matrix Wl+1 . Therefore, we
can conclude that layer l has row normalized weight matrix. When the batch normalization is applied to
layers, all the weight matrices will be row normalized. The exception is the weight matrix of the last
layer, which will be of the form N({zL−1
}m
i=1 , WL )WL .
i
May 24, 2017
DRAFT
33
F. Proof of Theorem 6
We begin by noting that f (x0 ) − f (x) = f (c(1)) − f (c(0)) and
f (c(1)) − f (c(0)) =
R1
0
df (c(t))
dt
dt =
R1
0
df (c(t)) dc(t)
dt
dc(t)
dt ,
where the first equality follows from the generalized fundamental theorem of calculus, following the idea
presented in the proof of Theorem 3. The second equality follows from the chain rule of differentiation.
Finally, we note that
df (c(t))
c(t)
= J(c(t)) and that the norm of the integral is always smaller or equal to the
integral of the norm and obtain
kf (x0 )
Z
1
− f (x)k2 =
J(c(t))
Z
≤
0
1
dc(t)
dt
dt
kJ(c(t))k2
0
dc(t)
dt
Z 1
≤ sup kJ(c(t))k2
t∈[0,1]
0
2
dt
2
dc(t)
dt
dt
2
= sup kJ(c(t))k2 dG (x, x0 ) ,
(69)
t∈[0,1]
where we have noted that
R1
0
dc(t)
dt
2
= dG (x, x0 ).
R EFERENCES
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances
in Neural Information Processing Systems (NIPS), pp. 1097–1105, 2012.
[2] G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. N. Sainath,
and B. Kingsbury, “Deep neural networks for acoustic modeling in speech recognition: the shared views of four research
groups,” IEEE Signal Processing Magazine, vol. 29, no. 6, pp. 82–97, Oct. 2012.
[3] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015.
[4] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), Dec. 2016.
[5] V. Nair and G. E. Hinton, “Rectified linear units improve restricted Boltzmann machines,” Proceedings of the 27th
International Conference on Machine Learning (ICML), pp. 807–814, 2010.
[6] J. Bruna, A. Szlam, and Y. LeCun, “Learning stable group invariant representations with convolutional networks,” International
Conference on Learning Representations (ICLR), 2013.
[7] Y.-L. Boureau, J. Ponce, and Y. LeCun, “A theoretical analysis of feature pooling in visual recognition,” Proceedings of the
27th International Conference on Machine Learning (ICML), pp. 111–118, 2010.
[8] G. Cybenko, “Approximation by superpositions of a sigmoidal function,” Mathematics of Control Signals and Systems,
vol. 2, no. 4, pp. 303–314, 1989.
[9] K. Hornik, “Approximation capabilities of multilayer feedforward networks,” Neural Networks, vol. 4, no. 2, pp. 251–257,
1991.
[10] G. Montúfar, R. Pascanu, K. Cho, and Y. Bengio, “On the number of linear regions of deep neural networks,” Advances in
Neural Information Processing Systems (NIPS), pp. 2924–2932, 2014.
May 24, 2017
DRAFT
34
[11] N. Cohen, O. Sharir, and A. Shashua, “On the expressive power of deep learning: a tensor analysis,” 29th Annual Conference
on Learning Theory (COLT), pp. 698–728, 2016.
[12] M. Telgarsky, “Benefits of depth in neural networks,” 29th Annual Conference on Learning Theory (COLT), pp. 1517–1539,
2016.
[13] S. Mallat, “Group invariant scattering,” Communications on Pure and Applied Mathematics, vol. 65, no. 10, pp. 1331–1398,
2012.
[14] J. Bruna and S. Mallat, “Invariant scattering convolution networks,” IEEE Transactions on Pattern Analysis and Machine
Intellignce, vol. 35, no. 8, pp. 1872–1886, Mar. 2012.
[15] T. Wiatowski and H. Bölcskei, “A mathematical theory of deep convolutional neural networks for feature extraction,”
arXiv:1512.06293, 2015.
[16] R. Giryes, G. Sapiro, and A. M. Bronstein, “Deep neural networks with random Gaussian weights: a universal classification
strategy?” IEEE Transactions on Signal Processing, vol. 64, no. 13, pp. 3444–3457, Jul. 2016.
[17] A. Choromanska, M. Henaff, M. Mathieu, G. B. Arous, and Y. LeCun, “The loss surfaces of multilayer networks,”
International Conference on Artificial Intelligence and Statistics (AISTATS), 2015.
[18] B. D. Haeffele and R. Vidal, “Global optimality in tensor factorization, deep learning, and beyond,” arXiv:1506.07540,
2015.
[19] R. Giryes, Y. C. Eldar, A. M. Bronstein, and G. Sapiro, “Tradeoffs between convergence speed and reconstruction accuracy
in inverse problems,” arXiv:1605.09232, 2016.
[20] A. M. Saxe, J. L. McClelland, and S. Ganguli, “Exact solutions to the nonlinear dynamics of learning in deep linear neural
networks,” International Conference on Learning Representations (ICLR), 2014.
[21] Y. Ollivier, “Riemannian metrics for neural networks I: feedforward networks,” Information and Inference, vol. 4, no. 2, pp.
108–153, Jun. 2015.
[22] B. Neyshabur and R. Salakhutdinov, “Path-sgd: Path-normalized optimization in deep neural networks,” Advances in Neural
Information Processing Systems (NIPS), pp. 2422–2430, 2015.
[23] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,”
Proceedings of the 32nd International Conference on Machine Learning (ICML), pp. 448–456, 2015.
[24] T. Salimans and D. P. Kingma, “Weight normalization: A simple reparameterization to accelerate training of deep neural
networks,” Advances in Neural Information Processing Systems (NIPS), pp. 901–909, 2016.
[25] S. An, M. Hayat, S. H. Khan, M. Bennamoun, F. Boussaid, and F. Sohel, “Contractive rectifier networks for nonlinear
maximum margin classification,” Proceedings of the IEEE International Conference on Computer Vision, pp. 2515–2523,
2015.
[26] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, “Dropout: A simple way to prevent neural
networks from overfitting,” Journal of Machine Learning Research (JMLR), vol. 15, no. 1, pp. 1929–1958, Jun. 2014.
[27] S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio, “Contractive auto-encoders: explicit invariance during feature
extraction,” Proceedings of the 28th International Conference on Machine Learning (ICML), pp. 833–840, 2011.
[28] J. Huang, Q. Qiu, G. Sapiro, and R. Calderbank, “Discriminative robust transformation learning,” Advances in Neural
Information Processing Systems (NIPS), pp. 1333–1341, 2015.
[29] V. N. Vapnik, “An overview of statistical learning theory,” IEEE Transactions on Neural Networks, vol. 10, no. 5, pp.
988–999, Sep. 1999.
[30] S. Shalev-Shwartz and S. Ben-David, Understanding machine learning: from theory to algorithms. Cambridge University
Press, 2014.
May 24, 2017
DRAFT
35
[31] P. L. Bartlett and S. Mendelson, “Rademacher and Gaussian complexities: risk bounds and structural results,” Journal of
Machine Learning Research (JMLR), vol. 3, pp. 463–482, 2002.
[32] H. Xu and S. Mannor, “Robustness and generalization,” Machine Learning, vol. 86, no. 3, pp. 391–423, 2012.
[33] B. Neyshabur, R. Tomioka, and N. Srebro, “Norm-based capacity control in neural networks,” Proceedings of The 28th
Conference on Learning Theory (COLT), pp. 1376–1401, 2015.
[34] S. Sun, W. Chen, L. Wang, and T.-Y. Liu, “Large margin deep neural networks: theory and algorithms,” arXiv:1506.05232,
2015.
[35] S. Zagoruyko and N. Komodakis, “Wide residual networks,” arXiv:1605.07146, 2016.
[36] C. Zhang, S. Bengio, M. Hardt, and B. Recht, “Understanding deep learning requires rethinking generalization,”
arXiv:1611.03530, 2016.
[37] J. Sokolić, R. Giryes, G. Sapiro, and M. R. D. Rodrigues, “Generalization error of invariant classifiers,” in International
Conference on Artificial Intelligence and Statistics (AISTATS), 2017.
[38] K. Q. Shen, C. J. Ong, X. P. Li, and E. P. V. Wilder-Smith, “Feature selection via sensitivity analysis of SVM probabilistic
outputs,” Machine Learning, vol. 70, no. 1, pp. 1–20, Jan. 2008.
[39] J. B. Yang, K. Q. Shen, C. J. Ong, and X. P. Li, “Feature selection via sensitivity analysis of MLP probabilistic outputs,”
IEEE International Conference on Systems, Man and Cybernetics, pp. 774–779, 2008.
[40] D. Shi, D. S. Yeung, and J. Gao, “Sensitivity analysis applied to the construction of radial basis function networks,” Neural
networks, vol. 18, no. 7, pp. 951–957, Mar. 2005.
[41] S. Mendelson, A. Pajor, and N. Tomczak-Jaegermann, “Uniform uncertainty principle for Bernoulli and subgaussian
ensembles,” Constructive Approximation, vol. 28, no. 3, pp. 277–289, Dec. 2008.
[42] N. Verma, “Distance preserving embeddings for general n-dimensional manifolds.” Journal of Machine Learning Research
(JMLR), vol. 14, no. 1, pp. 2415–2448, Aug. 2013.
[43] B. Neyshabur, R. Tomioka, R. Salakhutdinov, and N. Srebro, “Data-dependent path normalization in neural networks,”
International Conference on Learning Representations (ICLR), 2015.
[44] G. Watson, “Characterization of the subdifferential of some matrix norms,” Linear Algebra and its Applications, vol. 170,
pp. 33–45, Jun. 1992.
[45] K. B. Petersen and M. S. Pedersen, “The matrix cookbook,” Technical University of Denmark, 2012.
[46] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” arXiv:1603.05027, 2016.
[47] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, Nov.
1997.
[48] Y. Bengio, “Learning deep architectures for AI,” Foundations and trends® in Machine Learning, vol. 2, no. 1, pp. 1–127,
2009.
[49] A. Veit, M. Wilber, and S. Belongie, “Residual networks are exponential ensembles of relatively shallow networks,”
arXiv:1605.06431, 2016.
[50] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” Proceedings of
the IEEE, vol. 86, no. 11, pp. 2278–2324, Nov. 1998.
[51] A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” Computer Science Department,
University of Toronto, Tech. Rep, Apr. 2009.
[52] Y. S. Hsiao, J. Sanchez-Riera, T. Lim, K. L. Hua, and W. H. Cheng, “LaRED: a large RGB-D extensible hand gesture
dataset,” Proceedings of the 5th ACM Multimedia Systems Conference, pp. 53–58, 2014.
[53] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C.
May 24, 2017
DRAFT
36
Berg, and L. Fei-Fei, “ImageNet large scale visual recognition challenge,” International Journal of Computer Vision (IJCV),
vol. 115, no. 3, pp. 211–252, 2015.
[54] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller, “Striving for simplicity: the all convolutional net,”
International Conference on Learning Representations (ICLR - workshop track), Dec. 2015.
[55] Theano Development Team, “Theano: A Python framework for fast computation of mathematical expressions,”
arXiv:1605.02688, 2016.
May 24, 2017
DRAFT
| 9cs.NE
|
arXiv:1704.02191v2 [cs.NE] 19 Jun 2017
The (1+λ) Evolutionary Algorithm with Self-Adjusting
Mutation Rate∗
Benjamin Doerr
Laboratoire d’Informatique (LIX)
École Polytechnique
Palaiseau, France
Christian Gießen
DTU Compute
Technical University of Denmark
Kgs. Lyngby, Denmark
Carsten Witt
DTU Compute
Technical University of Denmark
Kgs. Lyngby, Denmark
Jing Yang
Laboratoire d’Informatique (LIX)
École Polytechnique
Palaiseau, France
June 20, 2017
Abstract
We propose a new way to self-adjust the mutation rate in population-based
evolutionary algorithms in discrete search spaces. Roughly speaking, it consists of
creating half the offspring with a mutation rate that is twice the current mutation
rate and the other half with half the current rate. The mutation rate is then updated
to the rate used in that subpopulation which contains the best offspring.
We analyze how the (1+λ) evolutionary algorithm with this self-adjusting mutation rate optimizes the OneMax test function. We prove that this dynamic version
of the (1+λ) EA finds the optimum in an expected optimization time (number of fitness evaluations) of O(nλ/log λ + n log n). This time is asymptotically smaller than
the optimization time of the classic (1 + λ) EA. Previous work shows that this performance is best-possible among all λ-parallel mutation-based unbiased black-box
algorithms.
This result shows that the new way of adjusting the mutation rate can find
optimal dynamic parameter values on the fly. Since our adjustment mechanism is
simpler than the ones previously used for adjusting the mutation rate and does not
have parameters itself, we are optimistic that it will find other applications.
1
Introduction
Evolutionary algorithms (EAs) have shown a remarkable performance in a broad range
of applications. However, it has often been observed that this performance depends crucially on the use of the right parameter settings. Parameter optimization and parameter
∗
An extended abstract of this report will appear in the proceedings of the 2017 Genetic and Evolutionary Computation Conference (GECCO 2017).
1
control are therefore key topics in EA research. Since these have very different characteristics in discrete and continuous search spaces, we discuss in this work only evolutionary
algorithms for discrete search spaces.
Theoretical research has contributed to our understanding of these algorithms with
mathematically founded runtime analyses, many of which show how the runtime of an
EA is determined by its parameters. The majority of these works investigate static
parameter settings, i. e., the parameters are fixed before the start of the algorithm and
are not changed during its execution. More recently, a number of results were shown
which prove an advantage of dynamic parameter settings, that is, the parameters of the
algorithm are changed during its execution. Many of these rely on making the parameters
functionally dependent on the current state of the search process, e.g., on the fitness
of the current-best individual. While this provably can lead to better performances,
it leaves the algorithm designer with an even greater parameter setting task, namely
inventing a suitable functional dependence instead of fixing numerical values for the
parameters. This problem has been solved by theoretical means for a small number of
easy benchmark problems, but it is highly unclear how to find such functional relations
in the general case.
A more designer-friendly way to work with dynamic parameters is to modify the
parameters based on simple rules taking into account the recent performance. A number
of recent results shows that such on the fly or self-adjusting parameter settings can
give an equally good performance as the optimal fitness-dependent parameter setting,
however, with much less input from the algorithm designer. For example, good results
have been obtained by increasing or decreasing a parameter depending on whether the
current iteration improved the best-so-far solution or not, e.g., in a way resembling the
1/5-th rule from continuous optimization.
Such success-based self-adjusting parameter settings can work well when there is a
simple monotonic relation between success and parameter value, e.g., when one speculates that increasing the size of the population in an EA helps when no progress was
made. For parameters like the mutation rate, it is not clear what a success-based rule
can look like, since a low success rate can either stem from a too small mutation rate
(regenerating the parent with high probability) or a destructive too high mutation rate.
In [17], a relatively complicated learning mechanism was presented that tries to learn
the right mutation strength by computing a time-discounted average of the past performance stemming from different parameter values. This learning mechanism needed
a careful trade-off between exploiting the currently most profitably mutation strength
and experimenting with other parameter values and a careful choice of the parameter
controlling by how much older experience is taken less into account than more recent
observations.
1.1
A New Self-Adjusting Mechanism for Population-Based EAs
In this work, we propose an alternative way to adjust the mutation rate on the fly
for algorithms using larger offspring populations. It aims at overcoming some of the
difficulties of the learning mechanism just described. The simple idea is to create half
2
the offspring with twice the current mutation rate and the other half using half the
current rate. The mutation rate is then modified to the rate which was used to create
the best of these offspring (choosing the winning offspring randomly among all best in
case of ambiguity). We do not allow the mutation rate to leave the interval [2/n, 1/4],
so that the rates used in the subpopulations are always in the interval [1/n, 1/2].
We add one modification to the very basic idea described in the first paragraph of this
section. Instead of always modifying the mutation rate to the rate of the best offspring,
we shall take this winner’s rate only with probability a half and else modify the mutation
rate to a random one of the two possible values (twice and half the current rate). Our
motivation for this modification is that we feel that the additional random noise will not
prevent the algorithm from adjusting the mutation rate into a direction that is more
profitable. However, the increased amount of randomness may allow the algorithm to
leave a possible basin of attraction of a locally optimal mutation rate. Observe that with
probability Θ(1/n2 ), a sequence of log2 n random modification all in the same direction
appears. Hence with this inverse-polynomial rate, the algorithm can jump from any
mutation rate to any other (with the restriction that only a discrete set of mutation
rates can appear). We note that the existence of random modifications is also exploited
in our runtime analysis, which will show that the new self-adjusting mechanism selects
mutation rates good enough to lead to the asymptotically optimal runtime among all
dynamic choices of the mutation rate for the (1+λ) EA.
In this first work proposing this mechanism, we shall not spend much effort finetuning it, but rather show in a proof-of-concept manner that it can find very good
mutation rates. In a real application, it is likely that better results are obtained by
working with three subpopulations, namely an additional one using (that is, exploiting)
the current mutation rate. Also, it seems natural that more modest adjustments of the
mutation rate, that is, multiplying and dividing the rate by a number F that is smaller
than the value F = 2 used by our mechanism, is profitable. We conduct some elementary
experiments supporting this intuition in Section 8.
1.2
Runtime Analysis for the Self-Adjusting (1+λ) EA on OneMax
To prove that the self-adjusting mechanism just presented can indeed find good dynamic
mutation rates, we analyse it in the purest possible setting, namely in the optimization
of the classic test function
OneMax : {0, 1}n → R; (x1 , . . . , xn ) 7→
n
X
xi
i=1
via the (1+λ) EA (see Algorithm 1).
The runtime of the (1+λ) EA with fixed mutation rates on OneMax is well understood [11, 22]. In particular, Gießen
show that the expected runtime
and Witt [22]
ln λ
er n ln n
(number of generations) is (1 ± o(1)) 12 · n ln
+
·
when a mutation rate of
ln λ
r
λ
r/n, r a constant, is used. Thus for λ not too large, the mutation rate determines the
leading constant of the runtime, and a rate of 1/n gives the asymptotically best runtime.
3
As a consequence of their work on parallel black-box complexities, Badkobeh, Lehre,
and Sudholt [1] showed that the (1+λ) EA with a suitable fitness-dependent mutation
rate finds the optimum of OneMax in an asymptotically better runtime of O( logn λ +
n log n
λ ),
where the improvement is by a factor of Θ(log log λ). This runtime is bestpossible among all λ-parallel unary unbiased black-box optimization algorithms. In
particular, no other dynamic choice of the mutation rate in the (1+λ) EA can achieve
an asymptotically better runtime. The way how the mutation rate depends on the
fitness in the above result, however, is not trivial. When the parent individual has
ln λ
fitness distance d, then mutation rate employed is p = max{ n ln(en/d)
, n1 }.
Our main technical result is that the (1+λ) EA adjusting the mutation rate according
to the mechanism described above has the same (optimal) asymptotic runtime. Consequently, the self-adjusting mechanism is able find on the fly a mutation rate that is
sufficiently close to the one proposed in [1] to achieve asymptotically the same expected
runtime.
Theorem 1. Let λ ≥ 45 and λ = nO(1) . Let T denote the number of generations of the
(1+λ) EA with self-adjusting mutation rate on OneMax. Then,
E(T ) = Θ
n
n log n
+
.
log λ
λ
λn
This corresponds to an expected number of functions evaluations of Θ( log
λ + n log n).
To the best of our knowledge, this is the first time that a simple mutation-based EA
achieves a super-constant speed-up via a self-adjusting choice of the mutation rate.
As an interesting side remark, our proofs reveal that a quite non-standard but fixed
mutation rate of r = ln(λ)/2 also achieves the Θ(log log λ) improvement as it implies the
bound of Θ(n/log λ) generations if λ is not too small. Hence, the constant choice r =
O(1) as studied in [22] does not yield the asymptotically optimal number of generations
unless λ is so small that the n log n-term dominates.
Lemma 1. Let λ ≥ 45 and λ = nO(1) , Let T denote the number of generations of the
(1+λ) EA with fixed mutation rate r = ln(λ)/2. Then,
E(T ) = O
n
n log n
+ √
.
log λ
λ
√
λn
This corresponds to an expected number of functions evaluations of O( log
λ + λn log n).
The paper is structured as follows: In Section 2 we give a overview over previous
analyses of the (1+λ) EA and of self-adjusting parameter control mechanism in EAs
from a theoretical perspective. In Section 3 we give the algorithm and the mutation
scheme. For convenience, we also state some key theorems that we will frequently use
in the rest of the paper. The next three sections deal with the runtime analysis of the
expected time spent by the (1+λ) EA on OneMax in each of three regions of the fitness
distance d. We label these regions the far region, middle region and near region, each
of which will be dealt with in a separate section. The proof of the main theorem and of
Lemma 1 is then given in Section 7. Finally, we conclude in Section 9.
4
2
Related Work
Since this is a theoretically oriented work on how a dynamic parameter choice speeds up
the runtime of the (1+λ) EA on the test function OneMax, let us briefly review what
is known about the theory of this EA and dynamic parameter choices in general.
2.1
The (1+λ) EA
The first to conduct a rigorous runtime analysis of the (1+λ) EA were Jansen, De Jong,
and Wegener [24]. They proved, among other results, that when optimizing OneMax a
linear speed-up exists up to a population size of Θ(log(n) log log(n)/log log log(n)), that
is, for λ = O(log(n) log log(n)/log log log(n)), finding the optimal solution takes an expected number of Θ(n log(n)/λ) generations, whereas for larger λ at least ω(n log(n)/λ)
generations are necessary. This picture was completed in [11] with a proof that the exn
n log log λ
pected number of generations taken to find the optimum is Θ( n log
λ + log λ ). The imln λ er n ln n
plicit constants were determined in [22], giving the bound of (1±o(1))( 12 n ln
ln λ + r
λ ),
for any constant r, as mentioned in the introduction.
Aside from the optimization behavior on OneMax, not too much is known for the
(1+λ) EA, or is at least not made explicit (it is easy to see that waiting times for an improvement which are larger than λ reduce by a factor of Θ(λ) compared to one-individual
offspring populations). Results made explicit are the Θ(n2 / log(n) + n) expected runtime (number of generations) on LeadingOnes [24], the worst-case Θ(n + n log(n)/λ)
expected runtime on linear functions [11], and the O(m2 (log n + log wmax )/λ) runtime
estimate for minimum spanning trees valid for λ ≤ m2 /n [31].
2.2
Dynamic Parameter Choices
While it is clear the EAs with parameters changing during the run of the algorithm (dynamic parameter settings) can be more powerful than those only using static parameter
settings, only recently considerable advantages of dynamic choices could be demonstrated
by mathematical means (for discrete optimization problems; in continuous optimization,
step size adaptation is obviously necessary to approach arbitrarily closely a target point).
To describe the different ways to dynamically control parameters, we use in the following the language proposed in Eiben, Hinterding, and Michalewicz [21] and its extension
from [10].
2.2.1
Deterministic Parameter Control
In this language, deterministic parameter control means that the dynamic choice of a
parameter does not depend on the fitness landscape. The first to rigorously analyze a
deterministic parameter control scheme are Jansen and Wegener [23]. They regard the
performance of the (1+1) EA which uses in iteration t the mutation rate 2k /n, where
k ∈ {1, 2, . . . , 2dlog2 ne−2 } is chosen such that log2 (k) ≡ t − 1 (mod dlog2 ne − 1). In other
words, they cyclically use the mutation rates 1/n, 2/n, . . . , K/n, where K is the largest
5
power of two that is less than n/2. Jansen and Wegener demonstrate that there exists an
example function where this dynamic EA significantly outperforms the (1+1) EA with
any static mutation rate. However, they also observe that for many classic problems,
this EA is slower by a factor of Θ(log n).
In [32], a rank-based mutation rate was analyzed for the (µ+1) EA. A previous experimental study [5] suggested that this is a profitable approach, but the mathematical
runtime analysis in [32] rather indicates the opposite. While there are artificial examples where a huge runtime gain could be shown and also the worst-case runtime of the
(µ+1) EA reduces from essentially nn to O(3n ), a rigorous analysis on the OneMax
function rather suggests that the high rate of offspring generated with a mutation rate
much higher than 1/n brings a significant risk of slowing down the optimization process.
For two non-standard settings in evolutionary computation, deterministic parameter
control mechanisms also gave interesting results. For problems where the solution length
is not known [4], more precisely, where the number or the set of bits relevant for the
solution quality is unknown, again random mutation rates gave good results [14, 19].
Here however, not a power-law scheme was used, but rather one based on very slowly
decreasing summable sequences. For problems where the discrete variables take many
values, e.g., the search space is {0, . . . , r − 1}n for some large r, the question is how to
change the value of an individual variable. The results in [15] suggest that a harmonic
mutation strength, that is, changing a variable value by ±i with i chosen randomly with
probability proportional to 1/i, can be beneficial. This distribution was analyzed earlier
in [7] for the one-dimensional case, where it was also shown to give the asymptotically
best performance on a OneMax type problem.
For randomized search heuristics outside evolutionary computation, Wegener [33]
showed that simulated annealing (using a time-dependent temperature) can beat the
Metropolis algorithm (using a static temperature).
2.2.2
Adaptive Parameter Control
A parameter control scheme is called adaptive if it used some kind of feedback from
the optimization process. This can be functionally dependent (e.g., the mutation rate
depends on the fitness of the parent) or success-based (e.g., a 1/5th rule).
The first to conduct a runtime analysis for an adaptive parameter control mechanism
(and show a small advantage over static choices) were Böttcher, Doerr, and Neumann [2].
They proposed to use the fitness-dependent mutation rate of 1/(LeadingOnes(x) + 1)
for the optimization of the LeadingOnes test function. They proved that with this
choice, the runtime of the (1+1) EA improves to roughly 0.68n2 compared to a time of
0.86n2 stemming from the classic mutation rate 1/n or a runtime of 0.77n2 stemming
from the asymptotically optimal static rate of approximately 1.59/n.
For the (1p+ (λ, λ)) GA, a fitness-dependent offspring population size of order λ = Θ( n/d(x) ) was suggested in [13], where d(x) is the fitness-distance
of the parent individual to the optimum. This choice improves the optimization
timep(number of fitness evaluations until the optimum is found) on OneMax from
Θ(n log(n) log log log(n)/log log(n) ) stemming from the optimal static parameter
6
choice [9] to O(n). Since in this adaptive algorithm the mutation rate p is functionally dependent on the offspring population size, namely via pp
= λ/n, the dynamic choice
of λ is equivalent to a fitness-dependent mutation rate of 1/ nd(x).
In the
aforementioned work by Badkobeh et al. [1], a fitness-dependent mutation rate
log λ
n
ln λ
of max n ln(en/d(x))
, n1 was shown to improve the classic runtime of O n log
+ n log
log λ
λ
n
to O logn λ + n log
. In [17], the (1+1) EA using a k-bit flip mutation operator together
λ
with a fitness-dependent choice of k was shown to give a performance on OneMax
that is very close to the theoretical optimum (among all unary unbiased black-box algorithms), however, this differs only by lower order terms from the performance of the
simple randomized local search heuristic (RLS). For nature-inspired algorithms other
than evolutionary ones, Zarges [34, 35] proved that fitness-dependent mutation rates
can be beneficial in artificial immune systems.
2.3
Self-adjusting and Self-adaptive Parameter Control
While all these results show an advantage of an adaptive parameter setting, it remains
questionable if an algorithm user would be able to find such a functional dependence of
the parameter on the fitness. This difficulty can be overcome via self-adjusting parameter
choices, where the parameter is modified according to a simple rule often based on the
success or progress of previous iterations, or via self-adaptation, where the parameter is
encoded in the genome and thus subject to variation and selection. The understanding of
self-adaptation is still very limited. The only theoretical work on this topic [6], however,
is promising and shows examples where self-adaptation can lead to significant speed-ups
for non-elitist evolutionary algorithms.
In contrast to this, the last years have produced a profound understanding of selfadjusting parameter choices. The first to perform a mathematical analysis were Lässig
and Sudholt [28], who considered the (1+λ) EA and a simple parallel island model together with two self-adjusting mechanisms for population size or island number, including
halving or doubling it depending on whether the current iteration led to an improvement or not. These mechanisms were proven to give significant improvements of the
“parallel” runtime (number of generations) on various test functions without increasing
significantly the “sequential” runtime (number of fitness evaluations).
In [10] it was shown that the fitness-dependent choice of λ for the (1 + (λ, λ)) GA
described above can also be found in a self-adjusting way. To this aim, another successbased mechanism was proposed, which imitates the 1/5-th rule from evolution strategies. With some modifications, this mechanism also works on random satisfiability problems [3]. For the problem of optimizing an r-valued OneMax function, a self-adjustment
of the step size inspired by the 1/5-th rule was found to find the asymptotically best
possible runtime in [16].
These results indicate that success-based dynamics work well for adjusting parameters when a monotonic relation like “if progress is difficult, then increase the population
size” holds. For adjusting a parameter like the mutation rate, it is less obvious how to do
this. For example, in the search space {0, 1}n both a too large mutation rate (creating
7
a stronger drift towards a Hamming distance of n/2 from the optimum) and a too small
mutation rate (giving a too small radius of exploration) can be detrimental. For this
reason, to obtain a self-adjusting version of their result on the optimal number k to optimize OneMax via k-bit flips [17], in [18] a learning mechanism was proposed that from
the medium-term past estimates the efficiency of different parameter values. As shown
there, this does find the optimal mutation strength sufficiently well to obtain essentially
the runtime stemming from the fitness-dependent mutation strength exhibited before.
In the light of these works, our result from the methodological perspective shows
that some of the difficulties of the learning mechanism of [18], e.g., the whole bookkeeping being part of it and also the setting of the parameters regulating how to discount
information over time, can be overcome by the mechanism proposed in this work. In
a sense, the use of larger populations enables us to adjust the mutation rate solely on
information learned in the current iteration. However, we do also use the idea of [18] to
intentionally use parameter settings which appear to be slightly off the current optimum
to gain additional insight.
3
3.1
Preliminaries
Algorithm
We consider the (1+λ) EA with self-adjusting mutation rate for the minimization of
pseudo-boolean functions f : {0, 1}n → R, defined as Algorithm 1.
The general idea of the mutation scheme is to adjust the mutation strength according
to its success in the population. We perform mutation by applying standard bit mutation
with two different mutation probabilities r/(2n) and 2r/n and we call r the mutation
rate. More precisely, for an even number λ ≥ 2 the algorithm creates λ/2 offspring with
mutation rate r/2 and with 2r each.
The mutation rate is adjusted after each selection. With probability a half, the
new rate is taken as the mutation rate that the best individual (i. e. the one with
the lowest fitness, ties broken uniformly at random) was created with (success-based
adjustment). With the other 50% probability, the mutation rate is adjusted to a random
value in {r/2, 2r} (random adjustment). Note that the mutation rate is adjusted in each
iteration, that is, also when all offspring are worse than the parent and thus the parent
is kept for the next iteration.
If an adjustment of the rate results in a new rate r outside the interval [2, n/4], we
replace this rate with the corresponding boundary value. Note that in the case of r < 2,
a subpopulation with rate less than 1 would be generated, which means flipping less
than one bit in expectation. At a rate r > n/4, a subpopulation with rate larger than
n/2 would be created, which again is not a very useful choice.
We formulate the algorithm to start with an initial mutation rate rinit . The only
assumption on rinit is to be greater than or equal to 2. The (1+λ) EA with this selfadjusting choice of the mutation rate is given as pseudocode in Algorithm 1.
8
Algorithm 1 (1+λ) EA with two-rate standard bit mutation
Select x uniformly at random from {0, 1}n and set r ← rinit .
for t ← 1, 2, . . . do
for i ← 1, . . . , λ do
Create xi by flipping each bit in a copy of x independently with probability
rt /(2n) if i ≤ λ/2 and with probability 2rt /n otherwise.
x∗ ← arg minxi f (xi ) (breaking ties randomly).
if f (x∗ ) ≤ f (x) then
x ← x∗ .
Perform one of the following two actions with prob. 1/2:
• Replace rt with the mutation rate that x∗ has been created with.
• Replace rt with either rt /2 or 2rt , each with probability 1/2.
Replace rt with min{max{2, rt }, n/4}.
Let us explain the motivation for the random adjustments of the rate. Without such
random adjustments, the rate can only be changed into some direction if a winning
offspring is generated with this rate. For simple functions like OneMax, this is most
likely sufficient. However, when the fitness of the best of λ/2 offspring, viewed as a
function of the rate, is not unimodal, then several adjustments into a direction at first
not yielding good offspring might be needed to reach good values of the rate. Here,
our random adjustments enable the algorithm to cross such a valley of unfavorable rate
values. We note that such ideas are not uncommon in evolutionary computation, with
standard-bit mutation being the most prominent example (allowing to perform several
local-search steps in one iteration to cross fitness valleys).
A different way to implement a mechanism allowing larger changes of the rate to cross
unfavorable regions would have been to not only generate offspring with rates r/2 and
2r, but to allow larger deviations from the current rate with some small probability. One
idea could be choosing for each offspring independently the rate r2−i with probability
2−|i|−1 for all i ∈ Z, i 6= 0. This should give similar results, but to us the process appears
more chaotic (e.g., with not the same number of individuals produced with rates r/2
and 2r).
The runtime, also called the optimization time, of the (1+λ) EA is the smallest t
such that an individual of minimum f -value has been found. Note that t corresponds
to a number of iterations (also called generations), where each generation creates λ
offspring. Since each of these offspring has to be evaluated, the number of function
evaluations, which is a classical cost measure, is by a factor of λ larger than the runtime as
defined here. However, assuming a massively parallel architecture that allows for parallel
evaluation of the offspring, counting the number of generations seems also a valid cost
measure. In particular, a speed-up on the function OneMax(x1 , . . . , xn ) := x1 +· · ·+xn
by increasing λ can only be observed in terms of the number of generations. Note that
9
for reasons of symmetry, it makes no difference whether OneMax is minimized (as in
the present paper) or maximized (as in several previous research papers).
Throughout the paper, all asymptotic notation will be with respect to the problem
size n.
3.2
Drift Theorems
Our results are obtained by drift analysis, which is also used in previous analyses of the
(1+λ) EA without self-adaptation on OneMax and other linear functions [11, 22].
The first theorems stating upper bounds on the hitting time using variable drift go
back to [25, 30]. We take a formulation from [29] but simplify it to Markov processes
for notational convenience.
Theorem 2 (Variable Drift, Upper Bound). Let (Xt )t≥0 , be random variables describing
a Markov process over a finite state space S ⊆ {0} ∪ [xmin , xmax ], where xmin > 0. Let
T be the random variable that denotes the earliest point in time t ≥ 0 such that Xt = 0.
If there exists a monotone increasing function h(x) : [xmin , xmax ] → R+ , where 1/h(x) is
integrable on [xmin , xmax ], such that for all x ∈ S with Pr(Xt = x) > 0 we have
E(Xt − Xt+1 | Xt = x) ≥ h(x)
then for all x0 ∈ S with Pr(X0 = x0 ) > 0
E(T | X0 = x0 ) ≤
xmin
+
h(xmin )
Z x0
xmin
1
dx.
h(x)
The variable drift theorem is often applied in the special case of additive drift in
discrete spaces: assuming E(Xt − Xt+1 | Xt = x; Xt > 0) ≥ for some constant , one
obtains E(T | X0 = x0 ) ≤ x0 /.
Since we will make frequent use of it in the following sections as well, we will also give
the version of the Multiplicative Drift Theorem for upper bounds, due to [12]. Again,
this is implied by the previous variable drift theorem.
Theorem 3 (Multiplicative Drift [12]). Let (Xt )t≥0 be random variables describing a
Markov process over a finite state space S ⊆ R+
0 and let xmin := min{x ∈ S | x > 0}. Let
T be the random variable that denotes the earliest point in time t ≥ 0 such that Xt = 0.
If there exist δ > 0 such that for all x ∈ S with Pr(Xt = x) > 0 we have
E(Xt − Xt+1 | Xt = x) ≥ δx ,
then for all x0 ∈ S with Pr(X0 = x0 ) > 0,
E(T | X0 = x0 ) ≤
10
1 + ln
δ
x0
xmin
.
3.3
Chernoff Bounds
For reasons of self-containedness and as a courtesy to the reader, we state two wellknown multiplicative Chernoff bounds and a lesser known additive Chernoff bound that
is also known in the literature as Bennett’s inequality.
Theorem 4 (Bennett’s inequality, Chernoff Bounds [8, Theorem 1.12, Theorem 1.10]).
P
Let X1 , . . . , Xn be independent random variables and let X = ni=1 Xi . Furthermore, let
P
b such that Xi ≤ E(Xi ) + b for all i = 1, . . . , n and σ 2 = n1 ni=1 Var(Xi ). Then, for all
γ>0
!!
!
nσ 2
bγ
γ
−1
.
1+
ln 1 +
Pr(X ≥ E(X) + γ) ≤ −
b
bγ
nσ 2
Moreover, if the Xi for all i = 1, . . . , n take values in [0, 1] then
• Pr(X ≤ (1 − δ)E(X)) ≤ exp(−δ 2 E(X)/2) for all 0 ≤ δ ≤ 1.
• Pr(X ≥ (1 + δ)E(X)) ≤ exp(−δ 2 E(X)/(2 + δ)) for all δ > 0.
3.4
Occupation Probabilities
As mentioned above, we will be analyzing two depending stochastic processes: the random decrease of fitness and the random change of the mutation rate. Often, we will prove
by drift analysis that the rate is drifting towards values that yield an almost-optimal
fitness decrease. However, once the rate has drifted towards such values, we would also
like the rates to stay in the vicinity of these values in subsequent steps. To this end, we
apply the following theorem from [27]. Note that in the paper a slightly more general
version including a self-loop probability is stated, which we do not need here.
Theorem 5 (Theorem 7 in [27]). Let a Markov process (Xt )t≥0 on R+
0 , where |Xt −
Xt+1 | ≤ c, with additive drift of at least d towards 0 be given (i. e., E(Xt − Xt+1 |
Xt ; Xt > 0) ≥ d), starting at 0 (i.e. X0 = 0). Then we have, for all t ∈ N and b ∈ R+
0,
2d
Pr(Xt ≥ b) ≤ 2e 3c (1−b/c) .
We can readily apply this theorem in the following lemma that will be used throughout the paper to bound the rate rt .
Lemma 2. If there is a point c ≥ 4 such that Pr(rt+1 < rt | rt > c) ≥ 1/2 +
for some constant > 0, then for all t0 ≥ min{t | rt ≤ c} and all b ≥ 4 it holds
Pr(rt0 ≥ c + 2b ) ≤ 2e−2b/3 .
Proof. Apply Theorem 5 on the process Xt := max{0, dlog2 (rt /c)e}. Note that this
process is on N0 , moves by an absolute value of at most 1 and has drift E(Xt − Xt+1 |
Xt ; Xt > 0) = 2. We use c := 1 and d := 2 in the theorem and estimate 1 − b ≤
−b/2.
11
4
Far Region
In this first of three technical sections, we analyze the optimization behavior of our selfadjusting (1+λ) EA in the regime where the fitness distance k is at least n/ln λ. Since
we are relatively far from the optimum, it is relatively easy to make progress. On the
other hand, this regime spans the largest number of fitness levels (namely Θ(n)), so we
need to exhibit a sufficient progress in each iteration. Also, this is the regime where
the optimal mutation rate varies most
we remark that the
√ drastically. Without proof, √
optimal rate is n for k ≥ n/2 + ω( n log λ), n/2 for k = n/2 ± o( n log λ), and then
quickly drops to r = Θ(log λ) for k ≤ n/2 − εn. Despite these difficulties, our (1+λ) EA
manages to find sufficiently good mutation rates to be able to reach a fitness distance of
k = n/ln λ in an expected number of O(n/ log λ) iterations.
Lemma 3. Let n be sufficiently large and 0 < k < n/2. We define c1 (k) = (2 ln(en/k))−1
and c2 (k) = 4n2 /(n − 2k)2 .
• If n/ln λ ≤ k and r ≤ c1 (k) ln λ, then the probability that a best offspring has been
created with rate 2r is at least 0.5005.
• Let λ ≥ 50. If n/2 ≥ r ≥ c2 (k) ln λ, then the probability that all best offspring have
been created with rate r/2 is at least 0.58.
• If r ≥ 2(1 + γ)c2 (k) ln λ, then the probability that all best offspring are worse than
the parent is at least 1 − λ−γ .
Proof. Let Q(k, i, r) be the probability that standard bit mutation with mutation rate
p = r/n creates from a parent with fitness distance k an offspring with fitness distance
at most k − i. Then
Q(k, i, r) =
x−i
k X
X
x=i y=0
k
x
!
!
n − k x+y
p (1 − p)n−x−y .
y
By comparing each component in Q(k, i, r/2) and Q(k, i, 2r) we obtain
Q(k, i, 2r)/Q(k, i, r/2) ≥ 4i
i −1.5r
(1 − 2r/n)n
≥
1
−
o(1)
4e
.
(1 − r/(2n))n
Here we notice that ln(1 − x) ≥ −x − x2 for all 0 ≤ x ≤ 1/2. Then
1 − 2r
n
r
1 − 2n
!n
1.5r
= 1−
n − 0.5r
n
−1.5rn
2(1.5r)2 n
≥ exp
−
n − 0.5r (n − 0.5r)2
!
≥ (1 − o(1))e−1.5r .
The above inequality applies λ = nO(1) . Therefore r < ln λ = O(ln n). Since Q(k, i, r)
is monotone decreasing in i, let i∗ be the largest i such that Q(k, i, 2r) ≥ 4/λ. We will
then have i∗ ≥ 2r because
!
Q(k, 2r, 2r) ≥
k
(2p)2r (1 − 2p)n
2r
12
2r · 2r (k − 1)(k − 2)
k 2r 2r
≥
·
·
1·2
k2
2r
n
2r
2c1 (k) ln λ
k
k
4
> 2r
≥
≥ .
en
en
λ
2r
· (1 − o(1))e−2r
The second inequality which involves (1−2p)n again uses the fact that r < ln λ = O(ln n).
√
This means p = o(1/ n) and then we have (1 − 2p)n ≥ exp(−2pn − 4p2 n) ≥ (1 −
o(1)) exp(−2r). The (1 − o(1)) factor and (k − 1)(k − 2)/k 2 is compensated by decreasing
(2r/2) to 1 if n is large enough. We notice that when r = 2 we have i∗ ≥ ln(λ)/(2 ln ln λ)
since
i
i
4k
1
4i e−4
ln λ
Q(k, i, 4/n) ≥
e−4 ≥
4i e−4 ≥
for i =
.
in
i ln λ
λ
2 ln ln λ
Let î = ln(λ)/(2 ln ln λ). We notice that
np − x − y
∂px+y (1 − p)n−x−y
= px+y−1 (1 − p)n−x−y
≥ 0 when x + y ≥ i ≥ np,
∂p
p−1
Q(k, i, r) is increasing in r when r ≤ i. We obtain i∗ ≥ î for all r ≥ 2 which results in
∗
4i e−1.5r ≥ exp(i∗ (ln 4 − 1.5/2)) ≥ Θ(λ1/ ln ln λ ). We also need an upper bound on i∗ .
Since r ≤ ln(λ)/2 and k ≤ n/2 implies kr/n ≤ ln(λ)/4, then Chernoff Bounds shows
that the probability that 1.65 ln λ ≥ 6.6kr/n bits being flipped from k bits is less than
exp(−(5.6/7.6)1.4 ln λ) < 1/λ. This means i∗ < 1.65 ln λ. We use this to compute the
upper bound on Q(k, i∗ , 2r). Let q(k, i, r) = Q(k, i, r) − Q(k, i + 1, r) be the probability
of that the fitness distance is decreased by i. We regard the terms in q(k, i, r) where
x − y = i. If x and y both increase by 1, the terms change by a factor of
k−x n−k−y
p2
kp
(n − k)p
r2
·
·
≤
(1
+
o(1))
·
≤
x+1
y+1
(1 − p)2
x+1
y+1
4xy
If we consider y ≥ r then xy > r2 and the factor r2 /(4xy) ≤ 1/4. The sum of these
factors for all y ≥ r is less than the geometric series with ratio 1/4. Therefore, the sum
from 0 ≤ y < r contributes to at least 2/3 of the total sum q(k, i, r). Consequently, if
we look at the first 2r terms in q(k, i∗ , 2r) and q(k, i∗ + 1, 2r) we have
q(k, i∗ + 1, 2r)
2 (k − 2r − i∗ + 1)2p
4kp
2rk
≥
·
≥
= ∗ .
∗
∗
∗
q(k, i , 2r)
3 (2r + i )(1 − 2p)
3(2/2 + 1)i
3i n
Since q(k, i∗ + 1, 2r) ≤ Q(k, i∗ + 1, 2r) and k ≥ n/ ln λ we further more have
q(k, i∗ , 2r)
3i∗ ln λ
Q(k, i∗ , 2r)
3i∗ ln λ
≤
and
≤
1
+
.
Q(k, i∗ + 1, 2r)
2r
Q(k, i∗ + 1, 2r)
2r
So finally Q(k, i∗ , 2r) ≤ 4(1 + 1.5i∗ ln λ/r)/λ and
∗
4(1 + 1.5i r ln λ ) 1
Q(k, i∗ , 2r)
Q(k, i , r/2) ≤ i∗ −1.5r ≤
· .
4 e
4i∗ e−1.5r
λ
∗
13
If λ = ω(1) then the factor before 1/λ is Θ(ln2 (λ)/λ1/ ln ln(λ) ) = o(1). Otherwise if
ln λ ≥ 130 or r ≥ 8 or r < 8 but i∗ ≥ 16 we prove Q(k, i∗ , r/2) < 0.8/λ so that with
probability less than 0.4 at least one offspring of r/2 achieve i∗ . Then with probability
at least (1 − (1 − 4/λ)λ/2 ) · (1 − 0.4) > 0.86 ∗ 0.6 > 0.51 all the best offspring have
been created with 2r. We first regard large λ. If ln λ ≥ 130, we define t = i∗ /r then
2 ≤ t ≤ (1.65 ln λ)/2 and the factor Q(k, i∗ , r/2)/λ becomes
4(1 + 1.5t ln λ)
4(1 + 1.5t ln λ)
.
≤
∗
exp(i (ln 4 − 1.5/t))
exp(î(ln 4 − 1.5/t))
If t ≤ 3 it is
4(1 + 1.5t ln λ)
4(1 + 4.5 ln λ)
≤
< 0.8
exp(î(ln 4 − 1.5/t))
exp(î(ln 4 − 1.5/2))
otherwise it is
4(1 + 1.5t ln λ)
4(1 + 1.5 ln2 λ)
≤
< 0.8.
exp(î(ln 4 − 1.5/t))
exp(î(ln 4 − 1.5/3))
If ln λ < 130 and r ≥ 8 we can bound
4(1 + 1.5 · 1.65 ln2 (λ)/r)
4(1 + 41828/r)
4(1 + 1.5i∗ ln(λ)/r)
≤
<
< 0.8.
∗
4i e−1.5r
exp(2r ln(4) − 1.5r)
exp(2r ln(4) − 1.5r)
If ln λ < 130, r < 8 but i∗ ≥ 16, then
4(1 + 1.5 · 130 · 16/2)
4(1 + 1.5i∗ ln(λ)/r)
≤
< 0.3.
4i∗ e−1.5r
416 e−12
For ln λ < 130 and r < 8, we prove that the probability that the best progress i attains
1.5r is at least 1 − 3 · 10−5 . Conditioning on 1.5r ≤ i ≤ i∗ it is obvious that i obtained
from 2r with probability at least 0.5 since q(k, i, 2r)/(q, i, r/2) = (1 − o(1))4i e−1.5r > 1.
For the remaining i ≥ i∗ + 1 which has probability no less than 1.2 · 10−3 , its very
likely that a best offspring is from 2r. This can cancel 3 · 10−5 . We first compute
Q(k, 1.5r, 2r) > 21/λ in the following way. Since 1.5r < 12 then for n large enough we
have
k
Q(k, 1.5r, 2r) ≥ (1 − o(1))
1.5r
!
2r
n
1.5r
e
−2r
(2r)1.5r
≥ (1 − o(1))
(1.5r)!
0.5r
n
k
k
en
2r
,
with (k/(en))2r ≥ (k/(en))2c1 (k) ln λ = 1/λ, (2r)1.5r /((1.5r)!) ≥ 43 /6 and (n/k)0.5r ≥
2. This means the best progress among λ/2 offspring attains 1.5r is at least 1 −
(1 − 21/λ)λ/2 > 1 − 3 · 10−5 . For i ≥ i∗ + 1 according to the definition of i∗ and
Q(k, i∗ , 2r)/Q(k, i∗ + 1, 2r) ≤ 1 + 1.5i∗ ln λ/r
4>
Q(k, i∗ + 1, 2r)
4
4
1
≥
≥
>
.
∗
λ
1 + 1.5i ln λ/r
1 + 1.5 · 16 · 130/2
391
14
Then for λ/2 offspring we have 1 − (1 − Q(k, i∗ + 1, 2r))λ/2 ≥ 1.2 · 10−3 . Moreover, use
the fact that i∗ + 1 ≥ 2r + 1 we also have
Q(k, i∗ + 1, 2r)
42r+1
45
≥
>
> 50.
Q(k, i∗ + 1, r/2)
e1.5r
e3
Therefore it is easy to bound
1 − (1 − Q(k, i∗ + 1, 2r))λ/2
1 − (1 − Q(k, i∗ + 1, 2r))λ/2
1 − (1 − Q(k, i∗ + 1, 2r))λ/2
>
>
.
Q(k, i∗ + 1, r/2) · λ/2
Q(k, i∗ + 1, 2r)/50 · λ/2
1 − (1 − Q(k, i∗ + 1, r/2))λ/2
We look at function fα (x) := (1 − (1 − x)α )/(αx) and gα (x) := (1 − x)α (1 + αx). Since
gα0 (x) < 0 for all α > 1 and 0 < x < 1, we have gα (x) < gα (0) = 1 when 0 < x < 1.
Therefore fα0 (x) = (gα−1 (x) − 1)/(αx2 ) < 0 and
fλ/2 (4/λ)
1 − (1 − Q(k, i∗ + 1, 2r))λ/2
1 − (1 − 4/λ)λ/2
1 − 1/e2
>
=
≥
> 21.
Q(k, i∗ + 1, 2r)/50 · λ/2
50
(4/λ)/50 · λ/2
1/25
This means that if the best offspring made a progress of i∗ + 1 or more, then the
conditional probability that it is from the 2r-subpopulation is at least 21/22. Finally we
can bound the probability that a best offspring is from the 2r-subpopulation by
(1 − 3 · 10−5 − 1.2 · 10−3 ) · 0.5 + 1.2 · 10−3 ·
21
> 0.5005.
22
For the second statement, let X(k, r) denote the random decrease of the fitness
distance when apply standard bit mutation with probability p = r/n to an individual
with k ones. Then E(X) = kp − (n − k)p = (2k − n)p. According to Bennett’s inequality
(Theorem 4), for any ∆ > 0 we have
Pr(X ≥ E(X) + ∆) ≤ exp −Var(X) · h
∆
Var(X)
where h(u) = (1 + u) ln(1 + u) − u. We compare h(u) with τ u2 for any constant factor τ .
Let g(u) = h(u) − τ u2 be the difference, then g(0) = g 0 (0) = 0 while g 00 (u) = 1/(1 + u) −
2τ ≥ 0 when 1/(1 + u) ≥ 2τ . This means h(u) ≥ u2 /(2u + 2). We now apply this bound
with X = X(k, 2r) and ∆ = E(X(k, r/2)) − E(X(k, 2r)) = (n − 2k)1.5r/n > 0. We
have Var(X(k, 2r)) = n(2p)(1 − 2p) = 2r(1 − 2r/n) and ∆/Var = (3/4)(n − 2k)/(n − 2r).
Then,
∆2
Pr(X(k, 2r) ≥ E(X(k, r/2))) ≤ exp −
(2 + 2u)Var(X(k, 2r))
9(n − 2k)2 r
= exp −
4n(7n − 8r − 6k)
!
9(n − 2k)2 c2 (k) ln λ
≤ exp −
28n2
!
!
=
1
λ9/7
.
We notice that we have 7n − 8r − 6k > 7n − 4n − 3n = 0 in the second inequality.
Therefore, with probability less than λ−9/7 (λ/2) < 50−2/7 /2 < 0.2 the best offspring
15
of rate 2r is better than the expectation of rate r/2. For rate r/2, let X + and X − be
the number of one-bits flipped and zero-bits flipped, respectively. Both X + and X −
follow a binomial distribution and X = X + − X − . We know E(X + ) = kr/(2n) and
E(X − ) = (n − k)r/(2n) ≥ ln λ. The median of X + is between bE(X + )c to dE(X + )e
by [26]. This means Pr(X + ≥ E(X + ) − 1) ≥ Pr(X + ≥ bE(X + )c) ≥ 1/2. For ln λ
large enough, we can use a normal distribution to approximate X − and have Pr(X − ≤
E(X − ) − 1) = 1/2 − o(1). Otherwise we notice that ln λ ≥ ln 50 > 3 and r/(2n) ≤ 1/4.
Let real number t denote E(X − ) and m denote n − k. It is clear that
Pr(X − ≤ t − 1) = Pr(X − ≤ bt − 1c) ≥ Pr(X − ≤ dte − 2)
= Pr(X − ≤ dte) − Pr(X − = dte) − Pr(X − = dte − 1).
We see that Pr(X − ≤ dte) ≥ 1/2 and
Pr(X − = dte)
(m − t + 1)t
t
1
4
=
<
<
<
Pr(X − = dte − 1)
dte · m · (1 − r/(2n))
dte(1 − r/(2n))
1 − r/(2n)
3
and
(m − t + 2)t
t
16
Pr(X − = dte − 1)
=
<
< .
Pr(X − = dte − 2)
(dte − 1) · m · (1 − r/(2n))
(dte − 1)(1 − r/(2n))
9
Then we obtain
Pr(X − ≤ E(X − ) − 1) >
1
1
·
> 0.097
2 1 + 16/9 + (16/9)(4/3)
Therefore Pr(X(k, r/2) ≥ E(X(k, r/2))) ≥ 0.048, with probability at least 1 − (1 −
0.048)λ/2 > 0.7 one offspring beats E(X(k, r/2)). Therefore all best offspring are from
r/2 with probability at least 0.7 · (1 − 0.2) > 0.56, this proves the second statement of
the lemma.
An offspring of mutation rate r/n is not worse than the parent if and only if X(k, r) ≥
0 and, again by using Bennett’s inequality, we have
−E(X)
Pr(X ≥ 0) ≤ exp −Var(X) · h
Var(X)
(n − 2k)2 r
≤ exp −
2n(2n − 2k − r)
!
(n − 2k)2 r
≤ exp −
4n2
!
.
Therefore if r = 2(1 + γ)c2 (k) ln λ, then the probability above for r/2 is 1/λ1+γ and for
2α is 1/λ4+4γ . This proves the third statement.
The lemma above shows that the rate r is attracted to the interval [c1 (k) ln n, c2 (k) ln n].
Unfortunately, we cannot show that we obtain a sufficient progress in the fitness for exactly this range of r-values. However, we can do so for a range smaller only by constant
factors. This is what we do now (for large values of k) and in Lemma 5 (for smaller
values of k). This case distinction is motivated by the fact that c2 (k) becomes very large
16
when k approaches n/2. Having a good drift only for such a smaller range of r-values
is not a problem since the random movements of r let us enter the smaller range with
constant probability, see Theorem 6 and its proof.
Let ∆ := ∆(λ/2, k, r) denote the fitness gain after selection among the best of λ/2
offspring generated with rate r from a parent with fitness distance k := OneMax(x).
Let x(i) , i ∈ {1, . . . , λ/2}, be independent offspring generated from x by flipping each
bit independently with probability r/n. Then the random variable ∆ is defined by
∆ := max{0, k − min{OneMax(x(i) ) | i ∈ {1, . . . , λ/2}}}.
We next show that a narrow region contained in c1 (k) ln λ and c2 (k) ln λ provides at
least logarithmic drift on fitness. We first do the proof for large k because c2 (k) will be
ω(1) when k is close to n/2.
Lemma 4. Let 2n/5 ≤ k < n/2 and n be large enough. Let c be such that c ≤
min{n2 /(50(n − 2k)2 ), n/(4 ln λ)} and c ≥ 1/2. Let r = c ln λ ≥ 1 and λ ≥ 2, then
E(∆) ≥ 10−3 ln λ.
Proof. Note that if k ≥ 2n/5 then n2 /(50(n − 2k)2 ) ≥ 1/(50 · 0.22 ) = 1/2. We look at
the number X of flips in k ones. This random variable follows
a binomial distribution
Bin(k, p) where p = r/n. Assume u = kp ≥ 1, B(x) = xk px (1 − p)k−x and F (x) =
Pk
i≥x B(i). If we use a normal distribution to approximate X, it’s not hard to see that
the probability of hitting the mean satisfies F (u) ≥ Ω(1). More specifically, if u < 1 then
F (u) = 1 − B(0) = 1 − (1 − p)n ≥ 1 − 1/e > 1/2. Otherwise if u ≥ 1 then the worst case
for F (u) is u = 1+o(1) and F (u) ≥ 1−(1−1/n)n −(1−1/n)n−1 > 1/4. We now estimate
F (u + δ). If F (u + δ) < F (u)/2 then F (u) − F (u + δ) ≥ F (u)/2 = 1/8. When comparing
p
F (u + δ) − F (u + 2δ) to F (u) − F (u + δ) we first notice that B(x + 1)/B(x) = k−x
x+1 · 1−p
which means B(0) < B(1) < · · · < B(u) > B(u + 1) > · · · > B(k). Comparing B(u) to
B(u + 2δ) we see that
B(u + 2δ)
(k − u) · · · (k − u − 2δ + 1)
p2δ
=
·
B(u)
(u + 1) · · · (u + 2δ)
(1 − p)2δ
≥
We take δ ≤
√
k − u − 2δ
k(1 − p)
2δ
u2δ
(u + 1) · · · (u + 2δ)
u/10 for u = Θ(k) and δ ≤ 1.4u otherwise for u = o(k) and then bound
k − u − 2δ
k(1 − p)
2δ
2δ
= 1−
k−u
2δ
≥ 0.98.
Using Stirling’s approximation for (u + 1) · · · (u + 2δ) we have
(u + 2δ)! ≤
u + 2δ
2π(u + 2δ)
e
q
and
u! ≥
√
u+2δ
u
2πu
u
e
exp
17
1
exp
,
12(u + 2δ)
1
.
12u + 1
Then
u!
≥ (1 − o(1))
(u + 2δ)!
r
u
uu e2δ
,
u + 2δ (u + 2δ)u+2δ
and therefore
B(u + 2δ)
≥ 0.98
B(u)
r
uu+2δ e2δ
u
u + 2δ (u + 2δ)u+2δ
u+2δ
2δ
u
1−
e2δ
u + 2.8u
u + 2δ
1
u
≥ exp ln
(u + 2δ) + 2δ
2
u + 2δ
2 !
1
2δ
u .
≥ exp −
2
u
r
≥ 0.98
√
√
For u = Θ(k) we take δ = u/10, otherwise let δ = 0.49u/ c satisfying δ ≤ 1.4u
for c > 0.5 which is required by assumption. Then (2δ/u)2 u ≤ 0.98 ln λ and B(u +
2δ)/B(u) ≥ 1/(2λ0.98 ). Then F (u + δ) > δB(u + 2δ) > δB(u)/(2λ0.98 ). Notice that
δB(u) ≥ F (u) − F (u + δ) therefore F (u + δ) ≥ 1/(16λ0.98 ). With probability at least
γ = 1 − (1 − 1/(16λ0.98 ))λ/2 > 1/32 > 0.03 we have that one offspring flips at least u + δ
ones which shows that E(∆) ≥ γ(u + δ − (n − k)p) ≥ γ(δ − (n − 2k)p). We compare δ
√
√
2
2
to (n − 2k)p. If u = o(k) then
p δ = 0.49kp/ c ≥ 0.19np/ c. If c ≤ n /(50(n − 2k) ) we
have δ/((n−2k)p) ≥ 0.19/ 1/50 ≥ 1.34 and δ −(n−2k)p ≥ 0.25δ. Notice that γ ≥ 0.03
√
then E(∆) ≥ 0.03 · 0.25δ ≥ 0.03 · 0.25 · 0.19 c ln λ ≥ 10−3 ln λ. Otherwise, ifp
u = Θ(k)
2
2
then r = Θ(n) will require
n /(n − 2k) ≥ r/ln λ. This means n − 2k = O( n/log λ)
p
√
and (n − 2k)r/n = O( n/log λ) which is of lower order compared to δ = u/10. Hence,
E(∆) ≥ ln λ.
We now extend the lemma to the whole region of n/ln λ ≤ k < n/2. If k < 2n/5
the situation becomes easier because 4 ≤ c2 (k) < 100 and every r in the smaller range
[c1 (k) ln λ, ln(λ)/2] provides at least an expected logarithmic fitness increase. Together
with the previous lemma, we obtain the following statement for the drift in the whole
region n/ln λ ≤ k < n/2.
Lemma 5. Let n/ln λ ≤ k < n/2 and n be large enough. Assume λ ≥ 2. If r ∈
[c1 (k) ln λ, c2 (k) ln(λ)/200] for k ≥ 2n/5 and r ∈ [c1 (k) ln λ, ln(λ)/2] for k < 2n/5 with
c1 (k), c2 (k) defined as in Lemma 3, then E(∆) ≥ 10−3 ln(λ)/ln(en/k).
Proof. For r ≤ ln(λ)/2 we consider the probability Q(k, i, r) of creating from a parent
with distance k an offspring with fitness distance at least k − i and i := c1 (k) ln λ =
0.5 ln(λ)/ln(en/k), via standard bit mutation with probability r/n.
!
Q(k, i, r) ≥
k
(p)i (1 − p)n ≥
i
k r
·
i n
i
e
−r
18
≥
k
en
i
e−r ≥ e− ln(λ)/2 e−r ≥ e− ln λ =
1
.
λ
In the second greater or equal symbol we apply (1 − p)n ≥ exp(−pn − p2 n) ≥ (1 −
o(1)) exp(−r) for p satisfying
p2 n = o(1). The (1 − o(1))
factor will be compensated by
k
i
using (k/i) to estimate i . Hence, Pr ∆ ≥ c1 (k) ln λ ≥ 1 − (1 − 1/λ)λ/2 ≥ 1 − e−1/2 >
0.3 and consequently E(∆) > 0.15 ln(λ)/ln(en/k). For r > ln(λ)/2, note that this occurs
only for k > 2n/5. In this case we apply Lemma 4 and obtain E(∆) > 10−3 ln λ ≥
10−3 ln(λ)/ln(en/k).
If we only consider generations that use a rate within the right region, we can bound
the expected runtime to reach k ≤ n/ ln λ by O(n/log λ) since the drift on the fitness is of
order log λ. The following theorem shows that the additional time spent with adjusting
the rate towards the right region does not change this bound on the expected runtime.
Theorem 6. The (1+λ) EA needs O(n/log λ) generations in expectation to reach a
OneMax-value of k ≤ n/ln λ after initialization.
√
Proof. We first argue quickly that it takes an expected number of at most O( n) iterations to reach a fitness distance of k < n/2. To this end, we note that if k ≥ n/2, then
the probability for an offspring to have a strictly better fitness than the parent is at least
Θ(1) for any 2 ≤ r ≤ n/4. Consequently, the expected fitness gain in one iteration is at
√
least constant. The initial fitness distance k deviates from n/2 by Ω( n) in expectation.
√
Hence, it takes O( n) generations to obtain k < n/2.
Without loss of generality we assume k < n/2 for the initial state. Our intuition is
that once we begin to use rate r bounded by c1 (k) ln λ and c2 (k) ln λ at some distance
level k, we will have a considerable drift on the OneMax-value and the strong drift on
the rate keeps r within or not far away from the bounds. After we make progress and k
decreases to a new level, the corresponding c1 and c2 decrease, and the algorithm takes
some time to readjust r into new bounds.
We consider the stochastic process Xt = log2 (rt ) and the current OneMax-value
Yt . According to Lemma 3 we have E(Xt − Xt+1 | Xt ; Xt > log2 (c2 (Yt ) ln λ)) ≥ =
Ω(1) and E(Xt+1 − Xt | Xt ; Xt < log2 (c1 (Yt ) ln λ)) ≥ = Ω(1). We pessimistically
assume that all the iterations adjusting rt make no progress. Let k0 > k1 > · · · > kN
be the OneMax-values taken by Yt until kN hits n/ln λ. According to the additive
drift analysis from Theorem 2 it takes at most O(log n) iterations to bound rt /ln λ by
c1 (k0 ) and c2 (k0 ), no matter how we set the initial rate. Consider ti to be the last
time that Yt = ki . This means rt makes a progress of ki − ki+1 > 0. Referring to
Lemma 3 we know that rt ≤ 2(1 + γ)c2 (ki ) ln λ with probability at least 1 − λ−γ . For
rt ≤ 2(1 + γ)c2 (ki ) ln λ, according to the additive drift theorem, it takes an expected
number of at most O(log(2(1 + γ)c2 (ki )/c2 (ki+1 )) iterations to reach rt+1 ≤ c2 (Yt ) ln λ.
When we sum up the different expectation on γ = 1, 2, . . . , the increase of log(1 + γ) can
be neglect comparing to the probability decrease on (1 − λ−γ )0 . Thus we can bound this
runtime by O(log(4c2 (ki )/c2 (ki+1 )). The total number of iterations of the readjustment
process for rt to satisfy the upper bound is
N
−1
X
i=0
O log
4c2 (ki )
c2 (ki+1 )
= O log 4N
19
c2 (k0 )
c2 (kN )
= O(N + log n).
We then consider the iterations of the readjustment process for rt to satisfy the lower
bound. Since c1 (ki ) decreases along with ki , once c1 (k0 ) is hit, the lower bound condition
is obtained for all the following ki .
Now we compute the expected number of generations for k0 to decrease to kN . We
choose b large enough such that 2e−2b/3 ≤ 1/2 − δ/2 holds for some positive constant
δ > 0 and note that b is constant. Applying Lemma 2 we obtain Pr(rt ≥ c2 (k) ln λ+2b ) ≤
2e−2b/3 and Pr(rt ≤ c1 (k) ln λ − 2b ) ≤ 2e−2b/3 . Once rt is between c1 (k) ln λ and
c2 (k) ln λ, before k decreases to another bound level, we have that c1 (k) ln λ − 2b ≤
rt ≤ c2 (k) ln λ + 2b happens with probability at least δ. We see that there are at
most log200
steps between range c1 (k) ln λ ≤ r ≤ c2 (k) ln λ and an even smaller range
2
c1 (k) ln λ ≤ r ≤ c2 (k) ln(λ)/200 which is described in Lemma 5. If rt reaches the wider
region, it takes at most a constant number of iterations α in expectation to reach the
narrow region because our mutation scheme employs a 50% chance to perform a random
step of the mutation rate. Based on Lemma 5 the narrow region for the rate ensures
0.05 ln(λ)/ln(en/k) drift on the fitness at distance k. This contributes to an average
drift of at least 0.05 ln(λ)/ln(en/k) · δ/(1 + α) = Ω(log(λ)/log(en/k) for all random rates
at distance k. Applying Theorem 2, we can estimate the runtime as
O
1
+
log(λ)/log(e log λ)
Z n/2
n/log λ
dk
log(λ)/log(en/k)
!
n
.
=O
log λ
Details about how to compute the above integral can be found in the proof of Theorem
4 of [1]. We notice that N is the number of different k values and N must be bounded
by the above runtime. Combining the expected number of O(N + log n) iterations to
√
adjust rt and the expected number of O( n) iterations to hit k < n/2, the total runtime
is O(n/log λ) in expectation.
5
Middle Region
In this section we estimate the expected number of generations until the number of onebits has decreased from k ≤ n/ln λ to k ≤ n/λ. We first claim that the right region for
r is 1 ≤ r ≤ ln(λ)/2. Hence, the (1+λ) EA is not very sensitive to the choice of r here.
Intuitively, this is due to the fact that a total fitness improvement of only O(n/log λ)
suffices to cross the middle region, whereas an improvement of Ω(n) is needed for the
far region.
We estimate the drift of the fitness in Lemma 6 and apply that result afterwards to
estimate the number of generations to cross the region.
Lemma 6. Let n/λ ≤ k ≤ n/ln λ, λ ≥ 26 and 1 ≤ r ≤ ln(λ)/2. Then
( √ )
1 λk
E(∆) ≥ min
,
.
8 32n
n−k ≥
Proof. The
√ probability that no zero-bit flipped in a single mutation is (1 − r/n)
−r
e ≥ 1/ λ. We regard the number Z of offspring that have no flipped zeros. The
20
√
√
expectation E(Z) is at least 1/ λ·λ/2
=
λ/2. Applying Chernoff bounds√(Theorem 4),
√
we observe that Z exceeds λ0 := λ/4 with probability at least 1 − exp(− λ/16) > 1/4
since λ ≥ 26. Assuming this to happen, we look at the first λ0 offspring without flipped
zeros. For i ∈ {1, . . . , λ0 } let Xi be the number of flipped ones in the i-th offspring.
Then Xi are i.i.d. with Xi ∼ Bin(k, r/n). Let X ∗ = max{Xi }. By applying a result
on order statistics for binomially distributed random variables by Gießen and Witt [22,
Lemma 4] we obtain the following:
If λ0 kr/n ≥ α then E(X ∗ ) ≥ α/(1 + α).
Therefore, λ0 kr/n ≥ 1 implies E(X ∗ ) ≥ 1/2, otherwise 1 + α < 2 implies E(X ∗ ) ≥
λ0 kr/(2n). Thus,
√
E(X ∗ ) ≥ min{1/2, λk/(8n)}.
Hence, using the law of total probability, we obtain the lower bound on the drift for the
middle region.
We now use our result on the drift to estimate the time spent in this region. We
notice that c2 (k) = 4 + o(1) when k = o(n). This means we will have frequently often
rt ∈ [1, ln(λ)/2] which provides the drift we need.
Theorem 7. Let λ ≥ 26. Assume k ≤ n/ln λ for the current OneMax-value of the
self-adjusting (1+λ) EA. Then the expected number of generations until k ≤ n/λ is
O(n/log λ).
Proof. For k ≤ n/ ln λ the upper bound from Lemma 3 is c2 (k) = 4/(1 − 2/ ln λ) < 11.
According to the lemma we have for Xt := dlog2 rt e that E(Xt − Xt+1 | Xt ; Xt >
log2 (c2 (k) ln λ)) ≥ = Ω(1). The additive drift theorem yields that in O(log n) time
we have rt ≤ c2 (k) ln λ. We choose b large enough such that 2e−2b/3 ≤ 1 − δ holds for
some positive constant δ > 0 and note that b is constant. Applying Lemma 2 we obtain
Pr(rt ≥ c2 (k) ln λ + 2b ) ≤ 2e−2b/3 and rt ≤ c2 (k) ln λ + 2b happens with probability at
least δ. Once rt < c2 (k) ln λ + 2b < 11 ln λ + 2b it takes at most a constant number of
iterations α in expectation to draw
√ rt to ln(λ)/2 or less. According to Lemma 6 this
ensures a√drift of (1/4) min{1/2, λk/(8n)}, which implies an average drift of at least
c min{1, λk/n} over all random rates at distance
√ k, where c > 0 is a constant.√The
minimum is taken on the first argument if k > n/ λ, and on the second if k < n/ λ.
We are interested in the expected time to reduce the OneMax-value to a most
n/λ. To ease the application of drift analysis, we artificially modify the process and
make it create the optimum when the state (OneMax-value) is strictly less than n/λ.
Clearly, the first hitting time of state at most n/λ does not change by this modification.
Applying the variable
√ drift theorem (Theorem 2) with xmin = n/λ, X0 = k ≤ n/ln λ
and h(x) = c min{1, λk/n}, the expected number of generations to reach state at most
n/λ is bounded from above by
n/λ
√
+
c λ(n/λ)/n
Z n/√λ
n/λ
n
√ dx +
c λx
Z n/ln λ
1
√
n/ λ
n
dx = O √
c
λ
21
n log λ
√
+O
λ
n
+O
,
log λ
which is O(n/log λ). The overall expected number of generations spent is O(log n +
n/log λ) = O(n/log λ) since λ = nO(1) by assumption.
6
Near region
In the near region, we have k ≤ n/λ. Hence, the fitness is so low that we can expect
only a constant number of offspring to flip at least one of the remaining one-bits. This
assumes constant rate. However, higher rates are detrimental since they are more likely
to destroy the zero-bits of the few individuals flipping one-bits. Hence, we expect the
rate to drift towards constant values, as shown in the following lemma.
Lemma 7. Let k ≤ n/λ, λ ≥ 45 and 4 ≤ rt ≤ ln(λ)/4. Then the probability that
rt+1 = rt /2 is at least 0.5099.
Proof. To prove the claim we exploit the fact that only few one-bits are flipped in
both subpopulations. Using r := rt , we shall argue as follows. With sufficiently high
(constant) probability, (i) the 2r-subpopulation contains no individual strictly better
than the parent, that is, with fitness less than k, and (ii) all 2r-offspring with fitness
r are identical to the parent. Conditional on this, either the r/2-population contains
individuals with fitness less than k and the winning individual surely stems from this
subpopulation, or the r/2-population contains no better offspring. In the latter case, we
argue that there are many more individuals with fitness exactly k in the r/2-population
than in the 2r-population, which gives a sufficiently high probability for taking the
winning individual from this side (as it is chose uniformly at random from all offspring
with fitness k).
Let Nr/2 and N2r be the number of offspring that did not flip any zero-bits using rate
r/2 and 2r, respectively. Then E(Nr/2 ) = (λ/2)(1 − r/(2n))n−k ≥ (1 − o(1))λe−r/2 /2,
since k ≥ 1 and
r
1−
2n
n−1
≥e
− r2
r
1−
2n
r −1
2
− r2
≥e
c ln n
1−
8n
c ln n −1
4
r
= (1 − o(1))e− 2 ,
where we used that r ≤ ln λ/4 and λ = nO(1) , i. e. ln λ ≤ c ln n for some constant c.
Using k ≤ n/λ we get E(N2r ) = (λ/2)(1 − 2r/n)n−k ≤ (λ/2)e−2r(1−1/λ) . In fact, we can
discriminate N1 and N2 by using Theorem 4 in the following way: we have
Pr Nr/2
λ r
≤ e− 2
4
λ −r
e 2
≤ exp −(1 − o(1))
16
1 7/8
≤ exp −(1 − o(1)) λ
16
< 0.175,
3
for sufficiently large n, since r ≤ (ln λ)/4 and λ ≥ 45. Similarly, we obtain
1
1
1
1
Pr N2r ≥ λe−2r(1− λ ) ≤ exp − λ1− 2 (1− λ ) < 0.312.
6
22
Note that e−2r(1−1/λ) < e−r/2 /4 holds for all r ≥ 4 and λ ≥ 45. Since the offspring
1
are generated independently, the events Nr/2 > λe−r/2 /4 and N2r < λe−2r(1− λ ) happen
together with probability at least (1−0.175)·(1−0.312) ≥ 0.567 =: 1−perr1 . Conditioning
on this and by using a union bound the probability perr2 that at least one of the N2r
offspring that do not flip any zero-bits flips at least one one-bit can be upper bounded
by
1
2kr
perr2 := N2r ·
≤ 2re−2r(1− λ ) ≤ 0.004
n
using k/n ≤ 1/λ in the first and r ≥ 4 and λ ≥ 45 in the last inequality. By using a
union bound, we find the probability perr3 that at least one 2r-offspring flips exactly one
one-bit and exactly one zero-bit to be at most
λ 2rk
2r k−1 2r(n − k)
2r n−k−1
:=
1−
1−
2 n
n
n
n
n−2
6
2r
≤ 2r2 1 −
≤ (1 + o(1))2e2(ln(r)−r) < (2 + o(1))e− 5 r < 0.017,
n
perr3
for sufficiently large n, using k/n ≤ 1/λ for the first inequality. The third inequality is
due to ln x − x ≤ −(1 − e−1 )x < −(3/5)x for all x > 0 and the last inequality stems
from r ≥ 4. The second inequality follows from
1−
2r
n
n−2
2
2r
≤ e− n (n−2) = e−2r(1− n ) = (1 + o(1))e−2r ,
using again r ≤ ln λ ≤ c ln n for some constant c. Let Mr be the number of such
offspring. Any other fitness-decreasing flip-combinations of zeroes and ones in the 2rsubpopulation require an offspring to flip at least two one-bits. The probability that
such an offspring is created is at most
perr4
λ k
:=
2 2
!
2r
n
2
≤
ln2 λ
< 0.021,
16λ
using k ≤ n/λ and r ≤ ln λ/4 and the fact that (ln2 x)/x is decreasing for x ≥ e2 and
λ ≥ 45 > e2 .
1
The events Nr/2 > λe−r/2 /4, N2r < λe−2r(1− λ ) , Mr = 0, and the event that no
fitness-decreasing offspring is created in the 2r-subpopulation are sufficient to ensure
that the best individual is either surely from the r/2-population or chosen uniformly at
random from the Nr/2 + N2r offspring. Conditioning on these events, we have that the
probability that the best offspring is chosen from the r/2-population is at least
Nr/2
1
≥
> 0.988.
1
−r(2(1− λ
)− 12 )
Nr/2 + N2r
1 + 4e
Hence, using a union bound for the error probabilities, the unconditional probability is
at least
(1 − perr1 − perr2 − perr3 − perr4 ) · 0.988 + 12
> 0.5099.
2
23
We note that the restriction rt ≥ 4 in the lemma above is not strictly necessary.
Also for smaller rt , the probability that the winning individual is chosen from the rt /2population is by an additive constant larger than 1/2. Showing this, however, would need
additional proof arguments as for smaller rt , the event that both subpopulations contain
individuals with fitness k − 1 becomes more likely. We avoid this additional technicality
by only arguing for rt ≥ 4, which is enough since any constant rt is sufficient for the
fitness drift we need (since we do not aim at making the leading constant precise).
In the following proof of the analysis of the near region, we use the above lemma
(with quite some additional arguments) to argue that the r-value quickly reaches 4 or
less and from then on regularly returns to this region. This allows to argue that in the
near region we have a speed-up of a factor of Θ(λ) compared to the (1+1) EA, since
every offspring only has a probability of O(1/λ) of making progress (see also [11, 22]).
Theorem 8. Assume k ≤ n/λ for the current OneMax-value of the self-adjusting
(1+λ) EA. Then the expected number of generations until the optimum is reached is
O(n log(n)/λ + log n).
Proof. The aim is to estimate the OneMax-drift at the points in time (generations) t
where rt = O(1). To bound the expected number of generations until the mutation rate has entered this region, we basically consider the stochastic process Zt :=
max{0, dlog2 (rt /c)e}, where c := 4, which is the lower bound on rt from Lemma 7.
However, as we do not have proved a drift of Zt towards smaller values in the region
L := (ln λ)/4 ≤ rt ≤ 16 ln λ =: U (where 16 is an upper bound on c2 (k) from Lemma 5),
we use the potential function
Z t
if c ≤ rt ≤ L
P 2 (rt /L)e −i
Xt (Zt ) := log2 (L/c) + dlog
4
i=1
Plog2 (U/L) −i
log2 (L/c) +
i=1
4
+
if L < rt < U
4− log2 (U/L) dlog
2 (rt /U )e
otherwise.
assuming that L and U have been rounded down and up to the closest power of 2,
respectively.
The potential function has a slope of 1 for c ≤ rt ≤ L. Lemma 7 gives us the drift
E(Xt − Xt+1 | Xt ; c ≤ rt ≤ L) ≥ 0.5099 − 0.4901 = 0.0198. The function satisfies
Xt (Zt ) − Xt (Zt − 1) ≥ 4(Xt (Zt ) − Xt (Zt + 1)) if L < rt < U , which corresponds
to the region where the probability of decreasing Zt by 1 has only be bounded from
below by 1/4 due to the random steps. Still, E(Xt − Xt+1 | Xt ; L < rt < U ) ≥
(1/4)4log2 (U/L)+1 − (3/4)4log2 (U/L) = Ω(1) in this region due to the concavity of the
potential function. Finally, E(Xt − Xt+1 | Xt ; rt ≥ U ) = 4− log2 (U/L) Ω(1) = Ω(1) by
Lemma 5. Hence, altogether E(Xt − Xt+1 | Xt ; rt ≥ c) ≥ κ for some constant κ > 0. As
X0 = O(log n), additive drift analysis yields an expected number of O(log n) generations
until for the first time Xt = 0 holds, corresponding to rt ≤ c. We denote this hitting
time by T .
We now consider an arbitrary point of time t ≥ T . The aim is to show a drift on the
OneMax-value, depending on the current OneMax-value Yt , which satisfies Yt ≤ n/λ
24
with probability 1. To this end, we will use Lemma 2. We choose b large enough such
that 2e−2b·κ/4 ≤ 1 − δ holds for some positive constant δ > 0 and note that b is constant.
We consider two cases for rt . If rt ≤ c + 2b , which happens with probability at least δ,
we have rt = O(1) and obtain a probability of at least
Yt
1− 1−
1
!
rt
n
rt
1−
n
n−1 !λ
≥1− 1−Θ
Yt
n
λ
= Ω(λYt /n)
to improve the OneMax-value by 1, using that Yt = O(n/λ). If rt > c + 2b , we bound
the improvement from below by 0. Using the law of total probability, we obtain
E(Yt − Yt+1 | Yt ; Yt ≤ n/λ) = δΩ(λYt /n) = Ω(λYt /n).
Now a straightforward multiplicative drift analysis (Theorem 3 using δ = Θ(λ/n)) gives
an expected number of O((n/λ) log Y0 ) = O(n log(n)/λ) generations until the optimum
is found. Together with the expected number O(log n) until the r-value becomes at
most c, this proves the theorem.
7
Putting Everything Together
In this section, we put together the analyses of the different regimes to prove our main
result.
Proof of Theorem 1. The lower bound actually holds for all unbiased parallel black-box
algorithms, as shown in [1].
We add up the bounds on the expected number of generations spent in the three
regimes, more precisely we add up the bounds from Theorem 6, Theorem 7 and Theorem 8, which gives us O(n/log λ+n log(n)/λ+log n) generations. Due to our assumption
λ = nO(1) the bound is dominated by O(n/log λ + n log(n)/λ) as suggested.
Proof of Lemma 1. We basically revisit the regions of different OneMax-values analyzed in this paper and bound the time spent in these regions under the assumption
r = ln(λ)/2. In the far region, Lemmas 4 and 5, applied with this value of r, imply a
fitness drift of Ω(log(λ)/log(en/k)) per generation, so the expected number of generations spent in the far region is O(n/log λ) as computed by variable drift analysis in the
proof of Theorem 6.
√
The middle region is shortened at the lower end. For k ≥ n/ λ, Lemma 6 gives a
fitness drift of Ω(1), implying
√ by additive drift analysis O(n/log λ) generations to reduce
the fitness to at most n/ λ.
√
In the near region, which now starts at n/ λ, we have to argue slightly differently.
n
− ln(λ)/2+O(1) = Ω(λ−1/2 )
Note that every offspring has a probability of at least
√ (1−r) ≥ e
of not flipping a zero-bit. Hence, we expect Ω( λ) such offspring. We pessimistically
assume that the other individuals√do not yield a fitness improvement; conceptually, this
reduces the population size to Ω( λ) offspring, all of which are guaranteed not to flip a
25
3000
2500
2000
1500
1000
Average runtime over 10000 runs
Self-adj. (1+λ) EA
Self-adj. (1+λ) EA (no random steps, F = 1.2)
Static (1+λ) EA (p = ln(λ)/(2n))
Static (1+λ) EA (p = 1/n)
(1+λ) EA using p = max(1/n, ln(λ)/(ln(en/d(x))))
200
400
600
800
1000
λ
Figure 1: Static and Self-adjusting (1+λ) EA average runtime comparison (n = 5000)
on OneMax
zero-bit. Adapting the arguments from the proof of Theorem 8, the probability that at
least of one of these individuals flips at least a one-bit is at least
Yt
1− 1−
1
!
rt
n
√
!Ω( λ)
Yt
≥1− 1−Θ
n
Ω(√λ)
√
= Ω( λYt /n),
which is a lower bound on the fitness drift. Using the multiplicative
drift analysis, the
√
expected number of generations in the near region is O(n log(n)/ λ). Putting the times
for the regions together, we obtain the lemma.
8
Experiments
Since our analysis is asymptotic in nature we performed some elementary experiments in
order to see whether besides the asymptotic runtime improvement (showing an improvement for an unspecified large problem size n) we also see an improvement for realistic
problem sizes. For this purpose we implemented the (1+λ) EA in C using the GNU
Scientific Library (GSL) for the generation of pseudo-random numbers.
The plot in Figure 1 displays the average runtime over 10000 runs of the self-adjusting
(1+λ) EA on OneMax for n = 5000 as given in Algorithm 1 over λ = 100, 200, . . . , 1000.
We set the initial mutation rate to 2, i. e., the minimum mutation rate the algorithm can
26
attain. Moreover, the plot displays the average runtime of the classic (1+λ) EA using a
static mutation probability of 1/n.
The average runtimes of both algorithms profit from higher offspring population sizes
λ leading to lower average runtimes as λ increases. Interestingly, the classic (1+λ) EA
outperforms the self-adjusting (1+λ) EA for small values of λ up to λ = 400. For
higher offspring population sizes the self-adjusting (1+λ) EA outperforms the classic
one, indicating that the theoretical performance gain of ln ln λ can in fact be relevant
in practice. Furthermore, we implemented the self-adjusting (1+λ) EA without the
random steps, that is, when the rate is always adjusted according to how the best
offspring are distributed over the two subpopulations. The experiments show that this
variant of the self-adjusting (1+λ) EA performs generally slightly better on OneMax.
Since the OneMax fitness landscape is structurally very simple, this result is not totally
surprising. It seems very natural that the fitness of the best of λ/2 individuals, viewed
as a function in the rate, is a unimodal function. In this case, the advantage of random
steps to be able to leave local optima of this function is not needed. On the other hand,
of course, this observation suggests to try to prove our performance bound rigorously
also for the case without random rate adjustments. We currently do not see how to do
this. Lastly, we implemented the (1+λ) EA using the fitness-depending mutation rate
ln λ
p = max{ n ln(en/d)
, n1 } as presented in [1]. The experiments suggest that this scheme
outperforms all other variants considered.
Additionally we implemented another variant of the self-adjusting (1+λ) EA using
three equally-sized subpopulations i. e. the additional one is using (that is, exploiting)
the current mutation rate. We compared this variant with the self-adjusting (1+λ) EA,
both with and without using random steps. The results are shown in Figure 2. The
experiments suggest that the variant using three subpopulations outperforms the selfadjusting (1+λ) EA slightly for small population sizes. For very high population sizes,
using just two subpopulations seems to be a better choice.
To gain some understanding on how the parameters influence the runtime, we implemented the self-adjusting (1+λ) EA using different mutation rate update factors, that
is, we consider the self-adjusting (1+λ) EA as given in Algorithm 1 where the mutation
rate rt is increased or decreased by some factor F (instead of the choice F = 2 made
in Algorithm 1). Note that we do not change the rule that we use the rates r/2 and
2r to create the subpopulations. Furthermore, after initialization, the algorithm starts
with rate F and the rate is capped below by F and above by 1/(2F ) during the run,
accordingly.
The results are shown in Figure 3. The plot displays the average runtime over 10000
runs of the self-adjusting (1+λ) EA on OneMax for n = 5000 over λ = 100, 200, . . . , 1000
using the update factors F = 2.0, 1.5, 1.01. The plot suggests that lower values of F yield
a better performance. This result is not immediately obvious. Clearly, a large factor F
implies that the rate changes a lot from generation to generation (namely by a factor
of F ). These changes prevent the algorithm from using a very good rate for several
iterations in a row. On the other hand, a small value for F implies that it takes longer
to adjust the rate to value that is far from the current one.
27
3000
EA
EA, 3 subpop.
EA (no random steps)
EA, 3 subpop. (no random steps))
2000
(1+λ)
(1+λ)
(1+λ)
(1+λ)
1000
Average runtime over 10000 runs
Self-adj.
Self-adj.
Self-adj.
Self-adj.
200
400
600
800
1000
1200
1400
λ
3000
2500
F = 2.0
F = 1.5
F = 1.2
1000 1500 2000
Average runtime over 10000 runs
Figure 2: Average runtime of the self-adjusting (1+λ) EA with two and three subpopulations each with and without random steps on OneMax (n = 5000)
200
400
600
800
1000
λ
Figure 3: Average runtime of the self-adjusting (1+λ) EA with different mutation rate
update factors F on OneMax (n = 5000)
28
600
300
0
100
200
rate
400
500
F = 1.1
F = 1.05
F = 1.01
0
10000
20000
30000
40000
50000
fitness
Figure 4: Development of the rate over the fitness of three example runs of the selfadjusting (1+λ) EA on OneMax (n = 100000, λ = (TODO, I forgot and have to check
at home)) using different factors F
29
Finally, to illustrate the nontrivial development of the rate during a run of the algorithm we plotted the rate of three single runs of the self-adjusting (1+λ) EA using
different factors F over the fitness in Figure 4. Since the algorithm initialized with the
rate F , the rate increases after initialization decreases again with decreasing fitnessdistance to the optimum. The plot suggests that for higher values of F the rate is more
unsteady due to the greater impact of the rate adjustments while smaller rate updates
yield a more stable development of the rate. Interestingly, for all three values of F , the
rates seem to correspond to the same rate after the initial increasing phase from F . Note
that this illustration does not indicate the actual runtime. In fact, the specific runtimes
are 19766 for F = 1.01, 19085 for F = 1.05 and 19857 for F = 1.1. A similar, more
pronounced behaviour can be seen for F = 2.0; we chose these particular values of F for
illustrative purposes since for F = 2.0 the variance in the rate can be visually confusing
for the reasons given above.
While we would draw from this experiment the conclusion that a smaller choice of F
is preferable in a practical application of our algorithm, the influence of the parameter
on the runtime is not very large. So it might not be worth optimizing it and rather view
Algorithm 1 as a parameter-less algorithm.
9
Conclusions
We proposed and analyzed a new simple self-adjusting mutation scheme for the (1+λ) EA.
It consists of creating half the offspring with a slightly larger and the rest with a slightly
smaller mutation rate. Based on the success of the subpopulations, the mutation rate
is adjusted. This simple scheme overcomes difficulties of previous self-adjusting choices,
e.g., the careful choice of the exploration-exploitation balance and the forgetting rate in
the learning scheme of [18].
We proved rigorously that this self-adjusting (1+λ) EA optimizes the OneMax
test function in an expected number of O(nλ/ log λ + n log n) fitness evaluations. This
matches the runtime shown in [1] for a careful fitness-dependent choice of the mutation
rate, which was also shown to be asymptotically optimal among all λ-parallel blackbox optimization algorithms. Hence our runtime result indicates that the self-adjusting
mechanism developed in this work is able to find very good mutation rates. To the best
of our knowledge, this is the first time that a self-adjusting choice of the mutation rate
speeds up a mutation-based algorithm on the OneMax test function by more than a
constant factor.
The main technical challenge in this work is to analyze the quality of the best offspring. In contrast to most previous runtime analyses, where only the asymptotic order
of the fitness gain was relevant, we needed a much higher degree of precision as we
needed to make statements about in which sub-population the best offspring is, or, in
case of multiple best offspring, how they are distributed over the two subpopulations.
Note that the quality of the best offspring is not as strongly concentrated around its
expectation as, e.g., the average quality.
30
As a side-result of our analyses, we have observed that using a fixed rate of r =
ln(λ)/2 gives the bound O(n/log λ+n log(n)/λ1/2 ), which is also asymptotically optimal
unless λ is small. However, this setting is far off the usual constant choice of r. It is the
first time that a significantly larger mutation rate was shown to be useful in a simple
mutation-based algorithm for a simple fitness landscape. Previously, it was only observed
that larger mutation rates can be helpful to leave local optima [20].
From this work, a number of open problems arise. A technical challenge is to prove
that our algorithm also without the random rate adjustments performs well. This requires an even more precise analysis of the qualities of the offspring in the two subpopulations, for which we currently do not have the methods. From the view-point of
understanding the mutation rate for population-based algorithms, two interesting question are (i) to what extent our observation that larger mutation rates are beneficial for
the (1+λ) EA on OneMax generalizes to other algorithms and problems, and (ii) for
which other problems our self-adjusting choice of the mutation rate gives an improvement
over the classic choice of 1/n or other static choices.
Acknowledgments This work was supported by a public grant as part of the Investissement d’avenir project, reference ANR-11-LABX-0056-LMH, LabEx LMH, and
by a grant by the Danish Council for Independent Research (DFF-FNU 4002–00542).
References
[1] Golnaz Badkobeh, Per Kristian Lehre, and Dirk Sudholt. Unbiased black-box complexity of parallel search. In Proc. of PPSN ’14, pages 892–901. Springer, 2014.
[2] Süntje Böttcher, Benjamin Doerr, and Frank Neumann. Optimal fixed and adaptive
mutation rates for the LeadingOnes problem. In Proc. of PPSN ’10, pages 1–10.
Springer, 2010.
[3] Maxim Buzdalov and Benjamin Doerr. Runtime analysis of the (1 + (λ, λ)) genetic
algorithm on random satisfiable 3-CNF formulas. In Proc. of GECCO ’17. ACM,
2017. To appear. Full version available at http://arxiv.org/abs/1704.04366.
[4] Stephan Cathabard, Per Kristian Lehre, and Xin Yao. Non-uniform mutation rates
for problems with unknown solution lengths. In Proc. of FOGA ’11, pages 173–180.
ACM, 2011.
[5] Jorge Cervantes and Christopher R. Stephens. Rank based variation operators for
genetic algorithms. In Proc. of GECCO ’08, pages 905–912. ACM, 2008.
[6] Duc-Cuong Dang and Per Kristian Lehre. Self-adaptation of mutation rates in
non-elitist populations. In Proc. of PPSN ’16, pages 803–813. Springer, 2016.
[7] Martin Dietzfelbinger, Jonathan E. Rowe, Ingo Wegener, and Philipp Woelfel. Tight
bounds for blind search on the integers and the reals. Combinatorics, Probability &
Computing, 19:711–728, 2010.
31
[8] Benjamin Doerr. Analyzing randomized search heuristics: tools from probability
theory. In Anne Auger and Benjamin Doerr, editors, Theory of Randomized Search
Heuristics, pages 1–20. World Scientific Publishing, 2011.
[9] Benjamin Doerr. Optimal parameter settings for the (1 + (λ, λ)) genetic algorithm.
In Proc. of GECCO ’16, pages 1107–1114. ACM, 2016.
[10] Benjamin Doerr and Carola Doerr. Optimal parameter choices through selfadjustment: applying the 1/5-th rule in discrete settings. In Proc. of GECCO ’15,
pages 1335–1342. ACM, 2015.
[11] Benjamin Doerr and Marvin Künnemann. Optimizing linear functions with the
(1+λ) evolutionary algorithm – different asymptotic runtimes for different instances.
Theoretical Computer Science, 561:3–23, 2015.
[12] Benjamin Doerr, Daniel Johannsen, and Carola Winzen. Multiplicative drift analysis. Algorithmica, 64:673–697, 2012.
[13] Benjamin Doerr, Carola Doerr, and Franziska Ebel. From black-box complexity to
designing new genetic algorithms. Theoretical Computer Science, 567:87–104, 2015.
[14] Benjamin Doerr, Carola Doerr, and Timo Kötzing. Solving problems with unknown
solution length at (almost) no extra cost. In Proc. of GECCO ’15, pages 831–838.
ACM, 2015.
[15] Benjamin Doerr, Carola Doerr, and Timo Kötzing. The right mutation strength for
multi-valued decision variables. In Proc. of GECCO ’16, pages 1115–1122. ACM,
2016.
[16] Benjamin Doerr, Carola Doerr, and Timo Kötzing. Provably optimal self-adjusting
step sizes for multi-valued decision variables. In Proc. of PPSN ’16, pages 782–791.
Springer, 2016.
[17] Benjamin Doerr, Carola Doerr, and Jing Yang. Optimal parameter choices via
precise black-box analysis. In Proc. of GECCO ’16, pages 1123–1130. ACM, 2016.
[18] Benjamin Doerr, Carola Doerr, and Jing Yang. k-bit mutation with self-adjusting k
outperforms standard bit mutation. In Proc. of PPSN ’16, pages 824–834. Springer,
2016.
[19] Benjamin Doerr, Carola Doerr, and Timo Kötzing. Unknown solution length problems with no asymptotically optimal run time. In Proc. of GECCO ’17. ACM, 2017.
To appear.
[20] Benjamin Doerr, Huu Phuoc Le, Régis Makhmara, and Ta Duy Nguyen. Fast
genetic algorithms. In Proc. of GECCO ’17. ACM, 2017. To appear. Full version
available at http://arxiv.org/abs/1703.03334.
32
[21] Agoston Endre Eiben, Robert Hinterding, and Zbigniew Michalewicz. Parameter
control in evolutionary algorithms. IEEE Transactions on Evolutionary Computation, 3:124–141, 1999.
[22] Christian Gießen and Carsten Witt. The interplay of population size and mutation
probability in the (1+λ) EA on OneMax. Algorithmica, 2017. In press, available
online at http://dx.doi.org/10.1007/s00453-016-0214-z.
[23] Thomas Jansen and Ingo Wegener. On the analysis of a dynamic evolutionary
algorithm. Journal of Discrete Algorithms, 4:181–199, 2006.
[24] Thomas Jansen, Kenneth A. De Jong, and Ingo Wegener. On the choice of the
offspring population size in evolutionary algorithms. Evolutionary Computation,
13:413–440, 2005.
[25] Daniel Johannsen. Random combinatorial structures and randomized search heuristics. PhD thesis, Saarland University, 2010.
[26] Rob Kaas and Jan M. Buhrman. Mean, median and mode in binomial distributions.
Statistica Neerlandica, 34:13–18, 1980.
[27] Timo Kötzing, Andrei Lissovoi, and Carsten Witt. (1+1) EA on generalized dynamic OneMax. In Proc. of FOGA ’15, pages 40–51. ACM, 2015.
[28] Jörg Lässig and Dirk Sudholt. Adaptive population models for offspring populations
and parallel evolutionary algorithms. In Proc. of FOGA ’11, pages 181–192. ACM,
2011.
[29] Per Kristian Lehre and Carsten Witt. Concentrated hitting times of randomized
search heuristics with variable drift. In Proc. of ISAAC ’14, pages 686–697. Springer,
2014.
[30] Boris Mitavskiy, Jonathan E. Rowe, and Chris Cannings. Theoretical analysis of
local search strategies to optimize network communication subject to preserving the
total number of links. International Journal of Intelligent Computing and Cybernetics, 2:243–284, 2009.
[31] Frank Neumann and Ingo Wegener. Randomized local search, evolutionary algorithms, and the minimum spanning tree problem. Theoretical Computer Science,
378:32–40, 2007.
[32] Pietro Simone Oliveto, Per Kristian Lehre, and Frank Neumann. Theoretical analysis of rank-based mutation - combining exploration and exploitation. In Proc. of
CEC ’09, pages 1455–1462. IEEE, 2009.
[33] Ingo Wegener. Simulated annealing beats Metropolis in combinatorial optimization.
In Proc. of ICALP ’05, pages 589–601. Springer, 2005.
33
[34] Christine Zarges. Rigorous runtime analysis of inversely fitness proportional mutation rates. In Proc. of PPSN ’08, pages 112–122. Springer, 2008.
[35] Christine Zarges. On the utility of the population size for inversely fitness proportional mutation rates. In Proc. of FOGA ’09, pages 39–46. ACM, 2009.
34
| 9cs.NE
|
arXiv:1705.08413v3 [math.ST] 17 Jul 2017
S TABLE L IMIT T HEOREMS FOR E MPIRICAL P ROCESSES
C ONDITIONAL N EIGHBORHOOD D EPENDENCE
UNDER
Ji Hyung Lee and Kyungchul Song
University of Illinois and University of British Columbia
A BSTRACT. This paper introduces a new concept of stochastic dependence among many
random variables which we call conditional neighborhood dependence (CND). Suppose
that there are a set of random variables and a set of sigma algebras where both sets are
indexed by the same set endowed with a neighborhood system. When the set of random
variables satisfies CND, any two non-adjacent sets of random variables are conditionally
independent given sigma algebras having indices in one of the two sets’ neighborhood.
Random variables with CND include those with conditional dependency graphs and a
class of Markov random fields with a global Markov property. The CND property is useful
for modeling cross-sectional dependence governed by a complex, large network. This
paper provides two main results. The first result is a stable central limit theorem for a
sum of random variables with CND. The second result is a Donsker-type result of stable
convergence of empirical processes indexed by a class of functions satisfying a certain
bracketing entropy condition when the random variables satisfy CND.
K EY WORDS . Conditional Neighborhood Dependence; Dependency Graphs; Markov Random Fields; Empirical Processes; Maximal Inequalities; Stable Central Limit Theorem
AMS MSC 2010: 60B10; 60F05; 60G57
JEL C LASSIFICATION: C12, C21, C31
Date: July 19, 2017.
We thank Denis Kojevnikov for his valuable comments. All errors are ours. Corresponding address:
Kyungchul Song, Vancouver School of Economics, University of British Columbia, 6000 Iona Drive, Vancouver, BC, Canada, V6T 1L4. Email address: [email protected]. Co-author address: Ji Hyung Lee,
Department of Economics, University of Illinois, 1407 W. Gregory Dr., 214 David Kinley Hall, Urbana, IL
61801. Email address: [email protected].
1
2
1. Introduction
Empirical processes indexed by a class of functions arise in many applications, in particular in developing asymptotic inference for nonparametric or semiparametric models
and for goodness-of-fit specification tests. (See, e.g., Andrews (1994) and Chapter 3 of
van der Vaart and Wellner (1996) for a review of applications of empirical process theory
in statistics and econometrics.) While a predominant body of the literature on empirical
process theory focuses on independent observations or time series observations, there
is relatively little research on empirical processes with spatial or cross-sectional dependence. This paper aims to contribute to the literature by providing limit theorems for
empirical processes which consist of random variables with a flexible, complex (crosssectional) dependence structure.
In this paper, we introduce a new notion of stochastic dependence among a set of
random variables. Suppose that we are given a set of random variables {Yi }i∈Nn indexed
by a set Nn , where the set Nn is endowed with a neighborhood system so that each
i ∈ Nn is associated with a subset νn (i) ⊂ Nn \{i} called the neighborhood of i. In this
paper, we call the map νn : Nn → 2Nn a neighborhood system. Given a neighborhood
system νn and a set of σ-fields M ≡ (Mi )i∈Nn , we say that {Yi }i∈Nn is conditionally
neighborhood dependent(CND) with respect to (νn , M) if for any two non-adjacent subsets
A and B of Nn , ((Yi )i∈A , (Mi )i∈A ) and ((Yi )i∈B , (Mi )i∈B ) are conditionally independent
given (Mi )i∈νn (A) , where νn (A) is the union of the neighborhoods of i ∈ A with the set A
itself removed.
Our CND property is a generalization of both dependency graphs and Markov random
fields with a global Markov property (Lauritzen, Dawid, Larsen, and Leimer (1990).) Dependency graphs were introduced by Stein (1972) in his study of normal approximation.
(See Chen and Shao (2004) and Rinott and Rotar (1996) for a general local dependence
notion that is different from ours.) A set of random variables have a graph as a dependency graph, if two sets of random variables are allowed to be dependent only when the
two sets are adjacent in the graph. This dependence can be viewed as restrictive in many
applications, as it requires that any random variables be independent even if their indices
are indirectly connected in the graph. In contrast, CND random variables are allowed to
be dependent even if they are not adjacent in a graph. The CND property captures the
notion that “any two random variables are independent once we condition on the source
of their joint dependence”. In this sense, the CND property is closely related to a Markov
property in the literature of random fields. However, in contrast to the Markov property,
the CND property does not require that the σ-fields Mi be generated by Yi itself.
3
This paper provides two main results. The first main result is a Berry-Esseen bound
for a sum of CND random variables. Our bound is comparable to Berry-Esseen bounds
established for a sum of random variables with a dependency graph in some generic situations in the literature. (Baldi and Rinott (1989), Chen and Shao (2004), and Penrose
(2003) to name but a few.) This latter literature typically uses Stein’s method to establish the bound, but to the best of our knowledge, the existing proofs using Stein’s method
for dependency graphs do not seem immediately extendable to a sum of CND random
variables, due to a more flexible form of conditioning σ-fields involved in the CND property. In this paper, we use a traditional characteristic function-based method to derive a
Berry-Esseen bound.
A typical form of a Berry-Esseen bound in these set-ups, including ours, involves the
maximum degree of the neighborhood system, so that when the maximum degree is high,
the bound is of little use. However, in many social networks observed, removing a small
number of high-degree vertices tends to reduce the maximum degree of the neighborhood
system substantially. Exploiting this insight, we provide a general version of a BerryEsseen bound which uses conditioning on random variables associated with high degrees.
The second main result in this paper is a stable limit theorem for an empirical process
indexed by a class of functions, where the empirical process is constituted by CND random variables. Stable convergence is a stronger notion of convergence than weak convergence, and is useful for asymptotic theory of statistics whose normalizing sequence
has a random limit.
To obtain a stable limit theorem, we first extend the exponential inequality of Janson
(2004) for dependency graphs to our set-up of CND random variables, and using this,
we obtain a maximal inequality for an empirical process with a bracketing-entropy type
bound. This maximal inequality is useful for various purposes, especially when one needs
to obtain limit theorems that are uniform over a given class of functions indexing the empirical process. Using this maximal inequality, we establish the asymptotic equicontinuity
of the empirical process which, in combination with the central limit theorem that comes
from our previously established Berry-Esseen bound, gives a stable limit theorem. This
enables stable convergence of the empirical process to a mixture Gaussian process.
As it turns out, our stable limit theorem for an empirical process requires that the maximum degree of the neighborhood system be bounded. However, in many real-life networks, the maximum degree can be substantial, especially when the networks behave like
a preferential attachment network of Barabási-Albert.(Barabási and Albert (1999)) Thus,
following the same spirit of extending the Berry-Esseen bound to the case conditional on
high degree vertices, we extend the stable limit theory to a set-up where it relies only on
those observations with relatively low degrees by conditioning on the random variables
4
associated with high degree vertices. This extension enables us to obtain a stable limit
theorem for empirical processes when the maximum degree of the neighborhood system
increases to infinity as the size of the system increases.
Stable convergence has been extensively studied in the context of martingale central
limit theorems. (See, e.g. Hall and Heyde (1980).) See Häusler and Luschgy (2010)
for stable limit theorems for Markov kernels and related topics. Recent studies by Kuersteiner and Prucha (2013) and Hahn, Kuersteiner, and Mazzocco (2016) established a
stable central limit theorem for a sum of random variables having both cross-sectional
dependence and time series dependence by utilizing a martingale difference array formulation of the random variables. Markov-type cross-sectional dependence on a graph
has received attention in the literature (Lauritzen (1996).) In particular, the pairwise
Markov property of random variables says that two non-adjacent random variables are
conditionally independent given all the other variables, and is captured by a precision
matrix in a high dimensional Gaussian model. (See Meinshausen and Bühlmann (2008)
and Cai, Liu, and Zhou (2016) for references.) This paper’s CND property is stronger
than the pairwise Markov property when the conditioning σ-fields, Mi ’s, are those that
are generated by the random variables. However, the CND property encompasses the
case where the latter condition does not hold, and thus includes dependency graphs as
a special case, unlike Markov-type dependence mentioned before. To the best of our
knowledge, stable convergence of empirical processes indexed by a class of functions has
not been studied under either dependency graphs or Markov random fields on a graph.
The remainder of the paper proceeds as follows. In Section 2, we formally introduce
the notion of conditional neighborhood dependence (CND) and study its basic properties.
In Section 3, we provide stable central limit theorems for a sum of CND random variables.
We also present the stable convergence of an empirical process to a mixture Gaussian
process. The mathematical proofs of the results are found in the appendix.
2. Conditional Neighborhood Dependence
2.1. Definition
Let N be an infinite countable set. For each n = 1, 2, ..., let Nn ⊂ N be a finite set such
that |Nn | = n and let 2Nn be the collection of all the subsets of Nn . We assume that Nn
is a proper subset of Nn+1 for each n ≥ 1. We will call each element of Nn a vertex, and
5
call any map νn : Nn → 2Nn a neighborhood system, if for each i ∈ Nn , i ∈
/ νn (i).1 Let us
define for each A ⊂ Nn ,
!
[
ν̄n (A) ≡
νn (i) ∪ A and νn (A) ≡ ν̄n (A) \ A.
i∈A
For {i1 , ..., im } ⊂ Nn , we simply write νn (i1 , ..., im ) = νn ({i1 , ..., im }) and ν̄n (i1 , ..., im ) =
ν̄n ({i1 , ..., im }), suppressing the curly brackets. Let us call ν̄n (A) the νn -closure of A and
νn (A) the νn -boundary of A. The νn -closure of A includes the vertices in A but the νn boundary around A excludes them.
If for each i, j ∈ Nn , i ∈ νn (j) implies j ∈ νn (i), we say that neighborhood system νn
is undirected. If there exists a pair i, j ∈ Nn such that i ∈ νn (j) but j ∈
/ νn (i), we say that
neighborhood system νn is directed.
It is often useful to compare different dependence structures governed by different
neighborhood systems. When we have two neighborhood systems νn and νn0 such that
νn (i) ⊂ νn0 (i) for each i ∈ Nn and νn (j) 6= νn0 (j) for some j ∈ Nn , we say that νn is strictly
finer than νn0 and νn0 is strictly coarser than νn . When νn (i) ⊂ νn0 (i) for each i ∈ Nn , we say
that νn is weakly finer than νn0 and νn0 is weakly coarser than νn .
Let us introduce the notion of dependence among a triangular array of σ-fields. Let
(Ω, F, P ) be a given probability space, and let {Fi }i∈Nn be a given triangular array of subσ-fields of F, indexed by i ∈ Nn . (Proper notation for the sub-σ-field in the triangular
array should be Fin , but we suppress the n subscript for simplicity.) For any A ⊂ Nn , we
let FA be the smallest σ-field that contains Fi , i ∈ A, and F−A be the smallest σ-field that
contains all the σ-fields Fi such that i ∈ Nn \A. When A = ∅, we simply take FA to be
the trivial σ-field. We apply this notation to other triangular arrays of σ-fields, so that if
{Mi }i∈Nn is a triangular array of sub-σ-fields of F, we similarly define MA and M−A for
any A ⊂ Nn . For given two σ-fields, say, G1 and G2 , we write G1 ∨ G2 to represent the
smallest σ-field that contains both G1 and G2 .
Given a triangular array of σ-fields, {Mi }∞
i=1 , let us introduce a sub σ-field G defined
by
\ \
(2.1)
G≡
Mi .
n≥1 i∈Nn
In many applications, G is used to accommodate random variables with a common shock.
For example, suppose that each Mi is generated by a random vector (εi , U ) from a set of
random variables {εi }i∈Nn that are conditionally independent given a common random
1
Equivalently, one might view the neighborhood system as a graph by identifying each neighborhood of i
as the neighborhood of i in the graph. However, it seems more natural to think of a stochastic dependence
structure among random variables in terms of neighborhoods rather than in terms of edges in a graph.
6
variable U . Then we can take G to be the σ-field generated by U . We will discuss examples
of CND random vectors in a later section, after we study their properties.
Let us introduce the notion of dependence of an array of σ-fields that is of central focus
in this paper.
Definition 2.1. (i) Given neighborhood system νn on Nn and an array of σ-fields, M ≡
{Mi }i∈Nn , we say that σ-fields {Fi }i∈Nn0 for a given subset Nn0 ⊂ Nn are conditionally
neighborhood dependent (CND) with respect to (νn , M), if for any A, B ⊂ Nn0 such that
A ⊂ Nn0 \ ν̄n (B) and B ⊂ Nn0 \ ν̄n (A), FA ∨MA and FB ∨MB are conditionally independent
given Mνn (A) .
(ii) If σ-fields generated by random vectors in {Yi }i∈Nn0 for a subset Nn0 ⊂ Nn are CND
with respect to (νn , M), we simply say that random vectors in {Yi }i∈Nn are CND with
respect to (νn , M).
Conditional neighborhood dependence specifies only how conditional independence
arises, not how conditional dependence arises. Conditional neighborhood dependence
does not specify independence or dependence between Yi and Yj if j ∈ νn (i) or i ∈
νn (j). Furthermore, conditional neighborhood dependence can accommodate the situation where the neighborhoods in the system νn are generated by some random graph on
Nn , as long as the random graph is G-measurable. In such a situation, the results of this
paper continue to hold with only minor modifications that take care of the randomness
of νn .
2.2. Monotonicity and Invariance to Conditioning
In general, a conditional independence property is not monotone in conditioning σfields. In other words, when X and Y are conditionally independent given a σ-field F 0 ,
this does not imply conditional independence between X and Y given any sub-σ-field of
F 0 or given any σ-field that contains F 0 . However, CND partially obeys monotonicity in
neighborhood systems. More specifically, the CND property with a finer neighborhood
system implies more conditional independence restrictions than what the CND property
with a coarser neighborhood system implies. We introduce a lemma which makes precise
this monotonicity property of CND. Suppose that we are given two neighborhood systems
νn and νn0 where νn is weakly finer than νn0 . The following lemma shows that the CND
property of a triangular array with respect to a given neighborhood system νn0 carries over
to that with respect to νn if νn is undirected.
Lemma 2.1. Given two neighborhood systems νn and νn0 , suppose that νn is weakly finer
than νn0 and that a triangular array of σ-fields {Fi }i∈Nn is CND with respect to (νn , M) for
7
F IGURE 1. Conditional Neighborhood Dependence
Notes: Suppose we are given random variables (Y1 , ..., Y13 ) and (M1 , ..., M13 ), where Mi = σ(Mi ). The
figure depicts the neighborhood system on Nn = {1, 2, ..., 13}. If Y1 , ..., Y13 are CND with respect to the
neighborhood system as illustrated above and (M1 , ..., M13 ), it implies that Y1 , ..., Y6 and Y10 , ..., Y13
are conditionally independent given M7 , M8 , M9 .
some σ-fields M ≡ {Mi }i∈Nn . Suppose further that νn is undirected. Then {Fi }i∈Nn is CND
with respect to (νn0 , M).
Proof: Take any A, B ⊂ Nn such that A ⊂ Nn \ ν̄n0 (B) and B ⊂ Nn \ ν̄n0 (A), so that
A ⊂ Nn \ ν̄n (B) and B ⊂ Nn \ ν̄n (A) as well. Let D ≡ νn0 (A) \ ν̄n (A), which, by the
undirectedness of νn , implies that A ⊂ Nn \ ν̄n (D). Therefore, A ⊂ Nn \ ν̄n (B ∪ D) and
B ∪ D ⊂ Nn \ ν̄n (A). By the CND property, FA ∨ MA and FB∪D ∨ MB∪D are conditionally
independent given Mνn (A) . By Lemma 4.2 of Dawid (1979), this implies that FA ∨ MA
and FB ∨ MB are conditionally independent given Mνn (A)∪D . This proves the lemma,
because
νn (A) ∪ D = νn0 (A),
which follows due to νn (A) ⊂ νn0 (A), νn (A) ⊂ ν̄n (A), and ν̄n (A) \ νn (A) = A.
In the above lemma, the requirement that νn be undirected cannot be eliminated. To
see this, consider the following counterexample of the lemma when νn is taken to be a
directed neighborhood system.
8
Example 1: Let us take Nn = {1, 2, 3, 4} with n = 4. Let (εi )i∈Nn be i.i.d. standard normal
random variables. Let us take
Y1 = ε1 , Y2 = ε2 , Y3 = ε1 + ε3 + ε4 , and Y4 = ε4 .
Let a neighborhood system νn be given as νn (i) = ∅ for all i = 1, 2, 4, and νn (3) = {1, 4}.
Hence νn is directed. Then we take M = F = (Fi )i∈Nn , where each Fi denotes the σ-field
generated by Yi . Then it is not hard to see that (Yi )i∈Nn is CND with respect to (νn , M).
Now let us introduce another neighborhood system νn0 that is weakly coarser than νn .
Let νn0 (1) = {3}, νn0 (2) = {1}, νn0 (3) = {1, 4}, and νn0 (4) = {3}. Then we take A = {1} and
B = {4} so that A ⊂ Nn \ ν̄n0 (B) and B ⊂ Nn \ ν̄n0 (A). Note that νn0 (A) = νn0 (1) = {3}.
Certainly Y1 and Y4 are not conditionally independent given Y3 , because Y3 involves both
ε1 and ε4 .
As we shall see later, using further conditioning in the CND property, one may obtain a
better normal approximation for a sum of CND random variables in some situations. Here
we give a preliminary result which addresses the question of whether the CND property
still holds after we increase the σ-fields in M in a certain way. More precisely, suppose
that Y ≡ {Yi }i∈Nn is CND with respect to (νn , M). Then let us choose Nn∗ ⊂ Nn , and
define a neighborhood system νn∗ on Nn such that for all A ⊂ Nn ,
νn∗ (A) ≡ νn (A) ∩ Nn∗ .
(2.2)
Let M∗ ≡ {M∗i }i∈Nn , where
(2.3)
M∗i ≡ Mi ∨ MNn \Nn∗ .
Then we obtain the following result.
Lemma 2.2. Suppose that Y ≡ {Yi }i∈Nn is CND with respect to (νn , M). Fix Nn∗ ⊂ Nn and
define νn∗ as in (2.2) and M∗ ≡ {M∗i }i∈Nn as in (2.3). Then Y ∗ ≡ {Yi }i∈Nn∗ is CND with
respect to (νn∗ , M∗ ).
Proof: Take A, B ⊂ Nn∗ such that A ⊂ Nn∗ \ν̄n (B) and B ⊂ Nn∗ \ν̄n (A). Note that for any
A0 ⊂ Nn∗ , Nn∗ \ν̄n (A0 ) = Nn∗ \ν̄n∗ (A0 ). Let
B ∗ ≡ B ∪ ((Nn \Nn∗ )\ν̄n (A)).
By the CND of Y , we have that σ(YA ) ∨ MA and σ(YB ∗ ) ∨ MB ∗ are conditionally independent given Mνn (A) . Hence σ(YA ) ∨ MA and σ(YB ) ∨ MB ∗ are conditionally independent
given Mνn (A) . From (2.2), we have
νn∗ (A)c ∩ νn (A) = νn (A) ∩ (Nn \Nn∗ ).
9
Hence we can write
νn (A) = νn∗ (A) ∪ (νn∗ (A)c ∩ νn (A)) = νn∗ (A) ∪ (νn (A) ∩ (Nn \Nn∗ )),
so that σ(YA ) ∨ MA and σ(YB ) ∨ MB ∗ are conditionally independent when we condition
on Mνn∗ (A)∪(νn (A)∩(Nn \Nn∗ )) . This implies that σ(YA )∨MA and σ(YB )∨MB are conditionally
independent given Mνn∗ (A)∪(Nn \Nn∗ ) by the definition of B ∗ , and by the choice of A ⊂ Nn∗ .
2.3. Examples
2.3.1. Conditional Dependency Graphs. Let Gn = (Nn , En ) be an undirected graph on
Nn , where En denotes the set of edges. Define νn (i) = {j ∈ Nn : ij ∈ En }. Suppose that
{Yi } has Gn as a conditional dependency graph, i.e., for each A ⊂ Nn , YA and YNn \ν̄n (A)
are conditionally independent given a σ-field C. Dependency graphs were introduced by
Stein (1972) and have received attention in the literature. (See, for example, Janson
(1988), Baldi and Rinott (1989) and Rinott and Rotar (1996) for Berry-Esseen bounds
for the sum of random variables with a dependency graph, and Janson (2004) for an
exponential inequality. See Song (2015) for an application to permutation inference.)
Then {Yi } is CND with respect to (νn , M), where Mi is taken to be C for each i ∈ Nn .
2.3.2. Functional Local Dependence. Let Gn = (Nn , En ) be a given directed graph, so
that ij ∈ En represents an edge from i to j. The neighborhood NnO (i) = {j ∈ Nn : ij ∈
En } is the out-neighborhood of i, the set of vertices j that there is an edge from vertex
i to vertex j. Similarly, NnI (i) = {j ∈ Nn : ji ∈ En } is the in-neighborhood of i. We
I
O
define N¯n (i) = NnO (i) ∪ {i} and N¯n (i) = NnI (i) ∪ {i}. Suppose that Yi is generated in
the following way:
(2.4)
Yi = γi (εN̄nI (i) , ηi ),
where εN̄nI (i) = (εj )j∈N̄nI (i) and εi ’s and ηi ’s are independent across i’s and (εi )i∈Nn and
(ηi )i∈Nn are independent from each other. Here the functions, γi ’s, are nonstochastic.
This way of modeling local dependence among Yi ’s through base random variables,
εi ’s, is useful in many contexts of applications. In particular, each outcome Yi may arise
as a consequence of local interactions among individual variables, where locality is determined by a given graph Gn . (See Leung (2016) and Canen, Schwartz, and Song (2017)
for applications in economics.)
Let us define a new (undirected) graph G0n = (Nn , En0 ) such that ij ∈ En0 if and only if
N̄nI (i) ∩ N̄nI (j) 6= ∅. In other words, i and j are adjacent in G0n , if their in-neighborhoods
overlap. Define Nn0 (i) = {j ∈ Nn : ij ∈ En0 } for each i ∈ Nn . Then it is not hard to see that
10
{Yi }i∈Nn is CND with respect to (Nn0 , M0 ), for any triangular arrays of σ-fields M0i , as long
as each M0i is the σ-fields generated by fi ((εj , ηj )j∈Nn ) for any nonstochastic measurable
map fi . Note that we can take M0i to be trivial σ-fields in which case the functional local
dependence is essentially equivalent to the dependency graph assumption.
However, functional local dependence has much richer implications than the dependency graph assumption alone, because it generates lots of conditional independence
restrictions that are not implied by the dependency graph assumption alone. Using such
restrictions, we can obtain conditional neighborhood dependence as follows: for each
i ∈ Nn , let Mi be the σ-field generated by (εi , ηi ). It is not hard to see that the triangular
array {Yi }i∈Nn is CND with respect to (NnI , M). Note that the neighborhood system NnI
is weakly finer than Nn0 , and hence when the graph Gn is undirected, by Lemma 2.1, the
CND with respect to (NnI , M) expresses richer conditional independence restrictions than
the CND with respect to (Nn0 , M).
The notion of functional local dependence is related to physical dependence of Wu
(2005). The main difference is that physical dependence in Wu (2005) is mainly intended
for time series dependence, where Yi involves only the past and present values of εi ’s
through a common function, whereas the functional local dependence captures local
dependence through a system of neighborhoods of random variables.
2.3.3. Markov Random Fields on a Neighborhood System. Suppose that we have a
triangular array of random vectors {Yi }i∈Nn , Yi ∈ Rd , where there is an undirected neighborhood system νn on Nn . A path in νn between two vertices i and j is defined to be a
sequence of distinct vertices i1 , i2 , ..., im such that i1 = i and im = j and it ∈ νn (it+1 ) for
each t = 1, ..., m − 1. Let us say that a set S ⊂ Nn separates sets A, B ⊂ Nn , if for any
i ∈ A and any j ∈ B, every path between a vertex in A and a vertex in B intersects set
S. Let us consider the following two notions of Markov properties (see Lauritzen (1996),
p.32).
Definition 2.2. (i) We say that {Yi }i∈Nn satisfies the the local Markov property if for any
vertex i ∈ Nn and any B ⊂ Nn \ν̄n (i), Yi and YB are conditionally independent given
Yνn (i) .
(ii) We say that {Yi }i∈Nn satisfies the the global Markov property if for any two subsets A
and B of Nn which are separated by set S ⊂ Nn , YA and YB are conditionally independent
given YS .
Suppose that Mi is the σ-field generated by Yi for each i ∈ Nn . It is not hard to see that
if {Yi }i∈Nn satisfies the global Markov property, it is CND with respect to (νn , M). And
if {Yi }i∈Nn is CND with respect to (νn , M), it satisfies the local Markov property. Hence
11
our notion of CND is an intermediate concept between the local and the global Markov
properties.2
3. Stable Limit Theorems
3.1. Stable Central Limit Theorems
3.1.1. The Basic Result. In this section, we give a Berry-Esseen bound conditional on G
for a triangular array of random variables that are CND with respect to (νn , M). Given a
neighborhood system νn on Nn , we define
1 X
dmx ≡ max |νn (i)| and dav ≡
|νn (i)|,
i∈Nn
n i∈N
n
where |νn (i)| denotes the cardinality of set νn (i). We call dmx the maximum degree and
dav the average degree of neighborhood system νn . We use dmx and dav to express the
conditions for the neighborhood system νn . For p = 1, ..., 4, define
1/p
p
Xi
,
µp ≡ max E
|G
i∈Nn
σn
P
where σn2 ≡ V ar( i∈Nn Xi |G). Let
2
X
X
1
rn2 ≡ 4 E
ξij |G ,
σn
i∈N
n
j∈ν̄n (i)
where ξij ≡ E[Xi Xj |Mνn (i,j) ] − E[Xi Xj |G].
Theorem 3.1. Suppose that a triangular array {Xi }i∈Nn is CND with respect to (νn , M).
Furthermore assume that E[Xi |Mνn (i) ] = 0, a.e. for each i ∈ Nn , and that ndmx dav µ33 ≤ 1,
a.e., for all n ≥ 1.
Then there exists an absolute constant C > 0 such that for each t ∈ R and each n ≥ 1,
q
q
3
4
3
2
2
ndmx dav µ3 − log(ndmx dav µ3 ) ndmx dav µ4 + rn , a.e.,
∆n (t; G) ≤ C
where
(
∆n (t; G) ≡
P
1 X
Xi ≤ t|G
σn i∈N
)
− Φ(t)
n
and Φ is the distribution function of the standard normal distribution.
2
Lauritzen, Dawid, Larsen, and Leimer (1990) proposed Markov fields over directed acyclic graphs. They
defined local Markov property and global Markov property and provided a sufficient condition under which
both are equivalent.
12
The use of the bound requires a good bound for rn . Observe that E[ξij |G] = 0. Therefore, when ξij ’s are locally dependent in a proper sense, we can expect that rn2 is at most
of the same order as the term nd2mx dav µ44 . The following corollary gives a set of conditions
under which this is true.
Lemma 3.1. Suppose that the conditions of Theorem 3.1 hold. Furthermore, Mi ’s are
conditionally independent given G. Then
rn2 ≤ 8nd2mx dav µ44 .
Focusing on a special case satisfying an additional condition below, we can obtain an
improved version of Theorem 3.1.
Condition A: For any A, B ⊂ Nn such that A ⊂ Nn \ ν̄n (B), Mνn (A) and FB are conditionally independent given G.
Condition A accommodates conditional dependency graphs but excludes Markov random fields.
Corollary 3.1. Suppose that a triangular array {Xi }i∈Nn is CND with respect to (νn , M),
and that Condition A holds. Suppose further that E[Xi |G] = 0, a.e. for each i ∈ Nn , and
that ndmx dav µ33 ≤ 1, a.e., for all n ≥ 1.
Then there exists an absolute constant C > 0 such that for each t ∈ R and each n ≥ 1,
q
q
3
3
4
2
ndmx dav µ3 − log(ndmx dav µ3 ) ndmx dav µ4 , a.e.,
∆n (t; G) ≤ C
where we define µp and ∆n (t; G) as in Theorem 3.1.
The improvement of the result due to Condition A is two fold. First, the condition
E[Xi |Mνn (i,j) ] = 0 is weakened to E[Xi |G] = 0. Second, the bound does not involve rn2 .
When it is cumbersome to compute a reliable bound for rn2 , the above corollary can be
useful.3
In the case of dependency graphs, there has been much research establishing a Berry√
Esseen bound. When we confine our attention to the special case of |Xi | ≤ 1/ n, i ∈ Nn ,
and dmx < D for all n ≥ 1 for some constant D > 0, our bound in Corollary 3.1 has
the same n−1/2 rate as in Baldi and Rinott (1989) (Corollary 2), Chen and Shao (2004)
(Theorem 2.7) and Penrose (2003) (Theorem 2.3), among others. These papers adopted
Stein’s method to obtain the bound. However, to the best of our knowledge, it is not
3
In the special case of conditional dependency graphs, one can follow the proof of Theorem 2.4 of Penrose
(2003) to obtain a slightly improved bound that does not have the logarithmic
p factor. It appears that this
improvement is
marginal
in
many
applications.
For
example,
the
quantity
nd2mx dav µ44 is asymptotically
p
dominated by ndmx dav µ33 , when µpp = OP (n−p/2 ) and dmx increases with n slower than the rate n1/2 .
13
straightforward to extend their results to our set-up of CND variables. The main reason is
that the conditioning σ-field conditional on which two sets of random variables XA and
XB are independent varies depending on the set A. Thus, for example, we cannot apply
Equation (2.4) in Penrose (2003), p.31, in our context. In this paper, we resort to a more
traditional Fourier analytic method in combination with Esseen’s inequality.
The Berry-Esseen bound gives stable convergence of a sum of random vectors to a
mixture normal distribution. More specifically, suppose that {Xi }i∈Nn is a triangular array
of random variables such that E[Xi |Mνn (i) ] = 0, a.e., for each i ∈ Nn , and for each t ∈ R,
∆n (t; G) →P 0, as n → ∞.
Then for each U ∈ G, and for each uniformly continuous and bounded map f on R, we
have
"
! #
1 X
E f
Xi 1U → E[f (Z)1U ], as n → ∞,
σn i∈N
n
where Z is a standard normal random variable that is independent of U and 1U denotes
the indicator of event U .
3.1.2. Conditional Neighborhood Dependence Conditional on High-Degree Vertices.
Let us extend Theorem 3.1 by considering normal approximation conditioning on random
variables associated with high degree vertices.
Let Nn∗ ⊂ Nn be a given subset, and νn∗ the neighborhood system on Nn given as in
(2.2). Let
1 X ∗
(3.1)
d∗mx ≡ max∗ |νn∗ (i)|, and d∗av ≡ ∗
|ν (i)|,
i∈Nn
n j∈N ∗ n
n
where n∗ ≡ |Nn∗ |. Hence, d∗mx and d∗av are the maximum and average degrees of the
restriction of νn to Nn∗ . Moreover, define M∗ = (M∗i )i∈Nn , where M∗i = Mi ∨ MNn \Nn∗ ,
and write M∗νn∗ (i) = Mνn∗ (i) ∨MNn \Nn∗ . We have in mind choosing Nn∗ so that the set Nn \Nn∗
consists only of high-degree vertices in the neighborhood system νn . Then by Lemma 2.2,
if (Xi )i∈Nn is CND with respect to (νn , M), then (Xi )i∈Nn∗ is CND with respect to (νn∗ , M∗ ).
The main idea is that if the difference between the two sums
X
X
(3.2)
(Xi − E[Xi |Mνn (i) ])
(Xi − E[Xi |M∗νn∗ (i) ]) and
i∈Nn∗
i∈Nn
is asymptotically negligible, we can use the Berry-Esseen bound for the first sum using
Theorem 3.1 and deal with the remainder term that comes from the difference. Now let
14
us present an extended version of Theorem 3.1. Let
1/p
p
Xi
Xi
∗
∗
∗
µp ≡ max∗ E
|Gn
, and µ̃p ≡ max ∗ E
∗
i∈Nn
i∈Nn \Nn
σn
σn∗
P
where σn∗2 ≡ V ar( i∈Nn∗ Xi |Gn∗ ), and
p
|Gn∗
1/p
,
Gn∗ ≡ G ∨ MNn \Nn∗ .
(3.3)
Note that the supremum in the definition of µ∗p is Nn∗ whereas that of µ̃∗p is Nn \Nn∗ . We
also define for i, j ∈ Nn∗ ,
2
X
X
1
rn∗2 ≡ ∗4 E
ξij |Gn∗ .
σn
∗
i∈N ∗
n
j∈ν̄n (i)∩Nn
Theorem 3.2. Suppose that a triangular array {Xi }i∈Nn is CND with respect to (νn , M)
and E[Xi |Mνn (i) ] = 0, a.e. for each i ∈ Nn as in Theorem 3.1, and n∗ d∗mx d∗av µ∗3
3 ≤ 1, a.e.,
for all n∗ ≥ 1.
Then there exists an absolute constant C > 0 such that for each t ∈ R, n ≥ 1, and for
each 1 ≤ r ≤ 4,
q
q
∗
∗
∗ ∗
∗
∗3
∗3
∗4
∗
∗
∗
∗
∗2
∆ (t; Gn ) ≤ C n dmx dav µ3 − C log(n dmx dav µ3 ) n∗ d∗2
mx dav µ4 + rn
+ C((n − n∗ )µ̃∗r + ρ∗r )r/(r+1) , a.e.,
where
(
∆∗ (t; Gn∗ ) ≡ P
1 X
Xi ≤ t|Gn∗
σn∗ i∈N
)
− Φ(t) ,
n
and
r 1/r
X
1
ρ∗r ≡ E ∗
E[Xi |M∗νn∗ (i) ] .
σn i∈N ∗
n
Compared to Theorem 3.1, the bound involves an additional term. This additional
term arises because the sum may not be centered around zero when we condition on
√
M∗νn∗ (i) . If |Xi | ≤ 1/ n and (σn∗ )2 ≥ c for some c > 0, the third term in the bound is of
√
order O(((n − n∗ )/ n + ρ∗r )r/(r+1) ).
If X has νn as a conditional dependency graph, we have ρ∗r = 0 (because Mi = G for
all i ∈ Nn in this case) and hence as long as
√
(n − n∗ )/ n → 0,
15
as n → ∞, the third term in the bound vanishes, and the same CLT as in Theorem 3.1
√
√
is restored. In this case, if (n∗ /n)1/2 (((n − n∗ )/ n)4/5 + (d∗mx / n∗ )1/2 ) converges to zero
√
faster than (dmx / n)1/2 , Theorem 3.2 has an improved rate over Theorem 3.1. Such an
approximation captures the situation where the neighborhood system νn has a very small
fraction of very high degree vertices.
Such an improvement can still arise generally, even if X does not have νn as a conditional dependency graph. To see this, note that for each i ∈ Nn \ν̄n (Nn \Nn∗ ), Xi is
conditionally independent of MNn \Nn∗ given Mνn (i) , we have
E[Xi |M∗νn∗ (i) ] = E[Xi |M(νn (i)∩Nn∗ )∪(Nn \Nn∗ ) ]
= E[Xi |Mνn (i)∪(Nn \Nn∗ ) ] = E[Xi |Mνn (i) ] = 0.
Hence
r 1/r
ρ∗r = E
1
σn∗
X
E[Xi |M∗νn∗ (i) ]
.
i∈ν̄n (Nn \Nn∗ )∩Nn∗
√
Suppose that |Xi | ≤ 1/ n, (σn∗ )2 ≥ c for some c > 0, n − n∗ = O(1), and dav = O(1) and
d∗av = O(1). Then we have
√
√
ρ∗r = O(|ν̄n (Nn \Nn∗ )|/ n) = O(dmx / n).
(3.4)
The rate in Theorem 3.2 improves on that in Theorem 3.1 because with p > 1, (ρ∗p )4/5 =
√
√
O((dmx / n)4/5 ) = o((dmx / n)1/2 ).
As we shall see later in Section 3.2.5, the approach of CND conditional on high-degree
vertices is useful for obtaining stable central limit theorem for empirical processes when
the random variables are CND with respect to a neighborhood system having a maximum
degree increasing with to infinity as n → ∞.
3.2. Empirical Processes and Stable Convergence
3.2.1. Stable Convergence in Metric Spaces. Let us first introduce some preliminary
results about stable convergence in metric spaces. Let (D, d) be a given metric space and
define B(D) to be the Borel σ-field of D. Let (Ω, F, P ) be the given probability space,
where G is a sub σ-field of F. Recall that a map L : B(D) × Ω → [0, 1] called a Markov
kernel if for each ω ∈ Ω, L(·, ω) is a Borel probability measure on B(D) and for each
B ∈ B(D), L(B, ·) is F-measurable. Following the notation in Häusler and Luschgy
(2010), let us define the marginal of L on B(D) as
Z
P L(B) ≡ L(B, ω)dP (ω).
16
For a given finite collection {h1 , ..., hk } ⊂ H, the finite dimensional projection of a
Markov kernel L is defined to be a Markov kernel Lh1 ,...,hk : B(Rk ) × Ω → [0, 1] such that
for any B ∈ B(Rk ),
Lh1 ,...,hk (B, ω) ≡ L({y ∈ D : (y(h1 ), ..., y(hk )) ∈ B}, ω).
In the spirit of the Hoffman-Jorgensen approach, we consider the following definition of
stable convergence for empirical processes. (See Berti, Pratelli, and Rigo (2012), p.2.,
for a similar definition.)
Definition 3.1. Suppose that we are given a sub σ-field G, a sequence of D-valued stochastic processes ζn , a Markov kernel L on B(D) × Ω, and a D-valued Borel measurable
random element ζ that has a Markov kernel L.
Suppose that for each U ∈ G, and each bounded Lipschitz functional f on D,
Z Z
∗
f (y)L(dy, ω)dP (ω),
E [f (ζn )1U ] →
U
D
∗
as n → ∞, where E denotes the outer-expectation and 1U is the indicator function of
the event U . Then we say that ζn converges to L, G-stably, (or equivalently, ζn converges
to ζ, G-stably), and write
ζn → L, G-stably (or equivalently, ζn → ζ, G-stably.)
Stable convergence according to Definition 3.1 implies weak convergence (in the sense
of Hoffman-Jorgensen). When ζn is Borel measurable, the above definition is equivalent
to the weak convergence of Markov kernels and many of the existing results on stable
convergence carry over. However, this equivalence does not extend to the case of ζn being
non-measurable, because there is no proper definition of Markov kernels for nonmeasurable stochastic processes. Nevertheless the above definition can still be useful when one
needs to deal with random norming, as shown in the following lemma which generalizes
Theorem 1’ in Aldous and Eagleson (1978) and part of Theorem 3.18 of Häusler and
Luschgy (2010).
Lemma 3.2. Suppose that ζn , ζ, ξn , ξ are D-valued random variables, and that ζn → ζ, Gstably, where ζ and ξ are Borel measurable, and let Pζ denote the distribution of ζ. Then the
following holds.
(i) If P ∗ {d(ξn , ξ) > ε} → 0 as n → ∞ for each ε > 0, and ξ is G-measurable, then
(ζn , ξn ) → (ζ, ξ), G-stably,
where P ∗ denotes the outer probability.
(ii) If f : D → D is Pζ -a.e. continuous, then f (ζn ) → f (ζ), G-stably.
17
The first result is a stable-convergence analogue of Cramér-Slutsky lemma. The second
result is a continuous mapping theorem.
3.2.2. Stable Convergence of an Empirical Process. Suppose that {Yi }i∈Nn is a given
triangular array of R-valued random variables which is CND with respect to (νn , M). Let
H be a given class of real measurable functions on R. Then, we consider the following
empirical process:
{Gn (h) : h ∈ H},
where, for each h ∈ H,
1 X
Gn (h) ≡ √
(h(Yi ) − E[h(Yi )|Mνn (i) ]).
n i∈N
n
The empirical process νn takes a value in l∞ (H), the collection of bounded functions on H
which is endowed with the sup norm so that (l∞ (H), k · k∞ ) forms the metric space (D, d)
with khk∞ ≡ supy∈R |h(y)|. In this section, we explore conditions for the class H and the
joint distribution of the triangular array {Yi }i∈Nn which delivers the stable convergence
of the empirical process. Stable convergence in complete separable metric spaces can be
defined as a weak convergence of Markov kernels. (See Häusler and Luschgy (2010).)
However, this definition does not extend to the case of empirical processes taking values
in D that is endowed with the sup norm, due to non-measurability.
Weak convergence of an empirical process to a Gaussian process is often established
in three steps. First, we show that the class of functions is totally bounded with respect
to a certain pseudo-metric ρ. Second, we show that each finite dimensional projection of
the empirical process converges in distribution to a multivariate normal random vector.
Third, we establish the asymptotic ρ-equicontinuity of the empirical process.
Let ρ be a given pseudo-metric on H such that (H, ρ) is a totally bounded metric space.
Then we define
Uρ (H) ≡ {y ∈ D : y is uniformly ρ-continuous on H}.
The following theorem shows that we can take a similar strategy in proving the stable
convergence of an empirical process to a Markov kernel. The structure and the proof of
the theorem is adapted from Theorem 10.2 of Pollard (1990).
Theorem 3.3. Suppose that the stochastic process ζn ∈ D is given, where (H, ρ) is a totally
bounded metric space. Suppose that the following conditions hold.
(i) For each finite set {h1 , ..., hk } ⊂ H, (ζn (h1 ), ..., ζn (hk )) → Lh1 ,...,hk , G-stably, where
Lh1 ,...,hk is a Markov kernel on B(Rk ) × Ω.
18
(ii) ζn is asymptotically ρ-equicontinuous on H, i.e., for each ε > 0, η > 0, there exists δ > 0
such that
(
)
limsupn→∞ P ∗
sup
|ζn (h) − ζn (h0 )| > η
< ε,
h,h0 ∈H:ρ(h,h0 )<δ
where P ∗ denotes the outer probability.
Then there exists a Markov kernel L on B(D) × Ω such that the following properties are
satisfied.
(a) The finite dimensional projections of L are given by Markov kernels Lh1 ,...,hk .
(b) P L(Uρ (H)) = 1.
(c) ζn → L, G-stably.
Conversely, if ζn converges to Markov kernel L on B(D)×Ω, G-stably, where P L(Uρ (H)) =
1, then (i) and (ii) are satisfied.
It is worth noting that for stable convergence of empirical processes, the conditions
for the asymptotic equicontinuity and the totally boundedness of the function class with
respect to a pseudo-norm are as in the standard literature on weak convergence of empirical processes. The only difference here is that the convergence of finite dimensional
distributions is now replaced by the stable convergence of finite dimensional projections.
3.2.3. Maximal Inequality. This subsection presents a maximal inequality in terms of
bracketing entropy bounds. The maximal inequality is useful primarily for establishing
asymptotic ρ-equicontinuity of the empirical process but also for many other purposes.
First, we begin with a tail bound for a sum of CND random variables. Janson (2004)
established an exponential bound for a sum of random variables that have a dependency
graph. The following exponential tail bound is crucial for our maximal inequality. The
result below is obtained by slightly modifying the proof of Theorem 2.3 of Janson (2004).
Lemma 3.3. Suppose that {Xi }i∈Nn is a triangular array of random variables that take
values in [−M, M ] and are CND with respect to (νn , M), and let σi2 ≡ V ar Xi |Mνn (i) and
P
Vn ≡ i∈Nn E[σi2 |G] with G as defined in (2.1).
Then, for any η > 0,
(
)
X
η2
P
Xi ≥ η|G
(3.5)
≤ 2 exp −
, a.e.
2(dmx + 1)(2(dmx + 1)Vn + M η/3)
i∈N
n
for all n ≥ 1.
Furthermore, if Condition A holds and the condition E[Xi |Mνn (i) ] = 0 is replaced by
E[Xi |G] = 0 and the σ-fields Mνn (i) in σi ’s are replaced by G, then the following holds: for
19
any η > 0,
(
(3.6)
P
)
X
i∈Nn
Xi ≥ η|G
≤ 2 exp −
8η 2
25(dmx + 1)(Vn + M η/3)
, a.e.
for all n ≥ 1.
The bound in (3.6) is the one obtained by Janson (2004) for the case of dependency
graphs. From this, the following form of maximal inequality for a finite set immediately
follows from Lemma A.1 of van der Vaart (1996).
Corollary 3.2. Suppose that {Yi }i∈Nn is a triangular array of random variables that are
P
CND with respect to (νn , M). Let for each h ∈ H, Vn (h) ≡ n−1 i∈Nn E[σi2 (h)|G] and
σi2 (h) ≡ V ar(h(Yi )|Mνn (i) ) with G as defined in (2.1).
Then there exists an absolute constant C > 0 such that
E max |Gn (hs )||G
1≤s≤m
q
J
≤ C(dmx + 1) √ log(1 + m) + log(1 + m) max Vn (hs ) , a.e.,
1≤s≤m
n
for any n ≥ 1 and any m ≥ 1 with a finite subset {h1 , ..., hm } of H such that for some
constant J > 0, max1≤s≤m supx∈R |hs (x)| ≤ J.
Let us now elevate the above inequality to a maximal inequality over function class H.
Recall that we allow the random variables Yi ’s to be idiosyncratically distributed across
i’s.
We define the following semi-norm on H:
s
1 X
E[h2 (Yi )].
ρ̄n (h) ≡
n i∈N
n
We denote N[] (ε, H, ρ̄n ) to be the ε-bracketing number of H with respect to ρ̄n , i.e., the
smallest number J of the brackets [hL,j , hU,j ], j = 1, ..., J, such that ρ̄n (hU,j − hL,j ) ≤ ε.
The following lemma establishes the maximal inequality in terms of a bracketing entropy
bound.
Lemma 3.4 (Maximal Inequality). Suppose that {Yi }i∈Nn is a triangular array of random
variables that are CND with respect to (νn , M). Suppose further that the class H of functions
have an envelope H such that ρ̄n (H) < ∞. Then, there exists an absolute constant C > 0
such that for each n ≥ 1,
Z ρ̄n (H) q
∗
1 + log N[] (ε, H, ρ̄n )dε.
E sup |Gn (h)| ≤ C(1 + dmx )
h∈H
0
20
The bracketing entropy bound in Lemma 3.4 involves the maximum degree dmx . Hence
the bound is useful only when the neighborhood system νn does not have a maximum
degree increasing with n.
3.2.4. Stable Central Limit Theorem. First, let us say that a stochastic process {G(h) :
h ∈ H} is a G-mixture Gaussian process if for any finite collection {h1 , ..., hm } ⊂ H, the
distribution of random vector [G(h1 ), G(h2 ), ..., G(hm )] conditional on G is a multivariate
normal distribution. Also, we call a Markov kernel K a G-mixture Gaussian Markov kernel
associated with a given G-mixture Gaussian process G if for any h1 , ..., hm ∈ H, the conditional distribution of [G(h1 ), G(h2 ), ..., G(hm )] given G is given by the finite dimensional
projection Kh1 ,...,hm of K. Let us summarize the conditions for the neighborhood system
as follows.
Assumption 3.1. (a) There exists C > 0 such that for all n ≥ 1,
Z ρ̄n (H) q
(1 + dmx )
1 + log N[] (ε, H, ρ̄n )dε < C.
0
(b) For any h1 , h2 ∈ H,
E[Gn (h1 )Gn (h2 )|G] →P K(h1 , h2 |G),
for some K(·, ·|G)(ω) : H × H → R, ω ∈ Ω, which is positive semidefinite a.e.
(c) dmx = O(1), and for each h ∈ H, −n−1/4 rn (h) log(n−1/2 dmx dav ) →P 0, as n → ∞,
where
2
X
X
1
rn2 (h) ≡ E √
ξij (h) |G ,
n i∈N
n
j∈ν̄n (i)
with
ξij (h) ≡ E[(h(Xi ) − E[h(Xi )|Mνn (i) ])(h(Xj ) − E[h(Xj )|Mνn (j) ])|Mνn (i,j) ]
−E[(h(Xi ) − E[h(Xi )|Mνn (i) ])(h(Xj ) − E[h(Xj )|Mνn (j) ])|G].
(d) For each h ∈ H, ρ(h) ≡ limn→∞ ρ̄n (h) exists and satisfies that whenever ρ(hn ) → 0 as
n → ∞, ρ̄n (hn ) → 0 as n → ∞ as well.
The following result gives a Donsker-type stable convergence of empirical processes.
21
Theorem 3.4. Suppose that {Yi }i∈Nn is a triangular array of random variables that are
CND with respect to (νn , M), satisfying Assumption 3.1. Suppose further that there exists
C > 0 such that for each n ≥ 1, maxi∈Nn E[H(Yi )4 ] < C, where H is an envelope of H.
Then νn converges to a G-mixture Gaussian process G in l∞ (H), G-stably, such that for
any h1 , h2 ∈ H, E[G(h1 )G(h2 )|G] = K(h1 , h2 |G), a.e..
Furthermore, we have P K(Uρ (H)) = 1, where K is the G-mixture Gaussian Markov kernel
associated with G.
The fourth moment condition maxi∈Nn E[H(Yi )4 ] < ∞ is used to ensure the convergence of finite dimensional distributions using Theorem 3.1. It is worth noting that Assumption 3.1(a) essentially requires that the maximum degree dmx to be bounded. It is
interesting that this condition was not required for the CLT in Theorem 3.1. This stronger
condition for the maximum degree is used to establish the asymptotic equicontinuity of
the process {Gn (h) : h ∈ H}. When the neighborhood system νn is generated according to a model of stochastic graph formation, this condition is violated for many existing
models of graph formation used, for example, for social network modeling. In the next
section, we utilize the approach of conditioning on high-degree vertices to weaken this
condition.
3.2.5. Conditional Neighborhood Dependence Conditional on High-Degree Vertices.
As mentioned before, Assumption 3.1(a) requires that dmx be bounded. Following the
idea of conditioning on high degree vertices as in Theorem 3.2, let us explore a stable
convergence theorem that relaxes this requirement. As we did prior to Theorem 3.2, we
choose Nn∗ ⊂ Nn to be a given subset and let d∗mx and d∗av be as defined in (3.1).
First, write
Gn (h) = G∗n (h) + Rn∗ (h) + ρ∗n (h),
(3.7)
where
1 X
G∗n (h) ≡ √
h(Yi ) − E[h(Yi )|M∗νn∗ (i) ]
n i∈N ∗
n
1
Rn∗ (h) ≡ √
n
X
h(Yi ) − E[h(Yi )|Mνn (i) ] , and
i∈Nn \Nn∗
1 X
ρ∗n (h) ≡ √
E[h(Yi )|Mνn (i) ] − E[h(Yi )|M∗νn∗ (i) ] .
n i∈N ∗
n
Note that
E
∗
sup |Rn∗ (h)|
h∈H
≤
p
2(n − n∗ )
√
max E[H 2 (Yi )].
i∈Nn
n
22
Since {Yi }i∈Nn∗ is CND with respect to (νn∗ , M∗ ) as defined in (2.2) and (2.3), we can
apply the previous results to G∗n (h). This gives the following extension of the maximal
inequality in Lemma 3.4. Since the maximal inequality is often of independent interest,
let us state it formally.
Lemma 3.5 (Maximal Inequality). Suppose that {Yi }i∈Nn is a triangular array of random
variables that are CND with respect to (νn , M). Suppose further that the class H of functions
have an envelope H such that ρ̄n (H) < ∞.
Then, there exists an absolute constant C > 0 such that for all n ≥ 1,
√
Z
C
n∗ (1 + d∗mx ) ρ̄n (H) q
∗
√
E sup |Gn (h)| ≤
1 + log N[] (ε, H, ρ̄n )dε
n
h∈H
0
p
n − n∗
∗
∗
√
+ C
max E[H 2 (Yi )] + E sup ρn (h) .
n i∈Nn
h∈H
√
If kHk∞ < C and n − n∗ = O(1), the second term in the bound is O(dmx / n) similarly
as we derived in (3.4). Thus this bound is an improvement over Lemma 3.4, whenever
O(d∗mx ) = o(dmx ) as n → ∞. Let us turn to the Donsker-type stable convergence of an
empirical process. We modify Assumption 3.1 as follows.
Assumption 3.2. (a) There exists C > 0 such that for all n ≥ 1,
Z ρ̄n (H) q
∗
1 + log N[] (ε, H, ρ̄n )dε < C.
(1 + dmx )
0
(b) For any h1 , h2 ∈ H,
E[Gn (h1 )Gn (h2 )|Gn∗ ] →P K(h1 , h2 |G ∗ ),
for some K(·, ·|G ∗ )(ω) : H × H → R, ω ∈ Ω, which is positive semidefinite a.e, and
non-constant at zero, and for some sub σ-field G ∗ of F, where Gn∗ is as defined in (3.3).
(c) d∗mx = O(1), and for each h ∈ H, −n∗−1/4 rn∗ (h) log(n∗−1/2 d∗mx d∗av ) →P 0, as n → ∞,
where
2
X X
1
ξij (h) |Gn∗ .
rn∗2 (h) ≡ E √
n i∈N ∗
∗
n
j∈ν̄n (i)∩Nn
(d) For each h ∈ H, ρ(h) ≡ limn→∞ ρ̄n (h) exists and satisfies that whenever ρ(hn ) → 0 as
n → ∞, ρ̄n (hn ) → 0 as n → ∞ as well.
(e) E∗ [suph∈H ρ∗n (h)] → 0, as n → ∞.
23
While Condition (a) essentially requires that d∗mx be bounded, Condition (c) allows dmx
√
to increase to infinity as n. The condition in (b) that K(·, ·|G ∗ ) be non-constant at zero
requires that
n − n∗ → λ, as n → ∞,
for some λ ∈ [0, ∞). Thus the number of the high degree vertices (n − n∗ ) selected when
we set Nn∗ ⊂ Nn should be bounded as n → ∞. In combination with (e), this implies that
√
we have (n − n∗ )/ n + E∗ [suph∈H ρ∗n (h)] → 0 as n → ∞, which makes it suffice to focus
on G∗n (h) in (3.7) for a stable limit theorem. We obtain the following extended version of
Theorem 3.4.
Theorem 3.5. Suppose that {Yi }i∈Nn is a triangular array of random variables that are
CND with respect to (νn , M), satisfying Assumption 3.2. Suppose further that there exists
C > 0 such that for each n ≥ 1, maxi∈Nn E[H(Yi )4 ] < C, where H is an envelope of H.
Then νn converges to a G ∗ -mixture Gaussian process G∗ in l∞ (H), G ∗ -stably, such that for
any h1 , h2 ∈ H, E[G∗ (h1 )G∗ (h2 )|G ∗ ] = K(h1 , h2 |G ∗ ), a.e..
Furthermore, we have P K(Uρ (H)) = 1, where K is the G ∗ -mixture Gaussian Markov
kernel associated with G∗ .
If we take Nn∗ to be identical to Nn , Theorem 3.5 is reduced to Theorem 3.4. However,
Theorem 3.5 shows that approximation of the distribution of an empirical process by a
mixture Gaussian process is possible even if dmx → ∞ as n → ∞.
4. Appendix: Mathematical Proofs
To simplify the notation, we follow van der Vaart (1996) and write an . bn for any
sequence of numbers, whenever an ≤ Cbn for all n ≥ 1 with some absolute constant
C > 0. The absolute constant can differ across different instances of ..
For any positive integer k and i ≡ (i1 , ..., ik ) ∈ Nnk and a triangular array of random
variables {Xi }i∈Nn , we define
(4.1)
X(i) ≡
k
Y
Xi r .
r=1
The following lemma is useful for the proofs of various results.
Lemma 4.1. Suppose that νn is a neighborhood system on Nn and {Xi }i∈Nn is a triangular
array of random variables that are CND with respect to (νn , M), where M = {Mi }i∈Nn .
Furthermore, for given positive integer k, let i = (i1 , ..., ik ) ∈ Nnk be such that it has two
24
partitioning subvectors Ik,1 (i) and Ik,2 (i) of i such that the entries of Ik,1 (i) are from Nn \
ν̄n (Ik,2 (i)) and the entries of Ik,2 (i) are from Nn \ ν̄n (Ik,1 (i)). Then,
E[X(i)|Mνn (i) ] = E[X(Ik,1 (i))|Mνn (Ik,1 (i)) ]E[X(Ik,2 (i))|Mνn (Ik,2 (i)) ].
Suppose further that Condition A holds. Then,
E[X(i)|G] = E[X(Ik,1 (i))|G]E[X(Ik,2 (i))|G].
Proof: By the choice of i, we have
(4.2)
νn (i) = νn (Ik,1 (i)) ∪ νn (Ik,2 (i)),
νn (Ik,1 (i)) ⊂ νn (i), and νn (Ik,2 (i)) ⊂ νn (i).
To see the second statement, note that whenever i ∈ νn (Ik,1 (i)), we have i ∈
/ Ik,1 (i) and
i∈
/ Ik,2 (i), and there must exist j ∈ Ik,1 such that i ∈ νn (j). Since Ik,1 ⊂ i, we find that
i ∈ νn (i).
As for the first statement of the lemma, we write
E[X(i)|Mνn (i) ] = E[E[X(Ik,1 (i))|Mνn (i) , X(Ik,2 (i))]X(Ik,2 (i))|Mνn (i) ]
= E[E[X(Ik,1 (i))|Mνn (Ik,1 (i))∪νn (Ik,2 (i)) , X(Ik,2 (i))]X(Ik,2 (i))|Mνn (i) ]
= E[E[X(Ik,1 (i))|Mνn (Ik,1 (i)) ]X(Ik,2 (i))|Mνn (i) ]
= E[X(Ik,1 (i))|Mνn (Ik,1 (i)) ]E[X(Ik,2 (i))|Mνn (i) ]
= E[X(Ik,1 (i))|Mνn (Ik,1 (i)) ]E[X(Ik,2 (i))|Mνn (Ik,2 (i)) ].
The second equality follows by (4.2). The third equality follows because νn (Ik,2 (i)) \
ν̄n (Ik,1 (i)) and Ik,2 (i) are outside ν̄n (Ik,1 (i)) and X(Ik,1 (i)) is conditionally independent
of (Mνn (Ik,2 (i))\ν̄n (Ik,1 (i)) , X(Ik,2 (i))) given Mνn (Ik,1 (i)) by the CND property of {Xi }. The
fourth equality follows because νn (Ik,1 (i)) ⊂ νn (i) . The fifth equality uses the fact that
νn (Ik,1 (i)) \ ν̄n (Ik,2 (i)) is outside of ν̄n (Ik,2 (i)) and the CND property of {Xi }i∈Nn .
Let us turn to the second statement of the lemma and now assume that Condition A
holds. We write
E[X(i)|G] = E[E[X(Ik,1 (i))|Mνn (Ik,1 (i)) , X(Ik,2 (i))]X(Ik,2 (i))|G]
= E[E[X(Ik,1 (i))|Mνn (Ik,1 (i)) ]X(Ik,2 (i))|G]
= E[X(Ik,1 (i))|G]E[X(Ik,2 (i))|G].
The second equality follows because Ik,2 (i) is outside ν̄n (Ik,1 (i)) and by the CND property.
The third equality follows by Condition A, i.e., Mνn (Ik,1 (i)) and X(Ik,2 (i)) are conditionally
independent given G.
25
Let us present the proof of Theorem 3.1. Recall the notation in the theorem, and define
P
P
≡ Xi /σn , W σ ≡ i∈Nn Xiσ , and Wiσ ≡ j∈ν̄n (i) Xjσ . Let us define for t ∈ R and ω ∈ Ω,
Xiσ
ϕn (t)(ω) ≡ E[exp(itW σ )|G](ω),
√
where i ≡ −1. Note that ϕn is uniformly continuous on R almost surely, and since
E[(W σ )2 |G] = 1, a.e., ϕn is twice continuously differentiable almost surely. (See Yuan and
Mei (2016).)
Lemma 4.2. Suppose that the conditions of Theorem 3.1 hold. Then for each t ∈ R,
q
0
2
3
4
2
2
|ϕn (t) + tϕn (t)| . nt dmx dav µ3 + t ndmx dav µ4 + rn , a.e.
Proof: First, as in the proof of Theorem 1 in Jenish and Prucha (2009), we decompose
(it − W σ ) exp(itW σ ) = h1,t (W σ ) − h2,t (W σ ) − h3,t (W σ ),
where
!
itW σ
h1,t (W σ ) ≡ ite
1−
X
Xjσ Wjσ
,
j∈Nn
h2,t (W σ ) ≡ e
itW σ
X
σ
Xjσ (1 − itWjσ − e−itWj ), and
j∈Nn
h3,t (W σ ) ≡ eitW
σ
X
σ
Xjσ e−itWj .
j∈Nn
Now, let us consider
!2
E[h1,t (W σ )2 |G] ≤ t2 E 1 −
X
Xiσ Wiσ
|G
i∈Nn
2
= t2 E
X X
Xiσ Xjσ |G − t2 ,
i∈Nn j∈ν̄n (i)
because
X X
i∈Nn j∈ν̄n (i)
by the definition of σn2 .
E Xiσ Xjσ |G = 1,
26
Define σij ≡ E[Xiσ Xjσ |G] and σij∗ ≡ E[Xiσ Xjσ |Mνn (i,j) ]. Note that for {i, j}, {k, l} ⊂ Nn
such that {k, l} ⊂ Nn \ ν̄n (i, j) and {i, j} ⊂ Nn \ ν̄n (k, l), we have by Lemma 4.1,
(4.3)
E[Xiσ Xjσ Xkσ Xlσ |Mνn (i,j,k,l) ] = E[Xiσ Xjσ |Mνn (i,j) ]E[Xkσ Xlσ |Mνn (k,l) ]
∗
= σij∗ σkl
.
Let
2
˜n ≡
∆
X X
2
σij −
X X
σij∗ .
i∈Nn j∈ν̄n (i)
i∈Nn j∈ν̄n (i)
Then we can write
2
(4.4)
1 =
X X
σij
i∈Nn j∈ν̄n (i)
2
=
X X
σij∗
˜n =
+∆
0
X
∗
σij∗ σkl
+
00
X
∗
˜ n,
σij∗ σkl
+∆
i∈Nn j∈ν̄n (i)
P
where the sum 0 is over (i, j, k, l) such that i ∈ Nn , j ∈ ν̄n (i), k ∈ Nn , l ∈ ν̄n (k) and
P
either {k, l} ∩ ν̄n (i, j) 6= ∅ or {i, j} ∩ ν̄n (k, l) 6= ∅, and the sum 00 is over (i, j, k, l) such
that i ∈ Nn , j ∈ ν̄n (i), k ∈ Nn , l ∈ ν̄n (k) and {k, l} ⊂ Nn \ ν̄n (i, j) and {i, j} ⊂ Nn \ ν̄n (k, l).
This implies that
2
X X
E
Xiσ Xjσ |G − 1
i∈Nn j∈ν̄n (i)
=
=
=
0
X
0
X
0
X
E[Xiσ Xjσ Xkσ Xlσ |G]
E[Xiσ Xjσ Xkσ Xlσ |G]
+
+
00
X
00
X
E[Xiσ Xjσ Xkσ Xlσ |G] − 1
∗
E[σij∗ σkl
|G] − 1
∗
˜ n |G].
|G] − E[∆
E[Xiσ Xjσ Xkσ Xlσ |G] − E[σij∗ σkl
The second equality is by (4.3) and the third equality is by (4.4). The leading sum is
P
bounded by Cnd2mx dav µ44 , because the number of the terms in the sum 0 is bounded by
C1 nd2mx dav for some constant C1 > 0.
˜ n . We write ∆
˜ n = Rn (2 − Rn ), by using (4.4), where
Let us focus on ∆
X X
Rn ≡
(σij − σij∗ ).
i∈Nn j∈ν̄n (i)
27
Since E[Rn |G] = 0, we have
˜ n |G] = −E[R2 |G] = −r2 .
E[∆
n
n
Hence,
E[h1,t (W σ )2 |G] . t2 nd2mx dav µ44 + rn2 .
Let us turn to h2,t (W σ ). Using series expansion of exp(−itx) (e.g. see (3.2) of Tikhomirov
(1980)), we bound
t2 X
t2 X X
E[h2,t (W σ )|G] ≤
E[|Xiσ |(Wiσ )2 |G] =
E[|Xiσ Xjσ Xkσ ||G].
2 i∈N
2 i∈N
n
n
j,k∈ν̄n (i)
Using arithmetic-geometric mean inequality, we can bound the last term by Ct2 ndmx dav µ33 .
Finally, let us turn to h3,t (W σ ). We write E[h3,t (W σ )|G] as
X
X
X
E[Xiσ exp(it(W σ − Wiσ ))|G] =
E Xiσ exp it
Xjσ |G .
i∈Nn
i∈Nn
j∈Nn \ν̄n (i)
The last conditional expectation is equal to
X
E E Xiσ exp it
Xjσ |Mνn (i) |G
j∈Nn \ν̄n (i)
= E E Xiσ |Mνn (i) E exp it
X
Xjσ |Mνn (i) |G = 0.
j∈Nn \ν̄n (i)
The first equality follows by CND and the second equality follows because E Xiσ |Mνn (i) =
0. Hence, it follows that
E[h3,t (W σ )|G] = 0.
Since we have
ϕ0n (t) + tϕn (t) = −i (E[(it − W σ ) exp(itW σ )|G]) ,
by collecting the results for h1,t (W σ ), h2,t (W σ ), and h3,t (W σ ), we obtain the desired result.
Proof of Theorem 3.1: For each t ∈ R,
ϕ0n (t) = −tϕn (t) − iE[(it − W σ ) exp(itW σ )|G] ≡ −tϕn (t) + γn (t), say.
28
Taking integral of both sides, we obtain the following expression:
2
2
Z t
u
t
γn (u) exp
ϕn (t) = exp −
1+
du
2
2
0
or
2
2
2Z t
u
t
t
|γn (u)| exp
(4.5)
ϕn (t) − exp −
≤ exp −
du.
2
2
2
0
Note that for all t ≥ 0,
t
Z
u2 exp(u2 /2)du ≤ exp(t2 /2)t,
0
Rt
and 0 u exp(u2 /2)du = exp(t2 /2) − 1. Applying Lemma 4.2, the last term in (4.5) for
t > 0 is bounded by tan + (1 − exp(−t2 /2))bn , where
q
3
an ≡ Cndmx dav µ3 , and bn ≡ C nd2mx dav µ44 + rn2 ,
for some absolute constant C > 0. Hence for any T ≥ 1,
Z
Z
Z
2
2
1 − e−t /2
ϕn (t) − e−t /2
dt ≤ an
dt + bn
dt ≤ 2T an + 2 log(T )bn .
t
t
[−T,T ]
[−T,T ]
[−T,T ]
Therefore, by Esseen’s inequality (see e.g. Theorem 1.5.2 of Ibragimov and Linnik (1971),
p.27), we obtain the following bound:
∆n (t; G) . T an + log(T )bn + T −1 . (an1/2 − log(an )bn ),
−1/2
by taking T = an
.
Proof of Lemma 3.1: For (i, j) and (i0 , j 0 ) such that either {i, j} ∩ ν̄n (i0 , j 0 ) = ∅ or
{i0 , j 0 } ∩ ν̄n (i, j) = ∅, E[ξij ξi0 j 0 |G] = E[ξij |G]E[ξi0 j 0 |G] = 0. Let A be the set of ((i, j), (i0 , j 0 ))
such that i ∈ Nn , j ∈ ν̄n (i), i0 ∈ Nn and j 0 ∈ ν̄n (i0 ). Then
2
4
X X
X
4 4
E
ξij
|G ≤ σ µ4
Aj ,
i∈Nn j∈ν̄n (i)
j=1
where A1 is the number of ((i, j), (i0 , j 0 )) ∈ A such that either i ∈ ν̄n (i0 ) or i0 ∈ ν̄n (i); A2
is the number of ((i, j), (i0 , j 0 )) ∈ A such that either i ∈ ν̄n (j 0 ) or j 0 ∈ ν̄n (i); A3 is the
number of ((i, j), (i0 , j 0 )) ∈ A such that either j ∈ ν̄n (i0 ) or i0 ∈ ν̄n (j); A4 is the number of
((i, j), (i0 , j 0 )) ∈ A such that either j ∈ ν̄n (j 0 ) or j 0 ∈ ν̄n (j). Thus, it is not hard to see that
4
X
j=1
completing the proof.
Aj ≤ 8nd2mx dav ,
29
Proof of Corollary 3.1: Similarly as in the proof of Lemma 4.2, we decompose
(it − W σ ) exp(itW σ ) = h1,t (W σ ) − h2,t (W σ ) − h3,t (W σ ).
The treatment of h2,t (W σ ) and h3,t (W σ ) is the same as that of the proof of Lemma 4.2.
The difference lies in the treatment of h1,t (W σ ). Using Condition A and Lemma 4.1, we
note that for {i, j}, {k, l} ⊂ Nn such that {k, l} ⊂ Nn \ ν̄n (i, j) and {i, j} ⊂ Nn \ ν̄n (k, l),
(4.6)
E[Xiσ Xjσ Xkσ Xlσ |G] = E[Xiσ Xjσ |G]E[Xkσ Xlσ |G] = σij σkl .
Following the same argument in the proof of Lemma 4.2, we find that
2
0
X X
X
σ σ
E
Xi Xj
|G − 1 =
E[Xiσ Xjσ Xkσ Xlσ |G] − σij σkl ,
i∈Nn j∈ν̄n (i)
which is bounded by Cnd2mx dav µ44 . Hence in the proof of Lemma 4.2, we do not need to
˜ n . Following the proofs of Lemma 4.2 and Theorem 3.1 for the rest of the
deal with ∆
terms, we obtain the desired result.
Lemma 4.3. Suppose that X and Y are random variables such that E[Y ] = 0, and F is the
CDF on R with density function f . Then for any r > 0 and t ∈ R,
1/(r+1)
r
r
|P {X + Y ≤ t} − F (t)| ≤ |P {X ≤ t} − F (t)| + 3 sup f (z) E[|Y | ]
.
z∈R
Proof: First, note that for any ε > 0,
(4.7) |P {X + Y ≤ t} − F (t)| ≤ |P {X + Y ≤ t, |Y | ≤ ε} − F (t)| + P {|Y | > ε}.
As for the probability inside the absolute value above, we note that
P {X + Y ≤ t, |Y | ≤ ε} ≥ P {X + ε ≤ t, |Y | ≤ ε}
≥ P {X + ε ≤ t} − P {|Y | > ε}.
Also, observe that
P {X + Y ≤ t, |Y | ≤ ε} ≤ P {X − ε ≤ t}.
Hence
|P {X + Y ≤ t, |Y | ≤ ε} − F (t)|
≤ max {|P {X + ε ≤ t} − P {|Y | > ε} − F (t)|, |P {X − ε ≤ t} − F (t)|}
≤ |P {X ≤ t} − F (t)| + sup f (z)ε + P {|Y | > ε}.
z∈R
30
From (4.7),
|P {X + Y ≤ t} − F (t)| ≤ |P {X ≤ t} − F (t)| + sup f (z)ε + 2P {|Y | > ε}
z∈R
≤ |P {X ≤ t} − F (t)| + sup f (z)ε + 2ε−r E[|Y |r ].
z∈R
By taking ε = (E[|Y |r ]/ supz∈R f (z))1/(1+r) , we obtain the desired result.
Proof of Theorem 3.2: We write ∆∗ (t; G) as
1 X
∗
(4.8)
≤
t|G
− Φ(t) ,
P
X
+
R
i
n
σn∗
∗
i∈N
n
where Rn∗ ≡
1
∗
σn
P
Xi . By Lemma 4.3, the term (4.8) is bounded by (for any r > 0)
1 X
∗
P
(X
−
E[X
|M
])
≤
t|G
− Φ(t)
∗
i
i
νn (i)
σn∗
∗
i∈Nn \Nn∗
(4.9)
i∈Nn
+
3 φ(0)(E[|Rn∗ |r |G])1/r + ρ∗r
r/(r+1)
,
where φ denotes the density of N (0, 1). By Lemma 2.2, {Xi }i∈Nn∗ is CND with respect to
(νn∗ , M∗ ), we apply Theorem 3.1 to the leading term to obtain its bound as
q
q
∗ ∗
∗
∗3
∗3
∗4
∗
∗
∗
∗
∗2
C n dmx dav µ3 − C log(n dmx dav µ3 ) n∗ d∗2
mx dav µ4 + rn
for some constant C > 0. Now let us seek a desired bound for the second term on the
right hand side of (4.9). We write
X Xi
X
Rn∗ =
=
ξi , say.
σn∗
∗
∗
i∈Nn \Nn
i∈Nn \Nn
Now, choose 1 ≤ r ≤ 4 and write
r
r
X
X
E
ξi |G ≤
(E[|ξi |r |G])1/r ≤ (|Nn \ Nn∗ |µ̃∗r )r .
i∈Nn \Nn∗
i∈Nn \Nn∗
Hence the last term in (4.9) is bounded by C(|Nn \ Nn∗ |µ̃∗r )r/(r+1) , delivering the desired
result.
Lemma 4.4. Suppose that (D, d) is a given metric space and for each n ≥ 1, ξn , ζn , ζ are
D-valued random variables. If for each ε > 0, P ∗ {d(ξn , ζn ) > ε} → 0 and ζn → ζ, G-stably,
as n → ∞, then
ξn → ζ, G-stably.
31
Proof: First note that ζn → ζ, G-stably if and only if for all event U ∈ G and any closed
set F ∈ B(D),
(4.10)
limsupn→∞ P ∗ {ζn ∈ F } ∩ U ≤ P {ζ ∈ F } ∩ U.
(This can be shown following the proof of Theorem 1.3.4 (iii) of van der Vaart and
Wellner (1996).) Using this and following the same arguments in the proof of Lemma
1.10.2 of van der Vaart and Wellner (1996), we deduce the lemma.
˜ n , ξn ), (ζn , ξ)) >
Proof of Lemma 3.2: (i) Since P ∗ {d(ξn , ξ) > ε} → 0, we have P ∗ {d((ζ
˜ 1 , f2 ), (g1 , g2 )) = d(f1 , g1 ) + d(f2 , g2 )
ε} → 0, where d˜ is a metric on D × D defined as d((f
for f1 , f2 , g1 , g2 ∈ D. Furthermore, note that (ζn , ξ) → (ζ, ξ), G-stably, because ζn → ζ,
G-stably, and ξ is G-measurable. Now the desired result follows by Lemma 4.4.
(ii) Note that ζn0 → ζ 0 , G-stably, if and only if for any event U ∈ G and any open set
G ∈ B(D),
(4.11)
liminfn→∞ P∗ {ζn0 ∈ G} ∩ U ≥ P {ζ 0 ∈ G} ∩ U,
where P∗ denotes the inner probability. Using this and following the same arguments in
the proof of Theorem 3.27 of Kallenberg (1997) for the continuous mapping theorem for
weak convergence, we can complete the proof of (ii).
For the proof of Theorem 3.3, we use the following lemma.
Lemma 4.5. If f : D 7→ R is bounded and continuous, and K ⊂ D is compact, then for
every > 0 there exists τ > 0 such that, if x ∈ K and y ∈ D with kx − yk < τ , then
|f (x) − f (y)| < .
Proof of Theorem 3.3: First, let us suppose that (i) and (ii) hold. To see that the
marginal P L of L is a tight Borel law, note that the stable finite dimensional convergence
of ζn implies the convergence of the finite dimensional distributions of ζn . Combining
this with the asymptotic ρ-equicontinuity and using Theorems 1.5.4 and 1.5.7 of van der
Vaart and Wellner (1996), we obtain that P L is a tight Borel law. The fact that P L is
concentrated on Uρ (H) follows from Theorem 10.2 of Pollard (1990).
Now let us show the G-stable convergence of ζn . We follow the arguments in the
proof of Theorem 2.1 of Wellner (2005). Let ζ ∈ Uρ (H) be a random element whose
distribution is the same as P L. Since (H, ρ) is totally bounded, for every δ > 0 there
exists a finite set of points Hδ that is δ-dense in H i.e. H ⊂ ∪h∈Hδ B (h, δ) where B (h, δ) is
the open ball with center h and radius δ. Thus, for each h ∈ H, we can choose πδ (h) ∈ Hδ
32
such that ρ(πδ (h), h) < δ. Define
ζn,δ (h) = ζn (πδ (h)) , and ζδ (h) = ζ(πδ (h)) for h ∈ H.
By the G-stable convergence of the finite dimensional projection of ζn , we have for each
U ∈ G and for each bounded and continuous functional f : `∞ (H) 7→ R,
(4.12)
E∗ [f (ζn,δ )1U ] → E [f (ζδ )1U ] .
Furthermore, the a.e. uniform continuity of the sample paths of ζ implies that
(4.13)
P lim sup |ζ(h) − ζδ (h)| = 0 = 1.
δ→0 h∈H
For each bounded and continuous functional f : `∞ (H) 7→ R, and for each U ∈ G such
that P (U ) > 0,
|E∗ [f (ζn )1U ] − E [f (ζ)1U ] |
≤ |E∗ [(f (ζn ) − f (ζn,δ ))1U ] | + |E [(f (ζn,δ ) − f (ζδ ))1U ] | + |E [(f (ζδ ) − f (ζ))1U ] |.
The last two absolute values vanish as n → ∞ and then δ → 0 by (4.12) and by
(4.13) combined with the Dominated Convergence Theorem. We use the asymptotic
ρ-equicontinuity of ζn and Lemma 4.5 and the fact that P L is a tight law, and follow
standard arguments to show that the leading difference vanishes as n → ∞ and then
δ → 0. (See the proof of Theorem 2.1 of Wellner (2005) for details.)
Since the G-stable convergence of ζn implies that of its finite dimensional distributions,
and the weak convergence of ζn , the converse can be shown using the standard arguments. (Again, see the proof of Theorem 2.1 of Wellner (2005).)
Proof of Lemma 3.3: The proof follows that of Theorems 2.3 and 3.4 of Janson (2004).
In particular (3.6) follows from Theorem 2.3. However, for (3.5), we need to modify
the proof of Theorem 3.4 because σi2 ’s are not necessarily G-measurable, and hence the
equations (3.9) and (3.10) on page 241 do not necessarily follow.
First, without loss of generality, we set M = 1. Following the proof of Theorem 3.4 in
Janson (2004) (see (3.7) there), we obtain that for any c ≥ 0,
(4.14)
E[exp(cXi )|Mνn (i) ] ≤ exp(σi2 g(c)),
where g(c) = ec − 1 − c. Let Nj∗ ⊂ Nn , j = 1, ..., J be disjoint subsets which partition Nn
such that for any i1 , i2 ∈ Nj∗ , i1 6= i2 , i1 ∈
/ νn (i2 ) and i2 ∈
/ νn (i1 ). Fix u ≥ 0, pj ≥ 0, j =
PJ
P
1, ..., J such that j=1 pj = 1 and wj ∈ [0, 1], j = 1, ..., J such that j∈Nn :i∈N ∗ wj = 1 for
j
all i ∈ Nn . Then using Lemma 4.1 and (4.14) and following the same argument in (3.8)
33
of Janson (2004),
"
E exp u
!
X
i∈Nn
Xi
#
|G
Y
wj u
≤
Xi |Mνn (i) |G
pj E
E exp
p
j
∗
j=1
i∈Nj
J
X
X
wj u
|G
≤
pj E exp
σi2 g
pj
j=1
i∈Nj∗
"
! #
J
X
X
w
u
j
|G ,
≤
pj E exp
σi2 g
p
j
j=1
i∈N
J
X
n
because g(·) ≥ 0. The last term above is bounded by
J
X
wj u
pj E exp (κn + Vn )g
(4.15)
|G
pj
j=1
where
"
!
κn = log E exp
X
σi2 − Vn
#
|G .
i∈Nn
As for κn , note that
κn ≤ log E [exp (|Vn |) |G] = Vn ,
P
because Vn is G-measurable. Let W = Jj=1 wj and take pj = wj /W to rewrite (4.15) as
(4.16)
E [exp ((κn + Vn )g (uW )) |G] .
Hence we have for each t ≥ 0,
(
)
X
P
Xi > t|G ≤ E [exp ((κn + Vn )g (uW ) − ut) |G] .
i∈Nn
If we take
1
u=
log
W
1
+1
(κn + Vn )W
and let ϕ(x) ≡ (1 + x) log(1 + x) − x, the last bound becomes
t
t2
exp −(κn + Vn )ϕ
≤ exp −
(κn + Vn )W
2W (W (κn + Vn ) + t/3)
t2
≤ exp −
,
2W (2W Vn + t/3)
where the first inequality follows by the inequality: ϕ(x) ≥ x2 /(2(1+x/3)), x ≥ 0, and the
last inequality follows by (4.16). Now, as in the proof of Theorem 2.3 of Janson (2004),
34
the rest of the proof can be proceeded by taking {(Nj∗ , wj )}Jj=1 as a minimal fractional
proper cover of Nn .
Proof of Lemma 3.4: We adapt the proof of Theorem A.2 of van der Vaart (1996) to
accommodate the CND property of {Yi }i∈Nn . Fix q0 so that
2−q0 ≤ ρ̄n (H) ≤ 2−q0 +1
N
q
Hqi such that
and for each q ≥ q0 , construct a nested sequence of partitions H = ∪i=1
!
1 X ∗
(4.17)
E
sup |h (Yj ) − g (Yj )|2 < 2−2q for every i = 1, ..., Nq .
n j∈N
h,g∈Hqi
n
By the definition of ρ̄n (h) and the bracketing entropy, Nq can be taken to satisfy
log Nq ≤
q
X
log 1 + N[] (2−r , H, ρ̄n ) .
r=q0
Choose for each q a fixed element hqi from each Hqi and set πq h = hqi and 4q h =
(suph,g∈Hqi |h − g|)∗ , whenever h ∈ Hqi , where (h)∗ defines the minimal measurable cover
of h (Dudley (1985).) Then (ρ̄n (4q h))2 < 2−2q from (4.17), and πq h and 4q h run through
a set of Nq functions as h runs through H. Define for each fixed n and q ≥ q0 , the following
numbers and indicator functions:
p
αq = 2−q / log Nq+1 ,
√
√
Aq−1 h = 1 4q0 h ≤ nαq0 , ..., 4q−1 h ≤ nαq−1 ,
√
√
√
Bq h = 1 4q0 h ≤ nαq0 , ..., 4q−1 h ≤ nαq−1 , 4q h > nαq ,
√
Bq0 h = 1 4q0 h > nαq0 .
Because the partitions are nested, Aq h and Bq h are constant in h on each of the partitioning sets Hqi at level q. Now decompose h = (h − πq0 h) + πq0 h ≡ I + II, say, with
I ≡ h − πq0 h = (h − πq0 h). Then we can write
I = (h − πq0 h) Bq0 h +
∞
X
q=q0 +1
(h − πq h) Bq h +
∞
X
q=q0 +1
≡ Ia + Ib + Ic , say .
We analyze the empirical process at each of Ia , Ib and Ic .
(πq h − πq−1 h) Aq−1 h
35
Control of Ia : Let us bound E∗ [suph∈H |Gn ((h − πq0 h)Bq0 h)|] by
E∗
1 X
(|h − πq0 h|Bq0 h)(Xi )
sup √
n i∈N
h∈H
n
!
1 X
+ sup √
E[(|h − πq0 h|Bq0 h)(Xi )|Mν(i) ] .
n i∈N
h∈H
n
√
√
Since |h − πq0 h|Bq0 h ≤ 2H1{2H > nαq0 } ≤ 4H 2 /( nαq0 ), we bound the last expression
by
−1
p
p
2
−q0
/
ρ̄
(H)
≤
8
2
8αq−1
log
N
2−2q0 +2 ≤ 32 · 2−q0 log Nq0 +1 ,
n
q0 +1
0
due to our choice of q0 satisfying that ρ̄n (H)2 ≤ 2−2q0 +2 .
P
Control of Ic : For Ic = ∞
q=q0 +1 (πq h − πq−1 h) Aq−1 h, there are at most Nq − 1 functions
πq h − πq−1 h and at most Nq−1 − 1 functions Aq−1 h. Since the partitions are nested, the
√
function |πq h − πq−1 h| Aq−1 h is bounded by 4q−1 hAq−1 h ≤ nαq−1 . Applying Corollary
√
3.2 (with J = nαq−1 and m = Nq − 1) to Gn (πq h − πq−1 h)Aq−1 h,
E max |Gn (πq h − πq−1 h) Aq−1 h||G
h∈H
!
r
. (dmx + 1) αq−1 log Nq + (log Nq ) max Vn (hs )
1≤s≤Nq −1
where {hs : s = 1, ..., Nq−1 } ≡ {(πq − πq−1 )hAq−1 h : h ∈ H}. From the law of the iterated
conditional expectations and Jensen’s inequality,
s
!
∞
X
p
E∗ sup |Gn Ic | . (dmx + 1)
(log Nq )E∗
max Vn (hs ) + 2−q log Nq
h∈H
1≤s≤Nq −1
q=q0 +1
. 2(dmx + 1)
∞
X
2−q
p
log Nq ,
q=q0 +1
where for the last inequality, we used (4.17) so that
!!
X
1
(4.18) E∗
max Vn (hs ) ≤ E∗
max
. 2−q+1 .
E E[h2s (Yi ) |Mi ]|G
1≤s≤Nq −1
1≤s≤Nq −1
n i∈N
n
Control of Ib and II: The proof of these parts are the same as that of Theorem A.2 of van
der Vaart (1996) except that we use ρ̄n (·) instead of k·kP,2 so that we have E∗ kGn Ib kH .
p
p
P∞
−q
log Nq and E∗ kGn IIkH . 2−q0 log Nq0 .
q=q0 +1 2
36
Now collecting the results for Ia , Ib , Ic and II, we have
∞
X
p
∗
2−q log Nq
E sup |Gn (h)| . (dmx + 1)
h∈H
q=q0 +1
Z
. (dmx + 1)
ρ̄n (H)
q
1 + log N[] (ε, H, ρ̄n )dε,
0
giving the required result.
Proof of Theorem 3.4: We prove conditions for Theorem 3.3. Let us first consider the
convergence of finite dimensional distributions. Without loss of generality, we consider
the CLT for Gn (h) for some h ∈ H such that V ar(Gn (h)|G) > 0, a.e.. Assumption 3.1 (a)
together with the moment condition for the envelope H implies that dmx < C for all n ≥ 1
for some C > 0. We apply Theorem 3.1 to obtain the convergence of finite dimensional
distributions. By the CND property of (Yi )i∈Nn and Assumption 3.1(b), Condition (i) in
Theorem 3.3 is satisfied.
Let us prove asymptotic ρ-equicontinuity (with ρ(h, h) ≡ ρ(h) ≡ limn→∞ ρ̄n (h).) Define
Hn,δ ≡ {h − g : ρ̄n (h − g) ≤ δ, h, g ∈ H}. Then, by Lemma 3.4,
#
"
Z δq
∗
E
sup |Gn (h)| . (1 + dmx )
1 + log N[] (ε, Hn,δ , ρ̄n )dε.
h∈Hn,δ
0
By noting that Hn,δ is contained in H − H and by Assumption 3.1(a), the last bound
vanishes as δ → 0 for each n ≥ 1. Thus the asymptotic ρ-equicontinuity follows, by the
condition for ρ(h).
Proof of Theorem 3.5: By Assumption 3.2, we have
p
n − n∗
∗
∗
2
√
max E[H (Yi )] + E sup ρn (h) → 0,
n i∈Nn
h∈H
as n → ∞. Hence the desired result follows applying Theorem 3.4 to G∗n in (3.7).
References
A LDOUS , D. J., AND G. K. E AGLESON (1978): “On Mixing and Stability of Limit Theorems,” Annals of Probability, 6, 325–331.
A NDREWS , D. W. K. (1994): “Empirical Process Methods in Econometrics,” in Handbook
of Econometrics, pp. 2247–2294. Elsevier.
B ALDI , P., AND Y. R INOTT (1989): “On Normal Approximations of Distributions in Terms
of Dependency Graphs,” Annals of Probability, 17, 1646–1650.
37
B ARAB ÁSI , A.-L., AND R. A LBERT (1999): “Emergence of Scaling in Random Networks,”
Science, 286, 509–512.
B ERTI , P., L. P RATELLI , AND P. R IGO (2012): “Limit Theorems for Empirical Processes
Based on Dependent Data,” Electronic Journal of Probability, 9, 1–18.
C AI , T. T., W. L IU , AND H. H. Z HOU (2016): “Estimating Sparse Precision Matrix: Optimal Rates of Convergence and Adaptive Estimation,” Annals of Statistics, 44, 455–488.
C ANEN , N., J. S CHWARTZ , AND K. S ONG (2017): “Estimating Local Interactions Among
Many Agents Who Observe Their Neighbors,” Working Paper.
C HEN , L. H. Y., AND Q.-M. S HAO (2004): “Normal Approximation Under Local Dependence,” Annals of Probability, 32, 1985–2028.
D AWID , P. A. (1979): “Conditional Independence in Statistical Theory,” Journal of the
Royal Statistical Society, B, 41, 1–31.
D UDLEY, R. M. (1985): “An Extended Wichura Theorem, Definitions of Donsker Class,
and Weighted Empirical Distributions,” in Probability in Banach Spaces V, ed. by
A. Beck, R. Dudley, M. Hahn, J. Kuelbs, and M. Marcus, pp. 141–178. Springer.
H AHN , J., G. KUERSTEINER , AND M. M AZZOCCO (2016): “Central Limit Theory for Combined Cross-Section and Time Series,” arXiv:1610.01697 [stat.ME].
H ALL , P., AND C. C. H EYDE (1980): Martingale Limit Theory and Its Application. Academic
Press, New York, USA.
H ÄUSLER , E., AND H. L USCHGY (2010): Stable Convergence and Stable Limit Theorems.
Springer Science+Business Media, New York, USA.
I BRAGIMOV, L. A., AND Y. V. L INNIK (1971): Independent and Stationary Sequences of
Random Variables. Wolters-Noordhoff, Groningen.
JANSON , S. (1988): “Normal Convergence by Higher Semiinvariants With Applications
to Sums of Dependent Random Variables and Random Graphs,” Annals of Probability,
16, 305–312.
(2004): “Large Deviations for Sums of Partly Dependent Random Variables,”
Random Structures and Algorithms, 24, 234–248.
J ENISH , N., AND I. R. P RUCHA (2009): “Central Limit Theorems and Uniform Laws of
Large Numbers for Arrays of Random Fields,” Journal of Econometrics, 150, 86–98.
KALLENBERG, O. (1997): Foundations of Modern Probability. Springer, New York.
KUERSTEINER , G. M., AND I. R. P RUCHA (2013): “Limit Theory for Panel Data Models
with Cross Sectional Dependence and Sequential Exogeneity,” Journal of Econometrics,
174, 107–126.
L AURITZEN , S. L. (1996): Graphical Models. Springer, New York.
L AURITZEN , S. L., A. P. D AWID , B. N. L ARSEN , AND H.-G. L EIMER (1990): “Independence
Properties of Directed Markov Fields,” Networks, 20, 491 – 505.
38
L EUNG, M. P. (2016): “Treatment and Spillover Effects under Network Interference,”
Working Paper.
M EINSHAUSEN , N., AND P. B ÜHLMANN (2008): “High-Dimensional Graphs and Variable
Selection With the LASSO,” Annals of Statistics, 34(3), 1436–1462.
P ENROSE , M. (2003): Random Geometric Graphs. Oxford University Press, Oxford, UK.
P OLLARD , D. (1990): Empirical Processes: Theorey and Applications. NSF-CBMS Regional
Conference Series in Probability and Statistics, Volume 2, Institute of Mathematical
Statistics, Hayward, USA.
R INOTT, Y., AND V. R OTAR (1996): “A Multivariate CLT for Local Dependence with
n1/2 log n Rate and Applications to Multivariate Graph Related Statistics,” Journal of
Multivariate Analysis, 56, 333–350.
S ONG, K. (2015): “Measuring the Graph Concordance of Locally Dependent Observations,” arXiv:1504.03712v2 [stat.ME].
S TEIN , C. (1972): “A Bound for the Error in the Normal Approximation to the Distribution
of a Sum of Dependent Random variables,” Proceedings in the Sixth Berkeley Symposium
on Mathematical Statistics and Probability, 2, 583–602.
T IKHOMIROV, A. N. (1980): “On the Convergence Rate in the Central Limit Theorem for
Weakly Dependent Random Variables,” Theory of Probability and Its Applications, 25,
790–809.
VAN DER VAART, A. W. (1996): “New Donsker Classes,” Annals of Statistics, 24, 2128–
2140.
VAN DER VAART , A. W., AND J. A. W ELLNER (1996): Weak Convergence and Empirical
Processes. Springer, New York, USA.
W ELLNER , J. A. (2005): Empirical Processes: Theory and Applications. Special Topics
Course Notes at Delft Technical University.
W U, W. B. (2005): “Nonlinear System Theory: Another Look at Dependence,” Proceedings of National Academy of Science USA, 102, 14150–14154.
Y UAN , D., AND L. M EI (2016): “Some Results Following from Conditional Characteristic
functions,” Communications in Statistics - Theory and Methods, 45, 3706–3720.
| 10math.ST
|
Tight Lower Bounds for Planted Clique
in the Degree-4 SOS Program
arXiv:1507.05136v3 [cs.DS] 11 Mar 2016
Prasad Raghavendra
∗
Tselil Schramm†
Abstract
√
We give a lower bound of Ω̃( n) for the degree-4 Sum-of-Squares SDP relaxation for the
planted clique problem. Specifically, we show that on an Erdös-Rényi graph G(n, 12 ), with high
probability there is a feasible
point for the degree-4 SOS relaxation of the clique problem with
√
an objective value of Ω̃( n), so that the program cannot
distinguish between a random graph
√
and a random graph with a planted clique of size Õ( n). This bound is tight.
We build on the works of Deshpande and Montanari and Meka et al., who give lower bounds
of Ω̃(n1/3 ) and Ω̃(n1/4 ) respectively. We improve on their results by making a perturbation to
the SDP solution proposed in their work, then showing that this perturbation remains PSD as
the objective value approaches Ω̃(n1/2 ).
In an independent work, Hopkins, Kothari and Potechin [HKP15] have obtained a similar
lower bound for the degree-4 SOS relaxation.
∗
UC Berkeley, [email protected]. Supported by NSF Career Award, NSF CCF-1407779 and the Alfred.
P. Sloan Fellowship.
†
UC Berkeley, [email protected]. Supported by an NSF Graduate Research Fellowship (NSF award no
1106400).
1
Introduction
In the Maximum Clique problem, the input consists of a graph G = (V, E) and the goal is to find
the largest subset S of vertices all of which are connected to each other. The Maximum Clique
problem is NP-hard to approximate within a n1−ε -factor for all ε > 0 [Hås96, Kho01].
Karp [Kar76] suggested an average case version of the Maximum Clique problem on random
graphs drawn from the Erdös-Rényi distribution G(n, 12 ). A heuristic argument shows that an
Erdös-Rényi graph G ∼ G(n, 12 ) has a clique of size (1 − o(1)) log n with high probability: given
such a graph, choose a random vertex, then choose one of its neighbors, then choose a vertex
adjacent to both, and continue this process until there is no vertex adjacent to the clique. After
log n steps, the probability that another vertex can be added is n1 , and so after about log n steps
this process terminates. This heuristic argument can be made precise, and one can show that this
greedy algorithm can find a clique of size (1 + o(1)) log n in an instance of G(n, 12 ) in polynomial
time.
Indeed, with some work it can be shown that the largest clique in an instance of G(n, 21 )
actually has size (2 ± o(1)) log n with high probability [GM75, Mat76, BE76]. But while some
clique of size (1 ± o(1)) log n can easily be found in polynomial time (using the heuristic from the
previous paragraph), an efficient algorithm for finding the clique of size 2 log n has been much more
elusive. In his seminal paper on the probabilistic analysis of combinatorial algorithms, Karp asked
whether there exists a polynomial-time algorithm for finding a clique of size (1 + ε) log n for any
fixed constant ε > 0 [Kar76]. Despite extensive efforts, there has been no algorithmic progress on
this question since.
The planted clique problem is a natural variant of this problem wherein the input is promised to
be either a graph drawn from G ∼ G(n, 12 ) or a graph G ∼ G(n, 12 ) with a clique of size k planted
within its vertices. The goal of the algorithm is to distinguish between the two distributions.
For k > (2 + ε) log n, there is a simple quasi-polynomial time algorithm that distinguishes the
two distributions. The algorithm simply tries all subsets of (2 + ε) log n vertices, looking for a
clique. For a random graph G(n, 21 ), there are no cliques of size (2 + ε) log n, but there is one in the
planted distribution. Clearly, the planted clique problem becomes easier as the planted clique’s size
√
k increases. Yet there are no polynomial-time algorithms known for this problem for any k < o( n).
√
For k = Ω( n), a result of Alon et al. uses random matrix theory to argue that looking at the
spectrum of the adjacency matrix suffices to solve the decision problem [AKS98].
The works of [FK08, BV09] show that, if one were able to efficiently calculate the injective tensor
norm of a certain random order-m tensor, then by extending the spectral algorithm of [AKS98] one
would have a polynomial-time algorithm for k > n1/m . However, there is no known algorithm that
efficiently computes the injective tensor norm of an order-m tensor; in fact computing the inective
tensor norm is hard to approximate in the general case [HM13].
While algorithmic progress has been slow, there has been success in proving strong lower bounds
for the planted clique problem within specific algorithmic frameworks. The first such bound was
given by Jerrum, who showed that a class of Markov Chain Monte Carlo algorithms require a superpolynomial number of steps to find a clique of size (1+ ε) log n, for any fixed ε > 0, in an instance of
G(n, 21 ) [Jer92]. Feige and Krauthgamer showed that r-levels of the Lovász-Schriver SDP hierarchy
√
are needed to find a hidden clique of size k ≥ Ω̃( n/2r ) [FK00, FK03]. Feldman et al. show
(for the planted bipartite clique problem) that any “statistical algorithm” cannot distinguish in a
√
polynomial number of queries between the random and planted cases for k < Õ( n) [FGR+ 12].
More recently, there has been an effort to replicate the results of [FK00, FK03] for the Sumof-Squares (or SOS) hierarchy, a more powerful SDP hierarchy. The recent work of [MPW15]
achieves a Ω̃(n1/2r )-lower bound for r-rounds of the SOS hierarchy, by demonstrating a feasible
1
solution for the level-r SDP relaxation with a large enough objective value in the random case.
The work of [DM15a] achieves a sharper Ω̃(n1/3 ) lower bound for the Meka-Potechin-Wigderson
SDP solution, but only for r = 2 rounds; a counterexample of Kelner (which may be found in
[Bar14]) demonstrates that the analysis of [DM15a] is tight for the integrality gap instance of
[DM15a, MPW15] within logarithmic factors.
This line of work brings to the fore the question: can a d = O(1)-degree SOS relaxation solve
√
the planted clique problem for some k < n? While lower bounds are known for Lovász-Schrijver
SDP relaxations for planted clique [FK00, FK03], SOS relaxations can in general be much more
powerful than Lovász-Schrijver relaxations. For example, while there are instances of unique games
that are hard for poly(log log n)-rounds of the Lovász-Schrijver SDP hierarchy [KS09, RS09], recent
work has shown that these instances are solved by degree-8 SOS hierarchy [BBH+ 12].
Moreover, even the degree-4 SOS relaxation proves to be surprisingly powerful in a few applications:
• First, the work of Barak et al. [BBH+ 12] shows that a degree 4 SOS relaxation can certify
2 − to − 4 hypercontractivity of low degree polynomials over the hypercube. This argument
is the reason that hard instances for Lovász-Schriver and other SDP hierarchies constructed
via the noisy hypercube gadgets are easily refuted by the SOS hierarchy.
• Second, a degree-4 SOS relaxation can certify that the 2-to-4 norm of a random subspace of
√
dimension at most o( n) is bounded by a constant (with high probability over the choice of
the subspace) [BBH+ 12]. This average-case problem has superficial similarities to the planted
clique problem.
In this work, we make modest progress towards a lower bound for SOS relaxations of planted
clique by obtaining a nearly tight lower bound for the degree-4 SOS relaxation (corresponding to
two rounds, r = 2). More precisely, our main result is the following.
Theorem 1.1. Suppose that G ∼ G(n, 21 ). Then with probability 1 − O(n−4 ), there exists a feasible
solution to the SOS-SDP of degree d = 4 (r = 2) with objective value
√
n
1
polylog n .
Note that by the work of [AKS98], this result is tight up to logarithmic factors. In an independent work, Hopkins, Kothari and Potechin [HKP15] have obtained a similar result.
Our work builds heavily on previous work by Meka, Potechin and Wigderson [MPW15] and
Deshpande and Montanari [DM15a]. Since the SDP solution constructed in these works is infeasible
√
for k > n1/3 , we introduce a modified SDP solution with objective value Ω̃( n), and prove that for
a random graph G the solution is feasible with high probability. At the parameter setting for which
the objective value becomes Ω(n1/3 ), the SDP solutions of [DM15a, MPW15] violate the PSDness
constraint, or equivalently, there exists a set of test vectors X such that xT M x < 0 for all x ∈ X.
Our feasible SDP solution is a perturbation of their solution–we add spectral mass to the solution
along the vectors from the set X, then enforce the linear constraints of the SDP program.
1.1
Notation
We use the symbol to denote the PSD ordering on matrices, saying that A 0 if A is PSD and
that A B if A − B 0. When we wish to hide constant factors for clarity, we use a . b to denote
that a ≤ C · b for some constant C.
1
We have made no effort to optimize logarithmic factors in this work; a more delicate analysis of the required
logarithmic factors is certainly possible.
2
= 1 ∀i ∈ [n], or the all-1’s vector. We denote
1n ∈ Rn the vector such that 1n (i) def
def
def
def
the normalized version of this vector by 1̃n = 1n /k1n k. Further, we use Jn = 11⊤ and Qn = 1̃1̃⊤ .
We denote by
We will drop the subscript when n is clear from context.
In our notation, we at times closely follow the notation of [DM15a], as our paper builds on their
results and we recycle many of their bounds.
n
For convenience, we will use the shorthand n = n log n.
We
will
abuse
notation
by
using
2
to refer to both the binomial coefficient and to the set n2 = {(a, b) | a, b ∈ [n], a 6= b}. We will
S
n
also use the notation ≤k
to refer to the union of sets ki=0 ni . Further, when we give a vector
n
v ∈ R( 2 ) , we will identify the entries of v by unordered pairs of elements of [n].
Throughout the paper, we will (unless otherwise stated) work with some fixed instance G of
G(n, 21 ), and denote by Ai ∈ Rn the “centered” ith row of the adjacency matrix of G, with jth
entry equal to 1 if the edge (i, j) ∈ E, equal to −1 if the edge (i, j) 6∈ E, and equal to 0 for j = i.
We will use Aij to denote the jth index of Ai .
1.2
Organization
In Section 2, we give background material on the degree-4 SOS relaxation for the max-clique problem, describe the integrality gap of Deshpande and Montanari for the planted clique problem, and
√
explain the obstacle they face to reach an integrality gap value of Ω̃( n). We then describe our integrality gap instance, motivating our construction using the obstacle for the Deshpande-Montanari
and Meka-Potechin-Wigderson witness, and give an overview of our proof that our integrality gap
instance is feasible. In Section 3, we prove that our witness is PSD, completing the proof of feasibility. Section 4 contains our concentration bounds for random matrices that arise within our proofs.
In our proof, we reuse several bounds proved by Deshpande and Montanari. As far as possible, we
restate the claims from [DM15a] as they are used; for convenience, in Appendix A, we list a few
other claims from Deshpande and Montanari that we use in this paper.
2
Preliminaries and Proof Overview
In this section, we describe the degree-4 SOS relaxation for the max-clique SDP and give background
on the Deshpande-Montanari witness. We then describe our own modified witness, and give an
overview of the proof that our witness is feasible (the difficult part being showing that our witness
is PSD). The full proof of feasibility is deferred to Section 3.
2.1
Degree-4 SOS Relaxation for Max Clique
The degree d = 4 SOS relaxation for the maximum clique problem is a semidefinite program whose
n
× n
variables are X ∈ R(≤2) (≤2) . For a subset S ⊆ V with |S| ≤ 2, the variable XS indicates whether
S is contained in the maximum clique. For a graph G on n vertices, the program can be described
as follows.
X
Maximize
X{i},{i}
(2.1)
i∈[n]
n
≤2
whenever S1 ∪ S2 = S3 ∪ S4
subject to XS1 ,S2 ∈ [0, 1]
∀S1 , S2 ∈
XS1 ,S2 = XS3 ,S4
if S1 ∪ S2 is not a clique in G
XS1 ,S2 = 0
3
X∅,∅ = 1
X0
It is instructive to think of the variable XS as a pseudoexpectation of the product of indicator
variables, or a pseudomoment:
#
"
Y
I(i ∈ clique) .
XS = Ẽ
i∈S
Intuitively, the constraints of the SDP force the solution to behave somewhat like the moments of a
probability distribution over integral solutions, although they needn’t correspond to the moments
of a true distribution, hence the term pseudomoment. For more background, see e.g. [Bar14].
The pseudmoment interpretation of the SDP solution motivates the choice of the witness in the
prior work. For example, we may notice that the objective
function in this view is simply the
P
pseudoexpectation of the size of the planted clique, Ẽ[ i∈[n] I(i ∈ clique)].
If sdpval(G, 4) denotes the optimum value of the SDP relaxation on graph G, then clearly
sdpval(G) is at least the size of the maximum clique in G. In order to prove a lower bound for
degree 4 SOS relaxation on G(n, 12 ), it is sufficient to argue that with overwhelming probability,
sdpval(G) is significantly larger than the maximum clique on a random graph. This amounts to
exhibiting a feasible SDP solution with large objective value, for an overwhelming fraction of graphs
sampled from G(n, 12 ). Formally, we will show the following:
Theorem 2.1 (Formal version of Theorem 1.1). There exists an absolute constant c ∈ N such that
√
n
P
≥ 1 − O(n−4 )
sdpval(G) ≥
c
1
log n
G∼G(n, 2 )
We obtain Theorem 2.1 by constructing a point, or witness, for each G ∼ G(n, 12 ), then proving that the point is feasible with high probability. We defer the description of our witness to
Definition 2.8 and Definition 2.9, as we spend Section 2.2 and Section 2.3 motivating our construction; however the curious reader may skip ahead to Definition 2.9 which does not require the
knowledge of additional notation.
2.2
Deshpande-Montanari Witness
Henceforth, fix a graph G that is sampled from G(n, 21 ). Both the work of Meka, Potechin and
Wigderson [MPW15] and that of Deshpande and Montanari [DM15a] construct essentially the same
SDP solution for the degree-4 SOS relaxation.
This SDP solution assigns to each clique of size 1, . . . , d, a value that depends only on its
size (in our case, d = 4). In essence, their solution takes advantage of the independence of the
G(n, p) instance. The motivating observation is that the variable XS can be thought of as a
pseudoexpectation of the indicator that S is a subclique of the planted clique. The idea is then
to make this pseudoexpectation of the indicator consistent with the true expectation under the
distribution where a clique of size k is planted uniformly at random within the instance of G(n, p).
Thus, every vertex is in the clique “with uniform probability:”
k
.
n
Then, the same principle is applied to edges, traingles, and 4-cliques, so that
k
Ẽ[X{i} ] ≈ E[I(i is in planted clique)] =
Ẽ[XS ] ≈ I(S is clique) · E[I(S is in planted clique)] =
4
|S|
n
|S|
·
1
2
(|S|)
2
.
This is the general idea of the SDP solution of [DM15a]. More formally, the SDP solution in
[DM15a] is specified by four parameters α = {αi }4i=1 as,
M (G, α) = α|A∪B| · GA∪B ,
where for a set of vertices A ⊆ V , GA is the indicator that the subgraph induced on A is a clique.
The parameters {αi }i∈[4] determine the value of the objective function, and the feasibility of the
solution. As a convention, we will define α0 = 1.
It is easy to check that the solution M (G, α) satisfies all the linear constraints of the SOS
program (2.1), since it assigns non-zero values only to cliques in G. The key difficulty is in showing
that the matrix M is PSD for an appropriate choice of parameters α.
In order to show that M (G, α) 0, it is sufficient to show that N (G, α) 0 where,
Y
NA,B = α|A∪B| ·
Gij ,
i∈A\B,j∈B\A
where Gij is the indicator for the presence of the edge (i, j). In words, N is the matrix where
the entry {a, b, c, d} is proportional not to the indicator of whether {a, b, c, d} is a clique, but to
the indicator of whether G has as a subgraph the bipartite clique with bipartitions {a, b} and
{c, d}. It is easy to see that
the matrix M is obtained by dropping from N the rows and columns
corresponding {a, b} ∈ n2 where (a, b) ∈
/ E(G). Hence N 0 =⇒ M 0.
Notice that N is a random matrix whose entries depend on the edges in the random graph G.
At the risk of over-simplification, the approach of both the previous works [MPW15] and [DM15a]
can be broadly summarized as follows:
1. (Expectation) Show that the expected matrix E[N ] has sufficiently large positive eigenvalues.
2. (Concentration) Show that with high probability over the choice of G, the noise matrix
N − E[N ] has bounded eigenvalues, so as to ensure that N = E[N ] + (N − E[N ]) 0
n
n
Here we will sketch a few key details of the argument in [DM15a]. The matrix N ∈ R(≤2)×(≤2)
n
n
can be decomposed into blocks {Nab }a,b∈{0,1,2} where Na,b ∈ R( a)×( b ) . Deshpande and Montanari
use the Schur complements to reduce the problem of proving that N 0 to facts about the blocks
{Nab }a,b∈{0,1,2} . Specifically, they show the following lemma:
n
× n
Lemma 2.2. Let A ∈ R(≤2) (≤2) be the matrix defined so that AA,B = α|A| α|B| . For a, b ∈ {0, 1, 2},
let Ha,b be the submatrix of N (G, α) − A corresponding to monomials XS with |S| = a + b. Then
N (G, α) is PSD if and only if
H22 −
H11
−1
T
H12 H11 H12
0,
0
(2.2)
(2.3)
The most significant challenge is to argue that (2.3) holds with high probability. In fact, the inequality only holds for the Deshpande-Montanari SDP solution with high probability for parameters
α for which the objective value is o(n1/3 ).
Expected matrix. The expected matrix E[H22 ] is symmetric with respect to permutations of the
vertices. It forms an association scheme (see [MPW15, DM15a]), by virtue of which its eigenvalues
and eigenspaces are well understood. In particular, the following proposition in [DM15a] is an
immediate consequence of the theory of association schemes.
5
Proposition 2.3 (Proposition 4.16 in [DM15a]). E[H22 ] has three eigenspaces, V0 , V1 , V2 such that
E[H22 ] = λ0 Π0 + λ1 Π1 + λ2 Π2 ,
where Π0 , Π1 , Π2 are the projections to the spaces V0 , V1 , V2 respectively. The eigenvalues are given
by,
(n − 2)(n − 3)
n(n − 1) 2
· α4 −
α2
32
2
(n − 4)
(n − 3)
def
λ1 (α) = α2 +
α3 −
α4
2
16
α4
def
λ2 (α) = α2 − α3 +
16
def
λ0 (α) = α2 + (n − 2)α3 +
(2.4)
(2.5)
(2.6)
Further the eigenspaces are given by,
V0 = span{1},
and
V1 = span{u | hu, 1i = 0, ui,j = xi + xj for x ∈ Rn },
n
V = R(≤2) \ (V ∪ V ) ,
2
0
1
n
where we have used R(≤2) to denote the space of vectors of real numbers indexed by subsets of n of
size at most 2.
Deviation from Expectation. Given the lower bound on eigenvalues of the expected matrix
E[H22 ], the next step would be to bound the spectral norm of the noise H22 − E[H22 ]. However,
since the eigenspaces of E[H22 ] are stratified (for the given α), with one large eigenvalue and several
much smaller eigenvalues, standard matrix concentration does not suffice to give tight bounds. To
T H −1 H along the eigenspaces of E[H ].
overcome this, Deshpande and Montanari split H22 and H12
12
22
11
More precisely, let us split H22 − E[H22 ] as
H22 − E[H22 ] = Q + K
where Q includes all multilinear entries, and K includes all non-multilinear entries, i.e., entries
K(A, B) where A ∩ B 6= ∅. Formally,
(
H22 (A, B) − E[H22 ](A, B) if A ∩ B = ∅
.
Q(A, B) =
0
otherwise
The spectral norm of the matrix Q over the eigenspaces V0 , V1 , V2 is carefully bounded in [DM15a].
Lemma 2.4. (Proposition 4.20, 4.25 in [DM15a]) With probability at least 1 − O(n−4 ), all of the
following bounds hold:
kΠa QΠb k . α4 n3/2
∀(a, b) ∈ {0, 1, 2}2
kΠ2 QΠ2 k . α4 n
1/2
kKk . α3 n
(2.7)
(2.8)
(2.9)
6
Proposition 2.3 and Lemma 2.4 are sufficient to conclude that H22 0 for parameter choices of
α that correspond to planted clique of size up to ω(n1/3 ). More precisely, to argue that with high
probability H22 0, it is sufficient to argue that, E[H22 ] E[H22 ] − H22 , i.e.,
λ0 0 0
kΠ0 QΠ0 k kΠ0 QΠ1 k kΠ0 QΠ2 k
0 λ1 0 kΠ1 QΠ0 k kΠ1 QΠ1 k kΠ1 QΠ2 k + α3 n1/2 · Id .2
0 0 λ2
kΠ2 QΠ0 k kΠ2 QΠ1 k kΠ2 QΠ2 k
Deshpande and Montanari fix α1 = κ, α2 = 4κ2 , α3 = 8κ3 and α4 = 512κ4 for a parameter κ.
Using Proposition 2.3 and Lemma 2.4, the above matrix inequality becomes,
2 4
3/2
n
n3/2 n3/2
n κ
0
0
0
nκ3 0 κ4 n3/2 n3/2 n3/2 ,
(2.10)
2
3/2
3/2
0
0 κ
n
n
n
which can be shown to hold for κ ≪ n−2/3 . Eventually, it is necessary to show (2.3), which is
−1
stronger than H22 0. This is again achieved by showing bounds on the spectra of H11
and H12 .
We refer the reader to [DM15a] for more details of the arguments.
2.3
Problematic Subspace
The SDP solution described above ceases to be PSD at κ ≃ n−2/3 which corresponds to an objective
value of O(n1/3 ). The specific obstruction to H22 0 arises out of (2.10). More precisely, the
bottom 2 × 2 principal minor which yields the constraint,
λ1
kΠ1 QΠ2 k
nκ3
−n3/2 κ4
0
≈
kΠ2 QΠ1 k
λ2
−n3/2 κ4
κ2
n
forcing κ ≪ n−2/3 . It is clear that the problematic vectors x ∈ R( 2 ) for which xT H22 x < 0
are precisely those for which xT Π2 QΠ1 x < 0 and |xT Π2 QΠ1 x| is large, i.e., Π2 x aligns with the
subspace Q(V1 ⊕ V0 ).
In fact, we identify a specific subspace W that is problematic for the [DM15a] solution. To
describe the subspace, let us fix some notation. Define the random variable Aij to be −1 if (i, j) 6∈ E,
and +1 otherwise. We follow the convention that Aii = 0.
n
def
Lemma 2.5. Let the vectors a1 , . . . , an ∈ R( 2 ) be defined so that ai (k, ℓ) = Aik Aiℓ , and let
def
W = span{a1 , . . . , an }. Then with probability at least 1 − O(n−4 ),
kΠ2 Q − Π2 ΠW Qk . α4 n
Proof. This is an immediate observation from the various matrix norm bounds in [DM15a]
(specifically Lemma A.2, Lemma A.3 and Observation A.5). We defer the detailed proof to
Appendix A.1.
Since kΠ2 QΠ1 k ≫ α4 n, the above lemma implies that all the vectors with large singular values
for Q are within the subspace W . Furthermore, we will show the following lemma which clearly
articulates that W is the sole obstruction to H22 0.
n
Here, we have identified the matrices E[H22 ] and E[H22 ] − H22 , which are matrices in R(≤2) , with the 3 × 3
matrices corresponding to diagonalizing E[H22 ] according to the three eigenspaces V0 , V1 , V2 of the expectation E[H22 ].
This is analagous to decomposing any quadratic form v ⊤ H22 v into v ⊤ (Π0 + Π1 + Π2 )H22 (Π0 + Π1 + Π2 )v.
2
7
Lemma 2.6. Suppose α ∈ R4+ satisfies
min(λ0 (α), λ1 (α), λ2 (α)) ≫ α3 n1/2 ,
λ0 (α) > λ1 (α) ≫ α4 n
3/2
,
λ2 (α) ≫ α4 n
(2.11)
(2.12)
(2.13)
then with probability 1 − O(n−4 ),
H22
Proof. Fix θ =
16kQk2
λ1 .
16kQk2
1
· E[H22 ] −
· Π2 ΠW Π2 .
4
λ1
Recall that H22 − E[H22 ] = Q + K. We can write the matrix
H22 + θ · Π2 ΠW Π2 = BW ⊥ + BW + BK +
1
E[H22 ] ,
4
where
BW ⊥
and
Π0 QΠ0
Π0 QΠ1
Π0 Q(I − ΠW )Π2
1
= E[H22 ] +
Π1 QΠ0
Π1 QΠ1
Π1 Q(I − ΠW )Π2 3
4
Π2 (I − ΠW )QΠ0 Π2 (I − ΠW )QΠ1
Π2 QΠ2
BW
1
= E[H22 ] +
4
0
0
0
0
Π2 ΠW QΠ0 Π2 ΠW QΠ1
Π0 QΠW Π2
Π1 QΠW Π2
θ · Π2 ΠW Π2
and BK = K + 41 E[H22 ].
It is sufficient to show that BW ⊥ , BW and BK 0. Using Proposition 2.3 and (2.9), BK
(λ0 − α3 n1/2 )Π0 + (λ1 − α3 n1/2 )Π1 + (λ2 − α3 n1/2 )Π2 0 when condidition (2.11) holds. Using
Proposition 2.3, Lemma 2.4 and Lemma 2.5 we can write,
BW ⊥
3/2
λ
0 0
n
n3/2 n
1 0
· 0 λ1 0 − α4 · n3/2 n3/2 n
4
0 0 λ2
n
n
n
which is PSD given the bounds on λ1 , λ2 , λ3 in conditions (2.12) and (2.13). To see this, one shows
that all the 2 × 2 principal minors are PSD.
n
On the other hand, for any x ∈ R( 2 ) , we can write
θ
xT BW x ≥ λ0 kΠ0 xk2 + kΠW Π2 xk2 − 2kQkkΠW Π2 xkkΠ0 xk
2
θ
+ λ1 kΠ1 xk2 + kΠW Π2 xk2 − 2kQkkΠW Π2 xkkΠ1 xk
2
Now we will appeal to the fact that a quadratic r(p, q) = ap2 + 2bpq + cq 2 ≥ 0 for all p, q ∈ R if
b2 < 4ac and a > 0. Since θλ1 , θλ0 ≥ 16kQk2 by condition (2.12), it is easily seen that the above
quadratic form is always non-negative, implying that BW 0.
3
Here again we diagonalize according to the subspaces V0 , V1 , V2 , as in (2.10)
8
An immediate corollary of the proof of the above lemma is the following.
Corollary 2.7. Under the hypothesis of Lemma 2.6, with probability 1 − O(n−4 ),
H22 − K
1
16kQk2
· E[H22 ] −
· Π2 ΠW Π2 .
2
λ1
The above corollary is a consequence of the fact that H22 − K = BW + BW ⊥ + 12 E[H22 ].
2.4
The Corrected Witness
Suppose we have an unconstrained matrix M that we wish to modify as little as possible so as to
ensure M 0. Given a test vector w so that wT M w < 0, the natural update to make is to take
M ′ = M + β · wwT forPa suitably chosen β. This would suggest creating a new SDP solution by
T
′ =H +β
setting H22
22
i∈[n] ai ai .
Unfortunately, the SOS SDP relaxation has certain hard constraints, namely that the non-clique
entries are fixed at zero. Moreover,
the entry XS1 ,S2 must depend only on S1 ∪ S2 . Setting the SDP
P
solution matrix to H22 + β i∈[n] ai aTi would almost certainly violate both these constraints. It is
thus natural to consider multiplicative updates to the entries of the matrix which clearly preserve
the zero entries of the matrix.
Specifically, the idea would be to consider an update of the form M ′ = M + βDw M Dw where
Dw is the diagonal matrix with entries given by the vector w. If the matrix M has a significantly
large eigenvalue along 1, i.e., M λ0 · 1̃1̃⊤ + ∆, for some matrix ∆ with k∆k ≪ λ1 , then this
multiplicative update has a similar effect as an additive update,
M ′ M + β · λ0 · wwT + βDw ∆Dw ,
where the norm of the final “error” term βDw ∆Dw is relatively small. Recall that, in our setting,
the Deshpande Montanari SDP solution matrix N does have a large eigenvalue along 1̃. We now
formally describe our SDP solution, first as a matrix according to the intuition given above, and
then as a set of pseudomoments.
n
Definition 2.8 (Corrected SDP Witness, matrix view). Let â1 , . . . , ân ∈ R(≤2) be defined so that
(
0
|A| < 2
âi (A) =
ai (A) |A| = 2.
n
Define D̂i ∈ R(≤2) to be the diagonal matrix with âi on the diagonal. Define K̂ to be the restriction
of N (G, α) to the non-multilinear entries. Also let
X
N ′ (G, α) = N (G, α) + β ·
D̂i N (G, α) − K̂ D̂i ,
i∈[n]
where β =
√1
.
100 n log n
Then our SDP witness is the matrix M ′ , defined so that
M ′ (G, α) = P N ′ (G, α) ,
where P is the projection that zeros out rows and columns corresponding to pairs (i, j) 6∈ E.
9
Definition 2.9 (Corrected SDP Witness, pseudomoments view). Let β =
√1
,
100 n log n
and let
R4+
α ∈
be a set of parameters, to be fixed later. For a subset S ⊆ [n], let G[S] be the graph
induced on G by S. For any subset of at most 4 vertices S ⊂ [n], |S| ≤ 4, we define
n
P
κ
#{edges from v to S in G} |S| = 4 and G[S] = clique,
v∈[n] (−1)
4 +β
c4 (α) · 4
κ
n
Ẽ[XS ](G, α) = c|S| (α) · |S|
|S| ≤ 3 and G[S] = clique,
|S|
0
otherwise ,
where c|S| (α) is some factor chosen for each |S| ∈ {0, . . . , 4} depending on the choice of α, which
we will set later to ensure that the final moments matrix is PSD.
Proposition 2.10. For β = 100√1n log n , and α4 ≤ 21 , with probability at least 1 − O(n−5 ), the
solution N ′ (G, α) does not violate any of the linear constraints of the planted clique SDP.
Proof. First, M ′ (S1 , S2 ) = M (S1 , S2 ) whenever |S1 ∪ S2 | < 4 so these entries satisfy the constraints
of the SDP. If |S1 ∪ S2 | = 4 then M ′ (S1 , S2 ) is given by,
X Y
Aij .
M ′ (S1 , S2 ) = α4 · I[S1 ∪ S2 is a clique] · 1 + β
i∈[n] j∈S1 ∪S2
M ′ (S
Notice P
that Q 1 , S2 ) is non-zero only if S1 ∪ S2 is a clique, and it depends only on S1 ∪ S2 .
Moreover, i∈[n] j∈S1 ∪S2 Aij is a sum over iid mean 0 random variables and therefore satisfies,
X Y
√
Aij ≤ 100 n log n ≥ 1 − O(n−10 ) .
P
i∈[n] j∈S1 ∪S2
A simple union bound over all subsets S1 ∪ S2 ∈
with probability at least 1 − O(n−5 ).
n
4
shows that M ′ (S1 , S2 ) ∈ [0, 1] for all of them
It now remains to verify that N ′ (G, α) 0. We will do this by verifying the Schur complement
conditions, as in [DM15a]. Analogous to the submatrix H22 , one can consider the corresponding
′ of N ′ . The expression for H ′ is as follows:
submatrix H22
X
′ def
1 2
H22
= H22 +
Di (H22 + 16
α2 J(n) − K)Di ,
2
i∈[n]
Here Di is the matrix with ai on the diagonal, and K is the matrix corresponding to the nonmultilinear entries (entries corresponding to monomials like x2a xb xc ), and J(n) is the all-1s matrix.
2
′ H ⊤ H −1 H
The matrices H12 and H11 are unchanged, and so we must simply verify that H22
12
12 11
′ 0.
and that H22
This concludes our proof overview. In Section 3, we verify the Schur complement conditions
and prove our main result, and in Section 4 we give the random matrix concentration results upon
which we rely throughout the proof.
3
Proof of the Main Result
′ 0, and that H ′ H ⊤ H −1 H . This will allow us
In this section, we will demonstrate that H22
12
22
12 11
to conclude that our solution matrix is PSD, and therefore is a feasible point for the degree-4 SOS
relaxation.
10
Parameters. Before we proceed further, it will be convenient to parametrize the choice of
α1 , α2 , α3 and α4 . In particular, it will be useful to fix,
def
α1 =
ρ
def
n1/2
α2 =
γρ2
n
def
α3 =
γ 3 ρ3
n3/2
def
α4 =
γ 6 ρ4
.
n2
(3.1)
for two parameters γ, ρ, which we will finally fix to γ = log4 n and ρ = log−20 n. For this setting of
parameters, the eigenvalues λ0 , λ1 , λ2 from Proposition 2.3 are bounded by,
λ0 ≥
α4 n2
γ 6 ρ4
=
64
64
λ1 ≥
α3 n
γ 3 ρ3
= 1/2
4
4n
λ2 ≥
def
α2
γρ2
=
2
2n
def
(3.2)
def
When convenient, we will also use the shorthand c1 = n1/2 α1 , c2 = nα2 , c3 = n3/2 α3 , and
def
c4 = n2 α4 .
3.1
′
Proving that H22
0
Here we will make a first step towards verifying the Schur complement conditions of Lemma 2.2 by
′ 0. Specifically, we will show the following stronger claim.
showing that H22
Theorem 3.1. For β =
at least 1 − O(n−4 ),
√1
100 n log n
and γ = log4 n, ρ < log−20 n, the following holds with probability
′
H22
Proof. Fix θ =
kQk2
λ1 .
βλ0
1
E[H22 ] +
· ΠW
8
16
′ , we have
By definition of H22
X
′
H22
= H22 + β ·
Di (H22 +
i∈[n]
− K)Di .
T
i∈[n] ai ai .
We can apply Lemma 2.6 to the H22 term and Corollary 2.7 for H22 − K,
X
1
1 2
1
E[H22 ] − θ · Π2 ΠW Π2 + β ·
Di
E[H22 ] − θΠ2 ΠW Π2 + α2 J Di .
4
2
16
i∈[n]
X
1
λ0
E[H22 ] − θ · Π2 ΠW Π2 + β ·
(dropping Π1 , Π2 , J)
Di
Π0 − θΠ2 ΠW Π2 Di .
4
2
i∈[n]
X
n
λ0
1
T
Di Π2 ΠW Π2 Di
(usingDi Π0 Di = ai ai /
)
E[H22 ] − θ · Π2 ΠW Π2 + β 2 PW − βθ
2
4
4n
Define PW =
′
H22
P
1 2
16 α2 J
i∈[n]
Now we will appeal to a few matrix concentration bounds that we show in Section 4. First,
with probability 1 − O(n−5 ), the vectors {a⊗2
i } are nearly orthogonal, and therefore form a wellconditioned basis for the subspace W .
PW
n2
· ΠW .
2
(see Lemma 4.3) .
Also, the vectors {a⊗2
i } have negligible projection on to the eigenspaces V0 , V1 which implies that
with overwhelming probability,
kΠ2 ΠW Π2 − ΠW k ≤
log2 n
· Id ,
n
11
(see Lemma 4.5) .
Finally, W is an n dimensional space. Each Di Π2 ΠW Π2 Di has only n non-zero singular values
each of which is O(1). Moreover, multiplying on the left and right by DP
i acts as a random linear
transformation/ random change of basis. Intuitively, this suggests that i Di Π2 ΠW Π2 Di has n2
eigenvalues all of which are roughly O(1). In fact, with probability 1 − O(n−5 ),
X
Di Π2 ΠW Π2 Di O(n) · Π0 + log2 n · Id
(see Lemma 4.6)
i
Substituting these bounds we get,
βλ0
θ log2 n
1
2
′
− θ · ΠW −
+ βθ log n · Id − βθ · O(n) · Π0
H E[H22 ] +
4
8
n
By Lemma 2.4, with probability at least 1 − O(n−4 ), kQk . α4 n3/2 . Substituting this bound for
2
θ = kQk
λ1 along with (3.2), finishes the proof for our choice of parameters. The details are presented
below for completeness.
√
α24 n3
n log n
θ
.
·
= log4 n · γ 3 ρ ≪ 1 ,
βλ0
α3 n
α4 n2
βθn
1
1
α2 n3
.√
. log2 n · γ 3 ρ ≪ 1
· 4 ·n·
λ0
α4 n2
n log n α3 n
Clearly λ0 > λ1 > λ2 and
because
3.2
log2 n
λ2 ≫ 100θ β log n +
n
2
,
θβ log2 n
α2 n3
1
log2 n
. 4 ·√
= log4 n · γ 8 ρ3 ≪ 1 .
·
λ2
α3 n
α2
n log n
Bounding singular values of H12
T H −1 H , Deshpande and Montanari [DM15a] observe the
Towards bounding the eigenvalues of H21
12
11
following properties of H21 with regards to the spaces V0 , V1 and V2 .
Lemma 3.2 (Consequence of Propositions 4.18 and 4.27 in [DM15a]). Let Qn ∈ Rn×n be the
orthogonal projector to the space spanned by 1n . Let p = 21 . For the matrix H21 , we have that for
sufficiently large n, with probability 1 − O(n−5 ),
k E[H21 ] − H21 k ≤ α3 n ,
and
kΠ0 E[H21 ]Qn k ≤ 14 n3/2 α3 + 2α2 n1/2
Π1 E[H21 ]Qn = 0
Π0 E[H21 ]Q⊥
n =0
1/2
kΠ1 E[H21 ]Q⊥
n k ≤ α2 n
Π2 E[H21 ]Q⊥
n = 0.
Π2 E[H21 ]Qn = 0
Unfortunately, the bound of [DM15a] on kE[H21 ] − H21 k is insufficient for our purposes, and
we require a more fine-grained bound on the deviation from expectation. In fact, outside the
problematic subspace W , we show that H21 − E[H21 ] is much better behaved.
12
Proposition 3.3. Let W = spani∈[n] {ai }, and let ΠW be the projector to that space. With proban
bility at least 1 − O(n−4 ), the following holds for every x ∈ R( 2 )
log2 n
2
2
T
2
2 2
kxk
kx Π2 (H21 − E[H21 ])k . α3 n kΠW xk +
n
Proof. From Lemma 3.2, we have that
Π2 E[H21 ] = 0.
Thus we may work exclusively with the difference from the expectation; for convenience, we let
U = H21 − E[H21 ]. Fix A = {a, b} and c, so that |{a, b}| = 2. By inspection, the entry (A, c) of U
is given by the polynomial
α3
UA,c =
(Aac Abc + (Aac + Abc )).
4
Thus, the columns of U are in W ∪ V1 . So we have that
Π2 U = Π2 (ΠW + ΠV1 \W + Π⊥
W ∪V1 )U
= Π2 ΠW U + Π2 ΠV1 \W U + Π⊥
W ∪V1 U
= Π2 ΠW U,
where the latter two terms were eliminated because V1 ⊥ V2 and W ∪ V1 ⊥ span{col(U )}.
2
In Lemma 4.5, we bound kΠ01 ΠW Π01 k ≤ O( logn n ). So we have that
kxT Π2 ΠW k2 ≤ xT (Id − Π01 )ΠW (Id − Π01 )x
= kΠW xk2 − 2xT Π01 ΠW x + xT Π01 ΠW Π01 x
≤ kΠW xk2 + 2kΠW xkkΠ01 ΠW xk + kxk2 kΠ01 ΠW Π01 k
. 2kΠW xk2 +
2 log2 n
· kxk2 ,
n
where we have applied the inequality a2 + b2 ≥ 2ab. The conclusion follows by noting that
kxT Π2 H21 k = kxT Π2 ΠW U k ≤ kxT Π2 ΠW k · kU k, and that by Lemma 3.2 kU k . α3 n.
3.3
−1
T
Bounding singluar values of H12
H11
H12
−1
′ , by splitting the matrix up according to the
We will bound H21 H11
H12 , as in our bounds on H22
eigenspaces Π0 , Π1 , Π2 .
Theorem 3.4. Let c1 , . . . , c4 be as defined in (3.1). For the choice of α in (3.1), we have that with
probability 1 − O(n−5 ),
T
−1
H12
H11
H12
c2 + c23 log2 n
c23
c3 log4 n
c23 log2 n
· Π0 + 2
·
Π
+
Π
+
· ΠW
1
2
c2
c1 n1/2
c1 n3/2
c1 n1/2
Proof. For each a, b ∈ {0, 12}, define the matrix Uab = Πa (E[H21 ] − H21 )Πb . We can verify that
for our choice of α the conditions of Lemma A.7 are met, and so we conclude that with probability
1 − O(n−5 ),
1
1
−1
H11
· Q⊥
· Qn +
n
2
α1
n(α2 − 2α1 )
n
For x ∈ R( 2 ) , let yU = (E[H12 ] − H12 )Π2 x, zU = (E[H12 ] − H12 )(Π0 + Π1 )x, xA = E[H12 ]Π0 x,
xB = E[H12 ]Π1 x and xC = E[H12 ]Π2 x. In order to simplify our computations, we will use the
following observation.
13
Observation 3.5. Given A ∈ Rm×m with A 0 and vectors x1 , . . . , xt ∈ Rm ,
T
X
X
X
xi A
xi ≤ t ·
xTi Axi
i∈[t]
i∈[t]
i∈[t]
Proof. If we set yi = A1/2 xi then the inequality reduces to,
X
X
k
yi k2 ≤ t
kyi k2 ,
i∈[t]
i∈[t]
which is an immediate consequence of triangle inequality for k·k and Cauchy-Schwartz inequality.
Using Observation 3.5,
−1
−1
−1
−1
−1
−1
T
xT H12
H11
H12 x ≤ 5 xTA H11
xA + xTB H11
xB + xTC H11
xC + yUT H11
yU + zUT H11
zU .
√
To simplify the calculations and make the dominant terms apparent, let us fix α1 = c1 / n,
α2 = c2 /n, α3 = c3 /n3/2 and α4 = c4 /n wherein each ci ∈ [1/ log 200 n, 1]. First, observe that
1
1
α1 ≫ n(α −2α2 ) for this setting of parameters.
2
1
−1
−1
−1
For the terms xTA H11
xA , xTB H11
xB and xTC H11
xC we can write,
1
1
⊥ 2
2
T −1
2
· kΠ0 E[H21 ]Qn k
· kΠ0 E[H21 ]Qn k +
xA H11 xA . kΠ0 xk
α1
n(α2 − 2α21 )
c2
(α2 + α3 n)2
· kΠ0 xk2 . 3 kΠ0 xk2
.
2
c2
(α2 − 2α1 )
1
1
2
⊥ 2
T
−1
2
· kΠ1 E[H21 ]Qn k +
· kΠ1 E[H21 ]Qn k
xB H11 xB . kΠ1 xk
α1
n(α2 − 2α21 )
c22
α22 n
kΠ1 xk2 .
· kΠ1 xk2
α1
c1 n1/2
1
1
−1
2
T
2
⊥ 2
· kΠ2 E[H21 ]Qn k +
xC H11 xC . kΠ2 xk
· kΠ2 E[H21 ]Qn k
α1
n(α2 − 2α21 )
.
=0
1
1
2
· kH21 − E[H21 ]k2 kQ⊥
· kH21 − E[H21 ]k2 kQn k2 +
nk
α1
n(α2 − 2α21 )
c23 log2 n
α23 n2
α23 n2
2
2
2
2
. kΠ0 xk + kΠ1 xk
.
kΠ
xk
+
kΠ
xk
·
+
0
1
α1
n(α2 − 2α21 )
c1 n1/2
−1
zUT H11
zU . k(Π0 + Π1 )xk2
Finally, we have
1
1
+
n(α2 − 2α21 ) α1
2 2
α3 n log2 n
α23 n2
α23 n2
α23 n2 log2 n
2
·
kΠ
xk
+
· kxk2
.
+
+
W
α1
nα1
n(α2 − 2α21 )
n2 (α2 − 2α21 )
−1
yUT H11
yU . k(H12 − E[H12 ])Π2 xk2 ·
.
c23 log2 n
c23 log4 n
2
·
kΠ
xk
+
· kxk2
W
c1 n1/2
c1 n3/2
The conclusion follows by grouping the projections, and taking the dominating terms as n grows.
14
3.4
−1
′
T
Proof of H22
H12
H11
H12
′ H T H −1 H
Theorem 3.6. For the choice of α given in (3.1), we have that H22
12 with probability
12 11
−4
1 − O(n ).
Proof. Recall that by Theorem 3.1 with probability at least 1 − O(n−4 ),
′
H22
λ0
λ1
λ2
βλ0
· Π0 +
· Π1 +
· Π2 +
ΠW .
4
4
4
16
By our choice of the parameters α1 , α2 , α3 , α4 , the conclusion of Theorem 3.4 implies that
1
1
c23
c22 + c23 log2 n
′
T
−1
H22 − H12 H11 H12
λ0 −
λ1 −
Π0 +
Π1
4
c2
4
c1 n1/2
1
c3 log4 n
c23 log2 n
1
Π2 +
ΠW
λ2 −
βλ0 −
+
4
16
c1 n3/2
c1 n1/2
0,
as desired. The details of the calculation are spelled out below for the sake of completeness; we
verify that the coefficient of each projector is non-negative.
For the space Π0 ,
c2
c23 1
·
= 3 = γ −1 ≪ 1
c2 λ0
c2 c4
For Π1 ,
c22 + c23 log2 n 1
c22 + c23 log2 n
1
·
.
= + γ 3 ρ2 log2 n ≪ 1 .
λ1
c3 c1
γ
c1 n1/2
For Π2 ,
c3 log4 n 1
1 c3 log4 n
·
≤
≪ 1.
c1 n3/2 λ2
n1/2 c1 c2
Finally for ΠW ,
1
c23 log3 n
c23 log2 n
·
=
= ρ log3 n ≪ 1 .
c1 c4
c1 n1/2 βλ0
This concludes the proof.
3.5
Proof of Main Theorem
We finally have the components needed to prove Theorem 2.1.
Proof of Theorem 2.1. First, we recall that independent of our choice of α, the SDP solution defined
in Definition 2.8 does not violate any of the linear constraints of (2.1), as shown in Proposition 2.10.
To meet the program constraints, it remains to show that for the choice of α given in (3.1), our
solution is PSD.
The solution matrix M ′ (G, α) is a principal submatrix of N ′ (G, α), and so N ′ 0 implies
′
M 0. We prove that N ′ satisfies the Schur complement conditions from Lemma 2.2 with high
′
′ = H
probability. Observing that H11
11 and H12 = H12 , we apply Theorem 3.6, which states that
15
′ H T H −1 H . For the H
for our choice of α, H22
21
11 term, we apply the lower bound on the
12 11
eigenvalues of H11 given by [DM15a] (Lemma A.7), which state that so long as α1 − α2 ≫ α2 n−1/2
and α2 − 2α21 ≥ 0, we have H11 0 with probability 1 − O(n−5 ). For our choice of α1 , α2 , we have
α2
γρ2
ρ
γρ2
=
≪
−
= α1 − α2 ,
n
n1/2
n3/2
n1/2
and
γρ2 − 2ρ2
≫ 0,
n
and so we may conclude that H11 0. Therefore by a union bound, the conditions of Lemma 2.2
are satisfied with probability 1−O(n−4 ), and N ′ (G, α) 0 and so our solution satisfies the PSDness
constraint.
√
P It remains only to1/2prove that the objective value is Ω̃( n). The objective value is simply
, concluding our proof.
i∈[n] α1 = nα1 = ρn
α2 − 2α21 =
4
Concentration of Projected Matrices
In this section, we give bounds on the spectra of random matrices that are part of the correction
term. Though we are able to recycle many of the spectral bounds of Deshpande and Montanari
[DM15a], in our modification to their witness, we introduce new matrices which also require description and norm bounds.
We obtain our bounds by employing the trace power method. The trace power P
method uses the
n×n
k
fact that if X ∈ R
is a symmetric matrix, then for even powers k, Tr(X ) = i∈[n] λi (X)k ≥
λmax (X)k . By bounding E[Tr(X k )]1/k for a sufficiently large k, we essentially obtain bounds on
the infinity norm of the vector of eigenvalues, i.e., a bound on the spectral norm of the matrix X.
A formal statement follows, and the proof is given in Appendix A.1 for completeness.
Lemma 4.1. Suppose an n × n random matrix M satisfies E[Tr(M k )] ≤ nαk+β · (γk)! for any even
integer k, where α, β, γ are constants. Then
P kM k . η −1/ log n · nα · log n ≥ 1 − η .
Our concentration proofs consist of, for each matrix in question, obtaining a bound on E[Tr(X k )].
The expression E[Tr(X k )] is a sum over products along closed paths of length k in the entries of X.
In our case, the entries of the random matrix X are themselves low degree polynomials in random
variables {Aij }i∈[n],j∈[n] where Aij is the centered random variable that indicates whether the edge
(i, j) is part of the random graph G. Thus Tr(X k ) can be written out as a polynomial in the random
variables {Aij }i,j∈[n]. Since the random variables {Aij }i,j∈[n] are centered (i.e., E[Aij ] = 0), almost
all of the terms in E[Tr(X k )] vanish to zero. The nonzero terms are precisely those monomials in
which every variable appears with even multiplicity.
For the purpose of moment calculations, we borrow much of our terminology from the work of
Deshpande and Montanari [DM15a]. Every monomial in random variables {Aij }i,j∈[n] corresponds
to a labelled graph (F = (V, E), ℓ) that consists of a graph F = (V, E) and a labelling ℓ : V → [n]
that maps its vertices to [n]. A labelling of F contributes (is nonzero in expectation), if and only
if every pair {i, j} appears an even number of times as a label of an edge in F . The problem of
bounding E[Tr(X k )] reduces to counting the number of the number of contributing labeled graphs.
For example, for a given matrix X, we may have a bound on the number of “vertices” and “edges”
in a term of E[Tr(X k )] as a function of k. In that case, we may use the following proposition, which
16
allows us to bound the number of such graphs in which every variable Aij , corresponding to an
edge between vertices i and j, appears at least twice.
Proposition 4.2. Let F = (V, E) be a multigraph and let ℓ : V → [n] be a labelling such that each
pair (i, j) appears an even number of times as the label of an edge in E. Then,
|{ℓ(v)|v ∈ V }| ≤
|E|
+ (# connected components of F)
2
Proof. From F , we form a new graph F ′ by identifying all the nodes with the same label; thus, the
number of nodes in F ′ is the number of labels in F . We then collapse the parallel edges in F ′ to
form the graph H; since each labelled edge appears at least twice, the number of edges in H is at
most half that in F . The number of nodes in H (and thus labels in F ) is at most the number of
edges in H plus the number of connected components; this is tight when H is a forest. Thus the
number of distinct labels in F is at most |E|/2 + c, where c is the number of components in F .
We apply this lemma, as well as simple inductive arguments, to bound the number of contributing terms in E[X k ] for the matrices X in question, and this allows us to bound their norms. We
give the concentration proofs the following subsection.
4.1
Proofs of Concentration
Let G be an instance of G(n, 12 ). As in the preceeding sections, define the vector Ai ∈ Rn so
that Ai (j) = 1 if (i, j) ∈ E(G), Ai (j) = −1 if (i, j) 6∈ E(G), and Ai (i) = 0. Again as in the
n
preceeding sections, define a1 , . . . , an ∈ R( 2 ) by setting ai to be the restriction of A⊗2
i to coordinates
corresponding to unordered pairs, i.e., ai ({c, d}) = Aic Aid for all {c, d} ∈ n2 . We will continue to
use the notation W = spani∈[n] (ai ), and the notation Di = diag(ai ).
We begin with a lemma that shows that the ai are close to an orthogonal basis for W .
def P
Lemma 4.3. If PW = i ai aTi then with probability at least 1 − O(n−5 ),
1
n2 (1
+ o(1)) · PW ΠW (1 − o(1)) n12 · PW
Proof. By definition, the vectors a1 , . . . , an form a basis for the subspace W .
Let R be the matrix whose ith row is ai . We will use matrix concentration to analyze the
eigenvalues of RRT , which are identical to the nonzero eigenvalues of PW = RT R.
⊗2
1
2
The (i, j)th entry of RRT is hai , aj i = 21 hA⊗2
i , Aj i = 2 hAi , Aj i . When i = j, this is precisely
1
2
T
2
2 (n − 1) , and so 2RR = (n − 1) · Idn + B, where B is a matrix that is 0 on the diagonal and
⊗2
⊗2
equal to hAi , Aj i in the (i, j)th entry for i 6= j.
Let M = B − E[B] = B − (n − 2)(Jn − Idn ). We will use the trace power method to prove that
kM k = O(n3/2 ). The (i, j)th entry of M is given by 0 for i = j, and when i 6= j
!
X
X
Aip Aiq Ajp Ajq .
Aip Aiq Ajp Ajq − (n − 2) =
M (i, j) = hAi , Aj i2 − (n − 2) =
p,q
p6=q
The expression Tr(M k ) is a sum over monomial products over variables {Aip }i,p∈[n] , where
each monomial product corresponds to a labelling L : F → [n] of a graph F . Each entry in Mij
corresponds to a sum over links, where each link is a cycle of length 4, with the vertices i, j on
opposite ends of the cycle, and the necessarily distinct vertices p, q are on the other opposite ends
of a cycle. We will refer to i, j as the center vertices and p, q as the peripheral vertices of the link.
17
Each edge (u, v) of the link is weighted by Auv . Since Aii = 0 for all i ∈ [n], for every contributing
labelling, it can never be the case that one of p, q = i. Each monomial product in the summation
Tr(M k ) corresponds to a labelling (F, L) of the graph F , where F is a cycle with k links. F has 4k
edges, and in total it has 3k vertices.
q ′′
j
i
p′′
q′
q
p
p′
The quantity Tr(M k ) is equal to the sum over all labellings of F . Taking the expectation, terms
in E[Tr(M k )] which contain a variable Auv with multiplicity 1 have expectation 0. Thus, E[Tr(M k )]
is equal to the number of labellings of F in which every edge appears an even number of times.
We prove that any such contributing labelling L : F → [n] has at most 3k/2 + 1 unique vertex
labels. We proceed by induction on k, the length of the cycle. In the base case, we have a cycle on
two links; by inspection no such cycle can have more than 5 labels, and the base case holds.
Now, consider a cycle of length k. If every label appears twice, then we are done, since there
are 3k vertices in F . Thus there must be a vertex that appears only once.
There can be no peripheral vertex whose label does not repeat, since the two center vertices neighboring a single peripheral vertex cannot have the same label in a contributing term, as M (i, i) = 0.
Now, if there exists a center vertex i whose label does not repeat, it must be that there is a matching
between its p, q neighbors so that every vertex is matched to a vertex of the same label; we identify
these same-label vertices and remove i and two of its neighbors from the graph, leaving us with a
cycle of length k − 1, having removed at most one label from the graph. The induction hypothesis
now applies, and we have a total of at most 3(k − 1)/2 + 2 ≤ 3k/2 + 1 labels, as desired.
Thus, there are at most 3k/2 + 1 unique labels in any contributing term of E[Tr(M k )]. We
may thus conclude that E[Tr(M k )] ≤ n3k/2+1 · (3k/2 + 1)3k , and applying Lemma 4.1, we have that
kM k . n3/2 with probability at least 1 − O(n−5 ).
Therefore, 2RRT = ((n − 1)2 − n + 2)Idn + (n − 2)Jn + M , and we may conclude that all
eigenvalues of RRT are (1 ± o(1)) · n2 , which implies the same of PW = RT R. Since the range of
PW and ΠW is the same, we finally have that with probability 1 − o(1)
(1 + o(1))/n2 · PW ΠW (1 − o(1))/n2 · PW ,
as desired.
The following lemma allows us to approximate the projector to V0 ∪ V1 by a matrix that is easy
to describe; we will use this matrix as an approximation to the projector in later proofs.
n
n
Lemma 4.4. Let Π01 be the projection to the vector space V0 ∪ V1 . Let P01 ∈ R( 2 ),( 2 ) be a matrix
defined as follows:
2
n−1 |{a, b, c, d}| = 2
1
P01 (ab, cd) = n−1
|{a, b, c, d}| = 3
0
|{a, b, c, d}| = 4.
Then
n−2
Π01 P01 ( n−1
) · Π01 ,
18
Proof. We will write down a basis for V0 ∪ V1 , take a summation over its outer products, and then
n
argue that this summation approximates Π01 . The vectors v1 , . . . , vn ∈ R( 2 ) are a basis for V1 ∪ V0 :
(
√1
{a, b} = {i, ·}
n−1
vi (a, b) =
0
otherwise.
2
1
. Let U ∈ Rn ×n be the matrix whose ith column is given
For any two vi , vj , we have hvi , vj i = n−1
P
by vi . Notice that the eigenvalues of i vi viT = U U T are equal to the eigenvalues of U T U , and
n−2
1
Jn + n−1
Idn . Therefore, as both matrices have the same column and row spaces,
that U T U = n−1
X
n−2
Π01
vi viT n−1
Π01 ,
i
Now, let P01 =
P
T
i vi vi ;
The conclusion follows.
we can explicitly calculate
2
n−1
1
P01 (ab, cd) = n−1
0
the entries of P01 ,
|{a, b, c, d}| = 2
|{a, b, c, d}| = 3
|{a, b, c, d}| = 4.
We will require the fact that W lies mostly outside of V0 ∪ V1 , which we prove in the following
lemma.
Lemma 4.5. With probability at least 1 − O(n−ω(log n) ),
2
log n
kΠ01 ΠW Π01 k ≤ O
.
n
Proof. Call M = Π01 ΠW Π01 . We will apply the trace
power method to M . By Lemma 4.4
(1+o(1)) P
T
′k =
and Lemma 4.3, we may exchange ΠW for
i ai ai and Π01 for P01 . Letting M
n2
P01 PW )k , we have by the cyclic property of the trace that E[Tr(M k )] ≤ Tr(M ′k ).
( (1+o(1))
n2
We consider the expression for E[Tr(M ′k )]. Let a chain consists of a set of quadruples
{aℓ , bℓ , cℓ , dℓ }ℓ∈[k] ∈ [n]4 such that for each ℓ ∈ [k], we have k{aℓ , bℓ } ∩ {cℓ−1 , dℓ−1 }| ≥ 1 (where
we identify aℓ with aℓ mod k ). Let Ck denote the set of all chains of size k. We have that,
Tr(M k ) ≤ Tr(M ′k ) =
k
Y
1 + o(1)
X
X
i1 ,...,ik {aℓ ,bℓ ,cℓ ,dℓ }ℓ∈[k] ∈Ck ℓ=1
n2
· rℓ · Aiℓ ,aℓ Aiℓ ,bℓ Aiℓ ,cℓ Aiℓ ,dℓ ,
1
2
where rℓ = n−1
or n−1
depending on whether one or both of aℓ , bℓ are common with the following
link in the chain. The quantity Tr(M k ) consists of cycles of k links, each link is a star on 4 outer
vertices aℓ , bℓ , cℓ , dℓ with center vertex iℓ , and the non-central vertices of the link must have at least
one vertex in common with the next link, so each link has 4 edges and the cycle is a connected
graph. See the figure below for an illustration (dashed lines indicate vertex equality, and are not
edges).
aℓ−1
cℓ−1
aℓ
iℓ−1
bℓ−1
cℓ
aℓ+1
iℓ+1
iℓ
dℓ−1
bℓ
dℓ
19
cℓ+1
bℓ+1
dℓ+1
Each term in the product has a factor of at most
P01 and PW . Thus we have
′k
E[Tr(M )] ≤
3
n3
k X
X
2(1+o(1))
,
n3
E
i1 ,...,ik {aℓ ,bℓ ,cℓ ,dℓ }ℓ∈[k] ∈Cn
"
k
Y
due to the scaling of the entries of
#
Aiℓ ,aℓ Aiℓ ,bℓ Aiℓ ,cℓ Aiℓ ,dℓ .
ℓ=1
The only contributing terms correspond to those for which every edge variable in the product has
even multiplicity. Each contributing term is a connected graph and has 4k edges and at most 5k
vertices where every labeled edge appears twice, so we may apply Proposition 4.2 to conclude that
there are at most 2k + 1 labels in any such cycle. We thus have that
′k
E[Tr(M )] ≤
3
n3
k
and applying Lemma 4.1, we conclude that kM k .
desired.
· n2k+1 · (5k)!,
log2 n
n
with probability 1 − O(n−ω(log n) ), as
We combine the above lemmas to bound the norm of one final matrix that arises in the computations in Theorem 3.1.
Lemma 4.6. With probability 1 − O(n−5 ),
X
Dw Π2 ΠW Π2 Dw O(n) · Π0 + O(log2 n) · Idn .
w
Proof. We begin by replacing Π2 with (1 − Π01 ), as by Lemma 4.5, Π01 can be replaced by P01
n
which has a convenient form. For any vector x ∈ R( 2 ) ,
!
!
!
!
X
X
X
X
Di Π01 ΠW Π01 Di x
Di Π01 ΠW Di x + xT
Di ΠW Di x − 2xT
Di Π2 ΠW Π2 Di x = xT
xT
≤
X
i
2
kΠW Di xk + 2kΠW Π01 Di xk · kΠW Di xk + kΠW Π01 Di xk2
T
X
Di ΠW Di
!
≤ 2xT
X
Di ΠW Di
!
≤ 2x
i
i
i
i
i
i
X
(Di x)T Π01 ΠW Π01 Di x
x+2
i
!
x + 2n kΠ01 ΠW Π01 k · kxk2 ,
where to obtain the second line we have applied Cauchy-Schwarz, to obtain the third line we have
used the fact that a2 +b2 ≥ 2ab, and to obtain the final line we have used the fact that kDi xk = kxk.
Now, the second term is O(log2 n) · kxk2 with overwhelming probability by Lemma 4.5. It
remains to bound the first term. To this end, we apply Lemma 4.3 to replace ΠW with 1+o(1)
·PW =
n2
1+o(1) P
1 P
T
· i ai ai . Let M = n2 · i Di PW Di . An entry of M has the form
n2
M (ab, cd) =
1
n+
n2
X
i6=j
Aia Aib Aic Aid Aja Ajb Ajc Ajb .
20
n
n
⊤
is the all-ones matrix in R( 2 )×( 2 ) and
Thus we can see that M = n1 J(n) + 1+o(1)
n2 BB , where J(n
2)
2
B is the matrix whose entries have the form
B(ab, ij) = Aia Aib Aja Ajb .
The matrix B is actually equal to the matrix J4,1 from [DM15a], and by Lemma A.3 has kBk . n
with probability 1−O(n−5 ). We can thus conclude that with probability 1−O(n−5 ), kM − n1 J(n) k ≤
1+o(1)
2
n2 kBk
≤ Õ(1), and so xT M x ≤
which gives the desired result.
1+o(1)
n hx,
2
1(n) i2 +xT (M −n−1J)x ≤ O(n)·kΠ0 xk2 + Õ(1)·kxk2 ,
2
Acknowledgements
We thank Satish Rao for many helpful conversations.
We also greatfully acknowledge the comments of anonymous reviewers in helping us improve
the manuscript.
References
[AKS98]
Noga Alon, Michael Krivelevich, and Benny Sudakov, Finding a large hidden clique in
a random graph, Random Struct. Algorithms 13 (1998), no. 3-4, 457–466. 1, 2
[Bar14]
Boaz Barak, Sum of squares: upper bounds, lower bounds, and open questions (lecture
notes, fall 2014), 2014. 2, 4
[BBH+ 12] Boaz Barak, Fernando G. S. L. Brandão, Aram Wettroth Harrow, Jonathan A. Kelner,
David Steurer, and Yuan Zhou, Hypercontractivity, sum-of-squares proofs, and their
applications, Proceedings of the 44th Symposium on Theory of Computing Conference,
STOC 2012, New York, NY, USA, May 19 - 22, 2012, 2012, pp. 307–326. 2
[BE76]
Bella Bollobas and Paul Erdös, Cliques in random graphs, Mathematical Proceedings of
the Cambridge Philosophical Society 80 (1976), 419–427. 1
[BV09]
S. Charles Brubaker and Santosh Vempala, Random tensors and planted cliques, Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques,
12th International Workshop, APPROX 2009, and 13th International Workshop, RANDOM 2009, Berkeley, CA, USA, August 21-23, 2009. Proceedings (Irit Dinur, Klaus
Jansen, Joseph Naor, and José D. P. Rolim, eds.), Lecture Notes in Computer Science,
vol. 5687, Springer, 2009, pp. 406–419. 1
[DM15a]
Yash Deshpande and Andrea Montanari, Improved sum-of-squares lower bounds for hidden clique and hidden submatrix problems, Proceedings of The 28th Conference on Learning Theory, COLT 2015, Paris, France, July 3-6, 2015 (Peter Grünwald, Elad Hazan,
and Satyen Kale, eds.), JMLR Proceedings, vol. 40, JMLR.org, 2015, pp. 523–562. 1, 2,
3, 4, 5, 6, 7, 10, 12, 16, 21, 23, 24
[DM15b]
, Improved sum-of-squares lower bounds for hidden clique and hidden submatrix
problems, CoRR abs/1502.06590 (2015). 23, 24
21
[FGR+ 12] Vitaly Feldman, Elena Grigorescu, Lev Reyzin, Santosh Vempala, and Ying Xiao, Statistical algorithms and a lower bound for planted clique, Electronic Colloquium on Computational Complexity (ECCC) 19 (2012), 64. 1
[FK00]
Uriel Feige and Robert Krauthgamer, Finding and certifying a large hidden clique in a
semirandom graph, Random Struct. Algorithms 16 (2000), no. 2, 195–208. 1, 2
[FK03]
, The probable value of the lovasz-schrijver relaxations for maximum independent
set, SIAM Journal on Computing 32 (2003), 2003. 1, 2
[FK08]
Alan M. Frieze and Ravi Kannan, A new approach to the planted clique problem, IARCS
Annual Conference on Foundations of Software Technology and Theoretical Computer
Science, FSTTCS 2008, December 9-11, 2008, Bangalore, India, 2008, pp. 187–198. 1
[GM75]
Geoffrey R. Grimmett and Colin J. H. McDiarmid, On colouring random graphs, Mathematical Proceedings of the Cambridge Philosophical Society 77 (1975), 313–324. 1
[Hås96]
Johan Håstad, Clique is hard to approximate within n1-epsilon , 37th Annual Symposium
on Foundations of Computer Science, FOCS ’96, Burlington, Vermont, USA, 14-16
October, 1996, 1996, pp. 627–636. 1
[HKP15]
Samuel B. Hopkins, Pravesh Kothari, and Aaron Potechin, SoS and planted clique: Tight
analysis of MPW moments at all degrees and an optimal lower bound at degree four. 0,
2
[HM13]
Aram Wettroth Harrow and Ashley Montanaro, Testing product states, quantum merlinarthur games and tensor optimization, J. ACM 60 (2013), no. 1, 3. 1
[Jer92]
Mark Jerrum, Large cliques elude the metropolis process., Random Struct. Algorithms
3 (1992), no. 4, 347–360. 1
[Kar76]
Richard Karp, The probabilistic analysis of some combinatorial search algorithms, Algorithms and Complexity: New Directions and Recent Results (1976), 1–19. 1
[Kho01]
Subhash Khot, Improved inaproximability results for maxclique, chromatic number and
approximate graph coloring, 42nd Annual Symposium on Foundations of Computer Science, FOCS 2001, 14-17 October 2001, Las Vegas, Nevada, USA, 2001, pp. 600–609.
1
[KS09]
Subhash Khot and Rishi Saket, SDP integrality gaps with local ell 1-embeddability, 50th
Annual IEEE Symposium on Foundations of Computer Science, FOCS 2009, October
25-27, 2009, Atlanta, Georgia, USA, 2009, pp. 565–574. 2
[Mat76]
David Matula, The largest clique size in a random graph, Tech. report, Southern
Methodist University, Dallas, 1976. 1
[MPW15] Raghu Meka, Aaron Potechin, and Avi Wigderson, Sum-of-squares lower bounds for
planted clique, Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory
of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015, 2015, pp. 87–96. 1,
2, 4, 5
[RS09]
Prasad Raghavendra and David Steurer, Integrality gaps for strong SDP relaxations of
UNIQUE GAMES, 50th Annual IEEE Symposium on Foundations of Computer Science,
FOCS 2009, October 25-27, 2009, Atlanta, Georgia, USA, 2009, pp. 575–585. 2
22
A
Matrix Norm Bounds from Deshpande and Montanari
In this appendix, we give for completeness a list of the bounds proven by Deshpande and Montanari
[DM15a] that were not included in the body above out of space or expository considerations.
Definition A.1.
Let A = {a, b} ⊂ [n] be disjoint from B = {c, d} ⊂ [n]. For η ∈ {1, . . . , 4} and
4
′
for ν(η) ∈ [ η ] we define the matrices J˜η,ν(η)
as follows:
′ (A, B) = A
J˜1,1
ac
′ (A, B) = A
J˜1,2
ad
′ (A, B) = A
J˜1,3
bc
′
˜
J (A, B) = Abd
1,4
′ (A, B) = A A
J˜2,1
ac bd
′ (A, B) = A A
J˜2,2
ac bc
′ (A, B) = A A
J˜2,3
ac ad
′
˜
J2,4 (A, B) = Aad Abd
′ (A, B) = A A
J˜2,5
bc bd
J˜′ (A, B) = Aad Abc
′ (A, B) = A A A
˜′
J˜3,1
ac ad bc J4,1 (A, B) = Aac Aad Abc Abd
′
J˜3,2 (A, B) = Aac Abc Abd
′ (A, B) = A A A
J˜3,3
ac ad bd
′
˜
J (A, B) = Aad Abc Abd
3,4
2,6
n
n
And further, letting P : R( 2 )×( 2 ) to be the matrix projector such that
(
MA,B |A ∪ B| = 4
(PM )A,B =
0
otherwise ,
def 2η
16 α4
′
we define Jη,nu
= PJη,ν , and finally we define Jη,ν =
P4 P(η4)
so that Q = η=1 ν=1 Jη,ν .
′
· Jη,ν
(as in Deshpande and Montanari),
Notice that since we have defined Aii = 0 and since |{a, b}| = |{c, d}| = 2, we have J˜4,1 = J4,1 .
For some of the terms, the J˜ is never considered; however for some terms it is cleaner to bound
the spectral norm of J˜ in the subspace V2 , and so Deshpande and Montanari provide trace power
method bounds on the difference in norm:
Lemma A.2
(Lemma 4.26 in [DM15b]). With probability at least 1 − 6n−5 , for each η ≤ 2 and for
4
each ν ≤ η ,
kJη,ν − J˜η,ν k . α4 n̄.
Deshpande and Montanari use the trace power method to bound the norm of Q by bounding
the norms of the Jη,ν individually. Some of the Jη,ν matrices have Wigner-like behavior.
Lemma A.3 (Lemmas 4.21, 4.22 in [DM15b]). With probability 1 − O(n−5 ), we have that for each
(η, ν) ∈ {(2, 1), (2, 6), (3, ·), (4, 1)},
kJη,ν k . α4 · n.
A select few of the Jη,ν have larger eigenvalues.
Lemma A.4 (Lemmas 4.23, 4.24 in [DM15b]). With probability 1 − O(n−4 ), we have that for each
(η, ν) ∈ {(1, ·), (2, 2), (2, 3), (2, 4), (2, 5)},
kJη,ν k . α4 · n3/2 .
We also give a short proof of an observation of Deshpande and Montanari, which states that
some of the Jη,ν vanish when projected to V2 :
23
Observation A.5 (Lemmas 4.23, 4.24 in [DM15a]). Let Π2 be the projector to V2 . Then always,
!
4
X
= 0, and similarly,
kΠ2 (J˜2,3 + J˜2,5 )k = 0.
Π2
J˜1,ν
ν=1
Proof. The proof follows from noting
Pthat therange of both of these sums of Jη,ν is in V1 . Consider
n
def
4
v. We will look at the entry of v ′ indexed by the
some vector v ∈ R( 2 ) ; let v ′ =
J
ν=1
1,ν
disjoint pair A = {a, b}. By definition of the J1,ν , we have that
X
′
vA =
(Aa,c + Aa,d ) + (Ab,c + Ab,d vc,d
c,d∈[n]
X
X
= (Aa,c + Aa,d )vc,d + (Ab,c + Ab,d )vc,d ,
c,d
c,d
and so by the characterization of V1 from Proposition 2.3 the vector v ′ ∈ V1 . The conclusion follows.
A similar proof holds for the matrix J˜2,3 + J˜2,5 .
Finally, we use a bound on the norm of the matrix K, which is the difference of H2,2 and the
non-multilinear entries.
Lemma A.6 (Lemma 4.25 in [DM15b]). Let K be the restriction of H2,2 − E[H2,2 ] to entries
indexed by sets of size at most 3. With probability at least 1 − n−5 ),
kKk ≤ Õ(α3 n1/2 ).
We also require bounds on the matrices used in the Schur complement steps. The bounds of
Deshpande and Montanari suffice for us, since we do not modify moments of order less than 4.
Lemma A.7 (Consequence of Proposition 4.19 in [DM15b]). Define Qn ∈ Rn,n to be the orthogonal
projection to the space spanned by ~1. Suppose that α satisfies α1 −α2 ≥ Ω(α2 n−1/2 ) and α2 −2α21 ≥ 0,
α1 ≥ 0. Then with probability at least 1 − n−5 ,
H1,1 0
−1
H1,1
A.1
2 ⊥
1
Qn +
Q .
2
α1 n
n(α2 p − α1 p)
Additional Proofs
We prove Lemma 2.5, which follows almost immediately from the bounds of [DM15a].
Proof of Lemma 2.5. Using the matrices from Definition A.1 and Observation A.5, we have that
!
4
X
J3,ν + J4,1 ,
Π2 Q = Π2 (J˜2,4 + J˜2,2 ) + Π2 J2,4 − J˜2,4 + J2,2 − J˜2,2 + Π2
ν=1
Π2 ΠW Q = Π2 ΠW (J˜2,4 + J˜2,2 ) + Π2 ΠW
J2,4 − J˜2,4 + J2,2 − J˜2,2 + Π2 ΠW
24
4
X
ν=1
J3,ν + J4,1
!
,
where we have used the fact that the columns of J2,4 and J2,2 lie in W . We thus have
Π2 Q − Π2 ΠW Q = Π2 (I − ΠW ) J2,2 − J˜2,2 + J2,4 − J˜2,4 +
4
X
ν=1
J3,ν + J4,1
!
,
And by the bounds kJ3,· k ≤ α4 ·n and kJ4,1 k ≤ α4 ·n from Lemma A.3 and the bounds kJ2,4 − J˜2,4 k ≤
α4 n and kJ2,2 − J˜2,2 k ≤ α4 n from Lemma A.2, and because Π2 (I − ΠW ) is a projection, the
conclusion follows.
Now, we prove that the trace power method works, for completeness.
Proof of Lemma 4.1. The proof follows from an application of Markov’s inequality. We have that
for even k,
P[kM k ≥ t] = P[kM k k ≥ tk ]
≤ P[Tr(M k ) ≥ tk ]
1
≤ k E[Tr(M k )]
t
k
γk
1p
nαk+β ,
≤ k πγk
t
e
where we have applied Stirling’s
approximation in the last step. Choosing k = O(log n) and
t = O η −1/k · γ · log n · nα completes the proof.
25
| 8cs.DS
|
Microwave Tomographic Imaging of Cerebrovascular
Accidents by Using High-Performance Computing
arXiv:1607.02573v1 [cs.CE] 9 Jul 2016
P.-H. Tourniera,b , I. Aliferisc , M. Bonazzolid , M. de Buhane , M. Darbasf ,
V. Doleand,g , F. Hechta,b , P. Joliveth , I. El Kanfoudc , C. Migliaccioc ,
F. Natafa,b , C. Pichotc , S. Semenovi
a
Laboratoire Jacques-Louis Lions, UMR CNRS 7598, Sorbonne Universités, UPMC,
Paris, France
b
INRIA-Paris, EPC Alpines, Paris, France
c
Laboratoire LEAT, UMR CNRS 7248, Université Nice Sophia Antipolis, Sophia
Antipolis, France
d
Laboratoire J.A. Dieudonné, UMR CNRS 7351, Université Nice Sophia Antipolis, Nice,
France
e
MAP5, UMR CNRS 8145, Université Paris-Descartes, Sorbonne Paris Cité, France
f
LAMFA, UMR CNRS 7352, Université de Picardie Jules Verne, Amiens, France
g
Dept of Maths and Stats, University of Strathclyde, Glasgow, UK
h
IRIT, UMR CNRS 5505, Toulouse, France
i
EMTensor GmbH, TechGate, 1220 Vienna, Austria
Abstract
The motivation of this work is the detection of cerebrovascular accidents by
microwave tomographic imaging. This requires the solution of an inverse
problem relying on a minimization algorithm (for example, gradient-based),
where successive iterations consist in repeated solutions of a direct problem. The reconstruction algorithm is extremely computationally intensive
and makes use of efficient parallel algorithms and high-performance computing. The feasibility of this type of imaging is conditioned on one hand
by an accurate reconstruction of the material properties of the propagation
medium and on the other hand by a considerable reduction in simulation
time. Fulfilling these two requirements will enable a very rapid and accurate
diagnosis. From the mathematical and numerical point of view, this means
solving Maxwell’s equations in time-harmonic regime by appropriate domain
decomposition methods, which are naturally adapted to parallel architectures.
Keywords: inverse problem, scalable preconditioners, Maxwell’s equations,
microwave imaging
Preprint submitted to Parallel Computing
July 12, 2016
1. Introduction
A stroke, also known as cerebrovascular accident, is a disturbance in the
blood supply to the brain caused by a blocked or burst blood vessel. As a
consequence, cerebral tissues are deprived of oxygen and nutrients. This results in a rapid loss of brain functions and often death. Strokes are classified
into two major categories: ischemic (85% of strokes) and hemorrhagic (15%
of strokes). During an acute ischemic stroke, the blood supply to a part of
the brain is interrupted by thrombosis - the formation of a blood clot in a
blood vessel - or by an embolism elsewhere in the body. A hemorrhagic stroke
occurs when a blood vessel bursts inside the brain, increasing pressure in the
brain and injuring brain cells. The two types of strokes result in opposite
variations of the dielectric properties of the affected tissues. How quickly
one can detect and characterize the stroke is of fundamental importance for
the survival of the patient. The quicker the treatment is, the more reversible
the damage and the better the chances of recovery are. Moreover, the treatment of ischemic stroke consists in thinning the blood (anticoagulants) and
can be fatal if the stroke is hemorrhagic. Therefore, it is vital to make a
clear distinction between the two types of strokes before treating the patient.
Moreover, ideally one would want to monitor continuously the effect of the
treatment on the evolution of the stroke during the hospitalization. The two
most used imaging techniques for strokes diagnosis are MRI (magnetic resonance imaging) and CT scan (computerized tomography scan). One of their
downsides is that the travel time from the patient’s home to the hospital is
lost. Moreover, the cost and the lack of portability of MRI and the harmful
character of CT scan make them unsuitable for a continuous monitoring at
the hospital during treatment.
This has motivated the study of an additional technique: microwave tomography. The measurement system is lightweight and thus transportable.
The acquisition of the data is harmless and faster than CT or MRI. Hence,
this imaging modality could be used by an emergency unit and for monitoring at the hospital. At frequencies of the order of 1 GHz, the tissues are
well differentiated and can be imaged on the basis of their dielectric properties. After the first works on microwave imaging in 1982 by Lin and Clarke
[1], other works followed, but almost always on synthetic simplified models
[2]. New devices are currently designed and studied by EMTensor GmbH
2
Figure 1: Left: Operating principle of the diagnosis apparatus. Middle:
imaging chamber prototype of EMTensor, by courtesy of EMTensor company.
Right: the corresponding simulation domain.
(Vienna, Austria) [3].
The purpose of this work is to solve in parallel the inverse problem associated with the time-harmonic Maxwell’s equations which model electromagnetic waves propagation. The dielectric properties of the brain tissues of
a patient yield the image that could be used for a rapid diagnosis of brain
strokes. Simulation results presented in this work have been obtained on the
imaging system prototype developed by EMTensor GmbH [3] (see Figure 1).
It is composed of 5 rings of 32 ceramic-loaded rectangular waveguides around
a metallic cylindrical chamber of diameter 28.5 cm and total height 28 cm.
The head of the patient is inserted into the chamber as shown in Figure 1
(left). The imaging chamber is filled with a matching solution and a membrane is used to isolate the head. Each antenna successively transmits a
signal at a fixed frequency, typically 1 GHz. The electromagnetic wave propagates inside the chamber and in the object to be imaged according to its
electromagnetic properties. The retrieved data then consist in the scattering
parameters measured by the 160 receiving antennas, which are used as input
for the inverse problem. These raw data can be wirelessly transferred to a remote computing center. The HPC machine will then compute the 3D images
of the patient’s brain. Once formed, these images can be quickly transmitted
from the computing center to the hospital, see Figure 2.
The paper is organized as follows. In Section 2 the direct problem and the
time-harmonic Maxwell’s equations in curl-curl form with suitable boundary
conditions are introduced. In Section 3, we briefly describe the discretization
method with edge finite elements. Section 4 is devoted to the introduction
of the domain decomposition preconditioner. In Section 5 we explain how
to compute the scattering coefficients. We also compare measurement data
3
Figure 2: Design concept of the diagnosis technology, by courtesy of EMTensor company.
obtained by EMTensor with the coefficients computed by the simulation.
We introduce the inverse problem in Section 6. Section 7 is dedicated to
numerical results. We first perform a strong scaling analysis to show the
effectiveness of the domain decomposition method. Then, we present results
obtained by solving the inverse problem in a realistic configuration, with noisy
synthetic data generated using a numerical brain model with a simulated
hemorrhagic stroke. Finally, we conclude this paper in Section 8 and give
directions for future research.
2. The direct problem
Let the domain Ω ⊂ R3 represent the imaging chamber (see Figure 1,
right). We consider in Ω a heterogeneous non-magnetic dissipative linear
isotropic dielectric medium, of dielectric permittivity ε(x) > 0 and electrical
conductivity σ(x) ≥ 0. For each transmitting antenna j = 1, . . . , N emitting
a time periodic signal at angular frequency ω, the complex amplitude Ej (x) of
the associated electric field Ej (x, t) = <(Ej (x)eiωt ) is solution to the following
second order time-harmonic Maxwell’s equation:
∇ × (∇ × Ej ) − µ0 (ω 2 ε − iωσ)Ej = 0
in Ω,
(1)
where µ0 is the permeability of free space. Note that the coefficient
κ =
µ0 (ω 2 ε − iωσ) in the equation can be written as κ = ω 2 µ0 ε − i ωσ , and in the
next sections we will consider the relative complex permittivity εr given by
the relation εr ε0 = ε − i ωσ , where ε0 is the permittivity of free space. Let n
be the unit outward normal to ∂Ω. Equation (1) is equipped with perfectly
conducting boundary conditions on the metallic walls Γm :
Ej × n = 0
4
on Γm ,
and with impedance boundary conditions on the outer section of the transmitting waveguide j and of the receiving waveguides i = 1, . . . , N , i 6= j
(see [4]):
(∇ × Ej ) × n + iβn × (Ej × n) = gj
(∇ × Ej ) × n + iβn × (Ej × n) = 0
on Γj ,
on Γi , i 6= j.
(2)
(3)
Here β is the propagation wavenumber along the waveguide, corresponding
to the propagation of the TE10 fundamental mode. Equation (2) imposes an
incident wave which corresponds to the excitation of the fundamental mode
E0j of the j-th waveguide, with gj = (∇ × E0j ) × n + iβn × (E0j × n). On
the other hand equation (3) corresponds to a first order absorbing boundary
condition of Silver–Müller approximating a transparent boundary condition
on the outer section of the receiving waveguides i = 1, . . . , N , i 6= j. The
bottom of the chamber is metallic, and we impose an impedance boundary
condition on the top of the chamber. We end up with the following boundary
value problem for each transmitting antenna j = 1, . . . , N : find Ej such that
∇ × (∇ × Ej ) − µ0 (ω 2 ε − iωσ)Ej = 0
in Ω,
Ej × n = 0
on Γm ,
(4)
(∇ × Ej ) × n + iβn × (Ej × n) = gj on Γj ,
(∇ × Ej ) × n + iβn × (Ej × n) = 0
on Γi , i 6= j.
Now, let V = {v ∈ H(curl, Ω), v × n = 0 on Γm }, where H(curl, Ω) = {v ∈
L2 (Ω)3 , ∇ × v ∈ L2 (Ω)3 } is the space of square integrable functions whose
curl is also square integrable. For each transmitting antenna j = 1, . . . , N ,
the variational form of problem (4) reads: find Ej ∈ V such that
Z
(∇ × Ej ) · (∇ × v) − µ0 (ω 2 ε − iωσ)Ej · v
Z Ω
Z
(5)
iβ(Ej × n) · (v × n) =
gj · v ∀v ∈ V.
+ S
N
i=1
Γi
Γj
3. Edge finite elements
Nédélec edge elements [5] are finite elements particularly suited for the
approximation of electromagnetic fields. Indeed, given a tetrahedral mesh T
of the computational domain Ω, the finite dimensional subspace Vh generated
5
by Nédélec basis functions is included in H(curl, Ω), since their tangential
component across faces shared by adjacent tetrahedra of T is continuous.
They thus match the continuity properties of the electric field. Nédélec elements are called edge elements because the basis functions are associated
with the edges of the mesh T . More precisely, for a tetrahedron T ∈ T , the
local basis functions are associated with the oriented edges e = {ni , nj } of T
as follows
we = λi ∇λj − λj ∇λi ,
where the λ` are the barycentric coordinates of a point with respect to the
node n` of T . Note that the polynomial degree of we is 1 since the barycentric
coordinates λ` are polynomials of degree 1 and their gradients are constant.
As these basis functions are vector functions, we only need one set of unknowns to approximate all the components of the field, and not three sets
of unknowns, one for each component of the field as is instead required for
usual nodal (scalar) finite elements.
The finite element discretization of the variational problem is obtained
by taking test functions v ∈ Vh , the edge finite element space on the mesh
T , and by looking for a solution Eh ∈ Vh in the same space: find Eh ∈ Vh
such that
Z
(∇ × Eh ) · (∇ × v) − µ0 (ω 2 ε − iωσ)Eh · v
Ω
Z
Z
(6)
+ S
iβ(Eh × n) · (v × n) =
gj · v ∀v ∈ Vh .
N
i=1
Γi
Γj
Locally,
over each tetrahedron T , we write the discretized field as Eh =
P
e
e∈T ce w , a linear combination with coefficients ce of the basis functions
associated with the edges e of T , and the coefficients ce will be the unknowns
of the resulting linear system. For edge finite elements (of degree 1) these
coefficients can be interpreted as the circulations of Eh along the edges of
the tetrahedra:
Z
1
Eh · te ,
ce =
|e| e
where te is the tangent vector to the edge e of length |e|, the length of e.
This is a consequence of the fact that the basis functions are in duality with
the degrees of freedom given by the circulations along the edges, that is:
(
Z
1 if e = e0 ,
1
0
we · te =
|e| e
0 if e 6= e0 .
6
4. Domain decomposition preconditioning
The finite element discretization (6) of the variational problem (5) produces linear systems
Auj = bj
for each transmitting antenna j. However, the matrix A can be ill-conditioned.
This, combined with the fact that the underlying PDE is indefinite, highlights
the need for a robust and efficient preconditioner. Here we employ domain
decomposition preconditioners, which are extensively described in [6], as they
are naturally suited to parallel computing. Our domain decomposition preconditioner is presented in the following.
Let T be the mesh of the computational domain Ω. First, T is partitioned
into NS non-overlapping meshes {Ti }16i6NS using standard graph partitioners
such as SCOTCH [7] or METIS [8]. If δ is a positive integer, the overlapping
decomposition {Tiδ }16i6NS is defined recursively as follows: Tiδ is obtained
by including all tetrahedra of Tiδ−1 plus all adjacent tetrahedra of Tiδ−1 ; for
δ = 0, Tiδ = Ti . Note that the number of layers in the overlap is then 2δ. Let
Vh be the edge finite element space defined on T , and {Viδ }16i6NS the local
edge finite element spaces defined on {Tiδ }16i6NS , δ > 0. Now consider the
restrictions {Ri }16i6NS from Vh to {Viδ }16i6NS , and a local partition of unity
{Di }16i6NS such that
NS
X
RiT Di Ri = In×n .
(7)
i=1
Algebraically speaking, if n is the global number of unknowns and {ni }16i6NS
are the numbers of unknowns for each local finite element space, then Ri is
a Boolean matrix of size ni × n, and Di is a diagonal matrix of size ni × ni ,
for all 1 6 i 6 NS . Note that RiT , the transpose of Ri , is a n × ni matrix
that gives the extension by 0 from Viδ to Vh .
Using these matrices, one can define the following one-level preconditioner, called Optimized Restricted Additive Schwarz preconditioner (ORAS)
[9]:
NS
X
−1
MORAS
=
RiT Di Bi−1 Ri ,
(8)
i=1
where {Bi }16i6NS are local operators corresponding to the subproblems with
impedance boundary conditions (∇ × E) × n + ikn × (E × n), where k =
√
ω µ0 εr ε0 is the wavenumber. These boundary conditions were first used as
7
transmission conditions at the interfaces between subdomains in [10]. The
local matrices {Bi }16i6NS of the ORAS preconditioner make use of more efficient transmission boundary conditions than the submatrices {Ri ARiT }16i6NS
of the original Restricted Additive Schwarz (RAS) preconditioner [11]. It is
important to note that when a direct solver is used to compute the action
of Bi−1 on multiple vectors, it can be done in a single forward elimination
and backward substitution. More details on the solution of linear systems
with multiple right-hand sides are given in Section 7. The preconditioner
−1
MORAS
(8) is naturally parallel since its assembly requires the concurrent
factorization of each {Bi }16i6NS , which are typically stored locally on different processes in a distributed computing context. Likewise, applying (8)
to a distributed vector only requires peer-to-peer communications between
neighboring subdomains, and a local forward elimination and backward substitution. See chapter 8 of [6] for a more detailed analysis.
4.1. Partition of unity
The construction of the partition of unity is intricate, especially for Nédélec
edge finite elements.
The starting point is the construction of partition of unity functions
{χi }16i6NS for the classical P1 linear nodal finite element, whose degrees
of freedom are the values at the nodes of the mesh. First of all, we define for
i = 1, . . . , NS the function χ
ei as the continuous piecewise linear function on
T , with support contained in Tiδ , such that
(
1 at all nodes of Ti0 ,
χ
ei =
0 at all nodes of Tiδ \ Ti0 .
The function χi can then be defined as the continuous piecewise linear function on T , with support contained in Tiδ , such that its (discrete) value for
each degree of freedom is evaluated by:
χi =
χ
ei
.
NS
X
χ
ej
(9)
j=1
P S
Thus, we have N
i=1 χi = 1 both at the discrete and continuous level. Remark
that if δ > 1, not only the function χi but also its derivative is equal to zero on
the border of Tiδ . This is essential for a good convergence if Robin boundary
8
conditions are chosen as transmission conditions at the interfaces between
subdomains. Indeed, if this property is satisfied, the continuous version of
the ORAS algorithm is equivalent to P. L. Lions’ algorithm (see [9] and
[6] §2.3.2). Note that in the practical implementation, the functions χ
ei and
ej in (9)
χi are constructed locally on Tiδ , the relevant contribution of the χ
being on Tjδ ∩ Tiδ . This removes all dependency on the global mesh T , which
could be otherwise problematic at large scales.
Now, the degrees of freedom of Nédélec finite elements are associated with
the edges of the mesh. For these finite elements, we can build a geometric
partition of unity based on the support of the degrees of freedom (the edges of
the mesh): the entries of the diagonal matrices Di , i = 1, . . . , NS are obtained
for each degree of freedom by interpolating the piecewise linear function χi at
the midpoint of the corresponding
edge. The partition of unity property (7)
PNS
is then satisfied since i=1 χi = 1.
4.2. Software stack
All operators related to the domain decomposition method can be easily generated using finite element Domain-Specific Languages (DSL). Here
we use FreeFem++ [12] (http://www.freefem.org/ff++/) since it has already been proven that it can enable large-scale simulations using overlapping
Schwarz methods [13] when used in combination with the library HPDDM
[14] (High-Performance unified framework for Domain Decomposition Methods, https://github.com/hpddm/hpddm). HPDDM implements several domain decomposition methods such as RAS, ORAS, FETI, and BNN. It uses
multiple levels of parallelism: communication between subdomains is based
on the Message Passing Interface (MPI), and computations in the subdomains can be executed on several threads by calling optimized BLAS libraries
(such as Intel MKL), or shared-memory direct solvers like PARDISO. Domain
decomposition methods naturally offer good parallel properties on distributed
architectures. The computational domain is decomposed into subdomains in
which concurrent computations are performed. The coupling between subdomains requires communications between computing nodes via messages. The
strong scalability of the ORAS preconditioner as implemented in HPDDM
for the direct problem presented in Section 2 will be assessed in Section 7.
9
5. Computing the scattering parameters
In order to compute the numerical counterparts of the reflection and
transmission coefficients obtained by the measurement apparatus of the imaging chamber shown in Figure 1, we use the following formula, which is appropriate in the case of open-ended waveguides:
Z
Ej · E0i
, i, j = 1, . . . , N,
(10)
Sij = ZΓi
0 2
|Ei |
Γi
where E0i is the TE10 fundamental mode of the i-th receiving waveguide and
Ej is the solution of the problem where the j-th waveguide transmits the
signal (Ej denotes the complex conjugate of Ej ). The Sij with i 6= j are the
transmission coefficients, and the Sjj are the reflection coefficients. They are
gathered in the scattering matrix, also called S-matrix.
Here we compare the coefficients computed from the simulation with a
set of measurements obtained by EMTensor. For this test case, the imaging chamber was filled with a homogeneous matching solution. The electric
permittivity ε of the matching solution is chosen by EMTensor in order to
minimize contrasts with the ceramic-loaded waveguides and with the different brain tissues. The choice of the conductivity σ of the matching solution is a compromise between the minimization of reflection artifacts from
metallic boundaries and the desire to have best possible signal-to-noise ratio.
Here the relative complex permittivity of the matching solution at frequency
f = 1 GHz is εgel
r = 44 − 20i. The relative complex permittivity inside the
ceramic-loaded waveguides is εcer
r = 59.
The set of experimental data at hand given by EMTensor consists in
transmission coefficients for transmitting antennas in the second ring from
the top. Figure 3 shows the normalized magnitude (dB) and phase (degree)
of the complex coefficients Sij corresponding to a transmitting antenna in
the second ring from the top and to the 31 receiving antennas in the middle ring (note that measured coefficients are available only for 17 receiving
antennas). The magnitude in dB is calculated as 20 log10 (|Sij |). The normalization is done by dividing every transmission coefficient by the transmission
coefficient corresponding to the receiving antenna directly opposite to the
transmitting antenna, which is thus set to 1. Since we normalize with respect to the coefficient having the lowest expected magnitude, the magnitude
10
70
simulation
measurements
60
magnitude (dB)
50
40
30
20
10
0
5
10
15
20
receiver number
25
30
simulation
measurements
0
phase (degree)
-500
-1000
-1500
-2000
5
10
15
20
receiver number
25
30
Figure 3: The normalized magnitude (top) and phase (bottom) of the transmission coefficients computed with the simulation and measured experimentally.
of the transmission coefficients displayed in Figure 3 is larger than 0 dB. We
can see that the transmission coefficients computed from the simulation are
in very good agreement with the measurements.
6. The inverse problem
The inverse problem that we consider consists in finding the unknown dielectric permittivity ε(x) and conductivity σ(x) in Ω, such that the solutions
Ej , j = 1, . . . , N of problem (4) lead to corresponding scattering parameters Sij (10) that coincide with the measured scattering parameters Sijmes ,
for i, j = 1, . . . , N . In the following, we present the inverse problem in the
continuous setting for clarity.
Let κ = µ0 (ω 2 ε − iωσ) be the unknown complex parameter of our inverse problem, and let us denote by Ej (κ) the solution of the direct problem
11
(4) with dielectric permittivity ε and conductivity σ. The corresponding
scattering parameters will be denoted by Sij (κ) for i, j = 1, . . . , N :
Z
Ej (κ) · E0i
, i, j = 1, . . . , N.
Sij (κ) = ΓiZ
0 2
|Ei |
Γi
The misfit of the parameter κ to the data can be defined through the following
functional:
Z
2
0
E
(κ)
·
E
N
N
N
N
j
i
1 XX
1 X X Γi
2
Z
J(κ) =
Sij (κ) − Sijmes =
− Sijmes .
2 j=1 i=1
2 j=1 i=1
|E0i |2
Γi
(11)
In a classical way, solving the inverse problem then consists in minimizing
the functional J with respect to the parameter κ. Computing the differential
of J in a given arbitrary direction δκ yields
Z
0
δE
(κ)
·
E
N X
N
i
X
Γi j
mes
, δκ ∈ C,
Z
DJ(κ, δκ) =
< Sij (κ) − Sij
0
2
j=1 i=1
|Ei |
Γi
where δEj (κ) is the solution of the following linearized problem:
∇ × (∇ × δEj ) − κδEj = δκEj in Ω,
δEj × n = 0
on Γm ,
(∇ × δE ) × n + iβn × (δE × n) = 0
on Γi , i = 1, . . . , N.
j
j
(12)
We now use the adjoint approach in order to simplify the expression of
DJ. This will allow us to compute the gradient efficiently after discretization,
with a number of computations independent of the size of the parameter
space. Considering the variational formulation of problem (12) with a test
12
function F and integrating by parts, we get
Z
Z
δκEj · F =
(∇ × (∇ × δEj ) − κδEj ) · F
Ω
Z
ZΩ
(∇ × (∇ × F) − κF) · δEj −
((∇ × δEj ) × n) · F
=
Ω
∂Ω
Z
+
((∇ × F) × n) · δEj
∂Ω
Z
(∇ × (∇ × F) − κF) · δEj +
=
Ω
N Z
X
i=1
Z
(∇ × δEj ) · (F × n) +
+
Γm
N Z
X
iβ(n × (F × n)) · δEj
Γi
((∇ × F) × n) · δEj .
Γi
i=1
Introducing the solution Fj (κ) of the following adjoint problem
∇ × (∇ × Fj ) − κFj = 0
in Ω,
Fj × n = 0
on Γm ,
mes
(Sij (κ) − Sij ) 0
Z
(∇
×
F
)
×
n
+
iβn
×
(F
×
n)
=
Ei on Γi , i = 1, . . . , N,
j
j
|E0i |2
Γi
(13)
we get
Z
Z
δκEj · Fj =
Ω
N
X
E0i · δEj
(Sij (κ) − Sijmes ) ΓZi
i=1
.
|E0i |2
Γi
Finally, the differential of J can be computed as
DJ(κ, δκ) =
N
X
Z
<
δκ Ej · Fj .
Ω
j=1
We can then compute the gradient to use in a gradient-based local optimization algorithm. The numerical results presented in Section 7 are obtained using a limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS)
algorithm. Note that every evaluation of J requires the solution of the state
problem (4) while the computation of the gradient requires the solution of (4)
13
as well as the solution of the adjoint problem (13). Moreover, the state and
adjoint problems use the same operator. Therefore, the computation of the
gradient only needs the assembly of one matrix and its associated domain
decomposition preconditioner.
Numerical results for the reconstruction of a hemorrhagic stroke from synthetic data are presented in the next section. The functional J considered in
the numerical results is slightly different from (11), as we add a normalization
term for each pair (i, j) as well as a Tikhonov regularizing term:
N
N
1 X X Sij (κ) − Sijmes
J(κ) =
2
2 j=1 i=1
Sijempty
2
α
+
2
Z
|∇κ|2 ,
(14)
Ω
where Sijempty refers to the coefficients computed from the simulation with
the empty chamber, that is the chamber filled only with the homogeneous
matching solution as described in the previous section, with no object inside.
In this way, the contribution of each pair (i, j) in the misfit functional is
normalized and does not depend on the amplitude of the coefficient, which
can vary greatly between pairs (i, j) as displayed in Figure 3. The Tikhonov
regularizing term aims at reducing the effects of noise in the data. For
now, the regularization parameter α is chosen empirically so as to obtain a
visually good compromise between reducing the effects of noise and keeping
the reconstructed image pertinent. All calculations carried out in this section
can be accommodated in a straightforward manner to definition (14) of the
functional.
As is usually the case with most medical imaging techniques, the reconstruction is done layer by layer. For the imaging chamber of EMTensor that
we study in this paper, one layer corresponds to one of the five rings of 32
antennas. This allows us to exhibit another level of parallelism, by solving
an inverse problem independently for each of the five rings in parallel. More
precisely, each of these inverse problems is solved in a domain truncated
around the corresponding ring of antennas, containing at most two other
rings (one ring above and one ring below). We impose absorbing boundary
conditions on the artificial boundaries of the truncated computational domain. For each inverse problem, only the coefficients Sij with transmitting
antennas j in the corresponding ring are taken into account: we consider 32
antennas as transmitters and at most 96 antennas as receivers.
14
Time to solution (in seconds)
500
(43)
Setup
Solve
Linear speedup
(53)
200
(64)
50
(81)
10
256
512
1 024
# of subdomains
2 048
Figure 4: Strong scaling experiment. Colors indicate the fraction of the
total time spent in the setup and solution phases. The number of GMRES
iterations is reported in parentheses.
7. Numerical results
Results in this paper were obtained on Curie, a system composed of
5,040 nodes made of two eight-core Intel Sandy Bridge processors clocked
at 2.7 GHz. The interconnect is an InfiniBand QDR full fat tree and the
MPI implementation used was BullxMPI version 1.2.8.4. Intel compilers and
Math Kernel Library in their version 16.0.2.181 were used for all binaries and
shared libraries, and as the linear algebra backend for dense computations.
One-level preconditioners such as (8) assembled by HPDDM require the use
of a sparse direct solver. In the following experiments, we have been using
either PARDISO [15] from Intel MKL or MUMPS [16]. All linear systems
resulting from the edge finite elements discretization are solved by GMRES
right-preconditioned with ORAS (8) as implemented in HPDDM. The GMRES algorithm is stopped once the unpreconditioned relative residual is lower
than 10−8 . First, we perform a strong scaling analysis in order to assess the
efficiency of our preconditioner. We then solve the inverse problem in a realistic configuration, with noisy synthetic data generated using a numerical
brain model with a simulated hemorrhagic stroke.
7.1. Scaling analysis
Using the domain decomposition preconditioner (8), we solve the direct
problem corresponding to the setting of Section 5 where the chamber is filled
15
NS
256
512
1,024
2,048
Setup
293.36
95.11
35.13
25.89
Solve
73.06
36.92
20.55
12.77
# of iterations
43
53
64
81
Speedup
1
2.8
6.6
9.5
Figure 5: Strong scaling experiment. Timings (in seconds) of the setup and
solution phases.
with a homogeneous matching solution. We consider a right-hand side corresponding to a transmitting antenna in the second ring from the top. Given
a fine mesh of the domain composed of 82 million tetrahedra, we increase
the number of MPI processes to solve the linear system of 96 million doubleprecision complex unknowns yielded by the discretization of Maxwell’s equation using edge elements. The global unstructured mesh is partitioned using
SCOTCH [7] and the local solver is PARDISO from Intel MKL. We use one
subdomain and two OpenMP threads per MPI process. Results are reported
in Figure 5 and illustrated in Figure 4 with a plot of the time to solution
including both the setup and solution phases on 256 up to 2048 subdomains.
The setup time corresponds to the maximum time spent for the factorization of the local subproblem matrix Bi in (8) over all subdomains, while the
solution time corresponds to the time needed to solve the linear system with
GMRES. We are able to obtain very good speedups up to 4096 cores (2048
subdomains) on Curie, with a superlinear speedup of 9.5 between 256 and
2048 subdomains.
7.2. Reconstruction of a hemorrhagic stroke from synthetic data
In this subsection, we assess the feasibility of the microwave imaging technique presented in this paper for stroke detection and monitoring through a
numerical example in a realistic configuration. We use synthetic data corresponding to a numerical model of a virtual human head with a simulated
hemorrhagic stroke as input for the inverse problem. The numerical model of
the virtual head comes from CT and MRI tomographic images and consists
of a complex permittivity map of 362 × 434 × 362 data points. Figure 6 (left)
shows a sagittal section of the head. In the simulation, the head is immersed
in the imaging chamber as shown in Figure 6 (right). In order to simulate a
hemorrhagic stroke, a synthetic stroke is added in the form of an ellipsoid in
which the value of the complex permittivity εr has been increased. For this
16
Figure 6: Left: sagittal section of the brain. Right: numerical head immersed in the imaging chamber, with a simulated ellipsoid-shaped hemorrhagic stroke.
test case, the value of the permittivity in the ellipsoid is taken as the mean
value between the relative permittivity of the original healthy brain and the
relative permittivity of blood at frequency f = 1 GHz, εblood
= 68 − 44i.
r
The imaging chamber is filled with a matching solution. The relative permittivity of the matching solution is chosen by EMTensor as explained in
Section 5 and is equal to εgel
r = 44 − 20i at frequency f = 1 GHz. In the real
setting, a special membrane fitting the shape of the head is used in order to
isolate the head from the matching medium. We do not take this membrane
into account in this synthetic test case. The synthetic data are obtained by
solving the direct problem on a mesh composed of 17.6 million tetrahedra
(corresponding to approximately 20 points per wavelength) and consist in
the transmission and reflection coefficients Sij calculated from the simulated
electric field as in (10). We subsequently add noise to the real and imaginary
parts of the coefficients Sij (10% additive Gaussian white noise, with different
values for real and imaginary parts). The noisy data are then used as input
for the inverse problem. Furthermore, we assume no a priori knowledge on
the input data, and we set the initial guess for the inverse problem as the
homogeneous matching solution everywhere inside the chamber. We use a
piecewise linear approximation of the unknown parameter κ, defined on the
same mesh used to solve the state and adjoint problems. For the purpose of
parallel computations, the partitioning introduced by the domain decomposition method is also used to compute and store locally in each subdomain
every entity involved in the inverse problem, such as the parameter κ and
the gradient.
17
Figure 7: Top row: imaginary part of the exact permittivity used to produce
noisy data as input for the inverse problem during time evolution of a simulated hemorrhagic stroke (indicated by the black arrow). The size of the
ellipsoid is 3.9 cm×2.3 cm×2.3 cm and 7.7 cm×4.6 cm×4.6 cm in the middle
and right column respectively. Bottom row: corresponding reconstructions
obtained by taking into account only the first ring of transmitting antennas.
Figure 7 shows the imaginary part of the exact and reconstructed permittivity for three steps of the evolution of the hemorrhagic stroke, from the
healthy brain (left column) to the large stroke (right column). Increasing the
size of the ellipsoid in which the value of the permittivity is raised simulates
the evolution of the stroke. Each of the three reconstructions in Figure 7
corresponds to the solution of an inverse problem in the truncated domain
containing only the first two rings of antennas from the top, and where only
the coefficients Sij corresponding to transmitting antennas j in the first ring
are taken into account. Each reconstruction starts from an initial guess consisting of the homogeneous matching solution and is obtained after reaching
a convergence criterion of 10−2 for the value of the cost functional, which
takes around 30 iterations of the L-BFGS algorithm.
Figure 8 gathers the results of a strong scaling experiment which consists
18
Time in minutes
16
Linear speedup
8
4
2
1
0.5
64
128 256 512 1 0 2 0 4 0
24 48 96
# of MPI processes
Figure 8: Strong scaling experiment: total time needed to obtain the third
reconstructed image shown in Figure 7.
in solving the same inverse problem corresponding to the third reconstructed
image of Figure 7 for an increasing number of MPI processes. We report
the total computing time needed to obtain the reconstructed image. For this
experiment we use one subdomain and one OpenMP thread per MPI process.
The mesh of the computational domain is composed of 674 580 tetrahedra,
corresponding to approximately 10 points per wavelength.
Note that evaluating the functional or its gradient requires the solution of
a linear system with 32 right-hand sides, one right-hand side per transmitter.
This introduces a trivial level of parallelism since the solution corresponding
to each right-hand side can be computed independently. However when considering a finite number of available processors, there is a tradeoff between
the parallelism induced by the multiple right-hand sides and the parallelism
induced by the domain decomposition method. Additionally, we solve for
multiple right-hand sides simultaneously using a pseudo-block method implemented inside GMRES which consists in fusing the multiple arithmetic
operations corresponding to each right-hand side (matrix-vector products,
dot products), resulting in higher arithmetic intensity. The scaling behavior
of this pseudo-block algorithm with respect to the number of right-hand sides
is nonlinear, as is the scaling behavior of the domain decomposition method
with respect to the number of subdomains. Thus, for a given number of
processors, we find the optimal tradeoff between parallelizing with respect
to the number of subdomains or right-hand sides through trial and error.
19
For example, the best computing time for 2048 MPI processes is achieved
by using 8 domain decomposition communicators (i.e. 8 concurrent direct
solves) with 256 subdomains treating 4 right-hand sides each.
Figure 8 shows that we can generate an image with a total computing
time of less than 2 minutes (94 seconds) using 4096 cores of Curie. These
preliminary results are very encouraging as we are already able to achieve a
satisfactory reconstruction time in the perspective of using such an imaging
technique for monitoring. This allows clinicians to obtain almost instantaneous images 24/7 or on demand. Although the reconstructed images do not
feature the complex heterogeneities of the brain, which is in accordance with
what we expect from microwave imaging methods, they allow the characterization of the stroke and its monitoring.
8. Conclusion
We have developed a tool that reconstructs a microwave tomographic image of the brain in less than 2 minutes using 4096 cores. This computational
time corresponds to clinician acceptance for rapid diagnosis or medical monitoring at the hospital. These images were obtained from noisy synthetic
data from a very accurate model of the brain. To our knowledge, this is the
first time that such a realistic study (operational acquisition device, highly
accurate three-dimensional synthetic data, 10% noise) shows the feasibility
of microwave imaging. This study was made possible by the use of massively
parallel computers and facilitated by the HPDDM and FreeFem++ tools
that we have developed. The next step is the validation of these results on
clinical data.
Regarding the numerical aspects of this work, we will accelerate the solution of the series of direct problems, which accounts for more than 80% of
our elapsed time. We explain here the three main avenues of research:
• The present ORAS solver for Maxwell’s equations is a one level algorithm, which cannot scale well over thousands of subdomains. The introduction of a two-level preconditioner with an adequate coarse space
would allow for very good speedups even for decompositions into a large
number of subdomains.
• Recycling information obtained during the convergence of the optimization algorithm will also enable us to improve the performance of the
method, see [17].
20
• Iterative block methods that allow for simultaneous solutions of linear
systems have not been fully investigated. Arithmetic intensity would
be increased since block methods may converge in a smaller number of
iterations while exploiting modern computer architectures effectively.
Acknowledgments
This work was granted access to the HPC resources of TGCC@CEA under
the allocations 2016-067519 and 2016-067730 made by GENCI. This work
has been supported in part by ANR through project MEDIMAX, ANR-13MONU-0012.
[1] J. C. Lin, M. J. Clarke, Microwave imaging of cerebral edema, Proceedings of the IEEE 70 (5) (1982) 523–524.
[2] S. Y. Semenov, D. R. Corfield, Microwave tomography for brain imaging: feasibility assessment for stroke detection, International Journal of
Antennas and Propagation.
[3] S. Semenov, B. Seiser, E. Stoegmann, E. Auff, Electromagnetic tomography for brain imaging: from virtual to human brain, in: 2014 IEEE
Conference on Antenna Measurements & Applications (CAMA), 2014.
[4] R. Beck, R. Hiptmair, Multilevel solution of the time-harmonic
Maxwell’s equations based on edge elements, International Journal for
Numerical Methods in Engineering 45 (7) (1999) 901–920.
[5] J.-C. Nédélec, Mixed finite elements in R3 , Numerische Mathematik
35 (3) (1980) 315–341.
[6] V. Dolean, P. Jolivet, F. Nataf, An Introduction to Domain Decomposition Methods: algorithms, theory and parallel implementation, SIAM,
2015.
[7] F. Pellegrini, J. Roman, SCOTCH: A Software Package for Static Mapping by Dual Recursive Bipartitioning of Process and Architecture
Graphs, in: High-Performance Computing and Networking, Springer,
1996, pp. 493–498.
21
[8] G. Karypis, V. Kumar, A fast and high quality multilevel scheme for
partitioning irregular graphs, SIAM Journal on Scientific Computing
20 (1) (1998) 359–392.
[9] A. St-Cyr, M. J. Gander, S. J. Thomas, Optimized multiplicative, additive, and restricted additive Schwarz preconditioning, SIAM Journal
on Scientific Computing 29 (6) (2007) 2402–2425).
[10] B. Després, P. Joly, J. E. Roberts, A domain decomposition method for
the harmonic Maxwell equations, in: Iterative methods in linear algebra
(Brussels, 1991), North-Holland, Amsterdam, 1992, pp. 475–484.
[11] X.-C. Cai, M. Sarkis, A restricted additive Schwarz preconditioner for
general sparse linear systems, SIAM Journal on Scientific Computing
21 (2) (1999) 792–797.
[12] F. Hecht, New development in FreeFem++, Journal of Numerical Mathematics 20 (3-4) (2012) 251–265.
[13] P. Jolivet, V. Dolean, F. Hecht, F. Nataf, C. Prud’homme, N. Spillane,
High-performance domain decomposition methods on massively parallel architectures with FreeFem++, Journal of Numerical Mathematics
20 (3-4) (2012) 287–302.
[14] P. Jolivet, F. Hecht, F. Nataf, C. Prud’homme, Scalable domain decomposition preconditioners for heterogeneous elliptic problems, in: Proc.
of the Int. Conference on High Performance Computing, Networking,
Storage and Analysis, IEEE, 2013, pp. 1–11.
[15] O. Schenk, K. Gärtner, Solving unsymmetric sparse systems of linear
equations with PARDISO, Future Generation Computer Systems 20 (3)
(2004) 475–487.
[16] P. Amestoy, I. Duff, J.-Y. L’Excellent, J. Koster, A fully asynchronous
multifrontal solver using distributed dynamic scheduling, SIAM Journal
on Matrix Analysis and Applications 23 (1) (2001) 15–41.
[17] M. L. Parks, E. De Sturler, G. Mackey, D. D. Johnson, S. Maiti, Recycling Krylov Subspaces for Sequences of Linear Systems, SIAM Journal
on Scientific Computing 28 (5) (2006) 1651–1674.
22
| 5cs.CE
|
Neural Processing Letters manuscript No.
(will be inserted by the editor)
An Ensemble Classification Algorithm Based on Information
Entropy for Data Streams
arXiv:1708.03496v1 [cs.DS] 11 Aug 2017
Junhong Wang · Shuliang Xu · Bingqian Duan ·
Caifeng Liu · Jiye Liang ·
Received: 02 Aug 2017 / Accepted: date
Abstract Data stream mining problem has caused widely concerns in the area of machine learning
and data mining. In some recent studies, ensemble classification has been widely used in concept drift
detection, however, most of them regard classification accuracy as a criterion for judging whether
concept drift happening or not. Information entropy is an important and effective method for
measuring uncertainty. Based on the information entropy theory, a new algorithm using information
entropy to evaluate a classification result is developed. It uses ensemble classification techniques,
and the weight of each classifier is decided through the entropy of the result produced by an
ensemble classifiers system. When the concept in data streams changing, the classifiers’ weight below
a threshold value will be abandoned to adapt to a new concept in one time. In the experimental
analysis section, six databases and four proposed algorithms are executed. The results show that the
proposed method can not only handle concept drift effectively, but also have a better classification
accuracy and time performance than the contrastive algorithms.
J. Wang
1. School of Computer and Information Technology, Shanxi University, Taiyuan, China
2. Key Laboratory of Computational Intelligence and Chinese Information Processing, Ministry of Education,
Taiyuan, China
Tel.: +86 0351-7010566
E-mail: [email protected]
S. Xu
School of Computer and Information Technology, Shanxi University, Taiyuan, China
E-mail: [email protected]
B. Duan
School of Computer and Information Technology, Shanxi University, Taiyuan, China
E-mail: [email protected]
C. Liu
School of Computer Science and Technology, Faculty of Electronic Information and Electrical Engineering, Dalian
University of Technology, Dalian, China
E-mail: [email protected]
J. Liang
Key Laboratory of Computational Intelligence and Chinese Information Processing, Ministry of Education, Taiyuan,
China
E-mail: [email protected]
2
Junhong Wang et al.
Keywords data streams · data mining · concept drift · information entropy · ensemble classification
1 Introduction
With the development of information society, many fields have produced a large amount of data
streams, such as: e-commerce, network monitoring, telecommunication, stock trading, etc. The data
in these fields are very different from the conventional static data, due to achieving fast, unlimited
number and concept drift in data streams, it makes the technologies of traditional data mining face
a big challenge [1].
Since data streams mining problem was proposed, it has received many attentions [2-6]. Especially, it has been widely used when the ensemble classification method was put forward [7]. Li et
al. proposed a data stream classification algorithm based on a random decision tree model [8]. The
algorithm creates a number of random decision trees, then randomly chooses the split attribute,
and determines the split value through the information gain; in addition, the algorithm can effectively distinguish noise data and concept drift. Elwell et al. proposed an incremental learning
algorithm for periodic concept drift called Learn++.NSE [9]. Learn++.NSE preserves the historical
classification models; when historical classifier to classify the data correctly, the classifier will be
improved the weight; after classifying correctly many times, the weight of the historical classifier
will reach the activation threshold which is predetermined by a user; the classifier is activated from
sleeping state, and then joins to the system to participate in deciding the labels of the unlabeled
data. Aiming at the problem of the unbalanced data stream classification, Rushing et al. proposed
an incremental algorithm based on the overlaying mechanism called CBEA [10]. While the algorithm training a classifier, the training examples are also saved. When the number of classifiers
reaches the threshold, the two most similar coverage sets are selected, and the oldest classifier will
be deleted. The final classification results are determined by KNN algorithm. Gomes et al. proposed
an adaptive algorithm based on the social network called SAE [11], the algorithm introduces the
related concepts of social network and each sub classifier seeing as a node on the network, so a
network consists of multiple classifiers. SAE algorithm sets a series of parameters to measure the
performances of the classifiers, and updates the classifiers to adapt to the new concept. Brzezinski
et al. proposed a data stream classification algorithm based on online data called OAUE [12], the
algorithm uses mean square error to determine the weights of classification models; when a period
of time for the detection coming, the replacement strategy is used to deal with concept drift. Farid
et al. proposed an ensemble classification algorithm based on instances weighting mechanism [13],
in order to detect outliers, the algorithm combines with the clustering algorithm, if a data point
does not belong to any existed cluster in the system, the class of the data will be thought as a new
concept, and then the algorithm counts the data information in leaf nodes to further confirm the
result. Liang et al proposed an online sequential learning algorithm called OS-ELM [14-17], OSELM is a development of the extreme learning machine (ELM) algorithm [18-22]; when a new data
block coming, OS-ELM uses new data to incrementally update the structures of the single hidden
feedforward neural networks; by the mean of the online sequential learning mechanism, OS-ELM
can effectively deal with concept drift in data stream environment.
For the above ensemble classification algorithms, most of them detect the concept drift based on
the accuracy of the classification result. In traditional data stream classification algorithm, accuracy
is an important indicator which is used to detect the concept drift, but by using accuracy, it can
only characterize current performance, and can not reflect the amount of information contained in
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
3
the information system. Information entropy can be used to measure the degree of uncertainty of a
system, and it has been proved to be an effective method to describe the information content in information systems [38-41]. Information entropy is a powerful tool to deal with uncertainty problem,
and it has been applied in many fields. Aiming at the problem of the data stream classification,
in this paper, we extended information entropy to measure the uncertainty of the concept in data
stream, and then an ensemble classification algorithm based on information entropy (ECBE) is proposed. The new algorithm is based on ensemble classification technique; the weighted voting rule
is adopted, and the weights of classifiers are determined according to the change of entropy values
before and after classification; by using Hoeffding bound, ECBE algorithm can estimate whether
concept drift appears or not. When concept drift appearing, the algorithm automatically adjusts the
classifiers according to their weights. Comparing with the existed algorithms, the ECBE algorithm
not only can effectively detect the concept drift, but also can get better classification results.
The rest of the paper is organized as follows: in section 2, we describe the related backgrounds;
section 3 introduces the ensemble classification algorithm based on entropy for data stream; section
4 is the experimental process and data analysis; finally, section 5 gives the conclusions and summary.
2 Backgrounds
2.1 Data stream and concept drift
We assume {· · · , dt−1 , dt , dt+1 , · · · } is a data stream generated by a system, where dt is the instance
generating at t moment; for every instance, we mark dt = {xt , yt }, where xt is the features vector,
and yt is the label. In order to illustrate related notions, we have the following definitions.
Definition 1. A certain number of instances are organized as a data set according to the time
sequence, so we call the data set as data block, which denoted by {d1 , d2 , · · · , dn }, where n is the
size of data block.
In the data stream classification, because in the face of massive data, the storage space of the
data is far beyond the capacity of the computer memory, in order to make the algorithm handle
massive data, the sliding window mechanism is widely used [23], in other words, each time, the
window is only allowed one data block coming into the system, only if the data blocks in the
current window are processed completely, the next data blocks can be acquired.
Definition 2. At t moment, the data in sliding window is used to train a classifier, and the
target concept we get is M ; ∆t time later, we use new data to train another classifier, and we
get the target concept is N ; if M 6= N , we can say concept drift has happened in data stream.
According to the different of ∆t, concept drift can be divided into two types: when ∆t is a short
time, concept drift is called gradual concept drift; when ∆t is a long time, concept drift is called
abrupt concept drift [24,42] 1
After concept drift appearing, the distribution of data in sliding window has changed, and
pt (y | x) 6= pt+1 (y | x); at this time, the performance of the classifiers will decline, if the corresponding measures do not be taken, the error rate of the classification results will continuously rise.
Therefore, in many applications, this property is used to detect concept drift [25].
1 http://kns.cnki.net/KCMS/detail/detail.aspx?dbcode=CDFD&dbname=CDFD1214&filename=1013198451
.nh&uid=WEEvREcwSlJHSldRa1FhcEE0NXdoZ2VUa09aRHFDZ2h2cVlMcHpDU3FSYz0=$9A4hF YAuvQ5ob
gVAqNKPCYcEjKensW4ggI8Fm4gTkoUKaID8j8gFw!!&v=MjE2MDRIYkt4RnRYSnJwRWJQSVI4ZVgxTHV4W
VM3RGgxVDNxVHJXTTFGckNVUkwyZllPUnRGeW5tVTczQVZGMjY=
4
Junhong Wang et al.
2.2 Ensemble classification techniques for data streams
Ensemble classification is an important classification method for data streams, which uses a number
of weak classifiers to combine into a strong classifier; so the method can effectively deal with the
problem of concept drift. In the process of classification, the ensemble classification gives different
classifiers with different weights, by adjusting the weights of the sub classifiers to update classifiers
system to adapt to the new concept of a data stream, classification result of the data eventually is
decided by voting mechanism [26]. In the data stream environment, the classification performance
of ensemble classifiers is better than that of single classifier [27]. Because of the many advantages
of ensemble classification, the method is widely used in data stream data mining [28-30].
3 Ensemble classification based on information entropy in data streams
3.1 Concept drift detection based on information entropy
Entropy is used to describe the disordered state of thermodynamic system. In physics, entropy is
used to indicate the disorder of a system. Shannon extended the concept of the entropy to the
information system, and utilized information entropy to represent the uncertainty of a source [31].
The entropy of a variable is greater, and it indicates that the uncertainty of this variable is larger,
so it needs more information when changing the state of the variable from uncertain to certain.
In the classification of data streams, when data blocks are in the sliding window, at t moment,
we assumes a data block in the sliding window is Bi , and Bi = {(x i1 , yi1 ), (x i2 , yi2 ), · · · , (x in , yin )},
where x is the feature matrix of the ith data block, x = {x i1 ; x i2 ; · · · ; x in }, and y is a label set
of the ith data block, {yi1 , yi2 , · · · , yin } ⊆ y . So at this time, the entropy of the data in the sliding
window is calculated as the equation (1):
H=−
|y|
X
pj log pj
(1)
j=1
Where |y | is the number of the labels in the y , and pk is the probability of the label yk in the
instances; pk can be calculated as the equation (2):
Pn
|yim = yij |
(2)
pj = m=1
n
In the equation (2), if yim = yj , |yim = yij | = 1; else if yim 6= yj , |yim = yij | = 0.
By utilizing the classifiers to classify the data block, the classification result can be obtained
with a weighted voting mechanism; at the same time, we can use the equations (1)-(2) to calculate
the entropy values of the classification result and the real result of the data block which are marked
as H1 and H2 respectively. The entropy’s deviation of the data block can be calculated:
Ψ (H) = |H1 − H2 |
(3)
Lemma (Hoeffding bound) [32, 33]. By independently observing a random variable r for n
times, the range of the random variable r is R, and the observed
q average of r is r̄; when confidence
level is 1 − α, the true value of r is at least r̄ − ε, where ε =
1
R2 ln( α
)
.
2n
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
5
Theorem 1. When the concept of a data stream (the distribution of the data) is stable, S1 and
S2 are the deviation of the two adjacent data blocks which are calculated according to the equation
(3), so it can be known that:
s
R2 ln( α1 )
|S1 − S2 | ≤ 2
(4)
2n
Proof: When the concept of a data stream is stable, the data distributions of the two adjacent
data blocks are consistent, and the difference of the observed entropy values is very small. Let S0
be the real entropy of the data block according to the equation (3), from the Hoeffding bound, it
is known that:
|S1 − S0 | ≤ ε and |S2 − S0 | ≤ ε
So it have:
−ε ≤ S1 − S0 ≤ ε; −ε ≤ S0 − S2 ≤ ε
If plus the two inequalities, the result is as follow:
−2ε ≤ S1 − S2 ≤ 2ε
So it can get |S1 − S2 | ≤ 2ε.
Therefore, it can use ∆H = |Ψ (H1 ) − Ψ (H2 )| as a measure to detect the concept drift in data
streams, where Ψ (H1 ) and Ψ (H2 ) are the deviation of the two adjacent data blocks. If ∆H > 2ε,
it is thought concept drift has appeared. But it is obvious that when a special situation occurring,
only using ∆H cannot work well: if the two adjacent data blocks represent different concepts, so
the classification accuracies of the two adjacent data blocks are very low and they are almost equal;
while calculating ∆H, it is still less than or equal to 2ε, and the algorithm will think no change
happening, but in fact, concept drift has appeared at this time. In order to solve the problem, we
correct the decision criterion of concept drift:
∆H > 2ε or Ψ (H1 ) > 2ε or Ψ (H2 ) > 2ε
(5)
we can say concept drift has appeared. From the equation (5), it is known that when the two
adjacent data blocks occuring two concept drifts, ∆H will be less than or equal to 2ε , however,
Ψ (H1 ) or Ψ (H2 ) will be still greater than 2ε, so concept drift will be detected.
For each sub classifier ensemble(j), (j = 1, 2, 3, · · · , k); according to the classification results of
the current data block, the entropy of the classification result and real result marked as hi and
Hi can be calculated by using the equation (1); so the change of the entropy before and after
classification is Ψ (hi ) = |hi − Hi |, and the weights of the classifiers are updated according to the
equation (6):
ensemble(j).weight(t) = δ(Ψ (hi )) · ensemble(j).weight(t − 1)
(6)
In the equation (6), ensemble(j).weight(t) is the new weight, and ensemble(j).weight(t − 1) is
the weight before updating. The function δ(Ψ (hi )) is showed as the equation (7) :
1
Ψ (hi ) > 0
δ(Ψ (hi )) = e1+Ψ (hi )
(7)
β, β is constant and β > 1 Ψ (hi ) = 0
6
Junhong Wang et al.
Theorem 2. If the concept of the data stream is stable, w1 , w2 , w3 , · · · , wn are the weights of
the ensemble classifiers which are updated on the basis of the equation (6), and the confidence level
is 1 − α, for each sub classifier, the weight must satisfy the following condition:
n
S
1X
wq − √ tα (n − 1) − 3S
wi ≥
n q=1
n
S2 =
n
m
1 X 1X
(
wq − wm )2 .
n − 1 m=1 n q=1
Proof: If the concept of a data stream is stable, each sub classifier in the system is adaptive to
the current concept; therefore, the difference of the classifier weights before after classifying is very
small. We assume that the weights of the ensemble classifiers according with a normal distribution,
so the mean and variance of the weights are calculated as:
µ=
n
n
1 X
1X
wq and S 2 =
(µ − wm )2
n q=1
n − 1 m=1
So we have:
(
P
µ−w
√S
n
)
6 tα (n − 1)
=1−α
From the inequality, it can be known
S
P w ≥ µ − √ tα (n − 1) = 1 − α
n
At the level of confidence 1 − α, we have
S
w ≥ µ − √ tα (n − 1)
n
We use S to approximately replace the standard deviation of the normal distribution δ, according
to the 3δ principle of the normal distribution, it can conclude:
n
wi ≥ w − 3S ≥
S
1X
wq − √ tα (n − 1) − 3S.
n q=1
n
From the theorem 2, we can see that when concept drift appearing, the algorithm will obtain
the statistic about the weights of all classifiers from this time to the last time when the last concept
drift happening, and the lower limit of the updating weight for each sub classifier is calculated as
follow:
n
θ weight =
1X
S
wq − √ tα (n − 1) − 3S
n q=1
n
(8)
According to the classification result of sub classifiers for the new data, the weights are updated
by using the equation (6). When the classifiers cannot adapt to the current concept, the weights
of the classifiers will sharply decreases below θ weight. Finally, the system will delete all classifiers
which does not satisfy the equation (8), so in the next process of the classification, the classifiers
with low performances do not continue to participate in the decision-making.
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
7
3.2 The execution of the ECBE algorithm
From the above knowledge, we can know that the implementation steps of ECBE are as follows:
Algorithm 1 ECBE
Input: Ensemble classifier: ensemble=NULL; data stream S ; the number of sub classifier: k ; the
size of data block: winsize; the array preserving the weights of the classifiers in the period of
two adjacent concept drifts: num=NULL;
Output: The trained classifiers: ensemble.
1: while S !=NULL do
2:
Read winsize instances to organize a data block;
3:
if size(ensemble)< k then
4:
Use Bi to train a new classifier Cj ;
5:
ensemble ← ensemble ∪ Cj and Cj .weight = 1;
6:
else
7:
Use the equation (3) to calculate the entropies Ψ (H1 ) and Ψ (H2 );
8:
for each ensemble(t) ∈ ensemble //updating the weights of the classifiers do
9:
Calculate the entropy ht of the result which is from the ensemble(t) classifying the data
block Bi , and update the weights of the classifiers according to the equation (6);
10:
num ← num ∪ ensemble(t).weight;
11:
end for
12:
if ∆H > 2ε or Ψ (H1 ) > 2ε or Ψ (H2 ) > 2ε // concept drift then
13:
Delete the classifier with the minimum weight;
14:
Calculate the weight θ weight according to the equation (8);
15:
num=NULL;
16:
end if
17:
Delete the classifier with weights less than θ weight;
18:
Use Bi to train a new classifier Cnew ;
19:
if size(ensemble)>=k then
20:
Delete the classifier with the minimum weight;
21:
end if
22:
ensemble = ensemble ∪ Cnew ;
23:
Use Bi to train each classifier in the system;
24:
end if
25: end while
In the ECBE algorithm, it uses entropy to detect concept drift; if the data block does not appear
concept drift, the classifiers system will have a good performance, so the classification results and
the actual labels are nearly consistent, and the difference of the two entropies is very small. When
concept drift appearing, the error rate of the classification result will increase; the difference of
the entropy will also increase, the increment will promote that concept drift in the data stream
can be detected. After concept drift is detected, some classifiers in the system are unable to adapt
to the current concept; if the defunct classifiers are not eliminated, it will not only decreases the
accuracy of the classification results, but also gives a false alarm about concept drift. In order to
solve this problem, ECBE algorithm preserves all the weights of the classifiers where the concept of
8
Junhong Wang et al.
the data is stable in a period of two adjacent concept drift. When concept drift happens, by saving
the statistical weights, the lower bound of the weight when the concept is stable can be computed.
Meanwhile, because of the concept changing, the weights of the classifiers will be dramatically
reduced, so all the classifiers whose weights are below the lower bound will be deleted; the measure
ensures that the classifier system can adapt to new concept at a very fast speed.
4 Experimental analysis
In order to validate the performance of ECBE algorithm, in this paper, we chosen SEA [34], AddExp [35], AWE [27], DCO [36],WEAP I[4], TOS-ELM [20] and OS-ELM [14] as the comparison
algorithms; C4.5, CART and Naive Bayesian et al. are used as base classifiers; all the algorithms
were tested on the three artificial datasets: waveform, Hyperplane and LED which were produced
by MOA platform [37] and the three practical datasets: sensor reading 24, shuttle and page-blocks.
The parameters of the experimental environment are as follows: Windows7 operating system, Intel
dual core CPU, 4 G memory. The algorithm is implemented by R2013a Matlab.
4.1 The descriptions of the datasets
The datasets using in the experiment are cited from UCI datasets. We noly give a brief explain
about the datasets, and the detail can be see from the website 2 .
waveform dataset: it is an artificial dataset and the data set has 50000 instances. Each instance
has 22 attributes; the first 21 attributes are numeric and there are 3 different labels in the dataset.
LED dataset: the dataset has 50000 instances, each of which contains 25 attributes, and the
values of the first 24 attributes are 0 or 1. There are 10 different labels in the dataset; the data
contains 5% of the noise.
Hyperplane dataset:
Pd hyperplane satisfies the following mathePd a sample X in a d dimensional
matical expression: i=1 ai xi = a0 , where a0 = 21 i=1 ai . The dataset contains 50000 instances
Pd
and 11 dimensions; the values of the first 10 dimensions are in [0, 1]. If i=1 ai xi ≥ a0 , the label
of the instance is marked as positive, otherwise, the label of the instance is marked as negative. In
addition, the dataset contains 10% noise.
sensor reading 24 dataset: the dataset contains 5456 instances and 25 properties. The first 24
properties are real; there are 4 different labels in the dataset.
shuttle dataset: the dataset contains 43500 instances and 11 attributes. The values of the first
10 attributes are continuous; there are 7 different labels in the dataset.
page-blocks dataset: the dataset contains 5473 instances and 11 attributes. In the first 10 attributes, some attributes values are continuous, and some of them are discrete. There are 5 different
labels in the dataset.
4.2 The experimental results
In order to verify the performance of the ECBE algorithm, SEA, AddExp, AWE, DCO, WEAP I,
TOSELM and ECBE are run on the three artificial datasets: waveform, Hyperplane and LED. For
2
http://archive.ics.uci.edu/ml/datasets.html
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
9
all the algorithms, k = 5, winsize=2000; the parameters of AddExp are as follows: β = 0.5, γ = 0.1;
the parameters of ECBE are as follows: α = 0.05, β = 0.5, γ = 0.1; the test results are showed in
Tables 1 and 2.
Table 1 Average accuracy on the artificial datasets
SEA
AddExp
AWE
DCO
WEAP I
TOSELM
ECBE
waveform
Hyperplane
LED
0.6777
0.3701
0.6653
0.6628
0.7090
0.5411
0.7583
0.6925
0.5404
0.6912
0.7086
0.6349
0.6349
0.7109
0.2878
0.1622
0.2715
0.3040
0.4165
0.3143
0.8145
Table 2 Time overhead on the artificial datasets (Unit: sec)
SEA
AddExp
AWE
DCO
WEAP I
TOSELM
ECBE
waveform
Hyperplane
LED
2743.422
1124.080
2272.473
1084.533
225.9743
0.01633
164.946
1429.004
1123.530
1004.433
407.561
99.3225
0.0338
85.381
6.2156
1249.691
4.8481
2.5660
4.0223
0.1407
26.551
From Tables 1 and 2, it is known that, on the three artificial datasets, for ECBE, the accuracy is
the highest and the time overhead is far less than the other compared algorithms on most datasets
except for TOSELM; the comprehensive performance of ECBE is the best of all.
From the data Table 1, comparing with the other algorithms, AddExp has the lowest accuracy,
and it is time consuming. The reason for the result is that AddExp updates the classifiers based on
a single data; on the three datasets, the data contains gradual concept drift, and the concept of the
data is constantly and slowly changing, which makes the accuracy of the old classifier decreasing;
when an instance is classified wrongly by a classifier, the algorithm would eliminate the classifier
to deal with the concept drift, however, the classifiers in the AddExp algorithm can get a better
performance only after trained many times; due to the training for the new classifiers is inadequate,
the error rate of the classification result is very higher, which will lead to the more classifier being
replaced, so all the classifiers filling in the system are not trained adequately and the AddExp
algorithm has a higher error rate on the three datasets.
From Table 1, for the LED dataset, the accuracies of the algorithms are very low except the
ECBE algorithm. On the LED dataset, there are 7 attributes relating to concept drift, and the others
are redundant attributes. When concept drift appearing, the speed of the concept changing on the
LED dataset is faster than that on the waveform and Hyperplane datasets; the three algorithms:
SEA, AddExp and AWE delete the classifiers with low performance on the basis of a single classifier,
so it will take a long training time to completely eliminate the weak classifiers; when the speed of
10
Junhong Wang et al.
concept drift is lower than the elimination speed, the performance of the classifiers will be always at
a low level. For the ECBE algorithm, when a concept drift is detected, the weights of the classifier
will be greatly reduced, however, ECBE will delete all weak classifiers according to their weights
at a time, so ECBE can adapt to the new concept at a fast speed. In this experiment, it shows the
performance of ECBE is the best of all.
In order to verify the effectiveness of the algorithm to deal with the practical datasets, SEA,
AddExp, AWE, DCO, WEAP I, TOSELM and ECBE are run on the sensor reading 24, shuttle
and page-blocks datasets, For all the algorithms, the maximum number of classifiers k =5; in the
sensor reading 24 and page-blocks datasets, the size of data block winsize=200 and on the others
datasets, winsize=2000; the parameters of AddExp are as follows: β = 0.5, γ = 0.1; the parameters
of ECBE are as follows: α = 0.05, β = 2, γ = 0.1; the obtained results are shown in Tables 3 and 4.
Table 3 Average accuracies on the practical datasets
SEA
AddExp
AWE
DCO
WEAP I
TOSELM
ECBE
sensor reading 24
shuttle
page-blocks
0.8168
0.5715
0.8673
0.8458
0.9500
0.5608
0.7939
0.8717
0.8127
0.8691
0.8615
0.9054
0.8333
0.9137
0.8923
0.7877
0.9235
0.9231
0.9035
0.8973
0.9215
Table 4 Average time overhead on the real datasets (Unit: sec)
SEA
AddExp
AWE
DCO
WEAP I
TOSELM
ECBE
sensor reading 24
shuttle
page-blocks
269.944
132.719
235.112
78.760
9.7250
2.1260
1.678
1163.610
1028.906
945.015
365.326
54.9151
2.1030
1.156
98.826
127.921
66.186
20.060
3.9338
1.2650
1.013
From Tables 3 and 4, it can be known that all the algorithms get good performances on the
practical datasets. The accuracy of ECBE is most prominent only on the shuttle dataset, but every
accuracy of the algorithm is more than 0.79 and the accuracies of ECBE on the other datasets
are very close to the best results; in addition, for the time overhead, ECBE is far less than the
comparison algorithms, so the advantage of ECBE is very obvious.
From the data in Tables 3 and 4, the performances of SEA, AddExp, AWE, DCO, WEAP I and
TOSELM testing on the three practical datasets are very close and most of them have achieved
good results. On the sensor reading 24, shuttle and page-blocks datasets, the data whose labels are
the same is highly concentrated; in most cases, the data distribution of two adjacent data blocks is
consistent, and the classification model trained by the data has a better adaptability for new data,
so the most performances of the eight algorithms are very good on the practical datasets. After
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
11
analyzing Tables 1-4, SEA, AddExp, AWE and DCO are effective for the datasets where the data
with the same labels is highly concentrated, but when dealing with the datasets containing gradual
concept drift, the performance of those algorithms are not satisfactory; if we analyze the results of
ECBE, it is very obvious that ECBE can deal with the two types of concept drift.
In the ECBE algorithm, we use the equation (8) to calculate the threshold which judges the
change of the entropy of data stream, and the threshold has an important influence on the performance of the algorithm; in the algorithm, the n in the equation (8) is 20 × winsize. To explore the
impact of the winsize values for the algorithm, we select the waveform dataset as the experimental
data, and the parameters of ECBE are as follows: α = 0.05, β = 2, γ = 0.1; we run the ECBE
algorithm, and the results are showed in Table 5 and Fig.1-6.
Table 5 Average accuracy with different winsize values
winsize
800
1000
1200
1400
1600
1800
2000
2200
2400
Hyperplan
LED
shuttle
waveform
0.7050
0.5961
0.8922
0.7602
0.6968
0.8284
0.9143
0.7565
0.7051
0.7403
0.9002
0.7559
0.7081
0.8245
0.9013
0.7556
0.7041
0.8165
0.9115
0.7561
0.7132
0.8151
0.9042
0.7561
0.7109
0.8145
0.9979
0.7583
0.7065
0.8166
0.9145
0.7519
0.7080
0.8131
0.9604
0.75
winsize
20
40
60
80
100
120
140
160
190
page
sensor
0.9409
0.9474
0.9220
0.8000
0.9275
0.7349
0.9303
0.7913
0.9096
0.6758
0.9150
0.7724
0.9109
0.8083
0.9287
0.7664
0.9193
0.7623
Fig. 1 Test result with different winsize values on Hyperplane
In the Fig.1-6 and Table 5, under the situation of different winsize values, it can be seen, from
the test result of the individual data blocks and the entire test results, there is no obvious linear
rule between the winsize values and the performance of ECBE. For the ECBE algorithm, a small
winsize value can lead to a large threshold, which can make ECBE is not too sensitive to the change
of data distribution, but a small data block will cause the classifiers trained by the data are under
12
Junhong Wang et al.
Fig. 2 Test result with different winsize values on LED
Fig. 3 Test result with different winsize values on shuttle
fitting, so the performance of the classifiers system will not be remarkably improved. If the winsize
value is very large, it will lead to a small threshold, which makes the algorithm is too sensitive
to the change of data distribution, so the classifiers will give a false alarm of concept drift and
affect the performance of ECBE. Therefore, on a practical application, the value of winsize needs
to consider the data distribution and the speed of concept changing; after trying many times, we
select the optimal winsize value from a series of candidate values.
In order to study the anti noise performance of ECBE, we select the waveform dataset as the
experimental data, and add 5%, 10%, 15% and 20% noise. The parameters of ECBE are set as
follows: winsize=2000, α = 0.05, β = 2, γ = 0.1. We run the ECBE algorithm on the datasets, and
the test results are showed in the Fig.7 and Table 6.
From the Fig.7, it can be known that the accuracy of ECBE also gradually reduces with the
increase of noise. From the table 6, we can see, comparing with the data containing 5% noise,
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
13
Fig. 4 Test result with different winsize values on waveform
Fig. 5 Test result with different winsize values on page
Table 6 Average accuracy of algorithm with different noise
Noise
5%
7%
10%
12%
15%
17%
20%
25%
Average accuracy
0.7348
0.7263
0.7167
0.7009
0.6929
0.6863
0.6842
0.6678
when the noise increasing 2%, 5%, 7%, 10%, 12% ,15% and 20%, the accuracy of ECBE only
decreases 0.0085, 0.0181, 0.0339, 0.0419, 0.0485, 0.0506 and 0.067 respectively and the magnitude
of the change is very small, thus it is known that ECBE has a good noise immunity. The good
noise immunity of ECBE comes from the elimination strategy; when the noise in the data stream is
increasing, the difference of the entropies of the classification results and the actual results increases,
and the weights of classifiers will decay at a fast speed. If the weights are below a specified threshold,
14
Junhong Wang et al.
Fig. 6 Test result with different winsize values in Hyperplane on sensor reading 24
Fig. 7 Test result with different noise level
the classifiers trained by the noise data will be deleted, so the weak classifiers will not affect the
classification of new data.
In order to further test the performance of ECBE, we compare ECBE with a representative
neural network algorithm called OS-ELM on the 6 datasets and test OS-ELM with different activation functions. In the experiment, we choose radbas, sigmoid, sine and hardlim functions as the
activation functions. The parameters of ECBE are as follows:α = 0.05, β = 2, γ = 0.1; the test
results are showed in Tables 7 and 8.
From Table 7, it can be seen, for the accuracy, ECBE is better than OS-ELM on 4 datasets; on
the waveform and Hyperplane datasets, when the activation function of OS-ELM is sigmoid, OSELM is better than ECBE, but in fact, the advantage is very weak, after calculating, we can know
the accuracies of OS-ELM are only more 0.0124 and 0.0194 than that of ECBE. In the aspect of
time overhead, ECBE is better than OS-ELM on 4 datasets and OS-ELM is only better than ECBE
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
15
Table 7 Average accuracy testing on datasets
ECBE
waveform
Hyperplane
LED
sensor reading 24
shuttle
page-blocks
0.7505
0.7080
0.8119
0.7938
0.9971
0.9321
radbas
OS-ELM
sigmoid
sine
0.3270
0.7020
0.2100
0.5645
0.7825
0.7888
0.7629
0.7274
0.7480
0.4366
0.8364
0.9127
hardlim
0.3595
0.6936
0.4737
0.4740
0.4618
0.4034
0.7116
0.6763
0.4673
0.6562
0.8584
0.8977
winsize
2500
2500
2500
300
1500
300
Table 8 Time overhead on datasets (Unit: sec)
radbas
OS-ELM
sigmoid
sine
hardlim
74.5529
72.9289
73.3113
0.3586
99.4194
0.6552
72.5873
74.4905
72.6981
0.2652
24.1985
0.2964
72.6185
73.6169
72.5395
0.3486
24.1802
0.3987
ECBE
waveform
Hyperplane
LED
sensor reading 24
shuttle
page-blocks
144.6957
68.7527
24.1768
0.8221
7.7122
0.8741
71.9789
773.6481
73.9757
0.3588
24.1958
0.2187
Table 9 The number of concept drift detected by ECBE
Dataset
Hyperplane
SEA
waveform
RBF
number
accuracy
winsize
6
0.6420
100
4
0.7460
100
6
0.746
100
7
0.2002
1650
on 2 datasets. If we synthetically consider the time overhead and accuracy of the two algorithms,
on the sensor reading 24, although ECBE costs more time, the accuracy of ECBE is far higher
than that of OS-ELM, so ECBE is still better than OS-ELM. From the analyses, it is obvious that
ECBE is the best of all.
In order to verify the ability of ECBE detecting the concept drift, we select Hyperplane, SEA,
waveform and RBF as experimental datasets which are produced by MOA. In consideration of
concept drift in the datasets unknown to us, so we reprocess Hyperplane, SEA and waveform, and
make the three dataset contain 5 concept drift in which the label changes form 1 to 2 or reverse
direction changing. The size of the four datasets are 3000, 3000, 3000 and 50000 respectively. The
results are shown in Table 9 and Fig.8.
From Fig.8, we can see when concept drift appearing, the accuracy of ECBE will degrade rapidly,
and then ECBE will delete classifiers with weak performance, so the accuracy will be restored to
the previous high level. In the experiment, ECBE detects concept drift for 15 times; from Table 9,
we can know, on the Hyperplane and waveform datasets, the algorithm has detected all concept
drifts, but producing a error alert; on the SEA dataset, no error alert has produced, but a real
changing is omitted; RBF is a gradual concept drift dataset which is generated by computer and
the number of concept drift in the datset is unknown; we use RBF to test the effectiveness of ECBE
handling gradual concept drift; from the result on RBF, it is obvious that ECBE can detect gradual
concept drift. So in summary, we can include that although there are some deviations for concept
16
Junhong Wang et al.
Fig. 8 Test result with different datasets of ECBE
drift detection in the algorithm, the experimental data indicates the mechanism of ECBE detecting
concept drift is effective and ECBE can cope with different types of concept drift.
5 Conclusion
In this paper, to solve the classification problem of the data streams, an ensemble classification
algorithm based on information entropy was proposed. The algorithm is on the basis of the ensemble
classification method, and utilizes the change of entropies before and after classification to detect
concept drift. The experimental results show that ECBE is effective. But it is obvious that the
ECBE algorithm is only suitable for the single labeled data, so how to apply the algorithm to the
classification of multi-label data is the focus of our future research.
Acknowledgments
This work was supported by National Natural Science Fund of China (Nos.U1435212, 61432011,
61202018, 61303008),National Key Basic Research and Development Program of China (973) (No.2013CB329404).
The authors are grateful to the editor and the anonymous reviewers for constructive comments that
helped to improve the quality and presentation of this paper.
References
[1] J. Gama, R. Sebasti, P.P. Rodrigues, Issues in evaluation of stream learning algorithms, Proceedings of the 15th ACM
SIGKDD international conference on Knowledge discovery and data mining, (ACM, Paris, France, 2009), pp. 329-338.
[2] L.L. Minku, X. Yao, DDD: A New Ensemble Approach for Dealing with Concept Drift, IEEE Transactions on Knowledge
& Data Engineering, 24 (2011) 619-633.
[3] A. Bifet, G. Holmes, B. Pfahringer, R. Kirkby, Gavald, R. , New ensemble methods for evolving data streams, ACM
SIGKDD International Conference on Knowledge Discovery & Data Mining (ACM 2009), pp. 139–148.
[4] Z Ouyang, M Zhou, et al. Mining concept-drifting and noisy data streams using ensemble classifiers. Artificial Intelligence
and Computational Intelligence, 2009. AICI’09. International Conference on. Vol. 4. IEEE, 2009.
An Ensemble Classification Algorithm Based on Information Entropy for Data Streams
17
[5] X. Wu, P. Li, X. Hu, Learning from concept drifting data streams with unlabeled data, Neurocomputing, 92 (2012)
145-155
[6] D. Liu, Y. Wu, H. Jiang, FP-ELM: An online sequential learning algorithm for dealing with concept drift, Neurocomputing, 207 (2016) 322-334.
[7] P. Jedrzejowicz, I. Czarnowski, R.J. Howlett, L.C. Jain, I. Czarnowski, P. J drzejowicz, Ensemble Classifier for Mining
Data Streams, Procedia Computer Science, 35 (2014) 397-406.
[8] P. Li, X. Wu, X. Hu, H. Wang, Learning concept-drifting data streams with random ensemble decision trees, Neurocomputing, 166 (2015) 68-83.
[9] R. Elwell, R. Polikar, Incremental Learning of Concept Drift in Nonstationary Environments, IEEE Transactions on
Neural Networks, 22 (2011) 1517-1531.
[10] J. Rushing, S. Graves, E. Criswell, A. Lin, A Coverage Based Ensemble Algorithm (CBEA) for Streaming Data, IEEE
International Conference on TOOLS with Artificial Intelligence (2004), pp. 106-112.
[11] H. Murilo Gomes, F. Enembreck, SAE: Social Adaptive Ensemble classifier for data streams, Computational Intelligence
and Data Mining (2013), pp. 199-206.
[12] D. Brzezinski, J. Stefanowski, Combining block-based and online methods in learning ensembles from concept drifting
data streams, Information Sciences An International Journal, 265 (2014) 50-67.
[13] D.M. Farid, L. Zhang, A. Hossain, C.M. Rahman, R. Strachan, G. Sexton, K. Dahal, An adaptive ensemble classifier
for mining concept drifting data streams, Expert Systems with Applications, 40 (2013) 5895-5906.
[14] N.Y. Liang, G.B. Huang, P. Saratchandran, N. Sundararajan, A Fast and Accurate Online Sequential Learning Algorithm
for Feedforward Networks, IEEE Transactions on Neural Networks, 17 (2006) 1411-1423.
[15] J.S. Lim, S. Lee, H.S. Pang, Low complexity adaptive forgetting factor for online sequential extreme learning machine
(OS-ELM) for application to nonstationary system estimations, Neural Computing & Applications, 22 (2013) 569-576.
[16] V. Kumar, P. Gaur, A.P. Mittal, Trajectory control of DC servo using OS-ELM based controller, Power India Conference,
2012 IEEE Fifth (2012), pp. 1-5.
[17] Z. Yang, Q. Wu, C. Leung, C. Miao, OS-ELM Based Emotion Recognition for Empathetic Elderly Companion (Springer
International Publishing, (2015).
[18] G.B. Huang, H. Zhou, X. Ding, R. Zhang, Extreme learning machine for regression and multiclass classification, IEEE
Transactions on Systems Man & Cybernetics Part B Cybernetics A Publication of the IEEE Systems Man & Cybernetics
Society, 42 (2012) 513-529.
[19] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: Theory and applications, Neurocomputing, 70 (2006)
489-501.
[20] Gu, Yang, et al., TOSELM: timeliness online sequential extreme learning machine. Neurocomputing 128 (2014): 119-127.
[21] Z. Ma, G. Luo, D. Huang, Short term traffic flow prediction based on on-line sequential extreme learning machine,
Eighth International Conference on Advanced Computational Intelligence (2016).
[22] P. Escandell-Montero, D. Lorente, J.M. Martnez-Martnez, E. Soria-Olivas, J. Vila-Frances, J.D. Martn-Guerrero, Online
fitted policy iteration based on extreme learning machines, Knowledge-Based Systems, 100 (2016) 200-211.
[23] H. Ryang, U. Yun, High utility pattern mining over data streams with sliding window technique, Expert Systems with
Applications, 57 (2016) 214-231.
[24] H. Becker, M. Arias, Real-time ranking with concept drift using expert advice, ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining (2007), pp. 86-94.
[25] J. Gama, P. Medas, P.P. Rodrigues, Learning decision trees from dynamic data streams, ACM Symposium on Applied
Computing (2005), pp. 685-688.
[26] D. Brzezinski, J. Stefanowski, Reacting to different types of concept drift: the Accuracy Updated Ensemble algorithm,
IEEE Transactions on Neural Networks & Learning Systems, 25 (2014) 81 - 94.
[27] H. Wang, W. Fan, P.S. Yu, J. Han, Mining concept-drifting data streams using ensemble classifiers, Kdd, (2003) 226-235.
[28] Z. Ouyang, M. Zhou, T. Wang, Q. Wu, Mining Concept-Drifting and Noisy Data Streams Using Ensemble Classifiers,
International Conference on Artificial Intelligence and Computational Intelligence (2009), pp. 360-364.
[29] Q. Wei, Z. Yang, Z. Junping, W. Yong, Mining multi-label concept-drifting data streams using ensemble classifiers,
International Conference on Fuzzy Systems & Knowledge Discovery (2009), pp. 789.
[30] S. Ramamurthy, R. Bhatnagar, Tracking Recurrent Concept Drift in Streaming Data Using Ensemble Classifiers, International Conference on Machine Learning and Applications (2007), pp. 404-409.
[31] C.E. Shannon, A mathematical theory of communication (McGraw-Hill, 1974).
[32] L. Rutkowski, L. Pietruczuk, P. Duda, M. Jaworski, Decision Trees for Mining Data Streams Based on the McDiarmid’s
Bound, Knowledge & Data Engineering IEEE Transactions on, 25 (2013) 1272-1279.
[33] P. Domingos, G. Hulten, Mining high-speed data streams, ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining (2000), pp. 71-80.
[34] W.N. Street, A streaming ensemble algorithm (SEA) for large-scale classification, ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2001), pp. 1152-1165.
[35] J.Z. Kolter, M.A. Maloof, Using additive expert ensembles to cope with concept drift, International Conference, Bonn,
Germany, August (2005), pp. 449–456.
[36] Y. Zhang, X. Jin, An automatic construction and organization strategy for ensemble learning on data streams, SIGMOD
Rec., 35 (2006) 28-33.
[37] A. Bifet, G. Holmes, R. Kirkby, B. Pfahringer, MOA: Massive Online Analysis, Journal of Machine Learning Research,
11 (2010) 1601-1604.
[38] Xie Z, Tie Y, Guan L. A new audiovisual emotion recognition system using entropy-estimation-based multimodal
information fusion, IEEE International Symposium on Circuits and Systems. IEEE, 2015:726-729.
[39] Zhang X, Mei C, Chen D, et al., Feature selection in mixed data: A method using a novel fuzzy rough set-based
information entropy, Pattern Recognition, 2016, 56(1):1-15.
18
Junhong Wang et al.
[40] Chen Y, Zhang Z, Zheng J, et al., Gene selection for tumor classification using neighborhood rough sets and entropy
measures, Journal of Biomedical Informatics, 2017, 67:59-68.
[41] Y Li,D Li,S Wang,Y Zhai, Incremental entropy-based clustering on categorical data streams with concept drift, KnowledgeBased Systems, 2014, 59:33-47.
[42] Y Yao, L Feng, The Research on Massive and Dynamic Data Stream Classification Method, Dalian University of
Technology, 2013. (in Chinese)
| 8cs.DS
|
arXiv:1401.2170v1 [math.AC] 9 Jan 2014
THE MULTIPLICATIVE STRUCTURE ON HOCHSCHILD
COHOMOLOGY OF A COMPLETE INTERSECTION
RAGNAR-OLAF BUCHWEITZ AND COLLIN ROBERTS
Dedicated to Hans-Bjørn Foxby in admiration
Abstract. We determine the product structure on Hochschild cohomology
of commutative algebras in low degrees, obtaining the answer in all degrees
for complete intersection algebras. As applications, we consider cyclic extension algebras as well as Hochschild and ordinary cohomology of finite abelian
groups.
Contents
1. Introduction
2. Statement of Results
3. Tate Models for the Multiplication Map of an Algebra
4. Diagonal Approximation and Graded Commutativity
5. Products in Low Degree
6. Proof of Theorem 2.5
7. Hochschild Cohomology of Homological Complete Intersections
8. The Case of One Variable
9. Cohomology of Finite Abelian Groups
References
1
2
7
9
12
17
18
20
24
27
1. Introduction
1
For any associative algebra A over a commutative ring K, the Hochschild cohomology ring HH(A/K, A) with its cup (or Yoneda) product is graded commutative
by Gerstenhaber’s fundamental result [20].
However, it is not necessarily strictly graded commutative, that is, the squaring
operation HH2i+1 (A/K, A) → 2 HH4i+2 (A/K, A), for i > 0, from Hochschild cohomology in odd degrees to, necessarily, the 2–torsion of the Hochschild cohomology
in twice that degree, might be nontrivial.
Indeed, this is already so in the simplest example. If A = K[x]/(x2 ) with 2 = 0 in
K, then the Hochschild cohomology ring of A over K is isomorphic to a polynomial
Date: January 13, 2014.
2000 Mathematics Subject Classification. 13D03 16E40 13C40 18G10 .
Key words and phrases. Hochschild cohomology, complete intersection, Tate model, diagonal
approximation, cup and Yoneda product, strictly graded commutative, group cohomology.
The authors were partly supported by NSERC grant 3-642-114-80 and this material is based
upon work supported by the National Science Foundation under Grant No. 0932078 000, while the
first author was in residence at the Mathematical Science Research Institute (MSRI) in Berkeley,
California, during the semester of Jan.–May 2013.
1
All our rings have a multiplicative identity and ring homomorphisms are unital.
1
2
R.-O. BUCHWEITZ AND C. ROBERTS
ring over A in one variable, HH(A/K, A) ∼
= A[z], where A ∼
= HH0 (A/K, A) is of
cohomological degree zero, and the variable z is of degree one, representing the
class ∂/∂x ∈ DerK (A, A) ∼
= HH1 (A/K, A). As we are in characteristic 2, this
ring is indeed graded commutative, but certainly not strictly so, the square of
z 2i+1 ∈ HH2i+1 (A/K, A) returning the non-zero element z 4i+2 ∈ HH4i+2 (A/K, A).
The purpose here is to clarify the structure of this squaring map in the simplest
case, when i = 0 and A is commutative. We will then deduce from this the entire
ring structure when A is a homological complete intersection over K.
The crucial point is that the square of a derivation with respect to cup product, as an element in HH2 (A/K, A), is determined by the Hessian quadratic forms
associated to the defining equations.
The key technical tool we use is the Tate model of the multiplication map
µ : Aev = A ⊗K A → A for a commutative K–algebra A. We show that the
Postnikov tower of such model carries a family of compatible co-unital diagonal
approximations and we give an explicit description of such approximation in low
degrees to obtain the desired results.
Acknowledgement. The present paper relies to a large extent on the second
author’s thesis [30].
The authors wish to thank Øyvind Solberg for originally asking how to determine
the squaring map on Hochschild cohomology of a complete intersection, Luchezar
L. Avramov and Alex Martsinkovsky for valuable help with references, and Lisa
Townsley for sharing with them a synopsis of her thesis [38]. We also thank the
referee of the paper for careful reading and valuable suggestions regarding presentation — aside from pointing out numerous typos that are hopefully taken care of
now.
2. Statement of Results
Our treatment requires some simple facts about differential forms and the calculus that goes along with them. While we cover all facts we will use, the reader
may gain the full picture by consulting [21].
2.1. If P = K[x ∈ X] is a polynomial ring over some commutative ring K, with
X a set of variables, then the module of Kähler differentials of P over
L K is a free
P –module, based on the differentials dx for x ∈ X, that is, Ω1P/K ∼
= x∈X P dx.
If now f ∈ P is any polynomial, then we may consider its Taylor expansion
f (x + dx) in SymP Ω1P/K ∼
= P ⊗K P .
= K[x, dx; x ∈ X] ∼
In concrete terms, the polynomial will only depend on finitely many variables,
say, x = x1 , ..., xn , and then
X
∂ (|a|) f (x) a
dx ,
f (x + dx) =
∂ ax
a=(a1 ,...,an )∈Nn
P
where |a| = i ai and the coefficient of dxa is the corresponding divided partial
derivative of f , in that the usual partial derivative satisfies
∂ (|a|) f (x)
∂ |a| f (x)
=
a
!
·
·
·
a
!
.
1
n
∂ a1 x1 · · · ∂ an xn
∂ a1 x1 · · · ∂ an xn
2.2. The linear part with respect to the dx in that expansion can be viewed as the
Pn
total differential of f , to wit, df = i=1 ∂f∂x(x)
dxi ∈ Ω1P/K .
i
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
3
It defines the P –linear Jacobian of the given polynomial,
gf : DerK (P, P )−→P ,
jac
D 7→ D(df )
on the P –module of K–linear derivations from P to P .
The quadratic part with respect to the dx in the Taylor expansion,
X ∂ (2) f (x)
dxi dxj ,
Hf (dx) =
∂xi ∂xj
16i6j6n
is the Hessian form defined by f . By definition it is an element of Sym2 Ω1P/K and
so defines a P –quadratic form
Pn
that sends D =
i=1
∂
pi ∂x
i
h̃f : DerK (P, P )−→P
P
∂
to
+ x∈X ′ px ∂x
X
∂ (2) f (x)
pi pj ,
∂xi ∂xj
16i6j6n
where we have set X ′ = X \ {x}. The P –quadratic module DerK (P, P ), h̃f
depends solely on f , not on the chosen coordinate system xi .
h̃f (D) = (D ⊗ D)(Hf (dx)) =
2.3. If we take a family of polynomials f = {fb }b∈Y , then the associated (linear)
Jacobian, respectively the (quadratic) Hessian map are given by
X
X
Y
gf =
]
jac
jac
h̃f =
h̃fb ∂fb : DerK (P, P ) →
P ∂fb ,
fb ∂fb ,
b∈Y
b∈Y
b∈Y
where the symbols ∂fb simply represent the corresponding δ–functions in the product of the free P –modules P ∂fb .
2.4. If A = P/I, with I = (fb )b∈Y , then A is a commutative K–algebra and the
module of Kähler differentials of A over K has a free A–module presentation
M
M
j
A[fb ] −−→ A ⊗P Ω1P/K ∼
Adx −−→ Ω1A/K −−→ 0 ,
=
x∈X
b∈Y
where the basis element L
[fb ] is mapped to j([fb ]) = 1 ⊗ dfb . The A–linear map j
factors as the surjection b∈Y A[fb ] → I/I 2 onto the conormal module I/I 2 that
sends [fb ] to fb mod I 2 , followed by the A–linear map d : I/I 2 → A ⊗P Ω1P/K that
is induced by the universal K–derivation, sending f mod I 2 to 1 ⊗ df .
It follows that the Jacobian and Hessian descend,
X
X
Y
jacf =
jacfb ∂fb , hf =
hfb ∂fb : DerK (P, A) →
(∗)
A∂fb ,
b∈Y
b∈Y
b∈Y
2
the A–linear map jacf taking its values in NA/P = HomQ
A (I/I , A), the normal
module of A with respect
L to P , viewed as a submodule of b∈Y A∂fb , the A–dual
of the free A–module b∈Y A[fb ]. The kernel of the Jacobian satisfies
Ker jacf ∼
= DerK (A, A) ∼
= HH1 (A/K, A) ∼
= Ext1Aev (A, A) .
Note that the last isomorphism holds for any associative K–algebra A, when Aev
denotes the enveloping algebra of A over K, while in higher degrees one only has
natural maps HHi (A/K, A) → ExtiAev (A, A) that are isomorphisms when A is projective as K–module; see also 2.6 below.
4
R.-O. BUCHWEITZ AND C. ROBERTS
Now we can formulate our first result.
Theorem 2.5. Let A = P/I be a presentation of a commutative K–algebra as
quotient of a polynomial ring P = K[x; x ∈ X] over some commutative ring K
modulo an ideal I = (fb )b∈Y ⊆ P .
(1) The Hessian map hf induces a well defined A–quadratic map
q : DerK (A, A) −→ HomA (I/I 2 , A)
from the module of derivations to the normal module. Explicitly,
!
X ∂ (2) f (x)
X
∂
q
(f mod I 2 ) =
ai aj mod I ,
ax
∂x
∂xi ∂xj
16i6j6n
x∈X
where, as above, x1 , ..., xn ∈ X are the variables that actually occur in a presentation of f , linearly ordered in some way.
(2) There is a canonical A–linear map
δ : HomA (I/I 2 , A)−→ Ext2Aev (A, A) ,
1
whose image is TA/K
⊆ Ext2Aev (A, A), the first tangent or André–Quillen cohomology of A.
(3) The map from DerK (A, A) = Ext1Aev (A, A) to Ext2Aev (A, A) that sends a derivation D to the class [D ◦ D] of its square under Yoneda product equals δ ◦ q and
1
takes its image in the 2–torsion of TA/K
.
Remark 2.6. If A as in the preceding result is further projective over K, then
there is a natural bijection HH• (A/K, A) → Ext•Aev (A, A) of graded commutative
K–algebras2.
1
Moreover, TA/K
⊆ HH2 (A/K, A) identifies with the A–submodule of classes of
symmetric Hochschild 2–cocycles.
2.7. In our second result, let the K–algebra A = P/I be presented as above, but
assume moreover that
(a) A is transversal to itself as K–module in that
K
TorK
+ (A, A) = ⊕i>0 Tori (A, A) = 0 .
(b) The ideal I ⊆ P is regular in the sense of Quillen [29, Def.6.10], that is, I/I 2
is a projective A–module and the canonical morphism of graded commutative
algebras
^
(I/I 2 )−→ TorP (A, A)
A
is an isomorphism. In (loc.cit.) Quillen points out that for P noetherian, I is
regular in P if, and only if, it is generated locally by a P –regular sequence.
2Simply using the definition of the algebra structure on source or target, one obtains an isomorphism from the opposite algebra of HH• (A/K, A) to the Yoneda Ext–algebra. However, as already
m
Gerstenhaber pointed out in his original paper, sending an element u of degree m to (−1) 2 u
provides an algebra isomorphism between a graded commutative algebra and its opposite algebra.
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
5
We will say, for short, that A is a homological complete intersection algebra over K
if conditions (a) and (b) are satisfied. Alternative terminology would be to say that
the kernel of the multiplication map µ : Aev → A has “free” or “exterior Koszul
homology,” a notion introduced and studied by A. Blanco, J. Majadas Soto, and
A. Rodicio Garcia in a series of articles following [7], or to call the multiplication
map a “quasi-complete intersection homomorphism” as in [4].
2.8. The key properties of such homological complete intersection algebras are that
the corresponding Tate model is small and the associated cotangent complex LA/K
is, by [29, Cor. 6.14], concentrated in homological degrees 0 and 1. More precisely,
LA/K ≡ 0
/ I/I 2 [1]
d
/ A ⊗P Ω1 [0]
P/K
/0
and its shifted A–dual, the shifted tangent complex tA := HomA (LA/K [1], A) is the
complex
tA ≡ 0
/ DerK (P, A)[−1]
jacf
/ NA/P [−2]
/ 0,
concentrated in cohomological degrees one and two.
2.9. The link between tangent and Hochschild cohomology was established in
Quillen’s fundamental paper [29, Sect. 8]. In essence, it says that the complex
returning Hochschild homology can be realized as a DG Hopf algebra whose primitive part is represented by the cotangent complex. The qualification here is two-fold:
there generally is only a spectral sequence reflecting this interpretation that generally only degenerates in characteristic zero, and, on the dual side, for Hochschild
cohomology vis-à-vis the enveloping algebra of the tangent complex, the algebra
structures in cohomology do not match up. However, for complete intersection
algebras, things are well controlled because the cotangent complex is so short. In
detail, we can complement the picture as follows.
2.10. The Hessian A–quadratic map q from 2.4(∗) above can be interpreted as
defining on tA the structure of a graded (super) Lie algebra, where we follow the
axiomatization of this notion in [2], and the Jacobian can be viewed as a Lie algebra
differential. The enveloping algebra of this DG Lie A–algebra; see [34]; is readily
identified with the Clifford algebra Cliff(q) on that quadratic map and the Jacobian
induces a DG algebra differential ∂jacf on it. In particular, the cohomology of this
DG algebra inherits an algebra structure this way.
Theorem 2.11. Let A = P/I be a homological complete intersection algebra over
K, with P = K[x1 , ..., xn ] and I = (f1 , ..., fc ) such that the fj induce an A–basis
of the A–module I/I 2 that we hence further assume to be free.
The cohomology of the DG algebra (Cliff(q), ∂jacf ) is then the Yoneda Ext–algebra
Ext•Aev (A, A) of self extensions of A as bimodule. In detail,
Cliff(q) ∼
= Aht1 , ..., tn ; s1 , ..., sc i ,
with the ti in degree 1, the sj in degree 2, the differential determined by
c
X
∂fj
mod I sj , ∂(sj ) = 0 .
∂(ti ) =
∂xi
j=1
6
R.-O. BUCHWEITZ AND C. ROBERTS
With respect to the multiplicative structure, the sj are central and
c (2)
X
∂ fj
2
mod I sj
ti =
∂x2i
j=1
c
X
∂ 2 fj
′
′
ti ti + ti ti =
mod I sj .
∂xi ∂xi′
j=1
If A is projective over K, then the natural homomorphism of graded commutative algebras HH• (A/K) → Ext•Aev (A, A) is an isomorphism, thus, the Hochschild
cohomology is given as cohomology of that DG Clifford algebra too.
Remark 2.12. In concrete terms, expanding on 2.9 above, the Poincaré–Birkhoff–
Witt theorem for enveloping algebras of (super) Lie algebras shows that as differential graded coalgebras,
in particular, as complexes of graded A–modules, one has
that Cliff(q), ∂jacf is isomorphic to the Koszul complex K over S = SymA (NA/P ) ∼
=
A[s1 , ..., sj ] on the sequence
c
X
∂fj
mod I sj ∈ S ; i = 1, ..., n.
∂xi
j=1
Bigrading the Koszul complex by placing the variables ti = ∂xi ∈ DerK (P, A)
into bidegree (1, 0) and sj = ∂fj into bidegree (0, 1), one regains the – well-known
[39, 22] – A–linear Hodge decomposition of the Hochschild cohomology for such a
homological complete intersection,
M
H i (K)j
HHp (A/K) ∼
=
i+2j=p
H (K)j ∼
= H i+2j HomA (Si+j (LA/K [1]), A) .
i
However, the preceding theorem says that, generally, this decomposition is not
compatible with the multiplicative structure, in that squaring sends HH1 (A/K) =
H 1 (K)0 to
1
= H 2 (HomA (LA/K [1], A)) ⊆ HH2 (A/K) ,
H 0 (K)1 ∼
= TA/K
but not to the other summand
H 2 (K)0 ∼
= HomA (Ω2A/K , A) = H 2 (HomA (S2 (LA/K [1]), A)) ⊆ HH2 (A/K) .
In other words, the “obvious” multiplication on the Hodge decomposition needs to
be twisted or “quantized”, here from the exterior multiplicative structure of the
Koszul complex to that of the (homogeneous) Clifford algebra.
This phenomenon is analogous; see [13] and the discussion therein; to the situation for non-affine smooth schemes in characteristic zero. Similarly, for arbitrary
commutative rings over a field of characteristic zero, the relation between Hodge
decomposition and multiplicative structure on Hochschild cohomology is studied
from a combinatorial point of view in [6].
We hasten to point out that for the affine (homological) complete intersections
considered here this multiplicative twist can only make a difference when 2 is not
a unit in A.
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
7
We end the paper by presenting some simple examples that highlight how the
treatment here encompasses in a unified manner various results scattered throughout the literature.
3. Tate Models for the Multiplication Map of an Algebra
Let S → R be a homomorphism between commutative rings. In his seminal
paper [37], John Tate presented in essence the following construction of a free S–
resolution of R; see also [3] for further details, and, in particular, [14] for details
on differential graded (DG) algebras with divided powers. The special case of the
multiplication map R ⊗K R → R for a commutative K–algebra R of finite type
that we will be mainly interested in has also been detailed in [5].
π
3.1. There exists a factorization of the given ring homomorphism as S → (T , ∂) −
→
R such that
(i) T is a strictly graded commutative S–algebra of the form
!
^
M
M
T• = SymS (F0 ) ⊗S
Fi ,
Fi ⊗S ΓS
S
i odd
i>0 , even
with divided powers (in even degrees), graded so that S is concentrated in
degree 0, each Fi , i > 0, a free S–module concentratedVin degree i.
Here Sym denotes the symmetric algebra functor,
the exterior algebra
functor, and Γ the divided power algebra functor.
(ii) The S–algebra differential ∂ : T• → T•−1 is compatible
L with divided powers, thus, uniquely determined by its restriction to i>1 Fi ⊆ T , vanishing
necessarily on F0 for degree reasons.
(iii) The morphism π is a quasiisomorphism of DG algebras, where one endows R,
concentrated in degree zero, with the, only possible, zero differential.
Such factorization of S → R is then a Tate model for R over S, but, with the
usual slight abuse, we also simply call (T , ∂) a Tate model, the remaining data
understood.
For given i, the module ⊕j6i Fj generates a DG subalgebra T (i) with divided
S
powers, and T (0) ⊆ · · · ⊆ T (i) ⊆ T (i+1) ⊆ · · · ⊆ T = i T (i) constitutes the
associated Postnikov tower. It follows from the description given that the inclusion
∼
=
T (i) → T induces isomorphisms Hj (T (i) ) −−→ Hj (T ) for j < i and that ∂|Fi+1 :
(i)
Fi+1 → T
can be interpreted as an S–linear attaching map that kills Hi (T (i) ).
Remark 3.2. Viewed just as a complex, such a Tate model constitutes in particular
a free resolution of R as S–module, thus, can be used to calculate TorS (R, −) and
ExtS (R, −).
If S is noetherian and S → R is surjective, one may choose F0 = 0 and each Fi ,
for i > 1, to be free of finite rank over S, and then T is in turn free of finite rank over
S in each degree. If R is only of finite type as an S–algebra, then one may choose
F0 to be a finite free S–module and each Fi a finite free T (0) = SymS (F0 )–module.
3.3. Now consider the particular case, where A is a commutative K–algebra over
some commutative ring K, and µ : Aev := A ⊗K A = S → A = R is the multiplication map, a K–algebra homomorphism from the enveloping algebra of A
e 1A = Ω
e 1A/K the kernel of the multiplication map and
over K to A. We denote Ω
8
R.-O. BUCHWEITZ AND C. ROBERTS
note that this ideal in Aev represents the functor of taking K–linear derivations
into arbitrary A–bimodules — in contrast to the A–module of Kähler differential
e 1 /(Ω
e 1 )2 that represents the functor of taking K–linear derivations
forms Ω1A/K = Ω
A
A
in symmetric A–bimodules. We use the convention that the universal derivation
e 1A sends a 7→ da = 1 ⊗ a − a ⊗ 1.
d:A→Ω
Assume now that A is presented as (SymK F )/I, for some free K–module F and
some ideal I in the polynomial ring P = SymK F over this free module.
The following result holds for an arbitrary presentation A = P/I, with neither
P nor A necessarily commutative.
Lemma 3.4 (cf. [18, Cor.2.11]). There exists an exact sequence of Aev –modules
(†)
I/I 2
d
e 1P ⊗P A
/ A ⊗P Ω
j
/ Aev
µ
/A
/0
e 1 to
where d is induced by restricting the universal K–linear derivation d : P → Ω
P
1
∼
e
I and j is obtained from the inclusion i : ΩP ⊆ P ⊗K P as j = A ⊗P i ⊗P A taking
into account Aev ∼
= A ⊗P (P ⊗K P ) ⊗P A.
Returning to the commutative situation, choose a P ev –linear surjection P ⊗K
G ⊗K P → I for some free K–module G. The preceding Lemma has the following
immediate consequence.
Corollary 3.5. One may construct a Tate model for the multiplication map µ :
A ⊗K A → A with F0 = 0, F1 = A ⊗K F ⊗K A, and F2 = A ⊗K G ⊗K A.
Proof. As the multiplication map is surjective, one may choose F0 = 0 in the
construction of the Tate model.
e 1 to F ⊆ P
The image of the restriction of the universal derivation d : P → Ω
P
e 1 ⊆ P ⊗K P and this shows that the induced map A ⊗K
generates the ideal Ω
P
e 1P ⊗P A is surjective, thus, the choice of F1 . The resulting DG
F ⊗K A → A ⊗PV Ω
algebra T (1) = Aev F1 is nothing but the Koszul complex over the composition
j
e 1 ⊗P A −
→ Aev whose first homology is isomorphic to
A ⊗K F ⊗K A → A ⊗P Ω
P
e 1 ⊗P A.
d(I/I 2 ) — this is exactness of the sequence (†) above at the term A ⊗P Ω
P
Applying on both sides the tensor product over P with A to the P ev –linear
surjection P ⊗K G ⊗K P → I yields an Aev –linear surjection first from F2 onto
I/I 2 and then via d onto that homology. As F2 is Aev –projective, one can find
a lifting of that surjection into the 1-cycles of T (1) and such lifting extends the
differential from T (1) to
^
T (2) =
(F1 ) ⊗Aev ΓAev (F2 ) ,
ev
A
whence the claim concerning F2 follows.
To be even more specific, Lemma 3.4 and the proof of the corollary indicate how
to describe the possible differentials on T (2) .
3.6. Let x = {xa }a be a basis of the free K–module F . Then the universal
derivation sends these algebra generators to
e 1P ⊆ P ⊗K P ∼
dxa = x′′a − x′a ∈ Ω
= SymK (F ⊕ F ) ,
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
9
where we abbreviate x′ = x ⊗ 1, x′′ = 1 ⊗ x. Below, we will use the same notation
for the respective residue classes in AevP
.
Representing a polynomial f (x) = ′A=(a1 ,...,ak ) αA xa1 · · · xak ∈ P , with αA ∈
K, as a finite sum of (ordered) monomials in (finitely many) basis elements xa of
F , the definition of d and then the product rule for this derivation yield
(∗)
df = 1 ⊗ f − f ⊗ 1 = f (x′′ ) − f (x′ )
X
=
k
X
αA x′a1 · · · x′aκ−1 dxaκ x′′aκ+1 · · · x′′ak
A=(a1 ,...,ak ) κ=1
=
X
k
X
αA x′a1 · · · x′aκ−1 (x′′aκ − x′aκ )x′′aκ+1 · · · x′′ak
A=(a1 ,...,ak ) κ=1
in
e1
Ω
P
= (x′′a − x′a )a ⊆ P ⊗K P .
3.7. Slightly abusing notation, we may then base the free Aev –module F1 in the
construction of T above on symbols dxa and define ∂dxa = x′′a − x′a ∈ Aev . Similarly, if {fb }b are polynomials in I whose classes generate I/I 2 , we may base the
free Aev –module F2 on symbols dfb and choose
(∗∗)
∂(dfb ) =
X
k
X
αA x′a1 · · · x′aκ−1 dxaκ x′′aκ+1 · · · x′′ak ∈ F1
A=(a1 ,...,ak ) κ=1
for the differential on those basis elements, where the right hand side results from
a chosen presentation of f = fb as above.
This time, the right hand side will depend on the chosen presentation, but those
choices are easily controlled, in that one may add ∂(ωb ) to the right hand side for
V
any two-form ωb ∈ 2Aev F1 to change the value of ∂(dfb ). These are also the only
choices.
If we collect coefficients of the terms dxa in (∗) for f = fb to write
X
dfb =
∆b,a (x′ , x′′ )dxa ,
a
with ∆b,a (x′ , x′′ ) in P ⊗K P , then these coefficients satisfy
µ∆b,a (x′ , x′′ ) = ∆b,a (x, x) =
∂fb (x)
∂xa
in P irrespective of the representation we chose.
4. Diagonal Approximation and Graded Commutativity
With T a Tate model of the multiplication map Aev → A for a commutative
K–algebra as before, note that T is naturally a complex of A–bimodules so that
we can form the tensor product T ⊗A T . This tensor product carries again the
structure of a differential strictly graded commutative algebra with divided powers.
Denote j1,2 : T → T ⊗A T the DG algebra homomorphisms j1 (t) = t ⊗A 1,
respectively j2 (t) = 1 ⊗A t, for t ∈ T .
4.1. The complex underlying T ⊗A T is one of free A–trimodules, that is, free
modules over A⊗3 . We make no claim about the homology of this complex, except
10
R.-O. BUCHWEITZ AND C. ROBERTS
to note that the algebra homomorphism ǫ ⊗A ǫ : T ⊗A T → A ⊗A A ∼
= A is the
natural map onto the zeroth homology.
Note, however, that if A is flat over K, then Aev is A–flat via both j1 or j2 , and
T constitutes a flat resolution of A over A via either j1 or j2 , thus, the homology
of T ⊗A T is trivial, equal to A in degree zero.
However, even without flatness we have the following.
Lemma 4.2. With notation as introduced above,
(1) The algebra homomorphisms
ǫ1 = ǫ ⊗ A T : T ⊗ A T → A ⊗ A T ∼
=T
∼T
ǫ2 = T ⊗ A ǫ : T ⊗ A T → T ⊗ A A =
are homotopic as morphisms of complexes of A–trimodules.
(2) If Hm (T ⊗A T ) = 0 for 1 ≤ m ≤ i for some i, then Ker ǫ1 ∩Ker ǫ2 has vanishing
homology up to and including degree i.
In particular, if A is flat over K, then the subcomplex Ker ǫ1 ∩ Ker ǫ2 ⊆
T ⊗A T has zero homology.
Proof. As ǫǫ1 = ǫǫ2 , the difference ǫ2 − ǫ1 takes its values in the acyclic complex
Ker(ǫ : T → A). As its source is a complex of projective, even free A–trimodules,
the claim (1) follows.
Concerning (2), consider the following commutative diagram of complexes with
exact rows and columns
0
0O
0O
/ Ker ǫ
O
/T
O
0O
ǫ
/A
O
ǫ2
0
/ Ker ǫ1
O
/ T ⊗A T
O
0
/ Ker ǫ1 ∩ Ker ǫ2
O
0
/0
ǫ
ǫ1
/T
O
/0
/ Ker ǫ2
O
/ Ker ǫ
O
/0
0
0
By construction, ǫ is a quasi isomorphism, whence Ker ǫ is acyclic. By assumption
ǫ1 , ǫ2 are quasiisomorphisms in degrees up to and including i, where we may take
i = ∞ if A is flat over K. The long exact homology sequence obtained from the
short exact sequence, say, at the bottom yields then the claims.
Definition 4.3 (cf. [32]). Let A be a complex of Aev –modules that comes with
ǫ
an augmentation A −
→ A. A co-unital diagonal approximation on A is a morphism
Φ : A → A ⊗A A of Aev –complexes so that the following diagram commutes,
A❍
✇✇✇✇ ❍❍❍❍❍❍❍❍
✇
✇
✇
❍❍❍❍
Φ
✇✇✇✇
❍❍❍❍
✇✇✇✇✇
❍❍❍
✇
✇
✇✇
/ A,
o
A
A ⊗A A
ǫ⊗A
A⊗ǫ
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
11
where we have identified A ⊗A A ∼
=A∼
= A ⊗A A.
As is well-known, the existence of a co-unital diagonal approximation on a Tate
model of the multiplication map implies graded commutativity of the Yoneda product on ExtAev (A, A); see, for example, [11, 32, 35].
Definition 4.4. Given a co-unital diagonal approximation Φ : T → T ⊗A T on
the Tate model of µ : Aev → A, define a cup product ∪ = ∪Φ on Hom•Aev (T , A)
through
f ⊗g
Φ
µ
f ∪ g = µ(f ⊗A g)Φ : T −−→ T ⊗A T −−−−→ A ⊗A A −−
→ A,
∼
=
for cochains f, g : T → A.
Proposition 4.5. Let A be a commutative K–algebra with a Tate model of µ :
Aev → A that admits a co-unital diagonal approximation. The corresponding
cup product then induces a graded commutative product in H(Hom•Aev (T , A)) =
Ext•Aev (A, A). This product coincides with the usual Yoneda or composition product on the Ext–algebra.
Proof. Given cocycles f, g ∈ Hom•Aev (T , A), lift them to morphisms of complexes
f˜, g̃ ∈ Hom•Aev (T , T ). In the homotopy category of complexes ǫ1 , ǫ2 : T ⊗A T → T
agree because of Lemma 4.2(1), and so the following diagram of complexes of Aev –
modules commutes in the homotopy category,
˜
id ⊗g̃
Φ
/ T ⊗A T f ⊗id / T ⊗A T ǫ1 ⊗A ǫ2/ A ⊗A A
T ❍❍❍ / T ⊗A T
❍❍❍❍
❍❍❍❍
❍❍❍❍ ǫ1
ǫ1 = ǫ2
ǫ2
❍❍❍❍
❍
g̃
f˜
∼
/T ❱
/T ▼
= µ
T ❨❨❨❨❨❨❨
▼▼▼
❨❨❨❨❨❨ ❱❱❱❱❱❱❱
❨❨❨❨❨❨ ❱❱❱❱f
▼
ǫ
❨❨❨❨❨❨ ❱❱❱❱ ▼▼▼▼
❨❨❨❨❨❱❨❱❱❱❱ ▼▼
f ◦g̃
❨❨❨❨❱❨❱❨❱▼&*
,A
By definition, f ◦ g̃ is a cocycle representing the Yoneda product of f with g, while
the composition across the top and down yields f ∪ g. Thus, these two products
coincide in ExtAev (A, A).
|g|
|f |
Further, for f˜ ∈ EndAev (T ), g̃ ∈ EndAev (T ) lifts of homogeneous cocycles as
above, one has
(f˜ ⊗ id)(id ⊗g̃) = f˜ ⊗ g̃ = (−1)|f ||g| (id ⊗g̃)(f˜ ⊗ id)
whence combining the above diagram with that corresponding to (id ⊗g̃)(f˜ ⊗ id)
yields graded commutativity of the Yoneda or cup product in cohomology, that is,
in Ext•Aev (A, A).
Remark 4.6. The preceding proof shows that a partial co-unital diagonal approximation Φ(i) : T (i) → T (i) ⊗A T (i) is enough to express the Yoneda product of two
cohomology classes whose degrees add up to at most i through the cup product on
Hom•Aev (T (i) , A). Moreover, these classes will commute in the graded sense with
respect to the product.
In the particular case that A is flat over K, one can indeed construct a co-unital
diagonal approximation that is furthermore a homomorphism of DG algebras with
divided powers.
12
R.-O. BUCHWEITZ AND C. ROBERTS
Proposition 4.7. If A is flat over K, then there exists a co-unital diagonal approximation Φ on the Tate model T of µ : Aev → A that induces such approximation
Φ(i) on each DG algebra T (i) in the Postnikov tower.
Proof. We construct a desired co-unital diagonal approximation Φ inductively along
the Postnikov tower. At the basis of the Postnikov tower, we have T (0) = T0 = Aev
and T (0) ⊗A T (0) = (T ⊗A T )0 ∼
= A⊗3 ⊆ T ⊗A T . Now
Φ(0) : T (0) ∼
= A⊗2 → T (0) ⊗A T (0) ∼
= A⊗3 ,
Φ(0) (a ⊗ b) = a ⊗ 1 ⊗ b
defines a homomorphism of algebras such that ǫ1 Φ(0) = idA⊗2 = ǫ2 Φ(0) . Via Φ(0) ,
we can, and will, view T ⊗A T as a DG Aev –algebra.
Now assume a co-unital diagonal approximation Φ(i) : T (i) → T (i) ⊗A T (i) has
been defined for some i > 0. With T (i+1) = T (i) hFi+1 i for some free Aev –module
Fi+1 , note that (j1 + j2 )|Fi+1 : Fi+1 → T (i+1) ⊗A T (i+1) satisfies
ǫ1 (j1 + j2 )|Fi+1 = idFi+1 = ǫ2 (j1 + j2 )|Fi+1 .
For m = 1, 2, by the induction hypothesis ǫm Φ(i) = idT (i) . As ǫm is a morphism of
complexes, it follows that
ǫm (Φ(i) ∂ − ∂(j1 + j2 )) = ∂ − ∂ǫm (j1 + j2 ) ,
vanishes on Fi+1 . Because Φ(i) is a morphism of complexes,
∂(Φ(i) ∂ − ∂(j1 + j2 )) = 0 ,
whence it follows that Φ(i) ∂ − ∂(j1 + j2 ) maps Fi+1 into the cycles of Ker ǫ1 ∩Ker ǫ2 .
As the latter complex is acyclic for A flat over K by Lemma 4.2(2), and as Fi+1 is
projective, even free, we can find Φ′i+1 : Fi+1 → Ker ǫ1 ∩ Ker ǫ2 such that Φ(i) ∂ =
∂(Φ′i+1 + j1 + j2 ) on Fi+1 .
The algebra homomorphism given by Φ(i) on T (i) and extended through
Φi+1 = Φ′i+1 + j1 + j2
first to Fi+1 and, then as algebra homomorphism to Φ(i+1) on T (i+1) satisfies
∂Φ(i+1) = Φ(i+1) ∂ and ǫm Φ(i+1) = id. In other words, we have found an extension
of Φ(i) to a co-unital diagonal approximation on T (i+1) , completing the inductive
argument.
5. Products in Low Degree
5.1. We can sharpen Proposition 4.7 a bit in low degrees in that we do not need
any flatness there.
Proceding as in the proof of Proposition 4.7, define Φ1 = (j1 + j2 )|F1 : F1 →
T (1) ⊗A T (1) and extend Φ(0) , Φ1 to a homomorphism Φ(1) : T (1) → T (1) ⊗A T (1)
of strictly commutative graded algebras with divided powers. It will indeed be
a co-unital approximation and homomorphism of DG algebras, as Φ(0) is such a
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
13
homomorphism and
∂Φ1 (dxa ) = ∂(j1 + j2 )(dxa ) = (j1 + j2 )∂(dxa )
= j1 (1 ⊗ xa − xa ⊗ 1) + j2 (1 ⊗ xa − xa ⊗ 1)
= (1 ⊗ xa ⊗ 1 − xa ⊗ 1 ⊗ 1) + (1 ⊗ 1 ⊗ xa − 1 ⊗ xa ⊗ 1)
= 1 ⊗ 1 ⊗ xa − xa ⊗ 1 ⊗ 1 = Φ(0) (∂dxa )
for any basis element dxa in F1 .
Put differently, a diagonal approximation Φ(1) : T (1) → T (1) ⊗A T (1) can be
chosen in such a way that
Φ(1) (a ⊗ b) = a ⊗ 1 ⊗ b
Φ(1) (dxa ) = dx′a + dx′′a
for a ⊗ b ∈ Aev ,
for dxa ∈ F1 ,
where we have shortened notation to dx′a = dxa ⊗1 = j1 (dxa ) and dx′′a = 1⊗dxa =
j2 (dxa ). We also use the algebra homomorphism Φ(0) : Aev → A⊗3 to identify
x′ = Φ(0) (x′ ) = x ⊗ 1 ⊗ 1 ,
x = 1⊗x⊗1,
x′′ = Φ(0) (x′′ ) = 1 ⊗ 1 ⊗ x .
5.2. It is noteworthy that the induced cup product on Hom•Aev (T (1) , A) is always
strictly graded commutative. Indeed, recall that T (1) is nothing but the Koszul
complex on (x′′a − x′a )a in Aev . Dualizing into A, the induced differential becomes
zero, and there is a quasi isomorphism
M
M
Hom•Aev (T (1) , A) ≃
HomAev (A ⊗P ΩiP/K ⊗P A, A) ∼
HomP (ΩiP/K , A)
=
i>0
i>0
to the graded module of alternating K–linear polyvector fields on P with values in
|f |
A. If now f ∈ HomAev (T (1) , A) is a cocycle of odd degree, then f ∪ f = 0. In fact,
(1)
if ω ∈ T2|f | , then Φ(1) (ω), in Sweedler notation, is of the form
X
Φ(1) (ω) =
(ω(1) ⊗A ω(2) + (−1)|ω(1) ||ω(2) | ω(2) ⊗A ω(1) )
for suitable homogeneous ω(k) ∈ T (1) , k = 1, 2, whose degrees add up to 2|f |, as the
explicit form of Φ(1) shows. Now f ⊗ f is zero on any summand with |ω(k) | 6= |f |,
while on summands where |ω(k) | = |f | is odd, one has
(f ⊗ f )(ω(1) ⊗A ω(2) − ω(2) ⊗A ω(1) ) = f (ω(1) )f (ω(2) ) − f (ω(2) )f (ω(1) ) = 0 ,
as A is commutative.
5.3. As just demonstrated, the construction of Φ(1) does not require A to be flat
over K, and we now show by explicit construction that one can find also without any
flatness assumption a co-unital diagonal approximation Φ(2) on T (2) that extends
the just constructed Φ(1) .
Recall that T (2) = T (1) hF2 i, with F2 a free Aev –module based on elements dfb
that correspond to polynomials fb ∈ I ⊆ P whose classes in I/I 2 generate that Aev –
module. An “attaching map” ∂|F2 : F2 → T (1) that killed the second homology
of T (1) was obtained in 3.7(∗∗) from some representation of each fb as K–linear
combination of monomials in the algebra generators of P over K.
In those terms, and extending the shorthand notation to dfb′ = dfb ⊗1 = j1 (dfb )
and dfb′′ = 1⊗dfb = j2 (dfb ), we now determine a diagonal approximation in degree
2.
14
R.-O. BUCHWEITZ AND C. ROBERTS
Proposition 5.4. The algebra homomorphism Φ(2) : T (2) → T (2) ⊗A T (2) that
extends Φ(1) by means of
Φ2 (dfb ) = dfb′ + dfb′′
X
−
X
αA x′a1 · · · x′aκ−1 dx′aκ xaκ+1 · · · xaλ−1 dx′′aλ x′′aλ+1 · · · x′′ak
A=(a1 ,...,ak ) 16κ<λ6k
P′
for some finite presentation fb (x) = A=(a1 ,...,ak ) αA xa1 · · · xak ∈ P of the given
polynomial fb defines a co-unital diagonal approximation.
Proof. Inspection reveals immediately that Φ(2) is co-unital, that is,
ǫ1 Φ(2) = idT (2) = ǫ2 Φ(2) .
It thus remains to verify that Φ(2) is compatible with the differentials, equivalently,
that ∂Φ2 (dfb ) = Φ(1) (∂dfb ) for each b.
This is a straightforward verification. First, we exhibit ∂(dfb′ ) and ∂(dfb′′ ) for
use below,
X
∂(dfb′ ) =
k
X
αA x′a1 · · · x′aκ−1 dx′aκ xaκ+1 · · · xak ,
A=(a1 ,...,ak ) κ=1
X
∂(dfb′′ ) =
k
X
αA xa1 · · · xaκ−1 dx′′aκ x′′aκ+1 · · · x′′ak .
A=(a1 ,...,ak ) κ=1
Next we apply the differential to the double sum in the definition of Φ2 (dfb ) and
obtain
X
X
∂
αA x′a1 · · · x′aκ−1 dx′aκ xaκ+1 · · · xaλ−1 dx′′aλ x′′aλ+1 · · · x′′ak
A=(a1 ,...,ak ) 16κ<λ6k
=
X
X
αA x′a1 · · · x′aκ−1 (xaκ − x′aκ )xaκ+1 · · · xaλ−1 dx′′aλ x′′aλ+1 · · · x′′ak
A=(a1 ,...,ak ) 16κ<λ6k
−
X
X
αA x′a1 · · · x′aκ−1 dx′aκ xaκ+1 · · · xaλ−1 (x′′aλ − xaλ )x′′aλ+1 · · · x′′ak
A=(a1 ,...,ak ) 16κ<λ6k
=
X
X
αA (xa1 · · · xaλ−1 − x′a1 · · · x′aλ−1 )dx′′aλ x′′aλ+1 · · · x′′ak
A=(a1 ,...,ak ) 16λ6k
−
X
X
αA x′a1 · · · x′aκ−1 dx′aκ (x′′aκ+1 · · · x′′ak − xaκ+1 · · · xak )
A=(a1 ,...,ak ) 16κ6k
= ∂(dfb′′ ) + ∂(dfb′ ) − Φ(1) (∂dfb ) ,
where the first equality uses that ∂ is a skew derivation of degree one, and the second
equality applies 3.6(∗) to the monomials xa1 · · · xaκ−1 , respectively, xaκ+1 · · · xak ,
while the final equality uses the form of ∂(dfb′ ) and ∂(dfb′′ ) as recalled above.
Reordering the terms, we obtain Φ(2) (∂dfb ) = ∂Φ(2) (dfb ) as required.
Note the following special property of the co-unital diagonal approximation just
constructed.
Corollary 5.5. The co-unital diagonal approximation Φ(2) is cocommutative in
the dfb , in that the automorphism of graded algebras σ : T (2) ⊗A T (2) → T (2) ⊗A
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
15
T (2) , ignoring differentials, that exchanges dfb′ ↔ dfb′′ , for each b, but leaves the
remaining variables unchanged, satisfies σ ◦ Φ(2) = Φ(2) .
Having determined one explicit co-unital diagonal approximation on T (2) , what
choices were involved? Provided T ⊗A T is exact in low degrees, we can easily
describe all co-unital diagonal approximations on T (2) that are algebra homomorphisms between algebras with divided powers..
Theorem 5.6. Let T be a Tate model of µ : Aev → A as before. The just exhibited
co-unital diagonal approximation Φ(2) can be modified to the DG algebra homomorphism Ψ : T (2) → T (2) ⊗A T (2) that respects divided powers, agrees in degree zero
with Φ(0) and is given on the algebra generators of T in degrees 1 and 2 through
Ψ(dxa ) = dx′a + dx′′a + ∂ωa
X
Ψ(dfb ) = Φ(2) (dfb ) +
∆b,a (x′ , x′′ )ωa + ∂ηb ,
a
where ∆b,a ∈ Aev are as in 3.6, the ωa are of degree 2 and the ηb are of degree 3 in
P
T (2) ⊗A T (2) , such that ∂ωa and a ∆b,a (x′ , x′′ )ωa + ∂ηb are in Ker ǫ1 ∩ Ker ǫ2 .
If T (2) ⊗A T (2) is exact in degrees 1 and 2, then these are the only possible
co-unital diagonal approximations that are homomorphisms of DG algebras with
divided powers and that agree with Φ(2) in degree zero. Moreover, in this case ωa
and ηb can already be chosen in Ker ǫ1 ∩ Ker ǫ2 .
Proof. That Ψ is a co-unital diagonal approximation along with Φ(2) is easily verified directly.
Conversely, if Ψ is some co-unital diagonal approximation on T (2) , then (Ψ −
(2)
Φ )(dxa ) is a cycle for each a and, if H1 (T (2) ⊗A T (2) ) = 0, it is a boundary,
thus, of the form ∂ωa for suitable ωa in degree two. Moreover, ∂ωa is necessarily a
cycle in Ker ǫ1 ∩ Ker ǫ2 , as both diagonal approximations are co-unital, and Lemma
4.2(2) shows that H1 of that complex vanishes too. Thus, ωa can be chosen in
Ker ǫ1 ∩ Ker ǫ2 .
P
If also H2 (T (2) ⊗A T (2) ) = 0, then the 2–cycles Ψ(dfb )−Φ(2) (dfb )− a ∆b,a ωa are
boundaries, thus, of the form ∂ηb for some form ηb of degree 3. Again, ∂ηb belongs
necessarily to Ker ǫ1 ∩Ker ǫ2 and H2 of that complex vanishes by the same argument
as before, whence we may replace ηb , if necessary, by an element in Ker ǫ1 ∩ Ker ǫ2
that has the same image under the differential.
To illustrate, we determine the cup product of two 1–cochains.
5.7. Assume T is a Tate model of µ : Aev → A for some commutative K–algebra
A. Let f, g : T → A be Aev –linear cochains of degree 1. Their cup product is then
a cochain of degree 2 with values in A, explicitly given by
(f ∪ g)(dxa1 ∧ dxa2 ) = µ(f ⊗A g)Φ(dxa1 ∧ dxa2 )
= µ(f ⊗A g)((dx′a1 + dx′′a1 ) ∧ (dx′a2 + dx′′a2 ))
= g(dxa1 )f (dxa2 ) − f (dxa1 )g(dxa2 )
16
R.-O. BUCHWEITZ AND C. ROBERTS
and
(f ∪ g)(dfb ) = µ(f ⊗A g)Φ(dfb )
X X
=
αA xa1 · · · xaκ−1 f (dxaκ )xaκ+1 · · · xaλ−1 g(dxaλ )xaλ+1 · · · xak
A 16κ<λ6k
=
X
X
A 16κ<λ6k
baλ xaλ+1 · · · xak f (dxaκ )g(dxaλ )
baκ xaκ+1 · · · xaλ−1 x
αA xa1 · · · xaκ−1 x
It remains to find a palatable form of the coefficients of f (dxaκ )g(dxaλ ) in that
expression.
Theorem 5.8. Given a commutative K–algebra A, represented as A ∼
= K[x; x ∈
X]/(fb )b . If for a given b, the polynomial fb depends on the variables x1 , ..., xn ∈ X,
one may choose a co-unital diagonal approximation Φ(2) on T (2) in such a way that
X ∂ (2) fb (x)
mod I f (dxi )g(dxj ) ∈ A ,
(f ∪ g)(dfb ) =
∂xi ∂xj
16i6j6n
ev
for any A –linear cochains f, g : T → A of degree one.
P
e1
en
Proof. Let fb =
e∈Nn αe x1 · · · xn with ae ∈ K be the presentation of fb as
polynomial in the variables x1 , ..., xn . Rewrite the occurring monomials as
αe x1e1 · · · xenn = αe xa1 · · · xa|e| ,
where
xa1 = · · · = xae1 = x1 ,
xae1 +1 = · · · = xae1 +e2 = x2 ,
xae1 +···+en−1 +1
..
.
= · · · = xae1 +···+en = xn .
In the corresponding choice of Φ2 (dfb ) in Proposition 5.4, this term contributes
X
αe x′a1 · · · x′aκ−1 dx′aκ xaκ+1 · · · xaλ−1 dx′′aλ x′′aλ+1 · · · x′′a|e| .
16κ<λ6|e|
Specializing to xa = x′a = x′′a for each index a, and sorting reduces this sum to
n
X
ei e1
ei−1 ei −2 ei+1
x1 · · · xi−1
xi xi+1 · · · xenn dx′i dx′′i
αe
2
i=1
X
ej−1 ej −1 ej+1
ei−1 ei −1 ei+1
+
xj
xj+1 · · · xenn dx′i dx′′j
ei ej xe11 · · · xi−1
xi xi+1 · · · xj−1
16i<j6n
The coefficient of dx′i dx′′j in this expression is indeed the corresponding divided
second derivative of the monomial, whence, putting the terms together again, it
follows that for this choice of presentation,
X ∂ (2) fb (x)
Φ2 (dfb )(x, x, x) = dfb′ + dfb′′ −
dx′i dx′′j
∂xi ∂xj
16i6j6n
and then the cup product takes the form claimed.
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
17
6. Proof of Theorem 2.5
Let A = P/I be as before a presentation of a commutative K–algebra as quotient
of a polynomial ring P = K[x; x ∈ X] over some commutative ring K modulo an
ideal I = (fb )b∈Y ⊆ P . We need to show that the Hessian quadratic map
X
Y
hf =
hfb ∂fb : DerK (P, A) →
A∂fb
b∈Y
b∈Y
as defined in 2.4 sends DerK (A, A) ⊆ DerK (P, A) to NA/P = HomA (I/I 2 , A) ⊆
Q
b∈Y A∂fb .
Q
an element in the normal
Now an element (ub ∂fb )b ∈ b∈Y A∂fb , represents
P
module NA/P if, and only if, for every relation P
b zb fb = 0, with zb ∈ P and only
finitely many of these elements nonzero, the sum b zb ub is zero in A.
Let X ′ = {x1 , ..., xN } ⊂ X be the finite subset of variables that are involved in
the finitely many polynomials fb that occur with nonzero coefficient in the given
relation. P
P
∂
∂
If D = x∈X ax ∂x
∈ DerK (A, A) is given, set D′ = x∈X ′ ãx ∂x
where ãx ∈ P
lifts ax ∈ A. One then has D′ (fb ) ≡ D(fb ) = 0 mod I, and
q(D)b =
X
16i6j6N
∂ (2) fb (x)
ai aj mod I ≡ h̃f (D′ )(fb ) mod I
∂xi ∂xj
for those fb that are involved in the relation. Next note that the second divided
P
(2)
power D′(2) = 16i6j6N ∂x∂i ∂xj ãi ãj of D′ satisfies
D′(2) (f g) = D′(2) (f )g + D′ (f )D′ (g) + f D′(2) (g)
for any polynomials f, g ∈ P .
P
Apply D′(2) to the relation 0 = b zb fb to obtain
!
X
X
0 = D′(2)
D′(2) (zb fb )
zb fb =
b
b
=
X
(D′(2) (zb )fb + D′ (zb )D′ (fb ) + zb D′(2) (fb ))
b
≡
X
zb q(D)b mod I
b
as in the middle equation, each term D′(2) (zb )fb is in I and D′ (fb ) ≡ D(fb ) =
0 mod I as D is a derivation on A. Therefore, q(D) = (q(D)b )b is indeed an
element of the normal module. This proves part (1) of Theorem 2.5.
Part (2) of Theorem 2.5 just restates a well-known fact, but we give another
deduction using the Tate model T of the multiplication map µ : Aev → A and
establish along the way as well Part (3) of Theorem 2.5.
To this end, recall, from 5.2 above, that T (1) is just the Koszul complex on (x′′a −
xa )a in Aev and that its dual HomAev (T (1) , A) carries the cup product induced by
Φ(1) with respect to which its cohomology H • (HomAev (T (1) , A)) ∼
= HomA (Ω•P/K , A)
is a strictly graded commutative algebra.
As the diagonal approximation Φ on T extends the diagonal approximation Φ(1)
on T (1) , the A–linear surjective restriction map p : T ∗ = Hom•Aev (T , A) → T (1)∗ =
Hom•Aev (T (1) , A) resulting from the inclusion T (1) ⊂ T is compatible with the
18
R.-O. BUCHWEITZ AND C. ROBERTS
cup products on source and target, thus, induces a K–algebra homomorphism in
cohomology.
If K is the kernel of the restriction map on the complexes, then the long exact
cohomology sequence contains
/ H 2 (K)
H 1 (T (1)∗ )
DerK (P, A)
jac
/ NA/K
/ H 2 (T ∗ )
δ
/ Ext2Aev (A, A)
H 2 (p)
/ H 2 (T (1)∗ )
/ HomA (Ω2 , A) .
P/A
If now D ∈ DerK (A, A) = H 1 (Hom•Aev (T , A)) then H 2 (p)(D ∪ D) = 0, as H • (p) is
an algebra homomorphism into a strictly graded commutative algebra. Thus, D ∪D
is in the image of the map δ : NA/P → Ext2Aev (A, A), and this image is known to
1
be TA/K
, the first André–Quillen cohomology of A over K. Moreover, the explicit
description of the cup product on HomAev (T (2) , A) shows that q(D) ∈ NA/P is
mapped via δ to D ∪ D.
7. Hochschild Cohomology of Homological Complete Intersections
Now we consider the case that A = P/I is a homological complete intersection
over K with I/I 2 free.
Proposition 7.1. (see [7] and [4, 25]) Assume the commutative K–algebra A admits a presentation A = P/I, with P = SymK F a symmetric K–algebra on a free
K–module F , and I ⊆ P an ideal.
K
(2)
If TorK
as constructed in Corol+ (A, A) = ⊕i>0 Tori (A, A) = 0, then T
ev
lary 3.5 is already a Tate model of µ : A → A if, and only if, I/I 2 is a
free
the natural homomorphism of strictly graded commutative algebras
V A–module,
P
2
I/I
→
Tor
(A, A) is an isomorphism, and the surjection P ⊗K G ⊗K P → I
A
was chosen to induce a bijection from a K–basis of G to an A–basis of I/I 2 .
Proof. In Corollary 3.5 we formed T (1) as the Koszul complex on (x′′a − x′a )a in Aev ,
where xa runs through a K–basis of F . Its homology can be identified as
∼ TorP ⊗K P (P, A ⊗K A) .
H(T (1) ) =
By [15, Chap. XVI§5(5a)], flatness of P over K together with TorK
+ (A, A) = 0
P ⊗K P
P
∼
yields an isomorphism Tor
(P, A ⊗K A) = Tor (A, A).
V
If I/I 2 is free and the canonical map A I/I 2 → TorP (A, A) ∼
= H(T (1) ) is an
(2)
= T , provided
isomorphism of algebras, then already Tate [37] showed that T
the map G → I/I 2 sends a K–basis to an A–basis. The main result of [7] yields
the converse.
Remark 7.2. This proposition applies in particular when I = (f1 , ..., fc ) ⊆ P =
K[x1 , ..., xn ] defines a complete intersection, in the sense that the fj form a Koszul–
regular sequence in P . In this case, the condition TorK
+ (A, A) = 0 is equivalent to
exactness of the Koszul complex on f1 (x′ ), ..., fc (x′ ), f1 (x′′1 ), ..., fc (x′′ ) in P ⊗K P ∼
=
K[x′ , x′′ ], in that this complex is simply the tensor product of the Koszul complex
on f1 , ..., fc in P , a free resolution of A over K, with itself over K.
The resolution of a complete intersection ring over its enveloping algebra through
the Tate model has a long history.
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
19
Wolffhardt [39] was the first to exhibit it, but he apparently was unaware of
Tate’s result and wrote down the resolution in case K is a field of characteristic
zero — in which case divided powers can be replaced by symmetric ones at the
expense of introducing denominators.
In explicit form, this resolution was also described in [10, 12, 22]. As the reviewers of [26] in MathSciNet note: “...the authors recover the calculation of
HH∗ (A/K, A) due originally to K. Wolffhardt [Trans. Amer. Math. Soc. 171
(1972), 5166; MR0306192 (46 #5319)], and more recently (since 1988) to a host of
other authors, including the reviewers.”
To see that the condition TorK
+ (A, A) = 0 is not void even in this special situation, contrary to what is implied in [12] or [22], one may just take c = 1 with, say,
K = Z, and f1 = 17x ∈ Z[x]. See also the next section for a detailed discussion of
this point.
Remark 7.3. Wolffhardt [39] treats the analytic case, when A is quotient of a
noetherian smooth analytic algebra in the broad sense of [1, Supplement]. We
leave it to the readers to retrace the foregoing results and to convince themselves
that everything works as well in that situation, provided one replaces the occurring
tensor products by their analytic counterparts.
We now turn to the
Proof of Theorem 2.11. In light of 5.7 and Theorem 5.8, we know explicitly the
form of the cup product on HomAev (T (2) , A), thus, on HomAev (T , A) in the case at
hand.
F . As a complex, Hom•Aev (T (2) , A)
Set F ∗ = HomAev (F, A) for any Aev –module
n
Pc ∂fj
is the Koszul complex on
in the polynomial ring A[∂ f1 , ..., ∂ fc ].
j=1 ∂xi ∂ fj
i=1
Here {∂ fj }j=1,..,c ∈ F2∗ form the dual A–basis to {dfj }j=1,..,c ⊆ F2 and we employ
the A–basis ∂ x1 , ..., ∂ xn ∈ F1∗ dual to {dxi }i=1,...,n ⊆ F1 . That Koszul complex
has then the compact description
^
Hom•Aev (T (2) , A) =
(F1∗ [−1]) ⊗A SymA (F2∗ [−2])
A
Pc
∂f
with differential ∂(∂ xi ) = j=1 ∂xji ∂ fj and ∂∂ fj = 0. The cohomological grading
on Hom•Aev (T (2) , A) is recovered from putting ∂ xi into degree 1 and ∂ fj into degree
2 as indicated.
It remains to determine the multiplicative structure.
V2
Identify {∂ xi1 ∧ ∂ xi2 }i2 <i1 ⊆ A F1∗ with the A–basis dual to {dxi2 ∧ dxi1 }i2 <i1
V2
in A F1 . The cup product resulting from the diagonal approximation in Theorem
5.8 is then explicitly given by
Pc
∂ (2) fj
∂ xi1 ∧ ∂ xi2 + j=1 ∂xi1 ∂xi2 ∂fb if i1 < i2 ,
(2)
∂ xi1 ∪ ∂ xi2 = Pc ∂ 2 fj ∂fj
if i = i1 = i2 ,
j=1 ∂ xi
∂
if i > i .
∧∂
xi1
xi2
1
2
Moreover, the elements ∂fj are central, the co-unital diagonal approximation Φ
being cocommutative in the dfj as pointed out in Corollary 5.5. These explicit
calculations show that Hom•Aev (T , A) with the cup product ∪ = ∪Φ from Theorem
20
R.-O. BUCHWEITZ AND C. ROBERTS
5.8 satisfies the relations
∂ xi ∪ ∂ xi =
∂ xi1 ∪ ∂ xi2 + ∂ xi2 ∪ ∂ xi1 =
c
X
∂ (2) fj
j=1
c
X
j=1
∂ 2 xi
∂fj ,
∂ 2 fj
∂fj ,
∂xi1 ∂xi2
∂ xi ∪ ∂ fj = ∂ fj ∪ ∂ xi ,
∂ fj1 ∪ ∂ fj2 = ∂ fj2 ∪ ∂ fj1 .
for i, i1 , i2 = 1, ..., n and j, j1 , j2 = 1, .., c.
Putting it together, this shows (Hom•Aev (T , A), ∪) ∼
= Cliff(q) as graded algebras
and the differential on Hom•Aev (T , A) readily identifies with ∂jacf on Cliff(q).
This completes the proof of Theorem 2.11.
8. The Case of One Variable
8.1. Here we consider cyclic extensions A = K[x]/(f ), with K some commutative
ring, f = a0 xd + · · · + ad ∈ P = K[x] a polynomial. We set I = (f ) and abbreviate
∂x f = ∂f
∂x . Let c = cf = (a0 , ..., ad ) ⊆ K be the content ideal of f . By [27,
5.3 Thm. 7], the polynomial f is a non-zero-divisor in K[x] if, and only if, the
annihilator of the content ideal in K is zero, equivalently, HomK (K/c, K) = 0.
To provide some context, we quote the following results from the literature and
refer to the references cited below for unexplained terminology.
Proposition 8.2. Let f = a0 xd + · · · + ad ∈ K[x] be a polynomial as above and
set A = K[x]/(f ).
(1) A is free as K–module if, and only if, the ideal I is generated by a monic
polynomial. A is then necessarily free of finite rank.
(2) A is finitely generated projective as K–module if, and only if, the ideal I is
generated by a quasi-monic polynomial.
(3) A is projective as K–module if, and only if, the ideal I is generated by an
almost quasi-monic polynomial.
(4) A is flat as K–module if, and only if, the content ideal c of f is generated in
K by an idempotent.
In cases (1) and (2), f is a non-zero-divisor in K[x], while in cases (3) and (4),
the content ideal is generated by an idempotent e ∈ K, cf = (e). With K1 =
K/(1 − e), K2 = K/(e), so that K = K1 × K2 , the polynomial ef is then a nonzero-divisor in K1 [x], while (1 − e)f = 0, and A ∼
= K1 [x]/(ef ) × K2 [x].
Proof. Parts (1), (2) and (3) can be found in [9], where also the terminology “(almost) quasi–monic” is introduced, while part (4) is contained in [28, Thm. 4.3].
Examples 8.3. The polynomial f = 15xd + 10x + 6, for d > 2, is quasi-monic over
K = Z/30Z. The polynomial f = 15xd + 6, for d > 1, is almost quasi-monic, but
not quasi-monic over K = Z/30Z. The ring A = Z[x]/(2x − 1) ∼
= Z[ 12 ] is flat as
module over K = Z, but not projective.
8.4. In each of the cases considered in Proposition 8.2, we have a decomposition
K = K1 × K2 and A ∼
= A1 × A2 , where A1 = K1 [x]/(f ) with f a non-zero-divisor
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
21
in K1 [x] and A2 = K2 [x] a polynomial ring. Moreover, as algebras,
∼ ExtAop ⊗ A (A1 , A1 ) × ExtAop ⊗ A (A2 , A2 ) ,
ExtAev (A, A) =
1
K1
1
2
K2
2
with the second factor satisfying
∂
(A2 , A2 ) ∼
ExtAop
[−1]
= K2 [x][0] ⊕ K2 [x] ∂x
2 ⊗K2 A2
with obvious multiplicative structure.
We will therefore now concentrate on the case that the given polynomial f ∈ K[x]
is a non-zero-divisor.
8.5. If f is a non-zero-divisor, the natural map A → I/I 2 , 1 7→ f mod I, is an
f
isomorphism, and 0 → P −
→ P → A → 0 is aVfree resolution of A both over P and
K, so that TorK
(A,
A)
=
0 for i > 2 and A I/I 2 → TorP (A, A) is trivially an
i
isomorphism.
In other words, I = (f ) ⊆ P is a regular ideal. The only missing piece for A to
be a homological complete intersection over K is thus the vanishing of TorK
1 (A, A).
That vanishing is equivalent to f ⊗ 1 ∈ P ⊗K A being a non-zero-divisor, in turn
equivalent to 1 ⊗ f, f ⊗ 1 forming a regular sequence in P ⊗K P .
In light of [27, 5. Exercise 9], the sequence 1 ⊗ f, f ⊗ 1 is regular in P ⊗K P if,
and only if, ExtiK (K/c, K) = 0 for i = 0, 1, that is, the depth of the content ideal c
on K is at least 2.
Example 8.6. Let R[s, t, u, v] be polynomial ring over a commutative ring R, and
set K = R[s, t, u, v]/(sv − tu). For d > 1, the polynomial f (x) = sxd + v ∈ K[x]
has content ideal cf = (s, v) of depth 2, while g(x) = sxd + u ∈ K[x] has content
ideal cg = (s, u) of depth 1.
Both f, g are thus non-zero-divisors in K[x], but only f defines a homological
complete intersection A = K[x]/(f ) over K.
8.7. Constructing the beginning of the Tate resolution of the multiplication map
µ : Aev ∼
= K[x′ , x′′ ]/(f (x′ ), f (x′′ )) → A, µ(x′ ) = µ(x′′ ) = x, as in Corollary 3.5, one
obtains T (2) as a 2–periodic complex, augmented versus A in degree 0,
0O o
AO o
0O o
0O o
0O o
···
≡
µ
0o
AO
ǫ
′′
′
ev o x −x
A
Aev o
∆
′′
′
ev o x −x
A
Aev o
···
≡
T (2)
where ∆ ∈ Aev is the residue class of the unique polynomial ∆(x′ , x′′ ) ∈ K[x′ , x′′ ]
that satisfies f (x′′ ) − f (x′ ) = (x′′ − x′ )∆(x′ , x′′ ).
Summarizing the discussion so far, we have the following characterizations, where
we refer to [4, 8] for the terminology and properties of exact zero divisors.
Theorem 8.8. If f = a0 xd + · · · + ad is a non-zero-divisor in K[x], then the
following are equivalent.
(1) A = K[x]/(f ) is a homological complete intersection over K.
(2) T (2) resolves A as Aev –module.
(3) (x′′ − x′ , ∆) is a pair of exact zero-divisors in Aev .
(4) The content ideal c of f has grade at least 2, in that HomK (K/c, K) = 0 =
Ext1K (K/c, K).
22
R.-O. BUCHWEITZ AND C. ROBERTS
Moreover, A will be flat over K if, and only if the content ideal of f is the unit
ideal in K.
Proof. The equivalence (1)⇐⇒ (2) is a special case of Proposition 7.1, while (2)⇐⇒
(3) is a tautology, in that (x′′ − x′ , ∆) is, by definition, a pair of exact zero-divisors
in Aev if, and only if, the ideals (x′′ − x′ ) and (∆) are each other’s annihilators
in Aev , if, and only if, the 2–periodic complex T (2) displayed above is exact. The
equivalence (3)⇐⇒ (4) has been discussed in 8.5 above.
The final claim follows from Proposition 8.2 in that flatness means that cf is
generated by an idempotent. If that idempotent is not 1, then f is not a non-zerodivisor in K[x].
Assume for the remainder of this section that f and A satisfy the equivalent
conditions in Theorem 8.8. We determine the algebra Ext•Aev (A, A).
8.9. Call A′ = K[x]/ (f (x), ∂x f ), the ramification algebra of f and introduce as
well the annihilator of ∂x f in A, given by
θ = (f :P ∂x f ) /(f ) ⊆ A .
Note that θ(∂x f ) = 0 in A, whence θ is naturally an A′ –module.
In fact, θ ∼
= HomA (A′ , A), as dualizing the exact sequence
0
/θ
/A
∂x f
/A
/ A′
/0
of A–modules shows. The same sequence can be read to identify θ ∼
= DerK (A, A)
1
, the first André–Quillen (tangent) cohomology of A over K.
and A′ ∼
= TA/K
Theorem 8.10. Assume f ∈ K[x] is a polynomial whose content ideal c is of depth
at least two.
The Yoneda Ext–algebra of A = K[x]/(f ) as Aev –module is then graded commutative and satisfies
Ext•Aev (A, A) ∼
= A × A′
(A′ ⊕ θt)[s]
,
(at)(bt) − abf (2) s; a, b ∈ θ
where A, A′ and θ are in degree zero, t is of degree 1, while s is central of degree 2,
and f (2) denotes the residue class of the second divided derivative of f in A′ .
2i
In particular, the even Yoneda Ext–algebra Exteven
Aev (A, A) = ⊕i>0 ExtAev (A, A)
′
′
is the fibre product of the ring A over A with the polynomial ring A [s] in one
variable of degree 2,
′
∼
Exteven
ev (A, A) = A ×A′ A [s] .
A
Proof. Just note that the algebra displayed is indeed the cohomology algebra of
the DG Clifford algebra Aht, si with relations t2 = f (2) s, ts = st, and differential
∂(t) = (∂x f )s, ∂s = 0. Now apply Theorem 2.11.
Remark 8.11. While we know from graded commutativity of the algebra that the
classes abf (2) , for a, b ∈ θ must be 2–torsion in A′ , we can do better, in that these
classes are already 2–torsion in A.
To see this, observe that 2f (2) = ∂ 2 f /∂x2 , the ordinary second derivative of f ,
and use that a∂x f = uf, b∂x f = vf for suitable polynomials u, v ∈ K[x], because
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
23
a, b ∈ θ. One then obtains
ab
∂2f
= a (∂x (b∂x f ) − ∂x b · ∂x f )
∂x2
= a∂x (vf ) − a∂x b · ∂x f
by the product rule,
using b∂x f = vf ,
= a(∂x v)f + av∂x f − a∂x b · ∂x f
by the product rule again,
= (a∂x v + uv − u∂x b) f
using a∂x f = uf ,
≡ 0 mod f.
Corollary 8.12. If 2 is a non-zero-divisor in A = K[x]/(f ), for f a polynomial in
K[x] with content ideal of depth at least two, then Ext•Aev (A, A) is strictly graded
commutative.
Example 8.13 (Inspired by [31, 19]). The foregoing Corollary fails in more than
one variable. Consider A = Z[x, y]/(x2 − 4xy + y 2 − 1) as a Z–algebra. As f =
x2 − 4xy + y 2 − 1 is monic in y of degree 2, the Z[x]–module A is free of rank 2, in
particular, 2 is a non-zero-divisor on A and A is projective over Z.
∂f
The Euler identity yields x ∂f
∂x +y ∂y = 2f +2 ≡ 2 in A, whence the Jacobi ideal of
1
∼
A over Z is (2x−4y, −4x+2y) = 2A, and so TA/Z
= A/(2A) ∼
= Z2 [x, y]/((x+y+1)2 )
is a nonzero 2–torsion
A–module.
∂f ∂
∂
∂
∂
For D = 12 ∂f
−
∂y ∂x
∂x ∂y = (2x−y) ∂x +(x−2y) ∂y one verifies that D(f ) = 0,
already in Z[x, y]. Using Theorem 2.5, one finds
∂2f
∂ (2) f
∂ (2) f
+ (2x − y)(x − 2y)
+ (x − 2y)2
2
∂x
∂x∂y
∂y 2
= (2x − y)2 − 4(2x − y)(x − 2y) + (x − 2y)2
q(D)(f ) = (2x − y)2
≡ y 2 + x2 ≡ 1 mod (f, 2) ,
whence q(D)(f ) is not 2–torsion in the normal module NA/P ∼
= A.
1
Further, D ∪ D ≡ 1 (mod 2A) does not vanish in TA/K
, in fact, the squaring
1
map sends Ext1Aev (A, A) ∼
, and one may verify
= DerZ (A, A) surjectively onto TA/K
2
2
1
∼ Ext ev (A, A) ∼
HH
(A/K,
A).
easily that further TA/K =
=
A
We finish this section by looking at some classical situations.
The Separable or Unramified Case. Recall that a polynomial f (x) ∈ K[x] is
separable over K, if its derivative ∂x f is a unit modulo f , equivalently, (f, ∂x f ) =
K[x]. In that case, the content ideal cf necessarily equals K, whence f is a nonzero-divisor, A = K[x]/(f ) is flat over K, and Theorem 8.10 applies. We retrieve
that all higher extension groups vanish, Ext>0
Aev (A, A) = 0, as it should be.
The Generically Unramified Case. The polynomial f is generically unramified,
if it is a non-zero-divisor in K[x] and its derivative ∂x f is a non-zero-divisor modulo
f . Assuming furthermore that A is flat over K, equivalently, that the content of f
is the unit ideal, Theorem 8.10 applies and shows that f is generically unramified if,
and only if, Ext1Aev (A, A) = 0, equivalently, all odd extension groups Ext2i+1
Aev (A, A)
vanish. In that case, the Yoneda Ext–algebra is thus reduced to its even part, the
fibre product of A over A′ with a polynomial ring A′ [z] generated in cohomological
degree 2.
24
R.-O. BUCHWEITZ AND C. ROBERTS
There is no room for a non-zero squaring map and so the Ext–algebra is strictly
graded commutative.
The Totally Ramified Case. The polynomial f is totally ramified, if it is a nonzero-divisor in K[x], but its derivative ∂x f is zero modulo f . Equivalently, A′ = A.
Assuming that the content of f is the unit ideal, Theorem 8.10 shows that the
Yoneda algebra Ext•Aev (A, A) is the Clifford algebra over the A–quadratic map
q : At → As, q(t) = f (2) (x)s. If f (2) (x) = 0 in A, as happens, for example, by
Remark 8.11 if 2 is a non-zero-divisor in A, then the Yoneda algebra is the tensor
product of the exterior algebra over A generated by t in degree 1 with the symmetric
algebra over A generated by s in degree 2.
In contrast, if f (2) (x) is a unit in A, whence A is necessarily of characteristic 2
by Remark 8.11, then the Yoneda algebra is the tensor algebra over A on a single
generator t in degree 1. This covers the case A = K[x]/(x2 ) with char K = 2
mentioned in the Introduction.
Polynomials over a field. Last, but not least, let us consider the case when
the coefficient ring K is a field. Then being a non-zero-divisor simply means that
f is not the zero polynomial. Trivially, A is projective over K, thus the Yoneda
Ext–algebra of self-extensions agrees with Hochschild cohomology.
Further, K[x] being a principal ideal domain, we obtain explicit descriptions of
A′ and θ. Set g = gcd(f, ∂x f ) and h = f /g in K[x], so that A′ ∼
= K[x]/(g) and
θ = (f (x) : ∂x f )/(f ) ∼
= (h)/(f ). Thus, θ is the principal ideal generated by h
in K[x]/(f ), isomorphic to A′ as A–module. Therefore, we obtain the following
presentation
HH• (A/K) ∼
=
K[x, y, z]
f, gy, gz, y 2 + f (2) h2 z
of the Hochschild cohomology ring of A = K[x]/(f ) over K as a quotient of a
graded polynomial ring, where x is of degree 0, y is the class of h∂x of degree 1,
and z is of degree 2.
If char K 6= 2, then necessarily f (2) h2 ≡ 0 (mod f ) by Remark 8.11 and the
description simplifies accordingly.
Remark 8.14. This last example encompasses Theorem 3.2, Lemma 4.1, Lemma
5.1, Theorem 5.2 and Theorem 6.2 from [23], completes the characteristic 2 cases
that were not handled there, and covers as well Theorem 3.9 from [36].
9. Cohomology of Finite Abelian Groups
9.1. As our final example, we consider the Hochschild and ordinary cohomology of
finite abelian groups. If G = µn1 × · · · µnr is such a group, with µn the multiplicatively written cyclic group of order n, we may assume that the orders of the factors
satisfy 2 6 n1 | n2 | · · · | nr , so that the ni are the elementary divisors or invariant
factors of the group.
9.2. The group algebra on G over a commutative ring K is then given by A =
KG ∼
= K[x1 , ..., xr ]/(xn1 1 − 1, ..., xnr r − 1). It is a homological complete intersection
over K, thus, Theorem 2.11 applies. However, because the variables xi become
units in A, one can simplify that result in this case, and recover as well the fact
established in [17] that there is an isomorphism of graded commutative algebras
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
25
HH• (A/K) ∼
= A⊗K Ext•A (K, K), where K is considered an A–module via the trivial
representation of G.
Theorem 9.3. Let G = µn1 × · · · µnr be a finite abelian group as above. The
Hochschild cohomology of KG over K is then the cohomology of the tensor product
of differential graded Clifford algebras,
!
r
O
Khτj , σj i
, ∂τj = nj σj , ∂σj = 0 ,
(C(G), ∂) = KG ⊗K
(τj2 − mj σj , τj σj − σj τj )
j=1
where
mj =
(
0
nj /2
if nj is odd,
if nj is even ,
and the (cohomological) degree of the τj is 1, that of the σj is 2.
Further, the cohomology of the second factor in the tensor product is the group
cohomology H • (G, K) = Ext•KG (K, K), where K is the trivial G–representation,
and one retrieves the isomorphism from [17] of graded commutative algebras
HH• (KG/K) ∼
= KG ⊗K H • (G, K) .
Proof. Note that A = KG ∼
= Kµn1 ⊗K · · · ⊗K Kµnr and that accordingly the
multiplication map µ : Aev → A is the tensor product over K of the multiplication
maps µj : (Kµj )ev → Kµj .
It follows that a Tate model for A over K is obtainable as tensor product of
the corresponding Tate models of the cyclic groups involved. Thus, it remains to
discuss the Tate model of a single cyclic group, say µn , for some natural number
n. The group algebra Kµn ∼
= K[x]/(xn − 1) is a cyclic extension of K, free as
K–module. Therefore, we can apply 8.7 with f = xn − 1 ∈ K[x] to obtain the Tate
model T (2) = T as a 2–periodic complex, augmented versus Kµn in degree 0,
0O o
0O o
KµO n o
0O o
0O o
···
≡
µ
0o
Kµ
n
O
ǫ
x
Kµn ev o
′′
−x
′
Kµn ev o
∆
x
Kµn ev o
′′
−x
′
Kµn ev o
···
≡
T (2)
where ∆ ∈ Kµn ev is the residue class of the unique polynomial
∆ = ∆(x′ , x′′ ) =
n−1
X
(x′′ )n − (x′ )n
(x′′ )i (x′ )n−1−i ∈ K[x′ , x′′ ]
=
′′
′
x −x
i=0
that satisfies f (x′′ )−f (x′ ) = (x′′ −x′ )∆(x′ , x′′ ) in K[x′ , x′′ ]. Written as DG algebra,
^
T = Kµn ev hdx, df i ∼
(dx) ⊗Kµn ev ΓKµn ev (df ) ,
=
ev
Kµn
with dx in (homological) degree 1 and df in degree 2 and differential ∂(df ) =
∆(x′ , x′′ )dx and ∂(dx) = x′′ − x′ .
Further, the co-unital diagonal approximation Φ : T → T ⊗Kµn T from Proposition 5.4 is determined by
Φ1 (dx) = dx′ + dx′′
Φ2 (df ) = dx′ + dx′′ −
X
(x′ )p (dx′ )xq (dx′′ )(x′′ )r
p+q+r=n−2
26
R.-O. BUCHWEITZ AND C. ROBERTS
in our earlier notation. Next we adapt this co-unital diagonal approximation using
Theorem 5.6. Setting x = x′ = x′′ , we have
n n−2 ′ ′′
′
′′
Φ2 (df ) ≡ dx + dx −
x
dx dx mod (x′′ − x, x − x′ ) .
2
Now the form dx′ df ′′ ∈ T ⊗Kµn T of degree 3 is in the kernel of both augmentation
maps ǫ1,2 and satisfies
∂(dx′ df ′′ ) = (x − x′ )df ′′ − dx′ ∆(x, x′′ )dx′′
≡ nxn−1 dx′ dx′′ mod (x′′ − x, x − x′ ) .
As x is a unit in Kµn , we can adjust Φ2 by adding ∂η, where
(
n−1
(x′ )−1 dx′ df ′′ if n is odd,
η = n 2 ′ −1 ′ ′′
dx df
if n is even,
2 (x )
to obtain the adapted co-unital diagonal approximation Ψ that agrees in degrees
0, 1 with Φ, but in degree 2 is given by
Ψ(df ) = Φ2 (df ) + ∂η ≡ dx′ + dx′′ mod (x′′ − x, x − x′ )
if n is odd,
n n−2 ′ ′′
x
dx dx mod (x′′ − x, x − x′ )
2
if n is even. (This adaptation of the multiplicative structure in case of a cyclic
group already occurs in [15, Chap. XII §7, p.252].)
At this stage, HomKµn ev (T , Kµn ) with the cup product ∪Ψ and original differential is the DG Clifford algebra
Ψ(df ) = Φ2 (df ) + ∂η ≡ dx′ + dx′′ −
HomKµn ev (T , Kµn ) ∼
=
Kµn ht, si
,
(t2 − mxn−2 s, st − ts)
where m = 0 for n odd, m = n/2 for n even, while the differential satisfies ∂(t) =
nxn−1 s and ∂s = 0.
For the final touch, set τ = xt, σ = xn s and note that this coordinate change is
indeed invertible, as x is a unit in Kµn . In terms of τ, σ we obtain
Kµn hτ, σi
(HomKµn ev (T , Kµn ), ∂) ∼
,
∂(τ
)
=
nσ,
∂σ
=
0
.
=
(τ 2 − mσ, στ − τ σ)
Finally, taking the tensor product over K of the Tate models for the cyclic groups
involved, their adapted diagonal approximations fit together to endow the tensor
product with a co-unital diagonal approximation. Dualizing into KG over KGev
results in the DG Clifford algebra (C(G), ∂) in the statement of the theorem.
To establish the relation to group cohomology, note first that
!
r
O
Khτ
,
σ
i
j
j
H(C(G), ∂) ∼
= KG ⊗K H
2 − m σ , τ σ − σ τ ) , ∂τj = nj σj , ∂σj = 0
(τ
j
j
j
j
j
j
j
j=1
as KG is flat over K. Secondly, observe that applying K ⊗KG (−) to the Tate model
T of µ : KGev → KG and then taking cohomology of HomKG (K ⊗KG T , K ⊗KG T )
yields a homomorphism of graded algebras
K⊗KG (−)
Ext•KGev (KG, KG) −−−−−−−−→ Ext•KG (K, K)
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
27
that factors through K ⊗KG Ext•KGev (KG, KG). Thus, one obtains a homomorphism of graded algebras
!
r
O
Khτ
,
σ
i
j
j
−→ Ext•KG (K, K) ,
(⋄)
H
2 − m σ , τ σ − σ τ ) , ∂j
(τ
j
j
j
j
j
j
j
j=1
where ∂j τj = nj σj , ∂j σj = 0.
One may check directly that this is an isomorphism by noting that one may
construct a Tate model for the augmentation KG → K that returns the argument
of H in the source of this homomorphism and that this is, at least, a K–linear
isomorphism, if one initially ignores the multiplicative structure.
Alternatively, one may use [33] to observe that the structure map K → KG
induces an algebra homomorphism
ExtKG (K, K)−→KG ⊗ ExtKG (K, K) ∼
= HH(KG/K)
that provides a splitting to the K–algebra homomorphism (⋄).
Remark 9.4. One obstacle to an even more transparent proof that (⋄) is an
isomorphism is that the co-unital diagonal approximation on the Tate model for
µ : KGev → KG does not obviously induce such an approximation when applying
K ⊗KG (−). This obstacle can be overcome by exhibiting directly such a co-unital
diagonal approximation on the Tate model for the augmentation KG → K, as is
done in [30], so that it will agree with the one on K ⊗KG C(G) exhibited above.
An explicit co-unital diagonal approximation on the Tate model for the augmentation KG → K has also been described in [24] and we refer to that reference and
[30] for explicit information on the structure of H(G, K), augmenting and reproducing results from [16, 38].
References
[1] André, M.: Homologie des algèbres commutatives. Die Grundlehren der mathematischen
Wissenschaften, Band 206. Springer-Verlag, Berlin-New York, 1974. xv+341 pp. 19
[2] Avramov, L. L.: Local algebra and rational homotopy. Algebraic homotopy and local
algebra (Luminy, 1982), 15–43, Astérisque, 113-114, Soc. Math. France, Paris, 1984. 5
[3] Avramov, L. L.: Infinite free resolutions. Six lectures on commutative algebra (Bellaterra,
1996), 1–118, Progr. Math., 166, Birkhäuser, Basel, 1998. 7
[4] Avramov, L. L.; Bonacho Dos Anjos Henriques, I. B.; Şega, L. M.: Quasi-complete intersection homomorphisms. preprint 2010, 20 pages, arXiv:1010.2143 5, 18, 21
[5] Avramov, L. L.; Vigué-Poirrier, M.: Hochschild homology criteria for smoothness. Internat. Math. Res. Notices 1992, no. 1, 17–25. 7
[6] Bergeron, N.; Wolfgang, H.L.: The decomposition of Hochschild cohomology and Gerstenhaber operations. J. Pure Appl. Algebra 104 (1995), no. 3, 243–265. 6
[7] Blanco, A.; Majadas, J.; Rodicio, A. G.: On the acyclicity of the Tate complex. J. Pure
Appl. Algebra 131 (1998), no. 2, 125–132. 5, 18
[8] Bonacho Dos Anjos Henriques, I.; Şega, L. M.: Free resolutions over short Gorenstein
local rings. Math. Z. 267 (2011), no. 3-4, 645–663. 21
[9] Brewer, J. W.; Montgomery, P. R.: The finiteness of I when R[X]/I is R–projective. Proc.
Amer. Math. Soc. 41 (1973), 407–414. 20
[10] Brüderle, S.; Kunz, E.: Divided powers and Hochschild homology of complete intersections.
With an appendix by Reinhold Hübl. Math. Ann. 299 (1994), no. 1, 57–76. 19
[11] Buchweitz, R.-O.; Flenner, H.: Global Hochschild (co-)homology of singular spaces.
Adv. Math. 217 (2008), no. 1, 205–242. 11
[12] Buenos Aires Cyclic Homology Group: Hochschild and cyclic homology of hypersurfaces.
Adv. Math. 95 (1992), no. 1, 18–60. 19
28
R.-O. BUCHWEITZ AND C. ROBERTS
[13] Calaque, D.; Van den Bergh, M.: Hochschild cohomology and Atiyah classes. Adv. Math.
224 (2010), no. 5, 1839–1889. 6
[14] Cartan, H.: Algèbre d’Eilenberg-Mac Lane et Homotopie. Séminaire Henri Cartan 7 (195455); available in electronic form through NumDam. 7
[15] Cartan, H. and Eilenberg, S.: Homological algebra, Princeton University Press, 1956. 18,
26
[16] Chapman, G. R.: The cohomology ring of a finite abelian group. Proc. London
Math. Soc. (3) 45 (1982), no. 3, 564–576. 27
[17] Cibils, C. and Solotar, A.: Hochschild cohomology algebra of abelian groups. Arch. Math.
(Basel) 68 (1997), no. 1, 17–21. 24, 25
[18] Cuntz, J. and Quillen, D.: Algebra extensions and nonsingularity. J. Amer. Math. Soc. 8,
no. 2 (1995), 251–289. 8
[19] Dumitrescu, T.; Ionescu, C.: Some examples of two-dimensional regular rings. preprint
2013, 6 pages, arXiv:1301.2615 23
[20] Gerstenhaber, M.: The cohomology structure of an associative ring. Ann. of Math. (2) 78
(1963), 267–288 1
[21] Grothendieck, A.: Eléments de géométrie algébrique. IV. Étude locale des schémas et des
morphismes de schémas. I. Inst. Hautes Études Sci. Publ. Math. No. 20 (1964), 259 pp. 2
[22] Guccione, J. A.; Guccione, J. J.: Hochschild homology of complete intersections. J. Pure
Appl. Algebra 74 (1991), no. 2, 159–176 6, 19
[23] Holm, T.: Hochschild cohomology rings of algebras k[X]/(f ). Beiträge Algebra Geom. 41
(2000), no. 1, 291–301. 24
[24] Hübschmann, J.: Cohomology of finitely generated abelian groups. Enseign. Math. (2) 37
(1991), no. 1-2, 61–71. 27
[25] Iyengar, S.: Acyclicity of Tate constructions. J. Pure Appl. Algebra 163 (2001), no. 3,
289–300. 18
[26] Majadas, J.; Rodicio, A. G.: The Hochschild (co-)homology of hypersurfaces. Comm.
Algebra 20 (1992), no. 2, 349–386. 19
[27] Northcott, D. G.: Finite free resolutions. Cambridge Tracts in Mathematics, No. 71.
Cambridge University Press, Cambridge-New York-Melbourne, 1976. xii+271 pp. 20, 21
[28] Ohm, J.; Rush, D. E.: The finiteness of I when R[X]/I is flat. Trans. Amer. Math. Soc.
171 (1972), 377–408. 20
[29] Quillen, D.: On the (co-)homology of commutative rings. Applications of Categorical
Algebra (Proc. Sympos. Pure Math., Vol. XVII, New York, 1968) pp. 65–87. Amer. Math.
Soc., Providence, R.I., 1970. 4, 5
[30] Roberts, C. D.: The Cohomology Ring of a Finite Abelian Group. Thesis (Ph.D.)– University of Waterloo. 2013. 141 pages. 2, 27
[31] Roberts, L. G.: Some examples of smooth and regular rings. Canad. Math. Bull. 23 (1980),
no. 3, 255–259. 23
[32] Sanada, K.: On the Hochschild cohomology of crossed products. Comm. Algebra 21 (1993),
no. 8, 2727–2748. 10, 11
[33] Siegel, S. F.; Witherspoon, S. J.: The Hochschild cohomology ring of a group algebra.
Proc. London Math. Soc. (3) 79 (1999), no. 1, 131–157. 27
[34] Sjödin, G.: Hopf algebras and derivations. J. Algebra 64 (1980), no. 1, 218–229. 5
[35] Suárez-Alvarez, M.: The Hilton-Heckmann argument for the anti-commutativity of cup
products. Proc. Amer. Math. Soc. 132 (2004), no. 8, 2241–2246 11
[36] Suárez-Alvarez, M.: Applications of the change-of-rings spectral sequence to the computation of Hochschild cohomology. preprint 2007, 27 pages, arXiv:0707.3210 24
[37] Tate, J.: Homology of Noetherian rings and local rings. Illinois J. Math. 1 (1957), 14–27.
7, 18
[38] Townsley Kulich, L. G.: Investigations of the integral cohomology ring of a finite group.
Thesis (Ph.D.)–Northwestern University. 1988. 83 pp. 2, 27
[39] Wolffhardt, K.: The Hochschild homology of complete intersections. Trans. Amer. Math.
Soc. 171 (1972), 51–66 6, 19
MULTIPLICATIVE STRUCTURE OF HOCHSCHILD COHOMOLOGY
29
Dept. of Computer and Mathematical Sciences, University of Toronto at Scarborough, 1265 Military Trail, Toronto, ON M1C 1A4, Canada
E-mail address: [email protected]
Dept. of Pure Mathematics, University of Waterloo, 200 University Avenue West,
Waterloo, ON N2L 3G1, Canada
E-mail address: [email protected]
| 0math.AC
|
JMLR: Workshop and Conference Proceedings 1–7
Unsupervised model compression for multilayer bootstrap
networks
Xiao-Lei Zhang
arXiv:1503.06452v1 [cs.LG] 22 Mar 2015
[email protected]
Department of Computer Science and Engineering, The Ohio State University, Columbus, OH 43210,
USA.
Abstract
Recently, multilayer bootstrap network (MBN) has demonstrated promising performance
in unsupervised dimensionality reduction. It can learn compact representations in standard data sets, i.e. MNIST and RCV1. However, as a bootstrap method, the prediction
complexity of MBN is high. In this paper, we propose an unsupervised model compression
framework for this general problem of unsupervised bootstrap methods. The framework
compresses a large unsupervised bootstrap model into a small model by taking the bootstrap model and its application together as a black box and learning a mapping function
from the input of the bootstrap model to the output of the application by a supervised
learner. To specialize the framework, we propose a new technique, named compressive
MBN. It takes MBN as the unsupervised bootstrap model and deep neural network (DNN)
as the supervised learner. Our initial result on MNIST showed that compressive MBN
not only maintains the high prediction accuracy of MBN but also is over thousands of
times faster than MBN at the prediction stage. Our result suggests that the new technique integrates the effectiveness of MBN on unsupervised learning and the effectiveness
and efficiency of DNN on supervised learning together for the effectiveness and efficiency
of compressive MBN on unsupervised learning.
Keywords: Model compression, multilayer bootstrap networks, unsupervised learning.
1. Introduction
Dimensionality reduction is a core problem of machine learning, where classification and
clustering can be regarded as its special cases that reduce high dimensional data to discrete
points. In this paper, we focus on unsupervised learning. Traditionally, dimensionality
reduction can be categorized to kernel methods, neural networks, probabilistic models, and
sparse coding. Kernel methods are too costly on large-scale problems. Although neural
networks are scalable to large scale data, they double the computational complexity by a
bottleneck structure and take the input as the output of the bottleneck structure at the
training stage which is slow, moreover, they learn data distribution globally which is not
very effective on learning local structures.
Multilayer bootstrap network (MBN) is a recently proposed bootstrap method (or unsupervised ensemble method). It has multiple nonlinear layers. Each layer is an ensemble
of k-centers clusterings. The centers of each k-centers clustering are only randomly sampled
data points (called a bootstrap sample) from the input. MBN is easily implemented and
trained, and scales well to large-scale problems as neural networks at the training stage.
Moreover, MBN learns a data distribution locally so that it can learn effective representa-
c
X.-L. Zhang.
Zhang
tions of data easily. However, MBN contains hundreds of clusterings, which is difficult to
be used for prediction.
Motivated by the aforementioned problem and the recent progress of compressing ensemble classifiers to a single small classifier in supervised learning (Bucilu et al. (2006); Hinton et al. (2015)), in this abstract paper, we propose an unsupervised model compression
framework. The framework uses a supervised model to approximate the mapping function
from the input of an unsupervised bootstrap method to the output of the application of
the unsupervised bootstrap method. We further specify the framework by taking MBN
as the unsupervised bootstrap method and DNN as the supervised model. The proposed
method is named compressive MBN. To our best knowledge, this is the first work of model
compression for bootstrap methods on unsupervised learning.
2. Methods
Compressive MBN is as follows:
• The first step trains MBN on a give training set, and outputs the low dimensional
representation of the training data points.
• [A step driven by applications] The second step applies the low dimensional representation to a given application in unsupervised learning, and outputs the prediction
result of the training data.
• The third step trains a DNN with the training set as the input and the prediction
result as the target. Finally, the DNN model will be used for prediction.
The algorithm is a very basic framework. We can easily extend compressive MBN to
other techniques by simply using other unsupervised bootstrap techniques to replace MBN
in the first step for potentially better performance. We can also use many other supervised
learners to replace DNN in the third step, but to our knowledge, DNN is currently already
a good choice.
We may also design a lot of new algorithms by simply specifying the second step for
different applications. Some examples are as follows. (i) When compressive MBN is used
for visualization, we may omit the second step, and simply take the input and output of the
MBN as the input and output of DNN respectively. (ii) When compressive MBN is used for
unsupervised prediction, we may run a hard clustering algorithm on the training set, and
get the predicted indicator vector of each training data point. For example, if a data point
is assigned to the second cluster, then its predicted indicator vector is [0, 1, 0, 0, . . . , 0]. We
may also get the probabilistic output of the clustering.
3. Experiments
We conducted an initial experiment on MNIST. We showed that the technique is very helpful
for reducing the high computational cost of MBN on unsupervised prediction problems. The
MNIST data was normalized by dividing its entries by 255.
2
Unsupervised model compression for multilayer bootstrap networks
Figure 1: MBN. Its prediction time on the 5000 images is 2333.24 seconds.
Figure 2: Compressive MBN. Its prediction time on the 5000 images is 0.58 seconds.
3.1. Experiment on visualizing small subsets of MNIST
In this subsection, we did not consider the generalization ability of MBN and compressive
MBN. Instead, we studied their visualization ability. A data set that contained 5000 unlabeled images randomly selected from the training set of MNIST was used for both training
and test.
For the MBN training, we trained MBN similarly as in the second experiment in (Zhang
(2014)). Specifically, the number of clusterings in each layer was set to 400. The parameters
k from layer 1 to layer 9 were set to 4000-2000-1000-500-250-125-65-30-15 respectively. As
we can see, MBN is a very large sparse model: (4000-2000-1000-500-250-125-65-30-15)×400.
The parameter a for random feature selection was set to 0.5. The parameter r for random
3
Zhang
reconstruction was set to 0.5. After getting the high-dimensional sparse representation from
MBN at the top hidden layer, we mapped it to two dimensional space by the expectationmaximization principle component analysis (EM-PCA) (Roweis (1998)).
For the training of compressive MBN, we omitted the second step, and used the
input and output of MBN as the input and output of a DNN model respectively. The
parameter settings are as follows. We trained a 786-2048-2048-2 DNN. The dropout rate
was set to 0.2. The rectified linear unit was used as the hidden unit, and the linear function
was used as the output unit. The number of the training epoches was set to 120. The batch
size was set to 32. The learning rate was set to 0.001.
The two dimensional visualizations produced by MBN and compressive MBN were
shown in Fig. 1 and Fig. 2 respectively. From the two figures, we found that the visualizations of both MBN and compressive MBN by DNN were equivalently good. When
we used the features for clustering, the NMIs of both the methods were around 81%. Amazingly, the prediction time of the compressive MBN on the 5000 images was only
0.58 seconds, which accelerated the prediction time of MBN by around 4000
times!1
3.2. Experiment on the full MNIST
We used all 60,000 training images for unsupervised model training and 10,000 test images
for test. We discussed the unsupervised generalization ability of the compressive MBN on
the test images.
3.2.1. Compressive MBN without random reconstruction (i.e. parameter
r = 0)
For the MBN training, we trained MBN similarly as in the third experiment in (Zhang
(2014)). Specifically, the number of clusterings in each layer was set to 400. The parameters
k from layer 1 to layer 9 were set to 4000-2000-1000-500-250-125-65-30-15 respectively. As
we can see, MBN is a very large sparse model: (4000-2000-1000-500-250-125-65-30-15)×400.
The parameter a for random feature selection was set to 0.5. The parameter r for random
reconstruction was set to 0. After getting the high-dimensional sparse representation from
MBN at the top hidden layer, we mapped it to 5 dimensional space by EM-PCA (Roweis
(1998)). We further encoded the 5-dimensional representations to 10-dimensional indicator
vectors by k-means clustering, which was a specialization of the second step of compressive
MBN.
For the training of compressive MBN, we took the raw feature of the training set as
the input of DNN, and took the 10-dimensional predicted indicator vectors as the training
target of DNN. The parameter settings of the DNN were as follows. We trained a 7862048-2048-10 DNN. The dropout rate was set to 0.2. The rectified linear unit was used as
the hidden unit, and the sigmoid function was used as the output unit. The number of the
training epoches was set to 50. The batch size was set to 128. The learning rate was set to
0.001.
Because k-means clustering suffers from local minima, we ran the aforementioned methods 10 times and recorded the average results. The experimental comparison between MBN
1. MBN did not enable parallel computing.
4
Unsupervised model compression for multilayer bootstrap networks
MBN & Compressive MBN without random reconstruction
0.9
0.85
Normalized mutual information
0.8
0.75
0.7
0.65
0.6
Training accuracy by MBN
0.55
Training accuracy by compressive MBN
Test accuracy by MBN
0.5
Test accuracy by compressive MBN
0.45
0.4
0
2
4
6
Number of layers
8
10
Figure 3: Comparison of the generalization ability of MBN and compressive MBN on clustering when the random reconstruction of MBN is not used (i.e. r = 0). The
clustering accuracy is evaluated by normalized mutual information. The prediction time of MBN on the 10,000 test images is 4699.69 seconds. The prediction
time of compressive MBN on the 10,000 test images is 1.15 seconds.
MBN & Compressive MBN with random reconstruction
0.9
0.85
Normalized mutual information
0.8
0.75
0.7
0.65
0.6
0.55
Training accuracy by MBN
0.5
Training accuracy by compressive MBN
Test accuracy by MBN
0.45
0.4
Test accuracy by compressive MBN
0
2
4
6
Number of layers
8
10
Figure 4: Comparison of the generalization ability of MBN and compressive MBN on clustering when the random reconstruction of MBN is not used (i.e. r = 0.5). The
prediction time of MBN on the 10,000 test images is 4857.45 seconds. The prediction time of compressive MBN on the 10,000 test images is 1.10 seconds.
and compressive MBN was summarized in Fig. 3. From the figure, we observed that the
curves of the training accuracy of MBN and compressive MBN were completely coincident;
5
Zhang
moreover, the curve of prediction of compressive MBN was even slightly better than that
of MBN; the highest prediction accuracy of compressive MBN reached 84% in terms of
NMI. The most advanced property of compressive MBN is that it needed only 1.15 seconds
to predict 10,000 images, while MBN needed 4699.69 seconds to predict 10,000 images. The
prediction time was accelerated by around 4000 times.
3.2.2. Compressive MBN with random reconstruction (i.e. parameter r = 0.5)
It is shown in (Zhang (2014)) that when the data is small scale (i.e. the training size
was similar to the largest parameter k), the random reconstruction operation can be quite
helpful, however, it is still unclear that whether random reconstruction will be helpful when
the data is large scale (i.e. the training size is much larger than the largest parameter
k), since the largest k in the third experiment of (Zhang (2014)) was only 1000 and the
experimental results of MBN with or without random reconstruction was not very exciting.
In this subsection, we enlarged k to 4000 as in Section 3.2.1.
The experimental settings of both MBN and compressive MBN were the same as in
Section 3.2.1 except that we set r = 0.5 and mapped the sparse features to 2 dimensional
space by EM-PCA. The experimental results were summarized in Fig. 4. From the figure,
we observed that all experimental conclusions in Section 3.2.1 could also be summarized
here, except that when random reconstruction was used, the performance of both MBN and
compressive MBN was not as good as that without random reconstruction.
4. Conclusions
In this paper, we proposed a general framework for unsupervised model compression. The
framework takes MBN (Zhang (2014)) as a case of study. The specialized technique, named
compressive MBN, uses DNN as an auxiliary model for modeling the mapping function
from the input of MBN to the prediction result of a given application that takes the low
dimension output of MBN as its input. The new technique aims to solve the problem
that although MBN is simple, effective, robust, and efficient-at-the-training-stage, it is time
consuming on prediction.
Our initial experimental result on MNIST showed that compressive MBN not only
inherited the generalization ability of MBN (and is even slightly better than MBN), but
also accelerated the prediction efficiency of MBN by over thousands of times.
Compressive MBN concatenates the effectiveness of MBN on unsupervised learning and
the effectiveness and efficiency of DNN on supervised learning together for both its effectiveness and its efficiency on unsupervised learning. Moreover, we can easily extend compressive
MBN to other unsupervised model compression techniques.
5. Acknowledgements
The author thanks Dr Yuxuan Wang for providing the well-designed DNN toolbox and Prof
DeLiang Wang for providing computing resources of the Ohio Supercomputing Center.
6
Unsupervised model compression for multilayer bootstrap networks
References
Cristian Bucilu, Rich Caruana, and Alexandru Niculescu-Mizil. Model compression. In
Proc. 12th Int. Conf. Knowl. Disc., Data Min., pages 535–541. ACM, 2006.
Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.
arXiv preprint arXiv:1503.02531, 2015.
Sam T Roweis. EM algorithms for PCA and SPCA. In Advances in Neural Information
Processing Systems 10, pages 626–632, Denver, CO, 1998.
Alexander Strehl and Joydeep Ghosh. Cluster ensembles—a knowledge reuse framework for
combining multiple partitions. J. Mach. Learn. Res., 3:583–617, 2003.
Xiao-Lei Zhang. Nonlinear dimensionality reduction of data by multilayer bootstrap networks. arXiv preprint arXiv:1408.0848, pages 1–18, 2014.
7
| 9cs.NE
|
Approximating the Largest Root and
Applications to Interlacing Families
arXiv:1704.03892v1 [cs.DS] 12 Apr 2017
Nima Anari1 , Shayan Oveis Gharan2 , Amin Saberi1 , and Nikhil Srivastava3
1
Stanford University, {anari,saberi}@stanford.edu
University of Washington, [email protected]
3
University of California, Berkeley, [email protected]
2
April 14, 2017
Abstract
We study the problem of approximating the largest root of a real-rooted polynomial of degree
n using its top k coefficients and give nearly matching upper and lower bounds. We present
algorithms with running time polynomial in k that use the top k coefficients to approximate
the maximum root within a factor of n1/k and 1 + O( logk n )2 when k ≤ log n and k > log n
respectively. We also prove corresponding information-theoretic lower bounds of nΩ(1/k) and
2n 2
log k
, and show strong lower bounds for noisy version of the problem in which one is
1+Ω
k
given access to approximate coefficients.
This problem has applications in the context of the method of interlacing families of polynomials, which was used for proving the existence of Ramanujan graphs of all degrees, the solution
of the Kadison-Singer problem, and bounding the integrality gap of the asymmetric traveling
salesman problem. All of these involve computing the maximum root of certain real-rooted
polynomials for which the top few coefficients are√accessible in subexponential time. Our results
3
yield an algorithm with the running time of 2Õ( n) for all of them.
Contents
1 Introduction
1.1 Motivation and Applications
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Preliminaries
2
3
5
3 Approximation of the Largest Root
8
3.1 Proof of Theorem 3.1: An Algorithm for Approximating the Largest Root . . . . . . 9
3.2 Proofs of Theorems 3.2 and 3.3: Matching Lower Bounds . . . . . . . . . . . . . . . 11
3.3 Lower Bound Given Approximate Coefficients . . . . . . . . . . . . . . . . . . . . . . 15
4 Applications to Interlacing Families
17
4.1 Oracle for Kadison-Singer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Oracle for ATSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1
A Appendix A: Proofs of Preliminary Facts
23
B Appendix B: Constructions of Graphs with Large Girth
24
1
Introduction
For a non-negative vector µ = (µ1 , . . . , µn ) ∈ Rn+ , let χµ denote the unique monic polynomial with
roots µ1 , . . . , µn :
n
Y
(x − µi ).
χµ (x) :=
i=1
Suppose that we do not know µ, but rather know the top k coefficients of χµ where 1 ≤ k < n.
In more concrete terms, suppose that
χµ (x) = xn + c1 xn−1 + c2 xn−2 + · · · + cn ,
and we only know c1 , . . . , ck . What do c1 , . . . , ck tell us about the roots and in particular maxi µi ?
Problem 1.1. Given the top k coefficients of a real rooted polynomial of degree n, how well can
you approximate its largest root?
This problem may seem completely impossible if k is significantly smaller than n. For example,
consider the two polynomials xn −a and xn −b. The largest roots of these two polynomials can differ
arbitrarily in absolute value and even by knowing the top n − 1 coefficients we can not approximate
the absolute value of the largest at all.
The key to approach the above problem is to exploit real rootedness. One approach is to
construct a polynomial with the given coefficients and study its roots. Unfortunately, even assuming
that the roots of the original polynomial are real and well-separated, adding an exponentially small
amount of noise to the (bottom) coefficients can lead to constant sized perturbations of the roots
in the complex plane — the most famous example is Wilkinson’s polynomial [Wil84]:
(x − 1)(x − 2) . . . (x − 20).
Instead, we use the given coefficients to compute a polynomial of the roots, e.g., the k-th
moment of the roots, and we use that polynomial to estimate the largest root. Our contributions
towards Problem 1.1 are as follows.
Efficient Algorithm for Approximating the Largest Root. In Theorem 3.1, we present
an upper bound showing that we can use the top k coefficients of a real rooted polynomial with
nonnegative roots to efficiently obtain an αk,n approximation of the largest root where
(
n1/k
k ≤ log n,
αk,n =
log n 2
1 + O( k ) k > log n.
Moreover
such an approximation can be done in poly(k) time. This implies that exact access to
log
n
O √ǫ coefficients is sufficient for (1 + ǫ)-approximation of the largest root.
2
Nearly Matching Lower Bounds. The main nontrivial part of this work is our informationtheoretic matching lower bounds. In Theorem 3.2, we show that when k < n1−ǫ and for some
constant c, there are no algorithms with approximation factor better than αck,n . Chebyshev polynomials are critical for this construction as well. We also use known constructions for large-girth
graphs, and the proof of [MSS13a] for a variant of Bilu-Linial’s conjecture [BL06]. Our bounds can
be made slightly sharper assuming Erdős’s girth conjecture.
1.1
Motivation and Applications
For many important polynomials, it is easy to compute the top k coefficients exactly, whereas it is
provably hard to compute all of them. One example is the matching polynomial of a graph, whose
coefficients encode the number of matchings of various sizes. For this polynomial, computing the
constant term, i.e. the number of perfect matchings, is #P-hard [Val79], whereas for small k, one
can compute the number of matchings of size k exactly, in time nO(k) , by simply enumerating all
possibilities. Roots of the matching polynomial, and in particular the largest root arise in a number
of important applications [HL72, MSS13a, SS13]. So it is natural to ask how well the largest root
can be approximated from the top few coefficients.
Another example of a polynomial whose top coefficients are easy to compute is the independence
polynomial of a graph, which is real rooted for claw-free graphs [CS07], and whose roots have
connections to the Lovász Local Lemma (see e.g. [HSV16]).
Subexponential Time Algorithms for Method of Interlacing Polynomials Our main
motivation for this work is the method of interlacing families of polynomials [MSS13a, MSS13b,
AO15b], which has been an essential tool in the development of several recent results including the
construction of Ramanujan graphs via lifts [MSS13a, MSS15, HPS15], the solution of the KadisonSinger problem [MSS13b], and improved integrality gaps for the asymmetric traveling salesman
problem [AO15a]. Unfortunately, all these results on the existence of expanders, matrix pavings,
and thin trees, have the drawback of being nonconstructive, in the sense that they do not give
polynomial time algorithms for finding the desired objects (with the notable exception of [Coh16]).
As such, the situation is somewhat similar to that of the Lovász Local Lemma (which is able to
guarantee the existence of certain rare objects nonconstructively), before algorithmic proofs of it
were found [Bec91, MT10].
√
3
In Section 4, we use Theorem 3.1 to give a 2Õ( m) time algorithm for rounding an interlacing
family of depth m, improving on the previously known running time of 2O(m) . This leads to
algorithms of the same running time for all of the problems mentioned above.
Lower Bounds Given Approximate Coefficients In the context of efficient algorithms, one
might imagine that computing a much larger number of coefficients approximately might provide
a better estimate of the largest root. In particular, we consider the following noisy version of
Problem 1.1:
Problem 1.2. Given real numbers a1 , . . . , ak , promised to be (1 + δ)−approximations of the first
k coefficients of a real-rooted polynomial p, how well can you approximate the largest root of p?
An important extension of our information theoretic lower bounds is that Problem 1.1 is extremely sensitive to noise: in Proposition 3.12 we prove that even knowing all but the k-th coefficient
3
exactly and knowing the k-th one up to a 1+1/2k error is no better than knowing only the first k−1
coefficients exactly. We do this by exhibiting two polynomials which agree on all their coefficients
except for the kth , in which they differ slightly, but nonetheless have very different largest roots.
This example is relevant in the context of interlacing families, because the polynomials in our
lower bound have a common interlacing and are characteristic polynomials of 2−lifts of a base
graph, which means they could actually arise in the proofs of [MSS13a, MSS13b]. To appreciate
this more broadly, one can consider the following taxonomy of increasingly structured polynomials:
complex polynomials ⊃ real-rooted polynomials ⊃ mixed characteristic polynomials
⊃ characteristic polynomials of lifts of graphs.
Our example complements the standard numerical analysis wisdom (as in Wilkinson’s example) that
complex polynomial roots are in general terribly ill-conditioned as functions of their coefficients,
and shows that this fact remains true even in the structured setting of interlacing families.
Proposition 3.12 is relevant to the quest for efficient algorithms for interlacing families for the
following reason. All of the coefficients of the matching polynomial of a bipartite graph can be
approximated to 1 + 1/ poly(n) error in polynomial time, for any fixed polynomial, using Markov
Chain Monte Carlo techniques [SJ89, JSV04, FL06]. One might imagine that an extension of these
techniques could be used to approximate the coefficients of the more general expected characteristic
polynomials that appear in applications of interlacing families. In fact for some families of interlacing polynomials (namely, the mixed characteristic polynomials of [MSS13b]) we can design Markov
chain Monte Carlo techniques to approximate the top half of the coefficients within 1 + 1/ poly(n)
error.
Our information theoretic lower bounds rule out this method as a way to approximate the largest
root, at least in full generality, since knowing all of the coefficients of a real-rooted polynomial up
to a (1 + 1/ poly(n)) error for any poly(n) is no better than just knowing the first log n coefficients
exactly, in the worst case, even under the promise that the given polynomials have a common
interlacing. In other words, even an MCMC oracle that gives 1 + 1/ poly(n) approximation of all
coefficients would not generically allow one to round an interlacing family of depth greater than
logarithmic in n, since the error accumulated at each step would be 1/polylog(n).
Connections to Poisson Binomial Distributions Finally, there is a probabilistic view of
Problem 1.1. Assume that X = B(p1 ) + · · · + B(pn ) is a sum of independent Bernoulli random
variables, i.e. a Poisson binomial, with parameters p1 , . . . , pn ∈ [0, 1]. Then Problem 1.1 becomes
the following: Given the first k moments of X how well can we approximate maxi pi ? In this view,
our paper is related to [DP15], where it was shown that any pair of such Poisson binomial random
variables with the same first k moments have total variation distance at most 2−Ω(k) . However, the
bound on the total variation distance does not directly imply a bound on the maximum pi .
Discussion Besides conducting a precise study of the dependence of the largest root of a realrooted polynomial on its coefficients, the results of this paper shed light on what a truly efficient
1/3
algorithm for interlacing families might look like. On one hand, our running time of 2Õ(m ) shows
that the problem is not ETH hard, and is unnatural enough to suggest that a faster algorithm
(for instance, quasipolynomial) may exist. On the other hand, our lower bounds show that the
polynomials that arise in this method are in general hard to compute in a rather robust sense:
namely, obtaining an inverse polynomial error approximation of their largest roots requires knowing
4
many coefficients exactly. This implies that in order to obtain an efficient algorithm for even
approximately simulating the interlacing families proof technique, one will have to exploit finer
properties of the polynomials at hand, or find a more “global” proof which is able to reason about
the error in a more sophisticated amortized manner, or perhaps track a more well-conditioned
quantity in place of the largest root, which can be computed using fewer coefficients and which still
satisfies an approximate interlacing property.
2
Preliminaries
We let [n] denote the set {1, . . . , n}. We use the notation [n]
to denote the family of subsets
k
T ⊆ [n] with |T | = k. We let Sn denote the set of permutations on [n], i.e. the set of bijections
σ : [n] → [n].
We use bold letters to denote vectors. For a vector µ ∈ Rn , we denote its coordinates by
µ1 , . . . , µn . We let µmax and µmin denote maxi µi and mini µi respectively.
For a symmetric matrix A, we denote the vector of eigenvalues of A, i.e. the roots of det(xI −A),
by λ(A). Similarly we denote the largest and smallest eigenvalues by λmax (A) and λmin (A). We
slightly abuse notation, and for a polynomial p we write λ(p) to denote the vector of roots of p.
We also write λmax (p) to denote the largest root of p.
For a graph G = (V, E) we let degmax (G) denote the maximum degree of its vertices and
degavg (G) denote the average degree of its vertices, i.e. 2|E|/|V |.
What follows are mostly standard facts; the proofs of Fact 2.4, Fact 2.13, Fact 2.13, and
Fact 2.22 are included in Appendix A for completness.
Facts from Linear Algebra For a matrix A ∈ Rn×n , the characteristic polynomial of A is
defined as det(xI − A). Letting σk (A) be the sum of all principal k-by-k minors of A, we have:
det(xI − A) =
n
X
xn−k σk (A).
k=0
There are several algorithms that for a matrix A ∈ Rn×n calculate det(xI − A) in time polynomial
in n. By the above identity, we can use any such algorithm to efficiently obtain σk (A) for any
1 ≤ k ≤ n.
The following proposition is proved in [MSS13b] using the Cauchy-Binet formula.
Proposition 2.1. Let v1 , . . . , vm ∈ Rn . Then,
!
n
m
X
X
X
σk
(−1)k xn−k
vi viT =
det xI −
i=1
k=0
[m]
k
S⊆(
)
X
i∈S
vi viT
!
.
Symmetric Polynomials We will make heavy use of the elementary symmetric polynomials,
which relate the roots of a polynomial to its coefficients.
Definition 2.2. Let ek ∈ R[µ1 , . . . , µn ] denote the k-th elementary symmetric polynomial defined
as
X Y
ek (µ) :=
µi .
i∈T
T ∈([n]
k )
5
Fact 2.3. Consider the monic univariate polynomial χ(x) = xn + c1 xn−1 + · · · + cn . Suppose that
µ1 , . . . , µn are the roots of χ. Then for every k ∈ [n],
ck = (−1)k ek (µ1 , . . . , µn ).
This means that knowing the top k coefficients of a polynomial is equivalent to knowing the
first k elementary symmetric polynomials of the roots. It also implies the following fact about how
shifting and scaling affect the elementary symmetric polynomials.
Fact 2.4. Let µ, ν ∈ Rn be such that ei (µ) = ei (ν) for i = 1, . . . , k. If a, b ∈ R then ei (aµ + b) =
ei (aν + b) for i = 1, . . . , k.
We will use the following relationship between the elementary symmetric polynomials and the
power sum polynomials.
Pn
k
Theorem 2.5 (Newton’s Identities). For 1 ≤ k ≤ n, the polynomial pk (µ) :=
i=1 µi can be
written as qk (e1 (µ), . . . , ek (µ)), where qk ∈ R[e1 , . . . , ek ]. Furthermore, qk can be computed at any
point in time poly(k).
One of the immediate corollaries of the above is the following.
Corollary 2.6. Let p(x) ∈ R[x] be a univariate polynomial with deg p ≤ k. Then
be written as
q(e1 (µ), . . . , ek (µ)),
Pn
i=1 p(µi )
can
where q ∈ R[e1 , . . . , ek ]. Furthermore, q can be computed at any point in time poly(k).
Theorem 2.5 shows how p1 , . . . , pk can be computed from e1 , . . . , ek . The reverse is also true.
A second set of identities, also known as Newton’s identities, imply the following.
Theorem 2.7 (Newton’s Identities). For each k ∈ [n], ek (µ) can be written as a polynomial of
p1 (µ), . . . , pk (µ) which can be computed in time poly(k).
A corollary of the above and Theorem 2.5 is the following.
Corollary 2.8. For two vectors µ, ν ∈ Rn , we have
(∀i ∈ [k] : ei (µ) = ei (ν)) ⇐⇒ (∀i ∈ [k] : pi (µ) = pi (ν)) .
Chebyshev Polynomials Chebyshev polynomials of the first kind, which we will simply call
Chebyshev polynomials, are defined as follows.
Definition 2.9. Let the polynomials T0 , T1 , · · · ∈ R[x] be recursively defined as
T0 (x) := 1,
T1 (x) := x,
Tn+1 (x) := 2xTn (x) − Tn−1 (x).
We will call Tk the k-th Chebyshev polynomial.
6
Notice that the coefficients of Tk can be computed in poly(k) time, by the above recurrence for
example. Chebyshev polynomials have many useful properties, some of which we mention below.
For further information, see [Sze39].
Fact 2.10. For k ≥ 0 and θ ∈ R, we have
Tk (cos(θ)) = cos(kθ),
Tk (cosh(θ)) = cosh(kθ).
Fact 2.11. The k-th Chebyshev polynomial Tk has degree k.
Fact 2.12. For any x ∈ [−1, 1], we have Tk (x) ∈ [−1, 1].
Fact 2.13. For any integer k ≥ 0, Tk (1 + x) is monotonically increasing for x ≥ 0. Furthermore
for x ≥ 0,
√
Tk (1 + x) ≥ (1 + 2x)k /2.
In our approximate lower bound we will use the following connection between Chebyshev polynomials and graphs, due to Godsil and Gutman [?].
Fact 2.14. If An is the adjacency matrix of a cycle on n vertices, then
det(2xI − An ) = 2Tn (x).
Graphs with Large Girth In order to prove some of our impossibility results, we use the
existence of extremal graphs with no small cycles.
Definition 2.15. For an undirected graph G, we denote the length of its shortest cycle by girth(G).
If G is a forest, then girth(G) = ∞.
The following conjecture by Erdős characterizes extremal graphs with no small cycles.
Conjecture 2.16 (Erdős’s girth conjecture [Erd64]). For every integer k ≥ 1 and sufficiently large
n, there exist graphs G on n vertices with girth(G) > 2k that have Ω(n1+1/k ) edges, or in other
words satisfy degavg (G) = Ω(n1/k ).
This conjecture has been proven for k = 1, 2, 3, 5 [Wen91]. We will use the following more
general construction of graphs of somewhat lower girth.
Theorem 2.17 ([LU95]). If d is a prime power and t ≥ 3 is odd, there is a d-regular bipartite
graph G on 2dt vertices with girth(G) ≥ t + 5.
Signed Adjacency Matrices Our lower bounds will also utilize facts about signings of graphs.
Definition 2.18. For a graph G = ([n], E), we define a signing to be any function s : E →
{−1, +1}. We define the signed adjacency matrix As ∈ Rn×n , associated with signing s, as follows
(
0
{u, v} ∈
/ E,
As (u, v) :=
s({u, v}) {u, v} ∈ E.
7
Note that by definition, As is symmetric and has zeros on the diagonal. The following fact is
immediate.
Fact 2.19. For a signed adjacency matrix As of a graph G, the eigenvalues λ(As ), i.e. the roots
of χ(x) := det(xI − As ), are real. If G is bipartite, the eigenvalues are symmetric about the origin
(counting multiplicities).
Signed adjacency matrices were used in [MSS13a] to prove the existence of bipartite Ramanujan
graphs of all degrees. We state one of the main results of [MSS13a] below.
Theorem 2.20 ([MSS13a]). For every graph G, there exists a signing s such that
p
λmax (As ) ≤ 2 degmax (G) − 1.
By Fact 2.19, we have the following immediate corollary.
Corollary 2.21. For every bipartite
p graph G, there exists a signing s such that the eigenvalues of
As have absolute value at most 2 degmax (G) − 1.
We note that trivially signing every edge with +1 is often far from achieving the above bound
as witnessed by the following fact.
Fact 2.22. Let A be the adjacency matrix of a graph G = ([n], E) (i.e. the signed adjacency matrix
where the sign of every edge is +1). Then the maximum eigenvalue of A is at least degavg (G).
3
Approximation of the Largest Root
In this section we give an answer to Problem 1.1. As witnessed by Fact 2.3, knowing the top k
coefficients of the polynomial χµ is the same as knowing e1 (µ), . . . , ek (µ). Therefore, without loss
of generality and more conveniently, we state the results in terms of knowing e1 (µ), . . . , ek (µ).
Theorem 3.1. There is an algorithm that receives n and e1 (µ), . . . , ek (µ) for some unknown
µ ∈ Rn+ as input and outputs µ∗max , an approximation of µmax , with the guarantee that
µ∗max ≤ µmax ≤ αk,n · µ∗max ,
where the approximation factor αk,n is
αk,n
(
n1/k
=
1 + O( logk n )2
k ≤ log n,
k > log n.
Furthermore the algorithm runs in time poly(k).
Note that there is a change in the behavior of the approximation factor in the two regimes
k ≪ log n and k ≫ log n. When k > log n, the expression n1/k is 1 + Θ( logk n ) which can be a
much worse bound compared to 1 + O( logk n )2 . When k is near the threshold of log n, n1/k and
1 + Θ( logk n )2 are close to each other and both of the order of 1 + Θ(1).
We complement this result by showing information-theoretic lower bounds.
8
Theorem 3.2. For every 1 ≤ k < n, there are two vectors µ, ν ∈ Rn+ such that ei (µ) = ei (ν) for
i = 1, . . . , k, and
νmax
≥ βk,n ,
µmax
where
nΩ(1/k)
k ≤ log n,
2n 2
βk,n =
1 + Ω log k
k > log n.
k
This shows that no algorithm can approximate µmax by a factor better than βk,n using e1 (µ), . . . , ek (µ).
Note that for k < n1−ǫ , βk,n = αck,n for some constant c bounded away from zero. For constant k,
it is possible to give a constant multiplicative bound assuming Erdős’s girth conjecture.
Theorem 3.3. Assume that k is fixed and Erdős’s girth conjecture (Conjecture 2.16) is true for
graphs of girth > 2k. Then for large enough n there are two vectors µ, ν ∈ Rn+ such that ei (µ) =
ei (ν) for i = 1, . . . , k and
νmax
≥ Ω(n1/k ).
µmax
3.1
Proof of Theorem 3.1: An Algorithm for Approximating the Largest Root
We consider two cases: if k ≤ log n we return (pk (µ)/n)1/k as the estimate for the maximum root.
It is not hard to see that in this case, (pk (µ)/n)1/k gives an n1/k approximation of the maximum
root (see Claim 3.4 below).
For k > log n we can still use (pk (µ)/n)1/k to estimate the maximum root, but this only guarantees a 1 + O( logk n ) approximation. We show that using the machinery of Chebyshev polynomials
we can obtain a better bound. The pseudocode for the algorithm can be see in Algorithm 1.
Algorithm 1 Algorithm For Approximating the Maximum Root From Top Coefficients
Input: n and e1 (µ), e2 (µ), . . . , ek (µ) for some µ ∈ Rn+ .
Output: µ∗max , an approximation of µmax .
if k ≤ log n then
P
Compute pk (µ) = ni=1 µki using Newton’s identities (Theorem 2.5).
return (pk (µ)/n)1/k .
else
t ← e1 (µ).
loop
P
Compute p(µ) := ni=1 Tk ( µti ) using Corollary 2.6.
if p(µ) > n then
return µ∗max ← t.
end if
t
t←
20 log n 2 .
1+(
end loop
end if
k
)
We will prove the following claims to show the correctness of Algorithm 1. Let us start with
the case k ≤ log n.
9
Claim 3.4. For any k ≥ 1 we have
p (µ) 1/k
k
≤ µmax ≤ pk (µ)1/k .
n
Proof. Observe,
pk (µ)/n =
n
X
µk
i
i=1
n
≤
n
X
µk
max
i=1
n
= µkmax ≤
Taking k1 -th root of all sides of the above proves the claim.
n
X
µki = pk (µ),
i=1
The rest of the section handles the case where k > log n. Our first claim shows that as long as
t ≥ µmax , the algorithm keeps decreasing t by a multiplicative factor of (1 − Ω(log(n)/k)2 ). Since
at the beginning we have t = e1 (µ) ≥ µmax , we will have µ∗max ≤ µmax .
Claim 3.5. For any t ≥ µmax ,
n
X
Tk (
i=1
µi
) ≤ n.
t
Proof. If t ≥ µmax , then µi /t ∈ [0, 1] for every i ∈ [n]. By Fact 2.12, we have
n
X
n
Tk (
X
µi
1 = n.
)≤
t
i=1
i=1
To finish the proof of correctness it is enough to show that µmax ≤ µ∗max (1 + O(log n/k)2 ). This
is done in the next claim. It shows that as soon as t gets lower than µmax , within one more iteration
of the loop, the algorithm terminates.
n 2
Claim 3.6. For k > log n and t > 0, if µmax > (1 + ( 20 log
) )t, then
k
n
X
i=1
Tk (
µi
) > n.
t
n 2
) )t, by Fact 2.13 we have
Proof. When µmax > (1 + ( 20 log
k
20 log n 2 1
√ 20 log n k
µmax
≥
) ≥ Tk 1 +
1+ 2·
t
k
2
k
k
1
3 log n
≥ exp
> 2n,
2
k
√
where we used the inequality 1 + 800x ≥ e3x for x ∈ [0, 1].
Now we have
Tk (
n
X
i=1
Tk (
µi
µmax
) = Tk (
)+
t
t
X
i6=argmaxj µj
Tk (
µi
) ≥ 2n − (n − 1) > n,
t
where we used Fact 2.12 and Fact 2.13 to conclude Tk ( µti ) ≥ −1 for every i.
10
The above claim also gives us a bound on the number of iterations in which the algorithm
terminates. This is because we start the loop with t = e1 (µ) ≤ nµmax and the loop terminates
within one iteration as soon as t < µmax . Therefore the number of iterations is at most
k 2
log n
= O log n · (
) = O(k2 ).
1+
n 2
log n
log 1 + ( 20 log
)
k
3.2
Proofs of Theorems 3.2 and 3.3: Matching Lower Bounds
The machinery of Chebyshev polynomials was used to prove Theorem 3.1. We show that this
machinery can also be used to prove a weaker version of Theorem 3.2.
Theorem 3.7. For every 1 ≤ k < n, there are µ, ν ∈ Rn+ such that ei (µ) = ei (ν) for i = 1, . . . , k
and
νmax
≥ 1 + Ω(1/k2 )
µmax
Proof. First let us prove this for k = n − 1. Let µ be the set of roots of Tn (x − 1) + 1 and ν the
set of roots of Tn (x − 1) − 1. These two polynomials are the same except for the constant term. It
follows that ei (µ) = ei (ν) for i = 1, . . . , n − 1. We use the following lemma to prove that µ, ν ∈ Rn+ .
Lemma 3.8. For θ ∈ R, the roots of Tn (x) − cos(θ), counting multiplicities, are cos( θ+2πi
n ) for
i = 0, . . . , n − 1.
Proof. We have
Tn
θ + 2πi
) = cos(θ + 2πi) = cos(θ).
cos(
n
For almost all θ these roots are distinct and since Tn has degree n, it follows that they are all
of the roots. When some of these roots collide, we can perturb θ and use the fact that roots are
continuous functions of the polynomial coefficients to prove the statement.
2πi
Using the above lemma for θ = π and θ = 0, we get that µi = 1+cos( π+2πi
n ) and νi = 1+cos( n ).
n
This proves that µ, ν ∈ R+ . Moreover we have
νmax
1+1
=
= 1 + Ω(1/n2 ).
µmax
1 + cos(π/n)
This finishes the proof for k = n − 1.
Now let us prove the statement for general k. By applying the above proof for n = k + 1, we
such that ei (µ̃) = ei (ν̃) for i = 1, . . . , k and
get µ̃, ν̃ ∈ Rk+1
+
ν̃max
≥ 1 + Ω(1/k 2 ).
µ̃max
Now construct µ, ν from µ̃, ν̃ by adding zeros to make the total count n. It is not hard to see, by
using Corollary 2.8 that ei (µ) = ei (ν) for i = 1, . . . , k. Moreover µmax = µ̃max and νmax = ν̃max .
This finishes the proof.
11
Note that the above lower bound is the same as the lower bound in Theorem 3.2 when k = Ω(n).
However, to prove Theorem 3.2 and Theorem 3.3 we need more tools. The crucial idea we use to
get the stronger Theorem 3.2 and Theorem 3.3 is the following observation about signed adjacency
matrices for graphs of large girth.
Lemma 3.9. Let G = ([n], E) be a graph and D ∈ Rn×n an arbitrary diagonal matrix. If girth(G) >
k, then the top k coefficients of the polynomial χ(x) = det(xI − (D + As )) are independent of the
signing s. In other words,
ei (λ(D + As1 )) = ei (λ(D + As2 )),
for any two signings s1 , s2 and i = 1, . . . , k.
We will apply the above lemma, with D = 0, to graphs of large girth constructed based on
Conjecture 2.16 or Theorem 2.17, in order to prove Theorem 3.3 and Theorem 3.2 for the regime
k ≤ Θ(log n). In order to prove Theorem 3.3 for the regime k ≥ Θ(log n), we marry these constructions with Chebyshev polynomials. We will prove the above lemma at the end of this section, after
proving Theorem 3.2 and Theorem 3.3.
First let us prove Theorem 3.3.
Proof of Theorem 3.3. We apply Lemma 3.9 to the following graph construction, the proof of
which we defer to Appendix B.
Claim 3.10. Let k be fixed and assume that Conjecture 2.16 is true for graphs of girth > 2k.
Then, for all sufficiently large n, there exist bipartite graphs G = ([n], E) with girth(G) > 2k,
degmax (G) = O(n1/k ), and degavg (G) = Ω(n1/k ).
Let G be the graph from the above claim. Let s1 be the trivial signing that assigns +1 to every
edge, and s2 the signing guaranteed by Corollary 2.21. Now let ν = λ(A2s1 ), i.e. the square of the
eigenvalues of As1 , and µ = λ(A2s2 ), i.e. the square of the eigenvalues of As2 . By Lemma 3.9 and
Corollary 2.8, we have
pi (ν) = p2i (λ(As1 )) = p2i (λ(As2 )) = pi (µ),
for i = 1, . . . , k. By another application of Corollary 2.8, we have ei (µ) = ei (ν) for i = 1, . . . , k.
On the other hand, by Corollary 2.21, we have
µmax = λmax (As2 )2 ≤ 4(degmax (G) − 1) = O(n1/k ),
and by Fact 2.22, we have
νmax = λmax (As1 )2 ≥ degavg (G)2 = Ω(n2/k ).
Therefore
Ω(n2/k )
νmax
= Ω(n1/k ).
=
µmax
O(n1/k )
Now let us prove Theorem 3.2 for k ≤ Θ(log n).
12
Proof of Theorem 3.2 for k ≤ Θ(log n). We apply Lemma 3.9 to the following graph construction,
the proof of which we defer to Appendix B.
Claim 3.11. Let d be a prime. For all sufficiently large n, there exist bipartite graphs G = ([n], E)
with girth(G) = Ω(log n), degmax (G) ≤ d, and degavg (G) ≥ d/2.
We will fix d to a specific prime later. Similar to the proof of Theorem 3.3, we let s1 be the
trivial signing with all +1s and s2 be the signing guaranteed by Corollary 2.21. Let t = Ω(log n/k)
be an even integer such that tk < girth(G). Such a t exists when k < c log n for some constant c.
Take ν = λ(Ats1 ) and µ = λ(Ats2 ). Then we have
√
µmax = λmax (As2 )t ≤ (2 d − 1)t ,
and
νmax = λmax (As1 )t ≥ (d/2)t .
This means that
νmax
≥
µmax
d
√
4 d−1
t
≥ eΩ(log n/k) = nΩ(1/k) ,
as long as 4√dd−1 > e, which happens for sufficiently large d (such as d = 127). It only remains
to show that ei (µ) = ei (ν) for i = 1, . . . , k. For every i ∈ [k] we have t · i < girth(G), which by
Lemma 3.9 gives us
pi (ν) = pt·i (λ(As1 )) = pt·i (λ(As2 )) = pi (ν),
and this finishes the proof because of Corollary 2.8.
The above method unfortunately does not seem to directly extend to the regime k ≥ Θ(log n),
since for large k, getting girth(G) > k requires many vertices of degree at most 2.1 Instead we use
the machinery of Chebyshev polynomials to boost our graph constructions.
Proof of Theorem 3.2 for k ≥ Θ(log n). Since Theorem 3.7 proves the same desired bound as
Theorem 3.2 when k = Ω(n), we may without loss of generality assume that n/k is at least a large
enough constant. Let m be the largest integer such that
c·
n
m
≤ ,
log m
k
where c is a large constant that we will fix later. It is easy to see that
n
n
log( ) .
m=Θ
k
k
(1)
(2)
We have already proved Theorem 3.2 for the small k regime. Using this proof (for n = m and
k = Θ(log m)), we can find µ̃, ν̃ ∈ Rm
+ such that ei (µ̃) = ei (ν̃) for i = 1, . . . , Ω(log m) and
ν̃max /µ̃max ≥ m1/Θ(log m) ≥ 2. Without loss of generality, by a simple scaling, we may assume that
ν̃max = 1 and µ̃max ≤ 1/2.
Let χµ̃ (x), χν̃ (x) ∈ R[x] be the unique monic polynomials whose roots are µ̃, ν̃ respectively. By
construction, the top Ω(log m) coefficients of these polynomials are the same. We boost the number
of equal coefficients by composing them with Chebyshev polynomials. Let p(x) := χµ̃ (Tt (x)) and
1
Unless all degrees are 2 in which case we can actually reprove Theorem 3.7; we omit the details here.
13
q(x) := χν̃ (Tt (x)) where t = ⌊n/m⌋. Note that deg p = deg q = tm ≤ n. We let µ, ν be the roots
of p(x), q(x) together with some additional zeros to make their counts n. First, we show that the
top k coefficients of p, q are the same. Then we show that they are real rooted, i.e., µ, ν are real
vectors. Finally, we lower bound νmax /µmax .
Note that p(x), q(x) have degree tm. They are not monic, but their leading terms are the same.
Besides the leading terms, we claim that they have the same top Ω(t log m) coefficients. This follows
from the fact that Tt (x) is a degree t polynomial. When expanding either χν̃ (Tt (x)) or χµ̃ (Tt (x)),
terms of degree ≤ m − Ω(log m) in χµ̃ , χν̃ produce monomials of degree at most tm − Ω(t log m),
which means that the top Ω(t log m) coefficients are the same. This shows that the first Ω(t log m)
elementary symmetric polynomials of µ, ν are the same. It follows that the first k elementary
symmetric polynomials to be the same, which follows from
Ω(t log m) = Ω(
n log m
) ≥ Ω(ck) ≥ k,
m
where for the first inequality we used (1) and for the second inequality we assumed c is large enough
that it cancels the hidden constants in Ω.
It is not obvious if µ, ν are even real. This is where we crucially use the properties of the
Chebyshev polynomial Tt (x). Note that the roots of χµ̃ (x) and χν̃ (x), i.e. µ̃i ’s and ν̃i ’s are all in
[0, 1] ⊆ [−1, 1]. Therefore each one of them can be written as cos(θ) for some θ ∈ R. By Lemma 3.8
the equation
Tt (x) = cos(θ),
has t real roots (counting multiplicities), and they are simply x = cos( θ+2πi
t ) for i = 0, . . . , t − 1.
So for each root of χµ̃ (x) we have t roots of χµ̃ (Tt (x)), all in [−1, 1]. This means that all of the
roots of p(x) are real and in [−1, 1]. By a similar argument, all of the roots of q(x) are real and in
[−1, 1].
The largest root of q(x) is 1, since
q(1) = χν̃ (Tt (1)) = χν̃ (1) = χν̃ (νmax ) = 0.
On the other hand, the largest root of p(x) is at most cos(π/3t), because for any x ∈ (cos(π/3t), 1]
there is θ ∈ [0, π/3t) such that x = cos(θ) and this means
p(x) = χµ̃ (Tt (cos(θ))) = χµ̃ (cos(tθ)) 6= 0,
because cos(tθ) > cos(π/3) = 1/2 ≥ µ̃max .
By the above arguments, µ, ν satisfy almost all of the desired properties, except that they could
be negative. However we know that µ, ν ∈ [−1, 1]n . So using Fact 2.4 we can easily make them
nonnegative. We simply replace µ, ν by µ + 1, ν + 1. Then µ, ν ∈ [0, 2]n and ei (µ) = ei (ν) for
i = 1, . . . , k. Finally, we have
νmax
1+1
= 1+Ω
≥
µmax
1 + cos(π/3t)
1
t2
=1+Ω
m 2
n
=1+Ω
log 2n
k
k
!2
,
where we used (2) for the last equality.
Having finished the proofs of Theorem 3.2, Theorem 3.3 we are ready to prove Lemma 3.9.
14
Proof of Lemma 3.9. By Corollary 2.8, it is enough to prove that
pk (λ(D + As1 )) = pk (λ(D + As2 )),
for k < girth(G). This is the same as proving
Tr (D + As1 )k = Tr (D + As2 )k .
(3)
For a matrix M ∈ Rn×n we have the following identity
X
Mv1 ,v2 Mv2 ,v3 . . . Mvk−1 ,vk Mvk ,v1 .
Tr(M k ) =
(v1 ,...,vk )∈[n]k
For ease of notation let us identify vk+1 with v1 . We apply the above formula to both sides of (3).
The sequence (v1 , . . . , vk ) can be interpreted as a sequence of vertices in the graph G. If for any i,
vi 6= vi+1 and {vi , vi+1 } ∈
/ E, then the term inside the sum vanishes. Therefore we can restrict the
sum to those terms (v1 , . . . , vk ) where for each i ∈ [k], either vi = vi+1 or vi and vi+1 are connected
in G. To borrow and abuse some notation from Markov chains, let us call such a sequence a lazy
closed walk of length k. In order to prove (3) it is enough to prove that for any such lazy closed
walk we get the same term for both s1 and s2 .
Let (v1 , . . . , vk ) be one such lazy closed walk. Consider a particle that at time i resides at vi .
In each step, the particle either does not move or moves to a neighboring vertex, and at time k + 1
it returns to its starting position. For each step that the particle does not move we get one of the
entries of D, corresponding to the current vertex, as a factor. This is clearly independent of the
signing. When the particle moves however, we get the sign of the edge over which it moved as a
factor. We will show that the particle must cross each edge an even number of times. Therefore
the signs for each edge cancel each other and we get the same result for s1 , s2 .
Consider the induced subgraph on v1 , . . . , vk . Because k < girth(G), this subgraph has no
cycles; therefore it must be a tree. A (lazy) closed walk crosses any cut in a graph an even number
of times. Each edge in this tree constitutes a cut. Therefore the lazy closed walk (v1 , . . . , vk ) must
cross each edge in the tree an even number of times.
3.3
Lower Bound Given Approximate Coefficients
The lower bounds proved in the previous section show that knowing a small number of the coefficients of a polynomial exactly is insufficient to obtain a good estimate of its largest root.
In this section we generalize the construction of Theorem 3.2 to provide a satisfying lower bound
to Problem 1.2.
Proposition 3.12. For every integer n > 1 and 1 < k < n there are degree n polynomials r(x), s(x)
such that
1. All of the coefficients of r and s except for the 2kth are exactly equal, and the 2kth coefficients
are within a multiplicative factor of 1 + 242k .
2. The largest root of r is at least 1 + Ω(1/k 2 ) of the largest root of s.
3. r and s have a common interlacing.
15
4. r and s are characteristic polynomials of graph Laplacians. Further, these Laplacians correspond to 2−lifts of a common base graph.
Proof. Let
r(x) := 2Tk2 (3/2 − x)
Since
and
s(x) := T2k (3/2 − x).
T2k = 2Tk2 − 1,
these polynomials differ only in their constant terms. Moreover, we have
r(0) = 2Tk2 (3/2) ≥ (2k−1 )2
and
s(0) ≥ 22k−1 /2
by Fact 2.13. Thus, they agree on the first 2k − 1 coefficients, and differ by a multiplicative factor
of at most
22k−2 + 1
= 1 + 4/22k
22k−2
in the 2kth coefficient, establishing (1).
To see (2), observe that r(x) has largest root 3/2 + cos(2π/k) whereas s(x) has largest root
3/2 + cos(2π/2k). Since the difference of these numbers is
4π 2 1
1
( −
+ o(1/k4 )),
2 k2 4k2
we conclude that their ratio is at least 1 + Ω(1/k 2 ).
To see (3), observe that the roots of r(x) are
cos(2π/k) − cos(2π/2k) =
3/2 − cos(2πj/k)
j = 0, . . . , k − 1 with multiplicity 2
and the roots of s(x) are
3/2 − cos(2πj/2k)
j = 0, . . . , 2k − 1 with multiplicity 1,
whence r and s have a common interlacing according to Definition 4.1.
For (4) we first apply Fact 2.14 to interpret both r and s as characteristic polynomials of cycles.
Let Ck denote a cycle of length k and let Ck ∪ Ck denote a union of two such cycles. Then we have:
det(2xI − ACk ∪Ck ) = det(2xI − ACk )2 = 4Tk (x)2 = 2r(3/2 − x),
and
det(2xI − AC2k ) = 2T2k (x) = 2s(3/2 − x),
whence
r(x) =
(−2)k
1
det(3I − ACk ∪Ck − 2xI) =
det(xI − ((3/2)I − (1/2)ACk ∪Ck )
2
2
and
(−2)k
1
det(3I − AC2k − 2xI) =
det(xI − ((3/2)I − (1/2)C2k ),
2
2
which are characteristic polynomials of graph Laplacians of weighted graphs with self loops. Note
that both graphs are 2−covers of Ck . Since multiplying by constants does not change any of the
properties we are interested in, we can ignore them.
Considering r̃(x) = xn−k r(x) and s̃(x) = xn−k s(x) yields examples of the desired dimension n;
note that multiplying by xn−k simply corresponds to adding isolated vertices to the corresponding
graphs.
s(x) =
16
4
Applications to Interlacing Families
√
3
In this section we use Theorem 3.1 to give an 2Õ( m) time algorithm for rounding an interlacing
family of depth m. Let us start by defining an interlacing family.
Qn−1
Definition 4.1 (Interlacing). We say that
a
real
rooted
polynomial
g(x)
=
α
0
i=1 (x − αi ) interQn
laces a real rooted polynomial f (x) = β0 i=1 (x − βi ) if
β1 ≤ α1 ≤ β2 ≤ α2 ≤ · · · ≤ αn−1 ≤ βn .
We say that polynomials f1 , . . . , fk have a common interlacing if there is a polynomial g such
that g interlaces all fi . The following key lemma is proved in [MSS13a].
Lemma 4.2. Let f1 , . . . , fk be polynomials of the same degree that are real rooted and have positive
leading coefficients. If f1 , . . . , fk have a common interlacing, then there is an i such that
λmax (fi ) ≤ λmax (f1 + · · · + fk ).
Definition 4.3 (Interlacing Family). Let S1 , . . . , Sm be finite sets. Let F ⊆ S1 × S2 × · · · × Sm be
nonempty. For any s1 , s2 , . . . , sm ∈ F, let fs1 ,...,sm (x) be a real rooted polynomial of degree n with
a positive leading coefficient. For s1 , . . . , sk ∈ S1 × · · · × Sk with k < m, let
Fs1 ,...,sk := {t1 , . . . , tm ∈ F : si = ti , ∀1 ≤ i ≤ k}.
Note that F = F∅ . Define
fs1 ,...,sk =
X
ft1 ,...,tm ,
t1 ,...,tm ∈Fs1 ,...,sk
and
f∅ =
X
ft1 ,...,tm .
t1 ,...,tm ∈F
We say polynomials {fs1 ,...,sm }s1 ,...,sm ∈F form an interlacing family if for all 0 ≤ k < m and all
s1 , . . . , sk ∈ S1 × · · · × Sk the following holds: The polynomials fs1 ,...,sk ,ti which are not identically
zero have a common interlacing.
In the above definition we say m is the depth of the interlacing families. It follows by repeated
applications of Lemma 4.2 that for any interlacing family {fs1 ,...,sm }s1 ,...,sm ∈F , there is a polynomial
fs1 ,...,sm such that the largest root of fs1 ,...,sm is at most the largest root of f∅ [MSS13b, Thm 3.4].
For an α > 1, an α-approximation rounding algorithm for an interlacing family {fs1 ,...,sm }s1 ,...,sm ∈F
is an algorithm that returns a polynomial fs1,...,sm such that
λmax (fs1 ,...,sm ) ≤ αλmax (f∅ ).
Next, we design such a rounding algorithm, given an oracle that computes the first k coefficients
of the polynomials in an interlacing family.
Theorem 4.4. Let S1 , . . . , Sm be finite sets and let {fs1 ,...,sm }s1 ,...,sm ∈F be an interlacing family of
degree n polynomials. Suppose that we are given an oracle that for any 1 ≤ k ≤ n and s1 , . . . , sℓ ∈
S1 , . . . , Sℓ with ℓ < m returns the top k coefficients of fs1 ,...,sm in time T (k). Then, there is an
algorithm that for any ǫ > 0 returns a polynomial fs1 ,...,sm such that the largest root of fs1 ,...,sm is
√
1/3
at most 1 + ǫ times the largest root of f∅ , in time T (O(log(n)m1/3 / ǫ)) max{|Si |}O(m ) poly(n).
17
Proof. Let M = m1/3 and k ≍ √1ǫ log(n)M . Then, by Theorem 3.1, for any polynomial fs1,...,sℓ we
can find a 1 + 2Mǫ 2 approximation of the largest root of fs1,...,sm in time T (k) poly(n). We round
the interlacing family in M 2 many steps and in each step we round M of the coordinates. We make
sure that each step only incurs a (multiplicative) approximation of 1 + 2Mǫ 2 so that the cumulative
approximation error is no more than
ǫ M2
≤1+ǫ
)
(1 +
2M 2
as desired.
Let us describe the algorithm inductively. Suppose we have selected s1 , . . . , sℓ for some 0 ≤
ℓ < m. We brute force over all polynomials fs1 ,...,sℓ ,tℓ+1 ,...,tℓ+M which are not identically zero for all
tℓ+1 , . . . , tℓ+M ∈ Sℓ+1 , . . . , Sℓ+M . Note that there are at most (maxi |Si |)M many such polynomials.
For any polynomial fs1 ,...,sℓ ,tℓ+1 ,...,tℓ+M (which is not identically zero) we compute µ∗s1 ,...,sℓ ,tℓ+1 ,...,tℓ+M ,
a 1 + 2Mǫ 2 approximation of its largest root using its top k coefficients. We let
sℓ+1 , . . . , sℓ+M = argmin µ∗s1 ,...,sℓ ,tℓ+1 ,...,tℓ+M .
tℓ+1 ,...,tℓ+M
It follows that the algorithm runs in time T (k) poly(n)(maxi |Si |)O(M ) . Because we have an
interlacing family, there is a polynomial fs1 ,...,sℓ ,tℓ+1 ,...,tℓ+M whose largest root is at most the largest
root of fs1 ,...,sℓ , in each step of the algorithm. Therefore,
λmax (fs1 ,...,sℓ+M ) ≤ (1 +
ǫ
)λmax (fs1 ,...,sℓ ).
2M 2
Therefore, by induction,
as desired.
λmax (fs1 ,...,sm ) ≤ (1 + ǫ)λmax (f∅ )
We remark that without the use of Chebyshev
polynomials and Theorem 3.1, one can obtain
√
the somewhat worse running time of 2Õ( m) by applying the same trick of rounding the vectors in
groups rather than one at a time.
To use the above theorem, we need to construct the aforementioned oracle for each application
of the interlacing families. Next, we construct such an oracle for several examples.
4.1
Oracle for Kadison-Singer
We start with interlacing families corresponding to the Weaver’s problem which is an equivalent
formulation of the Kadison-Singer problem. Marcus, Spielman, and Srivastava proved the following
theorem.
Theorem 4.5 ([MSS13b]). Given vectors v1 , . . . , vm in isotropic position,
m
X
vi viT = I,
i=1
such that maxi kvi k2 ≤ δ, there is a partitioning S1 , S2 of [m] such that for j ∈ {1, 2},
X
i∈Sj
√
vi viT ≤ 1/2 + O( δ).
18
1/3
We give an algorithm that finds the above partitioning and runs in time 2m /δ . It follows from
the proof of [MSS13b] that it is enough to design a (1 + δ)-approximation rounding algorithm for
the following interlacing family.
Let r1 , . . . , rm ∈ Rn be independent random vectors where for each i, Si is the support of ri .
For any r1 , . . . , rm ∈ S1 × . . . Sm let
fr1 ,...,rm (x) = det(xI −
m
X
ri rTi ).
i=1
Marcus et al. [MSS13b] showed that {fr1 ,...,rm }r1 ,...,rm ∈S1 ×···×Sm is an interlacing family. Next we
design an algorithm that returns the first k coefficients of any polynomial in this family in time
(m · maxi |Si |)k .
Theorem 4.6. Given independent random vectors r1 , . . . , rm ∈ Rd with support S1 , . . . , Sm . There
is an algorithm that for any r1 , . . . , rℓ ∈ S1 × · · · × Sℓ with 1 ≤ ℓ ≤ m and 1 ≤ k ≤ n returns the
top k coefficients of fr1,...,rℓ in time (m · maxi |Si |)k poly(n).
Proof. Fix r1 , . . . , rℓ ∈ S1 × · · · × Sℓ for some 1 ≤ ℓ ≤ m. It is sufficient to show that for any
0 ≤ k ≤ n, we can compute the coefficient of xn−k of fr1 ,...,rℓ in time (m · maxi |Si |)k poly(n). First,
observe that
"
!#
m
X
T
fr1 ,...,rℓ = Erℓ+1,...,rm det xI −
ri ri
.
i=1
So, by Proposition 2.1, the coefficient of
(−1)k
xn−k
in the above is equal to
!#
"
X
X
.
ri rTi
Erℓ+1,...,rm σk
i∈T
T ⊆([m]
k )
k
Note that there are at most
in the above summation. For any T ⊆ [m]
k , we can exactly
P m terms
compute Erℓ+1,...,rm σk ( i∈T ri rTi ) in time (maxi |Si |)k poly(n). All we need to do is brute force
over all vectors in the domain of {Si }i>ℓ,i∈T and average out σk (.) of the corresponding sums of
rank 1 matrices.
It follows by Theorem 4.4 and Theorem 4.6 that for any given set of vectors v1 , . . . , vm ∈ Rn
in isotropic position of squared norm at most δ we can find a two partitioning S1 , S2 such that
√
P
T ≤ 1/2 + O( δ) in time nO(m1/3 δ−1/4 ) .
i∈Sj vi vi
4.2
Oracle for ATSP
Next, we construct an oracle for interlacing families related to the asymmetric traveling salesman
problem. We say a multivariate polynomial p ∈ R[z1 , . . . , zm ] is real stable if it has no roots in the
upper-half complex plane, i.e., p(z) 6= 0 whenever Im(zi ) > 0 for all i. The generating polynomial
of µ is defined as
X
Y
gµ (z1 , . . . , zm ) =
µ(S)
zi .
S⊆[m]
i∈S
We say µ is a strongly Rayleigh probability distribution if gµ (z) is real stable. We say µ is homogeneous if all sets in the support of µ have the same size. The following theorem is proved by the
first and the second authors [AO15b].
19
Theorem 4.7 ([AO15b]). Let µ be a homogeneous strongly Rayleigh probability distribution on
subsets of [m] such that for each i, P [i] < ǫ1 . Let v1 , . . . , vm ∈ Rn be in isotropic position such
that for each i, kvi k2 ≤ ǫ2 . Then, there is a set S in the support of µ such that
X
i∈S
vi viT ≤ O(ǫ1 + ǫ2 ).
We give an algorithm that finds such a set S as promised in the above theorem assuming that
we have an oracle that for any z ∈ Rm returns gµ (z). It follows from the proof of [AO15b] that it is
enough to design a 1 + O(ǫ1 + ǫ2 )-approximation rounding algorithm for the following interlacing
family.
For any 1 ≤ i ≤ m, let Si = {0, vi }. For any S in the support of µ, let ri = vi if i ∈ S and
ri = 0 otherwise and we add r1 , . . . , rm to F. Then, define
!
m
X
T
ri ri .
fr1 ,...,rm = µ(S) · det xI −
i=1
It follows from [AO15b] that {fr1 ,...,rm }r1 ,...,rm ∈F is an interlacing family.
Next, we design an algorithm that returns the top k coefficients of any polynomial in this family
in time mk poly(n).
Theorem 4.8. Given a strongly Rayleigh distribution µ on subsets of [m] and a set of vectors
v1 , . . . , vm ∈ Rn , suppose that we are given an oracle that for any z ∈ Rm returns gµ (z). There is
an algorithm that for any r1 , . . . , rℓ ∈ S1 × · · · × Sℓ with 1 ≤ ℓ ≤ m and 1 ≤ k ≤ n returns the top
k coefficients of fr1 ,...,rℓ in time mk poly(n).
Proof. Fix r1 , . . . , rℓ ∈ S1 × · · · × Sℓ for some 1 ≤ ℓ ≤ m. First, note that if there is no such
element in F, then gµ (1, . . . , 1, zℓ+1 , . . . , zm ) = 0 and there is nothing to prove. So assume for some
rℓ+1 , . . . , rm , r1 , . . . , rm ∈ F.
It is sufficient to show that for any 0 ≤ k ≤ n, we can compute the coefficient of xn−k of
fr1 ,...,rℓ in time mk poly(n). Firstly, observe that since we are only summing up the characteristic
polynomials that are consistent with r1 , . . . , rℓ , we can work with the conditional distribution
µ̃ = {µ | i if 1 ≤ i ≤ ℓ and ri = vi , i if 1 ≤ i ≤ ℓ, ri = 0}.
Note that since we can efficiently compute gµ (z1 , . . . , zm ), we can also compute gµ̃ (zℓ+1 , . . . , zm ).
For any i that is conditioned to be in, we need to differentiate with respect to zi and for any i that
is conditioned to be out we let zi = 0. Also, note that instead of differentiating we can let zi = M
for a very large number M , and then divide the resulting polynomial by M . We note that when
µ is a determinantal distribution, which is the case in applications to the asymmetric traveling
salesman problem, this differentiation can be computed exactly and efficiently; in other cases, M
can be taken to be exponentially large, as we can tolerate an exponentially small error.
Now, we can write
!#
"
X
T
.
vi vi
fr1 ,...,rℓ = ET ∼µ̃ det xI −
i∈T
20
So, by Proposition 2.1, the coefficient of xn−k in the above is equal to
!
X
X
T
k
vi vi .
Pµ̃ [T ] · σk
(−1)
i∈T
T ∈([m]
k )
To compute the aboveP
quantity it is enough to brute force over all sets T ∈ [m]
k . For any such
T we can compute σk ( i∈T vi viT ) in time poly(n). In addition, we can efficiently compute Pµ̃ [T ]
using our oracle. It is enough to differentiate with respect to any i ∈ T ,
Y
i∈T :i>ℓ
∂
gµ̃ (zℓ+1 , . . . , zm )|zℓ+1 =···=zm =1 .
∂zi
Therefore, the algorithm runs in time mk poly(n).
It follows from Theorem 4.4 and Theorem 4.8 that for any homogeneous strongly Rayleigh
distribution µ with marginal probabilities ǫ1 with an oracle that computes gµ (z) for any z ∈ Rn ,
and for any vectors v1 , . . . , vm in isotropic
position with squared norm at most ǫ2 , we can find a
P
T ≤ O(ǫ + ǫ ) in time nO(m1/3 (ǫ1 +ǫ2 )1/2 ) .
set S in the support of µ such that
v
1
2
i∈S i vi
This is enough to get a polyloglog(m) approximation algorithm for asymmetric traveling sales1/3
man problem on a graph with m vertices that runs in time 2Õ(m ) [AO15a].
References
[AO15a]
Nima Anari and Shayan Oveis Gharan. Effective-Resistance-Reducing Flows, Spectrally
Thin Trees, and Asymmetric TSP. pages 20–39, 2015. FOCS. 3, 21
[AO15b]
Nima Anari and Shayan Oveis Gharan. The kadison-singer problem for strongly rayleigh
measures and applications to asymmetric tsp. 2015. 3, 19, 20
[Bec91]
József Beck. An algorithmic approach to the Lovász local lemma. I. Random Structures
Algorithms, 2(4):343–365, 1991. 3
[BL06]
Yonatan Bilu and Nathan Linial. Lifts, discrepancy and nearly optimal spectral gap.
Combinatorica, 26(5):495–519, 2006. 3
[Coh16]
Michael B Cohen.
Ramanujan graphs in polynomial time.
arXiv:1604.03544, 2016. 3
[CS07]
Maria Chudnovsky and Paul Seymour. The roots of the independence polynomial of a
clawfree graph. J. Combin. Theory Ser. B, 97(3):350–357, 2007. 3
[DP15]
Constantinos Daskalakis and Christos Papadimitriou. Sparse covers for sums of indicators. Probab. Theory Related Fields, 162(3-4):679–705, 2015. 4
[Erd64]
P. Erdős. Extremal problems in graph theory. In Theory of Graphs and its Applications
(Proc. Sympos. Smolenice, 1963), pages 29–36. Publ. House Czechoslovak Acad. Sci.,
Prague, 1964. 7
21
arXiv preprint
[FL06]
Shmuel Friedland and Daniel Levy. A polynomial-time approximation algorithm for the
number of k-matchings in bipartite graphs. arXiv preprint cs.CC/0607135, 28, 2006. 4
[HL72]
Ole J. Heilmann and Elliott H. Lieb. Theory of monomer-dimer systems. Comm. Math.
Phys., 25:190–232, 1972. 3
[HPS15]
Chris Hall, Doron Puder, and William F Sawin. Ramanujan coverings of graphs. arXiv
preprint arXiv:1506.02335, 2015. 3
[HSV16]
Nicholas JA Harvey, Piyush Srivastava, and Jan Vondrák. Computing the independence
polynomial in shearer’s region for the lll. arXiv preprint arXiv:1608.02282, 2016. 3
[JSV04]
Mark Jerrum, Alistair Sinclair, and Eric Vigoda. A polynomial-time approximation
algorithm for the permanent of a matrix with nonnegative entries. J. ACM, 51(4):671–
697, July 2004. 4
[LU95]
Felix Lazebnik and Vasiliy A. Ustimenko. Explicit construction of graphs with an arbitrary large girth and of large size. Discrete Appl. Math., 60(1-3):275–284, 1995. ARIDAM
VI and VII (New Brunswick, NJ, 1991/1992). 7
[MSS13a] Adam Marcus, Daniel A. Spielman, and Nikhil Srivastava. Interlacing Families I: Bipartite Ramanujan Graphs of All Degrees. In FOCS, pages 529–537, 2013. 3, 4, 8,
17
[MSS13b] Adam Marcus, Daniel A Spielman, and Nikhil Srivastava. Interlacing Families II: Mixed
Characteristic Polynomials and the Kadison-Singer Problem. 2013. 3, 4, 5, 17, 18, 19
[MSS15]
Adam W. Marcus, Nikhil Srivastava, and Daniel A. Spielman. Interlacing families iv:
Bipartite ramanujan graphs of all sizes. to appear in FOCS, 2015. 3
[MT10]
Robin A. Moser and Gábor Tardos. A constructive proof of the general Lovász local
lemma. J. ACM, 57(2):Art. 11, 15, 2010. 3
[SJ89]
Alistair Sinclair and Mark Jerrum. Approximate counting, uniform generation and
rapidly mixing Markov chains. Inf. Comput., 82(1):93–133, July 1989. 4
[SS13]
Alistair Sinclair and Piyush Srivastava. Lee-Yang theorems and the complexity of computing averages. In STOC’13—Proceedings of the 2013 ACM Symposium on Theory of
Computing, pages 625–634. ACM, New York, 2013. 3
[Sze39]
Gabor Szeg. Orthogonal polynomials, volume 23. American Mathematical Soc., 1939. 7
[Val79]
L. G. Valiant. The complexity of computing the permanent. Theoret. Comput. Sci.,
8(2):189–201, 1979. 3
[Wen91]
R. Wenger. Extremal graphs with no C 4 ’s, C 6 ’s, or C 10 ’s. J. Combin. Theory Ser. B,
52(1):113–116, 1991. 7
[Wil84]
James H. Wilkinson. The perfidious polynomial. In Studies in numerical analysis,
volume 24 of MAA Stud. Math., pages 1–28. Math. Assoc. America, Washington, DC,
1984. 2
22
A
Appendix A: Proofs of Preliminary Facts
Proof of Fact 2.4. If a = 0, the conclusion is trivial. Otherwise let ℓ(x) = (x − b)/a. Then aµ + b
is the vector of the roots of χµ (ℓ(x)) and similarly aν + b is the vector of the roots of χν (ℓ(x)).
It is easy to see that if χµ and χν have the same top k coefficients, then after expansion χµ (ℓ(x))
and χν (ℓ(x)) have the same top k coefficients as well, since ℓ(x) is linear.
Proof of Corollary 2.6. For p(x) = xk , the statement of the corollary becomes the same as Theorem 2.5.
For any other polynomial, we can write
p(x) = a0 + a1 x + · · · + ak xk .
Then we have
n
X
p(µi ) =
k
X
aj
j=1
i=1
n
X
i=1
µji
!
=
k
X
aj pj (µ).
j=1
Now because of Theorem 2.5, it follows that
n
X
p(µi ) =
i=1
k
X
aj qj (e1 (µ), . . . , ek (µ)).
j=1
The above is a polynomial of e1 (µ), . . . , ek (µ), and it can be computed in time poly(k), since each
qj can be computed in time poly(k).
Proof of Fact 2.13. For x ≥ 0, we have Tk (1+x) = cosh(k cosh−1 (1+x)); here cosh−1 is the inverse
of cosh when looked at as a function from [0, ∞) to [1, ∞). Both cosh and cosh−1 are monotonically
increasing over the appropriate ranges. Therefore Tk (1 + x) is monotonically increasing for x ≥ 0.
For x ≥ 0, we have
exp(k cosh−1 (1 + x))
√ k2
−1
k
(1 + 2x)
exp(cosh (1 + x))
≥
.
=
2
2
Tk (1 + x) = cosh(k cosh−1 (1 + x)) ≥
In the first inequality we used the fact
that cosh(x) ≥ ex /2, and in the last inequality we used the
√
fact that exp(cosh−1 (1 + x)) ≥ 1 + 2x.
Proof of Fact 2.22. The maximum eigenvalue of A is characterized by the so called Rayleigh quotient
x⊺ Ax
,
λmax (A) = max
x∈Rn −{0} x⊺ x
For x = 1, the all 1s vector, we have
λmax (A) ≥
1⊺ A1
2|E|
=
= degavg (G).
⊺
1 1
n
23
B
Appendix B: Constructions of Graphs with Large Girth
In this section we prove Claim 3.10 and Claim 3.11.
Proof of Claim 3.10. Assume that k is fixed and Conjecture 2.16 is true for graphs of girth > 2k.
The graphs promised by Conjecture 2.16 already have average degree Ω(n1/k ). We just need to
make them bipartite and make sure that their maximum degree is at most O(n1/k ).
Making the graph bipartite is easy. Given a graph G, the following procedure makes it bipartite:
Replace each vertex v by v1 , v2 . Replace each edge {u, v} with two edges {v1 , u2 } and {u1 , v2 }.
This procedure doubles the number of edges and the number of vertices, and it is easy to see that
it does not decrease the girth.
We can trim the maximum degree by the following procedure: If there is a vertex v where
deg v > n1/k , we introduce a new vertex v ′ and take an arbitrary subset of n1/k edges incident to
v and change their endpoint from v to v ′ . By repeating this procedure the graph will eventually
have maximum degree bounded by n1/k . This procedure increases the number of vertices, but the
number of new vertices is easily bounded by 2|E|/n1/k because each new vertex has degree n1/k
until the end and the number of edges |E| does not change.
Both of the above procedures change the number of vertices. To get graphs of arbitrary size
n, we can use the following trick: We start with a graph G0 promised by Conjecture 2.16 on n/c
vertices, where c is a large constant. We make sure that it has no more than (n/c)1+1/k edges
by removing edges if necessary (this can only increase girth). Now we make a bipartite graph G1
from G0 by the aforementioned procedure. From G1 we make G2 using the second procedure which
makes sure degmax (G2 ) = O(n1/k ). The number of new vertices added by this step is O(n/c). So
at the end the total number of vertices in G2 is O(n/c) and if we make c large enough we can make
sure it is less than n. We can add isolated vertices to G2 until the total number of vertices becomes
n and we get G. This procedure changes degavg by at most a constant factor and does not change
degmax , so both are Θ(n1/k ) at the end.
Proof of Claim 3.11. We build our graphs using Theorem 2.17. Let t ≥ 3 be the largest odd number
such that 2dt ≤ n. If n is large enough, such a t exists, and we have t = Ω(log n). By Theorem 2.17,
there exists a d-regular bipartite graph H with 2dt vertices and girth(H) ≥ t + 5 = Ω(log n). To
construct G, put as many copies of H as possible side by side making sure the number of vertices
does not grow more than n. At the end add some isolated vertices to make the total number of
vertices n and get G. It is easy to see that number of isolated vertices added at the end is at most
n/2. These vertices have degree 0 and the rest have degree d. Therefore degavg (G) ≥ d/2 and
degmax (G) = d.
24
| 8cs.DS
|
Reasoning in Systems with Elements that Randomly Switch Characteristics
Subhash Kak
Abstract.
We examine the issue of stability of probability in reasoning about complex systems with
uncertainty in structure. Normally, propositions are viewed as probability functions on an
abstract random graph where it is implicitly assumed that the nodes of the graph have stable
properties. But what if some of the nodes change their characteristics? This is a situation that
cannot be covered by abstractions of either static or dynamic sets when these changes take
place at regular intervals. We propose the use of sets with elements that change, and modular
forms are proposed to account for one type of such change. An expression for the dependence
of the mean on the probability of the switching elements has been determined. The system is
also analyzed from the perspective of decision between different hypotheses. Such sets are
likely to be of use in complex system queries and in analysis of surveys.
Introduction
For reasoning in complex systems, one often uses static data structures designed for situations
that do not change after they are constructed. Static sets allow only query operations on their
elements — such as checking whether a given value is in the set, or enumerating the values in
some arbitrary order. In dynamic sets, on the other hand, one can insert and delete elements.
But what if the elements stay the same in number, yet their properties change randomly?
Consider a very basic model where the set of N objects is viewed as consisting of nA objects of
Type A and the remaining of Type B, where the properties of the two types are distinct. The
observation process consists of choosing a random element and examining it and then
replacing it in the box. By repeated interaction with the system, the probability of A is easily
computed:
P( A) =
nA
N
(1)
Let the outcomes A and B be mapped to the random variable X with values 1 and -1,
respectively. We can pick the threshold of 0 to determine which probability is greater. If
∑ 𝑋𝑋𝑖𝑖 > 0, then P(A) > P(B).
Now assume that the objects are actually of 3 types: A, B, and C with numbers that are nA, nB,
and nC, nA+nB+nC=N, where C nodes are special in the sense that during query (interaction) they
project the properties of A and B in some random fashion? One could also view the situation as
1
consisting of nC objects of Type C that are non-classical or one could view Class C as a random
set whose elements change between Type A and B.
An example of this will be polling surveys. Assume there are three political parties but only two
candidates who belong to Parties A and B. The members of the third party must pick candidate
A or B, where their choice varies from day to day based on news and public opinion (for
simplicity, we assume here that members of the other two parties vote for their party
candidates, for in reality there will be some cross-voting from the other groups as well).
This situation represents unstable sets with a subset of floating elements and this fact should
be considered in determining system behavior.
Figure1. Elements of C switch characteristics randomly between A and B
The probability P(A) will not stabilize as the system is queried (say on different days) and its
value will fall in a range:
𝑛𝑛𝐴𝐴
𝑁𝑁
≤ 𝑃𝑃(𝐴𝐴) ≤
𝑛𝑛𝐴𝐴 +𝑛𝑛𝐶𝐶
(2)
𝑁𝑁
The C nodes may be seen as creating noise in the measurement and if indeed real systems
behave like this this property needs to be taken into consideration and its impact on reasoning
evaluated.
For purpose of illustration, assume nA=5, nB=2, and and nC=3. If the transitions occur at the end
of the day, then sampling of different systems will change the estimate of elements of Class A
as in Figure 2 below.
2
10
8
6
4
2
0
0
10
20
30
40
Series1
Figure 2. Changing count of the elements of A
As example of importance of such sets, consider that in medicine, a patient might respond to
different active ingredients in different ways due to a certain element of interchangeability in
the structure. The changeability will complicate analysis of Simpson’s and related paradoxes
[1][2]. The Simpson’s paradox arises from the fact that if a1 / b1 > c1 / d1 and a2 / b2 > c2 / d 2 there
will be situations so that ( c1 + c2 ) /(d1 + d 2 ) > ( a1 + a2 ) /(b1 + b2 ) , but if there is range associated
with the numerator values, we get a more complicated picture related to the ranges where
probability inversion occurs as seen in Table 1.
Table 1. Probability inversion
Week 1
Week 2
Lisa
0/1
3 or 4/5
Bart
1/4
2 or 3/3
Total
3 or 4/6
3 or 4/7
Lisa and Bart are assigned the task of editing articles. In Week 2, their work gets further edited
differently by supervisors. The performance of Bart is superior to Lisa’s both weeks (excepting
when Lisa has done 4/5 and Bart has done 2/3) but when the work is aggregated, Lisa is better.
Our consideration of interswitching may make it possible to do an indirect accounting of
interaction networks that exist in the body and the complicating influence of information and
regulatory networks to examine the placebo effect [3]. This effect is counterintuitive in that an
inert substance turns out to be an effective treatment for some patients, and it arises out of the
information flows and expectations associated with the treatment.
In some cases the interaction between elements may rightly be seen through the lens of
quantum decision theory, especially when there is evidence of superpositional behavior by the
nodes [4][5]. One major consequence of switching nodes in the quantum context is that
probability measurements are not commutative. Thus the result P(T1,T2)≠P(T2,T1) even when
T1 and T2 are apparently independent. The interaction disturbs the system that causes the
3
variables to change, so that the order has an influence on the expectations associated with the
system [6].
These ideas also apply in the field of belief vectors, which represent relations between states
underlying the reasoning process. A specific belief may be a numerical value or a Boolean
variable representing a choice between alternatives and in a basic form it may be seen as
representing a class such as Type A. The choice itself might fluctuate between two different
possibilities that may be viewed as being altered by the random influence projected by Class C.
The dynamics of this projection may be such so that it does go to zero or some other fixed
point.
In this paper, we consider the problem of reasoning with states of the system that change
dynamically with each interaction. Specifically, we consider a model where the internal states
change and a nonlinear mapping is proposed as a possible mechanism. Results on the statistical
characteristics of the observations have been obtained. The system has also been analyzed
from the perspective of noise theory.
Dynamics of the system model
Since the components of the system interact with each other, the system cannot be analyzed
based on classical probability. But since the nature of these interactions is potentially endless,
we must constrain our analysis by means of specific assumptions.
We assume the system to have nodes of type A, B, and C. The C nodes switch into A and B back
and forth just like the population of the predators and prey in an isolated habitat. Such complex
system dynamics will be characterized by fixed points and orbits. Different mechanisms may be
at work in different systems.
Rather than speak of the full system, we will now focus only on the elements of Type C. Let the
probability that an element of C will projected as A be n. Let each interaction with the system
change the proportion of A and B nodes. Assume dynamics determined by the function s: N →
N on the set of positive integers where 0 is taken to be an even number, and
(3)
𝑠𝑠(𝑖𝑖 + 1) = 𝑓𝑓(𝑠𝑠(𝑖𝑖))
where i is the interaction count with the system.
4
Various modular schemes for the dynamic change of probability will be considered. Let s(i) be
the number of Type C nodes that are observed as Class A nodes at ith measurement. Let the
value of i run from 0 to M-1.
Scheme 1
(4)
𝑠𝑠(𝑖𝑖 + 1) = 𝑠𝑠(𝑖𝑖) + 𝑘𝑘 mod 𝑀𝑀
This will make the probability change in a zig-zag manner.
Scheme 2
(5)
𝑠𝑠(𝑖𝑖 + 1) = 𝑘𝑘𝑘𝑘(𝑖𝑖) mod 𝑀𝑀
If M is prime then the nodes will go through a cycle and the cycle is full if k is a primitive
element of the prime.
Example 1. We have M=6, or a total of 5 nodes of Type C, and k = 2. The changes in the number
of Type A nodes on each inspection will be as follows (where the first row represents the
number in the first encounter):
1 2 3 4 5
2 4 0 2 4
4 2 0 4 2
So in this case we have a fixed point (which is 0) and a cycle 2 4.
If the initial number of nodes was not 3, then the probability will switch between the values 4/5
and 2/5. For initial number 3, the probability goes to zero.
Scheme 3
Consider a general modular mapping:
𝑠𝑠(𝑛𝑛) = �
𝑛𝑛
2
, if 𝑛𝑛 is even
(6)
𝑘𝑘𝑘𝑘 + 1 if 𝑛𝑛 is odd
In particular, we consider this transformation for the specific case of k=3:
5
n/2
s( n ) =
3n + 1 mod M
if n is even
(7)
if n is odd
This is the finite version of the 3n+1 – mapping associated with the names of L. Collatz and
other mathematicians [7][8]. The value of M determines the largest orbit that will be associated
with the probability evolution. The initial count of nodes of Type A will be taken to be s(0). If M
is even, the population of Type A nodes will never go down to zero excepting when 3i =1 mod
M.
As example, consider the mapping for mod 17. The numbers 1 through 16 are mapped in one
step as follows:
s(0) values
(9)
Figure 2. The evolution of the initial states with each observation.
The fixed point is 0 and the orbit is 1 4 2 1
If M is odd, the Type A nodes will either have a fixed point or be part of an orbit.
Example 2. If i(0) = 7, and M= 55, one gets the sequence 7, 22, 11, 34, 17, 52, 26, 13, 40, 20, 10,
5, 16, 8, 4, 2, 1; in other words, the orbit or trajectory of 7 has 17 elements and the path length
is 16. The maximum value generated is 52. Let N= 37, for the same initial distribution. We then
get 7, 22, 11, 34, 17, 15, 9, 28, 14, 7, …. which is an orbit of 9.
The trajectory for n = 3 for M=100 is 3, 10, 5, 16, 8, 4, 2, 1, that is it has 8 elements in it or it has
a path length of 7, with a maximum value of 16. The sequence for n = 25, takes 23 steps,
climbing to 88 before descending to 1: ( 25, 76, 38, 19, 58, 29, 88, 44, 22, 11, 34, 17, 52, 26, 13,
40, 20, 10, 5, 16, 8, 4, 2, 1 ).
6
The assumption in the modular mappings above is that the system is classical although highly
nonlinear. However there are situations where this is not true and the consideration of a
quantum model appears justified [9][10[11][12]. This takes us beyond classical sets in a
different manner.
Statistical analysis
Given that nA, nB, and nC are the number of elements belonging to the sets A, B, and C,
respectively, and 𝑁𝑁 = (𝑛𝑛𝐴𝐴 + 𝑛𝑛𝐵𝐵 + 𝑛𝑛𝐶𝐶 ), we can begin with the basic probability associated with
events A and B.
Figure 4. Probability of sets A and B
Let M be the variable representing the short-term average of X. Give the random switching of
the elements of C, M takes values that range over (𝑛𝑛𝐴𝐴 − 𝑛𝑛𝐵𝐵 − 𝑛𝑛𝐶𝐶 )/𝑁𝑁 to (𝑛𝑛𝐴𝐴 − 𝑛𝑛𝐵𝐵 + 𝑛𝑛𝐶𝐶 )/𝑁𝑁.
We will assume that M is a uniformly distributed random variable as shown below:
Figure 4. Distribution of M
The expected value of M, E(M)= E(X):
𝐸𝐸(𝑀𝑀) =
(𝑛𝑛𝐴𝐴 −𝑛𝑛𝐵𝐵 )
(9)
𝑁𝑁
The mean square value of M, E(M2), may be easily computed:
𝐸𝐸(𝑀𝑀2 ) =
And the variance of M is
2
2
2
(3𝑛𝑛𝐴𝐴
+3𝑛𝑛𝐵𝐵
+𝑛𝑛𝐶𝐶
−6𝑛𝑛𝐴𝐴 𝑛𝑛𝐵𝐵 )
(10)
𝑛𝑛2
(11)
3𝑁𝑁 2
𝑉𝑉𝑉𝑉𝑉𝑉(𝑀𝑀) = 3𝑁𝑁𝐶𝐶2
7
Thus the variance of the mean is one-third of the square of the proportion of switching
elements. This variance is independent of the actual values of nA and nB. An estimate of this
variance will be a measure of the size of C, which may be important in determining how many
individuals in the poll belong to a separate class where the choices are not firm.
Class C as Noise
We can also present a noise theory perspective on the problem. Assume, for the sake of
symmetry, that Classes A and B have the same number of elements and that they represent
two different Hypotheses A and B. A number of measurements, Xi, are taken and with A and B
mapped to 1 and -1, the decision taken is by threshold logic:
Choose Hypothesis A if ∑𝑖𝑖 𝑋𝑋𝑖𝑖 ≥ 0;
Otherwise, Hypothesis B.
(12)
Since we assume that the changes in the membership of C occurs with low rate (so that the
surveys on a specific day, when the membership is stable) can be as large as one pleases. One
might also assume now that elements of A and B can also change as in Figure 12 below. The
source probabilities are expressed by subscript “s” and the observation probabilities are
expressed by subscript “o”.
Figure 5. The observation process
We have
𝑝𝑝(𝐴𝐴𝑜𝑜 ) = 𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐴𝐴𝑠𝑠 )𝑝𝑝(𝐴𝐴𝑠𝑠 ) + 𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐵𝐵)𝑝𝑝(𝐵𝐵𝑠𝑠 ) + 𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐶𝐶𝑠𝑠 )𝑝𝑝(𝐶𝐶𝑠𝑠 )
and so on.
(13)
By Bayes’ theorem:
𝑝𝑝(𝐴𝐴𝑠𝑠 |𝐴𝐴𝑜𝑜 ) = 𝑝𝑝(𝐴𝐴
𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐴𝐴𝑠𝑠 )𝑝𝑝(𝐴𝐴𝑠𝑠
𝑜𝑜 |𝐴𝐴𝑠𝑠 )𝑝𝑝(𝐴𝐴𝑠𝑠 )+𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐵𝐵)𝑝𝑝(𝐵𝐵𝑠𝑠 )+𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐶𝐶𝑠𝑠 )𝑝𝑝(𝐶𝐶𝑠𝑠 )
(14)
If the task is to estimate the size of A (or B), then Bayes’ estimate may be used [13][14][15].
8
However, the random nature of the elements of set C implies that the probabilities 𝑝𝑝(𝐴𝐴𝑜𝑜 |𝐶𝐶)and
𝑝𝑝(𝐵𝐵𝑜𝑜 |𝐶𝐶)are not fixed. Under certain conditions, these probabilities may be taken to be fixed
and constant. But in general, the nature of the nonlinear process underlying the change of the
characteristics as well as the environment will determine what these probabilities are.
The query could be to ascertain the probability of various subsets associated with the system as
a means of determining its structure. But as information about the subsets comes in, the
sampling strategy will have to be changed to reflect the estimates obtained by Bayes’ testing.
This is a process that will have to be updated continuously.
Conclusions
We have considered sets where some of the elements switch characteristics in some random
manner over time. This is a situation that cannot be covered by abstractions of either static or
dynamic sets when these changes take place at regular intervals. We propose the use of sets
with elements that change, and modular forms are proposed to account for one kind of such
change. An expression for the dependence of the mean on the probability of the switching
elements was determined. The system was also analyzed from the perspective of decision
between different hypotheses. Such sets are likely to be of use in complex system queries and
in analysis of surveys.
References
1. Pearl, J. (2014) Understanding Simpson’s paradox. The American Statistician 88: 8-13
2. Rücker, G. (1989) A two-stage trial design for testing treatment, self-selection and treatment
preference effects. Statistics in Medicine 8: 477-485
3. Kaptchuk, T.J. and Miller, F.G. (2015) Placebo effect in medicine. N Engl J Med 373:8-9
4. Kak, S. (1996) The three languages of the brain: quantum, reorganizational, and associative. In
Learning as Self-Organization, K. Pribram and J. King (editors). Lawrence Erlbaum Associates,
Mahwah, NJ, 185-219
5. Lambert et al. (2013) Quantum biology. Nature Physics, 9(1), 10-18
6. Kak, S. (2016). Communication languages and agents in biological systems. In
Biocommunication: Sign-Mediated Interactions between Cells and Organisms. J. Seckbach & R.
Gordon (editors). London, World Scientific Publishing. Pp. 203-226
7. Lagarias, J.C. (1985) The 3x + 1 problem and its generalization. Amer. Math. Monthly, 92(1): 3-23
8. Wirsching, G.J. (1998) The Dynamical System Generated by the 3n+1 Function. Lecture Notes in
Mathematics 1681. Springer-Verlag.
9. Kak, S. (2000) Active agents, intelligence, and quantum computing. Information Sciences 128: 117
10. Kak, S. (1999) The initialization problem in quantum computing. Foundations of Physics 29: 267279
11. Kak, S. (2007) Quantum information and entropy. Int. Journal of Theoretical Physics 46: 860-876
9
12. Aerts, D. (2009) Quantum structure in cognition. Journal of Mathematical Psychology 53: 314348
13. Kumar, P.R. and Varaiya, P. (2015) Stochastic Systems: Estimation, identification, and adaptive
control. SIAM.
14. Scott, D.W. (2015) Multivariate Density Estimation: theory, practice, and visualization. John
Wiley.
15. Tuzlukov, V.P. (2001) Signal Detection Theory. Springer Science.
10
| 2cs.AI
|
A Local Approach for Information Transfer
arXiv:1801.09527v1 [cs.IT] 9 Jan 2018
P. Garcı́aa∗and R. Mujicab
a
Facultad de Ingenierı́a en Ciencias Aplicadas,
Universidad Técnica del Norte, Ibarra, Ecuador.
b
Centro de Fı́sica Teórica y Computacional,
Laboratorio de Fenómenos No Lineales,
Escuela de Fı́sica, Facultad de Ciencias,
Universidad Central de Venezuela,
Caracas, Venezuela.
Abstract
In this work, a strategy to estimate the information transfer between the elements
of a complex system, from the time series associated to the evolution of this elements,
is presented. By using the nearest neighbors of each state, the local approaches of the
deterministic dynamical rule generating the data and the probability density functions,
both marginals and conditionals, necessaries to calculate some measures of information
transfer, are estimated.
The performance of the method using numerically simulated data and real signals is
exposed.
1
Introduction
The estimation of directionality of coupling between electrocardiographic and respiratory signals[1] may help to decide on which system exerts control allowing a more
efficient medication; determination of the causal relations between shares of stock, to
select which one to monitor, in order to make decisions to buy or sell; the detection of
the elements that control the evolution of the others in a complex network, allows to
estimate its robustness and simplifies the control problem by reducing the number of
sites on which to take control actions[2]; these are some of the situations where it is
∗
Permanent address: Laboratorio de Sistemas Complejos, Departamento de Fı́sica Aplicada, Facultad de
Ingenierı́a, Universidad Central de Venezuela, Caracas, Venezuela.
useful to estimate how a subsystem sends and receives information from another in a
complex system.
From a more general perspective, the estimation of the extent to which a component
contributes to the production of information system and the rate at which it is shared
with the rest of the components, can provide important information about its structure.
For this reason, detecting the transfer of information and its directionality is a subject
of great interest due to the variety of practical applications, in areas ranging from
physics to marketing.
Since Wiener[3] proposed that an improvement of the prediction of the future of a
time series x, by the incorporation of information from the past of a second time series
y, can be seen as an indicator of a causal interaction from y to x, an operational way
to implement these ideas was proposed by Granger [4] and a further formalization of
the strategy by Schreiber [5]; the problem of the estimation of a cause-effect relation
between elements of a complex system, have been addressed under many different
points of views. From deterministic [6, 7, 8, 35, 10] to probabilistic [5, 11, 12] modeling
schemes have been proposed and the problem has been named as detection of causality
[7, 8] or directionality of the coupling estimation [6, 10, 11, 12] or detection of direct
links [13], among others.
In this work we propose to estimate the transmission of information between elements of complex systems, using the formal probabilistic scheme proposed by Schreiber
in [5], but approximating the probability density functions (PDF) involved, by using
a k-nearest neighbors approach. In this case, the neighborhood, to any point, is used
in a deterministic nonlinear local model to estimate the conditional PDF’s and in a
density estimator to approximate the marginals PDF’s associated with the data. This
strategy, although it is also based on the determination of the k nearest neighbors, is
conceptually simpler than the strategy presented in [14] and has the advantage that
the estimate can be made using only the nearest neighbor, which in principle makes it
computationally less expensive.
The use of deterministic models for the further determination of the information
transfer, is expected to offer an interesting relationship between the strategies based
on probabilistic and deterministic models, as well as between the parametric and nonparametric schemes.
In order to show the usefulness of the scheme, the work is organized as follows:
section 1 is devoted to present the idea behind of the information transfer and section 2
gives a methodology to estimate the conditional PDF’s involved in information transfer,
using local approaches. In section 3 we combine these ideas, into the approach of
Schreiber [5], to estimate the information transfer in the case of numerically simulated
and real signals. Finally in section 4, we give some concluding remarks.
2
Information transfer
There are many tools to estimate the information transfer in time series, see for example
[15] and references therein. Nevertheless an attempt to classify them in few categories
require a titanic effort, among other reasons, because of the unclear relations between
them. However, most popular time series tools appear to fall into one of two broad
classes, model-based strategies and parameter-free strategies or model-free strategies.
The representative of the first class is an approach called Granger Causality [4]
and the other class is represented by a scheme based on information theory, entropy
transfer, proposed by Schreiber in [5].
In Granger’s method, causality could be tested by measuring the ability to predict,
using a linear model, the future values of a time series using prior values of another
time series.
Using this scheme in the case of nonlinear subsystems can leads to erroneous conclusions. This problem is addressed in [16] where a nonlinear model is performed using
radial basis functions. Despite this, the scheme still has the weakness of being a static
measure.
In the case of parameter-free methods, the usual strategy to quantify the superimposing of information contained in interacting subsystems is the mutual information.
Unfortunately this is a symmetrical and static measure, so that does not give a sense
of direction or temporal evolution of information in the system. A strategy which
overcomes the two previously mentioned weaknesses is proposed in [5] to estimate the
transmission of information based on observations of the time series associated with
the system elements and information theory [17].
In [5], Schreiber proposes a measure of transfer of entropy which represents a rigorous derivation of Wiener causal measure and shares some of the desired properties,
to represent transfer of information, with Mutual Information but takes into account
the dynamics of the system.
This measure, with minimal assumptions about the dynamics of the system and
the nature of the coupling, is able to quantify the exchange of information between two
systems x and y as follows:
Ix→y
Iy→x
p(yn+1 |yn , xn )
=
p(yn+1 , yn , xn ) log2
p(yn+1 |yn )
yn+1 ,xn ,yn
X
p(xn+1 |xn , yn )
=
p(xn+1 , xn , yn ) log2
p(xn+1 |xn )
x
,x ,y
X
n+1
n
(1)
n
and it may be thought [18] as a conditional mutual information.
Where p(·, ·) and p(·|·) are the joint and conditional probabilities of occurrence of
the state or system. Determining the flow of information given by (1) requires the
calculation of probability densities associated with transitions between states of each
subsystem, and in this case of a brute-force estimation, requires a coarse graining of
the time series associated to both subsystems. This represents a large computational
cost that increases considerably when these probabilities should be calculated for all
pairs of elements of a large extended system.
3
Probability densities and local modeling
According to Wiener’s idea, if the improvement in prediction can be associated with a
reduction in uncertainty, it is natural that a measure of causality can be represented
in terms of information theory concepts. However, the determination of causal relationships or information transfer, may also be represented in terms of deterministic
models.
To show the feasibility of the idea, we start by representing the joint probabilities,
in the entropy transfer definition (1), in terms of conditional and marginal probability
densities using: p(x, y, z) = p(x|y, z)p(y|z)p(z), for later, we propose a strategy to estimate both, the conditional densities probabilities and marginal densities probabilities,
using local nonlinear modeling of the dynamical system [19, 20, 21] and nonparametric
method to estimate marginals densities functions [22] respectively.
Inspired in [23], we propose to estimate the conditional probabilities densities
p(xn+1 |xn ), p(xn+1 |xn , yn ), p(xn+1 |yn ) and p(yn+1 |yn ), p(yn+1 |yn , xn ), p(yn+1 |xn )
associated to the appearance of states xn+1 and yn+1 , given some of the predecessors
states of both subsystems, using the quality of predictions made from a deterministic
model. Specifically, we base the approach in the estimation of the nearest neighbors
of xn and yn . The marginal probability density functions, p(xn ) and p(yn ), can be
approached using the same strategy.
In order to do this, we start by constructing a Cumulative Distribution Function
to derive from it the necessary probability densities. This function is, by definition, a
real-valued and strictly increasing function of a random variable X, usually represented
as:
FX (x) = P (X ≤ x),
(2)
where the right-hand side is the probability that, the random variable X takes on a
value less than or equal to x. The probability that x lies in the semi-closed interval
(a, b], where a < b, is therefore
P (a < X ≤ b) = FX (b) − FX (a).
(3)
The Cumulative Density Function of a continuous random variable X can be expressed as the integral of its probability density function pX (x) as follows:
Z x
FX (x) =
pX (t) dt.
(4)
−∞
Similarly, for the case of Cumulative Conditional Distribution Function:
Z y
FXY (y|x) =
pXY (y|x) dy.
(5)
−∞
Thus, the conditional probability density (CPD) function can be obtained as:
pXY (y|x) =
∂F (y|x)
∂y
(6)
If we chose the Cumulative Conditional Distribution Function in (6) as a sigmoid
function, S(u) = 1/1 + e−ru , with u = f (x) − y, as in [23]:
F (y|x) = S(f (x) − y)
(7)
where f is a model for the system, constructed from the data; then we can distinguish
between two extremes for distribution: one strictly deterministic, i.e. without errors,
where r → ∞ and F is given by:
F (y|x) = θ(f (x) − y),
(8)
with θ the Heaviside function, or a probabilistic approach, where the parameter r < ∞.
Here r is, in some sense, a measure of the randomness of the process X.
It is clear, from equations (6) and (8), that if we have an approximation of the
dynamic rule that governs the evolution of the system, then it is possible to construct
an adequate cumulative distribution function, that allows to estimate the probability
densities, necessary to calculate the transfer of information. In this case we propose
to chose the parameter r, in the sigmoid function, proportional to 1/σ, where σ is
the standard deviation of the modeling errors, since the slope of the transition of the
sigmoid function determines the width of the distribution associated with it.
At this point, we can summarize the presentation into two main ideas: the Schreiber’s
information transfer, and some other estimators, can be written in terms of CPD’s and
the CPD’s can be estimated using deterministic models of the rule generating the data.
The remainder of the section, is intended to implement these models and show how they
can be incorporated into an estimator of information transfer, with a computational
cost similar to the determination of the k-nearest neighbors, with k = 1.
Let us suppose that we have a series of values of the states of a dynamical system
d
{xi }N
i=1 , with xi ∈ R , obtained either by measuring the d components of system’s
state at regular time intervals or by reconstructing the state space by using the Takens
theorem[24], from partial information about the states.
In our approach the dynamical rule is estimated using a local approximation for f .
This approach can be presented algorithmically as:
i. Given the i-th data value xi , with i = 1, 2, · · · , N .
ii. Determine the set of the m nearest states ({xvi }m
j=1 ), to the state i in the Euj
clidean metric, sorted in ascendant order with the distance from xi . Here vji ,
the label of the j-th neighbor of the i-th data value, is an integer in the interval
[1, N − 1].
iii. The approximation is obtained by Taylor expanding f to first order, around of
the nearest neighbor xvi , to obtain:
1
xi+1 ≈ f (xvi ) + Df (xvi )(xi − xvi )
1
1
1
(9)
where, Df (xvi ) is the Jacobian matrix of f .
1
iv. Finally, in zero order, the evolution of the system is approximated by the evolution
of the nearest neighbor, i.e., xi+1 ≈ xvi +1 . For a first order approximation it is
1
necessary to calculate Df (xvi ).
1
The marginal probabilities densities functions p(x) y p(y), can be estimated using
the k-th nearest neighbor density estimator[22]:
p(xi ) =
k
.
N ||x − xvi ||
(10)
k
To illustrate graphically the methodology to estimate the probability densities,
Figure 1 shows conditional and marginal probability density functions, for N = 500
data points of the skew tent map with a = 0.65,
xi
0≤x≤a
a
fa (xi ) =
,
(11)
1−xi
a<x≤1
1−a
Thus, for zero order approach of the dynamical rule f in (11) and using the equations
(6) and (8), the CPD function is given by:
−r(xvi +1 −x1+1 )
p(xi+1 |xi ) = h
re
1+e
1
i
−r(xvi +1 −xi+1 ) 2
(12)
1
Figure 1, illustrates the above mentioned ideas and gives a graphical insight of how
the proposed strategy associates the modeling errors with CPD’s.
4
Results
In order to show the performance of our strategy, we estimate the information transfer
in the case of signals from a numerical simulation and real data.
0.10
p ( xn )
0.08
0.06
0.04
0.02
0.00
0.0
0.2
0.4
0.6
0.8
1.0
xn
Figure 1: Left: CPD function. Doted line shows, as an example, the probability of the
occurrence of all values given the occurrence of the value 0.2. Right: Marginal probability
density function. The solid line represents the local estimation and the doted one represents
a histogram of the data. Both PDFs are estimated using 500 data points of the skew tent
map with parameter a = 0.65 and k = 1.
4.1
Numerical simulations
The data associated with the numerical simulations correspond to two coupled skew
tent maps in chaotic regime, connected by a nonlinear and bidirectional coupling function as in [25] and the Chuas’s system[26] the first case provides a controlled experiment
where the directionality of the transfer of information is known, which allows a performance testing strategy; the second example will allow to compare the performance
of the strategy using numerically simulated data with the results for the same system
but in the case of real data.
4.1.1
Coupled chaotic maps
Here, two chaotic tent maps (11) with parameter a = 0.5, are coupled according to:
xn+1 = f (xn + (yn − xn ))
yn+1 = f (yn + µ(xn − yn ))
(13)
where the parameters and µ define the intensity of the coupling. The transmission of
information, in this case, can be associated to the synchronization of the trajectories
of the systems [27]. Being understood by synchronization, the coincidence between the
states of the subsystems once sufficient time has elapsed. Figure 2 shows synchronization error and information transfer between the coupled maps (13) as a function of the
coupling parameters. It is worth to note as the flat zone, in the information transfer
surface, coincide with the synchronization zone, as suggested by Bollt in [27].
|x N - y N |
0.4
0.35
0.3
μ
0.25
0.2
0.15
0.1
0.05
0.0
0.0
0.1
0.2
0.3
0.4
ϵ
Figure 2: Left: Synchronization error |xN − yN | as a function of the coupling parameters.
Right: Information Transfer between the elements of the systems (13), estimated using 500
data points.
4.1.2
Chua’s system
The Chua’s system is a well-known chaotic system, representing a RLC circuit with
one or more nonlinear elements. The easy construction of the circuit has made it an
ubiquitous real-world example of a chaotic system.
dVC1
dt
dVC2
RC2
dt
diL
dt
= α[VC2 − VC1 − g(VC1 )]
= VC1 − VC2 + RiL
= −βVC2
(14)
where g(VC1 ), is the usual piece-wise linear function representing the Chua’s diode and
all parameters are chosen as in [28].
Let us rename the elements of the circuits indexing them as C1 → 1, C2 → 2 y
L → 3 and define the net flow of information associated to each component of the
circuit, as the difference among the sum of the information transfer between it and
the rest of the elements and the sum of the information transfer from the rest of the
elements to it, as:
VC 1
VC 2
iL
Ii→j
1.0
VC 1
0.8
0.6
VC 2
0.4
0.2
iL
0
Figure 3: Information transfer between all pairs formed with signals VC1 , VC2 and iL . The
data is obtained by numerical integration of the system (14) using a Runge-Kutta of 4-th
order. Here we chose the parameter r as two times the standard deviation of the modeling
errors at dimension three.
Ti =
3
X
Ii→j − Ij→i .
(15)
j=1
The evaluation of these quantities, see Figure 3, produces: T1 = −0.5684, T2 =
1.6476, T3 = −1.0792. This is consistent with some results about the control of Chua’s
circuit [28, 29, 30]. Here the authors prove that Chua’s system can be controlled by
using a linear feedback controller and say that their numerical simulations show that
simplest control can be achieved by perturbing only VC2 .
4.2
4.2.1
Real data
Chua’s circuit
In this case, an experimental implementation of the Chua’s circuit was done, using
operational amplifiers to simulate the Chua’s diode and the inductor L[31, 32]. The
current in the inductor was obtained by an indirect measure and the data, sampled
with rate of 5ms, is shown in the left side of Figure 4.
The right side of Figure 4 shows the information transfer estimated as explained
before for all pairs formed with signals VC1 , VC2 and iL .
In this case, T1 = −0.268914, T2 = 1.39525 and T3 = −1.09931, which is consistent
with the results obtained previously for the numerical simulation of this system.
VC 1
VC 2
iL
Ii→j
1.0
0.6
VC 1
0.8
0.2
0.6
VC
2
0.4
VC 2
0.0
0.4
- 0.2
0.2
- 0.4
iL
- 1.0
- 0.5
0.0
VC
0.5
1.0
0
1
Figure 4: Left: VC1 vs VC2 , for 1024 experimental data points from Chua’s circuit. Right:
Information transfer between all pairs formed with signals VC1 , VC2 and iL . Here we chose
the parameter r as two times the standard deviation of the modeling errors at dimension
three and the number of neighbors k = 1, but the numerical experiments suggests that the
results are not strongly dependent of this choice.
4.2.2
Physiological signals
This experimental data correspond to time series taken from the Massachusetts General
Hospital/Marquette Foundation (MGH/MF) Waveform Database in https://www.physionet.org/pn3/mghdb/
[34]. From there, the series mgh003, mgh005, mgh006, mgh087, mgh097, mgh148,
mgh160, mgh183, mgh190 and mgh202, corresponding to electrocardiographic and
blood pressure signals measured simultaneously, in patients in intensive care, were selected. To estimate the information transfer between both systems, 10000 data points
from each pair of series were used. In all cases a greater transfer of information
from arterial pressure or baro-reflex (br) system to the electro-cardiac (ec) system,
T = (Ibr→ec − Iec→br )/(Ibr→ec + Iec→br ), as obtained in [35].
5
Final remarks
Three aspects should be highlighted, regarding the performance of the proposed strategy, first, but not the most important, is referred to the low computational cost associated with the determination of the first nearest neighbors compared, at least, to the
cost to estimate CPD’s using coarse graining or calculating correlation integrals as in
[5, 33]. Second, we must note that in the face of the results using experimental data,
the estimation of the transfer of information based on nearest neighbors appears to
be robust to the presence of moderate noise, given the consistency with the results in
references [28, 29, 30]. Although a more detailed study of these aspects it is needed to
1.0
●
●
●
0.8
●
●
●
0.6
T
●
●
0.4
●
●
0.2
mgh202
mgh190
mgh183
mgh160
mgh148
mgh097
mgh087
mgh006
mgh005
0
mgh003
0.0
Data set
Figure 5: Information transfer between pairs formed with electrocardiographic and blood
pressure signals measured simultaneously. In this case T represent the information transfer
from the baro-reflex system to the cardiac system.
be conclusive, it is out of the scope of the present work and is subject of our current
research.
Finally, we emphasize that, although one of the advantages of Schreiber’s estimator
is that it is independent of models, and with this proposal, it will become a dependentmodel scheme, nevertheless we believe that the establishment of a relationship between
an useful probabilistic definition of information transfer and deterministic modeling,
may be a practical idea in the development of methodologies to determine causal
relationship between subsystems of complex systems or contribute in the design of
control schemes for extended systems.
References
[1] S. Schultz, F-C. Adochier, I-R. Edu, R. Schroeder, H. Costin, K-J. Bar and A.
Voss, Philos Trans A Math Phys Eng Sci, 371 (2013).
[2] Y-Y. Liu, J-J. Slotine, and A-L. Barabasi, Nature 473, 167 (2011).
[3] N. Wiener, The theory of prediction. In Modern Mathematics for Engineer,
McGraw-Hill, 1956.
[4] C. Granger, Econometrica 37, 424 (1969).
[5] T. Schreiber, Phys. Rev. Lett. 85, 461 (2000).
[6] B. Bezruchko, V. Ponomarenko, M. G. Rosenblum and A. Pikovsky, CHAOS 13,
179 (2003).
[7] H. Ma, K. Aihara and L. Chen, Scientific Reports 4, 1 (2014).
[8] K. Ishiguro, N. Otsu, M. Lungarella and Y. Kuniyoshi, Phys. Rev. E. 77, 026216
(2008).
[9] L. Cisneros, J. Jiménez, M. G. Cosenza and A. Parravano, Phys. Rev. E. 65,
045204R (2002).
[10] M. G. Rosenblum and A. S. Pikovsky, Phys. Rev. E. 64, 045202R (2001).
[11] M. Palus and A. Stefanovska, Phys. Rev. E. 67, 055201R (2003).
[12] M. Vejmelka and M. Palus, Phys. Rev. E. 77, 026214 (2008).
[13] N. Rubido, A. Martı́, E. Bianco-Martı́nez, C. Grebogi, M. S. Baptista and C.
Masoller, New Journal of Physics 16, 093010 (2014).
[14] J. Zhu, J-J. Bellanger, H. Shu and R. Le Bouquin Jeannés, Entropy 17, 4173
(2015).
[15] R. Vicente and M. Wibral, In: Efficient estimation of Information Transfer,
Springer-Verlag Berlin, 2014.
[16] N. Ancona, D. Marinazzo and S. Stramaglia, Phys. Rev. E. 70, 056221 (2004).
[17] C. Shannon, The Bell System Technical Journal 27, 379 (1948).
[18] M. Palus, Phys. Rev. E. 67, 046211 (2003).
[19] J. D. Farmer and J. J. Sidorowich, Phys. Rev. Lett. 58, 845 (1987).
[20] P. Garcı́a, J. Jiménez, A. Marcano and F. Moleiro, Phys. Rev. Lett. 76, 1449
(1996).
[21] P. Garcı́a and A. Merlitti. The European Physical Journal Special Topics 143,
261 (2007).
[22] D. Loftgaarden and C. Quesenberry, Annals of Mathematical Statistics 36, 1049
(1965).
[23] D. Husmeier, In: Perspectives in Neural Computing, Springer-Verlag, 1999.
[24] F. Takens, In: Dynamical Systems and Turbulence, Springer Berlin, 1980.
[25] M. Hasler and Y-L. Maistrenko. IEEE Trans. Circuits Syst. 44, 856 (1997).
[26] T. Matsumoto, IEEE Transactions on Circuits and Systems 31, 1055 (1984).
[27] E. M. Bollt, International Journal of Bifurcation and Chaos 22, 1250261 (2012).
[28] G. Chen and X. Dong, Journal of Circuits Systems and Computers 3, 139 (1993).
[29] G. Chen, Chaos, Solitons & Fractals 8, 1461 (1997).
[30] C. Chen and X. Dong, Proc. of IEEE Inr. Symp. on Circ. Syst., Chicago, IL.
2604–2607 (1983).
[31] K. Gopakumar, B. Premlet and K. G. Gopchandran, International Journal of
Electronics 98, 667 (2011).
[32] The implementation of this circuit is already an undergraduate exercise, in this
case it was done using an Educational Laboratory and Virtual Instrumentation
Suite (NI-ELVIS) from National Instruments.
[33] W. Liebert, Phys. Lett. A. 142, 107 (1989).
[34] J. P. Welch, P. J. Ford, R. S. Teplick and R. M. Rubsamen, The MGH/MF
Waveform Database, 1992.
[35] L. Cisneros and J. Jiménez, Revista Mexicana de Fsica 49, 17 (2003).
| 7cs.IT
|
arXiv:cs/0312031v1 [cs.DC] 16 Dec 2003
Theory and Practice of Logic Programming 1(3) 251–282, May 2001.
1
Distributed WWW Programming using
(Ciao-)Prolog and the PiLLoW library∗
DANIEL CABEZA and MANUEL HERMENEGILDO
CLIP Group
http://www.clip.dia.fi.upm.es – http://www.cliplab.org
Facultad de Informática
Universidad Politécnica de Madrid (UPM)
28660-Boadilla del Monte, Madrid, Spain
(e-mail: {dcabeza,herme}@fi.upm.es)
Abstract
We discuss from a practical point of view a number of issues involved in writing distributed
Internet and WWW applications using LP/CLP systems. We describe PiLLoW, a publicdomain Internet and WWW programming library for LP/CLP systems that we have
designed in order to simplify the process of writing such applications. PiLLoW provides
facilities for accessing documents and code on the WWW; parsing, manipulating and
generating HTML and XML structured documents and data; producing HTML forms;
writing form handlers and CGI-scripts; and processing HTML/XML templates. An important contribution of PiLLoW is to model HTML/XML code (and, thus, the content
of WWW pages) as terms. The PiLLoW library has been developed in the context of the
Ciao Prolog system, but it has been adapted to a number of popular LP/CLP systems,
supporting most of its functionality. We also describe the use of concurrency and a highlevel model of client-server interaction, Ciao Prolog’s active modules, in the context of
WWW programming. We propose a solution for client-side downloading and execution of
Prolog code, using generic browsers. Finally, we also provide an overview of related work
on the topic.
Keywords: WWW, HTML, XML, CGI, HTTP, Distributed Execution, (Constraint)
Logic Programming.
1 Introduction
The wide diffusion of the Internet and the popularity of the “World Wide
Web” (Berners-Lee et al., 1994) –WWW– protocols are effectively providing a
novel platform that facilitates the development of new classes of portable and
user-friendly distributed applications. Good support for network connectivity and
the protocols and communication architectures of this novel platform are obviously requirements for any programming tool to be useful in this arena. However, this alone may not be enough. It seems natural that significant parts
0
This paper is an expanded and improved version of (Cabeza & Hermenegildo, 1996a;
Cabeza & Hermenegildo, 1996b; Cabeza & Hermenegildo, 1997; Cabeza et al., 1996).
2
D. Cabeza and M. Hermenegildo
of network applications will require symbolic and numeric capabilities which
are not necessarily related with distribution. Important such capabilities are,
for example, high-level symbolic information processing, dealing with combinatorial problems, and natural language processing in general. Logic Programming (LP) (Kowalski, 1974; Colmerauer, 1975) and Constraint Logic Programming
(CLP) systems (Jaffar & Lassez, 1987; Van Hentenryck, 1989; Colmerauer, 1990;
M. Dincbas & Van Hentenryck, 1990; ECRC, 1993) have been shown particularly
successful at tackling these issues (see, for example, the proceedings of recent conferences on the “Practical Applications of Prolog” and “Practical Applications of
Constraint Technology”). It seems natural to study how LP/CLP technology fares
in developing applications which have to operate over the Internet.
In fact, Prolog, its concurrent and constraint based extensions, and logic programming languages in general have many characteristics which appear to set
them particularly well placed for making an impact on the development of practical networked applications, ranging from the simple to the quite sophisticated.
Notably, LP/CLP systems share many characteristics with other recently proposed
network programming tools, such as Java, including dynamic memory management,
well-behaved structure and pointer manipulation, robustness, and compilation to
architecture-independent bytecode. Furthermore, and unlike the scripting or application languages currently being proposed (e.g., shell scripts, Perl, Java, etc.),
LP/CLP systems offer a quite unique set of additional features including dynamic
databases, search facilities, grammars, sophisticated meta-programming, and well
understood semantics.
In addition, most LP/CLP systems also already offer some kind of low level support for remote communication using Internet protocols. This generally involves
providing a sockets (ports) interface whereby it is possible to make remote data connections via the Internet’s native protocol, TCP/IP. A few systems support higherlevel communication layers on top of this interface including linda-style blackboards (e.g., SICStus Prolog (Carlsson, 1988), Ciao (Carro & Hermenegildo, 1999;
Cabeza & Hermenegildo, 1995;
Hermenegildo & CLIP Group, 1994;
Hermenegildo et al., 1995a;
Hermenegildo et al., 1999a;
Hermenegildo et al., 1999b;
Bueno et al., 1997),
BinProlog/µ2Prolog (Tarau, 1996; De Bosschere, 1989), etc.) or shared variable-based communication (e.g., KL1 (Chikayama et al., 1994), AKL (Janson & Haridi, 1991), Oz
(Smolka, 1994), Ciao (Hermenegildo et al., 1995b; Cabeza & Hermenegildo, 1995),
etc.). In some cases, this functionality is provided via libraries, building on top of
the basic TCP/IP primitives. This is the case, for example, of the SICStus and Ciao
distributed linda-style interfaces. In fact, as we have shown in previous work, sharedvariable based communication can also be implemented in conventional systems
via library predicates, by using attributed variables (Hermenegildo et al., 1995b;
Cabeza & Hermenegildo, 1995). In addition to these communication primitives,
several systems offer concurrency and even higher-level abstractions (distributed
objects, mobile code, ...) which are very useful for developing general-purpose distributed applications.
Our concrete interest here is WWW applications. These applications generally
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
3
use specific high-level protocols (such as HTTP or FTP), data formats (such as HTML
or XML), and application architectures (e.g., the CGI interface) which are different from, e.g., the shared-variable or linda-based protocols typically used in other
types of distributed applications. In this paper we study how good support for
these WWW-related protocols, data formats, and architectures can be provided for
LP/CLP systems, building on the widely available interfaces to the basic TCP/IP
protocols. Our aim is to discuss from a practical point of view a number of the new
issues involved in writing WWW applications using LP/CLP systems, as well as
the architecture of some typical solutions. In the process, we will describe PiLLoW
(“Programming in Logic Languages on the Web”), a public domain Internet/WWW
programming library for LP/CLP systems which, we argue, significantly simplifies
the process of writing such applications. PiLLoW provides facilities for generating
HTML/XML structured documents by handling them as Herbrand terms, producing HTML forms, writing form handlers, processing HTML/XML templates,
accessing and parsing WWW documents, etc. We also describe the architecture of
some relatively sophisticated application classes, using a high-level model of clientserver interaction, active modules (Cabeza & Hermenegildo, 1995). Finally we describe an architecture for automatic LP/CLP code downloading for local execution,
using just the library and generic browsers.
Apart from the tutorial value of the paper, we present a number of technical
contributions which include the idea of representing HTML and XML code (and
structured text in general) as Prolog terms, the use of the logical variable in such
terms leading to a model of an “HTML template” (a pair comprising a term with
free variables and a dictionary associating names to those variables), the notion
of “active logic modules” and its application to solving efficiency issues in CGI
interaction in a very simple way, the idea of “Prolog scripts” and its application
to CGIs, and the identification of a number of features that should be added to
existing systems in order to facilitate the programming of WWW applications –
mainly concurrency.
The argument throughout the paper is that, with only very small limitations in
functionality (which disappear when concurrency is added, as in systems such as
BinProlog/µ2-Prolog, AKL, Oz, KL1, and Ciao Prolog), it is possible to add an
extremely useful Internet/WWW programming layer to any LP/CLP system without making any significant changes in the implementation. We argue that this layer
can simplify the generation of applications in LP/CLP systems including active
WWW pages, search tools, content analyzers, indexers, software demonstrators,
collaborative work systems, MUDs and MOOs, code distributors, etc.
The purpose of the paper is also to serve as a tutorial, containing sufficient
information for developing relatively complex WWW applications in Prolog and
other LP and CLP languages using the PiLLoW library. The PiLLoW library has
been developed in the context of the Ciao Prolog system, but it has been adapted to
a number of other popular LP/CLP systems, supporting most of its functionality.
The Ciao Prolog system and the PiLLoW library can be freely downloaded from
http://www.clip.dia.fi.upm.es and http://www.cliplab.org.
4
D. Cabeza and M. Hermenegildo
www.xxx.yyy
HTTP
Server
1
WWW
Browser
std. doc. request
4
output
2
3
hello_world
Fig. 1. The CGI interface
2 Writing basic cgi-bin applications
The simplest way of writing WWW applications is through the use of the “Common
Gateway Interface” (CGI). A CGI executable is a standard executable file but such
that the HTTP server (the program that responds to HTTP requests in a machine
which serves a WWW site) can tell that it in fact contains a program that is to be
run, rather than a document text that is to be sent to the client (the browser) as
usual. The file can be distinguished by belonging to a special directory, commonly
named cgi-bin, or by a special filename ending, such as .cgi. This is normally set
during configuration of the HTTP server. The basic idea behind the CGI interface
is illustrated in Figure 1. When the user selects an address of a CGI executable
in a document, such as http://www.xxx.yyy/cgi-bin/hello_world (or perhaps
http://www.xxx.yyy/foo/hello_world.cgi) the browser issues a standard document request (1). The HTTP server, recognizing that it is a CGI executable rather
than a document, starts the executable (2), and during such execution stores the
output of the executable in a buffer (3). Upon termination of the executable, the
contents of the buffer (which should be in a format that the browser can handle,
such as HTML) are returned to the browser as if a normal page with that content
had been accessed (4).
The following is an example of how a very simple such executable can be written
in an LP/CLP language. The source might be as follows:1
main :write(’Content-type: text/html’), nl, nl,
write(’<HTML>’),
write(’Hello world.’),
write(’</HTML>’).
And the actual executable could be generated as usual, for example in the Ciao
1
Note that in the examples presented, and in order to shorten them, the HTML code may
be slightly simplified, and as a result of this it may not be completely standard-conforming.
However, the examples can be used as is with all popular browsers.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
5
system, using the standalone compiler, by writing at a UNIX shell “ciaoc -o
hello_world.cgi hello_world”.2 The executable then has to be placed in an
appropriate place (accessible via an HTTP address by a browser) and have the
right permissions for being executed by the server (for example, in some systems
this means being executable by the user “nobody”).
In systems which make executables through saved states (which usually have the
disadvantage of their generally large size), at the system prompt one could create
the executable by writing something like:
?:- compile(’hello_world.pl’),
save(’hello_world.cgi’), main.
3 LP/CLP Scripts for CGI Applications
CGI executables are often small- to medium-sized programs that perform relatively
simple tasks. This, added to the slow speed of the network connection in comparison with that of executing a program (which makes program execution speed less
important) has made scripting languages (such as shell scripts or Perl) very popular
for writing these programs. The popularity is due to the fact that no compilation
is necessary (extensive string handling capabilities also play an important role in
the case of Perl), and thus changes and updates to the program imply only editing
the source file.
Logic languages are, a priori, excellent candidates to be used as scripting languages.3 However, the relative complication in making executables (needing in some
systems to start the top-level, compile or consult the file, and make a saved state)
and the often large size of the resulting executables may deter CGI application
programmers. It appears convenient to provide a means for LP/CLP programs to
be executable as scripts, even if with reduced performance.
It is generally relatively easy to support scripts with the same functionality in
most LP/CLP systems. In Ciao, the program ciao-shell –which has also been
adapted to SICStus (Hermenegildo, 1996)– accomplishes this task, by first loading
the file given to it as the first argument (but skipping the first lines and avoiding
loading messages) and then starting execution at main/1 (the argument provides
the list of command line options). Then, for example, in a Unix system, the following
program can be run directly as a script without any need for compilation:
#!/usr/local/bin/ciao-shell
main(_) :write(’Content-type: text/html’), nl, nl,
write(’<HTML>’),
2
3
It often convenient to use options (such as ciaoc’s -s or -S) which will generate a standalone
executable which is independent of any libraries.
For example, the built-in grammars and databases greatly simplify many typical script-based
applications.
6
D. Cabeza and M. Hermenegildo
www.xxx.yyy
<form action="http://www.xxx.yyy/handler.cgi">
2
WWW
Browser
1
HTTP
Server
form data
6
3
form
reply
4
5
handler.cgi
Fig. 2. The Forms interface
write(’Hello world.’),
write(’</HTML>’).
Note that in some UNIX versions either the program ciao-shell must be included in the /etc/shells listing or the first line should be replaced by these
two:
#!/bin/sh
exec ciao-shell $0 "$@"
The execution of Prolog scripts may be optimized in some systems. For example,
in Ciao, the first time a script is run it is also compiled and its bytecode is saved
to a file. At subsequent times, if the script has not changed, the object code is
retrieved from that file, avoiding compilation or interpretation overhead.
4 Form Handling in HTTP
So far we have shown CGI executables which produce output, but this output is not
a function of input coming from the request, which is obviously of limited interest.
CGI executables become most useful when combined with HTML forms. HTML
forms are HTML documents (or parts of HTML documents) which include special
fields such as text areas, menus, radio buttons, etc. which allow providing input to
CGI executables. The steps involved in the handling of the input contained in a
form are illustrated in Figure 2. When a document containing a form is accessed
via a form-capable browser (Mosaic, Netscape, Lynx, etc.), the browser displays the
input fields, buttons, menus, etc. indicated in the document, and locally allows the
user to perform input by modifying such fields. However, this input is not ultimately
handled by the browser. Instead, it will be sent to a “handler” CGI program, which
can be anywhere on the net, and whose address must be given in the form itself
(1). Forms generally have a “submit” button such that, when pressed, the input
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
7
provided through the menus, text areas, etc. is sent by the browser to the HTTP
server corresponding to the handler (2). Two methods for sending this input exist:
“GET” and “POST”. In the meantime, the sending browser waits for a response
from that program, which should come in the form of a new HTML document.
The handler program is invoked in much the same way as a cgi-bin application (3),
except that the information from the form is supplied to the handler (in different
ways depending on the system, the method of invocation, and the content type)
(4). This information is encoded in a predefined format, which relates each piece of
information to the corresponding field in the form, by means of a keyword associated
with each field. The handler then identifies the information corresponding to each
field in the original form, processes it, and then responds by writing an HTML
document to its standard output (5), which is forwarded by the server to the waiting
browser when the handler terminates (6). An important point to be noted is that,
as with simple cgi-bin applications, the handler is started and should terminate for
each transaction. The reader is referred, for example, to (Grobe & Naseer, 1998)
for a more complete introduction to CGI scripts and HTML forms.
5 Writing Form Handlers with Pi LLoW
The only complication in writing form handlers compared to writing simple CGI
applications is the need to capture and parse the form data. As we said before, this
data can be provided in several ways, depending on the system and the method
used to invoke the form, and is encoded with escape sequences. It is relatively easy
to write a Prolog program to parse such input (using, for example, definite clause
grammars –DCGs). The PiLLoW library provides some predicates which do this
and simplify the whole task, hiding the low-level protocol behind. The principal
predicates provided include:
get_form_input(Dic ) Translates input from the form (with either the POST or
GET methods, and even with CONTENT_TYPE multipart/form-data) to a dictionary Dic of attribute=value pairs. It translates empty values (which indicate
only the presence of an attribute) to the atom ’$empty’, values with more than
one line (from text areas or files) to a list of lines as strings, the rest to atoms or
numbers (using name/2). This is implemented using DCG parsers.
get_form_value(Dic,Var,Val ) Gets value Val for attribute Var in dictionary
Dic. Does not fail: value is ’’ if not found (this simplifies merging form producers
and form handlers, see later).
form_empty_value(V ) Useful to check that a value V from a text area is empty
(filters spaces, newlines, linefeeds, etc.).
form_default(Val,Default,NewVal ) Useful when a form is only partially filled
(and also in the first invocation of a combined form handler/producer – see
Section 7). If the value of Val is empty then NewVal=Default, else NewVal=Val.
my_url(URL ) Returns in URL the Uniform Resource Locator (WWW address) of
this cgi executable.
form_request_method(Method ) Returns in Method the method of invocation of
the form handler (“GET” or “POST”).
8
D. Cabeza and M. Hermenegildo
For example, suppose we want to make a handler which implements a database
of telephone numbers and is queried by a form including a single entry field with
name person_name. The handler might be coded as follows:
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
write(’Content-type: text/html’), nl, nl,
write(’<HTML><TITLE>Telephone database</TITLE>’), nl,
write(’<IMG SRC="phone.gif">’),
write(’<H2>Telephone database</H2><HR>’),
write_info(Name),
write(’</HTML>’).
write_info(Name) :form_empty_value(Name) ->
write(’You have to provide a name.’)
; phone(Name, Phone) ->
write(’Telephone number of <B>’),
write(Name),
write(’</B>: ’),
write(Phone)
; write(’No telephone number available for <B>’),
write(Name),
write(’</B>.’).
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
The code above is quite simple. On the other hand, the interspersion throughout
the text of calls to write with HTML markup inside makes the code somewhat
inelegant. Also, there is no separation between computation and input/output, as
is normally desirable. It would be much preferable to have an encoding of HTML
code as Prolog terms, which could then be manipulated easily in a more elegant way,
and a predicate to translate such terms to HTML for output. This functionality,
provided by the PiLLoW library, is presented in the next section.
6 Handling HTML as Prolog Terms
Since LP/CLP systems perform symbolic processing using Herbrand terms, it seems
natural to be able to handle HTML code directly as terms. Then, such structures
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib
9
only need to be translated by appropriate predicates to HTML code when they
need to be output. In general, this relationship between HTML code and Prolog
terms allows viewing any WWW page as a Herbrand term. The predicates which
provide this functionality in PiLLoW are:
output_html(F ) Accepts in F an HTML term (or a list of HTML terms) and
sends to the standard output the text which is the rendering of the term(s) in
HTML format.
html2terms(Chars,Terms ) (also, xml2terms/2) Relates a list of HTML
(resp. XML) terms and a list of ASCII characters which are the rendering of
the terms in HTML format. This predicate is reversible (but it normalizes in
the reverse direction –see later). output_html/2 uses this predicate to transform
HTML terms in characters. Again, this is implemented via DCG parsing.
In an HTML term certain atoms and structures represent special functionality
at the HTML level. An HTML term can be recursively a list of HTML terms. The
following are legal HTML terms:
hello
[hello, world]
["This is an ", em(’HTML’), " term"]
When converting HTML terms to characters, html2terms/2 translates special
structures into the corresponding format in HTML, applying itself recursively to
their arguments. Strings are always left unchanged. HTML terms may contain logic
variables, provided they are instantiated before the term is translated or output.
This allows creating documents piecemeal, back-patching of references in documents, etc.
In the following sections we list the meaning of the principal Prolog structures
that represent special functionality at the HTML level. Only special atoms are
translated, the rest are assumed to be normal text and will be passed through to
the HTML document.
6.1 General Structures
Basically, HTML has two kinds of components: HTML elements and HTML environments. An HTML element has the form “<NAME Attributes >” were NAME
is the name of the element and Attributes is a (possibly empty) sequence of attributes, each of them being either an attribute name or an attribute assignment
as name="Value".
An HTML environment has the form “<NAME Attributes > Text </NAME>”
were NAME is the name of the environment an Attributes has the same form as
before.
The general Prolog structures that represent these two HTML constructions are:
Name $Atts (‘$/2’ is defined as an infix, binary operator.) Represents an HTML
element of name Name and attributes Atts, were Atts is a (possibly empty) list
10
D. Cabeza and M. Hermenegildo
of attributes, each of them being either an atom or a structure name =value . For
example, the term
img$[src=’images/map.gif’,alt="A map",ismap]
is translated into the HTML source
<img src="images/map.gif" alt="A map" ismap>
Note that HTML is not case-sensitive, so we can use lower-case atoms.
name (Text ) (A term with functor name/1 and argument Text) Represents an
HTML environment of name name and included text Text. For example, the
term
address(’[email protected]’)
is translated into the HTML source
<address>[email protected]</address>
name (Atts,Text ) (This is a term with functor name/2 and arguments Atts and
Text) Represents an HTML environment of name name, attributes Atts and
included text Text. For example, the term
a([href=’http://www.clip.dia.fi.upm.es/’],"Clip home")
represents the HTML source
<a href="http://www.clip.dia.fi.upm.es/">Clip home</a>
env(Name,Atts,Text ) Equivalent to Name (Atts,Text ).
begin(Name,Atts ) It translates to the start of an HTML environment of name
Name and attributes Atts. There exists also a begin(Name ) structure. Useful,
in conjunction with the next structure, when including in a document output
generated by an existing piece of code (e.g. Name = pre). Its use is otherwise
discouraged.
end(Name ) Translates to the end of an HTML environment of name Name.
Now we can rewrite the previous example as follows (note how the use of the
logic variable Response allows injecting the result of the call to response/1 into
the output term, using unification):
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
’Content-type: text/html\n\n’,
html([title(’Telephone database’),
img$[src=’phone.gif’],
h2(’Telephone database’),
hr$[],
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 11
Response])]). %% Using the logic variable.
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
Any HTML construction can be represented with these structures (except comments and declarations, which could be included as atoms or strings), but the
PiLLoW library provides additional, specific structures to simplify HTML creation.
6.2 Specific Structures
In this section we will list some special structures for HTML which PiLLoW understands. While in many cases using the general structures (with the native HTML
names) is probably good practice, using specific structures such as these can sometimes be convenient. Also, some of these structures have special functionality (e.g.,
prolog_term/1). A predicate html_expansion/2 is provided which allows defining new structures (tables, layers, etc). Specific structures include (the reader is
referred to the PiLLoW manual for a full listing):
start Used at the beginning of a document (translates to <html>).
end Used at the end of a document (translates to </html>).
-- Produces a horizontal rule (translates to <hr>).
\\ Produces a line break (translates to <br>).
$ Produces a paragraph break (translates to <p>).
comment(Comment ) Used to insert an HTML comment (translates to
<!-- Comment -->).
declare(Decl ) Used to insert an HTML declaration – seldom used (translates to
<!Decl >).
image(Addr ) Used to include an image of address (URL) Addr (translates to an
<img> element).
image(Addr,Atts ) As above with the list of attributes Atts.
ref(Addr,Text ) Produces a hypertext link, Addr is the URL of the
referenced resource, Text is the text of the reference (translates to
<a href="Addr">Text</a>).
label(Label,Text ) Labels Text as a target destination with label Label (translates to <a name="Label">Text</a>).
heading(N,Text ) Produces a heading of level N (1 ≤ N ≤ 6), Text is the text to
be used as heading – useful when one wants a heading level relative to another
heading (translates to a <hN > environment).
12
D. Cabeza and M. Hermenegildo
itemize(Items ) Produces a list of bulleted items, Items is a list of corresponding
HTML terms (translates to a <ul> environment).
enumerate(Items ) Produces a list of numbered items, Items is a list of corresponding HTML terms (translates to an <ol> environment).
description(Defs ) Produces a list of defined items, Defs is a list whose elements
are definitions, each of them being a Prolog sequence (composed by ’,’/2 operators). The last element of the sequence is the definition, the other (if any) are
the defined terms (translates to an <dl> environment).
nice_itemize(Img,Items ) Produces a list of bulleted items, using the image Img
as bullet. The predicate icon_address/2 provides a colored bullet.
preformatted(Text ) Used to include preformatted text, Text is a list of HTML
terms, each element of the list being a line of the resulting document (translates
to a <pre> environment).
verbatim(Text ) Used to include text verbatim, special HTML characters
(<,>,&,") are translated into its quoted HTML equivalent.
prolog_term(Term ) Includes any prolog term Term, represented in functional
notation. Variables are output as _.
nl Used to include a newline in the HTML source (just to improve human readability).
entity(Name ) Includes the entity of name Name (ISO-8859-1 special character).
cgi_reply This is not HTML, rather, the CGI protocol requires this content
descriptor to be used by CGI executables (including form handlers) when replying
(translates to “Content-type: text/html”).
pr Includes in the page a graphical logo with the message “Developed using the
PiLLoW Web programming library”, which points to the manual and library
source.
With these additional structures, we can rewrite the previous example as follows
(note that in this example the use of heading/2 or h2/1 is equally suitable):
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 13
end]).
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
We have not included above the specific structures for creating forms. They are
included and explained in the following section.
6.3 Specific Structures for Forms
In this section we explain the structures which represent the various elements related
to forms:
start_form(Addr [,Atts ]) Specifies the beginning of a form. Addr is
the address (URL) of the program that will handle the form, and
Atts other attributes of the form, as the method used to invoke it.
If Atts is not present the method defaults to POST. (Translates to
<form action="Addr " Atts >.)
start_form Specifies the beginning of a form without assigning address to the
handler, so that the form handler will be the cgi-bin executable producing the
form.
end_form Specifies the end of a form (translates to </form>).
checkbox(Name,State ) Specifies an input of type checkbox with name Name,
State=on if the checkbox is initially checked (translates to an <input> element).
radio(Name,Value,Selected ) Specifies an input of type radio with name Name
(several radio buttons which are interlocked must share their name), Value is the
the value returned by the button, if Selected=Value the button is initially checked
(translates to an <input> element).
input(Type,Atts ) Specifies an input of type Type with a list of attributes Atts.
Possible values of Type are text, hidden, submit, reset, . . . (translates to an
<input> element).
textinput(Name,Atts,Text ) Specifies an input text area of name Name. Text
provides the default text to be shown in the area, Atts a list of attributes (translates to a <textarea> environment).
option(Name,Val,Options ) Specifies a simple option selector of name Name,
Options is the list of available options and Val is the initial selected option
(if Val is not in Options the first item is selected) (translates to a <select>
environment).
14
D. Cabeza and M. Hermenegildo
menu(Name,Atts,Items ) Specifies a menu of name Name, list of attributes Atts
and list of options Items. The elements of the list Items are marked with the
prefix operator ‘$’ to indicate that they are selected (translates to a <select>
environment).
For example, in order to generate a form suitable for sending input to the previously described phone database handler one could execute the following goal:
output_html([
start,
title(’Telephone database’),
heading(2,’Telephone database’),
$,
start_form(’http://www.clip.dia.fi.upm.es/cgi-bin/phone_db.pl’),
’Click here, enter name of clip member, and press Return:’,
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
Of course, one could have also simply written directly the resulting HTML document:
<html>
<title>Telephone database</title>
<h2>Telephone database</h2>
<p>
<form method="POST"
action="http://www.clip.dia.fi.upm.es/cgi-bin/phone_db.pl">
Click here, enter name of clip member, and press Return:
<br>
<input type="text" name="person_name" size="20">
</form>
</html>
7 Merging the Form Producer and the Handler
An interesting practice when producing HTML forms and handlers is to merge the
operation of the form producer and the handler into the same program. The idea is
to produce a generalized handler which receives the form input, parses it, computes
the answer, and produces a new document which contains the answer to the input,
as well as a new form. A special case must be made for the first invocation, in
which the input would be empty, and then only the form should be generated. The
following is an example which merges the producer and the handler for the phones
database:4
4
Notice that when only one text field exists in a form, the form can be submitted by simply
pressing “Return” inside the text field.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 15
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
start_form,
’Click here, enter name of clip member, and press Return:’,
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
response(Name, Response) :form_empty_value(Name) ->
Response = []
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone,$]
; Response = [’No telephone number available for ’,b(Name),’.’,$].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
This combination of the form producer and the handler allows producing applications that give the impression of being interactive, even if each step involves starting
and running the handler to completion. Note that forms can contain fields which
are not displayed and are passed as input to the next invocation of the handler.
This allows passing state from one invocation of the handler to the next one.
Finally, a note about testing and debugging CGI scripts: this is unfortunately
not as straightforward as it could be. Useful techniques include carefully checking
permissions, looking at the data logs of the server, replacing predicates such as
get_form_... with versions that print what is really being received, etc.
16
D. Cabeza and M. Hermenegildo
8 Templates
A problem in the previous programs is that the layout of the output page is not
easily configurable –it is hard-coded in the source and can only be changed by
modifying the program. This is something that a normal user (or even an expert
programmer if the size of the program is large) may not want to do. In order to
address this, PiLLoW provides a facility for reading in “HTML templates” (also
XML templates), and converting them into a term format in which it is very natural
to manipulate them. An HTML template is a file which contains standard HTML
code, but in which “slots” can be defined and given an identifier by means of a
special tag. These slots represent parts of the HTML code in which other HTML
code can be inserted. Once the HTML template is read by PiLLoW, such slots
appear as free logic variables in the corresponding PiLLoW terms. In this way, the
user can define a layout with an HTML editor of choice, taking care of marking the
“left out” parts with given names. These parts will then be filled appropriately by
the program. The functionality associated with parsing such terms is encapsulated
in the following predicate:
html_template(Chars, Terms, Dict ) Parses the string Chars as the contents
of an HTML template and unifies Terms with the list of HTML terms comprised
in the template, substituting occurrences of the special tag <V>name</V> with
prolog variables. Dict is instantiated to the dictionary of such substitutions, as a
list of name=Variable pairs.
In the following example a template file called TlfDB.html is assumed to hold the
formatting of the output page, defining an HTML variable called “response” which
will be substituted by the response of the CGI program. Note that the predicate
file_to_string/2 (defined in Ciao library file_utils) reads a file and returns
in its second argument the contents of the file as a list of character codes. Note also
that calling html_template/3 with the third argument instantiated to [response
= Response] has the effect of instantiating the “slot” in HTML_terms to the contents
of Response (this makes use of the fact that there is only one slot on the template;
normally, a call to member/2 is used to locate the appropriate name=Variable pair).
#!/usr/local/bin/ciao-shell
:- include(library(pillow)).
:- use_module(library(file_utils)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
file_to_string(’TlfDB.html’, Contents),
html_template(Contents, HTML_terms, [response = Response]),
output_html([cgi_reply|HTML_terms]).
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 17
response(Name, Response) :form_empty_value(Name) ->
Response = []
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone,$]
; Response = [’No telephone number available for ’,b(Name),’.’,$].
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
An example of the contents of the template file could be:
<HTML><HEAD><TITLE>Telephone database</TITLE></HEAD>
<BODY background="bg.gif">
<IMG src="phone.gif">
<H2>Telephone database</H2>
<HR>
<V>response</V>
<FORM method="POST">
Click here, enter name of clip member, and press Return:<BR>
<INPUT type="text" name="person_name" size="20"></FORM>
</BODY>
</HTML>
9 Accessing WWW documents
The facilities presented in the previous sections allow generating HTML documents,
including forms, and handling the input coming from forms. In many applications
such as search tools, content analyzers, etc., it is also desirable to be able to access
documents on the Internet. Such access is generally accomplished through protocols
such as FTP and HTTP which are built on top of TCP/IP. In LP/CLP systems which
have TCP/IP connectivity (i.e., a sockets/ports interface) the required protocols
can be easily coded in the source language using such facilities and DCG parsers.
At present, only the HTTP protocol is supported by PiLLoW. As with HTML code,
the library uses an internal representation of Uniform Resource Locators (URLs), to
be able to manipulate them easily, and provides predicates which translate between
the internal representation and the textual form. The facilities provided by PiLLoW
for accessing WWW documents include the following predicates:
url_info(URL,Info ) Translates a URL URL to an internal structure Info which
details its various components and vice-versa. For now non-HTTP URLs make
the predicate fail. E.g.
url_info(’http://www.foo.com/bar/scooby.txt’,Info)
gives Info = http(’www.foo.com’,80,"/bar/scooby.txt"),
url_info(URL, http(’www.foo.com’,2000,"/bar/scooby.txt")
gives URL = "http://www.foo.com:2000/bar/scooby.txt" (a string).
18
D. Cabeza and M. Hermenegildo
url_info_relative(URL,BaseInfo,Info ) Translates a relative URL URL which
appears in the HTML page referred to by BaseInfo (given as an url_info structure) to a complete url_info structure Info. Absolute URLs are translated as
with the previous predicate. E.g.
url_info_relative("/guu/intro.html",
http(’www.foo.com’,80,"/bar/scoob.html"), Info)
gives Info = http(’www.foo.com’,80,"/guu/intro.html")
url_info_relative("dadu.html",
http(’www.foo.com’,80,"/bar/scoob.html"), Info)
gives Info = http(’www.foo.com’,80,"/bar/dadu.html").
url_query(Dic,Args ) Translates a list of attribute=value pairs Dic (in the same
form as the dictionary returned by get_form_input/1) to a string Args for appending to a URL pointing to a form handler.
fetch_url(URL,Request,Response ) Fetches a document from the Internet. URL
is the Uniform Resource Locator of the document, given as a url_info structure.
Request is a list of options which specify the parameters of the request, Response
is a list which includes the parameters of the response. The request parameters
available include:
head To specify that we are only interested in the header.
timeout(Time ) Time specifies the maximum period of time (in seconds) to wait
for a response. The predicate fails on timeout.
if_modified_since(Date ) Get
document
only if newer than Date. An example of a structure that represents a date
is date(’Tuesday’,15,’January’,1985,’06:14:02’).
user_agent(Name ) Provide a user-agent field.
authorization(Scheme,Params ) Provides an authentication field when accessing restricted sites.
name (Param ) Any other functor translates to a field of the same name (e.g.
from(’user@machine’)).
The parameters which can be returned in the response list include (see the
HTTP/1.0 definition for more information):
content(Content ) Returns in Content the actual document text, as a list of
characters.
status(Type,Code,Phrase ) Gives the status of the response. Type can be any
of informational, success, redirection, request_error, server_error or
extension_code, Code is the status code and Phrase is a textual explanation
of the status.
pragma(Data ) Miscellaneous data.
message_date(Date ) The time at which the message was sent.
location(URL ) The document has moved to this URL.
http_server(Server ) Identifies the server responding.
allow(methods ) List of methods allowed by the server.
last_modified(Date ) Date/time at which the sender believes the resource was
last modified.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 19
expires(Date ) Date/time after which the entity should be considered stale.
content_type(Type,Subtype,Params ) Returns the MIME type/subtype of the
document.
content_encoding(Type ) Encoding of the document (if any).
content_length(Length ) Length is the size of the document, in bytes.
authenticate(Challenges ) Request for authentication.
html2terms(Chars,Terms ) We have already explained how this predicate transforms HTML terms to HTML format. Used the other way around it can parse
HTML code, for example retrieved by fetch_url. The resulting list of HTML
terms Terms is normalized: it contains only comment/1, declare/1, env/3 and
$/2 structures.
For example, a simple fetch of a document can be done as follows:
url_info(’http://www.foo.com’,UI), fetch_url(UI,[],R),
member(content(C),R), html2terms(C, HTML_Terms).
Note that if an error occurs (the document does not exist or has moved, for example)
this will simply fail. The following call retrieves a document if it has been modified
since October 6, 1999:
fetch_url(http(’www.foo.com’,80,"/doc.html"),
[if_modified_since(’Wednesday’,6,’October’,1999,’00:00:00’)],
R).
This last one retrieves the header of a document (with a timeout of 10 seconds) to
get its last modified date:
fetch_url(http(’www.foo.com’,80,"/last_news.html"),
[head,timeout(10)],R),
member(last_modified(Date),R).
The following is a simple application illustrating the use of fetch_url and
html2terms. The example defines check_links(URL,BadLinks ). The predicate
fetches the HTML document pointed to by URL and scours it to check for links
which produce errors when followed. The list BadLinks contains all the bad links
found, stored as compound terms of the form: badlink(Link,Error) where Link
is the problematic link and Error is the error explanation given by the server.
check_links(URL,BadLinks) :url_info(URL,URLInfo),
fetch_url(URLInfo,[],Response),
member(content_type(text,html,_),Response),
member(content(Content),Response),
html2terms(Content,Terms),
check_source_links(Terms,URLInfo,[],BadLinks).
check_source_links([],_,BL,BL).
check_source_links([E|Es],BaseURL,BL0,BL) :-
20
D. Cabeza and M. Hermenegildo
check_source_links1(E,BaseURL,BL0,BL1),
check_source_links(Es,BaseURL,BL1,BL).
check_source_links1(env(a,AnchorAtts,_),BaseURL,BL0,BL) :member((href=URL),AnchorAtts), !,
check_link(URL,BaseURL,BL0,BL).
check_source_links1(env(_Name,_Atts,Env_html),BaseURL,BL0,BL) :- !,
check_source_links(Env_html,BaseURL,BL0,BL).
check_source_links1(_,_,BL,BL).
check_link(URL,BaseURL,BL0,BL) :url_info_relative(URL,BaseURL,URLInfo), !,
fetch_url_status(URLInfo,Status,Phrase),
( Status \== success ->
name(P,Phrase),
name(U,URL),
BL = [badlink(U,P)|BL0]
; BL = BL0
).
check_link(_,_,BL,BL).
fetch_url_status(URL,Status,Phrase) :fetch_url(URL,[head,timeout(20)],Response), !,
member(status(Status,_,Phrase),Response).
fetch_url_status(_,timeout,"Timeout").
10 Providing Code Through the WWW
A facility which can be easily built on top of the primitives presented so far is that
of “remote WWW modules,” i.e., program modules which reside on the net at a
particular HTTP address in the same way that normal program modules reside
in a particular location in the local file system. This allows for example always
fetching the most recent version of a given library (e.g., PiLLoW) when a program
is compiled. For example, the form handler of Section 6.1, if rewritten as
#!/usr/local/bin/ciao-shell
:- use_module(’http://www.clip.dia.fi.upm.es/lib/pillow.pl’).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
...
would load the current version of the library each time it is executed. This generalized module declaration is just syntactic sugar, using expand_term, for a document
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 21
www.xxx.yyy
<form action="http://www.xxx.yyy/am_inter.cgi">
HTTP
Server
2
1
WWW
Browser
form data
8
3
form
reply
4
7
5
Active
Module
am_inter
6
predicate(Arg1,Arg2,...,ArgN)
Fig. 3. The Forms Interface using Active Modules
fetch, using fetch_url, followed by a standard use_module declaration. It is obviously interesting to combine this facility with caching strategies. An interesting
(and straightforward to implement) additional feature is to fetch remote byte-code
(as generally done by use_module), if available, but this is only possible if the two
systems use the same byte-code (this can normally be checked easily in the bytecode itself). Also, it may be interesting to combine this type of code downloading
with WWW document accesses, so that code is downloaded automatically when a
particular document is fetched. This issue is addressed in Section 12. Finally, there
are obvious security issues related to downloading code in general, which can be
addressed with standard techniques such as security signatures.
11 A High-Level Model of Client-Server Interaction: Active Modules
Despite its power, the cgi-bin interface also has some shortcomings. The most serious is perhaps the fact that the handler is started and expected to terminate for
each interaction. This has two disadvantages. First, no state is preserved from one
query to the next. However, as mentioned before, this can be fixed by passing the
state through the form (using hidden fields), by saving it in a temporary file at the
server side, by using “cookies”, etc. Second, and more importantly, starting and
stopping the application may be inefficient. For example, if the idea is to query
a large database or a natural language understanding system, it may take a long
time to start and stop the system. In order to avoid this we propose an alternative
architecture for cgi-bin applications (a similar idea, although not based on the idea
of active modules, has been proposed independently by Ken Bowen (Bowen, 1996)).
The basic idea is illustrated in Figure 3. The operation is identical to that of
standard form handlers, as illustrated in Figure 2, up to step 3. In this step, the
handler started is not the application itself, but rather an interface to the actual
application, which is running continuously and thus contains state. Thus, only the
22
D. Cabeza and M. Hermenegildo
interface is started and stopped with every transaction. The interface simply passes
the form input received from the server (4) to the running application (5) and then
forwards the output from the application (6) to the server before terminating, while
the application itself continues running. Both the interface and the application can
be written in LP/CLP, using the predicates presented. The interface can be a simple
script, while the application itself will be typically compiled.
An interesting issue is that of communication between interface and application.
This can of course be done through sockets. However, as a cleaner and much simpler alternative, the concept of active modules (Cabeza & Hermenegildo, 1995) can
be used to advantage in this application. An active module (or an active object,
if modularity is implemented via objects) is an ordinary module to which computational resources are attached (for example, a process on a UNIX machine), and
which resides at a given (socket) address on the network.5 Compiling an active
module produces an executable which, when running, acts as a server for a number of relations, which are the predicates exported by the module. The relations
exported by the active module can be accessed by any program on the network
by simply “loading” the module and thus importing such “remote relations.” The
idea is that the process of loading an active module does not involve transferring
any code, but rather setting up things so that calls in the local module are executed as remote procedure calls to the active module, possibly over the network.
Except for compiling it in a special way, an active module is identical from the
programmer point of view to an ordinary module. Also, a program using an active
module imports it and uses it in the same way as any other module, except that it
uses “use_active_module” rather than “use_module” (see below). Also, an active
module has an address (network address) which must be known in order to use it.
The address can be announced by the active module when it is started via a file or
a name server (which would be itself another active module with a fixed address).
We now present the constructs related to active modules in Ciao:
:- use_active_module(Module,Predicates ) A declaration used to import the
predicates in the list Predicates from the active module Module . From this
point on, the code should be written as if a standard use_module/2 declaration
had been used. The declaration needs the following predicate to be accessible
from the module.
module_address(Module,Address ) This predicate must return in Address the
address of Module , for any active module imported in the code. There are a
number of standard libraries defining versions of this predicate.
save_addr_actmod(Address ) This predicate should define a way to publish
Address , to be used in active modules (the name of the active module is taken
as the name of the current executable). There are a number of standard libraries
defining versions of this predicate, which are in correspondence with the libraries
which define versions of the previous predicate.
5
It is also possible to provide active modules via a WWW address. However, we find it more
straightforward to simply use socket addresses. In any case, this is generally hidden inside the
access method and can be thus made transparent to the user.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 23
make_actmod(ModuleFile,PublishModule ) Makes an active module executable
from the module residing in ModuleFile , using address publish module of
name PublishModule . When the executable is run (for example, at the operating system level by “Module &”), a socket is created and the hook predicate save_addr_actmod/1 mentioned above (which is supposed to be defined in
PublishModule ) is called in order to export the active module address as required. Then, a standard driver is run to attend network requests for the module
exported predicates. Note that the code of ModuleFile does not need to be
written in any special way.
This scheme is very flexible, allowing to completely configure the way active
modules are located. This is accomplished by writing a pair of libraries, one defining the way an active module address is published, and a second defining the way
the address of a given active module is found. For example, the Ciao standard libraries include as an example an implementation (libraries filebased_publish and
filebased_locate) which uses a directory accessible by all the involved machines
(via NFS) to store the addresses of the active modules, and the module_address/2
predicate examines this directory to find the required data. Other solutions provided
as examples include posting the address at a WWW address (webbased_publish
/ webbased_locate), and an implementation of a name server, that is, another
active module (this one with a known, fixed address) that records the addresses of
active modules and supplies this data to the modules that import it, serving as a
contact agency between servers and clients.
From the implementation point of view, active modules are essentially daemons:
Prolog executables which are started as independent processes at the operating
system level. In the Ciao system library, communication with active modules is
implemented using sockets (thus, the address of an active module is a UNIX socket
in a machine). Requests to execute goals in the module are sent through the socket
by remote programs. When such a request arrives, the process running the active
module takes it and executes it, returning through the socket the computed results.
These results are then taken by the remote processes.
Thus, when the compiler finds a use_active_module declaration, it defines the
imported predicates as remote calls to the active module. For example, if the predicate P is imported from the active module M , the predicate would be defined as
P :- module_address(M,A), remote_call(A,P )
Compiling the following code as an active module, by writing at the Ciao
toplevel “make_actmod(phone_db, ’actmods/filebased_publish’)” (or, using
the standalone compiler, by executing “ciaoc -a ’actmods/filebased_publish’
phone_db”), creates an executable phone_db which, when started as a process (for
example, by typing “phone_db &” at a UNIX shell prompt) saves its address (i.e.,
that of its socket) in file phone_db.addr and waits for queries from any module
which “imports” this module (it also provides a predicate to dynamically add information to the database):
:- module(phone_db,[response/2,add_phone/2]).
24
D. Cabeza and M. Hermenegildo
response(Name, Response) :form_empty_value(Name) ->
Response = ’You have to provide a name.’
; phone(Name, Phone) ->
Response = [’Telephone number of ’,b(Name),’: ’,Phone]
; Response = [’No telephone number available for ’,b(Name),’.’].
add_phone(Name, Phone) :assert(phone(Name, Phone)).
:- dynamic phone/2.
phone(daniel, ’336-7448’).
phone(manuel, ’336-7435’).
phone(sacha, ’543-5316’).
The following simple script can be used as a cgi-bin executable which will be the
active module interface for the previous active module. When started, it will process
the form input, issue a call to response/2 (which will be automatically handled
by the phone_db active module), and produce a new form before terminating. It
will locate the address of the phone_db active module via the module_address/2
predicate defined in library ’actmods/filebased_locate’.
#!/usr/local/bin/ciao-shell
:- use_active_module(phone_db,[response/2]).
:- use_module(library(’actmods/filebased_locate’)).
:- include(library(pillow)).
main(_) :get_form_input(Input),
get_form_value(Input,person_name,Name),
response(Name,Response),
output_html([
cgi_reply,
start,
title(’Telephone database’),
image(’phone.gif’),
heading(2,’Telephone database’),
--,
Response,
$,
start_form,
’Click here, enter name of clip member, and press Return:’,
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 25
\\,
input(text,[name=person_name,size=20]),
end_form,
end]).
There are many enhancements to this simple schema which, for brevity, are only
sketched here. One is to add concurrency to the active module (or whatever means of
handling the client-server interaction is being used), in order to handle queries from
different clients concurrently. This is easy to do in systems that support concurrency natively, such as Ciao, BinProlog/µ2-Prolog, AKL, Oz, and KL1. We feel that
Ciao can offer advantages in this area because it offers compatibility with Prolog
and CLP systems while at the same time efficiently supporting concurrent execution of clause goals via local or distributed threads (Carro & Hermenegildo, 1999).
Such goals can communicate at different levels of abstraction: sockets/ports, the
shared fact database (similarly to a blackboard), or shared variables. BinProlog/µ2Prolog also supports threads, with somewhat different communication mechanisms (Tarau, 1996; De Bosschere, 1989). Finally, as shown in (Szeredi et al., 1996),
it is also possible to exploit the concurrency present in or-parallel Prolog systems
such as Aurora for implementing a multitasking server.
It is also interesting to set up things so that a single active module can handle
different forms. This can be done even dynamically (i.e., the capabilities of the active
module are augmented on the fly, being able to handle a new form), by designating
a directory in which code to be loaded by the active module would be put, the
active module consulting the directory periodically to increase its functionalities.
Finally, another important issue that has not been addressed is that of providing
security, i.e., ensuring that only allowed clients connect to the active module. As
in the case of remote code downloading, standard forms of authentication based on
codes can be used.
12 Automatic Code Downloading and Local Execution
In this section we describe an architecture which, using only the facilities we have
presented in previous sections, allows the downloading and local execution of Prolog
(or other LP/CLP) code by accessing a WWW address, without requiring a special
browser. This is a complementary approach to giving WWW access to an active
module in the sense that it provides code which will be executed in the client
machine (à la Java). More concretely, the functionality that we desire is that by
simply clicking on a WWW pointer, and transparently for the user, remote Prolog
code is automatically downloaded in such a way that it can be queried via forms
and all the processing is done locally.
To allow this, the HTTP server on the server machine is configured to give a
specific mime.type (for example application/x-prolog) to the files which will
hold WWW-downloadable Prolog code (for example those with a special suffix, like
.wpl). On the other side, the browser is configured to start the wpl_handler helper
application when receiving data of type application/x-prolog. This wpl_handler
26
D. Cabeza and M. Hermenegildo
<form action="http://localhost/wpl_questioner.cgi">
www.xxx.yyy
WWW
Browser
http://www.xxx.yyy/app.wpl
form data
/tmp/x1349
form reply
1
app.wpl
2
11
wpl_handler /tmp/x1349
7
8
wpl_handler
loaded
wpl_questioner
3
?
6
4
10
9
5
FormReply
Active
Module
code
FormData
answerform(FormData, FormReply)
loadcode(’/tmp/x1349’)
Fig. 4. Automatic code downloading architecture
application is the interface to a Prolog engine which will execute the WWW downloaded code, acting as an active module. We now sketch the procedure (see figure
4):
1. The form that will be used to query the downloaded code (and which we
assume already loaded on the browser) contains a link which points to a
WWW-aware Prolog code file. Clicking on this link produces the download
as explained below. Note that for browsers that can handle multipart/mixed
mime types (such as most modern browsers), the form and the code file could
alternatively be combined in the same document. However, for brevity, we
will only describe the case when they are separate. The handler for the form
is specified as the local cgi-bin executable wpl_questioner.cgi.
2. As the server of the file tells the browser that this page is of type
application/x-prolog, the browser starts a wpl_handler and passes the
file to it (in this example by saving the file in a temporal directory and passing its name).
3. The wpl_handler process checks whether a Prolog engine is currently running
for this browser and, if necessary, starts one. This Prolog engine is configured
as an active module.
4. Then, through a call to a predicate of the active module “loadcode(File )”
the handler asks the active module to read the code.
5. The active module reads the code and compiles it.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 27
6. wpl_handler waits for the active module to complete the compilation and
writes a “done” message to the browser.
7. The browser receives the “done” message.
8. Now, when the “submit” button in the form is pressed, and following the
standard procedure for forms, the browser starts a wpl_questioner process,
sending it the form data.
9. The wpl_questioner process gets this form data, translates it to a dictionary
FormData and passes it to the active module through a call to its exported
predicate answerform(FormData,FormReply ).
10. The active module processes this request, and returns in FormReply a WWW
page (as a term) which contains the answer to it (and possibly a new form).
11. The wpl_questioner process translates FormReply to raw HTML and gives
it back to the browser, dying afterwards. Subsequent queries to the active
module can be accomplished either by going back to the previous page (using
the “back” button present in many browsers) or, if the answer page contains
a new query form, by using it. In any case, the procedure continues at 8.
The net effect of the approach is that by simply clicking on a WWW pointer,
remote Prolog code is automatically downloaded to a local Prolog engine. Queries
posed via the form are answered locally by the Prolog engine.
There are obvious security issues that need to be taken care of in this architecture. Again, standard authentication techniques can be used. However, since source
code is being passed around, it is comparatively easy to verify that no dangerous
predicates (for example, perhaps those that can access files) are executed. Note
again that it is also possible to download bytecode, since this is supported by most
current LP/CLP systems, using a similar approach.
13 Related Work
Previous general purpose work on WWW programming using computational logic
systems includes, to the best of our knowledge, the publicly available html.pl
library (Cabeza & Hermenegildo, 1996a) and manual, and the LogicWeb system
(Loke & Davison, 1996) (the PiLLoW library was also described previously in
(Cabeza et al., 1996)). The html.pl library was built by D. Cabeza and M.
Hermenegildo, using input from L. Naish’s forms code for NU-Prolog and M.
Hermenegildo and F. Bueno’s experiments building a WWW interface to the CHAT80 (Warren & Pereira, 1982) program. It was released as a publicly available WWW
library for LP/CLP systems and announced, among other places, in the Internet
comp.lang.prolog newsgroup (Cabeza & Hermenegildo, 1996b). The library has
since been ported to a large number of systems and adapted by several Prolog vendors, as well as used by different programmers in various institutions. In particular,
Ken Bowen has ported the library to ALS Prolog and extended it to provide group
processing of forms and an alternative to our use of active modules (Bowen, 1996).
The present work is essentially a significant extension of the html.pl library.
The main other previous body of work related to general-purpose interfacing of
28
D. Cabeza and M. Hermenegildo
logic programming and the WWW that we have knowledge of is the LogicWeb
(Loke & Davison, 1996) system, by S.W. Loke and A. Davison. The aim of LogicWeb is to use logic programming to extend the concept of WWW pages, incorporating in them programmable behavior and state. In this, it shares goals with Java.
It also offers rich primitives for accessing code in remote pages and module structuring. The aims of LogicWeb are different from those of html.pl/PiLLoW. LogicWeb
is presented as a system itself, and its implementation is done through a tight integration with the Mosaic browser, making use of special features of this browser.
In contrast, html.pl/PiLLoW is a general purpose library, meant to be used by a
general computational logic systems and is browser-independent. html.pl/PiLLoW
offers a wide range of functionalities, such as syntax conversion between HTML and
logic terms, access predicates for WWW pages, predicates for handling forms, etc.,
which are generally at a somewhat lower level of abstraction than those of LogicWeb. We believe that using PiLLoW and the ideas sketched in this paper it is
possible to add the quite interesting functionality offered by LogicWeb to standard
LP and CLP systems. We have shown some examples including access to passive
remote code (modules with an ftp or http address) from programs and automatic
remote code access and querying using standard browsers and forms. In addition,
we have discussed active remote code, where the functionality, rather than the code
itself, is exported.
More recently, a larger body of work on the topic was presented at the workshop held on the topic of Logic Programming and the Internet at the 1996
Joint International Conference and Symposium on Logic Programming (where
also a previous version of this paper was presented). The work presented in
(S.W. Loke & Sterling, 1996) is based on LogicWeb, and aims to provide distributed lightweight databases on the WWW. As with the basic LogicWeb system,
we believe that the PiLLoW library can be used to implement in other systems
the interesting ideas proposed therein. As briefly mentioned before, the work in
(Szeredi et al., 1996) proposes an architecture similar to that of our active modules in order to handle form requests. In this solution the handling multiple requests is performed by using or-parallelism. While we feel that and-parallelism (as
in &-Prolog’s or Ciao’s threads) is more natural for modeling this kind of concurrency, the ideas proposed are quite interesting. The ECLiPSe HTTP-library
(Ph. Bonnet & Thomsen, 1996), aimed at implementing INTERNET agents, offers
functionality that is in part similar to that of the Ciao html.pl/PiLLoW libraries,
including facilities that are similar to our active modules. The approach is different,
however, in several respects. The ECLiPSe library implements special HTTP servers
and clients. In contrast, PiLLoW uses standard HTTP servers and interfaces. Using special purpose servers may be interesting because the approach possibly allows
greater functionality. On the other hand this approach in general requires either
the substitution of the standard server on a given machine or setting the special
server at a different socket address from the standard one. The ECLiPSe library
also contains functionality that is related to our active modules, although the interface provided is at a lower level. Finally, other papers describing very interesting
WWW applications are being presented regularly, which underline the suitability of
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 29
computational logic systems for the task. We believe that the Ciao PiLLoW library
can contribute to making it even easier to develop such applications in the future.
Additional work on the topic of Logic Programming and the Internet can
be found in the proceedings of the workshop sponsored by the CompulogNet research network. The reader is referred to the tutorials and papers presented in these two workshops for more information on a number of applications, other libraries, and topics such as interfacing and compilation from
computational logic systems to Java. Examples of Prolog systems interfaced
with Java are BinProlog (see http://clement.info.umoncton.ca/BinProlog),
Ciao (Bueno et al., 1997), and others (Calejo, 1999). Experimental Prolog to
Java compilers have been built both in academia (see for example jProlog at
http://www.cs.kuleuven.ac.be/~bmd/PrologInJava/) and Commercially (see
for example the IF Prolog tools http://www.ifcomputer.com). This approach is
quite attractive, although the results cannot compete in performance with conventional Prolog compilers (it is open for research whether improvements in Java
performance or improved Prolog-to-Java compilation technology can bridge the
gap). Other commercial work on the topic of interfacing Prolog and the WWW (in
addition to that done on the ALS system mentioned above) include the Amzi! Prolog WebLS System (http://www.amzi.com/share.htm ) and the LPA PrologWeb
System (http://www.lpa.co.uk ).
Recent work using PiLLoW includes the “Web Integrator”(Davulcu et al., 1999)
–a webbase system that integrates data from various Web sources, and allows
users to query these Web sources as if they were a single database– and WebDB
(Cabeza & Hermenegildo, 1998) –a WWW-based database management interface.
Also, within the RadioWeb project (Partners, 1997), we have developed (in collaboration with the group of M. Codish at Ben Gurion University) a constraint-based
language for describing WWW page layout and style rules and an engine which,
by interpreting these rules, can generate WWW sites which dynamically adapt to
parameters such as user characteristics (Cederberg & CLIP Group, 1999).
Additional applications developed with the PiLLoW library can be accessed
from the PiLLoW WWW site (see later). A page with pointers to the proceedings of the previously mentioned workshops, as well as other information (including technical reports and tutorial) regarding the topic of Logic
Programming, Constraint Programming, and the Internet is maintained at
http://www.clip.dia.fi.upm.es/lpnet/.
14 Conclusions and Future Work
We have discussed from a practical point of view a number of issues involved in
writing Internet and WWW applications using LP/CLP systems. In doing so, we
have described PiLLoW, an Internet/WWW programming library for LP/CLP
systems. PiLLoW provides facilities for generating HTML/XML structured documents, producing HTML forms, writing form handlers, processing HTML/XML
templates, accessing and parsing WWW documents, and accessing code posted
at HTTP addresses. We have also described the architecture of some application
30
D. Cabeza and M. Hermenegildo
classes, including automatic code downloading, using a high-level model of clientserver interaction, active modules. Finally we have also described an architecture for
automatic LP/CLP code downloading for local execution, using generic browsers.
We believe that the Ciao PiLLoW library can ease substantially the process of
developing WWW applications using computational logic systems.
We have recently developed several extensions to the library (for example, for
setting and getting “cookies”), and sample applications which make extensive use
of concurrency (on those LP/CLP systems that support it) to overlap network
requests. We have also developed a complementary library for interfacing Prolog
with the Virtual Reality Modeling Language VRML (Smedbäck et al., 1999).
In addition to being included as part of the Ciao system, the PiLLoW library is
provided as a standard, standalone public domain library for SICStus Prolog and
other Prolog and CLP systems, supporting most of its functionality. Please contact
the authors or consult our WWW site http://www.clip.dia.fi.upm.es and the
PiLLoW
page
at http://www.clip.dia.fi.upm.es/Software/pillow/pillow.html for download details and an up-to-date online version of the PiLLoW manual. The Ciao
Prolog system is also freely available from http://www.clip.dia.fi.upm.es and
http://www.ciaoprolog.org.
Acknowledgments
The authors would like to thank Lee Naish, Mats Carlsson, Tony Beaumont, Ken
Bowen, Michael Codish, Markus Fromherz, Paul Tarau, Andrew Davison, and Koen
De Bosschere for useful feedback on previous versions of this document and the
PiLLoW code. The first versions of the Ciao system and the html.pl library were
developed under partial support from the ACCLAIM ESPRIT project. Subsequent
development has occurred in the context of MCYT projects “ELLA” and “EDIPIA”
(MCYT TIC99-1151), ESPRIT project RADIOWEB, and NSF/CICYT collaboration “ECCOSIC” (Fulbright 98059).
References
Berners-Lee, T., Cailliau, R., Luotonen, A., Nielsen, H.F., & Secret, A. (1994). The
World-Wide Web. Communications of the ACM, 37(8), 76–82.
Bowen, K. 1996 (March).
Personal communication.
Available from
http://www.als.com/als/html_pl.html.
Bueno, F., Cabeza, D., Carro, M., Hermenegildo, M., López-Garcı́a, P., & Puebla, G. 1997
(August). The Ciao Prolog System. Reference Manual. The Ciao System Documentation
Series–TR CLIP3/97.1. School of Computer Science, Technical University of Madrid
(UPM).
Cabeza, D., & Hermenegildo, M. (1995). Distributed Concurrent Constraint Execution
in the CIAO System. Proc. of the 1995 compulog-net workshop on parallelism and
implementation technologies. Utrecht, NL: U. Utrecht / T.U. Madrid. Available from
http://www.clip.dia.fi.upm.es/.
Cabeza, D., & Hermenegildo, M. 1996a (March). html.pl: An HTML package for (C)LP
systems. Spain. Available from http://www.clip.dia.fi.upm.es/miscdocs/.
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 31
Cabeza, D., & Hermenegildo, M. 1996b (February).
LP/CLP HTML and
WWW interface publicly available. Posting in comp.lang.prolog. Available from
http://www.clip.dia.fi.upm.es/.
Cabeza, D., & Hermenegildo, M. 1997 (April). WWW Programming using Computational
Logic Systems (and the PiLLoW/Ciao Library). Proceedings of the workshop on logic
programming and the WWW at WWW6.
Cabeza, D., & Hermenegildo, M. 1998 (June). A WWW Database Management Interface for Prolog. Technical Report CLIP2/98.0. School of Computer Science, Technical
University of Madrid (UPM), Facultad Informática UPM, 28660-Boadilla del Monte,
Madrid-Spain.
Cabeza, D., Hermenegildo, M., & Varma, S. 1996 (September). The PiLLoW/Ciao Library
for INTERNET/WWW Programming using Computational Logic Systems. Proceedings
of the 1st workshop on logic programming tools for internet applications. Available from
http://clement.info.umoncton.ca/~lpnet.
Calejo, M. (1999). Java+prolog: A land of opportunities. Pages 1–2 of: Proceedings of the
first international conference on the practical application of constraint technologies and
logic programming. The Practical Application Company Ltd. Also available at
http://dev.servisoft.pt/interprolog/paclp99/default.htm.
Carlsson, M. 1988 (February). Sicstus Prolog User’s Manual. Po Box 1263, S-16313
Spanga, Sweden.
Carro, M., & Hermenegildo, M. (1999). Concurrency in Prolog Using Threads and a Shared
Database. Pages 320–334 of: 1999 international conference on logic programming. MIT
Press, Cambridge, MA, U.S.A.
Cederberg, Per, & CLIP Group, The. 1999 (June). Flexible Layout and Styling – The LaSt
Language. Technical Report D2.2.M3 CLIP 3/99.0. RADIOWEB Project.
Chikayama, T., Fujise, T., & Sekita, D. (1994). A Portable and Efficient Implementation
of KL1. Tick, Evan (ed), Proc. of the 1994 icot/nsf workshop on parallel and concurrent
programming. U. of Oregon.
Colmerauer, A. (1975). Les gramaire de metamorphose. Tech. rept. Univ. D’aix-Marseille,
Groupe De Ia.
Colmerauer, A. (1990). An Introduction to Prolog III. Communications of the ACM,
28(4), 412–418.
Davulcu, Hasan, Freire, Juliana, Kifer, Michael, & Ramakrishnan, I.V.
1999 (June).
A layered architecture for querying dynamic web content.
ACM SIGMOD international conference on management of data.
URL:
http://www.acm.org/sigmod/sigmod99/eproceedings/.
De Bosschere, K. (1989). Multi–Prolog, Another Approach for Parallelizing Prolog. Pages
443–448 of: Proceedings of parallel computing. Elsevier, North Holland.
ECRC. (1993). Eclipse user’s guide. European Computer Research Center.
Grobe, M., & Naseer, Hasan. 1998 (July).
An instantaneous introduction
to
CGI
scripts
and
HTML
forms.
Available
from
http://www.cc.ukans.edu/~acs/docs/other/forms-intro.shtml.
Hermenegildo, M. 1996 (April). Writing “Shell Scripts” in SICStus Prolog. Posting in
comp.lang.prolog. Available from http://www.clip.dia.fi.upm.es/.
Hermenegildo, M., & CLIP Group, The. (1994). Some Methodological Issues in the Design of CIAO - A Generic, Parallel, Concurrent Constraint System. Pages 123–133 of:
Principles and practice of constraint programming. LNCS, no. 874. Springer-Verlag.
Hermenegildo, M., Bueno, F., Garcı́a de la Banda, M., & Puebla, G. 1995a (December).
The CIAO Multi-Dialect Compiler and System: An Experimentation Workbench for
32
D. Cabeza and M. Hermenegildo
Future (C)LP Systems. Proceedings of the ilps’95 workshop on visions for the future of
logic programming. Available from http://www.clip.dia.fi.upm.es/.
Hermenegildo, M., Cabeza, D., & Carro, M. (1995b). Using Attributed Variables in the
Implementation of Concurrent and Parallel Logic Programming Systems. Pages 631–
645 of: Proc. of the twelfth international conference on logic programming. MIT Press.
Hermenegildo, M., Bueno, F., Cabeza, D., Carro, M., M. Garcı́a de la Banda, LópezGarcı́a, P., & Puebla, G. (1999a). The CIAO Multi-Dialect Compiler and System: An
Experimentation Workbench for Future (C)LP Systems. Pages 65–85 of: Parallelism
and Implementation of Logic and Constraint Logic Programming. Commack, NY, USA:
Nova Science.
Hermenegildo, M., Puebla, G., & Bueno, F. (1999b). Using Global Analysis, Partial Specifications, and an Extensible Assertion Language for Program Validation and Debugging.
Pages 161–192 of: Apt, K. R., Marek, V., Truszczynski, M., & Warren, D. S. (eds), The
Logic Programming Paradigm: a 25–Year Perspective. Springer-Verlag.
Jaffar, Joxan, & Lassez, Jean-Louis. (1987). Constraint Logic Programming. Pages 111–
119 of: ACM symposium on principles of programming languages. ACM.
Janson, S., & Haridi, S. (1991). Programming Paradigms of the Andorra Kernel Language.
Pages 167–183 of: 1991 international logic programming symposium. MIT Press.
Kowalski, R. A. (1974). Predicate Logic as a Programming Language. Pages 569–574 of:
Proceedings ifips.
Loke, S.W., & Davison, A. (1996). Logic programming with the World-Wide Web.
Pages 235–245 of: 7th. ACM Conference on Hypertext. ACM Press. Available from
http://www.cs.unc.edu/~barman/HT96/P14/lpwww.html.
M. Dincbas, H. Simonis, & Van Hentenryck, P. (1990). Solving Large Combinatorial
Problems in Logic Programming. Journal of logic programming, 8(1 & 2), 72–93.
Partners, The RADIOWEB Project. 1997 (July). RADIOWEB EP25562: Automatic Generation of Web Sites for the Radio Brodcasting Industry – Project Description / Technical Annex. Technical Report. RADIOWEB Project.
Ph. Bonnet, S. Bressan, L. Leth, & Thomsen, B. 1996 (September).
Towards ECLiPSe Agents on the INTERNET.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet2.html.
Smedbäck, G., Carro, M., & Hermenegildo, M. (1999). Interfacing Prolog and VRML and
its Application to Constraint Visualization. Pages 453–471 of: The practical application
of constraint technologies and logic programming. The Practical Application Company.
Smolka, G. 1994 (November). The Definition of Kernel Oz. DFKI Oz documentation
series. German Research Center for Artificial Intelligence (DFKI).
S.W. Loke, A. Davison, & Sterling, L. 1996 (September).
Lightweight Deductive Databases on the World-Wide Web.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet10.html.
Szeredi, Péter, Molnár, Katalin, & Scott, Rob. 1996 (September). Serving Multiple HTML Clients from a Prolog Application.
Proceedings of the 1st workshop on logic programming tools for internet applications.
Available from
http://clement.info.umoncton.ca/~lpnet/lpnet9.html.
Tarau, P. 1996 (April). Binprolog 5.00. Posting in comp.lang.prolog. Available from
http://clement.info.umoncton.ca/~tarau.
Van Hentenryck, P. (1989). Constraint Satisfaction in Logic Programming. MIT Press.
Warren, D.H.D., & Pereira, F. C. N. (1982). An Efficient, Easily Adaptable System For
Distributed WWW Programming using (Ciao-)Prolog and the PiLLoW lib 33
Interpreting Natural Language Queries. American journal of computational linguistics,
8(3-4), 110–122.
| 6cs.PL
|
Call-by-need, neededness and all that⋆
Delia Kesner1 , Alejandro Rı́os2, and Andrés Viso3
arXiv:1801.10519v1 [cs.LO] 31 Jan 2018
1
3
IRIF, CNRS and Univ. Paris-Diderot
2
Universidad de Buenos Aires
CONICET and Universidad de Buenos Aires
Abstract. We show that call-by-need is observationally equivalent to
weak-head needed reduction. The proof of this result uses a semantical
argument based on a (non-idempotent) intersection type system called
V. Interestingly, system V also allows to syntactically identify all the
weak-head needed redexes of a term.
1
Introduction
One of the fundamental notions underlying this paper is the one of needed reduction in λ-calculus, which is to be used here to understand (lazy) evaluation
of functional programs. Key notions are those of reducible and non-reducible
programs: the former are programs (represented by λ-terms) containing nonevaluated subprograms, called reducible expressions (redexes), whereas the latter can be seen as definitive results of computations, called normal forms. It
turns out that every reducible program contains a special kind of redex known
as needed or, in other words, every λ-term not in normal form contains a needed
redex. A redex r is said to be needed in a λ-term t if r has to be contracted
(i.e. evaluated) sooner or later when reducing t to normal form, or, informally
said, if there is no way of avoiding r to reach a normal form.
The needed strategy, which always contracts a needed redex, is normalising [8], i.e. if a term can be reduced (in any way) to a normal form, then contraction of needed redexes necessarily terminates. This is an excellent starting
point to design an evaluation strategy, but unfortunately, neededness of a redex is not decidable [8]. As a consequence, real implementations of functional
languages cannot be directly based on this notion.
Our goal is, however, to establish a clear connection between the semantical
notion of neededness and different implementations of lazy functional languages
(e.g. Miranda or Haskell). Such implementations are based on call-by-need calculi, pioneered by Wadsworth [19], and extensively studied e.g. in [3]. Indeed,
call-by-need calculi fill the gap between the well-known operational semantics of
the call-by-name λ-calculus and the actual implementations of lazy functional
languages. While call-by-name re-evaluates an argument each time it is used –an
operation which is quite expensive– call-by-need can be seen as a memoized version of call-by-name, where the value of an argument is stored the first time it is
⋆
This work was partially founded by LIA INFINIS.
evaluated for subsequent uses. For example, if t = ∆ (I I), where ∆ = λx.x x and
I = λz.z, then call-by-name duplicates the argument I I, while lazy languages
first reduce I I to the value I so that further uses of this argument do not need
to evaluate it again.
While the notion of needed reduction is defined with respect to (full strong)
normal forms, call-by-need calculi evaluate programs to special values called
weak-head normal forms, which are either abstractions or arbitrary applications
headed by a variable (i.e. terms of the form x t1 . . . tn where t1 . . . tn are arbitrary
terms). To overcome this shortfall, we first adapt the notion of needed redex to
terms that are not going to be fully reduced to normal forms but only to weakhead normal forms. Thus, informally, a redex r is weak-head needed in a term t
if r has to be contracted sooner or later when reducing t to a weak-head normal
form. The derived notion of strategy is called a weak-head needed strategy, which
always contracts a weak-head needed redex.
This paper introduces two independent results about weak-head neededness,
both obtained by means of (non-idempotent) intersection types [13,12] (a survey
can be found in [9]). We consider, in particular, typing system V [14] and show
that it allows to identify all the weak-head needed redexes of a weak-head normalising term. This is done by adapting the classical notion of principal type [16]
and proving that a redex in a weak-head normalising term t is weak-head needed
iff it is typed in a principally typed derivation for t in V.
Our second goal is to show observational equivalence between call-by-need
and weak-head needed reduction. Two terms are observationally equivalent when
all the empirically testable computations on them are identical. This means that
a term t can be evaluated to a weak-head normal form using the call-by-need
machinery if and only if the weak-head needed reduction normalises t.
By means of system V mentioned so far we use a technique to reason about
observational equivalence that is flexible, general and easy to verify or even
certify. Indeed, system V provides a semantic argument: first showing that a
term t is typable in system V iff it is normalising for the weak-head needed
strategy (t ∈ WN whnd ), then by resorting to some results in [14], showing that
system V is complete for call-by-name, i.e. a term t is typable in system V iff
t is normalising for call-by-name (t ∈ WN name ); and that t is normalising for
call-by-name iff t is normalising for call-by-need (t ∈ WN need ). Thus completing
the following chain of equivalences:
t ∈ WN whnd
t typable in V
t ∈ WN name
t ∈ WN need
This leads to the observational equivalence between call-by-need, call-by-name
and weak-head needed reduction.
Structure of the paper : Sec. 2 introduces preliminary concepts while Sec. 3
defines different notions of needed reduction. The type system V is studied in
Sec. 4. Sec 5 extends β-reduction to derivation trees. We show in Sec. 6 how system V identifies weak-head needed redexes, while Sec. 7 gives a characterisation
of normalisation for the weak-head needed reduction. Sec. 8 is devoted to define
call-by-need. Finally, Sec. 9 presents the observational equivalence result.
2
Preliminaries
This section introduces some standard definitions and notions concerning the
reduction strategies studied in this paper, that is, call-by-name, head and weakhead reduction, and neededness, this later notion being based on the theory of
residuals [7].
2.1
The Call-By-Name Lambda-Calculus
Given a countable infinite set X of variables x, y, z, . . . we consider the following
grammar:
(Terms) t, u ::= x ∈ X | t u | λx.t
(Values) v ::= λx.t
(Contexts) C ::= | C t | t C | λx.C
(Name contexts) E ::= | E t
The set of λ-terms is denoted by Ta . We use I, K and Ω to denote the terms
λx.x, λx.λy.x and (λx.x x) (λx.x x) respectively. We use Chti (resp. Ehti) for the
term obtained by replacing the hole of C (resp. E) by t. The sets of free and
bound variables of a term t, written respectively fv(t) and bv(t), are defined
as usual [7]. We work with the standard notion of α-conversion, i.e. renaming
of bound variables for abstractions; thus for example λx.x y =α λz.z y.
A term of the form (λx.t) u is called a β-redex (or just redex when β is
clear from the context) and λx is called the anchor of the redex. The onestep reduction relation →β (resp. →name ) is given by the closure by contexts
C (resp. E) of the rewriting rule (λx.t) u 7→β t {x /u }, where { / } denotes
the capture-free standard higher-order substitution. Thus, call-by-name forbids
reduction inside arguments and λ-abstractions, e.g. (λx.II ) (II) →β (λx.II) I
and (λx.II) (II) →β (λx.I) (II) but neither (λx.II) (II) →name (λx.II) I nor
(λx.II) (II) →name (λx.I) (II) holds. We write ։β (resp. ։name ) for the reflexivetransitive closure of →β (resp. →name ).
2.2
Head, Weak-Head and Leftmost Reductions
In order to introduce different notions of reduction, we start by formalising
the general mechanism of reduction which consists in contracting a redex at
some specific occurrence. Occurrences are finite words over the alphabet {0, 1}.
We use ǫ to denote the empty word and notation an for n ∈ N concatenations
of some letter a of the alphabet. The set of occurrences of a given term is
def
def
{ǫ} ∪ {0p | p ∈ oc(t)} ∪
{ǫ}; oc(t u) =
defined by induction as follows: oc(x) =
def
{1p | p ∈ oc(u)}; oc(λx.t) = {ǫ} ∪ {0p | p ∈ oc(t)}.
Given two occurrences p and q, we use the notation p ≤ q to mean that p is a
prefix of q, i.e. there is p′ such that pp′ = q. We denote by t|p the subterm of t
at occurrence p, defined as expected [4], thus for example ((λx.y) z)|00 = y. The
def
set of redex occurrences of t is defined by roc(t) =
{p ∈ oc(t) | t|p = (λx.s) u}.
We use the notation r : t →β t′ to mean that r ∈ roc(t) and t reduces to t′ by
contracting the redex at occurrence r, e.g. 000 : (λx.(λy.y) x x) z →β (λx.x x) z.
This notion is extended to reduction sequences as expected, and noted ρ : t ։β
t′ , where ρ is the list of all the redex occurrences contracted along the reduction
sequence. We use nil to denote the empty reduction sequence, so that nil : t ։β t
holds for every term t.
Any term t has exactly one of the following forms: λx1 .. . . λxn .y t1 . . . tm
or λx1 .. . . λxn .(λy.s) u t1 . . . tm with n, m ≥ 0. In the latter case we say that
(λy.s) u is the head redex of t, while in the former case there is no head redex.
Moreover, if n = 0, we say that (λy.s) u is the weak-head redex of t. In terms
of occurrences, the head redex of t is the minimal redex occurrence of the form
0n with n ≥ 0. In particular, if it satisfies that t|0k is not an abstraction for every
k ≤ n, it is the weak-head redex of t. A reduction sequence contracting at each
step the head redex (resp. weak-head redex) of the corresponding term is called
the head reduction (resp. weak-head reduction).
Given two redex occurrences r, r′ ∈ roc(t), we say that r is to-the-left of
′
r if the anchor of r is to the left of the anchor of r′ . Thus for example, the
redex occurrence 0 is to-the-left of 1 in the term (I x) (I y), and ǫ is to-the-left
of 00 in (λx.(I I)) z. Alternatively, the relation to-the-left can be understood as
a dictionary order between redex occurrences, i.e. r is to-the-left of r′ if either
r′ = rq with q 6= ǫ (i.e. r is a proper prefix of r′ ); or r = p0q and r′ = p1q′ (i.e. they
share a common prefix and r is on the left-hand side of an application while r′
is on the right-hand side). Notice that in any case this implies r′ 6≤ r. Since this
notion defines a total order on redexes, every term not in normal form has a
unique leftmost redex . The term t leftmost reduces to t′ if t reduces to t′
and the reduction step contracts the leftmost redex of t. For example, (I x) (I y)
leftmost reduces to x (I y) and (λx.(I I)) z leftmost reduces to I I. This notion
extends to reduction sequences as expected.
3
Towards neededness
Needed reduction is based on two fundamental notions: that of residual, which
describes how a given redex is traced all along a reduction sequence, and that
of normal form, which gives the form of the expected result of the reduction
sequence. This section extends the standard notion of needed reduction [8] to
those of head and weak-head needed reductions.
3.1
Residuals
Given a term t, p ∈ oc(t) and r ∈ roc(t), the descendants of p after r in t,
written p/r, is the set of occurrences defined as follows:
∅ if
{p} if
{rq} if
{rkq | s|k = x} if
p = r or p = r0
r 6≤ p
p = r00q
p = r1q with t|r = (λx.s) u
For instance, given t = (λx.(λy.x) x) z, then oc(t) = {ǫ, 0, 1, 00, 000, 001, 0000},
roc(t) = {ǫ, 00}, 00/00 = ∅, ǫ/00 = {ǫ}, 00/ǫ = {ǫ} and 1/ǫ = {1, 00}.
Notice that p/r ⊆ oc(t′ ) where r : t →β t′ . Furthermore, if p is the occurrence
of a redex in t (i.e. p ∈ roc(t)), then p/r ⊆ roc(t′ ), and each position in p/r is
called a residual of p after reducing r. This notion is extended to S
sets of redex
def
occurrences, indeed, the residuals of P after r in t are P/r =
p∈P p/r. In
particular ∅/r = ∅. Given ρ : t ։β t′ and P ⊆ roc(t), the residuals of P
def
def
after the sequence ρ are: P/nil =
P and P/rρ′ =
(P/r)/ρ′ .
Stability of the to-the-left relation makes use of the notion of residual:
Lemma 1. Given a term t, let l, r, s ∈ roc(t) such that l is to-the-left of r, s 6≤ l
and s : t →β t′ . Then, l ∈ roc(t′ ) and l is to-the-left of r′ for every r′ ∈ r/s.
Proof. First, notice that s 6≤ l implies l/s = {l}, then it is immediate to see that
l ∈ roc(t′ ). If s 6≤ r, then r/s = {r} and the result holds immediately. Otherwise,
s ≤ r implies s ≤ r′ for every r′ ∈ r/s by definition. From l is to-the-left of r, we
may distinguish two cases: either r = lq with q 6= ǫ; or r = p1q and l = p0q′ .
Both cases imply that l and r share a common prefix, say p′ (resp. l or p on each
case). From s 6≤ l and s ≤ r we know that p′ is a proper prefix of s. Thus, it is a
proper prefix of r′ too, i.e. either r′ = lq′′ with q′′ 6= ǫ; or r′ = p1q′′ . Hence, l is
to-the-left of r′ for every r′ ∈ r/s.
Notice that this result does not only implies that the leftmost redex is preserved by reduction of other redexes, but also that the residual of the leftmost
redex occurs in exactly the same occurrence as the original one.
Corollary 1. Given a term t, and l ∈ roc(t) the leftmost redex of t, if the
reduction ρ : t ։β t′ contracts neither l nor any of its residuals, then l ∈ roc(t′ )
is the leftmost redex of t′ .
Proof. By induction on the length of ρ using Lem. 1.
3.2
Notions of Normal Form
The expected result of evaluating a program is specified by means of some appropriate notion of normal form. Given any relation →R , a term t is said to
be in R-normal form (NF R ) iff there is no t′ such that t →R t′ . A term t
is R-normalising (WN R ) iff there exists u ∈ NF R such that t ։R u. Thus,
given an R-normalising term t, we can define the set of R-normal forms of t as
def
nfR (t) =
{t′ | t ։R t′ ∧ t′ ∈ NF R }.
In particular, it turns out that a term in weak-head β-normal form
(WHNF β ) is of the form x t1 . . . tn (n ≥ 0) or λx.t, where t, t1 , . . . , tn are arbitrary terms, i.e. it has no weak-head redex. The set of weak-head β-normal
def
forms of t is whnfβ (t) =
{t′ | t ։β t′ ∧ t′ ∈ WHNF β }.
Similarly, a term in head β-normal form (HNF β ) turns out to be of the
form λx1 .. . . λxn .x t1 . . . tm (n, m ≥ 0), i.e. it has no head redex. The set of head
β-normal forms of t is given by hnfβ (t) def
= {t′ | t ։β t′ ∧ t′ ∈ HNF β }.
Last, any term in β-normal form (NF β ) has the form λx1 .. . . λxn .x t1 . . . tm
(n, m ≥ 0) where t1 , . . . , tm are themselves in β-normal form. It is well-known
that the set nfβ (t) is a singleton, so we may use it either as a set or as its unique
element.
It is worth noticing that NF β ⊂ HNF β ⊂ WHNF β . Indeed, the inclusions
are strict, for instance λx.(λy.y) z is in weak-head but not in head β-normal
form, while x ((λy.y) x) z is in head but not in β-normal form.
3.3
Notions of Needed Reduction
The different notions of normal form considered in Sec. 3.2 suggest different
notions of needed reduction, besides the standard one in the literature [8]. Indeed,
consider r ∈ roc(t). We say that r is used in a reduction sequence ρ iff ρ reduces
r or some residual of r. Then:
1. r is needed in t if every reduction sequence from t to β-normal form uses r;
2. r is head needed in t if every reduction sequence from t to head β-normal
form uses r;
3. r is weak-head needed in t if every reduction sequence of t to weak-head
β-normal form uses r.
Notice in particular that nfβ (t) = ∅ (resp. hnfβ (t) = ∅ or whnfβ (t) = ∅)
implies every redex in t is needed (resp. head needed or weak-head needed).
A one-step reduction →β is needed (resp. head or weak-head needed ),
noted →nd (resp. →hnd or →whnd ), if the contracted redex is needed (resp. head
or weak-head needed). A reduction sequence ։β is needed (resp. head or
weak-head needed ), noted ։nd (resp. ։hnd or ։whnd ), if every reduction step
in the sequence is needed (resp. head or weak-head needed).
For instance, consider the reduction sequence:
(λy.λx.I x (I I r1 )) (I I) →nd (λy.λx.I xr2 I) (I I) →nd (λy.λx.x I) (I I)r →nd λx.x I
3
which is needed but not head needed, since redex r1 might not be contracted to
reach a head normal form:
(λy.λx.I xr2 (I I)) (I I) →hnd (λy.λx.x (I I)) (I I) →hnd λx.x (I I)
r3
Moreover, this second reduction sequence is head needed but not weak-head
needed since only redex r3 is needed to get a weak-head normal form:
(λy.λx.I x (I I)) (I I) →whnd λx.I x (I I)
r3
Notice that the following equalities hold: NF nd = NF β , NF hnd = HNF β
and NF whnd = WHNF β .
Leftmost redexes and reduction sequences are indeed needed:
Lemma 2. The leftmost redex in any term not in normal form (resp. head or
weak-head normal form) is needed (resp. head or weak-head needed).
Proof. Since the existing proof [8] does not extend to weak-head normal forms,
we give here an alternative argument which does not only cover the standard case
of needed reduction but also the new ones of head and weak-head reductions.
Let t be a term not in normal form (resp. head normal form or weak-head
normal form) and let us consider ρ : t ։β t′ such that t′ is a normal form (resp.
head normal form or weak-head normal form). Assume towards a contradiction
that the leftmost redex l of t is not used in ρ. By Cor. 1 the occurrence l is still
the leftmost redex of t′ . This leads to a contradiction with t′ being a normal
form (resp. head normal form or weak-head normal form), in particular because
the leftmost redex of a term not in head normal form (resp. weak-head normal
form) is necessarily the head redex (resp. weak-head redex).
Theorem 1. Let r ∈ roc(t) and ρ : t ։β t′ be the leftmost reduction (resp.
head reduction or weak-head reduction) starting with t such that t′ = nfβ (t)
(resp. t′ ∈ hnfβ (t) or t′ ∈ whnfβ (t)). Then, r is needed (resp. head or weak-head
needed) in t iff r is used in ρ.
Proof. ⇒) Immediate by definition of needed (resp. head or weak-head needed).
⇐) Let ρ = ρ′ r′ ρ′′ with r′ ∈ r/ρ′ . By hypothesis r′ is the leftmost redex of its
corresponding term. By Lem. 2, r′ is needed (resp. head or weak-head needed).
Notice that, given a redex s not needed in t, it follows from the definition that
no residual of s is needed either. Therefore, r′ needed (resp. head or weak-head
needed) implies r needed (resp. head or weak-head needed) as well.
Notice that the weak-head reduction is a prefix of the head reduction, which
is in turn a prefix of the leftmost reduction to normal form. As a consequence,
it is immediate to see that every weak-head needed redex is in particular head
needed, and every head needed redex is needed as well. For example, consider:
r2
r3
r4
(λy.λx.I x (I I )) (I I )
r1
where r3 is a needed redex but not head needed nor weak-head needed. However,
r2 is both needed and head needed, while r1 is the only weak-head needed redex
in the term, and r4 is not needed at all.
4
The Type System V
In this section we recall the (non-idempotent) intersection type system V [14]
–an extension of those in [13,12]– used here to characterise normalising terms
w.r.t. the weak-head strategy. More precisely, we show that t is typable in system
V if and only if t is normalising when only weak-head needed redexes are contracted. This characterisation is used in Sec. 9 to conclude that the weak-head
needed strategy is observationally equivalent to the call-by-need calculus (to be
introduced in Sec. 8).
Given a constant type a that denotes answers and a countable infinite set B
of base type variables α, β, γ, . . ., we define the following sets of types:
(Types) τ, σ ::= a | α ∈ B | M → τ
(Multiset types) M, N ::= {{τi }}i∈I
where I is a finite set
The empty multiset is denoted by {{}}. We remark that types are strict [17],
i.e. the right-hand sides of functional types are never multisets. Thus, the general
form of a type is M1 → . . . → Mn → τ with τ being the constant type or a base
type variable.
Typing contexts (or just contexts), written Γ, ∆, are functions from variables to multiset types, assigning the empty multiset to all but a finite set of
def
variables. The domain of Γ is given by dom(Γ ) =
{x | Γ (x) 6= {{}}}. The union
def
of contexts, written Γ + ∆, is defined by (Γ + ∆)(x) =
Γ (x) ⊔ ∆(x), where ⊔
denotes multiset union. An example is (x : {{σ}}, y : {{τ }}) + (x : {{σ}}, z : {{τ }}) =
(x : {{σ, σ}}, y : {{τ }}, z : {{τ }}). This notion is extended to several contexts as
expected, so that +i∈I Γi denotes a finite union of contexts (when I = ∅ the
notation is to be understood as the empty context). We write Γ \\ x for the
context (Γ \\ x)(x) = {{}} and (Γ \\ x)(y) = Γ (y) if y 6= x.
Type judgements have the form Γ ⊢ t : τ , where Γ is a typing context, t
is a term and τ is a type. The intersection type system V for the λ-calculus is
given in Fig. 1.
x : {{τ }} ⊢ x : τ
(ax)
Γ ⊢t:τ
Γ \\ x ⊢ λx.t : Γ (x) → τ
Γ ⊢ t : {{σi }}i∈I → τ
⊢ λx.t : a
(val)
( → i)
(∆i ⊢ u : σi )i∈I
Γ +i∈I ∆i ⊢ t u : τ
( → e)
Fig. 1. The non-idempotent intersection type system V.
The constant type a in rule (val) is used to type values. The axiom (ax)
is relevant (there is no weakening) and the rule ( → e) is multiplicative. Note
that the argument of an application is typed #(I) times by the premises of rule
( → e). A particular case is when I = ∅: the subterm u occurring in the typed
term t u turns out to be untyped.
A (type) derivation is a tree obtained by applying the (inductive) typing
rules of system V. The notation ⊲V Γ ⊢ t : τ means there is a derivation of
the judgement Γ ⊢ t : τ in system V. The term t is typable in system V, or
V-typable, iff t is the subject of some derivation, i.e. iff there are Γ and τ
such that ⊲V Γ ⊢ t : τ . We use the capital Greek letters Φ, Ψ, . . . to name type
derivations, by writing for example Φ ⊲V Γ ⊢ t : τ . For short, we usually denote
with Φt a derivation with subject t for some type and context. The size of
the derivation Φ, denoted by sz(Φ), is defined as the number of nodes of
the corresponding derivation tree. We write RULE(Φ) ∈ {(ax), ( → i), ( → e)} to
access the last rule applied in the derivation Φ. Likewise, PREM(Φ) is the multiset
of proper maximal subderivations of Φ. For instance, given
Φ=
Φt
(Φiu )i∈I
Γ ⊢ tu : τ
( → e)
we have RULE(Φ) = ( → e) and PREM(Φ) = {{Φt }} ⊔ {{Φiu | i ∈ I}}. We also use
functions CTXT(Φ), SUBJ(Φ) and TYPE(Φ) to access the context, subject and type
of the judgement in the root of the derivation tree respectively. For short, we
also use notation Φ(x) to denote the type associated to the variable x in the
def
typing environment of the conclusion of Φ (i.e. Φ(x) =
CTXT(Φ)(x)).
Intersection type systems can usually be seen as models [11], i.e. typing is
stable by convertibility: if t is typable and t =β t′ , then t′ is typable too. This
property splits in two different statements known as subject reduction and subject
expansion respectively, the first one giving stability of typing by reduction, the
second one by expansion. In the particular case of non-idempotent types, subject
reduction refines to weighted subject-reduction, stating that not only typability is
stable by reduction, but also that the size of type derivations is decreasing. Moreover, this decrease is strict when reduction is performed on special occurrences
of redexes, called typed occurrences. We now introduce all these concepts.
Given a type derivation Φ, the set TOC(Φ) of typed occurrences of Φ, which
is a subset of oc(SUBJ(Φ)), is defined by induction on the last rule of Φ.
– If RULE(Φ) ∈ {(ax), (val)}, then TOC(Φ) def
= {ǫ}.
– If RULE(Φ) = ( → i) with SUBJ(Φ) = λx.t and PREM(Φ) = {{Φt }}, then
def
TOC(Φ) =
{ǫ} ∪ {0p | p ∈ TOC(Φt )}.
– If RULE(Φ) = ( → e) with SUBJ(Φ) = t u andSPREM(Φ) = {{Φt }} ⊔ {{Φiu | i ∈ I}},
def
then TOC(Φ) =
{ǫ} ∪ {0p | p ∈ TOC(Φt )} ∪ ( i∈I {1p | p ∈ TOC(Φiu )}).
Remark that there are two kind of untyped occurrences, those inside untyped
arguments of applications, and those inside untyped bodies of abstractions. For
instance consider the following type derivations:
x : {{a}} ⊢ x : a
ΦK =
(ax)
x : {{a}} ⊢ λy.x : {{}} → a
⊢ K : {{a}} → {{}} → a
( → i)
ΦKI =
( → i)
ΦKIΩ =
ΦK
⊢I:a
(val)
⊢ K I : {{}} → a
⊢KIΩ :a
( → e)
( → e)
Then, TOC(ΦKIΩ ) = {ǫ, 0, 00, 01, 000, 0000} ⊆ oc(KIΩ).
Remark 1. The weak-head redex of a typed term is always a typed occurrence.
For convenience we introduce an alternative way to denote type derivations.
We refer to AX(x, τ ) as the result of applying (ax) with subject x and type τ :
def
AX(x, τ ) =
x : {{τ }} ⊢ x : τ
(ax)
We denote with VAL(x, t) the result of applying (val) abstracting x and term t:
def
VAL(x, t) =
⊢ λx.t : a
(val)
We refer to ABS(x, Φt ) as the result of applying ( → i) with premise Φt and
abstracting variable x:
def
ABS(x, Φt ) =
Φt
CTXT(Φt ) \\ x ⊢ λx.t : Φt (x) → TYPE(Φt )
( → i)
Likewise, we write APP(Φt , u, (Φiu )i∈I ) for the result of applying ( → e) with
premises Φt and (Φiu )i∈I , and argument u (the argument u is untyped when
I = ∅). Note that this application is valid provided that TYPE(Φt ) = {{σi }}i∈I →
τ and (TYPE(Φiu ) = σi )i∈I . Then:
def
APP(Φt , u, (Φiu )i∈I ) =
Φt
(Φiu )i∈I
CTXT(Φt ) +i∈I CTXT(Φiu ) ⊢ t u : τ
( → e)
Given Φ and p ∈ TOC(Φ), the multiset Φ|p of all the subderivations of Φ
at occurrence p is inductively defined as follows:
def
{{Φ}}.
– If p = ǫ, then Φ|p =
′
– If p = 0p , i.e. RULE(Φ) ∈ {( → i), ( → e)} with SUBJ(Φ) ∈ {λx.t, t u} and
Φt ∈ PREM(Φ). Then, Φ|p def
= Φt |p′ .
– If p = 1p′ , i.e. RULE(Φ) = (F
→ e) with SUBJ(Φ) = t u and PREM(Φ) = {{Φt }} ⊔
def
i
{{Φiu | i ∈ I}}. Then, Φ|p =
i∈I Φu |p′ (recall ⊔ denotes multiset union).
Given type derivations Φ, (Ψi )i∈I and a position p ∈ oc(SUBJ(Φ)), replacing
the subderivations of Φ at occurrence p by (Ψi )i∈I , written Φ[(Ψi )i∈I ]p , is
a type derivation inductively defined as follows, assuming #(Φ|p ) = #(I) and
SUBJ(Ψi ) = SUBJ(Ψj ) for every i, j ∈ I (we call s the unique subject of all these
derivations):
def
– If p = ǫ, then Φ[(Ψi )i∈I ]p =
Ψi0 with I = {i0 }.
′
– If p = 0p , either:
• RULE(Φ) = ( → i) with SUBJ(Φ) = λx.t and PREM(Φ) = {{Φt }}. Then,
def
Φ[(Ψi )i∈I ]p =
ABS(x, Φt [(Ψi )i∈I ]p′ ); or
• RULE(Φ) = ( → e) with SUBJ(Φ) = t u and PREM(Φ) = {{Φt }}⊔{{Φju | j ∈ J}}.
def
Then, Φ[(Ψi )i∈I ]p =
APP(Φt [(Ψi )i∈I ]p′ , u, (Φju )j∈J ).
– If p = 1p′ , i.e. RULE(Φ)
U = ( → e) with SUBJ(Φ) = t u, PREM(Φ) = {{Φt }} ⊔
{{Φju | j ∈ J}} and I = j∈J Ij . Then,
def
Φ[(Ψi )i∈I ]p =
APP(Φt , u[s]p′ , (Φju [(Ψi )i∈Ij ]p′ )j∈J )
where s is the unique subject of all the derivations Ψi and r′ [r]q denotes the
replacement of the subterm r′ |q by r in r′ (variable capture is allowed). Remark that the decomposition of I into the sets Ij (j ∈ J) is non-deterministic,
thus replacement turns out to be a non-deterministic operation.
We can now state the two main properties of system V, whose proofs can be
found in Sec. 7 of [9].
Theorem 2 (Weighted Subject Reduction). Let Φ ⊲V Γ ⊢ t : τ . If r : t →β
t′ , then there exists Φ′ s.t. Φ′ ⊲V Γ ⊢ t′ : τ . Moreover,
1. If r ∈ TOC(Φ), then sz(Φ) > sz(Φ′ ).
2. If r ∈
/ TOC(Φ), then sz(Φ) = sz(Φ′ ).
Theorem 3 (Subject Expansion). Let Φ′ ⊲V Γ ⊢ t′ : τ . If t →β t′ , then there
exists Φ s.t. Φ ⊲V Γ ⊢ t : τ .
Note that weighted subject reduction implies that reduction of typed redex
occurrences turns out to be normalising.
5
Substitution and Reduction on Derivations
In order to relate typed redex occurrences of convertible terms, we now extend
the notion of β-reduction to derivation trees, by making use of a natural and
basic concept of typed substitution. In contrast to substitution and β-reduction
on terms, these operations are now both non-deterministic on derivation trees
(see [18] for discussions and examples). Given a variable x and type derivations
Φt and (Φiu )i∈I , the typed substitution of x by (Φiu )i∈I in Φt , written
Φt x (Φiu )i∈I by making an abuse of notation, is a type derivation inductively defined on Φt , only if Φt (x) = {{TYPE(Φiu )}}i∈I :
– If Φt = AX(y, τ ), then
Φt x (Φiu )i∈I
def
=
(
Φiu0
Φt
if y = x where I = {i0 }
if y =
6 x
– If Φt = VAL(y, t′ ), then
Φt x (Φiu )i∈I
= VAL(y, t′ {x /u })
def
if there is no capture of the variable y.
– If Φt = ABS(y, Φt′ ), then
def
Φt x (Φiu )i∈I =
ABS(y, Φt′ x (Φiu )i∈I )
if there is no capture of the variable y. U
– If Φt = APP(Φr , s, (Φjs )j∈J ) with I = I ′ ⊎( j∈J Ij ), where Φr (x) = {{TYPE(Φiu )}}i∈I ′
and (Φs (x) = {{TYPE(Φiu )}}i∈Ij )j∈J , then
Φt x (Φiu )i∈I
def
= APP(Φr x (Φiu )i∈I ′ , s {x /u }, (Φjs x (Φiu )i∈Ij )j∈J )
Remark that the decomposition of I into I ′ and the sets Ij (j ∈ J) is
non-deterministic, thus substitution of derivation trees turns out to be a
non-deterministic operation.
Intuitively, the typed substitution replaces typed occurrences of x in Φt by
a corresponding derivation Φiu matching the same type, where such a matching
is chosen in a non-deterministic way. Moreover, it also substitutes all untyped
occurrences of x by u, where this untyped operation is completely deterministic.
Thus, for example, consider the following substitution, where ΦKI is defined in
Sec. 4:
(ax)
ΦKI
x : {{{{}} → a}} ⊢ x : {{}} → a
( → e)
( → e) {x /ΦKI } = ⊢ (KI) (KI) : a
x : {{{{}} → a}} ⊢ x x : a
The following lemma relates the typed occurrences of the trees composing a
substitution and those of the substituted tree itself:
Lemma 3. Let Φt and (Φiu )i∈I be derivations such that Φt x (Φiu )i∈I is defined, and p ∈ oc(t). Then,
1. p ∈ TOC(Φt ) iff p ∈ TOC(Φt x (Φiu )i∈I ).
2. q ∈ TOC(Φku ) for some
k ∈ I iff there exists p ∈ TOC(Φt ) such that t|p = x
and pq ∈ TOC(Φt x (Φiu )i∈I ).
Proof. By induction on Φt .
Based on the previous notion of substitutions on derivations, we are now able
to introduce (non-deterministic) reduction on derivation trees. The reduction
relation →β on derivation trees is then defined by first considering the following
basic rewriting rules.
1. For typed β-redexes:
Φt ⊲V Γ ; x : {{σi }}i∈I ⊢ t : τ
Γ ⊢ λx.t : {{σi }}i∈I → τ
(Φiu ⊲V ∆i ⊢ u : σi )i∈I
Γ +i∈I ∆i ⊢ (λx.t) u : τ
7→β Φt x (Φiu )i∈I
2. For β-redexes in untyped occurrences, with u →β u′ :
Γ ⊢ t : {{}} → τ
Γ ⊢ tu : τ
7→ν
Γ ⊢ t : {{}} → τ
Γ ⊢ t u′ : τ
⊢ λx.u : a
7→ξ
⊢ λx.u′ : a
As in the case of the λ-calculus, where reduction is closed under usual term
contexts, we need to close the previous relation under derivation tree contexts.
However, a one-step reduction on a given subterm causes many one-step reductions in the corresponding derivation tree (recall Φ|p is defined to be a multiset).
Then, informally, given a redex occurrence r of t, a type derivation Φ of t, and
the multiset of minimal subderivations of Φ containing r, written M , we apply
the reduction rules 7→β,ν,ξ to all the elements of M , thus obtaining a multiset
M ′ , and we recompose the type derivation of the reduct of t.
To formalise this idea, given a type derivation Φ and an occurrence p ∈
oc(SUBJ(Φ)), we define the maximal typed prefix of p in Φ, written mtpΦ (p),
as the unique prefix of p satisfying
mtpΦ (p) ∈ TOC(Φ) ∧ ∀q ∈ TOC(Φ).(q ≤ p =⇒ q ≤ mtpΦ (p))
Notice that the multiset of subderivations of Φ at position mtpΦ (p) (i.e. Φ|mtpΦ (p) )
corresponds to the multiset of minimal subderivations of Φ containing p. For
instance, consider the type derivation ΦKIΩ presented in Sec. 4:
ΦK
ΦKIΩ =
ΦI =
⊢I:a
⊢ K I : {{}} → a
⊢KIΩ :a
(val)
( → e)
( → e)
where TOC(ΦKIΩ ) = {ǫ, 0, 00, 01, 000, 0000}. Indeed, mtpΦKIΩ (010) = 01 and the
minimal subderivation of ΦKIΩ containing this occurrence is ΦKIΩ |01 = {{ΦI }}.
Also, mtpΦKIΩ (1p) = ǫ where ΦKIΩ |ǫ = {{ΦKIΩ }}.
Then, given terms t and t′ , and type derivations Φ and Φ′ of t and t′ respectively, we say that Φ reduces to Φ′ (written Φ →β Φ′ ) iff there exists r ∈ roc(t)
such that
Φ|mtpΦ (r) β,ν,ξ M and Φ′ = Φ[M ]mtpΦ (r)
where β,ν,ξ denotes the lifting to multisets of the basic rewriting rules introduced above by applying the same rule to all the elements of the multiset.
This gives the reduction relation →β on trees. A reduction sequence on
derivation trees contracting only redexes in typed positions is dubbed a typed
reduction sequence.
Note that typed reductions are normalising by Thm. 2, yielding a special
kind of derivation. Indeed, given a type derivation Φ ⊲V Γ ⊢ t : τ , we say that
Φ is normal iff TOC(Φ) ∩ roc(t) = ∅. Reduction on trees induces reduction on
terms: when ρ : Φ ։β Φ′ , then SUBJ(Φ) ։β SUBJ(Φ′ ). By abuse of notation we
may denote both sequences with the same letter ρ.
6
Weak-Head Neededness and Typed Occurrences
This section presents one of our main results. It establishes a connection between
weak-head needed redexes and typed redex occurrences. More precisely, we first
show in Sec. 6.1 that every weak-head needed redex occurrence turns out to
be a typed occurrence, whatever its type derivation is. The converse does not
however hold. But, we show in Sec. 6.2 that any typed occurrence in a special
kind of typed derivation (that we call principal) corresponds to a weak-head
needed redex occurrence. We start with a technical lemma.
Lemma 4. Let r : Φt →β Φt′ and p ∈ oc(t) such that p 6= r and p 6= r0. Then,
p ∈ TOC(Φt ) iff there exists p′ ∈ p/r such that p′ ∈ TOC(Φt′ ).
Proof. If p = ǫ the result holds since ǫ/r = {ǫ} and ǫ ∈ TOC(Φ) for every possible
Φ by definition. Then, assume p 6= ǫ. We proceed by induction on r.
– r = ǫ. Then, r ≤ p and r ∈ roc(t) ∩ TOC(Φt ). Moreover, t = t|r = (λx.t1 ) t2
and
Φt1
Φt =
(Φit2 )i∈I
Γ ′ ⊢ λx.t1 : {{σi }}i∈I → τ
Γ ⊢ (λx.t1 ) t2 : τ
→β Φt1 x (Φit2 )i∈I
= Φt′
with t′ = t1 {x /t2 }. Then, there are two possibilities for p:
1. p = 00p′ . Then, p/ǫ = {p′ }. By definition, p ∈ TOC(Φt ) iff p′ ∈ TOC(Φt1 ).
Moreover, by Lem. 3 (1), p′ ∈ TOC(Φt1 ) iff p′ ∈ TOC(Φt′ ) (notice that
p′ ∈ oc(t1 )). Thus, we conclude.
2. p = 1p′′ . Then, p/ǫ = {qp′′ | t1 |q = x}. By definition p ∈ TOC(Φt ) iff
p′′ ∈ TOC(Φkt2 ) for some k ∈ I. By Lem. 3 (2), p′′ ∈ TOC(Φkt2 ) iff there
exists q ∈ TOC(Φt1 ) such that t1 |q = x and p′ = qp′′ ∈ TOC(Φt′ ). Thus,
we conclude.
– r = 0r′ . Then, we analyse the form of t:
• t = t1 t2 . Then, t′ = t′1 t2 with r′ : t1 →β t′1 . Moreover,
Φt =
Φt1
(Φit2 )i∈I
Γ ⊢ t1 t2 : τ
→β
Φt′1
(Φit2 )i∈I
Γ ⊢ t′1 t2 : τ
= Φt′
with Φt1 →β Φt′1 .
Now we have two possibilities for p:
∗ p = 0q. By definition, p ∈ TOC(Φt ) iff q ∈ TOC(Φt1 ). Since p 6= r
implies q 6= r′ , by inductive hypothesis, q ∈ TOC(Φt1 ) iff there exists
q′ ∈ q/r′ such that q′ ∈ TOC(Φt′1 ). Thus, by definition once again,
p ∈ TOC(Φt ) iff there exists p′ ∈ p/r such that p′ ∈ TOC(Φt′ ) (i.e. p′ =
0q′ ).
∗ p = 1q. Then, r 6≤ p and p/r = {p} (i.e. p′ = p). Thus, p ∈ TOC(Φt )
iff q ∈ TOC(Φkt2 ) for some k ∈ I iff p′ ∈ TOC(Φt′ ) by definition.
• t = λx.t1 . Then, t′ = λx.t′1 with r′ : t1 →β t′1 . If RULE(Φt ) = (val), the
result is immediate since TOC(Φt ) = {ǫ} = TOC(Φt′ ). If not,
Φt =
Φt1
Γ ⊢ λx.t1 : τ
→β
Φt′1
Γ ⊢ λx.t′1 : τ
= Φt′
with Φt1 →β Φt′1 . Then, p = 0q and, by definition, p ∈ TOC(Φt1 ) iff
q ∈ TOC(Φt1 ). By inductive hypothesis, q ∈ TOC(Φt1 ) iff there exists
q′ ∈ q/r′ such that q′ ∈ TOC(Φt′1 ). Thus, we conclude as in the previous
case.
– r = 1r′ . This case is symmetric to the one presented above. We have t = t1 t2
and t′ = t1 t′2 with r′ : t2 →β t′2 , and
Φt =
Φt1
(Φit2 )i∈I
Γ ⊢ t1 t2 : τ
→β
Φt1
(Φit′2 )i∈I
Γ ⊢ t1 t′2 : τ
= Φt′
with (Φit2 →β Φit′ )i∈I . Then, if p = 0q (i.e. r 6≤ p) we may conclude by
2
definition with p/r = {p}. Otherwise, p = 1q and p ∈ TOC(Φt ) iff q ∈ TOC(Φkt′ )
2
for some k ∈ I. Thus, the result follows by definition from the inductive
hypothesis.
6.1
Weak-Head Needed Redexes are Typed
In order to show that every weak-head needed redex occurrence corresponds
to a typed occurrence in some type derivation we start by proving that typed
occurrences do not come from untyped ones.
Lemma 5. Let ρ : Φt ։β Φt′ and p ∈ oc(t). If there exists p′ ∈ p/ρ such that
p′ ∈ TOC(Φt′ ), then p ∈ TOC(Φt ).
Proof. Straightforward induction on ρ using Lem. 4.
Theorem 4. Let r be a weak-head needed redex in t. Let Φ be a type derivation
of t. Then, r ∈ TOC(Φ).
Proof. By Thm. 1, r is used in the weak-head reduction from t to t′ ∈ WHNF β .
By Rem. 1, the weak-head reduction contracts only typed redexes. Thus, r or
some of its residuals is a typed occurrence in its corresponding derivation tree.
Finally, we conclude by Lem. 5, r ∈ TOC(Φ).
6.2
Principally Typed Redexes are Weak-Head Needed
As mentioned before, the converse of Thm. 4 does not hold: there are some typed
occurrences that do not correspond to any weak-head needed redex occurrence.
This can be illustrated in the following examples (recall ΦKIΩ defined in Sec. 4):
ΦKIΩ
⊢ λy.KIΩ : {{}} → a
( → i)
y : {{{{a}} → a}} ⊢ y : {{a}} → a
(ax)
y : {{{{a}} → a}} ⊢ y (KIΩ) : a
ΦKIΩ
( → e)
Indeed, the occurrence 0 (resp 1) in the term λy.KIΩ (resp. y (KIΩ)) is
typed but not weak-head needed, since both terms are already in weak-head
normal form. Fortunately, typing relates to redex occurrences if we restrict type
derivations to principal ones: given a term t in weak-head β-normal form, the
derivation Φ ⊲V Γ ⊢ t : τ is normal principally typed if:
n times
z
}|
{
– t = x t1 . . . tn (n ≥ 0), and Γ = {x : {{{{}} → . . . → {{}} → τ }}} and τ is a
type variable α (i.e. none of the ti are typed), or
– t = λx.t′ , and Γ = ∅ and τ = a.
Given a weak-head normalising term t such that Φt ⊲V Γ ⊢ t : τ , we say that
Φt is principally typed if Φt ։β Φt′ for some t′ ∈ whnfβ (t) implies Φt′ is
normal principally typed.
Note in particular that the previous definition does not depend on the chosen
weak-head normal form t′ : suppose t′′ ∈ whnfβ (t) is another weak-head normal
form of t, then t′ and t′′ are convertible terms by the Church-Rosser property [7]
so that t′ can be normal principally typed iff t′′ can, by Thm. 2 and 3.
Lemma 6. Let Φt be a type derivation with subject t and r ∈ roc(t) ∩ TOC(Φt ).
Let ρ : Φt ։β Φt′ such that Φt′ is normal. Then, r is used in ρ.
Proof. Straightforward induction on ρ using Lem. 4.
The notions of leftmost and weak-head needed reductions on (untyped) terms
naturally extends to typed reductions on tree derivations. We thus have:
Lemma 7. Let t be a weak-head normalising term and Φt be principally typed.
Then, a leftmost typed reduction sequence starting at Φt is weak-head needed.
Proof. By induction on the leftmost typed sequence (called ρ). If ρ is empty
the result is immediate. If not, we show that t has a typed weak-head needed
redex (which is leftmost by definition) and conclude by inductive hypothesis.
Indeed, assume t ∈ WHNF β . By definition Φt is normal principally typed and
thus it has no typed redexes. This contradicts ρ being non-empty. Hence, t has
a weak-head redex r (i.e. t ∈
/ WHNF β ). Moreover, r is both typed (by Rem. 1)
and weak-head needed (by Lem. 2). Thus, we conclude.
Theorem 5. Let t be a weak-head normalising term, Φt be principally typed and
r ∈ roc(t) ∩ TOC(Φt ). Then, r is a weak-head needed redex in t.
Proof. Let ρ : Φt ։β Φt′ be the leftmost typed reduction sequence where Φt′
is normal. Note that Φt′ exists by definition of principally typed. By Lem. 7, ρ
is a weak-head needed reduction sequence. Moreover, by Lem. 6, r is used in ρ.
Hence, r is a weak-head needed redex in t.
As a direct consequence of Thm. 4 and 5, given a weak-head normalising term
t, the typed redex occurrences in its principally typed derivation (which always
exists) correspond to its weak-head needed redexes. Hence, system V allows to
identify all the weak-head needed redexes of a weak-head normalising term.
7
Characterising Weak-Head Needed Normalisation
This section presents one of the main pieces contributing to our observational
equivalence result. Indeed, we relate typing with weak-head neededness by showing that any typable term in system V is normalising for weak-head needed
reduction. This characterisation highlights the power of intersection types. We
start by a technical lemma.
Lemma 8. Let Φ ⊲V Γ ⊢ t : τ . Then, Φ normal implies t ∈ WHNF β .
Proof. By induction on Φ analysing the last rule applied.
–
–
–
–
(ax). Then t = x ∈ WHNF β .
(val). Then t = λx.t′ ∈ WHNF β .
( → i). Then t = λx.t′ ∈ WHNF β .
( → e). Then t = r u with Γ = Γ ′ +i∈I ∆i , Φ′ ⊲V Γ ′ ⊢ r : {{σi }}i∈I → τ and
(∆i ⊢ u : σi )i∈I for some I. From Φ normal we have Φ′ normal too. Thus, by
inductive hypothesis, r ∈ WHNF β . Moreover, r 6= λy.r′ since ǫ ∈ TOC(Φ).
Then, r = x r1 . . . rn and we conclude with t = x r1 . . . rn u ∈ WHNF β .
Let ρ : t1 ։β tn . We say that ρ is a left-to-right reduction sequence iff
for every i < n if ri : ti →β ti+1 and li is to the left of ri then, for every j > i
such that rj : tj →β tj+1 we have that rj ∈
/ {li }/ρij where ρij : ti ։β tj is the
corresponding subsequence of ρ. In other words, for every j and every i < j,
rj is not a residual of a redex to the left of ri (relative to the given reduction
subsequence from ti to tj ) [7].
Left-to-right reductions define in particular standard strategies, which give
canonical ways to construct reduction sequences from one term to another:
Theorem 6 ([7]). If t ։β t′ , there exists a left-to-right reduction from t to t′ .
Theorem 7. Let t ∈ Ta . Then, Φ ⊲V Γ ⊢ t : τ iff t ∈ WN whnd .
Proof. ⇒) By Thm. 2 we know that the strategy reducing only typed redex
occurrences is normalising, i.e. there exist t′ and Φ′ such that t ։β t′ , Φ′ ⊲V
Γ ⊢ t′ : τ and Φ′ normal. Then, by Lem. 8, t′ ∈ WHNF β . By Thm. 6, there
exists a left-to-right reduction ρ : t ։β t′ . Let us write
ρ : t = t1 ։β tn ։β t′
such that t1 , . . . , tn−1 ∈
/ WHNF β and tn ∈ WHNF β .
We claim that all reduction steps in t1 ։β tn are leftmost. Assume towards
a contradiction that there exists k < n such that r : tk →β tk+1 and r is not the
leftmost redex of tk (written lk ). Since ρ is a left-to-right reduction, no residual
of lk is contracted after the k-th step. Thus, there is a reduction sequence from
tk ∈
/ WHNF β to tn ∈ WHNF β such that lk is not used in it. This leads to a
contradiction with lk being weak-head needed in tk by Lem. 2.
As a consequence, there is a leftmost reduction sequence t ։β tn . Moreover,
by Lem. 2, t ։whnd tn ∈ WHNF β = NF whnd . Thus, t ∈ WN whnd .
⇐) Consider the reduction ρ : t ։whnd t′ with t′ ∈ whnfβ (t). Let Φ′ ⊲V
Γ ⊢ t′ : τ be the normal principally typed derivation for t′ as defined in Sec. 6.2.
Finally, we conclude by induction in ρ using Thm. 3, Φ ⊲V Γ ⊢ t : τ .
8
The Call-by-Need Lambda-Calculus
This section describes the syntax and the operational semantics of the call-byneed lambda-calculus introduced in [1]. It is more concise than previous specifications of call-by-need [3,2,15,10], but it is operationally equivalent to them [6],
so that our results could also be presented by using alternative specifications.
Given a countable infinite set X of variables x, y, z, . . . we define different
syntactic categories for terms, values, list contexts, answers and need contexts:
(Terms) t, u ::= x ∈ X | t u | λx.t | t[x\u]
(Values) v ::= λx.t
(List contexts) L ::= | L[x\t]
(Answers) a ::= Lhλy.ti
(Need contexts) M, N ::= | N t | N[x\t] | Nhhxii[x\M]
We denote the set of terms by Te . Terms of the form t[x\u] are closures,
and [x\u] is called an explicit substitution (ES). The set of Te -terms without
ES is the set of terms of the λ-calculus, i.e. Ta . The notions of free and bound
def
variables are defined as expected, in particular, fv(t[x\u]) =
fv(t) \ {x} ∪ fv(u),
def
def
fv(λx.t) = fv(t)\{x}, bv(t[x\u]) = bv(t)∪{x}∪bv(u) and bv(λx.t) def
= bv(t)∪{x}.
We extend the standard notion of α-conversion to ES, as expected.
We use the special notation Nhhuii or Lhhuii when the free variables of u are not
captured by the context, i.e. there are no abstractions or explicit substitutions
in the context that binds the free variables of u. Thus for example, given N =
( x)[x\z], we have (y x)[x\z] = Nhyi = Nhhyii, but (x x)[x\z] = Nhxi cannot be
written as Nhhxii. Notice the use of this special notation in the last case of needed
contexts, an example of such case being (x y)[y\t][x\].
The call-by-need calculus, introduced in [1], is given by the set of terms
Te and the reduction relation →need , the union of →dB and →lsv , which are,
respectively, the closure by need contexts of the following rewriting rules:
Lhλx.ti u 7→dB Lht[x\u]i
Nhhxii[x\Lhvi] 7→lsv LhNhhvii[x\v]i
These rules avoid capture of free variables. An example of need-reduction sequence is the following, where the redex of each step is underlined for clearness:
(λx1 .I (x1 I)) (λy.I y)
x2 [x2 \x1 I][x1 \λy.I y]
x2 [x2 \(I x3 )[x3 \I]][x1 \λy.I y]
x2 [x2 \x4 [x4 \I][x3 \I]][x1 \λy.I y]
I[x2 \I][x4 \I][x3 \I][x1 \λy.I y]
→dB
→lsv
→dB
→lsv
(I (x1 I))[x1 \λy.I y]
x2 [x2 \(λx3 .I x3 ) I][x1 \λy.I y]
x2 [x2 \x4 [x4 \x3 ][x3 \I]][x1 \λy.I y]
x2 [x2 \I[x4 \I][x3 \I]][x1 \λy.I y]
→dB
→dB
→lsv
→lsv
As for call-by-name, reduction preserves free variables, i.e. t →need t′ implies
fv(t) ⊇ fv(t′ ). Notice that call-by-need reduction is also weak, so that answers
are not need-reducible.
9
Observational Equivalence
The results in Sec. 7 are used here to prove soundness and completeness of callby-need w.r.t weak-head neededness, our second main result. More precisely, a
call-by-need interpreter stops in a value if and only if the weak-head needed
reduction stops in a value. This means that call-by-need and call-by-name are
observationally equivalent.
Formally, given a reduction relation R on a term language T , and an associated notion of context for T , we define t to be observationally equivalent to
u, written t ∼
=R u, iff Chti ∈ WN R ⇔ Chui ∈ WN R for every context C. In order
to show our final result we resort to the following theorem:
Theorem 8 ([14]).
1. Let t ∈ Ta . Then, Φ ⊲V Γ ⊢ t : τ iff t ∈ WN name .
2. For all terms t and u in Ta , t ∼
=name u iff t ∼
=need u.
These observations allows us to conclude:
Theorem 9. For all terms t and u in Ta , t ∼
=need u.
=whnd u iff t ∼
∼
Proof. By Thm. 8:2 it is sufficient to show t =whnd u iff t ∼
=name u. The proof
proceeds as follows:
t∼
iff
(definition)
=name u
Chti ∈ WN name ⇔ Chui ∈ WN name iff
(Thm. 8:1)
Chti typable in V ⇔ Chui typable in V iff
(Thm. 7)
Chti ∈ WN whnd ⇔ Chui ∈ WN whnd iff
(definition)
t∼
=whnd u
10
Conclusion
We establish a clear connection between the semantical standard notion of neededness and the syntactical concept of call-by-need. The use of non-idempotent
types –a powerful technique being able to characterise different operational
properties– provides a simple and natural tool to show observational equivalence between these two notions. We refer the reader to [5] for other proof techniques (not based on intersection types) used to connect semantical notions of
neededness with syntactical notions of lazy evaluation.
An interesting (and not difficult) extension of our result in Sec. 6 is that
call-by-need reduction (defined on λ-terms with explicit substitutions) contracts
only dB weak-head needed redexes, for an appropriate (and very natural) notion
of weak-head needed redex for λ-terms with explicit substitutions. A technical
tool to obtain such a result would be the type system A [14], a straightforward
adaptation of system V to call-by-need syntax.
Given the recent formulation of strong call-by-need [6] describing a deterministic call-by-need strategy to normal form (instead of weak-head normal form), it
would be natural to extend our technique to obtain an observational equivalence
result between the standard notion of needed reduction (to full normal forms)
and the strong call-by-need strategy. This remains as future work.
References
1. Beniamino Accattoli, Pablo Barenbaum, and Damiano Mazza. Distilling abstract
machines. In Johan Jeuring and Manuel M. T. Chakravarty, editors, Proceedings
of the 19th ACM SIGPLAN international conference on Functional programming,
Gothenburg, Sweden, September 1-3, 2014, pages 363–376. ACM, 2014.
2. Zena M. Ariola and Matthias Felleisen. The call-by-need lambda calculus. J.
Funct. Program., 7(3):265–301, 1997.
3. Zena M. Ariola, Matthias Felleisen, John Maraist, Martin Odersky, and Philip
Wadler. The call-by-need lambda calculus. In Ron K. Cytron and Peter Lee,
editors, Conference Record of POPL’95: 22nd ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, San Francisco, California, USA,
January 23-25, 1995, pages 233–246. ACM Press, 1995.
4. Franz Baader and Tobias Nipkow. Term rewriting and all that. Cambridge University Press, 1998.
5. Thibaut Balabonski. La pleine paresse, une certaine optimalité. Ph.D. Thesis,
Université Paris-Diderot, 2012.
6. Thibaut Balabonski, Pablo Barenbaum, Eduardo Bonelli, and Delia Kesner. Foundations of strong call by need. PACMPL, 1(ICFP):20:1–20:29, 2017.
7. Hendrik P. Barendregt. The Lambda Calculus Its Syntax and Semantics, volume
103. North Holland, revised edition, 1984.
8. Hendrik P. Barendregt, Richard Kennaway, Jan Willem Klop, and M. Ronan Sleep.
Needed reduction and spine strategies for the lambda calculus. Inf. Comput.,
75(3):191–231, 1987.
9. Antonio Bucciarelli, Delia Kesner, and Daniel Ventura. Non-idempotent intersection types for the lambda-calculus. Logic Journal of the IGPL, 25(4):431–464,
2017.
10. Stephen Chang and Matthias Felleisen. The call-by-need lambda calculus, revisited. In Helmut Seidl, editor, Programming Languages and Systems - 21st European Symposium on Programming, ESOP 2012, Held as Part of the European Joint
Conferences on Theory and Practice of Software, ETAPS 2012, Tallinn, Estonia,
March 24 - April 1, 2012. Proceedings, volume 7211 of Lecture Notes in Computer
Science, pages 128–147. Springer, 2012.
11. Mario Coppo and Mariangiola Dezani-Ciancaglini. An extension of the basic functionality theory for the λ-calculus. Notre Dame Journal of Formal Logic, 21(4):685–
693, 1980.
12. Daniel de Carvalho. Sémantiques de la logique linéaire et temps de calcul. PhD
thesis, Université Aix-Marseille II, 2007.
13. Philippa Gardner. Discovering needed reductions using type theory. In Masami
Hagiya and John C. Mitchell, editors, Theoretical Aspects of Computer Software,
International Conference TACS ’94, Sendai, Japan, April 19-22, 1994, Proceedings, volume 789 of Lecture Notes in Computer Science, pages 555–574. Springer,
1994.
14. Delia Kesner. Reasoning about call-by-need by means of types. In Bart Jacobs
and Christof Löding, editors, Foundations of Software Science and Computation
Structures - 19th International Conference, FOSSACS 2016, Held as Part of the
European Joint Conferences on Theory and Practice of Software, ETAPS 2016,
Eindhoven, The Netherlands, April 2-8, 2016, Proceedings, volume 9634 of Lecture
Notes in Computer Science, pages 424–441. Springer, 2016.
15. John Maraist, Martin Odersky, and Philip Wadler. The call-by-need lambda calculus. J. Funct. Program., 8(3):275–317, 1998.
16. Simona Ronchi Della Rocca. Principal type scheme and unification for intersection
type discipline. Theor. Comput. Sci., 59:181–209, 1988.
17. Steffen van Bakel. Complete restrictions of the intersection type discipline. Theor.
Comput. Sci., 102(1):135–163, 1992.
18. Pierre Vial. Non-Idempotent Intersection Types, Beyond Lambda-Calculus. PhD
thesis, Université Paris-Diderot, 2017.
19. Christopher P. Wadsworth. Semantics and Pragmatics of the Lambda Calculus.
PhD thesis, Oxford University, 1971.
| 6cs.PL
|
A Differential Complex for CAT(0) Cubical Spaces
arXiv:1610.05069v1 [math.KT] 17 Oct 2016
J. Brodzki∗, E. Guentner†, N. Higson‡
1. Introduction
In the 1980’s Pierre Julg and Alain Valette [JV83, JV84], and also Tadeusz Pytlik
and Ryszard Szwarc [PS86], constructed and studied a certain Fredholm operator
associated to a simplicial tree. The operator can be defined in at least two ways:
from a combinatorial flow on the tree, similar to the flows in Forman’s discrete Morse
theory [For98], or from the theory of unitary operator-valued cocycles [Pim87, Val90].
There are applications of the theory surrounding the operator to C ∗ -algebra K-theory
[JV83, JV84], to the theory of completely bounded representations of groups that act
on trees [PS86], and to the Selberg principle in the representation theory of p-adic
groups [JV86, JV87].
The crucial property of the Fredholm operator introduced by Julg and Valette is that
it is the initial operator in a continuous family of Fredholm operators parametrized by
a closed interval. The applications all emerge from the properties of the family in the
circumstance where a group G acts properly on the underlying tree, in which case all
the operators in the family act on Hilbert spaces that carry unitary representations
of G. Roughly speaking, the family connects the regular representation of G to the
trivial representation within an index-theoretic context.
This calls to mind Kazhdan’s property T [Kaž67, BdlHV08], or rather the negation
of property T, as well as Haagerup’s property [Haa79, CCJ+ 01], which is a strong
negation of property T. Groups that act on trees are known to have the Haagerup
property (this is essentially due to Haagerup himself), and the Julg-Valette, PytlikSzwarc construction is perhaps best viewed as a geometric incarnation of this fact.
An immediate consequence is the K-theoretic amenability of any group that acts
property on a tree [Cun83, JV84], which is another strong negation of property T.
The main aim of this paper is to extend the constructions of Julg and Valette, and
Pytlik and Szwarc, to CAT(0) cubical spaces (a one-dimensional CAT(0) cubical
J. B. was supported in part by EPSRC grants EP/I016945/1 and EP/N014189/1.
E. G. was supported in part by a grant from the Simons Foundation (#245398).
‡
N. H. was supported in part by NSF grant DMS-1101382.
∗
†
1
space is the same thing as a simplicial tree). A secondary aim is to illustrate the
utility of the extended construction by developing an application to operator Ktheory and giving a new proof of K-amenability for groups that act properly on
bounded-geometry CAT(0)-cubical spaces. But we expect there will be other uses
for our constructions, beyond operator K-theory.
We shall associate to each bounded geometry CAT(0) cubical space not a Fredholm
operator but a differential complex with finite-dimensional cohomology. The construction is rather more challenging for general CAT(0) cubical spaces than it is for
trees. Whereas for trees there is a more or less canonical notion of flow towards a
distinguished base vertex in the tree, in higher dimensions this is not so, and for
example a vertex is typically connected to a given base vertex by a large number of
edge-paths. In addition, the need to consider higher-dimensional cubes, and the need
to impose the condition d2 = 0, oblige us to carefully consider orientations of cubes in
a way that is quite unnecessary for trees.
More interesting still is the problem of defining the final complex in the one-parameter
family of complexes that we aim to construct. To solve it we shall rely on the theory of
hyperplanes in CAT(0) cubical spaces [NR98a]. In the case of a tree the hyperplanes
are simply the midpoints of edges, but in general they have a nontrivial geometry all
of their own; in fact they are CAT(0) cubical spaces in their own rights.
We shall also introduce and study a related notion of parallelism among the cubes in
a CAT(0) cubical space. In a tree, any two vertices are parallel, while no two distinct
edges are parallel, but in higher dimensions parallelism is more subtle. For instance
in a finite tree the number of vertices is precisely one plus the number of edges (this
simple geometric fact is in fact an essential part of the Julg-Valette, Pytlik-Szwarc
construction). But the proof of the following generalization to higher dimensions is
quite a bit more involved.
Proposition. If X is finite CAT(0) cubical space, then the number of vertices of X
is equal to the number of parallelism classes of cubes of all dimensions.
We expect that parallelism and the other aspects of our constructions, will be of
interest and value elsewhere in the theory of CAT(0) cube complexes.
One last challenge comes in passing from CAT(0) cubical geometry to Fredholm
complexes and operator K-theory. There are two standard paradigms in operator
K-theory, of bounded cycles and unbounded cycles, but the geometry we are faced
with here forces us to consider a hybrid of the two. However once this is done we
shall arrive at our application:
Theorem. If a second countable and locally compact group G admits a proper action
on a bounded geometry CAT(0) cube complex, then G is K-amenable.
2
Groups that act properly on CAT(0) cube complexes are known to have the Haagerup
property [NR98b], and they were proved to be K-amenable in [HK01, Theorem 9.4].
The advantage of the present approach is that the constructions in the proof are all
tied to the finite-dimensional cube complex itself, whereas in [HK01] the authors rely
on an auxilliary action of the group on an infinite-dimensional Euclidean space that
is rather hard to understand directly.
Here is a brief outline of the paper. After reviewing the concept of hyperplane in
Section 2 we shall study orientations and define our Julg-Valette complex in Section
3. We shall introduce parallelism in Section 4 and define the final complex (we
shall call it the Pytlik-Szwarc complex) in Section 5. The one-parameter family of
complexes connecting the two will be constructed in stages, in Sections 6, 7 and 8,
and the application to operator K-theory will be the subject of Sections 9 and 10.
2. Cubes and Hyperplanes
We shall begin by fixing some basic notation concerning the cubes and hyperplanes
in a CAT(0) cube complex. We shall follow the exposition of Niblo and Reeves in
[NR98a], with some adaptations.
Throughout the paper X will denote a CAT(0) cube complex as in [NR98a, Section
2.2]. Though not everywhere necessary, we shall assume throughout that X is finitedimensional, and that it has bounded geometry in the sense that the number of cubes
intersecting any one cube C is uniformly bounded as C varies over all cubes.
Every q-cube contains exactly 2q codimension-one faces. Each such face is disjoint
from precisely one other, which we shall call the opposite face.
We shall use the standard terms vertex and edge for 0-dimensional and 1-dimensional
cubes.
The concept of a midplane of a cube is introduced in [NR98a, Section 2.3]. If we
identify a q-cube with the standard cube [− 21 , 12 ]q in Rq , then the midplanes are
precisely the intersections of the cube with the coordinate hyperplanes in Rq (thus
the midplanes of a cube C are in particular closed subsets of C). A q-cube contains
precisely q midplanes (and in particular a vertex contains no midplanes)
Niblo and Reeves describe an equivalence relation on the set of all midplanes in a
cube complex: two midplanes are (hyperplane) equivalent if they can be arranged as
the first and last members of a finite sequence of midplanes for which the intersection
of any two consecutive midplanes is again a midplane.
2.1 Definition. (See [NR98a, Definition 2.5].) A hyperplane in X is the union of the
set of all midplanes in an equivalence class of midplanes. A hyperplane cuts a cube
if it contains a midplane of that cube. When a hyperplane cuts an edge, we say that
3
the edge crosses the hyperplane. See Figure 1.
H
Figure 1: The hyperplane H is the union of three midplanes.
2.2 Examples. If X is a tree, then the hyperplanes are precisely the midpoints of
edges. If X is the plane, divided into cubes by the integer coordinate lines, then
hyperplanes are the half-integer coordinate lines.
Hyperplanes are particularly relevant in the context of CAT(0) cube complexes (such
as the previous two examples) for the following reason:
2.3 Lemma. (See [Sag95, Theorem 4.10] or [NR98a, Lemma 2.7].) If X is a CAT(0)
cube complex, then every hyperplane is a totally geodesic subspace of X that separates
X into two connected components.
The components of the complement of a hyperplane are the two half-spaces associated
to the hyperplane. The half-spaces are open, totally geodesic subsets of X. Moreover
the union of all cubes contained in a given half-space is a CAT(0) cube complex in
its own right, and a totally geodesic subcomplex of X.
Later on, it will be helpful to approximate an infinite complex by finite complexes,
as follows.
2.4 Lemma. Every bounded geometry CAT(0) cube complex X is an increasing union
of finite, totally geodesic CAT(0) subcomplexes Xn whose hyperplanes are precisely
the nonempty intersections of the hyperplanes in X with Xn .
Proof. Fix a base point in X and an integer n > 0. Form the set of all hyperplanes
whose distance to the base point is n or greater, and then form the intersection of all
4
the half-spaces for these hyperplanes that contain the base point . Denote by Xn the
union of all cubes that are included in this intersection; it is a totally geodesic subset
of X and so a CAT(0) cube complex. Moreover the intersection of any hyperplane in
X with Xn is connected. The union of all the Xn as n → ∞ is X and, since the set
of hyperplanes of distance less than n to the base point is finite, each Xn is a finite
subcomplex of X.
2.5 Definition. A hyperplane and a vertex are adjacent if the vertex is included in
an edge that crosses the hyperplane.
2.6 Lemma. If k hyperplanes in a CAT(0) cube complex intersect pairwise, then all
k intersect within some k-cube.
Proof. See [Sag95, Theorem 4.14].
2.7 Lemma. Assume that k distinct hyperplanes in a CAT(0) cube complex have a
non-empty intersection. If are they are all adjacent to a vertex, then they intersect
in a k-cube that contains that vertex.
Proof. See [NR98a, Lemma 2.14 and Proposition 2.15].
2.8 Lemma. If two hyperplanes H and K in a CAT(0) cube complex X are disjoint,
then one of the half-spaces of H is contained in one of the half-spaces of K.
Proof. See [GH10, Lemma 2.10].
3. The Julg-Valette Complex
Let X be a bounded geometry CAT(0) cube complex of dimension n. The aim of
this section is to define a differential complex
d
d
d
d
C[X 0 ] Ð→ C[X 1 ] Ð→ ⋯ Ð→ C[X n−1 ] Ð→ C[X n ]
which generalizes the complex introduced by Julg and Valette in the case of a tree
[JV83, JV84]. To motivate the subsequent discussion we recall their construction.
Let T be a tree with vertex set T 0 and edge set T 1 . Fix a base vertex P0 . The
Julg-Valette differential
d ∶ C[T 0 ] Ð→ C[T 1 ]
is defined by mapping a vertex P ≠ P0 to the first edge E on the unique geodesic path
from P to P0 ; P0 itself is mapped to zero. There is an adjoint differential
δ ∶ C[T 1 ] Ð→ C[T 0 ]
5
that maps each edge to its furtherest vertex from P0 . The composite dδ is the identity
on C[T 1 ], whereas 1−δd is the natural rank-one projection onto the subspace of C[T 0 ]
spanned by the base vertex. It follows easily that the cohomology of the Julg-Valette
complex is C in degree zero and 0 otherwise.
For the higher-dimensional construction we shall need a concept of orientation for the
cubes in X, and we begin there.
3.1 Definition. A presentation of a cube consists of a vertex in the cube, together
with a linear ordering of the hyperplanes that cut the cube. Two presentations are
equivalent if the edge-path distance between the two vertices has the same parity
as the permutation between the two orderings. An orientation of a cube of positive
dimension is a choice of equivalence class of presentations; an orientation of a vertex
is a choice of sign + or −.
3.2 Remark. Every cube has precisely two orientations, and if C is an oriented cube
we shall write C ∗ for the same underlying unoriented cube equiped with the opposite
orientation.
3.3 Definition. The space C[X q ] of oriented q-cochains on X is the vector space
comprising the finitely-supported, anti-symmetric, complex-valued functions on the
set of oriented q-cubes X q . Here, a function f is anti-symmetric if f (C) + f (C ∗ ) = 0
for every oriented cube C.
3.4 Remark. The space C[X q ] is a subspace of the vector space of of all finitely
supported functions on X q , which we shall call the full space of q-cochains. The
formula
f ∗ (C) = f (C ∗ )
defines an involution on the full space of q-cochains. We shall write C for both the
Dirac function at the oriented q-cube C and for the cube itself; in this way C belongs
to the full space of q-cochains. We shall write ⟨C⟩ for the oriented q-cochain
⟨C⟩ = C − C ∗ ∈ C[X q ],
which is the difference of the Dirac functions at C and C ∗ (the two possible meanings
of the symbol C ∗ agree).
Next, we introduce some geometric ideas that will allow us to define the Julg-Valette
differential in higher dimensions. The first is the following generalization of the notion
of adjacency introduced in Definition 2.5.
3.5 Definition. A q-cube C is adjacent to a hyperplane H if it is disjoint from H
and if there exists a (q+1)-cube containing C as a codimension-one face that is cut
by H.
6
3.6 Lemma. A q-cube C is adjacent to a hyperplane H if and only if it is not cut by
H and all of its vertices are adjacent to H.
Proof. Clearly, if the cube C is adjacent to H then so are all of its vertices. For the
converse, assume that all of the vertices of C are adjacent to H. By Lemma 2.6 it
suffices to show that every hyperplane K that cuts C must also cross H. For this,
let P and Q be vertices of C separated only by K, and denote by P op and Qop the
vertices separated from P and Q only by H, respectively. These four vertices belong
to the four distinct half-space intersections associated with the hyperplanes H and
K, so that by Lemma 2.8 these hyperplanes intersect.
We shall now fix a base vertex P0 in the complex X.
3.7 Definition. Let H be a hyperplane in X. Define an operator
H ∧ ∶ C[X q ] Ð→ C[X q+1 ]
as follows. Let C be an oriented q-cube in X.
(a) We put H ∧ C = 0 if C is not adjacent to H.
(b) In addition, we put H ∧ C = 0 if C is adjacent to H, but C lies in the same
H-half-space as the base point P0 .
(c) If C is adjacent to H, and is separated by H from the base point, then we define
H ∧ C to be the unique cube containing C as a codimension-one face that is cut
by H.
As for the orientations in (c), if C has positive dimension and is oriented by the vertex
P , and by the listing on hyperplanes H1 , . . . , Hq , then we orient H ∧ C by the vertex
that is separated from P by the hyperplane H alone, and by the listing of hyperplanes
H, H1 , . . . , Hq . If C is a vertex with orientation + then H ∧ C is oriented as above; if
C has orientation − then H ∧ C receives the opposite orientation.
3.8 Remark. The linear operator H ∧ of the previous definition is initially defined
on the full space of q-cochains by specifying its values on the oriented q-cubes C,
which form a basis of this space. We omit the elementary check that for an oriented
q-cube C we have
(3.1)
H ∧ C ∗ = (H ∧ C)∗ ,
which allows us to restrict H ∧ to an operator on the spaces of oriented q-cochains.
We shall employ similar conventions consistently throughout, so that all linear operators will be defined initially on the full space of cochains and then restricted to the
space of oriented cochains. Some formulas will hold only for the restricted operators
and we shall point these few instances out.
7
3.9 Definition. The Julg-Valette differential is the linear map
d∶ C[X q ] Ð→ C[X q+1 ]
given by the formula
d C = ∑ H ∧ C,
H
where the sum is taken over all hyperplanes in X. Note that only finitely many terms
in this sum are nonzero.
3.10 Example. In the case of a tree, if P is any vertex distinct from the base point
P0 , then H ∧ P is the first edge on the geodesic edge-path from P to P0 and our
operator d agrees the one defined by Julg and Valette. Once a base point is chosen
every edge (in any CAT(0) cube complex) is canonically oriented by selecting the
vertex nearest to the base point; vertices are canonically oriented by the orientation
+. Thus, because the original construction of Julg and Valette involves only vertices
and edges and assumes a base point, orientations do not appear explicitly.
3.11 Lemma. If H1 and H2 are any two hyperplanes, and if C is any oriented cube,
then
(a) H1 ∧ H2 ∧ C is nonzero if and only if H1 and H2 are distinct, they are both
adjacent to C, and they both separate C from P0 .
(b) H1 ∧ H2 ∧ C = (H2 ∧ H1 ∧ C)∗ .
3.12 Remark. Here, H1 ∧ H2 ∧ C means H1 ∧ (H2 ∧ C), and so on.
Proof. Item (a) follows from Lemmas 2.7 and 2.8. To prove (b), note first that
as a result of (a) the left hand side is nonzero if and only if the right hand side is
nonzero. In this case, both have the same underlying unoriented (q + 2)-cube, namely
the unique cube containing C as a codimension-two face and cut by H1 and H2 . As
for orientation, suppose C is presented by the ordering K1 , . . . , Kq and the vertex
P . The cube H1 ∧ H2 ∧ C is then presented by the ordering H1 , H2 , K1 , . . . , Kq and
the vertex Q, the vertex immediately opposite both H1 and H2 from P ; the cube
H2 ∧ H1 ∧ C is presented by the ordering H2 , H1 , K1 , . . . , Kq and the same vertex. The
same argument applies when C is a vertex with the orientation +, and the remaining
case follows from this and the identity (3.1).
3.13 Lemma. The Julg-Valette differential d, regarded as an operator on the space
of oriented cochains, satisfies d2 = 0.
Proof. Let C be any q-cube, so that
d2 ⟨C⟩ = ∑ H1 ∧ H2 ∧ ⟨C⟩,
H1 ,H2
8
As a consequence of Lemma 3.11 we have H1 ∧ H2 ∧ ⟨ C ⟩ + H2 ∧ H1 ∧ ⟨ C ⟩ = 0, and the
sum vanishes. It is important here that we work on C[X q ] and not on the larger full
space of q-cochains, where the result is not true. See Remarks 3.4 and 3.8.
3.14 Definition. Let H be a hyperplane and let q ≥ 1. Define an operator
H ⌟ ∶ C[X q ] Ð→ C[X q−1 ]
as follows. Let C be an oriented q-cube in X.
(a) If H does not cut C, then H ⌟ C = 0.
(b) If H does cut C then we define H ⌟ C to be the codimension-one face of C that
lies entirely in the half-space of H that is separated from the base point by H.
As for orientations in (b), if C is presented by the ordered list H, H1 , . . . , Hq−1 and the
vertex P , and P is not separated from the base point by H, then H ⌟ C is presented
by the ordered list H1 , . . . , Hq−1 and the vertex separated from P by H alone. If C
is an edge presented by the vertex P not separated from the base point by H then
H ⌟ C = P op , the vertex of C opposite to P , with the orientation +; if C is presented
by the vertex P and P is separated from the base point by H then H ⌟ C = P with
the orientation −.
3.15 Remark. For convenience we shall define the operator H ⌟
vertices.
to be zero on
3.16 Example. Let us again consider a tree T with a selected base vertex P0 . If E
is any edge then H ⌟ E is zero unless H cuts E. In this case H ⌟ E = P , where P is
the vertex of E which is farthest away from P0 ; we choose the orientation − if E was
oriented by the vertex P , and the orientation + otherwise.
3.17 Definition. Let q ≥ 0. Define an operator
δ∶ C[X q+1 ] Ð→ C[X q ]
by
δ C = ∑ H ⌟ C.
H
3.18 Definition. The oriented q-cubes are a vector space basis for the full space of
q-cochains. We equip this space with an inner product by√declaring this to be an
orthogonal basis and each oriented q-cube to have length 1/ 2. The subspace C[X q ]
of oriented q-cochains inherits an inner product in which
⎧
1,
if C1 = C2
⎪
⎪
⎪
⎪
⟨⟨C1 ⟩, ⟨C2 ⟩⟩ = ⎨−1, if C1 = C2∗
⎪
⎪
⎪
⎪
otherwise.
⎩0,
9
Thus, selecting for each unoriented q-cube one of its possible orientations gives a
collection of oriented q-cubes for which the corresponding ⟨C⟩ form an orthonormal
basis of the space of oriented q-cochains; this basis is canonical up to signs coming
from the relations −⟨C⟩ = ⟨C ∗ ⟩.
3.19 Proposition. The operators d and δ of Definitions 3.9 and 3.17 are formally
adjoint and bounded with respect to the inner products in Definition 3.18.
Proof. The fact that the operators are bounded follows from our assumption that
the complex X has bounded geometry. The fact that they are adjoint follows from
the following assertion: for a hyperplane H, an oriented q-cube C and an oriented
(q + 1)-cube D we have that H ∧ C = D if and only if H ⌟ D = C. See Definitions 3.7
and 3.14.
To conclude the section, let us compute the cohomology of the Julg-Valette complex.
We form the Julg-Valette Laplacian
(3.2)
∆ = (d + δ)2 = dδ + δd,
where all operators are defined on the space of oriented cochains (and not on the
larger full space of cochains), where we have available the formula d2 = 0 and hence
also δ 2 = 0.
3.20 Proposition. If C is an oriented q-cube then
∆⟨C⟩ = (q + p(C)) ⟨C⟩,
where p(C) is the number of hyperplanes that are adjacent to C and separate C from
P0 . In particular, the ⟨C⟩ form an orthonormal basis of eigenvectors of ∆, which is
invertible on the orthogonal complement of ⟨P0 ⟩ (and so also on the space of oriented
q-cochains for q > 0).
Proof. We shall show that each oriented q-cube C is an eigenvector of dδ + δd acting
on the full space of q-cochains, with eigenvalue as in the statement. If P is a vertex,
then dδP = 0 for dimension reasons while δdP = p(P )P , irrespective of the choice of
orientation. In higher dimensions, if q ≥ 1 and C is an oriented q-cube, then
δdC = ∑ H1 ⌟ H2 ∧ C
H1 ,H2
and similarly
dδC = ∑ H1 ∧ H2 ⌟ C.
H1 ,H2
10
Adding these, and separating the sum into terms where H1 = H2 and terms where
H1 ≠ H2 we obtain
(3.3) (dδ +δd)C = ∑ (H ⌟ H ∧ C + H ∧ H ⌟ C)+ ∑ (H1 ⌟ H2 ∧ C + H2 ∧ H1 ⌟ C) .
H
H1 ≠H2
It follows from Lemma 3.21 below that (each term of) the second sum in (3.3) is zero.
To understand the first sum in (3.3), observe that if H is any hyperplane and C is
any oriented cube, then
⎧
⎪
⎪C, if H cuts C
H ∧ (H ⌟ C) = ⎨
⎪
⎪
⎩0, otherwise,
and also
⎧
⎪
⎪C, if C is adjacent to H and is separated by H from P0
H ⌟ (H ∧ C) = ⎨
⎪
⎪
⎩0, otherwise.
The proposition now follows.
3.21 Lemma. If H1 and H2 are distinct hyperplanes, then
H1 ⌟ H2 ∧ C = H2 ∧ H1 ⌟ C ∗
for every oriented cube C.
Proof. If C is a vertex then both sides of the formula are zero. More generally, if C is
a q-cube and one of the following two conditions fails then both sides of the formula
are zero:
(a) H2 is adjacent to C, and separates it from the base point;
(b) H1 cuts C and crosses H2 .
Assume both of these conditions, and suppose that C may be presented by the listing
of hyperplanes H1 , K2 , . . . , Kq and vertex P , and that H1 separates P from the base
point ; if C is not an edge this is always possible. We shall leave the exceptional case
in which C is an edge oriented by its vertex closest to the base point to the reader.
Now, let Q be the vertex of C separated from P by H1 alone, and let P op and Qop
be the vertices directly opposite H2 from P and Q, respectively. The cube H ∧ C is
presented by the listing H2 , H1 , K2 , . . . , Kq together with the vertex P op , hence also
by the listing H1 , H2 , K2 , . . . , Kq and the vertex Qop . It follows that H1 ⌟ H2 ∧ C is
presented by the listing H2 , K2 , . . . , Kq and the vertex P op . As for the right hand
side, C ∗ is presented by the same listing as C but with the vertex Q, so that H1 ⌟ C ∗
is presented by the listing K2 , . . . , Kq and the vertex P . It follows that H2 ∧ H1 ⌟ C ∗
is presented by the listing H2 , K2 , . . . , Kq and the vertex P op , as required.
11
3.22 Corollary. The cohomology of the Julg-Valette complex is C in degree zero and
0 otherwise.
Proof. In degree q = 0 the kernel of d is one dimensional and is spanned by ⟨P0 ⟩. In
degrees q ≥ 1 proceed as follows. From d2 = 0 it follows that d∆ = dδd = ∆d, so that
also d∆−1 = ∆−1 d. Now the calculation
f = ∆∆−1 f = (dδ + δd)∆−1 f = d(δ∆−1 )f
shows that an oriented q-cocycle f is also an oriented q-coboundary.
We conclude the section with a slight generalization that will be needed later.
3.23 Definition. A weight function for X is a positive-real-valued function w on the
set of hyperplanes in X. The weighted Julg-Valette differential is the linear map
dw ∶ C[X q ] Ð→ C[X q+1 ]
given by the formula
dw C = ∑ w(H) H ∧ C.
H
In addition the adjoint operator
δw ∶ C[X q+1 ] Ð→ C[X q ]
is defined by
δw C = ∑ w(H) H ⌟ C.
H
3.24 Remark. We are mainly interested in the following examples, or small variations on them:
(a) w(H) ≡ 1.
(b) w(H) = the minimal edge-path distance to the base point P0 from a vertex
adjacent to H.
The calculations in this section are easily repeated in the weighted context: the operators dw and δw are formally adjoint, although unbounded in the case of an unbounded
weight function as, for example, in (b); both are differentials when restricted to the
spaces of oriented cochains; and the cohomology of either complex is C in degree zero
and 0 otherwise. We record here the formula for the weighted Julg-Valette Laplacian.
Compare Proposition 3.20.
3.25 Proposition. If C is an oriented q-cube then
∆w ⟨C⟩ = (qw (C) + pw (C))⟨C⟩,
where qw (C) is the sum of the squares of the weights of the hyperplanes that cut C and
pw (C) is the sum of the squares of the weights of the hyperplanes that are adjacent to
C and separate C from the base vertex.
12
4. Parallelism Classes of Cubes
The remaining aspects of our generalization of the Julg-Valette and Pytlik-Szwarc
theory to CAT(0) cube complexes all rest on the following geometric concept:
4.1 Definition. Two cubes D1 and D2 in a CAT(0) cube complex X are parallel if
they have the same dimension, and if every hyperplane that cuts D1 also cuts D2 .
H
Figure 2: The darker edges form a parallelism class determined by the
hyperplane H, see Definition 4.1.
Every parallelism class of q-cubes in X is determined by, and determines, a set of q
pairwise intersecting hyperplanes, namely the hyperplanes that cut all the cubes in
the parallelism class. Call these the determining hyperplanes for the parallelism class.
4.2 Proposition. The intersection of the determining hyperplanes associated to a
parallelism class of q-cubes carries the structure of a CAT(0) cube complex in which
the p-cubes are the intersections of this space with the (p+q)-cubes in X that are cut
by every determining hyperplane.
Proof. The case when q = 0 is the assertion that X itself is a CAT(0) cube complex.
The case when q = 1 is the assertion that a hyperplane in CAT(0) cube complex X is
itself a CAT(0) cube complex in the manner described above, and this is proved by
Sageev in [Sag95, Thm. 4.11].
For the general result, we proceed inductively as follows. Suppose given k distinct
hyperplanes K1 , . . . , Kk in X. The intersection Z = K2 ∩ ⋅ ⋅ ⋅ ∩ Kk is then a CAT(0)
cube complex as described in the statement, and the result will follow from another
application of [Sag95, Thm. 4.11] once we verify that K1 ∩Z is a hyperplane in Z. Now
the cubes, and so also the midplanes of Z are exactly the non-empty intersections
of the cubes and midplanes of X with Z. So, we must show that if two midplanes
13
belonging to the hyperplane K1 of X intersect Z non-trivially then their intersections
are hyperplane equivalent in Z. But this follows from the fact that Z is a totally
geodesic subspace of X.
4.3 Proposition. Let X be a CAT(0) cube complex and let P be a vertex in X.
In each parallelism class of q-cubes there is a unique cube that is closest to P , as
measured by the distance from closest point in the cube to P in the edge-path metric.
Before beginning the proof, we recall that the edge-path distance between two vertices
is equal to the number of hyperplanes separating the vertices; see for example [Sag95,
Theorem 4.13]. In addition, let us make note of the following simple fact:
4.4 Lemma. A hyperplane that separates two vertices of distinct cubes in the same
parallelism class must intersect every determining hyperplane.
Proof. This is obvious if the hyperplane is one of the determining hyperplanes. Otherwise, the hyperplane must in fact separate two cubes in the parallelism class, and so
it must separate two midplanes from each determining hyperplane. Since hyperplanes
are connected the result follows.
Proof of Proposition 4.3. Choose a vertex R from among the cubes in the parallelism
class such that
(4.1)
d(P, R) ≤ d(P, S)
for every other such vertex S. We shall prove the addition formula
(4.2)
d(P, S) = d(P, R) + d(R, S),
and this will certainly prove the uniqueness of R.
The addition formula (4.2) is a consequence of the following hyperplane property of
any R satisfying (4.1): every hyperplane that separates P from R is parallel to (that
is, it does not intersect) at least one determining hyperplane. Indeed, it follows from
Lemma 4.4 and the hyperplane property that no hyperplane can separate R from both
P and S, so that (4.2) follows from the characterization of the edge path distance
given above.
It remains to prove the hyperplane property for any R satisfying (4.1). For this we
shall use the notion of normal cube path from [NR98a, Section 3]. There exists a
normal cube path from R to P with vertices
R = R1 , . . . , Rl = P.
14
This means that every pair of consecutive Ri are diagonally opposite a cube, called a
normal cube, all of whose hyperplanes separate R from P , and every such separating
hyperplane cuts exactly one normal cube. It also means that every hyperplane K
separating Ri from Ri+1 is parallel to at least one of the hyperplanes H separating
Ri−1 from Ri (so each normal cube is, in turn, as large as possible). Note that the
hyperplane K is contained completely in the half-space of H that contains P .
No hyperplane H separating R = R1 from R2 can intersect every determining hyperplane, for if it did, then it would follow from Lemma 2.7 that H and the determining
hyperplanes would intersect in a (q+1)-cube having R as a vertex. The vertex S
separated from R by H alone would then belong to a cube in the parallelism class,
and would be strictly closer to P than R.
Consider the second normal cube, with opposite vertices R3 and R2 . Any hyperplane
K separating R3 from R2 is parallel to some hyperplane H separating R2 from R1 ,
and this is in turn parallel to some determining hyperplane. But K is contained
completely in the half-space of H that contains P , while the determining hyperplane
is contained completely in the half-space of H that contains R. So K does not meet
this determining hyperplane.
Continuing in this fashion with successive normal cubes, we find that every hyperplane that separates P from R is indeed parallel to some determining hyperplane, as
required.
We can now verify the formula mentioned in the introduction:
4.5 Proposition. If X is finite CAT(0) cubical space, then the number of vertices of
X is equal to the number of parallelism classes of cubes of all dimensions.
Proof. Fix a base vertex P and associate to each vertex Q the first cube in the normal
cube path from Q to P . This correspondence induces a bijection from vertices to
parallelism classes of cubes.
Indeed it follows from the hyperplane property that if C is the nearest cube to P
within its parallelism class, and if Q is the vertex of C furthest from P , then C is
the first cube in the normal cube path from Q to P . So our map is surjective. On
the other hand it follows from the addition formula that if C is not nearest to P
within its equivalence class, and if Q is the vertex of C furthest from P , then any
hyperplane that separates Q from the nearest cube also separates Q from P . Choosing
a hyperplane that is adjacent to Q but not does not cut C, we find that C is not the
first cube in the normal cube path from Q to P , and our map is injective.
4.6 Proposition. Let X be a CAT(0) cube complex and let P and Q be vertices in X
that are separated by a single hyperplane H. The nearest q-cubes to P and Q within
15
a parallelism class are either the same, or are opposite faces, separated by H, of a
(q+1)-cube that is cut by H.
Proof. Denote by R and S the nearest vertices to P and Q, respectively, among the
vertices of cubes in the equivalence class, and suppose that a hyperplane K separates
R from S. Then it must separate P from S by the addition formula (4.2) applied
to the nearest point R, and also separate Q from R, by the addition formula applied
to the nearest point S. So it must separate P from Q, and hence must be H. So
either there is no hyperplane separating R from S, in which case of course R = S
and the nearest cubes to P and Q are the same, or R is opposite S across H. If
H is a determining hyperplane, then R and S are vertices of the same q-cube in
the parallelism class; if H is not a determining hyperplane, then R and S belong to
q-cubes that are opposite to one another across H, as required.
5. The Pytlik-Szwarc Complex
As described in the introduction, our ultimate goal involves deforming the JulgValette complex into what we call the Pytlik-Szwarc complex , a complex with the same
cohomology but which is equivariant in the case of a group acting on the CAT(0) cube
complex. In this short section we describe the (algebraic) Pytlik-Szwarc complex.
As motivation for what follows we consider how to compare orientations on parallel
cubes. The key observation is that a vertex in a q-cube is uniquely determined by
its position relative to the cutting hyperplanes K1 , . . . , Kq . Thus, there is a natural
isometry between (the vertex sets of) any two parallel q-cubes. We shall say that
parallel q-cubes of positive dimension are compatibly oriented if their orientations are
presented by vertices P1 and P2 which correspond under this isometry and a common
listing of the cutting hyperplanes K1 , . . . , Kq ; vertices are compatibly oriented if they
are oriented by the same choice of sign.
We shall now generalize these considerations to pairs comprising a cube and one of
its faces.
5.1 Definition. A cube pair is a pair (C, D) in which C is a cube containing D as
a face. Two cube pairs (C1 , D1 ) and (C2 , D2 ) are parallel if the cubes C1 and C2 are
parallel, and the cubes D1 and D2 are parallel too. When D is a q-cube, and C is a
(p + q)-cube, we shall call (C, D) a (p, q)-cube pair, always keeping in mind that in
this notation p is the codimension of D in C.
We may describe the parallelism class of a (p, q)-cube pair (C, D) by grouping the
determining hyperplanes of the parallelism class of C into a symbol
(5.1)
{ H1 , . . . , Hp ∣ K1 , . . . , Kq } ,
16
in which the K1 , . . . , Kq determine the parallelism class of D. The hyperplanes
H1 , . . . , Hp which cut C but not D are the complementary hyperplanes of the cube
pair, or of the parallelism class.
An orientation of a cube pair (C, D) is an orientation of the face D. In order to
compare orientations of parallel cube pairs (Ci , Di ) we can compare the orientations
on the faces Di , which are themselves parallel cubes, but must also take into account
the position of the faces within the ambient cubes Ci . For this we introduce the
following notion.
5.2 Definition. Two parallel cube pairs (C1 , D1 ) and (C2 , D2 ) have the same parity
if the number of complementary hyperplanes that separate D1 from D2 , is even.
Otherwise they have the opposite parity.
5.3 Definition. Let (C1 , D1 ) and (C2 , D2 ) be parallel cube pairs, each with an orientation. The orientations are aligned if one of the following conditions holds:
(a) (C1 , D1 ) and (C2 , D2 ) have the same parity, and D1 and D2 are compatibly
oriented; or
(b) (C1 , D1 ) and (C2 , D2 ) have the opposite parity, and D1 and D2 are not compatibly oriented.
In the symbol (5.1) describing the parallelism class of a cube pair (C, D), the hyperplanes are not ordered; the only relevant data is which are to the left, and which to the
right of the vertical bar. If the cube pair (C, D) is oriented, then the symbol receives
additional structure coming from the orientation of D. We group the determining
hyperplanes as before, and include a vertex R of D into a new symbol
(5.2)
{ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R } .
Here, in the case q > 0, the hyperplanes K1 , . . . , Kq form an ordered list which, together
with the vertex R are a presentation of the oriented cube D. In the case q = 0 this
list is empty and we replace it by the sign representing the orientation of the vertex
D = R, obtaining a symbol of the form
(5.3)
{ H1 , . . . , Hp ∣ + ∣ R }
or
{ H1 , . . . , Hp ∣ − ∣ R } .
In either case the hyperplanes H1 , . . . , Hp remain an unordered set. Conversely, a
formal expression as in (5.2) or (5.3) is the symbol of some oriented (p, q)-cube pair
precisely when the hyperplanes H1 , . . . , Kq are distinct and have nonempty (pairwise)
intersection, and the vertex R is adjacent to all of them.
The following definition captures the notion of alignment of orientations in terms of
the associated symbols.
17
5.4 Definition. Symbols
{ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R }
and
{ H1′ , . . . , Hp′ ∣ K1′ , . . . , Kq′ ∣ R′ }
of the form (5.2) are equivalent if
(a) the sets { H1 , . . . , Hp } and { H1′ , . . . , Hp′ } are equal;
(b) the K1 , . . . , Kq are a permutation of the K1′ , . . . , Kq′ ; and
(c) the number of hyperplanes among the H1 , . . . , Kq separating R and R′ has the
same parity as the permutation in (b).
In the case of symbols of the form (5.3) we omit (b) and replace (c) by
(c′ ) the number of hyperplanes among the H1 , . . . , Hp separating R and R′ is even
if the orientation signs agree, and odd otherwise.
An oriented (p, q)-symbol is an equivalence class of symbols. We shall denote the
equivalence class of the symbol (5.2) by
[ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ],
or simply by [ H ∣ K ∣ R ] when no confusion can arise, and we use similar notation
in the case of symbols of the form (5.3). We shall denote the set of oriented (p, q)symbols by Hqp , and the (disjoint) union Hq0 ∪ ⋯ ∪ Hqn−q by Hq .
5.5 Proposition. The oriented symbols associated to oriented (p, q)-cube pairs agree
precisely when the orientations of the cube pairs are aligned.
Our generalization of the Pytlik-Szwarc complex will be a differential complex designed to capture the combinatorics of oriented, aligned cube pairs:
(5.4)
d
d
d
d
C[H0 ] Ð→ C[H1 ] Ð→ ⋯ Ð→ C[Hn−1 ] Ð→ C[Hn ].
5.6 Definition. The space of oriented q-cochains of type p in the Pytlik-Szwarc
complex is the space of finitely supported, anti-symmetric, complex-valued functions
on Hqp . Here, a function is anti-symmetric if
f ([ H ∣ K ∣ R ]) + f ([ H ∣ K ∣ R ]∗ ) = 0,
where we have used the involution on Hqp defined by reversing the orientation of the
symbol. We shall denote this space by C[Hqp ]. The space of oriented q-cochains is
defined similarly using the oriented symbols of type (p, q) for all 0 ≤ p ≤ n − q. It splits
as the direct sum
C[Hq ] = C[Hq0 ] ⊕ ⋅ ⋅ ⋅ ⊕ C[Hqn−q ].
18
5.7 Remark. As with the Julg-Valette cochains, the space of oriented Pytlik-Szwarc
q-cochains of type p is a subspace of the full space of Pytlik-Szwarc q-cochains of type
p, which is the vector space of all finitely supported functions on the set Hqp . We shall
follow conventions similar to those in Section 3: we write
[ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ]
or
[H ∣K ∣R]
for both the Dirac function at an oriented symbol and the symbol itself, and
⟨ H ∣ K ∣ R ⟩ = [ H ∣ K ∣ R ] − [ H ∣ K ∣ R ] ∈ C[Hqp ]
∗
for the difference of the Dirac functions. Further, linear operators will be defined on
the full space of cochains by specifying their values on the basis of Dirac functions at
the oriented symbols. We shall typically omit the elementary check that an operator
commutes with the involution and so restricts to an operator on the spaces of oriented
cochains.
We now define the differential in the Pytlik-Szwarc complex (5.4).
5.8 Definition. The Pytlik-Szwarc differential is the linear map d ∶ C[Hq ] → C[Hq+1 ]
which is 0 on oriented symbols of type (0, q) and which satisfies
p
̂i , . . . , Hp ∣ Hi , K1 , . . . , Kq ∣ Ri ]
d [ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ] = ∑ [ H1 , . . . , H
i=1
for oriented (p, q)-symbols with p, q ≥ 1. Here, Ri is the vertex separated from R by
Hi alone and, as usual, a ‘hat’ means that an entry is removed. When q = 0 the same
formula is used for symbols of the form [H ∣+∣ R] which, together with the requirement
that d commute with the involution, determines d on symbols of the form [H ∣ − ∣ R].
Since d maps an oriented symbol of type (p, q) to a linear combination of oriented
symbols of type (p − 1, q + 1) in all cases, it splits as the direct sum of linear maps
p−1
]
d ∶ C[Hqp ] Ð→ C[Hq+1
for 0 < p ≤ n − q, and is 0 on the C[Hq0 ].
5.9 Lemma. The Pytlik-Szwarc differential d, regarded as an operator on the space
of oriented cochains, satisfies d2 = 0.
5.10 Example. Let T be a tree. The Pytlik-Szwarc complex has the form
d ∶ C ⊕ C[H01 ] Ð→ C[H10 ],
where d is 0 on C and, after identifying each of C[H01 ] and C[H10 ] with the space of
finitely supported functions on the set of edges of T , the identity C[H01 ] → C[H10 ]. For
the identifications, note that both H10 and H01 are identified with the set of oriented
edges in T and that the involution acts by reversing the orientation. So the space
of anti-symmetric functions on each identifies with the space of finitely supported
functions on the set of edges.
19
Our goal for the remainder of this section is to analyze the Pytlik-Szwarc complex.
Emphasizing the similarities with the Julg-Valette complex we begin by providing a
formula for the formal adjoint of the Pytlik-Szwarc differential.
5.11 Definition. Let δ ∶ C[Hq ] → C[Hq−1 ] be the linear map which is 0 on oriented
symbols of type (p, 0) and which satisfies
q
̂j , . . . , Kq ∣ R ] ,
δ [ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ] = ∑(−1)j [ H1 , . . . , Hp , Kj ∣ K1 , . . . , K
j=1
for oriented symbols of type (p, q) with q ≥ 1. Again a ‘hat’ means that an entry is
removed. Since δ maps an oriented symbol of type (p, q) to a linear combination of
oriented symbols of type (p + 1, q − 1) it splits as a direct sum of linear maps
p+1
δ ∶ C[Hqp ] → C[Hq−1
]
for 0 < q ≤ n − p, and is 0 on the C[H0p ].
5.12 Definition. We define an inner product on the full space of Pytlik-Szwarc qcochains √
by declaring that the elements of Hq are orthogonal, and that each has
length 1/ 2. The subspace C[Hq ] of oriented Pytlik-Szwarc q-cochains inherits an
inner product in which
⎧
1,
[ H ∣ K ∣ R ] = [ H ′ ∣ K ′ ∣ R′ ]
⎪
⎪
⎪
⎪
⟨⟨ H ∣ K ∣ R ⟩ , ⟨ H ′ ∣ K ′ ∣ R′ ⟩⟩ = ⎨−1, [ H ∣ K ∣ R ] = [ H ′ ∣ K ′ ∣ R′ ]∗
⎪
⎪
⎪
⎪
otherwise
⎩0,
5.13 Lemma. The operators d and δ of Definitions 5.8 and 5.11 are formally adjoint
and bounded with respect to the inner products in Definition 5.12.
5.14 Proposition. The Pytlik-Szwarc Laplacian
∆ = (d + δ)2 = dδ + δd ∶ C[Hq ] Ð→ C[Hq ]
acts on the summand C[Hqp ] as scalar multiplication by p + q.
Proof. We prove the above statement for the operator dδ + δd defined on the full
space of cochains. This operator equals ∆ when restricted to the subspace of oriented
cochains. The proof is a direct calculation. The result of applying δd to an oriented
symbol [ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ] of type (p, q) is the sum
p
∑[ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ] +
i=1
p
q
̂i , . . . , Hp , Kj ∣ Hi , K1 , . . . , K
̂j , . . . , Kq ∣ R ],
+ ∑ ∑(−1)j [ H1 , . . . , H
i=1 j=1
20
whereas the result of applying dδ is
q
̂ j , . . . , Kq ∣ R ] +
∑(−1)j+1 [ H1 , . . . , Hp ∣ Kj , K1 , . . . , K
j=1
q
p
j=1
i=1
̂i , . . . , Hp , Kj ∣ Hi , K1 , . . . , K
̂j , . . . , Kq ∣ R ].
+ ∑(−1)j+1 ∑[ H1 , . . . , H
When these are added, the second summands cancel and the first summands combine
to give (p + q)[ H1 , . . . , Hp ∣ K1 , . . . , Kq ∣ R ].
5.15 Corollary. The cohomology of the Pytlik-Szwarc complex is C in dimension
zero and 0 otherwise.
6. Continuous Fields of Hilbert Spaces
Our objective over the next several sections is to construct a family of complexes that
continuously interpolates between the Julg-Valette complex and the Pytlik-Szwarc
complex. We shall construct the interpolation within the Hilbert space context, using
the concept of a continuous field of Hilbert spaces.
We refer the reader to [Dix77, Chapter 10] for a comprehensive treatment of continuous fields of Hilbert spaces. In brief, a continuous field of Hilbert spaces over a
topological space T consists of a family of Hilbert spaces parametrized by the points
of T , together with a distinguished family Σ of sections that satisfies several axioms,
of which the most important is that the pointwise inner product of any two sections
in Σ is a continuous function on T . See [Dix77, Definition 10.1.2]. The following
theorem gives a convenient means of constructing continuous fields.
6.1 Theorem. Let T be a topological space, let {Ht } be a family of Hilbert spaces
parametrized by the points of T , and let Σ0 be a family of sections that satisfies the
following conditions:
(a) The pointwise inner product of any two sections in Σ0 is a continuous function
on T .
(b) For every t ∈ T the linear span of {σ(t) ∶ σ ∈ Σ } is dense in Ht .
There is a unique enlargement of Σ0 that gives {Ht }t∈T the structure of a continuous
field of Hilbert spaces.
Proof. The enlargement Σ consists of all sections σ such that for every t0 ∈ T and
every ε > 0 there is a section σ0 in the linear span of Σ0 such that
∥σ0 (t) − σ(t)∥t < ε
for all t in some neighborhood of t0 . See [Dix77, Proposition 10.2.3].
21
6.2 Definition. We shall call a family Σ0 , as in the statement of Theorem 6.1, a
generating family of sections for the associated continuous field of Hilbert spaces.
Ultimately we shall use the parameter space T = [0, ∞], but in this section we shall
concentrate on the open subspace (0, ∞], and then extend to [0, ∞] in the next
section. In both this section and the next we shall deal only with the construction of
continuous fields of Hilbert spaces; we shall construct the differentials acting between
these fields in Section 8.
We begin by completing the various cochain spaces from Section 3 in the natural way
so as to obtain Hilbert spaces.
6.3 Definition. Denote by `2 (X q ) the Hilbert space completion of the Julg-Valette
oriented cochain space C[X q ] in the inner product of Definition 3.18 in which the
basis comprised of the oriented cochains ⟨ C ⟩ is orthonormal.
6.4 Remark. As was the case in Section 3, we shall also consider the full cochain
space comprised of the square-summable functions on the set of oriented q-cubes. This
is the completion of the full space of Julg-Valette q-cochains in the inner product
of Definition 3.18, and contains the space `2 (X q ) of the previous definition as the
subspace of anti-symmetric functions.
We shall now construct, for every q ≥ 0, families of Hilbert spaces parametrized by
the topological space (0, ∞]. These will be completions of the spaces of Julg-Valette
q-cochains, both full and oriented, but with respect to a family of pairwise distinct
inner products. Considering the oriented cochains, we obtain a family of Hilbert
spaces `2t (X q ) each of which is a completion of the corresponding C[X q ]. The Hilbert
space `2∞ (X q ) will be the space `2 (X q ) just defined.
6.5 Definition. If D1 and D2 are q-cubes in X, and if D1 and D2 are parallel and
have compatible orientations, then denote by d(D1 , D2 ) the number of hyperplanes
in X that are disjoint from D1 and D2 and that separate D1 from D2 . If D1 and D2
are q-cubes in X, but are not parallel, or have incompatible orientations, then set
d(D1 , D2 ) = ∞.
If D1 and D2 are (compatibly oriented) vertices, then d(D1 , D2 ) is the edge-path
distance from D1 to D2 . In higher dimensions, if D1 and D2 are parallel then they
may be identified with vertices in the CAT(0) cube comples which is the intersection
of the determining hyperplanes for the parallelism class. If in addition they are compatibly oriented, then d(D1 , D2 ) is the edge-path distance in this complex. Compare
Theorem 4.2.
6.6 Definition. Let t > 0 and q ≥ 0. For every two oriented q-cubes D1 and D2 define
⟨D1 , D2 ⟩t = 21 exp(− 12 t2 d(D1 , D2 )),
22
where of course we set exp(− 12 t2 d(D1 , D2 )) = 0 if d(D1 , D2 ) = ∞, and then extend by
linearity to a sesqui-linear form on the full space of Julg-Valette q-cochains.
Note that the formula in the definition makes sense when t = ∞, where
1
2
exp(− 12 t2 d(D1 , D2 ))
⎧
⎪
⎪ 21 , D1 = D2
=⎨
⎪
⎪
⎩0, D1 ≠ D2 .
In particular, the form ⟨ , ⟩∞ is the one underlying Definition 3.18 that we used to
define `2 (X q ).
6.7 Theorem. The sesqui-linear form ⟨ , ⟩t is positive semi-definite.
Proof. Consideration of oriented, as opposed to unoriented, cubes merely gives two
(orthogonal copies) of each space of functions. Aside from this, the result is proved
in [NR98a, Technical Lemma, p.6] in the case q = 0. See also [GH10, Prop. 3.6]. The
case q > 0 reduces to the case q = 0 using Theorem 4.2.
6.8 Definition. For t ∈ (0, ∞] denote by `2t (X q ) the Hilbert space completion of the
Julg-Valette oriented cochain space C[X q ] in the inner product ⟨ , ⟩t .
6.9 Remark. The Hilbert spaces of the previous definition are completions of the
quotient of C[X q ] by the elements of zero norm. We shall soon see that every nonzero
linear combination of oriented q-cubes has nonzero `2t -norm for every t, so the natural
maps from C[X q ] into the `2t (X q ) are injective.
Next, we define a generating family of sections, using either one of the following
lemmas; on the basis of Theorem 6.1, it is easy to check that the continuous fields
arising from the lemmas are one and the same.
6.10 Lemma. Let t ∈ (0, ∞]. The set of all sections of the form
t ↦ f ∈ C[X q ] ⊆ `2t (X q ),
indexed by all f ∈ C[X q ], is a generating family of sections for a continuous field.
6.11 Lemma. The set of all sections of the form
t ↦ f (t) ⟨C⟩ ∈ `2t (X q ),
where f is a continuous scalar function on (0, ∞] and C is an oriented q-cube, is a
generating family of sections for a continuous field.
23
The continuous fields that we have constructed are not particularly interesting as
continuous fields. In fact they are isomorphic to constant fields (they become much
more interesting when further structure is taken into account, as we shall do later in
the paper). For the sequel it will be important to fix a particular isomorphism, and
we conclude this section by doing this.
The required unitary isomorphism will be defined using certain cocycle operators
Wt (C1 , C2 ), which are analogues of those studied by Valette in [Val90] in the case
of trees. In the case q = 0 the cocycle operators for general CAT(0) cube complexes
were constructed in [GH10]. The case where q > 0 involves only a minor elaboration
of the q = 0 case, and so we shall refer to [GH10] for details in what follows.
6.12 Definition. If D is a q-cube that is adjacent to a hyperplane H, then define
Dop to be the opposite face to D in the unique (q + 1)-cube that is cut by H and
contains D as a q-face (such a cube exists by Lemma 3.6). In the case D is oriented,
we orient Dop compatibly. In either case, we shall refer to a pair such as D and Dop
as being adjacent across H.
6.13 Definition. Let C and C op be adjacent across a hyperplane H, as in the previous
definition. If D is any oriented q-cube that is adjacent to H, then for t ∈ (0, ∞] we
define
1 2
⎧
2
⎪
⎪(1 − e−t )1/2 D − e− 2 t Dop , if D is separated from C by H
Wt (C , C)D = ⎨ − 1 t2 op
2
⎪
2 D
+ (1 − e−t )1/2 D, if D is not separated from C by H;
⎪
⎩e
op
in addition we define
Wt (C op , C)D = D
if D is not adjacent to H.
We extend Wt (C op , C) by linearity to a linear operator on the spaces of (full and
oriented) Julg-Valette q-cochains.
For example
W0 (C op , C)C = C op
and W0 (C op , C)C op = −C,
W∞ (C op , C)C = C
and W∞ (C op , C)C op = C op ,
while
and indeed W∞ (C op , C) is the identity operator. More generally, when restricted to
the two-dimensional space spanned by the ordered basis (D, Dop ) with D adjacent
to H but not separated from C by H, the operator Wt (C op , C) acts as the unitary
matrix
1 2
2
(1 − et )1/2
−e− 2 t
[
].
1 2
2
e− 2 t
(1 − et )1/2
24
In particular, Wt (C op , C) extends to a unitary operator on the completed cochain
spaces of Definition 6.3 and subsequent remark.
Let us now assume that two q-cubes C1 and C2 are parallel, but not necessarily
adjacent across a hyperplane. It follows from Theorem 4.2 that that there exists a
path of q-cubes E1 , E2 , . . . , En , with E1 = C1 and En = C2 , where each consecutive
pair Ei , Ei+1 consists of parallel and adjacent q-cubes. For all t ≥ 0 let us define
(6.1)
Wt (C1 , C2 ) = Wt (E1 , E2 )Wt (E2 , E3 ) . . . Wt (En−1 , En ).
This notation, which omits mention of the path, is justified by the following result:
6.14 Proposition. The unitary operator Wt (C1 , C2 ) is independent of the path from
C1 to C2 .
Proof. Let γ and γ ′ be two cube paths connecting cubes C1 and C2 . As the cubes
C1 and C2 are parallel, by Theorem 4.2 they can be thought of as vertices in the
CAT(0) cube complex created from their parallelism class. The paths γ and γ ′ then
give rise to vertex paths in this CAT(0) cube complex with common beginning and
end vertices. In this way we reduce the general case of the proposition to the zero
dimensional case, which has been proved in [GH10, Lemma 3.3].
In what follows we shall use the base vertex P0 that was selected during the construction of the Julg-Valette complex.
6.15 Definition. Let t ∈ (0, ∞]. For every oriented q-cube D let
Ut D = Wt (D0 , D)D,
where D0 is the cube nearest to the base vertex P0 in the parallelism class of D (see
Proposition 4.3). Extend Ut by linearity to a linear operator on the spaces of full and
oriented Julg-Valette q-cochains; in particular, on oriented cochains we have
Ut ∶ C[X q ] Ð→ C[X q ].
6.16 Lemma. The linear operator Ut is a vector space isomorphism.
Proof. Consider the increasing filtration of the cochain space, indexed by the natural
numbers, in which the nth space is spanned by those cubes whose nearest vertex to
P0 in the edge-path metric is of distance n or less from P0 . The operator Ut preserves
this filtration. In fact, a simple direct calculation (see [GH10, Lemma 4.7]) shows
that
Ut D = Wt (D0 , D)D
= constant ⋅ D + linear combination of cubes closer to P0 than D.
This formula shows that the induced map on associated graded spaces is an isomorphism. So Ut is an isomorphism.
25
6.17 Lemma. If D1 and D2 are any two oriented q-cubes in X, then
⟨Ut D1 , Ut D2 ⟩ = ⟨D1 , D2 ⟩t ,
where the inner product on the left hand side is that of `2 (X q ).
6.18 Remark. The lemma implies that the sesqui-linear form ⟨ , ⟩t is positive definite for each t > 0, since ⟨ , ⟩ is positive-definite and Ut is an isomorphism.
Proof of the lemma. We can assume that the q-cubes D1 and D2 are parallel and
compatibly oriented since otherwise both sides of the formula are zero. Let D0 denote
the q-cube in the parallelism class that is nearest to the base vertex P0 . Then the
unitarity of Wt and Proposition 6.14 give
⟨Ut D1 , Ut D2 ⟩ = ⟨Wt (D0 , D1 )D1 , Wt (D0 , D2 )D2 ⟩
= ⟨Wt (D0 , D2 )∗ Wt (D0 , D1 )D1 , D2 ⟩
= ⟨Wt (D2 , D0 )Wt (D0 , D1 )D1 , D2 ⟩
= ⟨Wt (D2 , D1 )D1 , D2 ⟩.
But, by an elaboration of [GH10, Proposition 3.6] we have
1
2 d(D
(6.2) Wt (D2 , D1 )D1 = e− 2 t
2 ,D1 )
D2 + multiples of oriented cubes other than D2 .
Hence we conclude that
1
2 d(D
⟨Wt (D2 , D1 )D1 , D2 ⟩ = 21 e− 2 t
2 ,D1 )
= ⟨D1 , D2 ⟩t ,
as required.
The following results are immediate consequences of the above:
6.19 Theorem. For all t ∈ (0, ∞] the map
Ut ∶ C[X q ] Ð→ C[X q ]
extends to a unitary isomorphism
Ut ∶ `2t (X q ) Ð→ `2∞ (X q ).
6.20 Theorem. The unitary operators Ut determine a unitary isomorphism from the
continuous field {`2t (X q )}t∈(0,∞] generated by sections in Lemmas 6.10 and 6.11 to the
constant field with fiber `2 (X q ).
26
7. Extension of the Continuous Field
In this section we shall extend the continuous fields over (0, ∞] defined in Section 6
by adding the following fibers at t = 0.
7.1 Definition. We shall denote by `20 (X q ) the completion of the space of oriented
Pytlik-Szwarc q-cochains in the inner product of Definition 5.12. It is the subspace
of anti-symmetric functions in the Hilbert space of all square-summable functions on
the set of oriented symbols Hq .
The following two definitions focus on the particular continuous sections that we shall
extend.
7.2 Definition. Let p, q ≥ 0 and let (C, D) be an oriented (p, q)-cube pair. The
associated basic q-cochain of type p is the linear combination
fC,D = ∑ (−1)d(D,E) E
E∥C D
in the full cochain space. Here, the sum is over those q-cubes E in C that are parallel
to D, each of which is given the orientation compatible with the orientation of D.
The associated basic oriented cochain is
f⟨C,D⟩ = fC,D − fC,D∗ = ∑ (−1)d(D,E) ⟨ E ⟩,
E∥C D
belonging to the space C[X q ] of oriented q-cochains.
7.3 Example. For q ≥ 0, a basic q-cochain of type p = 0 is just a single oriented
q-cube. A basic 0-cochain of type 1 is a difference of vertices across an edge. Finally,
if p + q > dim(X) then there are no basic q-cochains of type p, since there are no
(p + q)-cubes in X.
7.4 Definition. A basic section of type p of the continuous field { `2t (X q ) }t∈(0,∞] is a
continuous section of the form
(0, ∞] ∋ t z→ t−p f⟨C,D⟩ ∈ `2t (X q ),
where (C, D) is an oriented (p, q)-cube pair.
We shall extend the basic sections to sections over [0, ∞] by assigning to each of
them a value at t = 0 in the Hilbert space `20 (X q ), namely the Pytlik-Szwarc symbol
associated to the cube pair (C, D), as in Section 5. We shall write it as
⟨ C, D ⟩ = [C, D] − [C, D]∗ ∈ `20 (X q ).
Compare Definition 5.4 and Remark 5.7. We shall prove the following result.
27
7.5 Theorem. Let q ≥ 0.
(a) The pointwise inner product
⟨t−p1 fC1 ,D1 , t−p2 fC2 ,D2 ⟩t
of any two basic sections (of possibly different types) extends to a continuous
function on [0, ∞].
(b) The value of this continuous function at 0 ∈ [0, ∞] is equal to the inner product
⟨[C1 , D1 ], [C2 , D2 ]⟩0 .
7.6 Example. Suppose that X is a tree. When q = 1, the only basic sections are
those of type p = 0, and they are the functions t ↦ E, where E is an oriented edge in
X. Theorem 7.5 is easily checked in this case. When q = 0 there are basic sections
t ↦ Q of type p = 0, which are again easily handled, but also basic sections of type
p = 1. These have the form
t ↦ t−1 (P − Q),
where P and Q are adjacent vertices in the tree. One calculates that
1 2
⟨t−1 (P − Q), t−1 (P − Q)⟩t = 2t−2 (1 − e− 2 t ),
which converges to 1 as t → 0, in agreement with Theorem 7.5. In addition if t−1 (R−S)
is a second, distinct basic cochain, and if the vertices P, Q, R, S are arranged in
sequence along a path in the tree, then a short calculation reveals that if d is the
distance between Q and R, then
1 2
1 2
⟨t−1 (P − Q), t−1 (R − S)⟩t = −t−2 e−d 2 t (1 − e− 2 t )2 = O(t2 ).
In particular the inner product converges to 0 as t ↘ 0, again in agreement with
Theorem 7.5.
7.7 Definition. An extended basic section of type p of the continuous field of Hilbert
spaces { `2t (X q ) }t∈[0,∞] is a section of the form
⎧
⎪
t=0
⎪⟨C, D⟩,
t z→ ⎨ −p
⎪
⎪
⎩t f⟨C,D⟩ , t > 0,
where (C, D) is an oriented (p, q)-cube pair.
The basic sections form a generating family of sections for the continuous field
{`2t (X q )}t∈(0,∞] , and of course the symbols ⟨C, D⟩ span `20 (X q ). So it follows from
the theorem that the extended basic sections form a generating family of sections for
a continuous field over [0, ∞] with fibers `2t (X q ), whose restriction to (0, ∞] is the
continuous field of the previous section.
28
We shall prove Theorem 7.5 by carrying out a sequence of smaller calculations. The
following formula is common to all of them, and it will also be of use in Section 8.
Here, and subsequently, we shall write O(tp ) for any finite sum of oriented q-cubes
times coefficient functions, each of which is bounded by a constant times tp as t ↘ 0.
7.8 Lemma. If (C, D) is an oriented (p, q)-cube pair then
∑ (−1)d(D,E) Wt (D, E)E = (−t)p Dop + O(tp+1 ),
(7.1)
E∥C D
where Dop is the q-face of C separated from D by the complementary hyperplanes of
the pair (C, D), with compatible orientation.
Proof. We shall prove the lemma by induction on p. The case p = 0 is clear. As for
the case p > 0, let H be a hyperplane that cuts C but not D. Our aim is to apply the
induction hypothesis to the codimension-one faces of C separated by H. Denote these
faces by C± with C+ being the face containing D; denote D+ = D and D− the face of
C− directly across H from D; and finally denote by D±op the face in C± separated from
D± by all the complementary hyperplanes of the pair (C, D) except H. We have, in
particular, Dop = D−op .
Now, the expression on the left hand side of (7.1) depends on the cube pair (C, D)
and for the course of the proof we shall denote it by gC,D . We compute the summand
of gC,D corresponding to a face E that belongs to C− using the path from D+ to D−
and on to E. Doing so, we see that
gC,D = gC+ ,D+ − Wt (D+ , D− )gC− ,D−
1 2
1
2
= (1 − e− 2 t ) gC+ ,D+ − (1 − e−t ) 2 gC− ,D− .
Here, we have used that the coefficient of gC+ ,D+ at a face E of C+ equals the coefficient
of gC− ,D− at the face of C− which is directly across H from E. By the induction
1 2
hypothesis, gC+ ,D+ = (−t)p−1 D+op + O(tp ), which is O(tp−1 ). Since 1 − e− 2 t is O(t2 ) the
first term in this expression is O(tp+1 ). As for the second term, again by induction
we have gC− ,D− = (−t)p−1 D−op + O(tp ), which is O(tp−1 ). It follows that
2
1
2
1
−(1 − e−t ) 2 gC− ,D− = −t gC− ,D− + (t − (1 − e−t ) 2 ) gC− ,D−
2
1
= (−t)p D−op + O(tp+1 ) + (t − (1 − e−t ) 2 )O(tp−1 )
= (−t)p Dop + O(tp+1 ),
1
where we have used that t − (1 − e−t ) 2 is O(t3 ). Putting things together, the lemma
is proved.
2
In the previous section we defined unitary isomorphisms Ut ∶ `2t (X q ) → `2 (X q ). While
these were defined using a specific choice of base point within each parallelism class
29
of q-cubes, the choice is not important as far as the unitarity of Ut is concerned.
We shall exploit this by making judicious choices of base point to calculate the inner
products in Theorem 7.5.
7.9 Lemma. Let (C, D) be an oriented (p, q)-cube pair, and let fC,D be the associated
basic q-cochain of type p. The pointwise inner product
⟨t−p fC,D , t−p fC,D ⟩t
converges to
1
2
as t ↘ 0.
Proof. Choose D as the base point for defining the unitary isomorphisms Ut . Then
Ut fC,D is exactly the expression (7.1) in the previous lemma. It follows from the
lemma that
⟨t−p fC,D , t−p fC,D ⟩t = ⟨t−p Ut fC,D , t−p Ut fC,D ⟩∞
= ⟨(−1)p Dop + O(t), (−1)p Dop + O(t)⟩∞
= 21 + O(t),
and the result follows.
7.10 Lemma. Let (C1 , D1 ) and (C2 , D2 ) be parallel (p, q)-cube pairs of the same
parity, in which the q-dimensional faces are compatibly oriented. The pointwise inner
product
⟨t−p fC1 ,D1 , t−p fC2 ,D2 ⟩t
converges to
1
2
as t ↘ 0.
Proof. We may assume that D2 lies on the same side of each of the complementary
hyperplanes of the parallelism class as D1 ; indeed replacing D2 by this face, if necessary, does not change the corresponding basic cochain. Choose D1 as the base point
for defining the unitary isomorphisms Ut , so that by Lemma 7.8 we have
Ut fC1 ,D1 = (−t)p D1op + O(tp+1 )
and also, using the identity Wt (D1 , E) = Wt (D1 , D2 )Wt (D2 , E) for the q-dimensional
faces E of C2 ,
Ut fC2 ,D2 = (−t)p Wt (D1 , D2 )D2op + O(tp+1 ).
But, the hyperplanes separating D1 and D2 are precisely those separating D1op and
D2op , so that by (6.2) we have
Wt (D1 , D2 )D2op = Wt (D1op , D2op )D2op
1
2
= e− 2 d(D1 ,D2 )t D1op + terms orthogonal to D1op .
30
Putting everything together we get
⟨t−p fC1 ,D1 , t−p fC2 ,D2 ⟩t = ⟨t−p Ut fC1 ,D1 , t−p Ut fC2 ,D2 ⟩∞
1
2
= e− 2 d(D1 ,D2 )t ⟨(−1)p D1op , (−1)p D1op ⟩∞ + O(t)
and the result follows from this.
7.11 Lemma. Let (C1 , D1 ) and (C2 , D2 ) be oriented cube pairs of types (p1 , q) and
(p2 , q), respectively, and let fC1 ,D1 and fC2 ,D2 be the associated basic q-cochains. If
(C1 , D1 ) and (C2 , D2 ) are not parallel, or if D1 and D2 are not compatibly oriented,
then the pointwise inner product
⟨t−p1 fC1 ,D1 , t−p2 fC2 ,D2 ⟩t
converges to 0 as t ↘ 0. In particular, this is the case if p1 ≠ p2 .
Proof. If D1 and D2 fail to be parallel or have incompatible orientations, then fC1 ,D1
and fC2 ,D2 are orthogonal in the full cochain space for all t > 0, and the lemma is
proved. So we can assume that D1 and D2 are parallel and compatibly oriented, and
therefore that C1 and C2 are not parallel. There is then, after reindexing if necessary,
a hyperplane H that passes through C2 but not C1 , and through neither D1 nor D2 .
Choose as a base point for the unitary Ut a q-dimensional face D of C2 which is
parallel to the Di , compatibly oriented, and on the same side of H as the cube C1 .
So fC2 ,D2 = ±fC2 ,D and also
⟨t−p1 fC1 ,D1 , t−p2 fC2 ,D ⟩t = ⟨t−p1 Ut fC1 ,D1 , t−p2 Ut fC2 ,D ⟩∞
= ⟨t−p1 Ut fC1 ,D1 , (−1)p2 Dop + O(t)⟩∞ ,
where Dop is the face of C2 separated from D by all the complementary hyperplanes
of the pair (C2 , D). In particular, D and Dop are on opposite sides of H. Now,
it follows from the definition of Ut and basic properties of the cocycle Wt that all
cubes appearing in the support of Ut fC1 ,D1 are on the same side of H as D. Further,
from Lemma 7.8 we have that Ut fC1 ,D1 is O(tp1 ), so that the inner product above is
O(t).
Proof of Theorem 7.5. The possible values of the inner product in (b) are 0 and ±1/2:
the positive value occurs when the oriented cube pairs (C1 , D1 ) and (C2 , D2 ) are
parallel and aligned; the negative value occurs when they are parallel and not aligned;
and 0 occurs when they are not parallel. The result now follows from Lemmas 7.10
and 7.11.
As we already pointed out, Theorem 7.5 allows us to extend our continuous field
to [0, ∞]. In the sequel it will be convenient to work with the following generating
family of continuous bounded sections.
31
7.12 Definition. A (not necessarily continuous) section σ of the continuous field
{`2t (X q )}t∈[0,∞] is geometrically bounded if there is a finite set A ⊆ Xq such that s(t)
is supported in A for all t ∈ (0, ∞].
7.13 Proposition. The space of geometrically bounded, continuous sections of the
continuous field {`2t (X q )}t∈[0,∞] is spanned over C[0, ∞] by the extended basic continuous sections.
Proof. Every basic continuous section is certainly geometrically bounded. If X is a
finite complex, then the converse is true since the fiber dimension of the continuous
field is finite and constant in this case, and so the continuous field is a vector bundle,
while the basic continuous sections span each fiber of the bundle. In the general
case, we can regard any geometrically bounded continuous section as a section of the
continuous field associated to a suitable finite subcomplex, as in Lemma 2.4, and so
express it as a combination of basic continuous sections.
8. Differentials on the Continuous Field
The purpose of this section is to construct differentials
dt
dt
dt
dt
`2t (X 0 ) Ð→ `2t (X 1 ) Ð→ ⋯ Ð→ `2t (X n−1 ) Ð→ `2t (X n )
that continuously interpolate between the Julg-Valette differentials at t = ∞ and the
Pytlik-Szwarc differentials at t = 0. For later purposes it will be important to use
weighted versions of the Julg-Valette differentials, as in Definition 3.23. But first we
shall proceed without the weights, and then indicate at the end of this section how
the weights are incorporated.
Recall that the operators
Ut ∶ C[X q ] Ð→ C[X q ]
from Definition 6.15 were proved to be isomorphisms in Lemma 6.16.
8.1 Definition. For t ∈ (0, ∞] we define
dt = Ut−1 dUt ∶ C[X q ] Ð→ C[X q+1 ],
where d is the Julg-Valette differential from Definition 3.9. In addition, we define
d0 ∶ C[Hq ] Ð→ C[Hq+1 ]
to be the Pytlik-Szwarc differential from Definition 5.8.
We aim to prove the following continuity statement concerning these operators:
32
8.2 Theorem. If {σ(t)} is any continuous and geometrically bounded section of the
continuous field {`2t (X q )}t∈[0,∞] , then the pointwise differential {dt σ(t)} is a continuous and geometrically bounded section of {`2t (X q+1 )}t∈[0,∞] .
According to Proposition 7.13, the space of continuous and geometrically bounded
sections is generated as a module over C[0, ∞] by the extended basic sections, so it
suffices to prove Theorem 8.2 for such a section. This we shall now do, following two
preliminary lemmas.
8.3 Lemma. Let (C, D) be an oriented (p, q)-cube pair and assume that all the
complementary hyperplanes of the pair (C, D) separate D from the base point P0 .
The associated basic q-cochain of type p satisfies
t−p Ut fC,D = Wt (D0 , D1 )D + O(t),
where D0 is the q-cube in X that is closest to the base point P0 among cubes parallel
to D, and D1 is the face of C that is parallel to D and separated from D by all the
complementary hyperplanes.
Proof. According to our definitions, fC,D = (−1)p fC,D1 and
Ut fC,D1 = ∑ (−1)d(D1 ,E) Wt (D0 , E)E
E∥C D
= Wt (D0 , D1 ) ∑ (−1)d(D1 ,E) Wt (D1 , E)E
E∥C D
= (−t) D + O(tp+1 ),
p
where we have applied Lemma 7.8. The result follows.
8.4 Lemma. Let (C, D) and D1 be as in the previous lemma. Let C0 be the nearest
cube to P0 in the parallelism class of C, let F be the face of C0 which is parallel to
D and separated from the base point P0 by the complementary hyperplanes, and let
F1 be the face of C0 that is parallel to D and separated from F by the complementary
hyperplanes. Then
(a) H ∧ F is nonzero if and only if H is a complementary hyperplane, in which case
H ∧ F ⊆ C0 ;
(b) d(C0 , C) = d(F, D) = d(F1 , D1 );
(c) Wt (D0 , D1 )D = F + O(t).
Proof. Consider first the case q = 0. In this case, D0 = P0 and the vertex F1 is characterized by the following hyperplane property from the proof of Proposition 4.3: every
hyperplane separating P0 and F1 is parallel to at least one determining hyperplane of
parallelism class of C (and C0 ).
33
For (a), H ∧ F is nonzero exactly when H is adjacent to F and separates it from P0 .
The hyperplanes cutting C0 certainly satisfy this condition. Conversely, a hyperplane
satisfying this condition must intersect all determining hyperplanes by Lemma 2.8,
so cannot separate F1 from P0 and so must cut C0 .
For (b), no determining hyperplane (of the parallelism class of C) separates F and
D. It follows easily that a hyperplane separates C and C0 if and only if it separates
F and D. The same argument applies to F1 and D1 .
For (c), from the cocycle property we have
Wt (D0 , D1 )D = Wt (D0 , F1 )Wt (F1 , D1 )D.
To evaluate this, observe that a hyperplane appearing along (a geodesic) path from F1
to D1 must cross every determining hyperplane. It follows that Wt (F, F1 ) = W (D, D1 )
commutes with Wt (F1 , D1 ) and we have
Wt (F1 , D1 )D = Wt (F, F1 )W (D1 , D)Wt (F1 , D1 )D
= Wt (F, D)D
1
2
= e 2 dt F + O(t),
where d = d(F, D), and the last equality follows from an elaboration of [GH10, Proposition 3.6]. Finally, no hyperplane separating D0 and F1 is adjacent to F so that
Wt (D0 , F1 )F = F . Putting things together, the result follows.
We reduce the general case to the case q = 0 using Proposition 4.2, according to which
the set of q-cubes parallel to D is the vertex set of a CAT(0) cube complex in such
a way that the (p + q)-cubes in X correspond to the p-cubes in this complex. The
key observation is that the p-cube in this complex corresponding to the (p + q) cube
C0 in the statement of the lemma is the p-cube closest to the vertex corresponding
to D0 .
Proof of Theorem 8.2. Let (C, D) be an oriented (p, q)-cube pair, with associated
extended basic q-cochain
⎧
⎪
⎪[C, D], t = 0
σC,D (t) = ⎨ −p
⎪
⎪
⎩t fC,D , t > 0.
We shall show that the section {dt σ(t)}t∈[0,∞] is a linear combination of extended
basic cochains, plus a term that is geometrically bounded and O(t).
After possibly changing a sign, we can assume that D is the furthest from the base
point among the q-dimensional faces of C parallel to D. In other words, we can
assume that the complementary hyperplanes H1 , . . . , Hp of the pair (C, D) separate
34
D from the base point. Each Hi ∧ D is therefore a (q + 1)-dimensional face of C, and
we shall show that
p
dt (σC,D (t)) = ∑ σC,Hi ∧D (t) + O(t).
i=1
We have equality when t = 0, so it suffices to show that
p
dt (t−p fC,D ) = t−(p−1) ∑ fC,Hi ∧D + O(t)
i=1
for t > 0, or equivalently that
p
(8.1)
d Ut (t−p fC,D ) = ∑ Ut (t−(p−1) fC,Hi ∧D ) + O(t).
i=1
As for the left hand side of (8.1), applying Lemmas 8.3 and 8.4 we have
p
d Ut (t−p fC,D ) = d F + O(t) = ∑ Hi ∧ F + O(t),
i=1
where F is as in the statement of Lemma 8.4. So, to complete the verification of (8.1)
it suffices to check that
Ut (t−(p−1) fC,Hi ∧D ) = Hi ∧ F + O(t).
But this follows from Lemmas 8.3 and 8.4, applied to the (p − 1, q + 1)-cube pair
(C, Hi ∧ D) (although a little care must be taken here since the base cube D0 that is
nearest to P0 within the parallelism class of D should be replaced by an analogous
base cube for the parallelism class of Hi ∧ D).
Consider now the adjoint operators
(8.2)
δt = Ut−1 δUt ∶ C[X q ] Ð→ C[X q+1 ]
for t > 0, together with the adjoint Pytlik-Szwarc differential
δ0 ∶ C[Hq ] Ð→ C[Hq+1 ]
8.5 Theorem. If {σ(t)} is any continuous and geometrically bounded section of the
continuous field {`2t (X q+1 )}t∈[0,∞] , then {δt σ(t)} is a continuous and geometrically
bounded section of the continuous field {`2t (X q )}t∈[0,∞] .
Proof. While this could be approached through computations similar to those used to
prove Theorem 8.2, there is a shortcut. Each continuous and geometrically bounded
section can be viewed as associated to a finite subcomplex of X as in Lemma 2.4. In
the case of a finite complex the differentials {dt } constitute a map of vector bundles,
and their pointwise adjoints {δt } automatically give a map of vector bundles too.
35
Finally, we return to the issue of weights, which will be important in the next section
when we work in the context of Kasparov theory. Let wt the function on hyperplanes
defined by the formula
(8.3)
⎧
⎪
⎪1 + t dist(H, P0 ), 0 < t ≤ 1
wt (H) = ⎨
⎪
1 ≤ t ≤ ∞.
⎪
⎩1 + dist(H, P0 )
In the next section we shall work with the weighted operators
(8.4)
dt = Ut−1 dwt Ut ∶ C[X q ] Ð→ C[X q+1 ],
for t > 0, where as before Ut is the isomorphism from Definition 6.15, and where dwt
is the weighted Julg-Valette differential described in Definition 3.23.
If t > 0, then operator in (8.4) does not extend from C[X q ] to a bounded operator
between `2t -spaces. But since the pointwise values of a geometrically bounded section
lie in C[X q ], Theorem 8.2 makes sense in the weighted case without extending the
domains of the operators dt in (8.4) beyond C[X q ]. Moreover the theorem remains
true for the weighted family of operators. The proof reduces immediately to the
unweighted case because the weighted and unweighted differentials, applied to a continuous and geometrically bounded section, differ by an O(t) term. The same applies
to Theorem 8.5.
9. Equivariant Fredholm Complexes
We shall assume from now on that a second countable, locally compact Hausdorff
topological group1 G acts on our CAT(0) cube complex X (preserving the cubical
structure). We shall not assume that G fixes any base point in X.
Our goal in this section to place the Julg-Valette and Pytlik-Szwarc complexes within
the context of equivariant Fredholm complexes, and we need to begin with some
definitions.
9.1 Definition. A Fredholm complex of Hilbert spaces is a bounded complex of
Hilbert spaces and bounded operators for which the identity morphism on the complex
is chain homotopic, through a chain homotopy consisting of bounded operators, to a
morphism consisting of compact Hilbert space operators.
In other words, a Fredholm complex of Hilbert spaces is a complex of the form
d
d
d
H0 Ð→ H1 Ð→ ⋯ Ð→ Hn ,
1
The topological restrictions on the group G are not really necessary, but they will allow us to
easily fit the concept of equivariant Fredholm complex into the context of Kasparov’s KK-theory in
the next section.
36
with each Hp a Hilbert space and each differential a bounded operator. Moreover
there exist bounded operators
h ∶ Hp Ð→ Hp−1
(p = 1, . . . , n)
such that each operator
dh + hd ∶ Hp Ð→ Hp
(p = 0, . . . , n)
is a compact perturbation of the identity operator.
The Fredholm condition implies that the cohomology groups of a Fredholm complex
are all finite-dimensional, which is the main reason for the definition. But we are
interested in the following concept of equivariant Fredholm complex, for which the
cohomology groups are not so relevant.
9.2 Definition. Let G be a second countable Hausdorff locally compact topological
group. A G-equivariant Fredholm complex of Hilbert spaces is a bounded complex of
separable Hilbert spaces and bounded operators for which
(a) Each Hilbert space carries a continuous unitary representation of G.
(b) The differentials d are not necessarily equivariant, but the differences d − gdg −1
are compact operator-valued and norm-continuous functions of g ∈ G.
(c) The identity morphism on the complex is chain homotopic, through a chain
homotopy consisting of bounded operators, to a morphism consisting of compact
Hilbert space operators.
(d) The operators h in the chain homotopy above are again not necessarily equivariant, but the differences h − ghg −1 are compact operator-valued and normcontinuous functions of g ∈ G.
9.3 Remark. Because the differentials are not necessarily equivariant, the cohomology groups of an equivariant Fredholm complex of Hilbert spaces do not necessarily
carry actions of G, and so are not of direct interest themselves as far as G is concerned.
Nevertheless the above definition, which is due to Kasparov (in a minor variant form),
has played an important role in a number of mathematical areas, most notably the
study of the Novikov conjecture in manifold topology [Kas88] (see [BCH94] for a
survey of other topics).
We are going to manufacture equivariant Fredholm complexes from the Julg-Valette
and Pytlik-Szwarc complexes. The Julg-Valette complex is the more difficult of the
two to understand. Disregarding the group action, the Julg-Valette differentials from
Definition 3.9 extend to bounded operators on the Hilbert space completions of the
cochain spaces associated to the inner products in (3.18), and the resulting complex
of Hilbert spaces and bounded operators is Fredholm, as in Definition 9.1. Moreover
37
the group G certainly acts unitarily. But the Julg-Valette differentials typically fail
to be G-equivariant, since they are defined using a choice of base point in the complex
X which need not be fixed by G. This means that the technical items (b) and (d) in
Definition 9.2 need to be considered carefully.
In fact to handle these technical items it will be necessary to finally make use of
the weight functions w(H) that we introduced in Definition 3.23. The following
computation will be our starting point. Assemble together all the Julg-Valette cochain
spaces so as to form the single space
dim(X)
C[X ] = ⊕ C[X q ],
●
q=0
and then form the Hilbert space completion
dim(X)
` (X ) = ⊕ `2 (X q ).
2
●
q=0
9.4 Lemma. For any weight function w(H) the Julg-Valette operator
D = d + δ ∶ `2 (X ● ) Ð→ `2 (X ● ),
viewed as a densely-defined operator with domain C[X ● ], is essentially self-adjoint.
Proof. The operator D is formally self-adjoint in the sense that
⟨Df1 , f2 ⟩ = ⟨f1 , Df2 ⟩
for all f1 , f2 ∈ C[X ● ]. The essential self-adjointness of D is a consequence of the fact
that the range of the operator
I + ∆ = I + D2
is dense in `2 (X ● ), and this in turn is a consequence of the fact that the Julg-Valette
Laplacian is a diagonal operator, as indicated in Proposition 3.25.
Since D is an essentially self-adjoint operator, we can study the resolvent operators
(D ± iI)−1 , which extend from their initial domains of definition (namely the ranges
of (D ± iI) on C[X ● ]) to bounded operators on `2 (X ● ).
9.5 Lemma. If w is a weight function that is proper in the sense that for every d > 0
the set { H ∶ w(H) < d } is finite, then the resolvent operators
(D ± iI)−1 ∶ `2 (X ● ) Ð→ `2 (X ● )
are compact Hilbert space operators.
38
Proof. The two resolvent operators are adjoint to one another, and so it suffices to
show that the product
(I + ∆)−1 = (D + iI)−1 (D − iI)−1
is compact. But the compactness of (I + ∆)−1 is clear from Proposition 3.25.
Let us now examine the dependence of the Julg-Valette operator D on the initial
choice of base point in X.
9.6 Lemma. If w is a weight function that is G-bounded in the sense that
sup ∣w(H) − w(gH)∣ < ∞
H
for every g ∈ G, then
∥D − g(D)∥ < ∞.
That is, the difference D − g(D), which is a linear operator on C[X ● ], extends to a
bounded linear operator on `2 (X ● ).
Proof. It suffices to prove the estimate for d in place of D = d + δ, since d and δ are
adjoint to one another. Now
dC − g(d)C = ∑ w(H)H ∧P0 C − ∑ w(g(H))H ∧g(P0 ) C,
H
H
where ∧P0 and ∧g(P0 ) denote the operators of Definition 3.7 associated to the two
indicated choices of base points. Since w(H) − w(gH) is uniformly bounded we can
replace w(g(H)) by w(H) in the second sum, and change the overall expression only
by a term that defines a bounded operator. So it suffices to show that for any pair of
base points P0 and P1 the expression
∑ w(H)(H ∧P0 C − H ∧P1 C)
H
defines a bounded operator. But the expression in parentheses is only non-zero when
H separates P0 from P1 , and there are only finitely many such hyperplanes. So the
lemma follows from the fact that for any hyperplane H the formula
H ∧P0 C − H ∧P1 C
defines a bounded operator, as long as the cube complex X has bounded geometry.
From now on we shall assume that the Julg-Valette complex is weighted using a proper
and G-bounded weight function. In fact, in the next section we shall work with the
specific weight function w∞ in (8.3), and so let us do the same here, even though it
39
is not yet necessary. Since the weighted Julg-Valette differential is not bounded, we
shall need to make an adjustment to fit the weighted complex into the framework of
Fredholm complexes of Hilbert spaces and bounded operators. We do this by forming
the normalized differentials
1
d′ = d(I + ∆)− 2 ∶ `2 (X q ) Ð→ `2 (X q+1 )
(where, strictly speaking, by d in the above formula we mean the closure of d in the
sense of unbounded operator theory). The normalized Julg-Valette complex is the
complex
(9.1)
d′
d′
d′
`2 (X 0 ) Ð→ `2 (X 1 ) Ð→ ⋯ Ð→ `2 (X n ).
1
It is indeed a complex because d and (I + ∆)− 2 commute with one another, and it is
a Fredholm complex because the adjoints d′∗ constitute a chain homotopy between
the identity and a compact operator-valued cochain map. In fact
d′ d′∗ + d′∗ d′ = D2 (I + D2 )−1 = I − (I + D2 )−1 ,
and (I + D2 )−1 is compact by Lemma 9.5.
We shall use the following computation from the functional calculus to show that the
normalized complex is an equivariant Fredholm complex of Hilbert spaces.
9.7 Lemma (Compare [BJ83]). If T is a positive, self-adjoint Hilbert space operator
that is bounded below by some positive constant, then
1
T −2 =
2 ∞ 2
(λ + T )−1 dλ
∫
π 0
The integral converges in the norm topology.
9.8 Theorem. The normalized Julg-Valette complex
d′
d′
d′
`2 (X 0 ) Ð→ `2 (X 1 ) Ð→ ⋯ Ð→ `2 (X n )
that is defined using the proper and G-bounded weight function w∞ in (8.3) is an
equivariant Fredholm complex.
Proof. It suffices to show that the normalized operator
D′ = D(I + D2 )−1/2 = d′ + d′∗
has the property that g(D′ ) − D′ is a compact operator-valued and norm-continuous
function of g ∈ G. For this we use Lemma 9.7 and the formula
1
D(λ2 + 1 + D2 )−1 = ((D + iµ)−1 + (D − iµ)−1 ),
2
40
where µ = (λ2 + 1)1/2 , to write the difference g(D′ ) − D′ as a linear combination of
two integrals
∞
−1
−1
∫ ((g(D) ± iµ) − (D ± iµ) ) dλ.
0
The integrand is
(9.2)
(g(D) ± iµ)−1 (D − g(D))(D ± iµ)−1 ,
which is a norm-continuous, compact operator valued function of λ ∈ [0, ∞) whose
norm is O(λ−2 ) as λ ↗ ∞. So the integrals converge to compact operators, as required.
Let us now examine the Pytlik-Szwarc complex. The inner products on the PytlikSzwarc cochain spaces given in Definition 5.12 are G-invariant, and the Pytlik-Szwarc
differentials given in Definition 5.8 are bounded and G-equivariant, so the story here
is much simpler.
9.9 Theorem. The Pytlik-Szwarc complex
`20 (X 0 ) Ð→ `20 (X 1 ) Ð→ ⋯ Ð→ `20 (X n )
is an equivariant Fredholm complex.
Proof. It follows from Proposition 5.14 that the formula
h=
1
p+1
]
δ ∶ C[Hqp ] Ð→ C[Hq−1
p+q
(we set h = 0 when p = q = 0) defines an exactly G-equivariant and bounded chain
homotopy between the identity and a compact operator-valued cochain map, namely
the orthogonal projection onto C[H00 ] ≅ C in degree zero, and the zero operator in
higher degrees.
To conclude this section we introduce the following notion of (topological, as opposed to chain) homotopy between two equivariant Fredholm complexes. In the next
section we shall construct a homotopy between the Julg-Valette and Pytilik-Szwarc
equivariant Fredholm complexes we constructed above using the continuous field of
complexes constructed in Section 8.
9.10 Definition. Two equivariant complexes of Hilbert spaces (H●0 , d0 ) and (H●1 , d1 )
are homotopic if there is a bounded complex of continuous fields of Hilbert spaces
over [0, 1] and adjointable families of bounded differentials for which
(a) Each continuous field carries a continuous unitary representation of G.
41
(b) The differentials d = {dt } are not necessarily equivariant, but the differences
d − gdg −1 are compact operator-valued and norm-continuous functions of g ∈ G.
(c) The identity morphism on the complex is chain homotopic, through a chain homotopy consisting of adjointable families of bounded operators, to a morphism
consisting of compact operators between continuous fields.
(d) The operators h = {ht } in the homotopy above are again not necessarily equivariant, but the differences h − ghg −1 are compact operator-valued and normcontinuous functions of g ∈ G.
(e) The restrictions of the complex to the points 0, 1 ∈ [0, 1] are the complexes
(H●0 , d0 ) and (H●1 , d1 ).
We need to supply definitions for the operator-theoretic concepts mentioned above.
These are usually formulated in the language of Hilbert modules, as for example in
[Lan95], but for consistency with the rest of this paper we shall continue to use the
language of continuous fields of Hilbert spaces.
9.11 Definition. An adjointable family of operators (soon we shall contract this to
adjointable operator ) between continuous fields {Ht } and {H′t } over the same compact
space T is a family of bounded operators
At ∶ Ht Ð→ H′t
that carries continuous sections to continuous sections, whose adjoint family
A∗t ∶ H′t Ð→ Ht
also carries continuous sections to continuous sections. An adjointable operator is
unitary if each At is unitary.
9.12 Definition. A representation of G as unitary adjointable operators on a continuous field {Ht } is continuous if the action map
G × { continuous sections } Ð→ { continuous sections }
is continuous. We place on the space of continuous sections the topology associated
to the norm ∥σ∥ = max ∥σ(t)∥.
9.13 Definition. An adjointable operator A = {At } between continuous fields of
Hilbert spaces over the same compact base space T is compact if it is the norm limit,
as a Banach space operator
A ∶ { continuous sections } Ð→ { continuous sections },
of a sequence of linear combinations of operators of the form
σ z→ ⟨σ1 , σ⟩σ2 ,
42
where σ1 and σ2 are continuous sections (of the domain and range continuous fields,
respectively). The compact operators form a closed, two-sided ideal in the C ∗ -algebra
of all adjointable operators.
Here, then, is the theorem that we shall prove in the next section:
9.14 Theorem. The equivariant Fredholm complexes obtained from the Julg-Valette
and Pytlik-Szwarc complexes in Theorems 9.8 and 9.9 are homotopic (in the sense of
Definition 9.10).
10. K-Amenability
The purpose of this section is to prove Theorem 9.14. But before giving the proof, we
shall explain the K-theoretic relevance of the theorem. To do so we shall need to use
the language of Kasparov’s equivariant KK-theory [Kas88], but we emphasize that
the proof of Theorem 9.14 will involve only the definitions from the last section our
work ealier in the paper. We shall assume familiarity with Kasparov’s theory.
A G-equivariant complex of Hilbert spaces, as in Definition 9.2, determines a class in
Kasparov’s equivariant representation ring
R(G) = KKG (C, C),
in such a way that
(a) homotopic complexes, as in Definition 9.10, determine the same element,
(b) a complex whose differentials are exactly G-equivariant determines the same
class as the complex of cohomology groups (these are finite-dimensional unitary
representations of G) with zero differentials, and
(c) a complex with the one-dimensional trivial representation in degree zero, and
no higher-dimensional cochain spaces, determines the multiplicative identity
element 1 ∈ R(G).
10.1 Definition. See [JV84, Definition 1.2]. A second countable and locally compact
Hausdorff topological group G is K-amenable if the multiplicative identity element
1 ∈ R(G) is representable by an equivariant Fredholm complex of Hilbert spaces
H0 Ð→ H1 Ð→ ⋯ Ð→ Hn
in which the each cochain space Hp , viewed as a unitary representation of G, is weakly
contained in the regular representation of G.
10.2 Theorem (See [JV84, Corollary 3.6].). If G is K-amenable, then the natural
homomorphism of C ∗ -algebras
∗
∗
Cmax
(G) Ð→ Cred
(G)
43
induces an isomorphism of K-theory groups
∗
∗
K∗ (Cmax
(G)) Ð→ K∗ (Cred
(G)).
10.3 Remarks. The C ∗ -algebra homomorphism in the theorem is itself an isomorphism if and only if the group G is amenable; this explains the term K-amenable. Not
every group is K-amenable; for example an infinite group with Kazhdan’s property
T is certainly not K-amenable, because the K-theory homomorphism is certainly not
an isomorphism.
After having quickly surveyed this background information, we can state the main
result of this section:
10.4 Theorem. If a second countable and locally compact group G admits a proper
action on a bounded geometry CAT(0) cube complex, then G is K-amenable.
The theorem was proved by Julg and Valette in [JV84] in the case where the cube
complex is a tree. They used the Julg-Valette complex, as we have called it, for a
tree, and showed that the continuous field of complexes that we have constructed in
this paper is a homotopy connecting the Julg-Valette and Pytlik-Szwarc complexes.
We shall do the same in the general case. The construction of this homotopy proves
the theorem in view of the following simple result, whose proof we shall omit.
10.5 Lemma. Assume that a second countable and locally compact group G acts
proper action on a CAT(0) cube complex. The Hilbert spaces in the Julg-Valette
complex are weakly contained in the regular representation of G.
10.6 Remark. Theorem 10.4 is not new; it was proved by Higson and Kasparov in
[HK01, Theorem 9.4] using a very different argument that is both far more general
(it applies to a much broader class of groups) and far less geometric.
To prove Theorem 10.4 it therefore suffices to prove Theorem 9.14, and this is what
we shall now do.
We shall construct the homotopy that the theorem requires by modifying the constructions in Section 8 in more or less the same way that we modified the Julg-Valette
complex to construct the complex (9.1). We shall therefore be applying the functional
calculus to the family of operators
(10.1)
Dt = Ut∗ (dwt + δwt )Ut ∶ `2t (X ● ) Ð→ `2t (X ● ),
where dwt is the Julg-Valette differential associated to the weight function in (8.3),
and of course δwt is the adjoint differential. To apply the functional calculus we shall
need to know that the family of resolvent operators
(Dt + iλ)−1 ∶ `2t (X ● ) Ð→ `2t (X ● )
44
carries continuous sections to continuous sections. This is a consequence of the following result:
10.7 Proposition. Let λ be a nonzero real number. The family of operators
{ (Dt + iλI)−1 ∶ `2t (X ● ) → `2t (X ● ) }t∈[0,∞]
carries the space of continuous and geometrically bounded sections to a dense subspace
of the space of continuous and geometrically bounded sections in the norm ∥s∥ =
supt∈[0,∞] ∥s(t)∥`2t (X ● ) .
Actually we shall need a small variation on this proposition:
10.8 Definition. Denote by P = {Pt } the operator that is in each fiber the orthogonal
projection onto the span of the single basic q-cochain fP0 ,P0 of type p = 0 (of course
this basic cochain is just P0 ).
It follows from the formula for the Julg-Valette Laplacian in Proposition 3.20 that
the operators Pt + ∆t are essentially self-adoint and bounded below by 1. So we can
form the resolvent operators (Dt + Pt + iλI)−1 for any λ ∈ R, including λ = 0.
10.9 Proposition. Let λ be any real number (possibly zero). The family of operators
{ (Dt + Pt + iλI)−1 ∶ `2t (X ● ) → `2t (X ● ) }t∈[0,∞]
carries the space of continuous and geometrically bounded sections to a dense subspace
of the space of continuous and geometrically bounded sections.
Both propositions will be proved by examining action of the Laplacians
(10.2)
∆t = Dt2 = Ut∗ (dwt + δwt )2 Ut
on continuous and geometrically controlled sections of the field {`2t (X ● }t∈[0,∞] .
Proof of Propositions 10.7 and 10.9. The family of operators {Dt } maps the space of
continuous, geometrically bounded sections into itself, so we can consider the compositions
∆t + λ2 I = (Dt + iλI)(Dt − iλI)
and
∆t + Pt + λ2 I = (Dt + Pt + iλI)(Dt + Pt − iλI),
and it suffices to show that the families of these operators map the space continuous
and geometrically bounded sections into a dense subspace of itself.
45
Let fC,D ∈ C[X q ] be a basic q-cochain of type p. Lemmas 8.3 and 8.4 tell us that
t−p Ut fC,D = (−1)p F + O(t),
where the q-cube F has the property that there are precisely p hyperplanes adjacent
to it that separate it from the base point P0 . So according to our formula for the
Julg-Valette Laplacian in Proposition 3.20,
(dwt + δwt )2 Ut ∶ t−p fC,D z→ (p + q) ⋅ (−1)p F + O(t)
and so, by applying Ut∗ to both sides we get
(∆2t + λ2 I) ∶ t−p fC,D z→ ((p + q) + λ2 ) ⋅ t−p fC,D + O(t).
Similarly
(∆t +Pt +λ2 I) ∶ t−p fC,D z→ (max{1, (p + q)} + λ2 ) ⋅ t−p fC,D + O(t).
So the ranges of the families {∆t +λ2 I} and {∆t +Pt +λ2 I} contain O(t) perturbations
of every basic section. The propositions follow from this.
Now form the bounded self-adjoint operators
1
Ft = Dt (Pt + Dt2 )− 2 .
By the above and Lemma 9.7 the family {Ft }t∈[0,∞] maps continuous sections to
continuous sections. So we can consider the bounded complex of continuous fields of
Hilbert spaces over [0, 1] and bounded adjointable operators
(10.3)
{`2t (X 0 )}t∈[0,∞]
{d′t }t∈[0,∞]
/ {`2 (X 1 )}t∈[0,∞]
t
{d′t }t∈[0,∞]
/
⋯
{d′t }t∈[0,∞]
/
{`2t (X n )}t∈[0,∞]
in which each differential {d′t } is the component of {Ft } mapping between the indicated continuous fields.
10.10 Proposition. Disregarding the G-action, the complex (10.3) is a homotopy of
Fredholm complexes.
Proof. If we set ht = d′t ∗ , then
ht d′t + d′t ht = ∆t (Pt + ∆t )−1 = I − Pt (Pt + ∆t )−1 ,
and {Pt (Pt +∆t )−1 }, is compact operator on the continuous field {`2t (X ● )}t∈[0,∞] .
46
It remains show that (10.3) is an equivariant homotopy. If the resolvent families
{(Dt + Pt + iλI)−1 } were compact, then we would be able to follow the route taken
in the previous section to prove equivariance of the Fredholm complex associated to
the Julg-Valette complex. But compactness fails at t = 0, and so we need to be a bit
more careful. The following two propositions will substitute for the Lemmas 9.5 and
9.6 that were used to handle the Julg-Valette complex in the previous section.
10.11 Proposition. For every ε > 0 and for every λ ∈ R the restricted family of
operators
{(Dt + Pt ± iλ)−1 }t∈[ε,∞]
is a compact operator on the continuous field {`2t (X ● )}t∈[ε,∞] . Moreover
∥(Dt + Pt ± iλ)−1 ∥ ≤ ∣1 + iλ∣−1
for all t and all λ.
10.12 Proposition. For every g ∈ G the operators Dt − g(Dt ) are uniformly bounded
in t:
sup ∥Dt − g(Dt )∥ < ∞
t∈[0,∞]
Moreover
∥Dt − g(Dt )∥ = O(t).
as t → 0.
Taking these for granted, for a moment, here is the result of the calculation:
10.13 Theorem. The complex (10.3) is a homotopy of equivariant Fredholm complexes in the sense of Definition 9.10.
Proof. We need to check that the families of differentials {d′t } in the complex (10.3)
are G-equivariant modulo compact operators, and also that {g(d′t )} varies normcontinuously with g ∈ G.
Let us discuss norm-continuity first. If g is sufficiently close to the identity in G, then
g fixes the base point P0 , and for such g we have g(d′t ) = d′t for all t. So {g(d′t )} is
actually locally constant as a function of g.
The proof of equivariance modulo compact operators is a small variation of the proof
of Theorem 9.8. It suffices to show that the family of operators {g(Ft )−Ft } is compact.
Since
1
Ft = Dt (Pt + ∆t )− 2
1
= (Pt + Dt )(Pt + ∆t )− 2 + compact operator,
47
it suffices to prove that the operator
1
Et = (Pt + Dt )(Pt + ∆t )− 2
is equivariant modulo compact operators. Applying Lemma 9.7 we find that
2 ∞
(Pt + Dt )(λ2 I + Pt + ∆t )−1 dλ
π ∫0
1 ∞
= ∫ ((Dt + Pt − iλ)−1 + (Dt + Pt + iλ)−1 ) dλ
π 0
Et =
So the difference g(Et ) − Et is the sum of the two integrals
(10.4)
1 ∞
((g(Dt ) + g(Pt ) ± iλ)−1 − (Dt + Pt ± iλ)−1 ) dλ
∫
π 0
Now the integrands in (10.4) can be written as
(10.5) (g(Pt ) + g(Dt ) ± iλ)−1 (Dt − g(Dt ))(Pt + Dt ± iλ)−1
+ (g(Pt ) + g(Dt ) ± iλ)−1 (Pt − g(Pt ))(Pt + Dt ± iλ)−1
Both terms in (10.5) are norm-continuous, compact operator valued functions of λ ∈
[0, ∞), the first by virtue of Proposition 10.12 and the second because Pt is compact.
Moreover the norms of both are O(λ−2 ) as λ → ∞. So the integrals in (10.4) converge
to compact operators, as required.
It remains to prove Propositions 10.11 and 10.12. The first is easy and we can deal
with it immediately.
Proof of Proposition 10.11. We want to show that the family of operators
{Kt }t∈[ε,∞] = {(Dt + Pt ± iλ)−1 }t∈[ε,∞]
is compact. Since the compact operators form a closed, two-sided ideal in the C ∗ algebra of all adjointable families of operators it suffices to show that the family
{Kt∗ Kt }t∈[ε,∞] = {(∆t + Pt + λ2 )−1 }t∈[ε,∞]
is compact; compare [Ped79, Proposition 1.4.5]. Conjugating by the unitaries Ut it
suffices to prove that the family
{(dwt δwt + δwt dwt + Pt + λ2 )−1 }t∈[ε,∞]
on the constant field of Hilbert spaces with fiber `2 (X ● ) is compact; this is one of the
things that restricting to t ∈ [ε, ∞] makes possible. But this final assertion is a simple
48
consequence of the explicit formula for the Julg-Valette Laplacian in Proposition 3.20,
together with the fact that the weight functions wt are uniformly proper in t ∈ [ε, ∞]
in the sense that for every N , all but finitely many hyperplanes H satisfy wt (H) ≥ N
for all t ∈ [ε, ∞].
As for the norm estimate in the proposition, this holds not just for ∆t + Pt but for
any self-adjoint operator bounded below by 1, and is elementary.
Let us turn now to Proposition 10.12. A complicating factor is that G not only fails to
preserve the Julg-Valette differential, but also fails to preserve the unitary operators
Ut that appear in the definitions of the differentials dt . The proposition is only correct
because the two failures to a certain extent cancel one another out.
10.14 Definition. Let P and Q be vertices in X. Define a unitary operator
̂t (Q, P )∶ `2 (X q ) Ð→ `2 (X q )
W
̂t (Q, P ) to be the cocycle operator Wt (Q, P )
as follows. When q = 0, we define W
̂t (Q, P ) respects the decomposition of `2 (X q )
of Definition 6.13. On higher cubes, W
according to parallelism classes, and on a summand determined by a given class we
̂t (Q, P ) = Wt (CQ , CP ), where CQ and CP are the cubes in the equivalence class
set W
nearest to Q and P .
It is immediate from the definition of the unitary operator Ut in Definition 6.15 that
(10.6)
g(Ut ) = Wt (Q0 , P0 )Ut ∶ `2t (X ● ) → `2 (X ● )
From this and the definition of Dt we find that
(10.7)
̂t (Q, P )Ut .
̂t (Q, P )∗ (g(dwt ) + g(δwt ))W
g(Dt ) = Ut∗ W
̂t ∶= W
̂t (Q, P ) and write
Now let us use the abbreviation W
̂t∗ (g(dwt ) + g(δwt ))W
̂t )Ut
Dt − g(Dt ) = Ut∗ ((dwt + δwt ) − W
The right-hand side can be rearranged as
Ut∗ Wt∗ (Wt dwt − g(dwt )Wt )Ut + Ut∗ (δwt Wt∗ − Wt∗ g(δwt ))Wt Ut
and the norm of this expression is no more than
̂t dwt − g(dwt )W
̂t ∥ + ∥δwt W
̂t∗ − W
̂t∗ g(δwt )∥
∥W
So it suffices to show that the operators
(10.8)
̂t dwt − g(dwt )W
̂t
W
̂t∗ − W
̂t∗ g(δwt )
and δwt W
49
satisfy the conclusions of Proposition 10.12. The second operator is adjoint to the
first. So in fact it suffices to prove the conclusions of Proposition 10.12 for the first
operator alone. This is what we shall do.
Before we proceed, let us adjust our notation a bit, as follows. Given a vertex P
in X, we shall denote by dP,wt the Julg-Valette differential that is defined using the
base vertex P and the weight function (8.3), for whose definition we also use the base
vertex P rather than P0 . With this new notation we can drop further mention of the
group G: Proposition 10.12 is a consequence of the following assertion:
10.15 Proposition. The operator
̂t (Q, P )∶ C[X q ] Ð→ C[X q+1 ]
̂t (Q, P )dP,wt − dQ,wt W
W
is bounded for all t > 0, and moreover
̂t (Q, P )dP,wt − dQ,wt W
̂t (Q, P )∥ = 0.
lim∥W
t→0
Recall now that the Julg-Valette differential is defined using the operation H ∧ C
between hyperplanes and cubes. Since the operation depends on a choice of base
vertex, we shall from now on write H ∧P C to indicate that choice, as we did earlier.
To prove Proposition 10.15 it suffices to consider the case where P and Q are at
distance 1 from one another (so they are separated by a unique hyperplane). We
shall make this assumption from now on.
10.16 Lemma. If a hyperplane H fails to separate P from Q, then
̂t (P, Q)D = W
̂t (P, Q)(H ∧Q D)
H ∧P W
for all oriented q-cubes D.
Proof. First, if H fails to separate P from Q, then the operators H∧P and H∧Q are
equal to one another. We shall drop the subscripts for the rest of the proof.
Next, if H cuts D, then it cuts all the cubes parallel to D, and therefore it cuts all
̂t (P, Q)D. So both sides of the equation in the lemma are
the cubes that make up W
zero. So can assume from now on that H is disjoint from D.
Let K be the hyperplane that separates Q from P . According to Proposition 4.6
the nearest q-cubes to P and Q in the parallelism class of D are either equal or are
̂t (P, Q)D is either
opposite faces, across K, of a (q+1)-cube that is cut by K. So W
just D or is a combination
(10.9)
̂t (P, Q)D = aD + bE
W
50
of D and another cube E that is an opposite face from D in a (q+1)-cube that is cut
by K.
We see that if H fails to separate D from P , or equivalently, if it fails to separate D
̂t (P, Q)D from P or Q,
from Q, then it also fails to separate any of the terms in W
and accordingly both sides of the equation in the lemma are zero. So we can assume
from now on that H does separate D from P and Q.
Suppose now that K fails to be adjacent to D, either because it cuts D or because
some vertex of D is not adjacent to K. The left-hand side of the equation is then
H ∧ D. This is either zero, in which case the equation obviously holds, or it is a
(q+1)-cube to which K also fails to be adjacent, in which case the right-hand side of
the equation is simple H ∧ D. So we can assume that K is adjacent to D.
Let E be the q-cube that is separated from D by K alone, as in (10.9). Since H fails
to separate D from E, or P from Q, but separates D and E from P and Q, we see
from Lemma 2.8 that H and K intersect. By Lemma 2.7, if H is adjacent to either
of D or E, then there is a (q+2)-cube that is cut by H and K and contains both D
and E as faces. In this case both sides of the equation in the lemma are
aH ∧ D + bH ∧ E
with a and b as in (10.9). Finally, if H is adjacent to neither D nor E, then both
sides of the equation are zero.
10.17 Lemma. If H separates P from Q, then
̂t (P, Q)D − W
̂t (P, Q)(H ∧Q D) = f (t)H ∧Q D − g(t)H ∧P D,
H ∧P W
where f and g are smooth, bounded functions on [0, ∞) that vanish at t = 0.
Proof. If D fails to be adjacent to H, then both sides in the displayed formula are
zero. So suppose D is adjacent to H. In this case
̂t (P, Q)(H ∧Q D) = H ∧Q D.
W
Now according to the definitions
̂t (P, Q)D = ±e− 21 t2 E + (1 − e−t2 ) 21 D
W
where E is the q-cube opposite D across H, and where the sign is +1 if D is separated
from P by H, and −1 if it is not. We find then that
̂t (P, Q)D = ±e− 12 t2 H ∧P E + (1 − e−t2 ) 12 H ∧P D.
H ∧P W
51
But H ∧P E = 0 if E is not separated from P by H, which is to say if D is separated
from P by H. So we can write
̂t (P, Q)D = −e− 12 t2 H ∧P E + (1 − e−t2 ) 12 H ∧P D.
H ∧P W
In addition
H ∧P E = −H ∧Q D
so that
̂t (P, Q)D = e− 21 t2 H ∧Q D + (1 − e−t2 ) 21 H ∧P D.
H ∧P W
Finally we obtain
̂t (P, Q)(H ∧Q D) − H ∧P W
̂t (P, Q)D = (e 21 t2 − 1)H ∧Q D − (1 − e−t2 ) 21 H ∧P D,
W
as required.
Proof of Proposition 10.15. We shall use the previous lemmas and the formula
dP,wt D = ∑ wP,t (H) H ∧P D,
H
for the Julg-Valette differential. We get
̂t (Q, P )dP,wt − dQ,wt W
̂t (Q, P )
(10.10) W
̂t (Q, P )D).
̂t (Q, P ) (H ∧P D) − wQ,t (H)H ∧Q W
= ∑(wP,t (H)W
H
Let us separate the sum into a part indexed by hyperplanes that do not separate P
from Q, followed by the single term indexed by the hyperplane H0 that does separate
P from Q. According to Lemma 10.16 the first part is
̂t (Q, P ) (H ∧P D) .
∑ (wP,t (H) − wQ,t (H)) W
H≠H0
Inserting the definition of the weight function, we obtain
(10.11)
̂t (Q, P ) (H ∧P D) ,
t ∑ (dist(H, P ) − dist(H, Q)) W
H≠H0
and moreover
∣dist(H, P ) − dist(H, Q)∣ ≤ 1.
As for the part of (10.10) indexed by H, keeping in mind that
dist(H0 , P ) =
1
2
= dist(H0 , Q),
we obtain from Lemma 10.17 the following formula for it:
(10.12)
(1 + 21 t)f (t)H0 ∧Q D − (1 + 21 t)g(t)H0 ∧P D,
where f and g are bounded and vanish at 0. The required estimates follow, because
the terms in (10.11) and (10.12) are uniformly bounded in number, are supported
uniformly close to D, are uniformly bounded in size, and vanish at t = 0.
52
References
[BCH94]
P. Baum, A. Connes, and N. Higson. Classifying space for proper actions
and K-theory of group C ∗ -algebras. In C ∗ -algebras: 1943–1993 (San Antonio, TX, 1993), volume 167 of Contemp. Math., pages 240–291. Amer.
Math. Soc., Providence, RI, 1994.
[BdlHV08] B. Bekka, P. de la Harpe, and A. Valette. Kazhdan’s property (T), volume 11 of New Mathematical Monographs. Cambridge University Press,
Cambridge, 2008.
[BJ83]
S. Baaj and P. Julg. Théorie bivariante de Kasparov et opérateurs non
bornés dans les C ∗ -modules hilbertiens. C. R. Acad. Sci. Paris Sér. I
Math., 296(21):875–878, 1983.
[CCJ+ 01]
P.-A. Cherix, M. Cowling, P. Jolissaint, P. Julg, and A. Valette.
Groups with the Haagerup property.
Modern Birkhäuser Classics.
Birkhäuser/Springer, Basel, 2001. Gromov’s a-T-menability, Paperback
reprint of the 2001 edition [ MR1852148].
[Cun83]
J. Cuntz. K-theoretic amenability for discrete groups. J. Reine Angew.
Math., 344:180–195, 1983.
[Dix77]
J. Dixmier. C ∗ -algebras. North-Holland Publishing Co., Amsterdam,
1977. Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15.
[For98]
R. Forman. Morse theory for cell complexes. Adv. Math., 134(1):90–145,
1998.
[GH10]
E. Guentner and N. Higson. Weak amenability of CAT(0)-cubical groups.
Geom. Dedicata, 148:137–156, 2010.
[Haa79]
U. Haagerup. An example of a nonnuclear C ∗ -algebra, which has the
metric approximation property. Invent. Math., 50(3):279–293, 1978/79.
[HK01]
N. Higson and G. G. Kasparov. E-theory and KK-theory for groups
which act properly and isometrically on Hilbert space. Invent. Math.,
144(1):23–74, 2001.
[JV83]
P. Julg and A. Valette. K-moyennabilité pour les groupes opérant sur les
arbres. C. R. Acad. Sci. Paris Sér. I Math., 296(23):977–980, 1983.
[JV84]
P. Julg and A. Valette. K-theoretic amenability for SL2 (Qp ), and the
action on the associated tree. J. Funct. Anal., 58(2):194–215, 1984.
53
[JV86]
P. Julg and A. Valette. Twisted coboundary operator on a tree and the
Selberg principle. J. Operator Theory, 16(2):285–304, 1986.
[JV87]
P. Julg and A. Valette. L’opérateur de co-bord tordu sur un arbre, et le
principe de Selberg. II. J. Operator Theory, 17(2):347–355, 1987.
[Kas88]
G. G. Kasparov. Equivariant KK-theory and the Novikov conjecture.
Invent. Math., 91(1):147–201, 1988.
[Kaž67]
D. A. Každan. On the connection of the dual space of a group with the
structure of its closed subgroups. Funkcional. Anal. i Priložen., 1:71–74,
1967.
[Lan95]
E. C. Lance. Hilbert C ∗ -modules: A toolkit for operator algebraists, volume 210 of London Mathematical Society Lecture Note Series. Cambridge
University Press, Cambridge, 1995.
[NR98a]
G. A. Niblo and L. D. Reeves. The geometry of cube complexes and the
complexity of their fundamental groups. Topology, 37(3):621–633, 1998.
[NR98b]
G. A. Niblo and M. A. Roller. Groups acting on cubes and Kazhdan’s
property (T). Proc. Amer. Math. Soc., 126(3):693–699, 1998.
[Ped79]
G. K. Pedersen. C ∗ -algebras and their automorphism groups, volume 14
of London Mathematical Society Monographs. Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], London-New York, 1979.
[Pim87]
M. V. Pimsner. Cocycles on trees. J. Operator Theory, 17(1):121–128,
1987.
[PS86]
T. Pytlik and R. Szwarc. An analytic family of uniformly bounded representations of free groups. Acta Math., 157(3-4):287–309, 1986.
[Sag95]
M. Sageev. Ends of group pairs and non-positively curved cube complexes.
Proc. London Math. Soc. (3), 71(3):585–617, 1995.
[Val90]
A. Valette. Cocycles d’arbres et représentations uniformément bornées.
C. R. Acad. Sci. Paris Sér. I Math., 310(10):703–708, 1990.
Department of Mathematical Sciences, University of Southampton, Southampton
SO17 1BJ, UK
E-mail address: [email protected]
Department of Mathematical Sciences, University of Hawaii at Manoa, Honolulu,
HI 96822, USA
54
E-mail address: [email protected]
Department of Mathematics, Penn State University, University Park, PA 16802, USA
E-mail address: [email protected]
55
| 4math.GR
|
arXiv:1711.10968v1 [cs.CV] 29 Nov 2017
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
1
Colour Constancy: Biologically-inspired
Contrast Variant Pooling Mechanism
Arash Akbarinia1
1
Centre de Visió per Computador
Universitat Autònoma de Barcelona
Barcelona, Spain
2
Department of Information and
Communication Technologies
Universitat Pompeu Fabra
Barcelona, Spain
www.cvc.uab.es/people/sakbarinia/
Raquel Gil Rodríguez2
http://ip4ec.upf.edu/user/53
1
C. Alejandro Parraga
www.cvc.uab.es/people/aparraga/
1
Abstract
Pooling is a ubiquitous operation in image processing algorithms that allows for
higher-level processes to collect relevant low-level features from a region of interest.
Currently, max-pooling is one of the most commonly used operators in the computational
literature. However, it can lack robustness to outliers due to the fact that it relies merely
on the peak of a function. Pooling mechanisms are also present in the primate visual cortex where neurons of higher cortical areas pool signals from lower ones. The receptive
fields of these neurons have been shown to vary according to the contrast by aggregating
signals over a larger region in the presence of low contrast stimuli. We hypothesise that
this contrast-variant-pooling mechanism can address some of the shortcomings of maxpooling. We modelled this contrast variation through a histogram clipping in which the
percentage of pooled signal is inversely proportional to the local contrast of an image.
We tested our hypothesis by applying it to the phenomenon of colour constancy where a
number of popular algorithms utilise a max-pooling step (e.g. White-Patch, Grey-Edge
and Double-Opponency). For each of these methods, we investigated the consequences
of replacing their original max-pooling by the proposed contrast-variant-pooling. Our
experiments on three colour constancy benchmark datasets suggest that previous results
can significantly improve by adopting a contrast-variant-pooling mechanism.
Introduction
Many computer vision frameworks contain a pooling stage that combines local responses
at different spatial locations [8]. This operation is often a sum- or max-pooling mechanism
implemented by a wide range of algorithms, e.g. in feature descriptors (such as SIFT [37] and
HOG [12]) or convolutional neural networks [36, 43]. Choosing the correct pooling operator
can make a great difference in the performance of a method [8]. Current standard pooling
mechanisms lack the desired generalisation to find equilibrium between frequently-occurring
and rare but informative descriptors [39]. Therefore, many computer vision applications can
benefit from a more dynamic pooling solution that takes into account the content of pooled
signals.
c 2017. The copyright of this document resides with its authors.
It may be distributed unchanged freely in print or electronic forms.
2
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
Such pooling operators are commonly used in modelling the phenomenon of colour
constancy (i.e. a visual effect that makes the perceived colour of surfaces remain approximately constant under changes in illumination [19, 28]) both in biological [10] and computational [34] solutions. Despite decades of research, colour constancy still remains as an
open question [19], and solutions to this phenomenon are important from a practical point
of view, e.g. camera manufacturers need to produce images of objects that appear the same
as the actual objects in order to satisfy customers. Motivated by above, in this article we
propose a contrast-variant-pooling mechanism and investigate its feasibility in the context of
computational colour constancy.
1.1
Computational Models of Colour Constancy
Mathematically, the recovery of spectral reflectance from a scene illuminated by light of
unknown spectral irradiance is an ill-posed problem (it has infinite possible solutions). The
simplest and most popular solution has been to impose some arbitrary assumptions regarding
the scene illuminant or its chromatic content. Broadly speaking colour constancy algorithms
can be divided into two categories: (i) low-level driven, which reduce the problem to solving a set of non-linear mathematical equations [23, 34, 46], and (ii) learning-based, which
train machine learning techniques on relevant image features [1, 18, 21]. Learning-based
approaches may offer the highest performance results, however, they have major setbacks
which make them unsuitable in certain conditions: (a) they rely heavily on training data that
is not easy to obtain for all possible situations, and (b) they are likely to be slow [24] and unsuitable for deployment inside limited hardware. A large portion of low-level driven models
can be summarised using a general Minkowski expression [15, 46] such as:
Z
Lc (p) =
p
[ fc (x)] dx
1
p
= kec ,
(1)
Ω
where L represents the illuminant at colour channel c ∈ {R, G, B}; f (x) is the image’s pixel
value at the spatial coordinate x ∈ Ω; p is the Minkowski norm; and k is a multiplicative
constant chosen such that the illuminant colour, e, is a unit vector.
Distinct values of the Minkowski norm p results into different pooling mechanisms. Setting p = 1 in Eq. 1 reproduces the well known Grey-World algorithm (i.e. sum-pooling), in
which it is assumed that all colours in the scene average to grey [9]. Setting p = ∞ replicates
the White-Patch algorithm (i.e. max-pooling), which assumes the brightest patch in the image corresponds to the scene illuminant [34]. In general, it is challenging to automatically
tune p for every image and dataset. At the same time inaccurate p values may corrupt the
results noticeably [24].
The Minkowski framework can be generalised further by replacing f (x) in Eq. 1 with its
higher-order derivatives [46]. These non-linear solutions are analogous to centre-surround
mechanisms of visual cortex [35], which is typically modelled by a Difference-of-Gaussians
(DoG) operators where a narrower, positive Gaussian plays the role of the “centre” and
a broader, negative Gaussian plays the role of the “surround” [14, 38]. Recently, a few
biologically-inspired models of colour constancy grounded on DoG have offered promising
results [23, 41], however their efficiency largely depends on finding an optimum pooling
strategy for higher cortical areas. In short, pooling is a crucial component of many colour
constancy models driven by low-level features (or even in deep-learning solutions [7, 20]).
In the primate visual systems the size of the receptive field varies according to the local
3
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
contrast of the light falling on it [4, 45] presenting a dynamic solution dependent on the
region of interest.
The low-level models mentioned above are related to the early stages of visual processing, i.e. the primary visual cortex (area V1), that are likely to be involved in colour constancy.
Physiological measures suggest that although receptive fields triple in size from area V1 to
area V2 [48], their basic circuity with respect to surround modulation is similar, i.e. keeping
the same size dependency with contrast properties found in V1 [45]. This is consistent with
the large body of physiological and psychophysical literature highlighting the significance
of contrast in the visual cortex. In computer vision, contrast-dependent models have also
shown encouraging results in various applications such as visual attention [29], tone mapping [42], and boundary detection [2, 3], to name a few. From these we can hypothesise the
convenience and explanatory value of various “pooling strategies” such as those proposed by
previous colour constancy methods. In the rest of this work we will explore the advantages
of replacing the different feed-forward (pooling) configurations of some successful colour
constancy models [23, 34, 46] by that of the visual system (as described by current neurophysiological literature [4, 45]). Our aim in doing so is dual, on the one hand we want to
explore the technological possibilities of creating a more efficient algorithm and on the other
hand we would like to test the idea that contrast-variant-pooling might play an important role
in colour constancy.
1.2
Summary of the Contributions
1
1
0
1 0
Figure 1: Flowchart of the proposed contrast-variant-pooling (CVP) mechanism in the context of colour constancy. We implemented CVP through a top-x-percentage-pooling. Given
an input image or a feature map: (i) value of the x is computed according to the inverse
of local contrast for each channel, and (ii) we estimate the scene illuminant as the average
value of the top-x-percentage of pooled pixels (those on the right side of the depicted dashed
lines).
In the present article we propose a generic contrast-variant-pooling (CVP) mechanism
that can replace standard sum- and max-pooling operators in a wide range of computer vision applications. Figure 1 illustrates the flowchart of CVP, which is based on local contrast and therefore it offers a dynamic solution that adapts the pooling mechanism accord-
4
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
ing to the content of region of interest. We tested the feasibility of CVP in the context
of colour constancy by substituting the pooling operation of four algorithms: (a) WhitePatch [34], (b) first-order Grey-Edge [46], (c) second-order Grey-Edge [46], and (d) DoubleOpponency [23] in three benchmark datasets. Results of our experiments show the quantitative and qualitative benefits of CVP over max-pooling.
2
2.1
Model
Max-pooling Colour Constancy
One of the earliest computational models of colour constancy (White-Patch) is grounded on
the assumption that the brightest pixel in an image corresponds to a bright spot or specular
reflection containing all necessary information about the scene illuminant [34]. Mathematically this is equivalent to a max-pooling operation over the intensity of all pixels:
Lc = arg max Ic (x, y),
(2)
x,y
where L represents the estimated illuminant at each chromatic channel c ∈ {R, G, B}; I is the
original image and (x, y) are the spatial coordinates in the image domain.
One important flaw of this simple approach is that a single bright pixel can misrepresent
the whole illuminant. Furthermore, the White-Patch algorithm may fail in the presence of
noise or clipped pixels in the image due to the limitations of the max-pooling operator [22].
One approach to address these issues is to account for a larger set of “white” points by pooling a small percentage of the brightest pixels (e.g. the top 1%) [13], an operation referred as
top-x-percentage-pooling. In this manner, the pooling mechanism is collectively computed
considering a group of pixels rather than a single one. This small variant might be a crucial
factor in the estimation of the scene illuminant [32]. A similar mechanism has also been
deployed successfully in other applications such as shadow removal [17].
In practice, given the chosen x-percentage, the top-x-percentage-pooling can be implemented through a histogram-based clipping mechanism [13, 17]. Let H be the histogram of
the input image I, and let Hc (i) represents number of pixels in colour channel c with intensity
i ∈ [0, · · · , Kc ] (histogram’s domain). The scene illuminant Lc is computed as:
Lc =
1
Nkc
Kc
∑ i · Hc (i),
(3)
i=kc
where Nkc is the total number of pixels within the intensity range kc to Kc . The values of
kc={R,G,B} are determined from the chosen x-percentage such that:
Kc
Nkc =
∑ Hc (i) = xc · P,
(4)
i=kc
where P is the total number of pixels in the image and xc is the chosen percentage for colour
channel c. Within this formulation it is very cumbersome to define a universal, optimallyfixed percentage of pixels to be pooled [13] and consequently the free variable x requires
specific tuning for each image or dataset individually. Naturally, this limitation restricts the
usability of the top-x-percentage-pooling operator. In the following sections we show how
to automatically compute this percentage, based on the local contrast of each image.
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
2.2
5
Pooling Mechanisms in the Visual Cortex
We know from the physiology of the cerebral cortex that neurons in higher cortical areas pool
information from lower areas over increasingly larger image regions. Although the exact
pooling mechanism is yet to be discovered, “winner-takes-all” and “sparse coding” kurtotical
behaviour are common to many groups of neurons all over the visual cortex [10, 40], and
it is conceivable that a mechanism analogous to max-pooling might be present within the
cortical layers. Indeed such behaviour has been discovered across a population of cells in the
cat visual cortex [33] and the activation level of cells with max-like behaviour was reported
to vary depending on the contrast of visual stimuli.
Results reported by [33] suggest an inverse relationship between the contrast of a stimulus and the percentage of the signal pooled. When pooling neurons were exposed to low
contrast stimuli, their responses shifted slightly away from pure max-pooling (selecting the
highest activation response within a region) towards integrating over a larger number of
highly activated neurons. In the language of computer vision, this can be regarded as topx-percentage-pooling, where x assumes a smaller value in high contrast and a larger value
in low contrast. Interestingly, the pooling of those neurons remained always much closer
to max-pooling than to the linear integration of all neurons (sum-pooling) [33]. Mathematically, this can be interpreted as having a very small (top-x-percentage-pooling) x value.
It does not come as a great surprise that the pooling mechanism in the visual cortex
depends on the stimulus’ contrast. There is a large body of physiological studies showing
that receptive fields (RF) of neurons are contrast variant (for a comprehensive review refer
to [4]). Quantitative results suggest that RFs in visual area one (V1) of awake macaques
double their size when measured at low contrast [45]. Similar expansions have also been
reported for the RFs of neurons in extrastriate areas, such as V2 and V4. This implies that a
typical neuron in higher cortical areas that normally pool responses from its preceding areas
over about three neighbouring spatial locations [48] can access a substantially larger region
to pool from in the presence of low contrast stimuli. This is in line with the reported pooling
mechanism in the cat visual cortex [33].
2.3
Contrast Variant Pooling
In order to model this contrast-variant-pooling mechanism, we first computed local contrast
C of the input image I at every pixel location by means of its local standard deviation defined
as:
s
2
1
Cc (x, y; σ ) =
Ic (x0 , y0 ) − µ(x, y)
(5)
∑
#Nσ (x, y) (x0 ,y0 )∈N (x,y)
σ
where c indexes each colour channel; (x, y) are the spatial coordinates of a pixel; µ is an
isotropic average kernel with size σ ; and Nσ (x, y) represents the neighbourhood centred on
pixel (x, y) of radius σ .
To simulate this inverse relation between stimulus contrast and percentage of signal
pooled [33] in the top-x-percentage-pooling operator, we determined the percentage xc in
Eq. 4 as the average of inverse local contrast signals:
xc =
1
1
, ∀(x, y) ∈ Ω,
P∑
C
(x,
y; σ )
c
x,y
(6)
where Cc is computed from Eq. 5, and Ω is the spatial image domain. In this fashion, instead
of defining a fix percentage of signal (pixels) to be pooled (as in [17]), we chose an adaptive
6
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
percentage according to the contrast of image. In terms of colour constancy, this effectively
relates the number of pooled pixels to compute the scene illuminant to the average contrast
of an image. We illustrated this mechanism of contrast-variant-pooling in the central panel
of Figure 1, where red, green and blue signals correspond to the histogram of each chromatic
channel. Pixels on the right side of the dashed lines (kc ) are pooled. In this example, contrast
is higher for the red signal and therefore a smaller percentage of cells are pooled in the red
channel.
Bearing in mind that “contrast” is just a fraction in the range [0, 1] – with 0 characterising
an absolutely uniform area and 1 representing points with the highest contrast, e.g. edges –
we can predict that the percentage x will be a very small number for natural images where
homogeneous regions are likely to form the majority of the scene. Consequently in our topx-percentage-pooling operator we always pool a small percentage. This is in agreement with
observations of [33] which indicate that such pooling mechanism is always much closer to
max-pooling than to sum-pooling.
2.4
Generalisation to Other Colour Constancy Models
There is a number of colour constancy models in the literature which are driven by low-level
features that require a pooling mechanism on top of their computed feature maps in order to
estimate the scene illuminant. In the Double-Opponency [23] algorithm this feature map is
computed by convolving a colour-opponent representation of the image with a DoG kernel
followed by a max-pooling operation. In the Grey-Edge [46] model, higher-order derivatives
of the image are calculated through its convolution with the first- and second-order derivative
of a Gaussian kernel. This is complemented by a Minkowski summation, which oscillates
between sum- and max-pooling depending on its norm.
Similar to the White-Patch algorithm, the pooling mechanism of these models can also
be replaced with our top-x-percentage-pooling operator, where x is computed according to
the local contrast of image as explained above. The only difference is that instead of pooling
from an intensity image (as in case of the White-Patch algorithm), Double-Opponency and
Grey-Edge pool over their respective feature maps. This means that Eq. 3 receives a feature
map M instead of an intensity image I as input.
3
Experiments and Results
In order to investigate the efficiency of our model, we applied the proposed contrast-variantpooling (CVP) mechanism to four different colour constancy algorithms whose source code
were publicly available: White-Patch [34], first-order Grey-Edge [46], second-order GreyEdge [46], and Double-Opponency [23]. We simply replaced their max-pooling operator
with our proposed pooling mechanism. To evaluate each method we used the recovery angular error defined as:
< ee , et >
◦
ε (ee , et ) = arccos
,
(7)
kee kket k
where < . > represents the dot product of the estimated illuminant ee and the ground truth et ;
and k.k stands for the Euclidean norm of a vector. It is worth mentioning that this error measure might not correspond precisely to observers’ preferences [47], however, it is the most
commonly used comparative measure in the literature. We also computed the reproduction
7
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
angular error [16] in all experiments (due to lack of space these results are not reported here).
Readers are encouraged to check the accompanying supplementary materials.
We conducted experiments on three benchmark datasets1 , (i) SFU Lab (321 images) [6],
(ii) Colour Checker (568 images) [44], and (iii) Grey Ball (11,346 images) [11]. In Table 1
we have reported the best median and trimean angular errors for each of the considered
methods (these metrics were proposed by [27] and [25] respectively to evaluate colour constancy algorithms since they are robust to outliers). Mean angular errors are reported in the
supplementary materials.
SFU Lab [6]
Colour Checker [44]
Grey Ball [11]
Median Trimean Median Trimean Median Trimean
White-Patch
[34]
6.44
7.51
5.68
6.35
6.00
6.50
Grey-Edge 1st -order [46]
3.52
4.02
3.72
4.76
5.01
5.80
Grey-Edge 2nd -order [46]
3.22
3.65
4.27
5.19
5.72
6.39
Double-Opponency [23]
2.37
3.27
3.46
4.38
4.62
5.28
CVP White-Patch
2.99
3.42
2.97
3.45
5.02
5.15
CVP Grey-Edge 1st -order
3.29
3.72
2.48
2.79
4.70
5.17
CVP Grey-Edge 2nd -order 2.85
3.13
2.59
2.93
4.65
5.05
CVP Double-Opponency
2.02
2.56
2.39
2.84
4.00
4.24
Table 1: Recovery angular errors of four colour constancy methods under max- and contrastvariant-pooling (CVP) on three benchmark datasets. Lower figures indicate better performance.
Method
Figure 2 illustrates three exemplary results obtained by the proposed contrast-variantpooling (CVP) operator for two colour constancy models: White-Patch and the first-order
Grey-Edge. Qualitatively, we can observe that CVP does a better job than max-pooling in
estimating the scene illuminant. This is also confirmed quantitatively for the angular errors,
shown at the right bottom side of each computed output.
9.73◦
1.83◦
6.73◦
2.59◦
11.39◦
1.52◦
10.31◦
2.34◦
8.45◦
1.45◦
7.18◦
1.69◦
Original
Ground Truth
WP
CVP WP
GE1
CVP GE1
Figure 2: Qualitative results of White-Patch (WP) and the first-order Grey-Edge (GE1) under
max- and contrast-variant-pooling (CVP). Angular errors are indicated on the bottom right
corner of each panel. Images are from the SFU Lab, Colour Checker and Grey Ball dataset
respectively.
1 All
source code and materials are available in the supplementary submission.
8
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
3.1
Influence of the Free Parameters
For each free variable of the tested models we compared the performance of max- to contrastvariant-pooling. White-Patch does not have any free variable, therefore it is exempted from
this analysis. In Figure 3 we have reported the impact of different σ s (receptive filed size) on
Double-Opponency algorithm for the best and the worst results obtained by free variable k in
each dataset (results for all ks are available in the supplementary material). We can observe
that almost in all cases contrast-variant-pooling outperforms max-pooling. The improvement
is more tangible for the Colour Checker and Grey Ball datasets and in low σ s.
10
10
10
9
9
9
8
8
8
7
7
7
6
6
6
5
5
5
4
4
4
3
3
3
2
2
2
1
1
1
0
1
2
3
4
5
6
7
0
8
1
2
3
4
5
6
7
0
8
1
2
3
4
5
6
7
8
SFU Lab [6]
Colour Checker [44]
Grey Ball [11]
Figure 3: The best and the worst results obtained by max- and contrast-variant-pooling for
the free variables of Double-Opponency [23] algorithm (k and σ ).
Figure 4 illustrates the impact of different σ s (Gaussian size) on the first- and secondorder Grey-Edge algorithm. We can observe similar patterns as with Double-Opponency
(contrast-variant-pooling outperforms max-pooling practically in all cases). This improvement is more significant for low σ s, for the Colour Checker dataset and for the second-order
derivative. It must be noted that the objective of this article was merely to study the performance of max-pooling and CVP on top of the Grey-Edge algorithm. However, the angular
errors of our CVP Grey-Edge happen to be in par with the best results reported for Grey-Edge
(obtained by using the optimum Minkowski norm for each dataset [46]), with the important
caveat that CVP has no extra variables to be tuned, whereas in the Minkowski norm optimisation the value of p must be hand-picked for each dataset.
9
9
9
8
8
8
7
7
7
6
6
5
5
5
4
4
4
3
3
6
3
2
2
1
2
3
4
5
6
1
2
3
4
5
6
2
1
2
3
4
5
6
SFU Lab [6]
Colour Checker [44]
Grey Ball [11]
Figure 4: Comparison of max- and contrast-variant-pooling for the free variable σ of GreyEdge [46] algorithm (both the first- and second-order derivatives).
From Figures 3 and 4 we can observe that the greatest improvement occurs in the Colour
Checker dataset. We speculate that one of the reasons for this is the larger range of intensity
values in the Colour Checker dataset (16-bit) in comparison to the other two datasets that
contain 8-bit images, therefore, an inaccurate max-pooling is more severely penalised.
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
3.2
9
Discussion
We would like to emphasise that the objective of this article is not to improve state-of-the-art
in colour constancy, but to show that contrast-variant-pooling (CVP) almost always produces
improvements over max-pooling. Surprisingly, the results we obtained are even competitive
with the state-of-the-art. For instance, in the SFU Lab dataset, the lowest reported angular
error is 2.1 (obtained by an Intersection-based Gamut algorithm [5]). This means that our
CVP Double-Opponency with an angular error of 2.0 outperforms the state-of-the-art in this
dataset. In the Colour Checker and Grey Ball datasets there are a few learning-based models
(e.g. Exemplar-based method [31]) that obtain lower angular errors in comparison to CVP
Double-Opponency, nevertheless our results are comparable with theirs.
Physiological evidence besides, the better performance of CVP can be explained intuitively by the fact that max-pooling relies merely on the peak of a function (or a region of
interest), whereas in our model, pooling is defined collectively based on a number of elements near the maximum. Consequently, those peaks that are outliers and likely caused by
noise get normalised by other pooled elements. The rationale within our model is to pool a
larger percentage at low contrast since in those conditions, peaks are not informative on their
own, whereas at high contrast peaks are likely to be more informative and other irrelevant
details must be removed (therefore a smaller percentage is pooled).
Although, the importance of choosing an appropriate pooling type has been demonstrated
both experimentally [30, 49], and theoretically [8], current standard pooling mechanisms
lack the desired generalisation [39]. We believe that contrast-variant-pooling can offer a
more dynamic and general solution. In this article, we evaluated CVP on the colour constancy phenomenon as a proof-of-concept, however our formulation of CVP is generic (and
based on local contrast) and in principle it can be applied to a wider range of computer vision
algorithms, such as deep-learning, where pooling is a decisive factor [43].
In our implementation of CVP, we approximated local contrast through local standard
deviation (see Eq. 5). There are at least two factors that require a more profound analysis: (i)
incorporating more sophisticated models of contrast perception [26] accounting for extrema
in human contrast sensitivity; and (ii) analysing the role of kernel size in the computation of
local contrast.
4
Conclusion
In this article, we presented a novel biologically-inspired contrast-variant-pooling (CVP)
mechanism grounded in the physiology of the visual cortex. Our main contribution can be
summarised as linking the percentage of pooled signal to the local contrast of the stimuli, i.e.
pooling a larger percentage at low contrast and a smaller percentage at high contrast. Our
CVP operator remains always closer to max-pooling rather than to sum-pooling since natural
images generally contain more homogeneous areas than abrupt discontinuities.
We tested the efficiency of our CVP model in the context of colour constancy by replacing the max-pooling operator of four algorithms with the proposed pooling. After that, we
conducted experiments in three benchmark datasets. Our results show that contrast-variantpooling outperforms the commonly used max-pooling operator nearly in all cases. This can
be explained by the fact that our model allows for more informative peaks to be pooled while
suppressing less informative peaks and outliers.
We further argued that the proposed CVP is a generic operator, thus its application can
10
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
be extended to a wider range of computer vision algorithms by offering a dynamic (and
automatic) framework that is based on the local contrast of an image or a pixel. This opens a
multitude of possibilities for future lines of research and it remains an open question whether
our model can reproduce its excellent results in other domains as well. Therefore, it certainly
is interesting to investigate whether our CVP can improve convolutional neural networks.
Acknowledgements
This work was funded by the Spanish Secretary of Research and Innovation (TIN201341751-P and TIN2013-49982-EXP).
References
[1] Vivek Agarwal, Andrei V Gribok, and Mongi A Abidi. Machine learning approach to
color constancy. Neural Networks, 20(5):559–563, 2007.
[2] Arash Akbarinia and C Alejandro Parraga. Biologically-inspired edge detection
through surround modulation. In Proceedings of the British Machine Vision Conference, pages 1–13, 2016.
[3] Arash Akbarinia and C Alejandro Parraga. Feedback and surround modulated boundary
detection. International Journal of Computer Vision, pages 1–14, 2017.
[4] Alessandra Angelucci and S Shushruth. Beyond the classical receptive field: Surround
modulation in primary visual cortex. The new visual neurosciences, pages 425–444,
2013.
[5] Kobus Barnard. Improvements to gamut mapping colour constancy algorithms. In
Computer Vision–ECCV 2000, pages 390–403. Springer, 2000.
[6] Kobus Barnard, Lindsay Martin, Brian Funt, and Adam Coath. A data set for color
research. Color Research & Application, 27(3):147–151, 2002.
[7] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379–387, 2015.
[8] Y-Lan Boureau, Jean Ponce, and Yann LeCun. A theoretical analysis of feature pooling in visual recognition. In Proceedings of the international conference on machine
learning (ICML), pages 111–118, 2010.
[9] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal
of the Franklin institute, 310(1):1–26, 1980.
[10] Matteo Carandini and David J Heeger. Normalization as a canonical neural computation. Nature Reviews Neuroscience, 13(1):51–62, 2012.
[11] Florian Ciurea and Brian Funt. A large image database for color constancy research.
In Color and Imaging Conference, volume 2003, pages 160–164, 2003.
[12] Navneet Dalal and Bill Triggs. Histograms of oriented gradients for human detection.
In Computer Vision and Pattern Recognition (CVPR), volume 1, pages 886–893, 2005.
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
11
[13] Marc Ebner. Color constancy, volume 6. John Wiley & Sons, 2007.
[14] Christina Enroth-Cugell and John G Robson. The contrast sensitivity of retinal ganglion cells of the cat. The Journal of physiology, 187(3):517–552, 1966.
[15] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In
Color and Imaging Conference, volume 2004, pages 37–41, 2004.
[16] Graham D Finlayson and Roshanak Zakizadeh. Reproduction angular error: An improved performance metric for illuminant estimation. perception, 310(1):1–26, 2014.
[17] Graham D Finlayson, Steven D Hordley, and Mark S Drew. Removing shadows from
images. In Computer Vision–ECCV 2002, pages 823–836. Springer, 2002.
[18] David A Forsyth. A novel algorithm for color constancy. International Journal of
Computer Vision, 5(1):5–35, 1990.
[19] David H Foster. Color constancy. Vision research, 51(7):674–700, 2011.
[20] Damien Fourure, Rémi Emonet, Elisa Fromont, Damien Muselet, Alain Trémeau, and
Christian Wolf. Mixed pooling neural networks for color constancy. In Image Processing (ICIP), 2016 IEEE International Conference on, pages 3997–4001. IEEE, 2016.
[21] Brian Funt and Weihua Xiong. Estimating illumination chromaticity via support vector
regression. In Color and Imaging Conference, volume 2004, pages 47–52. Society for
Imaging Science and Technology, 2004.
[22] Brian Funt, Kobus Barnard, and Lindsay Martin. Is machine colour constancy good
enough? In Computer Vision–ECCV’98, pages 445–459. Springer, 1998.
[23] Shao-Bing Gao, Kai-Fu Yang, Chao-Yi Li, and Yong-Jie Li. Color constancy using
double-opponency. Pattern Analysis and Machine Intelligence, IEEE Transactions on,
37(10):1973–1985, 2015.
[24] Arjan Gijsenij and Theo Gevers. Color constancy using natural image statistics and
scene semantics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33
(4):687–698, 2011.
[25] Arjan Gijsenij, Theo Gevers, and Marcel P Lucassen. Perceptual analysis of distance
measures for color constancy algorithms. JOSA A, 26(10):2243–2256, 2009.
[26] Andrew M Haun and Eli Peli. Perceived contrast in complex images. Journal of vision,
13(13):3–3, 2013.
[27] Steven D Hordley and Graham D Finlayson. Reevaluation of color constancy algorithm
performance. JOSA A, 23(5):1008–1020, 2006.
[28] Paul M Hubel. The perception of color at dawn and dusk. Journal of Imaging Science
and Technology, 44(4):371–375, 2000.
[29] Laurent Itti and Christof Koch. Computational modelling of visual attention. Nature
reviews neuroscience, 2(3):194–203, 2001.
12
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
[30] Kevin Jarrett, Koray Kavukcuoglu, Yann LeCun, et al. What is the best multi-stage
architecture for object recognition? In International Conference on Computer Vision
(ICCV), pages 2146–2153, 2009.
[31] Hamid Reza Vaezi Joze and Mark S Drew. Exemplar-based color constancy and multiple illumination. Pattern Analysis and Machine Intelligence, IEEE Transactions on,
36(5):860–873, 2014.
[32] Hamid Reza Vaezi Joze, Mark S Drew, Graham D Finlayson, and Perla Aurora Troncoso Rey. The role of bright pixels in illumination estimation. In Color and Imaging
Conference, volume 2012, pages 41–46, 2012.
[33] Ilan Lampl, David Ferster, Tomaso Poggio, and Maximilian Riesenhuber. Intracellular
measurements of spatial integration and the max operation in complex cells of the cat
primary visual cortex. Journal of neurophysiology, 92(5):2704–2713, 2004.
[34] Edwin H Land. The Retinex Theory of Color Vision. Scientific American offprints.
W.H. Freeman Company.
[35] Edwin H Land. An alternative technique for the computation of the designator in the
retinex theory of color vision. Proceedings of the National Academy of Sciences, 83
(10):3078–3080, 1986.
[36] Yann LeCun, Bernhard E Boser, John S Denker, Donnie Henderson, Richard E
Howard, Wayne E Hubbard, and Lawrence D Jackel. Handwritten digit recognition
with a back-propagation network. In Advances in neural information processing systems, pages 396–404, 1990.
[37] David G Lowe. Distinctive image features from scale-invariant keypoints. International
journal of computer vision, 60(2):91–110, 2004.
[38] David Marr and Ellen Hildreth. Theory of edge detection. Proceedings of the Royal
Society of London B: Biological Sciences, 207(1167):187–217, 1980.
[39] Naila Murray and Florent Perronnin. Generalized max pooling. In Computer Vision
and Pattern Recognition (CVPR), pages 2473–2480, 2014.
[40] Bruno A Olshausen et al. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381(6583):607–609, 1996.
[41] C. Alejandro Parraga and Arash Akbarinia. Colour constancy as a product of dynamic
centre-surround adaptation. Journal of Vision, 16(12):214–214, 2016.
[42] Erik Reinhard, Michael Stark, Peter Shirley, and James Ferwerda. Photographic tone
reproduction for digital images. In ACM Transactions on Graphics (TOG), volume 21,
pages 267–276. ACM, 2002.
[43] Dominik Scherer, Andreas Müller, and Sven Behnke. Evaluation of pooling operations in convolutional architectures for object recognition. Artificial Neural Networks–
ICANN 2010, pages 92–101, 2010.
[44] Lilong Shi and Brian Funt. Re-processed version of the gehler color constancy dataset
of 568 images. http://www.cs.sfu.ca/~colour/data/.
ARASH AKBARINIA: CONTRAST VARIANT POOLING MECHANISM
13
[45] S Shushruth, Jennifer M Ichida, Jonathan B Levitt, and Alessandra Angelucci. Comparison of spatial summation properties of neurons in macaque v1 and v2. Journal of
neurophysiology, 102(4):2069–2083, 2009.
[46] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy.
IEEE Transactions on image processing, 16(9):2207–2214, 2007.
[47] Javier Vazquez-Corral, C. Alejandro Parraga, Ramon Baldrich, and Maria Vanrell.
Color constancy algorithms: Psychophysical evaluation on a new dataset. Journal of
Imaging Science and Technology, 53(3):31105–1, 2009.
[48] Hugh R Wilson and Frances Wilkinson. Configural pooling in the ventral pathway. The
new visual neurosciences, pages 617–626, 2014.
[49] Jianchao Yang, Kai Yu, Yihong Gong, and Thomas Huang. Linear spatial pyramid
matching using sparse coding for image classification. In Computer Vision and Pattern
Recognition (CVPR), pages 1794–1801, 2009.
| 1cs.CV
|
1
Data Discovery and Anomaly Detection Using
Atypicality: Theory
arXiv:1709.03189v1 [cs.IT] 10 Sep 2017
Anders Høst-Madsen, Fellow, IEEE, Elyas Sabeti, Member, IEEE, Chad Walton
Abstract
A central question in the era of ’big data’ is what to do with the enormous amount of information. One possibility
is to characterize it through statistics, e.g., averages, or classify it using machine learning, in order to understand
the general structure of the overall data. The perspective in this paper is the opposite, namely that most of the value
in the information in some applications is in the parts that deviate from the average, that are unusual, atypical. We
define what we mean by ’atypical’ in an axiomatic way as data that can be encoded with fewer bits in itself rather
than using the code for the typical data. We show that this definition has good theoretical properties. We then develop
an implementation based on universal source coding, and apply this to a number of real world data sets.
Index Terms
Big Data, atypicality, minimum description length, data discovery, anomaly.
I. I NTRODUCTION
One characteristic of the information age is the exponential growth of information, and the ready availability
of this information through networks, including the internet – “Big Data.” The question is what to do with this
enormous amount of information. One possibility is to characterize it through statistics – think averages. The
perspective in this paper is the opposite, namely that most of the value in the information is in the parts that deviate
from the average, that are unusual, atypical. The rest is just background noise. Take art: the truly valuable paintings
are those that are rare and atypical. The same could be true for scientific research and entrepreneurship. Take online
collections of photos, such as Flickr.com. Most of the photos are rather pedestrian snapshots and not of interest
to a wider audience. The photos that of interest are those that are unique. Flickr has a collection of photos rated
for ’interestingness,’ and one can notice that those photos are indeed very different from typical photos. They are
atypical.
The aim of our approach is to extract such ’rare interesting’ data out of big data sets. The central question is
what ’interesting’ means. A first thought is to focus on the ’rare’ part. That is, interesting data is something that
is unlikely based on prior knowledge of typical data or examples of typical data, i.e., training. This is the way
A. Høst-Madsen and E. Sabeti are with the Department of Electrical Engineering, University of Hawaii Manoa, Honolulu, HI 96822 (e-mail:
{ahm,sabeti}@hawaii.edu). C. Walton is with the Department of Surgery, University of Hawaii, Honolulu, HI, 96813. Email: [email protected].
This work was supported in part by NSF grants CCF 1017823, 1017775, and 1434600. The paper was presented in part at IEEE Information
Theory Workshop 2013, Seville.
September 12, 2017
DRAFT
2
an outlier is usually defined. Unlikeliness could be measured in terms of likelihood, in terms of codelength [1],
[2] – called ’surprise’ in [3] – or according to some distance measure. This is also the most common principle in
anomaly detection [4]. However, perhaps being unlikely is not sufficient for something to be ’interesting.’ In many
cases, outliers are junk that are eliminated not to contaminate the typical data. What makes something interesting
is maybe that it has a new unusual structure in itself that is quite different from the structure of the data we
have already seen. Return to the example of paintings: what make masterworks interesting is not just that they are
different than other paintings, but that they have some ’structure’ that is intriguing. Or take another example. Many
scientific discoveries, like the theory of relativity and quantum mechanics, began with experiments that did not fit
with prevailing theories. The experiments were outliers or anomalies. What made them truly interesting was that it
was possible to find a new theory to explain the data, be it relativity or quantum mechanics. This is the principle
we pursue: finding data that have better alternative explanations than those that fit the typical data.
Something being unlikely is not even necessary for the data to be ’interesting.’ Suppose the typical data is
iid uniform {0, 1}. Then any sequence of bits are equally likely. Therefore, a sequence consisting of purely 1,
1111111 . . . is in no way ’surprising.’ Yet, it should catch our interest.
When we look for new interesting data, a characteristic is that we do not know what we are looking for. We are
looking for “unknown unknowns” [5]. Instead of looking at specific statistics of data, we need to use a universal
approach. This is provided by information theory.
This idea of finding alternative explanations for data rather than measuring some kind of difference from typical
data is what separates our method from usual approaches in outlier detection and anomaly detection. As far as
we can determine from reading hundreds of papers, our approach has not been explored previously. Obviously,
information theory and coding have been used in anomaly detection, data mining, and knowledge discovery before,
and we will discuss how this compares to our approach later. Our methodology also has connections to tests for
randomness, e.g., the run length test and [6], [7], but our aim is different.
A. Applications
Atypicality is relevant in large number of various applications. We will list a few applications here.
ECG. For electrocardiogram (ECG) recordings there are patterns in heart rate variability that are known to indicate
possible heart disease [8], [9], [10], [11]. With modern technology it is possible for an individual to wear and
unobtrusive heart rate monitor 24/7. If atypical patterns occur, it could be indicative of disease, and the individual
or a doctor could be notified. But perhaps a more important application is to medical research. One can analyze
a large collection of ECG recordings and look for individuals with atypical patterns. This can then potentially be
used to develop new diagnostic tools.
Genomics. Another example of application is interpretation of large collections of genomics data. Given that all
mammals have essentially the same set of genes, there must exist some significant differences that distinguish
the obvious distinct attributes between species, as well as more subtle differences within a species. Although the
genome has been mined by exhaustive studies applying a panoply of approaches, regions once thought to be
“uninteresting” have recently come under increased study for their potential role in defined morphological and
September 12, 2017
DRAFT
3
physiological differences between individuals [12]. Applying an atypical evaluation tool to genomic data from
individuals of known pathophysiological/morphological irregularities may provide valuable insight to the genetic
mechanisms underlying the condition.
Ocean Monitoring. In passive acoustic monitoring (PAM) [13] of oceans, one or more hydrophones is towed
behind a ship or deployed in a fixed bottom-mounted or suspended array in order to record vocalizations of marine
mammals. One major focus is to detect, and perhaps count, rare or endangered species. It would be highly interesting
to scan the data for any unusual patterns, which can then be further examined by a researcher.
Plant Monitoring. In for example nuclear plants, atypical monitoring data may be indicative of something about
to go wrong.
Computer Networks. Atypical network traffic could be indicative of a cyberattack. This is already being used
through anomaly detection [14]. However, an abstract atypicality approach can be used to find more subtle attacks
– the unknown unknowns.
Airport Security. Already software is being used to flag suspicious flyers, likely based on past attacks. Atypical
detection could be used to find innovative attackers.
Stock Market. Atypicality could be used to detect insider trading. It could also be used by investors to find unusual
stocks to invest in, promising outstanding returns – or ruin.
Astronomy. Atypicality can be used to scan huge databases for new kinds of cosmological phenomena.
Credit Card Fraud. Unusual spending patterns could be indicative of fraud. This is already used by credit card
companies, but obviously in a simple, and annoying way, as anyone who’s credit card has been blocked on an
overseas trip can testify to.
Gambling. Casinos are constantly fighting fraudsters. This is a game of cat and mouse. Fraudsters constantly find
new ways to trick the casinos (one such inventor was Shannon himself). Therefore, an abstract atypicality approach
may be the best solution to catch new ways of fraud.
B. Notation
We use x to denote a sequence in general, and xl when we need to make the length explicit; xi denotes a single
sample of the sequence. We use capital letters Xi to denote random variables rather than specific outcomes. Finally
X denotes a subsequence. All logarithms are to base 2 unless otherwise indicated.
II. ATYPICALITY
Our starting point is the in theory of randomness developed by Kolmogorov and Martin-Löf [15], [7], [16].
Kolmogorov divides (infinite) sequences into ’typical’ and ’special.’ The typical sequences are those that we
can call random, that is, they satisfy all laws of probability. They can be characterized through Kolmogorov
complexity. A sequence of bits {xn , n = 1, . . . , ∞} is random (i.e, iid uniform) if the Kolmogorov complexity of
the sequence satisfies K(x1 , . . . , xn ) ≥ n − c for some constant c and for all n[15]. The sequence is incompressible
if K(x1 , . . . , xn |n) ≥ n for all n, and a finite sequence is algorithmically random if K(x1 , . . . , xn |n) ≥ n [16]. In
terms of coding, an iid random sequence is also incompressible, or, put another way, the best coder is the identity
September 12, 2017
DRAFT
4
function. Let us assume we draw sequences xn from an iid uniform distribution. The optimum coder is the identity
function, and the code length is n. Now suppose that for one of these sequences we can find a (universal) coder
so that the code length is less than n; while not directly equivalent, one could state this as K(x1 , . . . , xn |n) < n.
With an interpretation of Kolmogorov’s terms, this would not be a ’typical’ sequence, but a ’special’ sequence. We
will instead call such sequences ’atypical.’ Considering general distributions and general (finite) alphabets instead
of iid uniform distributions, we can state this in the following general principle
Definition 1. A sequence is atypical if it can be described (coded) with fewer bits in itself rather than
using the (optimum) code for typical sequences.
This definition is central to our approach to the atypicality problem.
In the definition, the “(optimum) code for typical sequences,” is quite specific, following the principles in for
example [16]. We assume prefix free codes. Within that class the coding could be done using Huffman codes,
Shannon codes, Shannon-Fano-Elias codes, arithmetic coding etc. We care only about the code length, and among
these the variation in length is within a few bits, so that the code length for typical encoding can be quite accurately
calculated.
On the other hand, “described (coded) with fewer bits in itself” is less precise. In principle one could use
Kolmogorov complexity, but Kolmogorov complexity is not calculable and it is only given except for a constant,
and comparison with code length therefore is not an “apples-to-apples” comparison. Rather, some type of universal
source coder should be used. This can be given a quite precise meaning in the class of finite state machine sources,
[17] and following work, and is strongly related to minimum description length (MDL) [18], [19], [20], [17]. What
is essential is that we adhere to strict decodability at the decoder. The decoder only sees a stream of bits, and from
this it should be able to accurately reconstruct the source sequence. So, for example, if a sequence is atypical, there
must be a type of “header” telling the decoder to use a universal decoder rather than the typical decoder. Or, if
atypical sequences can be encoded in multiple ways, the decoder must be informed through the sequence of bits
which encoder was used. One could argue that such things are irrelevant for for example anomaly detection, since
we are not actually encoding sequences. The problem is that if such terms are omitted, it is far too easy to encode
a sequence “in itself.” This is like choosing a more complex model to fit data, without accounting for the model
complexity in itself, which is exactly what MDL sets out to solve, although also in this case actual encoding is
not done. We therefore try to account for all factors needed to describe data, and we believe this is one of the key
strengths of the approach.
A major difference between atypical data and anomalous data is that atypicality is an axiomatic property of data,
defined by Definition 1 based on Kolmogorov-Martin-Löf randomness. On the other hand, as far as we know, an
anomaly is not something that can be strictly defined. Usually, we think of an anomaly as something caused by an
outside phenomenon: an intruder in a computer network, a heart failure, a gambler playing tricks. This influences
how we think of performance. If a detector fails to give an indication of an anomaly, we have a miss (or type
II error), but if it gives an indication when such things are not happening we have a false alarm (type I error).
September 12, 2017
DRAFT
5
Atypicality, on the other hand, is purely a property of data. Ideally, there are therefore no misses or false alarms:
data is atypical or not. Here is what we mean. If there is an anomaly that expresses itself through the observed data,
that must mean that there is some structure in the data, and in theory a source coder would discover and exploit
such structure and reduce code length. Thus, if the data is not atypical that means there is simply no way to detect
the anomaly through the observations – again in theory. We therefore cannot really call that a miss. On the other
hand, suppose that in a casino a gambler has a long sequence of wins. This could be due to fraud, but it could
also be simply due to randomness. But casino security would be interested in either case for further scrutiny. Thus,
the reason for the atypicality does not really matter, the atypicality itself matters. Still, to distinguish the two cases
we call a sequence intrinsically atypical if it is atypical according to Definition 1 while being generated from the
typical probability model, while it is extrinsically atypical if it is in fact generated by any other probability law.
Definition 1 has two parts that work in concert, and we can write it simplified as Ct (x) − Ca (x) > 0 where Ct
is the typical codelength and Ca the atypical codelength. The typical code length Ct (x) is simply an expression
of the likelihood of seeing a particular sequence. If Ct (x) is large it means that the given sequence is unlikely to
happen, and detecting sequences by Ct (x) > τ would catch many outliers. As an extreme example, if a sequence
is impossible according to the typical distribution, Ct (x) = ∞, and it would always be caught. But it would
not work universally. If, as we started out with, typical sequences are iid uniform, any sequence is equally likely
and Ct (x) > τ would not catch any sequences. In this case, if a test sequence has some structure, it is possible
that Ca (x) < Ct (x), and such sequences would be caught by atypicality; thus calculating Ca (x) is essential.
Calculating Ct (x) is also essential. Suppose that we instead use E[Ct ] − Ca (x), where E[Ct ] is the code length
used to encode typical sequences “on average,” essentially the entropy rate. Again, this will catch some sequences:
if a test sequence has more or less structure than typical sequences, E[Ct ] − Ca (x) 6= 0. But again, it will omit
very obvious examples: if as test sequence we use a typical sequence with 0 and 1 swapped, E[Ct ] ≈ Ca (x), while
on the other hand Ct (x) > Ca (x). And impossible sequences with Ct (x) = ∞ would not be caught with absolute
certainty. Now, to declare something an outlier, we have to find a coder with Ca (x) < Ct (x). It is not sufficient
that Ct (x) is large, i.e., that the sequence is unlikely to happen. However, we can always use the trivial coder that
transmits data uncoded. If the sequence is unlikely to happen according to the typical distribution, then it is likely
that Ct (x) > (length of x).
Thus, it can be seen that the two parts work in concert to catch sequences. Each part might catch some sequences,
but to catch all “anomalies,” both parts have to be used.
Another point of view is the following. Suppose again the typical model is binary uniform iid. We look at a
collection of sequences, and now we want to find the most atypical sequences, i.e., the most “interesting” sequences.
Without a specification of what “interesting” is, it seems reasonable to choose those sequences that have the most
structure, and again this can reasonably be measured by how much the sequence can be compressed. This is what
Rissanen [17] calls “useful information,” U (x) = n − Ca (x). But again, we need to take into account the typical
model if it is not uniform iid. For example, if typical sequences have much structure, then sequence with little
structure might be more interesting. We therefore end up with that Ct (x) − Ca (x) is a reasonable measure of how
interesting sequences might be.
September 12, 2017
DRAFT
6
A. Alternative approaches
While, as argued in the introduction, and outlined above, what we are aiming for is not anomaly detection in
the traditional sense, there are still many similarities. And certainly information theory and universal source coding
has been used previously in anomaly detection, e.g., [4], [21], [22], [23], [24], [25], [26], [27], [28], [29], [30].
The approaches have mostly been heuristic. A more fundamental and systematic approach is Information Distance
defined in [31]. Without being able to claim that this applies to all of the perhaps hundreds of papers, we think the
various approaches can be summarized as using universal source coding as a type of distance measure, whether it
satisfies strict mathematical metric properties as in [31] or is more heuristic. On the other hand, our methodology in
Definition 1 cannot be classified as a distance measure in a traditional sense. We are instead trying to find alternative
explanations for data. We will comment on how our approach contrasts with a few other approaches.
While the similarity distance developed in [31] is not directly applicable to the problem we consider, we can to
some extent adapt it, which is useful for contrast. The similarity distance is
d=
min{K(y|x∗ ), K(x|y ∗ )}
max{K(x), K(y)}
Instead of being given the typical distribution, we can imagine that we are given a very long typical sequence x
which is used for “training.” In that case
d=
K(x|y ∗ )
K(x, y) − K(y)
=
K(x)
K(x)
within a certain approximation. Suppose, as was our starting point above, that the typical distribution is binary iid
uniform. If y is also binary iid uniform, within a constant K(x, y) = K(x) + K(y), and d = 1. But if y is drawn
from some other distribution, x cannot help describing x either, and still d = 1. That makes sense: two completely
random sequences are not similar, whether they are from the same distribution or not. Thus, similarity distance
cannot be used for ’anomaly’ detection as we have have defined it: looking for ’special’ sequences in the words of
Kolmogorov. This is not a problem of the similarity metric; it does exactly what it is designed for, which is really
deterministic similarity between sequences, appropriate for classification. The reason similarity distance still gives
results for anomaly detection [32] is actually that universal source coders approximate Kolmogorov complexity
poorly.
Heuristic methods using for anomaly detection using universal source coding [4], [21], [22], [23], [24], [25],
[26], [27], [28], [29], [30], [1], [2] are mostly based on comparing code length. Let C(x) be the code length to
encode the sequence x with a universal source coder. Let x be a training string and y a test sequence. We can
then compare
C(x)
|x|
with
C(y)
|y|
(which could be seen as a measure of entropy rate) or compare C(xy) with C(x) to
detect change. The issue with this is that there are many completely dissimilar sources that have the same entropy
rate. As an example, let the data be binary iid with the original source having P (X = 1) =
1
3
and the new source
P (X = 1) = 23 . Then the optimum code for the original source and the optimum code for the new source have the
same length. On the other hand, atypicality will immediately distinguish such sequences.
September 12, 2017
DRAFT
7
III. B INARY IID C ASE
In order to clarify ideas, at first we consider a very simple model. The typical model is iid binary with P (Xn =
1) = p. The alternative model class also binary iid but with P (Xn = 1) = θ, where θ is unknown. We want to
decide if a given sequence xl is typical or atypical. This can be stated as the hypothesis test problem
H0 :θ = p
H1 :θ 6= p
This problem does not have an UMP (universal most powerful) test. However, a common approach to solving this
type of problem is the GLRT (generalized likelihood ratio test) [33]. Let
P (b) = P (Xn = b)
P̂ (b) =
N (b|xl )
l
where l is the sequence length and N (b|xl ) is the number of xn = b ∈ {0, 1}. The GLRT is
Q1
l
P̂ (b)N (b|x )
L = log Qb=0
1
N (b|xl )
b=0 P (b)
=
1
X
b=0
=l
1
X
1
N (b|x ) log N (b|xl ) −
N (b|xl ) log P (b)
l
1
X
b=0
l
b=0
1
P̂ (b) log N (b|xl ) − l
l
1
X
P̂ (b) log P (b)
b=0
= lD(p̂kp)
1 L > t
φ(xl ) =
0 L ≤ t
Where D(p̂kp) =
P1
b=0
P̂ (b) log
P̂ (b)
P (b)
(1)
is the relative entropy [16] and t some threshold. While the GLRT is a
heuristic principle, it satisfies some optimality properties, and in this case it is equal to the invariant UMP test [34],
which can be considered an optimum solution under certain constraints. Thus, it is reasonably to take this as the
optimum solution for this problem, and we do not need to appeal to Kolmogorov or information theory to solve
the problem.
The complications start if we consider sequences of variable length l. The test (1) depends on the sequence length.
We need to choose a threshold t(l) as a function of l, which will then result in a false alarm probability PF A (t(l))
and detection probability PD (t(l)). There is no obvious argument for how to choose t(l) from a hypothesis testing
point of view; we could choose t independent of l, but that is just another arbitrary choice.
We will consider this problem in the context of Definition 1. In order to do so, we need to model the problem
from a coding point of view. We assume we have an (infinite) sequence of sequences of variable length li , and
these need to be encoded. We need to encode each bit, and also to encode whenever a new sequence starts. For
September 12, 2017
DRAFT
8
typical encoding of the bits we can use a Shannon code, Huffman code, arithmetic coding etc. The code length for
a sequence of length l is
1
1
Lt = N (1|xl ) log + N (0|xl ) log
p
1−p
1
1
= l p̂ log + (1 − p̂) log
(2)
p
1−p
P
xi . We also need to encode where a sequence ends
except for a small constant factor; here p̂ = P̂ (1) = 1l
and a new one starts. For simplicity let us for now assume lengths are geometrically distributed. We can then
model the problem as one with three source symbols ’0’, ’1’ and ’,’ with an iid distribution with P (0 ,0 ) = ,
P (0 00 ) = p − 2 , P (0 10 ) = (1 − p) − 2 . If we assume is small, the expression (2) is still valid for the content
part, and to each sequence is added a constant − log to encode separators. To decide if a sequence is atypical
according to Definition 1, we can use the universal source coder from [16]: the source encodes first the number of
ones k; then it enumerates the sequences with k ones, and transmits the index of the given sequence. For analysis
it is important to have a simple expression for the code length. We can therefore use La = lH(p̂) +
1
2
log l. This
is an approximation which is good for reasonably large l and it also reaches the lower bound in [17], [35]. The
source-coder also needs to inform the decoder that the following is an atypical sequence (so that it knows to use the
atypical decoder rather than the typical encoder), and where it ends. For the former we can use a ’.’ to indicate the
start of an atypical sequence rather than the ’,’ for typical sequences. If the probability that a sequence is atypical
is δ 1, P (0 .0 ) = δ and P (0 ,0 ) = (1 − δ) ≈ . The code length for a ’.’ now is − log − log δ. To mark the
end of the atypical sequence we could again insert a ’.’ or a ’,’. But the code for either is based on the distribution
of lengths of typical sequences, which we assume known, whereas we would have no knowledge of the length of
atypical sequences. Instead it seems more reasonable to encode the length of the specific atypical sequence. As
argued in [18], [36] this can be done with log∗ l + log c, where c is a constant and
log∗ (l) = log l + log log l + log log log l + · · ·
(3)
where the sum continues as long as the argument to the log is positive. To summarize we have
1
1
Lt = l p̂ log + (1 − p̂) log
− log
p
1−p
1
La = lH(p̂) + log l + log∗ l + log c − log − log δ
2
3
≈ lH(p̂) + log l − log + τ
2
τ = − log δ + log c
(4)
The criterion for a sequence to be atypical is La < Lt , which easily seen to be equivalent to
D(p̂kp) >
τ+
3
2
log l
(5)
l
If the lengths are fixed, this reduces to (1). But if the lengths are variable, (5) provides a threshold as a function
of l. The term
September 12, 2017
3
2
log l ensures that liml→∞ PF A (l) = 0, which seems reasonable. If instead D(p̂kp) >
τ
l
is used, it
DRAFT
9
is easy to see that liml→∞ PF A (l) > 0. Except for this property, the term
3
2?
3
2
log l might seem arbitrary, e.g., why
But it is based on solid theory, and as will be seen later it has several important theoretical properties.
We will examine the criterion (5) in more detail. The inequality (5) gives two thresholds for p̂,
p̂ > p+
p̂ < p−
Where 0 < p− < p < p+ < 1. It is impossible to find explicit expressions for p± , but it is clear that
p± → p
as l → ∞.
Therefore, for l large, we can replace D(p̂kp) with a series expansion. We then end up with the more explicit
criterion
(p − p̂)2
1
3
> (τ + log l)
pq ln 4
l
2
r
r
pq ln 4
3
·
τ + log l
|p̂ − p| > ∆τ =
l
2
(6)
In the following we will use this as it is considerably simpler to analyze. We can also write this as
Pl
√
i=1 xi − p
√
> 2τ ln 2 + 3 ln l
pql
(7)
Now, if not for the term 3 ln l, this would be a central limit type of statement, and the probability that a sequence
is classified as (intrinsically) atypical would be
PA ≈ 2Q
√
2 ln 2τ
(8)
independent of l. Our main interest is exactly the the dependency on l, which is given by the following Theorem
Theorem 2. Consider an iid {0, 1}-sequence. Let PA (l) be the probability that a sequence of length l is classified
as intrinsically atypical according to (6). Then PA (l) is bounded by
PA (l) ≤2−τ +1
1
K(l, τ )
l3/2
(9)
∀τ : lim K(l, τ ) =1
l→∞
For p =
1
2
this can be strengthened to
1
l3/2
(10)
ln PA (l)
=1
l→∞ − 3 ln l
2
(11)
PA (l) ≤2−τ +1
These bounds are tight in the sense that
lim
Proof: The Chernoff bound (e.g., [37]) states
PA (l) = P (|p̂ − p| > ∆τ ) =2P
l
X
!
Xi ≥ lp + b
i=1
≤2 inf e−lsp−sb MX (s)l
s>0
September 12, 2017
DRAFT
10
Where (as usual, q = 1 − p)
r
b=
r
lpq ln 4 τ +
3
log l
2
and MS (s) is the moment generating function of Xi , which for a Bernoulli random variable is
MX (s) =pes + q
Then
n
o
1
l
PA (l) ≤ inf exp (−s (pl + b)) (pes + q)
s>0
2
Minimizing over s gives
1
PA (l) ≤
2
lq
lq − b
l
q(lp + b)
p(lq − b)
−lp−b
or
q(lp + b)
+ (−lp − b) ln
p(lq − b)
b
b
=l ln 1 +
+ (−lp − b) ln 1 +
lq − b
p(lq − b)
2
2 2
2
2
b 3l q p + lb(7p − 6p − 3) + b (6p + 3)
≤
6p2 (b − lq)3
1
ln PA (l) ≤l ln
2
≤−
lq
lq − b
For p =
for p =
1
2
1
2
x2
2
b2
b3
+ O (1) 2
2lpq
l
= − τ ln 2 −
where we have used x −
≤ ln(1 + x) ≤ x −
3
ln3/2 l
ln l + O (1) √ τ 3/2 ,
2
l
x2
2
+
x3
3
(12)
for x ≥ 0. The equation (12) directly leads to (9).
Hoeffding’s inequality [38] gives the bound
b2
PA (l) ≤ 2 exp −2
l
3
= 2 exp − ln 2 τ + log l
2
(13)
this is tighter than (12).
For the lower bound we use moderate deviations from [39]. Define X̃i =
Pl
i=1 X̃i
p
>1
l(2τ ln 2 + 3 ln l)
We define al =
1
2τ ln 2+3 ln l ,
Xi −p
√
pq .
We can then rewrite (7) as
which satisfies liml→∞ al = 0, liml→∞ lal = ∞. Using this as al in [39, Theorem
3.7.1] gives
1
lim inf
ln P
l→∞ 2τ ln 2 + 3 ln l
= lim inf
l→∞
≥−
September 12, 2017
1
ln P
3 ln l
Pl
i=1
p
X̃i
l(2τ ln 2 + 3 ln l)
!
Pl
X̃
i
i=1
p
>1
l(2τ ln 2 + 3 ln l)
!
>1
1
2
DRAFT
11
Together with the upper bound, this gives (11).
Figure 1 compares the upper bound with simulations.
−1
10
Upper bound
Simulation
−2
10
−3
PA
10
−4
10
−5
10
−6
10
−7
10
1
10
2
3
10
4
10
10
n
Fig. 1. Simulated PA and the Upper bound for τ = 1, p = 0.3.
We can also bound the miss probability for extrinsically atypical sequences as follows
Theorem 3. Suppose that the typical sequence is iid {0, 1}-sequence with P (Xn = 1) = p. Let the test sequence
by iid with P (Xn = 1) = pa . The probability that the test sequence is missed according to criterion (6) is upper
bounded by
−τ
PM (l) ≤2
1
qa p
pa q
√lpq(2τ ln 2+3 ln l)
l3/2
p−1 p+1 −l
qa q
×
K(l, τ )
ppa pp
(14)
∀τ : lim K(l, τ ) =1
l→∞
Proof: We may assume that pa < p. Similarly to the proof of Theorem 2 the Chernoff bound is
PM (l) ≤ inf
s>0
n
o
l
exp (−s (pl − b)) (pa es + qa )
Minimizing over s gives
PM (l) ≤
September 12, 2017
lqa
lq + b
l
qa (lp − b)
pa (lq + b)
−lp+b
DRAFT
12
or
lqa
qa (lp − b)
ln PM (l) ≤l ln
+ (−lp + b) ln
lq + b
pa (lq + b)
qa
qa
q
≤l ln
− lp ln
− lp ln
q
pa
p
3
p
qa
b2
b
+ b ln
+ ln
−
+O 2
q
pa
2lpq
l
using series expansions.
A. Hypothesis testing interpretation
The solution (5) may seem arbitrary, but it has a nice interpretation in terms of hypothesis testing [40]. Return
to the solution (1). That solution gives a test for a given l. However, the problem is that it does not reconcile tests
for different l. One way to solve that issue is to consider l a random variable, i.e., introducing a prior distribution
in the Bayesian sense. Let the prior distribution of l be PL (l). The equation (1) now becomes
Q1
N (b|xl )
PL (l)
b=0 P̂ (b)
L = log Q1
l)
N
(b|x
PL (0)
b=0 P (b)
=l
1
X
b=0
−l
1
X
1
P̂ (b) log N (b|xl )
l
P̂ (b) log P (b) + log PL (l) − log PL (0)
b=0
= lD(p̂kp) + log PL (l) − log PL (0)
The hypothesis test now is
D(p̂kp) >
τ + log PL (0) − log PL (l)
l
(15)
Of course, the problem is that we don’t know P (l). Still, compare that with (5) without the approximations,
D(p̂kp) >
τ+
1
2
log l + c + log∗ l
l
(16)
To the term c + log∗ l corresponds a distribution on the integers, namely Q(l) in [18, (3.6)]. Except for the term
1
2
log l, the equations (15) and (16) are identical if we use the prior distribution PL (l) = Q(l). Rissanen [18] argues
that the distribution Q(l) is the most reasonable distribution on the integers when we have really no prior knowledge,
mainly from a coding point of view. This therefore seems a reasonable distribution for P (l). What about the term
1
2
log l? The model for the non-null hypothesis has one unknown parameter, p, so that it is more complex than the
null hypothesis. We have to account for this additional complexity. Our goal is to find an explanation for atypical
sequences among a large class of explanations, not just the distribution of zeros and ones. If there is no “penalty”
for finding a complex explanation, any data can be explained, and all data will by atypical. This is Occam’s razor
[16]. The “penalty” for one unknown parameter as argued by Rissanen is exactly
1
2
log l. We therefore have the
following explanation for (5),
September 12, 2017
DRAFT
13
Fact 4. The criterion (5) can be understood as a hypothesis test with prior distribution Q(l) [18] and penalty
1
2
log l for the unknown parameter.
Seen in this light, Theorem 2 is not surprising. In (5) we have replaced
implicitly corresponds to the prior distribution PL (l) ∼ l
−3/2
1
2
log l + log∗ l with
3
2
log l, which
, which is exactly the distribution seen in (9).
B. Atypical subsequences
One problem where we believe our approach excels is in finding atypical subsequences of long sequences. The
difficulty in find atypical subsequences is that we may have short subsequences that deviate much from the typical
model, and long subsequences that deviate little. How do we choose among these? Definition 1 gives a precise
answer. For the formal problem statement, consider a sequence {xn , n = −∞, . . . , ∞} from a finite alphabet A
(where in this section A = {0, 1}). The sequence is generated according to a probability law P, which is known.
In this sequence is embedded (infrequent) finite subsequences Xi = {xn , n = ni , . . . , ni + li − 1} from the finite
alphabet A, which are generated by an alternative probability law P̃θ . The probability law P̃θ is unknown, but it
might be known to be from a certain class of probability distributions, for example parametrized by the parameter
θ. Each subsequence Xi may be drawn from a different probability law. The problem we consider is to isolate these
subsequences, which we call atypical subsequences. In this section, as above, we will assume both P and P̃θ are
binary iid.
The solution is very similar to the one for variable length sequences above. The atypical subsequences are encoded
with the universal source coder from [16] with a code length La = lH(p̂) +
1
2
log l. The start of the sequence is
encoded with an extra symbol ’.’ which has a code length − log P ( . ) and the length is encoded in log∗ l bits. In
0 0
conclusion we end up with exactly the same criterion as (5), repeated here
D(p̂kp) >
τ+
3
2
log l
(17)
l
The only difference is that τ has a slight different meaning.
For the subsequence problem, a central question is what the probability is that a given sample xn is part of an
(intrinsically) atypical subsequence. Notice that there are infinitely many subsequences that can contain xn , and
each of these have a probability of being atypical given by Theorem 2.
We can obtain an upper bound as follows. Let us say that Xn has been determined to be part of an atypical
sequence Xi . It is clear that the sequence Xi must also be atypical according to (17). Therefore, we can upper
bound the probability PA (Xn ) that Xn is part of an atypical sequence with the probability of the event (17), using
the approximate criterion (6),
Pn1 +l−1
∃n1 ≤ n < n1 + l :
September 12, 2017
i=n1
√
Xi − p
pql
>
√
2τ ln 2 + 3 ln l
DRAFT
14
We can rewrite this as
Pn1 +l−1
i=n1
√
∃n1 ≤ n < n1 + l :
pql
n1X
+l−1
∃n1 ≤ n < n1 + l :
Xi − p
>
Xi − p >
√
2τ ln 2 + 3 ln l
p
lpq ln 2(2τ + 3 log l)
i=n1
We could upper bound this with a union bound using Theorem 2. However, it is quickly seen that this does not
converge. The problem is that the events in the union bound are highly dependent, so we need a slightly more
refined approach; this results in the following Theorem
Theorem 5. Consider the case p =
1
2.
The probability PA (Xn ) that a given sample Xn is part of an atypical
subsequence is upper bounded by
√
PA (Xn ) ≤ (K1 τ + K2 )2−τ
(18)
for some constants K1 , K2 .
Proof: Without loss of generality we can assume n = 0. For some l0 > 0 let Il0 be the set of subsequences
containing X0 of length l ≤ l0 . For i ∈ Il0 let l(i) be the length of the subinterval. From Theorem 2 we know that
1
PA (i) ≤ 2−τ +1 l3/2
K(l, τ ) and therefore
X
PA (i) ≤ K2−τ
i∈Il0
for some constant K. This argument does not work if we allow arbitrarily long subsequences, because the sum is
divergent. However, we can write
PA (X0 ) ≤
X
PA (i) ≤ K2−τ + PA,l0 (X0 )
i∈Il0
where PA,l0 (X0 ) is the probability that X0 is in an atypical subsequence of at least length l0 . The proof will be to
bound PA,l0 (X0 ).
Define the following events
(n
)
1 +l−1
X
A(n1 , l) =
Xi − p >
p
lpq ln 2(2τ + 3 log l)
i=n1
(n
)
1 +l−1
X
A(n1 , l) =
Xi − p < −
p
lpq ln 2(2τ + 3 log l)
i=n1
For p =
1
2
we can rewrite
n1X
+l−1
Xi − p >
p
lpq ln 2(2τ + 3 log l)
i=n1
=
n1X
+l−1
(2Xi − 1) >
p
l ln 2(2τ + 3 log l)
(19)
i=n1
For ease of notation define
υ(l) =
September 12, 2017
lp
m
l ln 2(2τ + 3 log l)
DRAFT
15
Then using the union bound we can write
−1
X
PA (X0 ) ≤
∞
[
P
n1 =−∞
∞
[
P
n1 =−∞
−1
X
≤2
=2
A(n1 , l)
∞
[
P
!
A(n1 , l)
l=−n1 +1
∞
\
1−P
n1 =−∞
c
!
A (n1 , l)
l=−n1 +1
−1
X
=2
!
l=−n1 +1
n1 =−∞
−1
X
A(n1 , l)
l=−n1 +1
−1
X
+
!
1−
n1 =−∞
∞
Y
P
c
A (n1 , l)
l=−n1 +1
l−1
\
c
!!
A (n1 , `)
`=n−n1 +1
c
Tl−1
c
where we have excluded the length one sequence consisting of X0 itself. Now consider P A (n1 , l) `=−n1 +1 A (n1 , `) =
Tl−1
c
1 − P A(n1 , l) `=−n1 +1 A (n1 , `) .
Pn1 +l−1
We can think of Sl =
(2Xi − 1) as a simple random [41] , and we will use this to upper bound
i=n1
Tl−1
c
the probability P A(n1 , l) `=−n1 +1 A (n1 , `) . This probability can be interpreted as the probability that the
random walk passes υ(l) given that it was below υ(`) at times −n1 < ` < l. But since the random walk can
increase by at most one, and since the threshold is increasing with l, that means that at time l we must have
Sl = υ(l). Furthermore, it is easy to see that the probability is upper bounded by the probability that Sl = υ(l)
given that the random walk is below υ(l) at times −n1 < ` < l. Thus
!
l−1
\
c
P A(n1 , l)
A (n1 , `)
`=−n1 +1
≤ P (Sl = υ(l) |S` < υ(l), −n1 < ` < l )
P (Sl = υ(l), S` < υ(l), −n1 < ` < l)
P (S` < υ(l), −n1 < ` < l)
P (Sl = υ(l), S` < υ(l), −n1 < ` < l)
≤
P (S` < υ(l), 0 ≤ ` < l)
=
(20)
The denominator can be interpreted as the probability that the maximum of the random walk stays below υ(l),
which by Theorem 2 can be expressed by
PD (l) = P (S` < υ(l), 0 ≤ ` < l)
= 1 − 2P (Sl−1 ≥ υ(l + 1) − 1) − P (Sl−1 = υ(l) − 1)
≥ 1 − 2−τ +c l−3/2
≥
September 12, 2017
1
2
DRAFT
16
for τ and l sufficiently large, and where c is some constant. Since, as discussed at the start of the proof, we can
assume that l ≥ l0 , we can choose l0 large enough that this is satisfied; furthermore, since PN (l) is increasing in
τ , we can choose l0 independent of τ as long as τ is sufficiently large.
We will next upper bound the numerator in (20). This is the probability that we have a path that has stayed
below υ(l) at steps −n1 < ` < l, but then at step l hits υ(l). We will count such paths. We divide them into two
groups that we count separately. The first group are all paths that start at zero and hit υ(l) first time after l steps.
The second group is more easily described in reverse time. Those are paths that start at υ(l) at step l, then stay
below υ(l) until time ñ < 0, when they hit υ(l) again, and finally hit 0 at time n1 . According to [41, Section 3.10]
we can count all these paths by
N=
−n
1 +1
X
υ(l)
1
Nl (0, υ(l)) +
Nl−t−1 (1, 0)Nt (0, υ(l))
l
l−t−1
(21)
t=υ(l)
Where Nn (a, b) are the number of length n paths between a and b.
We need to upper bound the probability P (Sn = k) that a path starting a 0 hits k after n steps. We use [41,
Section 3.10] and [16, 13.2] to get
P (Sn = k) =Nn (0, k)2−n
n
2−n
=
1
2 (n + k)
≤
r
s
=
nH
n
2
1
π 4 (n + k)(n − k)
nH
4n
2
2
2
π(n − k )
1 (n+k)
2
n
1 (n+k)
2
n
2−n
2−n
We can bound the power of the exponent to 2 as follows
1
2 (n + k)
−n
nH
n
1 1k
=n H
+
−1
2 2n
2
2
1k
≤−
n
ln 2
2n
=−
1 k2
2 ln 2 n
(22)
Thus,
2
1 k2
P (Sn = k) ≤ √ e2 −
2 ln 2 n
πn
where e2 (x) = 2x .
We will use this to bound the probability of set of paths in the second term in (21). We can bound
September 12, 2017
DRAFT
17
P2 (n1 , l) =
−n
1 +1
X
t=υ(l)
≤
≤
Here the sum
P−n1 +1
t=υ(l)
1
Nl−t−1 (1, 0)Nt (0, υ(l))2−(l−1)
l−t−1
−n
1 +1
X
2
1
p
l − t − 1 π(l − t − 1)
t=υ(l)
1
1
× e2 −
Nt (0, υ(l))2−t
2 ln 2 l − t − 1
−n
1 +1
X
4
π(l + n1 − 2)3/2
P (St = υ(l))
t=υ(l)
P (St = υ(l)) when looked at in reverse time can be interpreted as the probability of a path
starting at υ(l) hits zero before time −n1 + 1. We can the write this as (See [41, Section 3.10])
−n
1 +1
X
P (St = υ(l)) = P (M−n1 +1 ≥ υ(l))
t=υ(l)
≤ 2P (S−n1 +1 ≥ υ(l))
We can use the proof of Theorem 2, specifically (13) to bound this by
2υ(l)2
P (M−n1 +1 ≥ υ(l)) ≤ exp −
−n1 + 1
Then
P2 (n1 , l) ≤
√
K
υ(l)2
exp
−
−n1 + 1
3/2
2(−n1 + 1)
(l + n1 − 2)
We will next bound the probability of the paths in the first term in (21). We have
s
4l
1 υ(l)2
P (Sl = υ(l)) ≤
e2 −
π(l2 − υ(l)2 )
2 ln 2 l
v
u
1 υ(l)2
u 4
e
−
t
2
2
2 ln 2 l
πl 1 − υ(l)
2
l
2
=r
π 1−
−τ −2
υ(l)2
2
l
l2
and
P1 (l) =
r
≤
υ(l)
P (Sl = υ(l))
l
ln 2(2τ + 3 log l)
2
r
l
π 1−
υ(l)2
l2
2
−τ −2
l
v
u
8τ ln 2
u
2−τ l−5/2
≤t
υ(l)2
π 1 − l2
v
r
u
u 12 ln l 4 −τ −5/2
+t
2 l
2
π
π 1 − υ(l)
2
l
September 12, 2017
DRAFT
18
Thus
∞
Y
ln
l−1
\
c
A (n1 , l)
P
≥
l=−n1 +1
≥K
!!
A (n1 , `)
`=−n1 +1
l=−n1 +1
∞
X
c
P1 (l) − P2 (n1 , l)
ln 1 −
PD (l)
∞
X
−P1 (l) − P2 (n1 , l)
l=−n1 +1
.
=S(−n1 , τ )
and
−1
X
PA (X0 ) ≤2
1 − eS(−n1 ,τ )
n1 =−∞
−1
X
≤2K
∞
X
P1 (l)
n1 =−∞ l=−n1 +1
−1
X
+ 2K
∞
X
P2 (n1 , l)
(23)
n1 =−∞ l=−n1 +1
where K > 0 is some constant.
First we evaluate the sum of P1 (l). The term
υ(l)2
l2
is decreasing in l, so for sufficiently large l1 ,
υ(l)2
l2
≤ 21 . We
can evaluate the sum separately for l0 ≤ l ≤ l1 and for l > l1 . Convergence depends only on the latter tail. The
threshold l1 is increasing with τ . If for example we put l1 = 8τ ln 2 , i.e., proportional to τ , we have
for τ > 10. Therefore
l1
X
υ(l)2
l2
≤
1
2
P1 (l) ≤ Kτ 2−τ
l=l0
For l > l1 we can write
r
P1 (l) =
16τ ln 2 −τ −5/2
2 l
+
π
r
24 ln l
π
r
4 −τ −5/2
2 l
π
Then (for −n1 + 1 > l1 )
∞
X
K
P1 (l)
l=−n1 +1
s
ln(n − n1 )
(−n1 )3
!
r
3
−τ
+ k2 2 erfc
ln(−n1 )
2
r
τ
+ k3 2−τ
(−n1 )3
−τ
≤k1 2
September 12, 2017
(24)
DRAFT
19
where ki > 0 are some constants and where we have used
Z ∞
∞
X
−5/2
l
≤
x−5/2 dx =
k−1
l=k
∞ √
X
ln ll−5/2 ≤
∞
Z
√
2
3(k − 1)3/2
ln xx−5/2 dx
k−1
l=k
√
1
=
9
"r
6πerfc
#
!
√
3
6 ln k − 1
ln k − 1 +
2
(k − 1)3/2
as it can be verified that all three sums, when (24) is inserted in (23), are convergent, using
R∞
f (x)dx.
1
P∞
k=1
f (k) ≤ f (1) +
We bound the second sum in (23),
−1
X
∞
X
P2 (n1 , l)
n1 =−∞ l=−n1 +1
−1
X
=
√
8 −n1 + 1
υ(l)2
exp
−
2(−n1 + 1)
π(l + n1 − 2)3/2
+1
∞
X
n1 =−∞ l=−n1
We can ignore the small constants and write
P =
−1
X
∞
X
n1 =−∞ l=−n1
√
8
υ(l)2
exp −
−n1
3/2
2(−n
)
π(l + n1 )
1
+1
√
lτ
3l
8 −n1
=
2− −n1 l− −n1
3/2
π(l + n1 )
n1 =−∞ l=−n1 +1
√
Z ∞Z ∞
lτ
3l
8 t
≤
2− t l− t dldt
3/2
π(l
−
t)
1
t
√
Z ∞Z ∞
8 t
2−l̃τ ˜l−3l̃ t−3l̃ td˜ldt
=
πt3/2 (˜l − 1)3/2
1
1
Z ∞Z ∞
8
2−l̃τ ˜l−3l̃ t−3l̃ d˜ldt
=
π(˜l − 1)3/2
1
1
Z ∞ Z ∞
8
−3l̃
=
2−l̃τ ˜l−3l̃ d˜l
t dt
3/2
˜
π(l − 1)
1
1
Z ∞
8
1
2−l̃τ ˜l−3l̃ d˜l
=
3˜l − 1 π(˜l − 1)3/2
1
Z ∞
1
8
= 2−τ
2−(l̃−1)τ ˜l−3l̃ d˜l
3˜l − 1 π(˜l − 1)3/2
1
−1
X
∞
X
The remaining integral is clearly convergent, and decreasing in τ . Therefore P ≤ K2−τ
There are two important implications of Theorem 5. First is that for τ sufficiently large, PA (Xn ) < 1, and in fact
PA (Xn ) can be made arbitrarily small for large enough τ . This is an important theoretical validation of Definition 1
and the resulting criterion (5) and (6). If the theory had resulted in PA (Xn ) = 1 then everything would be atypical,
and atypicality would be meaningless. That this is not trivially satisfied is shown by Proposition 6 just below. What
that Proposition says is that if in the above equation instead of
have been atypical. Now,
1
2
3
2
log l we had had
1
2
log l, then everything would
log l corresponds to “forgetting” that the length of an atypical sequence also needs to
be encoded for the resulting sequence to be decodable. Thus, it is the strict adherence to decodability that has
September 12, 2017
DRAFT
20
lead to a meaningful criterion. So, although decodability at first seems unrelated to detection, it turns out to be
of crucial importance. Similarly, at first the term
3
2
log l may have seen arbitrary. However, this is just (within a
margin) sufficient to ensure that not everything becomes atypical.
The second important implication of Theorem 5 is that it validates the meaning of τ . The way we introduced
τ was as the number of bits needed to encode the fact that an atypical sequence starts, and therefore we should
put τ = − log P (atypical sequence starts). Theorem 5 confirms that τ has the desired meaning for purely random
sequences. And the reasons is this is not trivial is that τ was chosen from the probability of an atypical sequence,
while Theorem 5 gives the probability of a sample being atypical.
Proposition 6. Consider the case p = 12 . Suppose instead of (7) we use the criterion
Pl
√
Xi − p
q
> 2τ ln 2 + α ln l
1
4l
i=1
(25)
(with α = 3 giving (7)). Then if α ≤ 1, the probability that a given sample Xn is part of an atypical subsequence
is PA (Xn ) = 1.
Proof: We can assume that n = 0. We will continue with the random walk framework from the proof of
Theorem 5. Define the event
(
Āl =
0
X
)
p
(2Xi − 1) >
l ln 2(2τ + α log l)
i=−l+1
(
Al =
0
X
)
p
(2Xi − 1) <
l ln 2(2τ + α log l)
i=−l+1
and
υ(l) =
lp
l ln 2(2τ + α log l)
Then
PA (X0 ) ≥ P
∞
[
∞
[
Āl ∪
l=0
m
!
Al
l=0
Namely, we declare that X0 is atypical if it is the endpoint of an atypical sequence {x[−l], x[−l + 1], . . . , x[0]}
for some l. Clearly, X0 could be the start or midpoint of an atypical sequence, so this a rather loose lower bound.
Now we can write
P
∞
[
Āl ∪
l=0
∞
[
!
Al
l=0
=1−P
∞
\
Ācl
∩
l=0
=1−
∞
Y
P
=1−
l=0
September 12, 2017
!
Acl
l=0
Ācl
l=0
∞
Y
∞
\
∩
Acl
l−1
\
Āck
∩
k=0
"
1−P
Āl ∪ Al
l−1
\
!
Ack
k=0
l−1
\
k=0
Āck
∩
l−1
\
!#
Ack
k=0
DRAFT
21
Tl−1
Tl−1
Consider the probability P Āl ∪ Al k=0 Āck ∩ k=0 Ack . The only way the conditional event can happen is if
Sl−1 = υ(l) − 1 and Xl = 1 or Sl−1 = −υ(l) + 1 and Xl = −1. Here we have
P (Sl−1 = υ(l) − 1)
= Nl−1 (0, υ(l) − 1)2−l+1
l−1
2−l+1
=
1
2 (l + υ(l) − 2)
s
1
(l+υ(l)−2)
(l−1)H 2 l−1
l−1
2−l+1
≥
2
2(l + υ(l) − 2)(l − υ(l))
s
1
(l+υ(l)−2)
(l−1)H 2 l−1
l−1
2
=
2−l+1
2((l − 1)2 − (υ(l) − 1)2 )
1
r
(l+υ(l)−2)
1 (l−1)H 2 l−1
2−l+1
≥
2
2l
Here
1
(l + υ(l) − 2)
(l − 1)H 2
−l+1
l−1
1 1 υ(l) − 1
=(l − 1) H
+
−1
2 2 l−1
2
3 !
1 υ(l) − 1
υ(l) − 1
2
(l − 1)
+ (l − 1)o
≥−
ln 2
2 l−1
l−1
2
1
1 (υ(l) − 1)
(υ(l) − 1)3
=−
+
2 ln 2
l−1
(l − 1)2
l
2
2
1
υ(l)
υ(l)
1
υ(l)3
1
≥−
+
+
+ 2
2 ln 2
l
(l − 1)2
l−1
l
l
2
3
1 υ(l)
υ(l)
1
=−
+ 2
2 ln 2 l
l
l
≥−
1 υ(l)2
υ(l)3
− 2
2 ln 2 l
l
Where the last inequality is only true for l sufficiently large, as for some l0 we have ∀l > l0 : |(l−1 )| < 1. Then
r
1 −τ − α − υ(l)2 3
2 l 22 l
P (Sl−1 = υ(l) − 1) ≥
2l
r
1 −τ − α+1 − υ(l)2 3
=
2 l 2 2 l
2
And
ln 1 − P
∞
[
l=0
≤
≤
∞
X
l=1
∞
X
l=0
September 12, 2017
r
ln 1 −
r
−
Āl ∪
∞
[
!!
Al
l=0
1 −τ − α+1 − υ(l)2 3
2 l 2 2 l
2
!
1 −τ − α+1 − υ(l)2 3
2 l 2 2 l
2
DRAFT
22
Here liml→∞
υ(l)3
l2
υ(l)3
l2
≤ 1. Then
!!
∞
∞
[
[
Āl ∪
Al
= 0. So, for example, for l sufficiently large,
ln 1 − P
l=0
∞
X
l=0
r
1 −τ − α+1
2 l 2
2
l=l0
S∞
S∞
This is divergent for α ≤ 1 proving that P
l=0 Āl ∪
l=0 Al = 1.
≤
Theorem 5 states that for α =
3
2
−
1
2
PA (Xn ) < 1 (convergence), while Proposition 6 shows that for α =
1
2
PA (Xn ) = 1 (divergence). There is a gap between those values of α that is hard to fill in theoretically. We have
therefore tested it out numerically, see Fig. 2. Of course, testing convergence numerically is not quite well-posed.
Still the figure indicates that the phase transitions between divergence and convergence happens right around α = 1.
Probability of atypicality of samle Xn for different values of τ
when length=105+1 and n = length/2 (at the middle)
1
τ = 0.5
τ=1
τ=2
0.9
0.8
Probability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2
α
2.5
3
3.5
4
Fig. 2. Transition between divergence and convergence as a function of α
C. Recursive coding
Instead of using Definition 1 directly, we could approach the problem as follows. First the sequence is encoded
with the typical code. Now, if the distribution of the sequence is in agreement with the typical code, the results
should be a sequence of iid binary bits with P (Xi = 1) =
1
2
[16], i.e., a purely random sequence; and this sequence
cannot be further encoded. We can now try if we can further encode the sequence with a (universal) code. If so,
we categorize the sequence as atypical. Let l∗ be the length of the sequence after typical coding. In (4) the typical
and atypical codelengths are therefore
Lt = l∗ − log
La = l∗ H(p̂∗ ) +
September 12, 2017
3
log l∗ − log + τ ∗
2
(26)
DRAFT
23
Here p̂∗ is the estimated p for the encoded sequence. Now
1
1
∗
l = l p̂ log + (1 − p̂) log
p
1−p
(27)
l∗ H(p̂∗ ) ∼ lH(p̂)
(28)
log l∗ = log l + log p̂ log
1
1
+ (1 − p̂) log
p
1−p
∼ log l
(29)
(30)
The argument for (28) is as follows (without doing detailed calculations): if we encode a sequence with a “wrong”
code and then later re-encode with the “correct” code (for the induced statistic), the result is the same as originally
encoding with the correct code. Thus the criterion (4) and (26) are approximately equivalent. We can state this as
follows
Proposition 7. Definition 1 can be applied to encoded sequences instead of the original data.
This of course ignores all integer constraints, block boundaries etc. But the importance of this statement is that
it is sometimes easier to operate on (partially) encoded sequences simply because the amount of data has already
been reduced, and the problem has been standardized: as such, we do not need to know the typical codebook or
even the model of typical data since everything under the typical model has been reduced to a stream of iid binary
digits, and atypicality algorithms can therefore be applied to data streams without knowledge of what is the original
data. It also means that theoretical results such as Theorem 5 where we assume typical data is iid uniform has
general applicability.
However, first encoding the sequence and then doing atypicality detection also has disadvantages in a practical,
finite length setting. Atypical subsequences become embedded in typical sequences in unpredictable ways. For
example, it could be difficult to determine where exactly an atypical sequence starts and ends. Our practical
implementation therefore uses Definition 1 directly.
IV. G ENERAL C ASE
Return to the problem considered at the start of Section III where we are given a sequence x of fixed length l
and we need to determine if it is atypical. In the iid case this is a simple hypothesis test problem and the solution
is given by (1). In the general case we would like find to alternative explanations from a large abstract class of
models. The issue is that it is often possible to fit an alternative model very well to the data if we just allow complex
enough models – the well known Occam’s razor problem [16]. Rissanen’s MDL [42], [18], [19] is a solution to
this problem. Therefore, in the general case, even for fixed length sequences, the problem is not a straightforward
hypothesis test problem, and we have to resort to information theory.
September 12, 2017
DRAFT
24
A. Finite State Machines
On possible class of models in the general case is the class of finite state machines (FSM). Rissanen [17] defines
the complexity of a sequence xl in the class of FSM sources by
I(xl ) = min{− log P̂ (xl |fj ) + log∗ j + c}
(31)
where f1 , f2 , . . . is a sequence of state machines, and where we have used P̂ (xl |fj ) to emphasize that the probability
is estimated. Rissanen uses Laplace’s estimator, but the KT-estimator [43], [16] could also be used. Except for
integer constraints, this is a valid descriptive length, and can therefore be used in Definition 1. This is a natural
extension of the iid case considered in Section III. As opposed to Kolmogorov complexity, this complexity could
actually be calculated, although with high complexity. Because of the complexity, it is mostly useful for theoretical
considerations, and one result is the following generalization of Theorem 2
Theorem 8. Assume that the typical distribution is iid uniform. If the atypical descriptive length is given by (31)
with a maximum number of states independent of l, the probability of an intrinsically atypical sequence PA (l)
satisfies
ln PA (l)
=1
l→∞ − 3 ln l
2
lim
(32)
Proof: Since we consider all state machines with the number of states up to a certain maximum, this must also
include the state machine with a single state. This is equivalent to the iid model in Section III, and we therefore
get the lower bound in (32). The proof will be to upper bound the probability. As in Section III we use log l + τ
bits to indicate beginning and end of atypical sequences. The probability that a sequence xl is atypical therefore is
PA (l) = P (I(xl ) + log l + τ > l)
[
= P ( − log P̂ (xl |fj ) + log∗ j + c + log l + τ > l)
fj
≤
[
P (− log P̂ (xl |fj ) + log l + τ > l)
fj
We will prove that P (− log P̂ (xl |fj ) + log l + τ > l) ≤ Kj l−(k+2)/2 for constants Kj and k the number of states
in the state machine, and since the slowest decay dominates, we get the upper bound for (32).
For a fixed state machine f the code length according to [17, (3.6)] is
l
X
X
n
(x
)
s
+
L(xl |f ) =
log
log(ns (xl ) + 1)
l
n
(x
)
s
s
0|s
where ns (xl ) denotes the number of occurrences of state s in xl and n0|s (xl ) the number of times the next symbols
is 0 at this state. Further, from [16, 13.2]
l
L(x |f ) ≥
X
s
l
ns (x )H
n0|s (xl )
ns (xl )
!
n0|s (xl ) n1|s (xl )
1
− log 8
2
ns (xl ) ns (xl )
+ log(ns (xl ) + 1)
September 12, 2017
−
1
log ns (xl )
2
!
(33)
DRAFT
25
We want to upper bound the probability of the event L(xl |f ) + log l + τ < l. We can write
2 !
1
1
1 ns (xl ) + 1
l
l
log(ns (x ) + 1) − log(ns (x )) = log l + log
.
2
2
l
ns (xl )
Let
l
r(x ) =
X
l
ns (x ) H
s
n0|s (xl )
ns (xl )
!
!
−1
and let R(xl ) be the remaining small terms in (33) dependent on xl ,
!
n0|s (xl ) n1|s (xl )
1
R(x ) =
− log 8
2
ns (xl ) ns (xl )
s
2 !
1 ns (xl ) + 1
+ log
.
l
ns (xl )
P
Then we have to upper bound (notice that s ns (xl ) = l),
k+2
l
l
log l
P −r(x ) − R(x ) ≥ τ +
2
X
l
The Chernoff bound is
P
−r(xl ) − R(xl ) ≥ τ +
≤ exp(−t(τ +
k+2
log l
2
k+2
log l))M (t)
2
or
ln P
−r(xl ) − R(xl ) ≥ τ +
≤ −t(τ +
k+2
log l
2
k+2
log l) + ln M (t)
2
where
M (t) = E exp(−t(r(xl ) + R(xl ))
In order to get a valid bound, we need to show that M (t) < K < ∞ independent of l for t < ln 2. Now it’s easy
to see that exp(−tR(xl )) ≤ K < ∞ for all t and l. So, we have to show
E exp(−t(r(xl )) ≤ K < ∞
We have to show that this is true for all state machines in the class of finite state machines with k states, which
can be done by showing
max
E exp(−t(r(xl )) ≤ K
FSM with k states
It turns out it easier to prove this if we expand the class over which we take the maximum, and clearly expanding
the class does not decrease the maximum. A FSM with k states is a function f (xl ) ∈ {1, . . . , k} that satisfies
that if f (xm ) = f (x̃m ) = s then f (xm b) = f (x̃m b) for any bit b [17], i.e., if the FSM is in state s after m
steps, the next state transition is only dependent on the next bit, not how it got to state s. We extend the class by
dispensing with this requirement. We can then describe the ’program’ we run as follows. Based on xm we choose
a state sm ∈ {1, . . . , k} without having any knowledge about xm+1 , except that it is independent and uniformly
September 12, 2017
DRAFT
26
distributed (by the assumption on the typical distribution). We can think of this slightly differently. The program
puts xm+1 into bucket s ∈ {1, . . . , k} and updates ns (m) and ni|s (m), in order to maximize E exp(−t(r(xl )) .
It does so based on past data xm . Now, as opposed to the state machine setup, the choice of sm in no way restricts
the choices of states (or buckets) sn , n > m. Since the program has no knowledge of xm+1 the program cannot
optimize sm based on the values of xm . Rather, it is sufficient to look at ns (m). It is now easy to see that the
worst case is obtained if the bits are distributed evenly in the states. Thus, the worst case of r(xl ) is
!
!
X l
n0|s (xl )
l
H
−1
r(x ) =
k
l/k
s
where the n0|s (xl ) are independent of s. Thus, the problem is reduced to the case of a single state, which is showing
that
k
≤K<∞
E exp tl 1 − H
l
(34)
Here we have
k
E exp tl 1 − H
l
l
X
k
t
k
2−l
=
2 ln 2 l(1−H( l ))
l
k=0
l/2
X
k
k
t
2−l
=1+2
2 ln 2 l(1−H( l ))
l
k=1
s
l/2
X
t
k
k
l(1−H(
))
−l
lH(
)
l
≤1+2
2 ln 2
2 2 l
k=1
l/2
X
l
πk(l − k)
s
l
πk(l − k)
k=1
s
l/2
X
t
2
k
1 2
l
≤1+2
2−(1− ln 2 ) ln 2 l( l − 2 )
πk(l − k)
=1+2
2
(1− lnt 2 )l(1−H( kl ))
k=1
where we have used [16, 13.2] and (22). The sum is actually decreasing as a function of l, but this seems hard to
prove. Instead we upper bound the sum by
l/2
X
s
l
πk(l − k)
k=1
s
Z l/2
t
2
k
1 2
l
≤
2−(1− ln 2 ) ln 2 l( l − 2 )
dk
πk(l − k)
1
September 12, 2017
2−(
1− lnt 2
)
2
ln 2 l
(
k
1
l −2
)
2
DRAFT
27
4k
Here we can upper bound
q
√1
l
−1
+2
for 1 ≤ k ≤ 2l . Then
s
Z l/2
2
l
−(1− lnt 2 ) ln22 l( kl − 12 )
2
dk
πk(l
− k)
1
q
1
Z 1/2
4x
−1 +2
t
2
1 2
l
√
2−(1− ln 2 ) ln 2 l(x− 2 )
≤
ldx
π
1/l
q
1
Z 1/2
−
1
+2
4x
2
t
2
1
l
√
2−(1− ln 2 ) ln 2 l(x− 2 )
ldx
≤
π
−∞
K1
= √ + K2
l
l
πk(l−k)
≤
k
√
π
for some constants K1 , K2 , using Gaussian moments. This proves (34).
10 -1
The probability of intrinsically atypical sequence
Probability
Upper Bound
10 -2
PA (l)
10 -3
10 -4
10 -5
10 -6
10 1
10 2
10 3
Length
Fig. 3. Probability of an intrinsically atypical sequence. The typical distribution is iid uniform, and for detection of atypical sequences the
CTW algorithm has been used (Section IV-B).
While the Theorem is for the typical model iid uniform, as outlined in Section III-C in principle it also applies
to general sources, since we can first encode and then look for atypical sequences.
The theorem shows that looking for more complex explanations for data does not essentially increase the
probability of intrinsically atypical sequences. Fig. 3 (compare with Fig. 1) confirms this experimentally. The
atypical detection is based on CTW, which as explained in Section IV-B below, is a good approximation of FSM
modeling. On the other hand, if one of the FSM models do in fact fit the data, the chance of detecting the sequence
is greatly increased, although hard to quantify. If we think of intrinsically atypical sequences as false alarms, this
shows the power of the methodology.
Since FSM sources has the same PA (l) as in the iid case, it seems reasonable to conjecture that Theorem 5 is still
valid, that is PA (Xn ) < 1 for sufficiently large τ , which is clearly an essential theoretical property of atypicality.
September 12, 2017
DRAFT
28
However, as Theorem 5 does not follow directly from Theorem 2, to verify the conjecture requires a formal proof
which we do not have at present.
B. Atypical Encoding
In terms of coding, Definition 1 can be stated in the following form
C(x|P) − C(x) > 0
Here C(x|P) is the code length of x encoded with the optimum coder according to the typical law, and C(x) is x
encoded ’in itself.’ As argued in Section III, we need to put a ’header’ in atypical sequences to inform the encoder
that an atypical encoder is used. We can therefore write C(x) = τ + C̃(x), where τ is the number of bits for the
’header,’ and C̃(x) is the number of bits used for encoding the data itself. For encoding the data itself an obvious
solution is to use a universal source coder. There are many approaches to universal source coding: Lempel-Ziv
[16], [44], [45], Burrows-Wheeler transform [46], partial predictive mapping (PPM) [47], [48], or T-complexity
[49], [6], [50], [51], [52], [53], [54], and anyone of them could be applied to the problem considered in this paper.
The idea of atypicality is not linked to any particular coding strategy. In fact a coding strategy does not need to be
decided. We could try several source coders and choose the the one giving the shortest code length; or they could
even be combined as in [55]. However, to control complexity, we choose a single source coder. The most popular
and simplest approach to source coding is perhaps Lempel-Ziv [16], [44], [45]. The issue with this is that while
l
)
= H(X) wp 1, the convergence is very slow. According to [56]
it is optimum in the sense that lim supl→∞ C(x
l
h
i
h
i
l
C(xl )
C(x
)
1
E
− H(X) ∼ log
∼ 1l . Thus, Lempel-Ziv is poor for short sequences, which is exactly
l
l while var
l
what we are interested in for atypicality.
We have therefore chosen to use the Context Tree Weighing (CTW) algorithm [43]. The CTW approach has
some advantages in our setup: it is a natural extension of the simple example considered in Section III, it allows
estimation of code length without actually encoding, there is flexibility in how to estimate probabilities. Importantly,
it can be seen as a practical implementation of the FSM based descriptive length used in Section IV-A.
C. Typical Encoding and Training
In Definition 1 and the example in Section III we have assumed that the typical model of data is exactly known.
If that is the case, typical encoding is straightforward, using for example arithmetic coding – notice that we just
need codelength, which can be calculated for arithmetic coding without actually encoding. However, in many cases
the typical model is not known exactly. In the simplest case, the typical model is from a small class parametrized
with a few parameters (e.g., binary iid with P (X = 1) unknown). In that case, finding the typical model is a simple
parameter estimation problem, and we will not discuss this further. We will focus on the case where the typical
model is not given by a specific model. In that case, it seems obvious to also use universal source coding for typical
data. However, this is not straightforward if we want to stay faithful to the idea of Definition 1 as we will argue
below.
September 12, 2017
DRAFT
29
Let P be the typical model and P̂ be the estimated typical model. When no specific model is given, in principle
P̂ could be given by an estimate of the various joint probability mass functions. However, a more useful approach
is to estimate the conditional probabilities p(xn = 1|xn−1 , xn−2 , . . . , xn−N ), where s = xn−1 , xn−2 , . . . , xn−N is
called the context. If the source has finite memory these probabilities characterize the source, and otherwise they
could give a good approximate model. The issue is that there are 2N possible contexts, so for N even moderately
large the amount of training data required to even observe every context is large, and to get good estimates for
every context it is even larger. Realistically, therefore not every probability p(xn = 1|s) can be estimated. This is
an issue universal source coding is designed to deal with, and we therefore turn to universal source coding.
Let us assume we are given a single long sequence t for training – rather than a model P – and based on this
we need to encode a sequence x. Let us denote this coder as C(x|t). To understand what this means, we have to
realize that when x is encoded according to C(x|P) with a known P, the coding probabilities are fixed; they are
not affected by x. As discussed in Section II, this is an important part of Definition 1 that reacts to ’outliers,’ data
that does not fit the typical model. If the coding probabilities for C(x|P) were allowed to depend on x, in extreme
case we would always have C(x) = C(x|P).
The issue with universal source coders is that they often easily adapts to new types of data, a desirable property
of a good universal source coder, but problematic in light of the above discussion. We therefore need to ’freeze’ the
source coder, for example by not updating the dictionary. However, because the training data is likely incomplete
as discussed above, the freezing should not be too hard. The resulting encoder C(x|t) is not a universal source
coder, but rather a training based fixed source coder, and the implementation can be quite different from a universal
source coder, requiring careful consideration.
We will suggest one algorithm based on the principle of the CTW algorithm. This naturally complements using
the CTW algorithm for atypical encoding, but could also be used with other atypical encoders. The following
discussion requires a good knowledge of the CTW algorithm, which is most easily obtained from [57].
The algorithm is based on estimating P (xn = 1|s) for contexts s. The estimate for a given s is done with the
KT-estimator [43], [16] P (xn = 1|s) =
bs +1/2
as +bs +1 ,
where as and bs are the number of 0s and 1s respectively seen
with context s in the training data t, but unaffected by the test sequence x. As discussed above, the complication
is that not every context s might be seen and that in particular long contexts s are rarely seen so that the estimates
P (Xn = 1|s) might be more accurate for shorter contexts. We solve that with the weighting idea of the CTW
algorithm; the weights can be thought of as a prior distribution on different models. We can summarize this as
follows. For every context s, the subsequence associated with s could either be memoryless or it could have memory
[57]. We call the former model M1 and the latter M2 . The CTW algorithm uses a prior distribution, weights, on
the models P (M1 ) = P (M2 ) = 21 . Our basic idea is to weigh with P (M1 |t) and P (M2 |t) instead of 12 .
The algorithm is best described through an example. We have trained the algorithm with t, resulting in the context
tree seen in Fig. 4. Now suppose we want to find a coding distribution P (1|010) for the actual data. We begin at
September 12, 2017
DRAFT
30
as, bs
s=11
1
as, bs
s=1
1
0
as, bs
s=λ
as, bs
s=01
0
as, bs
s=0
Fig. 4. Example context tree.
the root and calculate
P (1|010, t) = P (1|010, M1 )P (M1 |t) + P (1|010, M2 )P (M2 |t)
P (Mi |t) =
P (t|Mi )P (Mi )
P (t|Mi )
=
P (t)
P (t|M1 ) + P (t|M2 )
P (t|M1 ) = Pe (as , bs )
P (t|M2 ) = Pw0s (t)Pw1s (t)
s=λ
(empty context)
under model M1 the data is memoryless, so
P (1|010, M1 ) = P (1|M1 ) =
bs + 1/2
;
as + bs + 1
s=λ
To find P (1|010, M2 ) we look in the 0-node of the context tree. Here we calculate similarly
P (1|010, t) = P (1|010, M1 )P (M1 |t) + P (1|010, M2 )P (M2 |t)
P (Mi |t) =
P (t|Mi )P (Mi )
P (t|Mi )
=
P (t)
P (t|M1 ) + P (t|M2 )
P (t|M1 ) = Pe (as , bs )
P (t|M2 ) = Pw0s (t)Pw1s (t)
s=0
and again under model M1 the data is iid, so
P (1|10, M1 ) = P (1|0, M1 ) =
September 12, 2017
bs + 1/2
;
as + bs + 1
s=0
DRAFT
31
and so on. From the context tree it is seen that the context 10 has not been seen before; then the context 010 has
not been seen either. Then we have
P (1|010, M2 ) =
1
.
2
No more look-up is needed, and the calculation has completed.
The algorithm can be implemented as follows. We run the standard CTW algorithm on the training data. To
freeze, in each node corresponding to the context s we can pre-compute P (M1 |t) and P (1|s, M1 ). This is the only
data that needs to be stored. While the algorithm is described from the root and up, implementation is simpler (no
recursion) from the top and down to the root.
Often the source coder might be trained with many separate sequences, rather than one long sequence. This is
not an issue, but care has to be taken with the startup for each sequence. The original CTW paper [43] assumes that
a context of length at least D is available prior to the start of the sequence, which is not true in practice. The paper
[58] solves this by introducing an indeterminate context. A context may start with an indeterminate context, but at
most once. With multiple training sequences this could happen more than once. A better approach is therefore to
use the start of each training sequence purely as a context (i.e., not code it). This wastes some training bits, but
if the sequences are long the loss is minor. A different case would be if we need to find short atypical sequences
rather than subsequences. In that case a more careful treatment of start of sequences would be needed.
Freezing the encoder is essential in implementing atypicality. A simulation confirming this is shown in Fig. 5.
The CTW algorithm is trained with a three-state Markov chain with transition probability [.05 .95 0;0 .05 .95;.95 0
.05] while generating {0,1} according to [0 1 x;x 1 0;1 x 0], so this Markov chain mostly generates the following
pattern: [1 0 1]. The test sequence is generated by another three-state Markov chain with the same transition
probability but generating {0,1} according to [0 1 x;x 1 0;0 x 1], i.e., generating mostly [1 0 0] as pattern. With
the non-frozen algorithm the code length difference between typical and atypical encoding is so small that it can
easily be missed, although the difference in the patterns themselves in the raw data is clearly visible to the naked
eye. The reason the non-frozen algorithm does not work is that it quickly learns the new [1 0 0] pattern. Any good
source coder would do that including LZ. This is advantageous to source coding, but in this case it means missing
a very obvious atypical pattern.
If a very large amount of data is used for training, the complexity can become very high, mainly in terms of
memory. Namely, all contexts might be observed, and the context tree completely filled out. For example, suppose
the typical data is actually iid. That means that every string x1 , . . . , xN is seen with equal probability. For the CTW
algorithm that means that every node of the context tree will be filled out, and the number of nodes with depth D
is 2D+1 . For dictionary based algorithms, it means that the dictionary size becomes huge. What is needed is some
algorithm that not only estimates the unknown parameters, but also the model (e.g., iid). One way could be to trim
the context tree (or dictionary), but we have not looked into this in detail.
D. Atypical subsequences
For finding atypical subsequences of a a long sequences, the same basic setup as in the previous sections can
be used. Let X (n, l) = (xn , xn+1 , . . . , xn+l−1 ) be a subsequence of {xn , n = 0, . . . , ∞} that we want to test for
September 12, 2017
DRAFT
32
The Effect of Freezing: Non-Frozen Typical Coder
Will Learn The Atypical Sequences
800
Non-Frozen Typical Code Length
Frozen Typical Code Length
Atypical Code Length
700
600
500
400
300
200
100
0
0
50
100
150
200
250
300
350
400
450
500
Fig. 5. The importance of freezing source coding when testing for atypicality.
atypicality. As in Section III-B the start of a sequence needs to be encoded as well as the length. Additionally the
code length is minimized over the maximum depth D of the context tree. The atypical code length is then given
by
LA (X (n, l)) = min − log Pwλ (D) + log∗ D + log∗ l
D
except for the τ . Here
Pwλ (D)
denotes the probability at the root of the context tree [43] of depth D. Since we are
also interested in finding short sequences, how the encoding is initialized is of importance, and for atypical coding
we therefore use the algorithm in Section II of [58].
For typical coding we use either a known fixed model and Shannon codes, or the algorithm in Section IV-C
when the model is not known; when we encode X (n, l) we use xn−D , . . . xn−1 as context for xn (we can assume
n > D). Equivalently, we can encode the total sequence {xn , n = 0, . . . , ∞} (with the algorithm from Section
IV-C); let L(n) be the codelength for the sequence x0 , . . . , xn . Then we can put LT (X (n, l)) = L(n+l−1)−L(n).
We need to test every subsequence of every length, that is, we need to test subsequences X (n, l) for every value
of n and l. For atypical coding this means that a new CTW algorithm needs to be started at every sample time. So,
if the maximum sequence length is L, L separate CTW trees need to maintained at any time. These are completely
independent, so they can be run on parallel processors.
The result is that for every bit of the data we calculate
∆L(n) = LA (X (n, l)) − min LT (X (n, l))
l
(35)
and we can state the atypicality criterion as ∆L(n) < −τ . The advantage of stating it like this is that we do not
need to choose τ prior to running the algorithm. We can sort according to ∆L(n), and first examine the data with
smallest ∆L(n), which should be the most atypical parts of data. Thus, the algorithm is really parameter free,
September 12, 2017
DRAFT
33
which is one advantage of our approach. In many anomaly detection algorithms, there are multiple parameters that
need to be adjusted.
This implementation clearly is quite complex, but still feasible to implement for medium sized data sets due to
the speed of the source coder. In order to process larger data sets, a faster approximate search algorithm is needed,
and we are working on such algorithms, but leave that as a topic of later papers.
V. E XPERIMENTAL R ESULTS
In order to verify the performance of our algorithm, we used three different experiments. In the first we evaluated
randomness of sources, in accordance with our starting point of Kolmogorov-Martin-Löf randomness. In the second,
we looked for infection in human DNA, and in the third we looked for arrhythmia in ECG.
A word about presentation of the results. For the outcome of our method we plot ∆L(n) given by (35). At the
same time, we would like to illustrate the raw data. The source in all cases is a stream of bits x[n] ∈ {0, 1}. We
PN
convert this to y[n] ∈ {−1, 1} (i.e, y[n] = (−1)x[n]−1 ), and then plot S[N ] = n=1 y[n]; we call this the random
walk representation. In our experience, this allows one to quickly assess if there is any obvious pattern in data. If
the data is random, the results will look like a typical random walk: both small fluctuations and large fluctuations.
All experimental data and software used is available at http://itdata.hostmadsen.com.
A. Coin Tosses
In this experiment the typical data is iid binary random. As source of typical data we used experimental coin
tosses from [59]. This data consists of 40,000 tosses by two Berkeley undergraduates of a fair coin and the result
has 20,217 heads (Xi = 1), so Pr{Xi = 1} = 0.505425 ≈
1
2.
Therefore we can consider it as a real binary
IID experiment, indeed it is an example of pure random data. In our experiments with this data, we examine the
randomness of other types of data.
One type of data one might think is purely random are word length changes in a text. In the first experiment,
we generate binary data using consecutive word length comparison of part of M. B. Synge’s “On the shores of the
great sea” in the following manner: If the next word is longer than the current word, 1 is assigned to the binary
data, otherwise 0. In the case of two consecutive words with same length, a random 0 or 1 is generated (with a
good random number generator). We then insert this data in the coin toss data. Since we assume coin tosses data is
IID, there is no need to train the CTW and the the code length of the IID case (2) can be used for typical coding.
Fig. 6 illustrates the result of the algorithm on the mixed data. Thus, word length changes are not iid random.
Perhaps this is because word length is bounded from above and below, so that there are limits to how long runs of
0s or 1s are possible.
In second experiment, we generated random data with the infamous RANDU random number generator. This
was a random number generator that was widely used until it was discovered that it has some clear deviation from
randomness. RANDU generates random numbers in the interval [0, 231 − 1], so each number needs 31 bits for
binary representation. But instead of using 31 bits for each number, we sum up all the 31 bits and compare it with
September 12, 2017
DRAFT
34
Mixture of Coin Tosses and Consecutive Word Length Comparison
Data from "On The Shores Of The Great Sea" by M. B. Synge
100
Random Walk of Mixed Data
"L(n) (bits)
80
Cummulative Bits
60
40
20
Text Word Length
Comparison
0
-20
-40
500
1000
1500
2000
2500
3000
3500
4000
Samples
Fig. 6. Random walk of mixed coin tosses and consecutive word comparison.
Mixture of Coin Tosses and RANDU Data
300
Random Walk of Mixed Data
"L(n) (bits)
250
Cummulative Bits
200
150
100
50
RANDU
0
-50
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Samples
Fig. 7. Random walk of mixed coin tosses and RANDU.
15.5 to generate either 0 or 1. Then this data is inserted in the part of coin tosses data. Fig. 7 shows that the most
atypical segment is where we have inserted data from RANDU random number generator.
In the third experiment, we generate binary data using consecutive heart rate comparison of part of normal sinus
rhythm downloaded from MIT- BIH database [60] in the same way as for the text. Fig. 7 represents the result
of the algorithm on the mixed data. As can be seen the atypicality measure shows a huge difference between iid
randomness and randomness of consecutive heart beats. We don’t know why this is the case.
September 12, 2017
DRAFT
35
Mixture of Coin Tosses and Consecutive
Normal Heart Rate Comparison
150
Normal Heart
Rate Comparison
Cummulative Bits
100
50
0
-50
Random Walk of Mixed Data
"L(n) (bits)
-100
500
1000
1500
2000
2500
3000
3500
4000
4500
5000
Samples
Fig. 8. Random walk of mixed coin tosses and HRV.
B. DNA
In this collection of experiments, we detect viral and bacterial insertion into human genomic DNA. DNA from
foreign species can be inserted into the human genome either through natural processes [61], typically though viral
infections, bacterial infections, or through genetic engineering [62]. The inverse also occurs in genetic engineering
experiments during the creation of “transgenic” organisms, with the insertion of human DNA into bacteria, yeast,
worms, or mice. In the experiments we show here, we have focused on the former case. We train the CTW algorithm
on pure human genomic DNA.
The data that we have used was comprised of ~20 kilobases of human genomic DNA (each sequence from a
different chromosome) with either bacterial or viral random DNA sequences (~2 kilobases per insertion) inserted.
Since our software is too slow to find atypical sequences of length more than a few hundreds, we removed the
middle of the insertions. Notice that this actually makes detection harder. We used some of the human DNA for
training, but not the same as the test sequences.
In the first experiment we tried to detect short sequences from Streptococcus Pneumoniae (a bacterial infection
with a high fatality rate, and a frequent cause of death in the elderly) randomly inserted into larger segments of
human genomic DNA. Fig. 9 illustrates the result of the experiment. Based on the figure, the inserted Streptococcus
Pneumoniae DNA fragment was detected by our algorithms.
In the second experiment we tried to detect HIV inserted into human genomic DNA to mimic viral infection,
which is a more realistic experiment since viruses typically insert their DNA into the host genome every time a
human obtains a viral infection. Fig. 10 illustrates the result of the experiment. As can be seen, the infected viral
fragment was detected by our algorithms.
September 12, 2017
DRAFT
36
Primarily Human Genomic DNA Infected with Streptococcus
Pneumoniae (Fatal Bacterial Infection In the Elderly)
80
60
Cummulative Bits
40
20
0
-20
-40
-60
Streptococcus Pneumoniae
Infection
-80
-100
Random Walk of the Infected Human DNA
"L(n) (bits)
-120
1000
2000
3000
4000
5000
6000
7000
DNA samples
Fig. 9. Random walk of human DNA with bactrial infection.
50
Primarily Human Genomic DNA Infected with HIV
(Viral Infection)
Random Walk of the Infected Human DNA
"L(n) (bits)
0
Cummulative Bits
-50
-100
-150
-200
-250
HIV Infection
-300
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
DNA samples
Fig. 10. Random walk of human DNA with viral infection.
C. HRV
While HRV (Heart Rate Variability) can be a powerful indicator for arrhythmia [8], the common issue is that it is
not known exactly what to look for in the data. Our aim for this application is to use atypicality to localize signs of
subtle and complex arrhythmia. In [63] based on our modest goal of localizing a simple type of known arrhythmia,
we managed to find premature beats using HRV signal, but here we attempted to detect more subtle arrhythmia.
The HRV signals that we used were downloaded from MIT- BIH database [60]. We used “MIT-BIH normal sinus
rhythm database (nsrdb)” and “MIT-BIH supraventricular arrhythmia database (svdb)”. Encoding of HRV signals
September 12, 2017
DRAFT
37
were done by same manner as the text word length comparison of subsection V-A. In this experiment, CTW was
trained with HRV of normal sinus rhythm, then applied to a HRV signal that has supraventricular rhythms. Fig.
11 shows the result of the simulation. The algorithm was able to localize the segment that suffers from abnormal
rhythms.
200
HRV Random Walk and Atypicality Measure
150
Supraventricular
Arrhythmia
Cummulative Bits
100
50
0
-50
-100
Random Walk of the HRV
"L(n) (bits)
-150
1000
2000
3000
4000
5000
6000
7000
8000
HRV samples
Fig. 11. Random walk of HRV.
VI. C ONCLUSION
In this paper we have developed a criterion for finding atypical (sub)sequences in large datasets. The criterion is
based on a solid theoretical foundation, and we have shown that the criterion is amenable to theoretical analysis.
In particular, we have shown that the probability that a sample is intrinsically atypical is less than 1, an important
theoretical requirement that is not trivially satisfied.
We have also shows in a few examples with real-world data that the method is able to find known atypical
subsequences. Our aim set out in the introduction was more ambitiously to find ’interesting’ data in large datasets.
In that context, the purpose of the current paper is only to introduce the methodology, and show some theoretical
properties. In order to analyze really big datasets, we need much faster (probably approximate) algorithms, much
more efficient software (say in Python or C), and faster computers. We are working on that as future work.
R EFERENCES
[1] L. Akoglu, H. Tong, J. Vreeken, and C. Faloutsos, “Fast and reliable anomaly detection in categorical data,” in Proceedings of the 21st
ACM international conference on Information and knowledge management.
ACM, 2012, pp. 415–424.
[2] K. Smets and J. Vreeken, The Odd One Out: Identifying and Characterising Anomalies, 2011, ch. 69, pp. 804–815. [Online]. Available:
http://epubs.siam.org/doi/abs/10.1137/1.9781611972818.69
[3] W. Liu, I. Park, and J. Principe, “An information theoretic approach of designing sparse kernel adaptive filters,” Neural Networks, IEEE
Transactions on, vol. 20, no. 12, pp. 1950–1961, Dec 2009.
September 12, 2017
DRAFT
38
[4] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection for discrete sequences: A survey,” Knowledge and Data Engineering, IEEE
Transactions on, vol. 24, no. 5, pp. 823 –839, may 2012.
[5] D. Rumsfeld, Known and Unknown: A Memoir.
Penguin, 2011.
[6] K. Hamano and H. Yamamoto, “A randomness test based on t-codes,” in Information Theory and Its Applications, 2008. ISITA 2008.
International Symposium on, dec. 2008, pp. 1 –6.
[7] A. Nies, Computability and Randomness.
Oxford University Press, 2009.
[8] M. Malik, “Heart rate variability,” Annals of Noninvasive Electrocardiology, vol. 1, no. 2, pp. 151–181, April 1996.
[9] M. F. Hilton, R. A. Bates, K. R. Godfrey, and et al., “Evaluation of frequency and time-frequency spectral analysis of heart rate variability
as a diagnostic marker of the sleep apnea syndrome,” Med. Biol. Eng. Comput., vol. 37, no. 6, pp. 760–769, November 1999.
[10] N. V. Thakor and Y. S. Zhu, “Application of adaptive filtering to ECG analysis: Noise cancellation and arrhythmia detection,” IEEE Trans.
on Biomed. Eng., vol. 38, pp. 785–794, August 1991.
[11] J. F. Thayer, S. S. Yamamoto, and J. F. Brosschot, “The relationship of autonomic imbalance, heart rate variability and cardiovascular
disease risk factors,” International Journal of Cardiology, vol. 141, no. 2, pp. 122–131, May 2009.
[12] J. Fondon and H. Garner, “Probing human cardiovascular congenital disease using transgenic mouse models,” Proc Natl Acad Sci U S A,
vol. 101, no. 52, pp. 18 058–63, 2004.
[13] D. K. Mellinger, K. M. Stafford, S. E. Moore, R. P. Dziak, and H. Matsumoto, “An overview of fixed passive acoustic observation methods
for cetaceans,” Oceanograpy, vol. 20, no. 4, pp. 36–45, 2007.
[14] M. Thottan and C. Ji, “Anomaly detection in ip networks,” Signal Processing, IEEE Transactions on, vol. 51, no. 8, pp. 2191 – 2204, aug.
2003.
[15] M. Li and P. Vitányi, An Introduction to Kolmogorov Complexity and Its Applications, 3rd ed.
[16] T. Cover and J. Thomas, Information Theory, 2nd Edition.
Springer, 2008.
John Wiley, 2006.
[17] J. Rissanen, “Complexity of strings in the class of markov sources,” Information Theory, IEEE Transactions on, vol. 32, no. 4, pp. 526 –
532, jul 1986.
[18] ——, “A universal prior for integers and estimation by minimum description length,” The Annals of Statistics, no. 2, pp. 416–431, 1983.
[19] ——, “Universal coding, information, prediction, and estimation,” Information Theory, IEEE Transactions on, vol. 30, no. 4, pp. 629 –
636, jul 1984.
[20] ——, “Stochastic complexity and modeling,” The Annals of Statistics, no. 3, pp. 1080–1100, Sep. 1986.
[21] S. Evans, B. Barnett, S. Bush, and G. Saulnier, “Minimum description length principles for detection and classification of ftp exploits,” in
Military Communications Conference, 2004. MILCOM 2004. 2004 IEEE, vol. 1, oct.-3 nov. 2004, pp. 473 – 479 Vol. 1.
[22] N. Wang, J. Han, and J. Fang, “An anomaly detection algorithm based on lossless compression,” in Networking, Architecture and Storage
(NAS), 2012 IEEE 7th International Conference on, 2012, pp. 31–38.
[23] W. Lee and D. Xiang, “Information-theoretic measures for anomaly detection,” in Security and Privacy, 2001. S P 2001. Proceedings.
2001 IEEE Symposium on, 2001, pp. 130–143.
[24] I. Paschalidis and G. Smaragdakis, “Spatio-temporal network anomaly detection by assessing deviations of empirical measures,” Networking,
IEEE/ACM Transactions on, vol. 17, no. 3, pp. 685–697, 2009.
[25] C.-K. Han and H.-K. Choi, “Effective discovery of attacks using entropy of packet dynamics,” Network, IEEE, vol. 23, no. 5, pp. 4–12,
2009.
[26] P. Baliga and T. Lin, “Kolmogorov complexity based automata modeling for intrusion detection,” in Granular Computing, 2005 IEEE
International Conference on, vol. 2, 2005, pp. 387–392 Vol. 2.
[27] H. Shahriar and M. Zulkernine, “Information-theoretic detection of sql injection attacks,” in High-Assurance Systems Engineering (HASE),
2012 IEEE 14th International Symposium on, 2012, pp. 40–47.
[28] Y. Xiang, K. Li, and W. Zhou, “Low-rate ddos attacks detection and traceback by using new information metrics,” Information Forensics
and Security, IEEE Transactions on, vol. 6, no. 2, pp. 426–437, 2011.
[29] F. Pan and W. Wang, “Anomaly detection based-on the regularity of normal behaviors,” in Systems and Control in Aerospace and
Astronautics, 2006. ISSCAA 2006. 1st International Symposium on, 2006, pp. 6 pp.–1046.
[30] E. Eiland and L. Liebrock, “An application of information theory to intrusion detection,” in Information Assurance, 2006. IWIA 2006.
Fourth IEEE International Workshop on, 2006, pp. 16 pp.–134.
[31] M. Li, X. Chen, X. Li, B. Ma, and P. Vitanyi, “The similarity metric,” Information Theory, IEEE Transactions on, vol. 50, no. 12, pp.
3250 – 3264, dec. 2004.
September 12, 2017
DRAFT
39
[32] E. Keogh, S. Lonardi, and C. A. Ratanamahatana, “Towards parameter-free data mining,” in Proceedings of the tenth ACM SIGKDD
international conference on Knowledge discovery and data mining.
ACM, 2004, pp. 206–215.
[33] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume I: Estimation Theory.
Prentice-Hall, 1993.
[34] L. L. Scharf, Statistical Signal Processing: Detection, Estimation, and Time Series Analysis.
Addison-Wesley, 1990.
[35] G. Shamir, “On the mdl principle for i.i.d. sources with large alphabets,” Information Theory, IEEE Transactions on, vol. 52, no. 5, pp.
1939–1955, May 2006.
[36] P. Elias, “Universal codeword sets and representations of the integers,” Information Theory, IEEE Transactions on, vol. 21, no. 2, pp. 194
– 203, mar 1975.
[37] H. V. Poor, An Introduction to Signal Detection and Estimation.
Springer-Verlag, 1994.
[38] W. Hoeffding, “Probability inequalities for sums of bounded random variables,” Journal of the American statistical association, vol. 58,
no. 301, pp. 13–30, 1963.
[39] A. Dembo and O. Zeitouni, Large Deviations Techniques and Applications.
[40] E. L. Lehmann, Testing Statistical Hypotheses.
Springer, 1998.
Springer, 2005.
[41] G. R. Grimmett and D. R. Stirzaker, Probability and Random Processes, Third Edition.
Oxford University Press, 2001.
[42] J. Rissanen, “Modeling by shortest data description,” Automatica, pp. 465–471, 1978.
[43] F. M. J. Willems, Y. Shtarkov, and T. Tjalkens, “The context-tree weighting method: basic properties,” Information Theory, IEEE
Transactions on, vol. 41, no. 3, pp. 653–664, 1995.
[44] J. Ziv and A. Lempel, “A universal algorithm for sequential data compression,” Information Theory, IEEE Transactions on, vol. 23, no. 3,
pp. 337 – 343, may 1977.
[45] ——, “Compression of individual sequences via variable-rate coding,” Information Theory, IEEE Transactions on, vol. 24, no. 5, pp. 530
– 536, sep 1978.
[46] M. Effros, K. Visweswariah, S. Kulkarni, and S. Verdu, “Universal lossless source coding with the burrows wheeler transform,” Information
Theory, IEEE Transactions on, vol. 48, no. 5, pp. 1061 –1081, may 2002.
[47] J. Cleary and I. Witten, “Data compression using adaptive coding and partial string matching,” Communications, IEEE Transactions on,
vol. 32, no. 4, pp. 396 – 402, apr 1984.
[48] A. Moffat, “Implementing the ppm data compression scheme,” Communications, IEEE Transactions on, vol. 38, no. 11, pp. 1917 –1921,
nov 1990.
[49] M. Titchener, “Deterministic computation of complexity, information and entropy,” in Information Theory, 1998. Proceedings. 1998 IEEE
International Symposium on, aug 1998, p. 326.
[50] K. Kawaharada, K. Ohzeki, and U. Speidel, “Information and entropy measurements on video sequences,” in Information, Communications
and Signal Processing, 2005 Fifth International Conference on, 0-0 2005, pp. 1150 –1154.
[51] U. Speidel, “A note on the estimation of string complexity for short strings,” in Information, Communications and Signal Processing,
2009. ICICS 2009. 7th International Conference on, dec. 2009, pp. 1 –5.
[52] U. Speidel, R. Eimann, and N. Brownlee, “Detecting network events via t-entropy,” in Information, Communications Signal Processing,
2007 6th International Conference on, dec. 2007, pp. 1 –5.
[53] U. Speidel and T. Gulliver, “An analytic upper bound on t-complexity,” in Information Theory Proceedings (ISIT), 2012 IEEE International
Symposium on, july 2012, pp. 2706 –2710.
[54] J. Yang and U. Speidel, “String parsing-based similarity detection,” in Information Theory Workshop, 2005 IEEE, aug.-1 sept. 2005, p. 5
pp.
[55] P. A. J. Volf and F. M. J. Willems, “Switching between two universal source coding algorithms,” in Data Compression Conference, 1998.
DCC ’98. Proceedings, 1998, pp. 491–500.
[56] P. Jacquet and W. Szpankowski, “Limiting distribution of lempel ziv’78 redundancy,” in Information Theory Proceedings (ISIT), 2011
IEEE International Symposium on, 31 2011-aug. 5 2011, pp. 1509 –1513.
[57] F. Willems, Y. Shtarkov, and T. Tjalkens, “Reflections on "the context tree weighting method: Basic properties",” Newsletter of the IEEE
Information Theory Society, vol. 47, no. 1, 1997.
[58] F. Willems, “The context-tree weighting method: extensions,” Information Theory, IEEE Transactions on, vol. 44, no. 2, pp. 792–798, Mar
1998.
[59] “Department of Statistics, UC Berkeley,” http://www.stat.berkeley.edu/.
[60] “PhysioBank ATM,” http://physionet.org/cgi-bin/atm/ATM/.
September 12, 2017
DRAFT
40
[61] R. J. Britten, “DNA sequence insertion and evolutionary variation in gene regulation,” Proceedings of the National Academy of Sciences,
vol. 93, no. 18, pp. 9374–9377, 1996. [Online]. Available: http://www.pnas.org/content/93/18/9374.abstract
[62] J. Z. K. Khattak, S. Rauf, Z. Anwar, H. M. Wahedi, and T. Jamil, “Recent advances in genetic engineering - a review,” Current Research
Journal of Biological Sciences, vol. 4, no. 1, 2012.
[63] A. Høst-Madsen, E. Sabeti, and C. Walton, “Information theory for atypical sequences,” in IEEE Information Theory Workshop (ITW’13),
Seville, Spain, 2013.
September 12, 2017
DRAFT
| 7cs.IT
|
arXiv:1611.06245v1 [cs.NE] 18 Nov 2016
Spikes as Regularizers
Anders Søgaard∗
Department of Computer Science
University of Copenhagen
Copenhagen, DK-2200
[email protected]
Abstract
We present a confidence-based single-layer feed-forward learning algorithm S PI RAL (Spike Regularized Adaptive Learning) relying on an encoding of activation
spikes. We adaptively update a weight vector relying on confidence estimates and
activation offsets relative to previous activity. We regularize updates proportionally
to item-level confidence and weight-specific support, loosely inspired by the observation from neurophysiology that high spike rates are sometimes accompanied
by low temporal precision. Our experiments suggest that the new learning algorithm S PIRAL is more robust and less prone to overfitting than both the averaged
perceptron and A ROW.
1
Confidence-weighted Learning of Linear Classifiers
The perceptron [Rosenblatt, 1958] is a conceptually simple and widely used discriminative and
linear classification algorithm. It was originally motivated by observations of how signals are passed
between neurons in the brain. We will return to the perceptron as a model of neural computation, but
from a more technical point of view, the main weakness of the perceptron as a linear classifier is that it
is prone to overfitting. One particular type of overfitting that is likely to happen in perceptron learning
is feature swamping [Sutton et al., 2006], i.e., that very frequent features may prevent co-variant
features from being updated, leading to catastrophic performance if the frequent features are absent or
less frequent at test time. In other words, in the perceptron, as well as in passive-aggressive learning
Crammer et al. [2006], parameters are only updated when features occur, and rare features therefore
often receive inaccurate values.
There are several ways to approach such overfitting, e.g., capping the model’s supremum norm,
but here we focus on a specific line of research: confidence-weighted learning of linear classifiers.
Confidence-weighted learning explicitly estimates confidence during induction, often by maintaining
Gaussian distributions over parameter vectors. In other words, each model parameter is interpreted as
a mean, and augmented with a covariance estimate. Confidence-Weighted Learning CWL [Dredze
et al., 2008] was the first learning algorithm to do this, but Crammer et al. [2009] later introduced
Adaptive Regularization of Weight Vectors (A ROW), which is a simpler and more effective alternative:
A ROW passes over the data, item by item, computing a margin, i.e., a dot product of a weight vector
µ and the item, and updating µ and a covariance matrix Σ in a standard additive fashion. As in
CWL, the weights – which are interpreted as means – and the covariance matrix form a Gaussian
distribution over the weight vectors. Specifically, the confidence is x> Σx. We add a smoothing
constant r(= 0.1) and compute the learning rate α adaptively:
∗
This research is funded by the ERC Starting Grant LOWLANDS No. 313695, as well as by the Danish
Research Council.
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
α=
max(0, 1 − yx> µ)
x> Σx + r
(1)
We then update µ proportionally to α, and update the covariance matrix as follows:
Σ←
Σ − Σxx> Σ
x> Σx + r
(2)
CWL and A ROW have been shown to be more robust than the (averaged) perceptron in several
studies [Crammer et al., 2012, Søgaard and Johannsen, 2012], but below we show that replacing
binary activations with samples from spikes can lead to better regularized and more robust models.
2
2.1
Spikes as Regularizers
Neurophysiological motivation
Neurons do not fire synchronously at a constant rate. Neural signals are spike-shaped with an onset,
an increase in signal, followed by a spike and a decrease in signal, and with an inhibition of the
neuron before returning to its equilibrium. Below we simplify the picture a bit by assuming that
spikes are bell-shaped (Gaussians).
The learning algorithm (S PIRAL) which we will propose below, is motivated by the observation that
spike rate (the speed at which a neuron fires) increases the more a neuron fires [Kawai and Sterling,
2002, Keller and Takahashi, 2015]. Futhermore, Keller and Takahashi [2015] show that increased
activity may lead to spiking at higher rates with lower temporal precision. This means that the more
active neurons are less successful in passing on signals, leading the neuron to return to a more stable
firing rate. In other words, the brain performs implicit regularization by exhibiting low temporal
precision at high spike rates. This prevents highly active neurons from swamping other co-variant,
but less active neurons. We hypothesise that implementing a similar mechanism in our learning
algorithms will prevent feature swamping in a similar fashion.
Finally, Blanco et al. [2015] show that periods of increased spike rate lead to a smaller standard
deviation in the synaptic weights. This loosely inspired us to implement the temporal imprecision at
high spike rates by decreasing the weight’s standard deviation.
2.2
The algorithm
In a single layer feedforward model, such as the perceptron, sampling from Gaussian spikes only
effect the input, and we can therefore implement our regularizer as noise injection [Bishop, 1995].
The variance is the relative confidence of the model on the input item (same for all parameters), and
the means are the parameter values. We multiply the input by the inverse of the sample, reflecting the
intuition that highly active neurons are less precise and more likely to drop out, before we clip the
sample from 0 to 1.
We give the pseudocode in Algorithm 1, following the conventions in Crammer et al. [2009].
3
3.1
Experiments
Main experiments
We extract 10 binary classification problems from MNIST, training on odd data points, testing on
even ones. Since our algorithm is parameter-free, we did not do explicit parameter tuning, but during
the implementation of S PIRAL, we only experiment with the first of these ten problems (left, upper
corner). To test the robustness of S PIRAL relatively to the perceptron and A ROW, we randomly
corrupt the input at test time by removing features. Our set-up is inspired by Globerson and Roweis
[2006]. In the plots in Figure 2, the x-axis presents the number of features kept (not deleted).
We observe two tendencies in the results: (i) S PIRAL outperforms the perceptron consistently with up
to 80% of the features, and sometimes by a very large margin; except that in 2/10 cases, the perceptron
2
Figure 1: Sampling activations from Gaussian spikes.
A ROW
1.00
1.00
1.00
0.95
0.95
0.95
0.90
0.90
0.90
1.00
1.00
0.95
0.95
0.90
0.90
0.85
0.85
0.85
0.85
0.85
0.80
acc
0.80
acc
acc
acc
acc
0.80
0.80
0.80
0.75
0.75
0.75
0.75
0.75
0.70
0.70
0.70
Perceptron
0.65
0.70
Perceptron
0.65
Arow
0.60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.60
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.6
0.7
0.8
0.9
0.60
0.1
1.0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.55
0.1
1.0
0.2
0.3
0.4
0.6
0.7
0.8
0.9
Arow
0.60
0.1
1.0
0.2
0.3
0.4
0.6
0.7
0.8
0.6
0.7
0.8
0.6
0.7
0.8
0.6
0.7
0.8
0.9
1.0
1.00
0.95
0.95
0.90
0.90
0.5
noise
1.00
0.95
0.95
0.5
noise
1.00
0.90
Perceptron
0.65
Arow
noise
1.00
0.95
Perceptron
0.60
Arow
noise
1.00
0.70
0.65
Perceptron
0.65
Arow
0.90
0.90
0.85
0.85
0.85
0.85
0.80
acc
acc
0.80
acc
0.80
acc
acc
0.85
0.80
0.75
0.80
0.75
0.75
0.75
0.70
0.75
0.70
0.70
0.70
Perceptron
0.65
Perceptron
Arow
0.60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Arow
0.65
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.6
0.7
0.8
0.9
0.70
0.65
Perceptron
0.65
Perceptron
0.60
Arow
0.60
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.6
0.7
0.8
0.9
Perceptron
0.65
Arow
0.55
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.6
0.7
0.8
0.9
Arow
0.60
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.9
1.0
noise
S PIRAL
1.00
1.00
0.95
0.95
0.90
0.90
1.00
1.00
0.95
0.95
0.90
0.90
0.85
0.80
acc
acc
acc
0.85
acc
0.80
0.80
0.80
0.75
0.80
0.75
0.75
0.70
0.70
0.75
0.70
0.75
0.70
Perceptron
Perceptron
Spikes
0.65
0.1
0.85
0.85
acc
0.85
1.00
0.95
0.90
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Perceptron
0.65
Spikes
0.65
0.1
1.0
0.2
0.3
0.4
noise
0.5
0.6
0.7
0.8
0.9
0.2
0.3
0.4
noise
1.00
0.5
0.6
0.7
0.8
0.9
Perceptron
Spikes
0.60
0.1
1.0
0.2
0.3
0.4
noise
1.00
0.70
Perceptron
0.65
Spikes
0.60
0.1
1.0
0.5
0.6
0.7
0.8
0.9
Spikes
0.65
0.1
1.0
0.2
0.3
0.4
noise
1.00
0.5
0.9
1.0
noise
1.00
1.00
0.95
0.95
0.95
0.95
0.95
0.90
0.90
0.90
0.90
0.90
0.85
acc
acc
0.85
0.80
0.85
0.80
acc
acc
0.85
acc
0.85
0.75
0.80
0.80
0.80
0.70
0.75
0.75
0.75
0.65
0.75
0.70
Perceptron
0.2
0.3
0.4
0.5
noise
0.6
0.7
0.8
0.9
0.70
Perceptron
Spikes
0.65
0.1
Perceptron
Spikes
1.0
0.70
0.1
0.2
0.3
0.4
0.5
noise
0.6
0.7
0.8
0.9
1.0
0.2
0.3
0.4
0.5
noise
0.6
0.7
0.8
0.9
0.70
Perceptron
0.60
Spikes
0.65
0.1
Perceptron
Spikes
1.0
0.55
0.1
0.2
0.3
0.4
0.5
noise
0.6
0.7
0.8
0.9
Spikes
1.0
0.65
0.1
0.2
0.3
0.4
0.5
0.9
1.0
noise
Figure 2: Performance over noise levels (percentage of features kept). First two lines compare the
perceptron (blue) and A ROW (green); the third and fourth compare the perceptron (blue) and S PIRAL
(green).
3
Algorithm 1 S PIRAL. Except for lines 8–10, this is identical to A ROW.
1: r = 0.1, µ0 = 0, Σ0 = I, {xt | xt ∈ Rd }, v = 0
2: for t < T do
3:
xt = xt ·
4:
if vt > v then
5:
v = vt
6:
end if
7:
vt = x>
t Σt−1 xt (sampling activations from Gaussian spikes)
8:
vt = 0 < vt < 1 (clipping values outside the [0,1] window)
9:
νt ∼ N (µt−1 , vvt )
10:
xt = xt · (1 − νt )
11:
mt = µt−1 · xt
12:
if mt yt < 1 then
13:
14:
max(0,1−y x> µ
)
t t
t−1
αt =
x>
t Σt−1 xt +r
µt = µt−1 + αt Σt−1 yt xt
Σ
15:
Σt = t−1
16:
end if
17: end for
18: return µT , ΣT
−Σt−1 xt x>
t Σt−1
x>
t Σxt +r
is better with only 10% of the features. (ii) In contrast, A ROW is less stable, and only improves
significantly over the perceptron under mid-range noise levels in a few cases. The perceptron is
almost always superior on the full set of features, since this is a relatively simple learning problem,
where overfitting is unlikely, unless noise is injected at test time.
3.2
Practical Rademacher complexity
We compute S PIRAL’s practical Rademacher complexity as the ability of S PIRAL to fit random
re-labelings of data. We randomly label the above dataset ten times and compute the average
error reduction over a random baseline. The perceptron achieves a 5% error reduction over a
random baseline, on average, overfitting quite a bit to the random labelling of the data. In contrast,
S PIRAL only reduces 0.6% of the errors of a random baseline on average, suggesting that it is almost
resilient to overfitting on this dataset.
4
Conclusion
We have presented a simple, confidence-based single layer feed-forward learning algorithm S PI RAL that uses sampling from Gaussian spikes as a regularizer, loosely inspired by recent findings in
neurophysiology. S PIRAL outperforms the perceptron and A ROW by a large margin, when noise is
injected at test time, and has lower Rademacher complexity than both of these algorithms.
Acknowledgments
References
Cristopher Bishop. Training with noise is equivalent to tikhonov regularization. Neural Computation,
7(1):108–116, 1995.
Wilfredo Blanco, Catia Pereira, Vinicius Cota, Annie Souza, Cesar Renno-Costa, Sharlene Santos,
Gabriella Dias, Ana Guerreiro, Adriano Tort, Adriao Neto, and Sidarta Ribeiro. Synaptic homeostasis and restructuring across the sleep-wake cycle. PLoS Computational Biology, 11:1–29,
2015.
Koby Crammer, Ofer Dekel, Joseph Keshet, Shai Shalev-Shwartz, and Yoram Singer. Online
passive-agressive algorithms. Journal of Machine Learning Research, 7:551–585, 2006.
Koby Crammer, A Kulesza, and Mark Dredze. Adaptive regularization of weighted vectors. In NIPS,
2009.
4
Koby Crammer, Mark Dredze, and Fernando Pereira. Confidence-weighted linear classification for
text categorization. Journal of Machine Learning Research, 13:1891–1926, 2012.
Mark Dredze, Koby Crammer, and Fernando Pereira. Confidence-weighted linear classification. In
ICML, 2008.
Amir Globerson and Sam Roweis. Nightmare at test time: robust learning by feature deletion. In
ICML, 2006.
Fusao Kawai and Peter Sterling. cgmp modulates spike responses of retinal ganglion cells via a
cgmp-gated current. Visual Neuroscience, 19:373–380, 2002.
Clifford Keller and Terry Takahashi. Spike timing precision changes with spike rate adaptation in the
owl’s auditory space map. Journal of Neurophysiology, 114:2204–2219, 2015.
Frank Rosenblatt. The perceptron: a probabilistic model for information storage and organization in
the brain. Psychological Review, 65(6):386–408, 1958.
Anders Søgaard and Anders Johannsen. Robust learning in random subspaces: equipping NLP
against OOV effects. In COLING, 2012.
Charles Sutton, Michael Sindelar, and Andrew McCallum. Reducing weight undertraining in
structured discriminative learning. In NAACL, 2006.
5
| 9cs.NE
|
arXiv:1801.04486v5 [cs.AI] 19 Mar 2018
Can Computers Create Art?
Aaron Hertzmann
Adobe Research∗
March 20, 2018
Abstract
This essay discusses whether computers, using Artificial Intelligence (AI), could create art. The first part concerns AI-based tools for assisting with art making. The history
of technologies that automated aspects of art is covered, including photography and animation. In each case, we see initial fears and denial of the technology, followed by a
blossoming of new creative and professional opportunities for artists. The hype and reality
of Artificial Intelligence (AI) tools for art making is discussed, together with predictions
about how AI tools will be used. The second part speculates about whether it could ever
happen that AI systems could conceive of artwork, and be credited with authorship of an
artwork. It is theorized that art is something created by social agents, and so computers
cannot be credited with authorship of art in our current understanding. A few ways that
this could change are also hypothesized.
1
Introduction
Artificial Intelligence (AI) research has made staggering advances recently, including many
publicly-visible developments in web search, image recognition, conversational agents, and
robotics. These developments have stoked fear about Artificial Intelligence’s effect on many
aspects of society. In the context of art, news media hype presents new image and video creation algorithms as if they are automating the creation of art. Perhaps they will empower
everyday users while putting artists out of work... and, while they’re at it, rob us of our humanity?
Beyond the hype, confusion about how technology influences art pervades serious discussions. Professional artists often are concerned that computers might put them out of their jobs
∗
This essay expresses my own opinions and not those of my employer. This manuscript was submitted to
the journal Arts, Special Issue on The Machine as Artist (for the 21st Century) ( http://www.mdpi.com/
journal/arts/special_issues/Machine_Artist ), by invitation. Please send comments to [email protected]. However, I cannot promise to respond to all feedback.
1
[48] — a concern I’ve heard for decades. Some practitioners present their algorithms as themselves, potentially, artists [10, 21, 58], as do some journalists [51, 19]. And, I was recently
contacted by a prominent social psychologist who, inspired by recent results with neural networks, wished to conduct experiments to assess whether ordinary people might be willing to
buy artwork made by a computer, and, if so, why. It was assumed that computers were already
happily making their own artwork.
On the other hand, whenever I have informally asked friends or colleagues the question of
whether computers can create art, the answer is usually a decisive “No.” Art requires human
intent, inspiration, a desire to express something. Thus, by definition, there is no such thing as
art created by a computer... why would anyone worry? The concepts of art and inspiration are
often spoken of in mystical terms, something special and primal beyond the realm of science
and technology; as if humans can only create art because only humans have “souls.” Surely,
there should be a more scientific explanation.
In this essay, I tackle the question of “Can Computers Create Art?” This might seem
like a simple question, but it hides substantial complications. People sometimes approach
this question as one of technological capability: are computers “smart” enough or “creative”
enough; analogous to asking whether computers can perform well-defined tasks like solving
mathematical problems or accurately resolving search queries. But, I argue, this is not the right
way to look at the question, and I do not attempt any sort of survey of the existing technology.
In my opinion, whether computers can create art is not a computer science question so much
as it is a question of philosophy of art and human psychology.
In the first part of this essay, I discuss the history and current state of automation for art.
I begin with some historical perspective: previous moments in history when new technologies
automated image and film creation, particularly the invention of photography. In each case,
we see that these new technologies caused fears of displacing artists, when, in fact, the new
technology both created new opportunities for artists while invigorating the traditional media.
I argue that new technologies benefit art and artists, creating new tools and modes of expression, and new styles of expression. Sadly, art and science are often viewed as being separate,
or in even opposition or competition [61]. Rather than being a foe, technological development
stimulates so much of the continued vitality of art, and new artistic technologies create new
job opportunities. Often, artists are important contributors to these innovations; artists and
technologists have much in common, as tinkerers, experimenters, and explorers. Our new AI
technologies follow this trend, and will for the forseeable future: new AI algorithms will provide new tools for expression and transform our art in culture in positive ways, just as so many
other technologies have in the past.
Computers do not create art, artists using computers create art. Despite many decades of
procedural and computer-generated art, there has never been a computer which has widely
accepted as the author of an artwork.
In the second part, I discuss whether this could ever change: would we ever agree to assign
authorship to a computer? I argue that artistic creation is primarily a social act, an action that
people primarily as an interaction with other humans in society. This means that computers
cannot create art, just as people do not give gifts to their coffeemakers or marry their cars.
2
However, the boundaries of art are fluid, and I discuss scenarios where AI algorithms could
come to be accepted as artists, along with some of the dangers of too eagerly accepting software
as artists.
This essay expresses my point of view on these topics, as someone who has developed various kinds of technology for art, has followed the development of new technologies affecting art
and culture over the years, and, in a few small instances, exhibited my own work as artworks.
I have written it in a style that is meant to be accessible to a general audience, since these are
topics that many kinds of people care about.
2
History and Current Practice
Before making general claims about how computer tools relate to art, I begin by first describing
several useful historical examples, including the inventions of photography, film, and computer
animation. These examples show previous situations in which a new technology appeared
poised to displace artists, but, in reality, provided new opportunities and roles for artists. I also
discuss how procedural art and computer art employs automation. In so many of these fields,
outsiders view the computer or the technology as doing the work of creativity. Yet, in each of
these cases, the human artist or artists behind the work are the true authors of the work.
I argue that, throughout history, technology has expanded creative and professional opportunities for artists dramatically, by providing newer and more powerful tools for artists. The
advent of new technologies often cause fears of displacement among traditional artists. In fact,
these new tools ultimately enable new artistic styles and inject vitality into art forms that might
otherwise grow stale. These new tools also make art more accessible to wider sections of society, both as creators and as consumers. These trends are particularly visible in the past two
centuries since the Industrial Revolution.
2.1
How photography became an artform
“From today, painting is dead!” – Paul Delaroche, painter, at a demonstration of the daguerreotype in 18391
For lessons from the past about AI and art, perhaps no invention is more significant than
photography.2 Prior to the invention of photography, realistic images of the world could only be
produced by artists. In today’s world, we are so swamped with images that it is hard to imagine
just how special and unique it must have felt to see a skillfully-executed realistic painting. The
technical skills of realism were inseparable from other creative aspects. This changed when
photography automated the task of producing images of the real world.
In 1839, the first two commercially-practical photographic processes were invented: LouisJacques-Mandé Daguerre’s dagguereotype, and William Henry Fox Talbot’s negative-positive
process. They were mainly presented as ways to produce practical records of the world. Of the
1
2
The historical information from this section is distilled from two texts: [54] and [57].
This connection has been made previously [32, 1], but I expand upon it much more here.
3
(a)
(b)
Figure 1: (a) A traditional silhouette portrait. (b) Daguerreotype portrait of Abraham Lincoln,
1846. Photographic techniques like this completely displaced previous portraiture techniques.
two, the daguerreotype was more popular for several decades, because Talbot’s process was
restricted by patents. Improvements to Talbot’s method eventually made the daguerreotype
obsolete and evolved into modern film processes.
Portraiture was a main driver for early adoption. Then, as today, people enjoyed possessing
pictures of their friends, loved-ones, and ancestors. Portrait painting was only available to
aristocrats and the very wealthy; even today, portrait painting is a symbol of great wealth and
status. In the 18th century, a number of inexpensive alternatives were developed, such as the
silhouette, a tiny black representation of an individual’s outline (Figure 1(a)), typically handcut by an artisan out of black paper. The daguerreotype offered an economical way to create
a realistic portrait (Figure 1(b)). It was very slow and required locking the subject’s head in
place with a head brace for several minutes, while the subject gripped their chair tightly, so
as not to flutter their fingers. Nonetheless, numerous daguerreotype studios arose and became
commonplace as technologies improved, and many portraitists switched to this new technology.
By 1863, a painter-photographer named Henri Le Secq said “One knows that photography has
harmed painting considerably, and has killed portraiture especially, once the livelihood of the
artist.” Photography largely replaced most older forms of portraiture, such as the silhouette,
and no one seems to particularly regret this loss. As much as I appreciate the mystery and
beauty at looking at old etchings and portraits, I’d rather use my mobile phone camera for my
own pictures than to try to paint them by hand.
Another early use for the daguerreotype was to produce souvenirs for tourists: by 1850,
daguerreotypes of Roman ruins completely replaced the etchings and lithographs that tourists
had previously purchased. As the technology improved, photography became indispensable as
a source of records for engineering projects and disappearing architectural ruins, as well as for
4
Figure 2: The Unhappy Painter, by Theodor Hosemann, 1843, satirizes the painter, a victim of
progress, made obsolete by a daguerreotype.
documentary purposes, such as Matthew Brady’s photographs of the horrors of the American
civil war.
“Is photography art?” This question was debated for many decades, coalescing into three
main positions. Many people believed that photography could not be art, because it was made
by a mechanical device rather than by human creativity. Many artists were dismissive of photography, and saw it as a threat to “real art.” A second view was that photography could be
useful to real artists, such as for reference, but should not be considered as equal to drawing
and painting. Finally, a third group, relating photography to established forms like etching and
lithography, felt that photography could eventually be as significant an art form as painting.
Photography ultimately had a profound and unexpected effect on painting. Painters’ mimetic
abilities had been improving over the centuries. Many painters of the 19th century, such as the
Pre-Raphaelites like John Everett Millais and Neoclassicists like Ingres, painted depictions of
the world with dazzling realism, more than had ever been seen before. However, cameras
became cheaper, lighter, and easier to use, and grew widespread among both amateurs and
professionals. Realistic photographs became commonplace by the end of the 19th century. If
photorealism could be reduced to a mechanical process, then what is the artist’s role?
This question drove painters away from visual realism toward different forms of abstraction. James McNeill Whistler’s Tonalist movement created atmospheric, moody scenes; he
wrote: “The imitator is a poor kind of creature. If the man who paints only the tree, or the
flower, or other surface he sees before him were an artist, the king of artists would be the photographer. It is for the artist to do something beyond this.” The Impressionists, who sought
to capture the perceptions of scenes, were likely influenced by the perceptual “imperfections”
of photographs. In contrast, Symbolists and post-Impressionist artists moved away from perceptual realism altogether. Edvard Munch wrote “I have no fear of photography as long as it
cannot be used in heaven and in hell. ... I am going to paint people who breathe, feel, love,
and suffer.” Vincent Van Gogh, describing his artistic breakthroughs around 1888, wrote to his
brother: “You must boldly exaggerate the effects of either harmony or discord which colors
produce. It is the same thing in drawing — accurate drawing, accurate color, is perhaps not the
5
essential thing to aim at, because the reflection of reality in a mirror, if it could be caught, color
and all, would not be a picture at all, no more than a photograph.” Photography continued to
influence modern art of the 20th century; one can infer a significant influence of Étienne-Jules
Marey’s multiple-exposure photography on Futurism and Cubism, e.g., in Duchamp’s Nude
Descending A Staircase.
One can argue, in fact, that photography was one of the major catalysts of the Modern Art
movement: its influence led to decades of vitality in the world of painting, as artists were both
inspired by photographic images and pushed to go beyond realism.
Meanwhile, the Pictorialist movement, begun around 1885, was an attempt to firmly establish photography as an art form. Pictorialists introduced much more artistic control over the
photographs, often using highly-posed subjects as in classical painting, and manipulating their
images in the darkroom. Many of their works had a hazy, atmospheric work, similar to Tonalism, that deliberately softened the realism of high-quality photography. They seemed to be
deliberately mimicking the qualities of the fine art painting of the time, and today seem rather
affected. They pursued various strategies toward legitimization of their work as art form, such
as the organization of photographic societies, periodicals, and juried photography exhibitions
[64]. Their works and achievements made it harder and harder to deny the artistic contributions of photography; culminating in the “Buffalo Show,” organized by Alfred Stieglitz at the
Albright Gallery in Buffalo, NY, the first photography exhibition at an American art museum,
in 1910. Photography was firmly established as an art, and free to move beyond the pretensions
of Pictorialism.
This story provides several lessons that are directly relevant for AI as an artistic tool. At
first, photography, like AI, was seen by many as non-artistic, because it was a mechanical
process. Some saw photography as a threat and argued against its legitimacy. Photography did
displace old technologies that had fulfilled non-artistic functions, such as portraiture’s social
function. Some artists enthusiastically embraced the new technology, and began to explore its
potential. As the technology improved, and became more widespread over nearly a century,
artists learned to better control and express themselves with the new technology, until there was
no more real controversy over the status of photography. The new technology made imagemaking much more accessible to non-experts and hobbyists; today, everyone can experiment
with photography. Furthermore, the new technology breathed new life into the old art form,
provoking it toward greater abstraction. Wherever there is controversy in AI as an artistic tool,
I predict the same trajectory. Eventually, new AI tools will be fully recognized as artists’ tools;
AI tools may stimulate traditional media as well, e.g., the New Aesthetic [63].
2.2
The technology of live-action cinema
The story of filmmaking and technology has important lessons about how artists and technologists can work together, each pushing the other further. Most of the early photographers were,
by necessity, both artists and technologists, experimenting with new techniques driven by their
art or to inspire their art. But, in film and animation, the interaction has been much more central
to the art form.
6
Figure 3: The interplay of painting and early photography. (a) By the 19th century, Western
painters had achieved dazzling levels of realism. (b) Early cameras took low-quality (though
evocative) pictures. This daguerreotype took over ten minutes to expose. (c) However, camera
technology steadily improved, capturing greater and greater realism, in much faster exposures.
(d) This challenged painters to create works that were not about hyper-realistic depiction, such
as Whistler’s Tonalist Nocturne. (e) The Pictorialist photographers attempted to establish photography as an art form by mimicking the styles and abstraction of painting. Works: (a)
Ophelia, John Everett Millais, 1851 (b) Boulevard du Temple, Daguerre, 1838; (c) Portrait of
Sarah Bernhardt, Félix Nadar, 1864; (d) Nocturne in Blue and Gold: Old Battersea Bridge,
James MacNeill Whistler, c. 1872-1875; (e) Morning, Clarence H. White, 1908
7
(a)
(b)
(c)
Figure 4: Technological developments in the art of filmmaking: (a) the first captured film, of
workers leaving the Lumiere Brothers’ factory (1895), (b) George Méliès’ 1908 A Trip to The
Moon filmed like a stage play but with fantastical special effects, (c) Citizen Kane, which used
numerous experimental camera and lens effects to tell the story.
The history of film is filled with artist-tinkerers, as well as teams of artists and technologists. The Lumière Brothers created the first film, a simple recording of workers leaving their
factory, but also experimented with a wide range of camera technologies, color processing,
and artistic ways to use them. The stage magician George Méliès filmed fantastical stories
like A Trip to the Moon, employing a wide range of clever in-camera tricks and techniques
to create delightfully inventive and beguiling films. Walt Disney employed and pushed new
technologies of sound and color recording, and drove other innovations along the way, such as
the multiplane camera. Many of Orson Welles’ innovative film techniques were made possible
by new camera lenses employed by his cinematographer Gregg Toland. The introduction of
more portable camera and audio equipment enabled the experiments of the French New Wave,
who, in turn, influenced young American directors like Francis Ford Coppola and George Lucas. George Lucas’ team for Star Wars was an early developer of many new visual effects
on a shoestring budget (think of Ben Burtt hitting telephone guy-wires to create the “blaster”
sound effect), as well as an early innovation in digital film editing and compositing [56]. Digital and computer graphics technology, have, obviously, revolutionized film storytelling since
then, with directors like Michel Gondry and James Cameron pushing the technology further
into unforeseen directions. In each case, we see technologies rapidly adopted by directors to
create new storytelling techniques and styles, transforming the medium over and over.
2.3
3D computer animation: a collaboration
3D computer animation as an artform was pioneered by Pixar Animation Studios, and that
success is due to the close collaboration of artists and engineers [52]. It all began with Ed
Catmull, an animation enthusiast who received a PhD in computer science in 1974. In his
thesis, he invented several core techniques that every major 3D computer graphics system uses
today. During his time in graduate school, he quietly set a goal for himself: to make the world’s
first computer-animated film [5]. Consequently, he founded the Lucasfilm Computer Division,
and hired a team of brilliant engineers to invent computer systems to be used for film-making.
However, none of this group could animate, that is, bring a character to life through movement.
Hence, Catmull recruited John Lasseter, an animator trained deeply in the Disney tradition.
8
Through tight collaboration between Lasseter and the technical staff, they were able to invent
new technologies and discover together how computer animation could start to become its own
art form [41]. This group spun out as Pixar, and, over the following years, invented numerous
technical innovations aimed at answering the needs set out by Pixar’s artists; in turn, the artists
were inspired by these new tools, and pushed them to new extremes, and so on. One of their
mantras was “Art challenges technology, technology challenges art” [5].
Pixar, by design, treats artists and engineers both as crucial to the company’s success and
minimizes any barriers between the groups. When I worked there during a sabbatical, despite
my technical role, I had many energizing conversations with different kinds of artists, attended
many lectures on art and storytelling, sketched at an open life drawing session, watched a performance of an employee improv troupe, and participated in many other social and educational
events that deliberately mixed people from different parts of the company. This is the culture
that, though it still has some flaws to address, achieved so many years of technical and creative
innovation, and, ultimately, commercial and artistic success.
Computer animation is another technology that scared traditional artists. In the early, prePixar days, Catmull’s group made many attempts to interest Disney animators in their work
[52]. Alvy Ray Smith later said: “Animators were frightened of the computer. They felt that
it was going to take their jobs away. We spent a lot of time telling people, No, it’s just a
tool-it doesn’t do the creativity!’ That misconception was everywhere” [50]. It is a common
misconception that computer animation just amounts to the computer solving everything; a
programmer presses a button and the characters just move on their own. In reality, computer
animation is extraordinarily labor-intensive, requiring the skills of talented artists (especially
animators) for almost every little detail. Character animation is an art form of extreme skill
and talent, requiring laborious effort using the same fundamental skills of performance — of
bringing a character to life through pure movement — as in conventional animation [41].
Traditional cel animation jobs did not last at Disney, for various reasons. Disney Feature
Animation underwent a renaissance in the early ’90s, starting with The Little Mermaid. Then,
following some changes in management, the Disney animation began a slow, sad decline. After
releasing duds like Brother Bear and Home on the Range, management shut down all traditional 2D animation at Disney, and converted the studios entirely to 3D computer animation.
Many conventional animators were retrained in 3D animation, but Disney’s first 3D animation,
Chicken Little was still a dud. Following Disney’s acquisition of Pixar several years later, they
revived Disney’s beloved 2D animation productions. The result, a charming and enjoyable
film called The Princess and The Frog, performed so-so at the box office, and, moreover, the
animators’ creative energy was focused on the newer 3D art form [5]. Today, traditional 2D
animation at Disney is dead.3 Today, computer animation is a thriving industry, and it thrives
in many more places than cel animation ever did: at many different film studios, in visual effects for live-action films, in video games, television studios, web startups, independent web
animators, and many more. There are now more types of opportunities for animators than ever
3
Traditional animation styles are still vital in countries like Japan and France that, unlike America, do not
believe that animation is just for kids. Even so, their visual styles have evolved considerably due to computer
technology.
9
Figure 5: Procedural artworks in the fine art world of galleries and art museums. Left: Electric
Sheep, by Scott Draves, evolves dazzling procedural abstract animations based on thousands
of votes. Right: The Top Grossing Film of All Time, 1 x 1 (2000) by Jason Salavon, shows the
average color of each frame of the movie Titanic.
before. The story here is not the destruction of jobs, but the evolution and growth of an art
form through technology. This is another story that contradicts the popular notion of art and
technology operating in conflict, when, in fact, the opposite is usually true.
2.4
Procedural artwork
In the art world, there is a long tradition of procedural artwork. Jean Arp created artworks
governed by laws of chance in the 1910s (or so he claimed), and, beginning in the 1950s, John
Cage used random rules to compose music. The term Generative Art appears to have originated
in the 1960s. Sol LeWitt’s wall drawings are provided as lists of precise instructions; people
are still drawing new versions of his paintings after his death [11]. Starting in the 1970’s,
classically-trained painter Harold Cohen began exhibiting paintings generated by a program
he wrote called AARON [6]. Since the 1980s, many current artists, such as Karl Sims, Scott
Snibbe, Golan Levin, Scott Draves, and Jason Salavon, produce abstract artworks by writing
computer programs that produce either static images, or create interactive artistic experiences
and installation works (Figure 5). In Sims’ and Draves’ work, the artwork “evolves” according
to audience input. The popularity of the Processing computer language for artists speaks to the
growth of procedural art.
In each of these cases, despite the presence of procedural, emergent, and/or crowdsourced
elements, the human behind it is credited as the author of the artwork, and it would seem
perverse to suggest otherwise. The human has done all of the creative decision-making around
the visual style, of designing a framework and process, of testing and evaluating alternative
algorithms, and so on.
2.5
State of the art in computer science research
Recent developments in computational artistic image synthesis are quite spectacular. But they
should not be mistaken for AI artists.
10
(a)
(b)
Figure 6: Painterly rendering algorithms that process an input photograph, using hand-coded
rules and algorithms. (a) Paul Haeberli’s interactive painting system [30]. (b) My automatic
painting system [31] processes a photograph without user input aside from selecting some
parameter settings.
Non-Photorealistic Rendering (NPR) is a subfield of computer graphics research [55] that
I have worked in for many years. NPR research develops new algorithms and artistic tools
for creating images inspired by the look of a conventional media, such as painting or drawing.
Paul Haeberli’s groundbreaking 1990 paper [30] introduced a paint program that began with a
user-selected photograph. Whenever the user clicked on the canvas (initially blank), the system
placed a brush stroke with color and orientation based on the photograph. In this way, a user
could quickly create a simple painting without any particular technical skill (Figure 6(a)). In a
follow-up paper, Pete Litwinowcz automated the process entirely, by placing brush strokes on
a grid [43]. My own first research paper arose from experimenting with modifications to his
algorithm: the method that I came up with places long curved strokes, beginning with a large
strokes that were then refined by small details [31] (Figure 6(b)). The algorithm was inspired
by my experience with real painting, and the way artists often start from a rough sketch and
then refine it.
This type of artistic algorithm design reflects the majority of computer graphics research
in this area [33, 55]. The algorithms are automated, but we can explain in complete detail why
the algorithm works and the intuitions about artistic process it embodies. This mathematical
modeling of artistic representation continues the tradition begun in the Renaissance with Filippo Brunelleschi’s invention of linear perspective, a viewpoint I have written about elsewhere
[34].
At some point, I found it very difficult to embody richer intuitions about artistic process
11
Figure 7: Our Image Analogies algorithm [35], which stylizes a photograph in the style of a
given artwork; in this case, Van Gogh’s Starry Night over the Rhône.
Figure 8: The Neural Style Transfer Algorithm [24], which stylizes a photograph in the style of
a given artwork; in this case, Van Gogh’s The Starry Night. This algorithm has led to numerous
new apps and research in stylization.
into source code. Instead, inspired by recent results in computer vision [20], I began to develop
a method for working from examples. My collaborators and I published this method in 2001,
calling it “Image Analogies” [35]. We presented the work as learning artistic style from example. But the “learning” here was quite shallow. It amounted to rearranging the pixels of the
source artwork in a clever way, but not generalizing to radically new scenes or style (Figure 7).
Since then, other researchers have improved the method substantially, making it much more
robust [23].
In 2016, Leon Gatys and his colleagues published a new breakthrough in this space, Neural
Style Transfer [24]. Based on recent advances in neural networks, their method transfers certain
neural network correlation statistics from a painting to a photograph, thus producing a new
painting of the input photograph (Figure 8). The method is still “shallow” in a sense — there
is no “understanding” of the photograph or the artwork — but the method seems to be more
robust than the original Image Analogies algorithm. This paper led to a flurry of excitement
and new applications, including the popular Prisma app and Facebook’s Live Video stylization,
as well as many new research papers improving upon these ideas. This work is ongoing today.
Another development which received considerable attention in 2015 was the invention of
DeepDreams by Alexander Mordvintsev [46], who, developing a visualization tool for neural
networks, discovered that a simple activation excitation procedure produced striking, hallu12
Figure 9: Images that we generated using a neural network trained on different subsets of a
database of artistic imagery [71]. Each one is meant to typify a single stylistic category or a
medium, though the biases of the convolutional neural network architecture are also evident.
cinatory imagery of a type we had never seen before. There are many other current projects,
particularly those around Generative Adversarial Networks [28] and Project Magenta at Google
[44], that also show promise as new artistic tools. For example, Figure 9 shows images that we
generated by visualizing trends learned by a neural network from a large collection of artistic
images in different styles. A variety of related images are produced by Creative Adversarial
Networks [21]; it is a visual style that seems familiar but not the same as what we are familiar with (and is probably driven, in part, by the biases of the convolutional neural network
representation).
In each of these cases, the artworks are produced by a human-defined procedure, and the
human is the author of the imagery in each case.
2.6
Artificial Intelligence is not intelligent
Unfortunately, there has been a considerable amount of media hype around AI techniques; in
the news media, algorithms are often anthropomorphized, as if they have the same consciousness as humans (e.g., [72]), sometimes described as artists [51, 19]. In fact, we do not really
know what consciousness is (despite many theories), or what it would mean to embody it in an
algorithm.
Today, the most-successful AI and machine learning algorithms are best thought of as glorified data-fitting procedures [4]. That is, these algorithms are basically like fitting a curve to a
set of datapoints, except with very sophisticated ways to fit high-dimensional curves to millions
of datapoints. When we as researchers speak of “training” an algorithm, or an algorithm that
“learns,” it is easy to misinterpret this as being the same thing as human learning. But these
words mean quite different things in the two contexts. In general, “training” a model to learn a
task involves careful human effort to formulate the problem, acquire appropriate data, and test
13
different formulations. It is laborious and requires considerable expertise and experimentation.
When a new task needs to be solved, the human starts over.
Compared to human intelligence, these algorithms are brittle and bespoke. For example,
image recognition algorithms, which has undergone breathtaking breakthroughs in the past
decade, are now widely used in consumer products. Yet they often fail on inputs that suggest bizarre misunderstandings; the existence and robustness of adversarial examples [66] and
procedurally-evolved images [47] demonstrates that these algorithms haven’t really learned
anything like human-level understanding. The algorithm has learned no general purpose understanding of the world. It is like a tourist in a foreign country that can repeat and combine
phrases from the phrasebook, but lacks true understanding of the language or culture that it
is visiting. These systems have no autonomy except within the narrow scope for which they
were trained, and typically failsafes must be put in place as well (e.g., Google Photos no longer
classifies anything as a “gorilla” because of one high-profile failure [60]).
There are, indeed, fascinating parallels between human learning and machine learning, and
it does seem likely that humans are, in some way, machines for optimization of evolutionary
principles [2, 39, 67]. But going from these high-level analogies to actual machine intelligence
is a problem for which the solution is not even on the horizon.
3
How technology changes art
Based on this history, I now make several specific claims about how technology changes art.
Far from replacing artists, new technologies become new tools for artists, invigorating and
changing art and culture. These claims apply equally to current developments in AI as they did
to previous developments like photography and animation.
3.1
Algorithms are artists’ tools
In every technology that we currently employ — whether photography, film, or software algorithm — the technologies and algorithms we use are basic tools, just like brushes and paint.4
The same is true for the new AI-based algorithms that are appearing. They are not always
predictable, and the results are often surprising and delightful — but the same could be said
for the way watercolor flows on the page. There is no plausible sense in which current systems
reflect “true” artificial intelligence: there is always a human behind the artwork.
Applying the same standard to the current research in neural networks and neural style
transfer, it would seem equally perverse to assign authorship of their outputs to the software
itself. The DeepDream software was authored by a human; another human then selected an
4
This view reflects, I think, the conventional wisdom in the computer graphics research field, which is my
research background. This field has always had close ties with certain artistic communities, especially computer
animation and visual effects, and, based on this experience, the field is often resistant to attempts to automate
creative tasks. In contrast, artificial intelligence researchers use terminology much more aspirationally, historically
using words like “intelligence,”” “learning,” and “expert systems” in ways that far simpler than the human versions
of these things.
14
input image, and experimented with many parameter settings, running the software over and
over until obtaining good results. Indeed, in a recent art exhibition meant to promote these
methods and their exploration [68], human artists were credited for each of the individual
works. The same process of selection of tools and inputs, adjusting settings and even modifying
code, and iterating until a desirable output is produced, occurs in all current computer artworks.
There have been a few cases where AI algorithms have been presented as artists or potential
artists [10, 58, 19, 51]. For the above reasons, I think this claim misunderstands the nature of
procedural art.
In short, in our present understanding, all art algorithms, including methods based on
machine learning, are tools for artists; they are not themselves artists.
3.2
Technology helps art stay vital
Rather than being afraid of the new technologies, we should be enthusiastic about the new artworks that it will enable artists to produce. When we think of art of having external influences,
we normally think of social or political influences, but ignore the effect of new tools. In contrast, I argue that, especially in the 19th and 20th Centuries, technological developments have
played a pivotal role in advancing art, in keeping it vital and injecting fresh ideas. The stories
I gave of photography and cinema include many of examples of this. However, the effect is far
more widespread.
One of the most important breakthroughs in the history of Western art was the invention
of oil paint by Flemish painters such as Jan van Eyck in the 15th century [13]. Previously,
painting had been done primarily with tempera, which lack subtle coloration, and fresco, which
was very cumbersome to paint. Oil paint had existed in some form for centuries before, but
van Eyck and others found new techniques and compounds that gave them a very practical
new medium that was fast drying and allowed them to create richer colors, sharp edges, hard
surfaces, and a much wider gamut of colors. The rich light and color that we associate with the
Northern Renaissance and the Italian Renaissance are due to this technology (Figure 10).
In each decade since the 1950s, many of our culturally-important works used technology
that had only been invented within the previous ten years. For example, most technology used
in today’s feature films did not exist ten years ago (e.g., widespread use of HD digital cameras;
facial performance capture); the same goes for artworks using smartphones and crowdsourcing;
artworks involving white LEDs and Arduino controllers; DJs performing on stage behind their
laptops; and so on. Even the most vérité-seeming romantic comedies frequently involve recent
digital video editing and digital backdrops.
Conversely, artistic styles that fail to change become stale and lose their cultural relevance;
the adoption and exploration of new technology is one of the main ways that art stays vibrant.
For example, the introduction of synthesizer music into 1980s pop music created a new sound
that was exciting and modern. The sound diversified as the tools improved, until grunge became
popular and made the 80s synthpop sound seem superficial and old-fashioned. Nowadays, a
recent revival of 70s and 80s instruments by bands like Daft Punk and LCD Soundsystem
seems most exciting at times when they are creating new types of music using old instruments.
15
(a)
(b)
Figure 10: The development of the oil painting technology changed painting as an art form.
(a) Fresco painting by Michaelangelo on the Sistine Chapel (ca. 1508). The fresco process
was difficult and achieved limited tonal range. Today, fresco is defunct as a medium. (b) Oil
painting by Jan van Eyck for the Ghent Alterpiece (ca. 1430). Much richer colors and lighting
are possible with oil paint.
In contrast, the swing music revival of the 1990s never went anywhere (from bands like Big
Bad Voodoo Daddy and Squirrel Nut Zippers) in my opinion, because the bands aped classic
styles with classic instruments, without inventing anything particular original themselves.
In each era, radical technological innovations are met by artists with both enthusiasm and
rejection. For example, when the Moog synthesizer became popular, it was adopted by bigname bands like Emerson, Lake, and Palmer. Other bands felt that twisting knobs to make
music was cheating: Queen’s album covers proudly state that the band did not use synthesizers.
Robert Moog described one New York musician who said of the instrument “This is the end
of the world” [62]. It now seems silly to imagine that people might have ever categorically
objected to synthesized music, or to the scratching and sampling of hip-hop DJs, just as it now
seems silly that people once rejected waltzing, the Impressionists, and the Rite of Spring as
invalid or immoral.
In addition to stimulating professional artists, new tools make art more accessible to larger
portions of society. Photography was once accessible only to the most determined early
adopters, but has continually become easier, faster, and more compact, to the point where
nearly everyone carries a mobile phone camera in their pocket or purse. The same goes for the
tools of cinematography (from hand-cranked to heavy cameras to Steadicams to handycams to
iPhones), and so on. Modern computers give nearly everyone access to digital equivalents of
darkrooms, mixing studios, painting studios, and so on; these were formerly highly-specialized
technologies requiring laborious effort.
16
Figure 11: Les Paul, inventor the solid-body electric guitar in 1943; Robert Moog, pioneer of
the electronic synthesizer, respectively. Their technologies transformed popular music in ways
they could not have foreseen.
3.3
The general story of jobs and technology
Concerns of how technology displaces jobs have been around since at least the 19th century,
when Luddite protesters destroyed mechanical weaving machines, and, in folk songs, John
Henry competed against a steam drilling machine. These fears are real and understandable. Yet,
despite many technological disruptions, we do not live in a world of massive unemployment
caused by technology. This is because, as old roles are erased, many more arise in their stead.
But these fears keep recurring, because, at any given time, it is easy to imagine losing specific
jobs but it requires superhuman imagination to forecast what new opportunities will be created
by transformative new technologies. Nowadays, most of us do jobs that would be hard to even
explain in detail to a 19th-century worker.
The real workforce concerns should not be about the technology itself, but whether the economic system shares the benefits of new productivity fairly across society versus concentrating
wealth only among the very richest [65] and whether machine learning systems are misused to
magnify existing forms of inequality [49]. When displacements due to new technology occur,
their effects can be eased by social safety nets and better educational foundations (for employment flexibility and retraining). Conversely, a society which fails to distribute wealth and
economic gains fairly has much bigger problems than just the impact of AI.
Fears of new technology seem to be human nature. I suspect many people view the “normal” state of things as being how they were when they came of age, and they view any significant change as scary. Yet, nearly all of our familiar modern technologies were viewed as
threatening by some previous generation.
The fear of human-created life has been with us for a long time. Notably, 18th-century
scientists discovered electricity. As they searched to understand it, they discovered the life-like
effect of galvanism, that the muscles of dead frogs could be stimulated by electrical currents.
Had the secret of life been discovered? This inspired Mary Shelley’s novel Frankenstein;
or The Modern Prometheus, in which a student uses modern science to create new life [59].
Today, the story is vivid and evocative, but, intellectually, we recognize it as preposterous. The
fear of AI is essentially the same irrational fear; SkyNet is Frankenstein’s monster, but with
17
neural networks as the Promethean spark instead of galvanism.5 At present, the Terminator’s
autonomous AI is only slightly more plausible than its ability to travel backwards in time.
3.4
Future of AI as tools for artists
Some general trends around the evolution of technology and art seem quite robust. As discussed above, current AI algorithms are not autonomous creators, and will not be in the foreseeable future. They are still just tools, ready for artists to explore and exploit. New developments in these tools will be enthusiastically adopted by some artists, leading to exciting new
forms and styles that we cannot currently foresee. It is possible that some tasks performed
by human artists will gradually fade out, but these will generally be mechanical tasks that do
not require much creativity because they fill societal functions other than artistic expression.
Some traditional arts may fade simply due to seeming old-fashioned. This is the nature of art:
nothing is fresh forever, which is not to be blamed on technology. But, conversely, new technologies will enable new styles, aesthetics, and job descriptions. Novices will have access to
new simplified tools for expression. Artistic technology is a “imagination amplifier” [8] and
better technology will allow artists to see even further than before.
Aside from general trends, it is hard to make specific predictions about the art of the future.
Les Paul, who invented the solid-body guitar in the 1940s, himself primarily performed light
pop, country, and showtunes, and could hardly have predicted how the electric guitar would
be used by, say, Led Zeppelin, just as it’s hard to imagine Daguerre predicting Instagram.
More generally, making predictions about how AI technologies might transform society is
very hard because we have so little understanding of what these technologies might actually be
[4]. Even the science fiction writers of the 1950s and 1960s completely failed to imagine the
transformative power of the Internet and mobile computing [40]; for them, the computers of
the future would still be room-scale monstrosities that one had to sit in front of to operate. But
they did predict moon colonies and replicants by 2018.
In short, we cannot predict what new inventions and ideas artists will come up with in the
future, but we can predict that they will be amazing, and they will be amazing because they
make use of technology in new, unpredictable ways.
4
What is an Artist?
So far, I have described how computer technologies are currently accepted as tools for artists,
not as artists themselves. Why is this? After all, computers can do other human tasks like
speak, search, print, navigate, and, to some extent, drive cars. There are several obvious reasons
why computers don’t make art, including tradition, the incentives involved, and the relatively
predictable nature of existing automation. But, still one could imagine an alternate history in
5
In fact, Frankenstein is presented as a cautionary tale about the quest for knowledge in general. Victor
Frankenstein tells his story as a warning when he learns that Captain Walton is himself driven by an obsessive
quest for knowledge that is entirely unrelated to Frankenstein’s.
18
which some machines or computer programs had already been called artists. I believe there is
a more fundamental reason, which explains not just why this has not happened, but why it is
unlikely to happen anytime soon.
In this section, I theorize about what are the prerequisites for being an artist. I will then
apply this discussion to AI in the next section.
4.1
Art is social
“What an artist is trying to do for people is bring them closer to something, because of course
art is about sharing: you wouldn’t be an artist if you didn’t want to share an experience, a
thought.” – David Hockney [36]
Why do we create and consume art? I argue that art is primarily a social behavior: art is
about communication and displays between people. For example, people often speak of art as
being about personal expression, which is an act of communication.
I am directly inspired by the theory, going back to Charles Darwin, that art-making is
an adaptive product of our biological evolution. [17] sets out a persuasive argument for this
theory, which I briefly summarize, though I cannot do it full justice. Creating art served several
functions for our Pleistocene ancestors. Art-making served as a fitness signal for mating and
sexual selection. Art can also be used as displays of wealth and status. Storytelling, music, and
dance strengthen social bonds within a group. Storytelling additionally plays a very important
role of communicating information that would otherwise be hard to share.
I observe that each of these functions of art is social: art arose as forms of communication,
displays, and sharing between people. Although art takes many different forms in different
cultures today, each of these forms serves one or more of the same basic social functions that
it did in the Pleistocene.
I generalize this theory beyond humans to hypothesize: art is an interaction between social
agents. A “social agent” is anything that has a status akin to personhood; someone worthy
of empathy and ethical consideration. Many of our other behaviors are interactions between
social agents, such as gifts, conversation, and social relationships like friendship, competition,
and romance.
In contrast, while we can get emotionally attached to our computers and other possessions,
we feel no real empathy for their needs, and no ethical duty toward them. Possessions can participate only in shallow versions of these interactions. For example, we frequently talk about
our possessions with statements like “The brakes on my bicycle were complaining so I gave
it a new pair as a gift and now it’s much happier. I love my bike.” This statement indicates
emotional attachment but not true empathy with the bike’s feelings, despite the anthropomorphizing language. We don’t live in a social hierarchy with our possessions: we do not compete
with them for status, or try to impress them. We care about what other people have to say because we care about other people; we care about what computers have to say only insofar as it
is useful to us. We generally treat conversational agents (like Siri and Alexa) as user interfaces
to software, not like people.
Art-making does also have non-social benefits to the artist. For example, art-making can
19
help practice skills like dexterity and problem solving. Creating art is often pleasurable or
meditative in itself. But these benefits are secondary to their social benefits: they are not the
reasons that evolution has produced art as a human activity. Similarly, one may also talk or
sing to oneself while alone, but talking and singing are still fundamentally social activities.
Note that the evolutionary argument here is optional; one can discuss whether art is fundamentally social without it. But I believe that the evolutionary view of art gives some additional
understanding.
4.2
Non-Human Authors
As we have seen, despite many technological advances, current algorithms are not accepted
as artists. There are a few other existing examples where objects are created by authors or
processes that are not human-driven. These give some support to this theory.
Natural processes. Natural processes, including landscapes like the Grand Canyon or the
HuangShan Mountains, are not considered art, even though they may be extraordinarily beautiful and change one’s perspective immensely. Beautiful structures made instinctively by animals, such as honeycombs and coral, are not considered art. This indicates that simply creating
complex and beautiful outputs is not itself sufficient for art, since there is no creative social
communication in these cases.
Animals. Some higher mammals, including chimpanzees, elephants, and dolphins, have been
trained to paint [14, 17]. Many writers are skeptical of animal-made art. Typically, the animal’s
owner or handler steers the process, letting the animal throw paint on the canvas, then stopping
the painting when they believe it is done, then selecting which works to show; the animals
seem to show no interest in the artwork afterward. Animal artwork has not had any significant
cultural impact or popularity; it seems to have been largely the product of media stunts. (People for the Ethical Treatment of Animals have recently tried to claim copyright in favor of a
monkey, but failed [45], as US copyright law only allows humans to claim copyright.)
The most interesting aspect of this discussion is not whether animals can create art, but
how we decide. Discussions of whether animal-made artifacts are art are not based on a priori
rules whether animals can be artists. Instead, they are attempts to study the evidence of the
animal’s behavior around the artwork, and, from that, to infer whether the artwork is some
form of inner expression [14], or an artifact that the animal has a special appreciation for [17].
In other words, we are open to the idea of animals creating art, because they can have social
relationships with us. It’s just that we haven’t found any other creature that satisfies our criteria
for creating art, whatever they are.
4.3
Judging the work instead
It is tempting to judge whether a computer is an artist based solely on the merits of the work that
it produces. In this hypothesis, whether or not a computer can be an artist is a judgment of the
20
Figure 12: Not art. (a) Beautiful landscapes are not considered art. (b) Pierre Brassau, the
monkey painter. He was part of a 1963 hoax that satirized modern art.
quality of the work that it produces, independent of the properties of the computer itself. If an
algorithm outputs a continual stream of diverse, stimulating, beautiful, and/or skillful outputs,
without many duds, we might be quite tempted to call this algorithm an artist. The better
computer-generated art becomes, the more we will hear questions about whether computers
are artists.
Skill is clearly not the real requirement for someone or something to be an artist. Any
human can make art, including unskilled amateurs and children. Conversely, computers can
already be programmed to create infinite sequences of dazzling realistic or abstract imagery,
exhibiting technical proficiency way beyond the typical human’s.
I suspect that, when we look at a computer’s output and ask “is this work good enough
to call the computer is an artist?”, we are not actually judging the quality of the work per se.
Instead, we are really looking for evidence that the system itself is intelligent, conscious, and
feeling: traits that we associate with social agents. No matter how skillful and surprising a
computer’s output is, we will not accept it as an artist until we infer some sort of social being
inside.
4.4
An Intent Machine
Another hypothesis is as follows: in the modern art world, the role of the artist is to supply the
“intent” and the “idea” for the work; it is not necessary for the artist to execute on the work,
other than coordinating in its production. For humans, this is clearly true in numerous examples, such as those in Section 2.4, as well as appropriation works like Duchamp’s Readymades
and Richard Prince’s questionable Instagram reproductions; artists can also employ helpers or
crowdworkers, such as in Scott Draves’ Electric Sheep and Aaron Koblin’s Sheep Market [38].
Consequently, for a computer to be an artist it simply needs to supply an intent.
It is easy to imagine designing a system that creates intent and even coordinates the labor of
producing a work. For example, one could write a simple procedural algorithm to generate very
basic intents (e.g., “portray ominous landscape”), or sample intents from a Recursive Neural
21
Network trained on artists’ statements scraped from the web. Or the method could randomly
select some news item, photograph, or historical event, and randomly sample some attitude
toward that thing. Starting from this intent, crowdworkers could be used to refine the idea and
convert it into a new image, similar to systems like Soylent [3]. One could also automate steps
of the process, e.g., using GANs [28] to generate entirely new starting images from scratch.
Crowdworkers could also be used to rate and evaluate the outputs of the system, selecting just
the best results and discarding the rest. Final steps of the process could also automatically hire
professional designers, e.g., from sites like Upwork or 99Designs. This system could then run
continuously, generating new images over time (with payment being automatically made to the
crowdworkers involved). Workers could group images with common themes and intents, and
create separate collections around these themes. Artist’s statements could generated around
these themes. The system’s preferences could grow and adapt over time as more data are
gathered or external data streams (e.g., photography blogs) change.
Suppose someone were to build this system, calling it, say, The Intent Machine, and exhibits its work in an art show or gallery. Suppose, moreover, they convinced the curator or
gallery owner to credit The Intent Machine with authorship of the works it’d created, but fully
disclosed the procedure by which it worked. Would people credit it as the artist who had
authored its works, or would they say that the system-builder is the real artist here?
I believe that, in general, the consensus would be that the system-builder is the real artist
here, and that this is really an artwork about probing the nature of computer-generated art, or
the nature of the commercial artworld. Note that the procedure used to define the system is
not fundamentally different than any other procedural computer-generated art algorithm (as in
Sections 2.4 and 2.5). Even if the work itself ended up being quite good, viewers would ask
why it is good, and it is doubtful that the computer’s own contribution would be judged as
significant beyond those of the humans involved.
Are artists just “intent machines”? If art is a social act, then the answer is no.
4.5
Creativity and Growth
Finally, two related criteria that have intuitive appeal are the notions of creativity and growth.
They are important for human artists, so perhaps they should be for AI artists as well.
Several authors have proposed energy terms or criteria for creativity [47, 21]. These are often attempts to express the idea that the system’s output should somehow “surpass” the human
programmer and/or the training data. For example, [9] proposes that we judge the creativity of
a system, in part, by whether the system’s output surprises the system’s author. I believe that
this is too weak a criterion, since many mechanical or algorithmic phenomena may be surprising to their own discoverer or author at first. For example, the basic algorithm that produces
Mandelbrot set images can be specified in a single sentence,6 yet produces dazzling animations of infinite complexity.7 The Mandelbrot set is very surprising and produces beautiful,
Color the image location (x, y) proportional to the number of steps the iteration zn+1 = zn2 + x + yi requires
to reach |z| ≥ τ , starting from z1 = 0, where τ is a large constant.
7
For example: https://www.youtube.com/watch?v=PD2XgQOyCCk
6
22
unprecedented images, but we do not call its iteration equation creative, or an artist.
All current procedural art systems, such as the Mandelbrot set, have a recognizable style,
and, after awhile, lose their novelty. I believe that the same will be true for systems specifically
designed with “creativity” objectives [21, 47]. Unlike human artists, these systems do not grow
or evolve over time.
Perhaps an AI artist would need to exhibit some form of growth. For example, Harold
Cohen, the fine-art painter who began, in 1968, to write software to generate art, described the
evolution of his views: “Ten years after [1968], I would have said ‘look, the program is doing
this on its own.’ ... Another ten years on and I would have said ‘the fact that it is the program
who is doing this on its own is the central issue here,’ denoting my belief in the program’s
potential and growing autonomy over the whole business of art-making. ... It was producing
complex images of a high quality and I could have had it go on forever without rewriting a
single line of code. How much more autonomous than that can one get? ... [But] it’s virtually
impossible to imagine a human being in a similar position. The human artist is modified in the
act of making art. For the program to have been similarly self-modifying would have required
not merely that it be capable of assessing its own output but that it had its own modifiable
worldview ...” [7]
In any existing system, it is easy to think of trivial ways for the system to evolve and
change over time, e.g., subtly change the color palette or the training data over the years.
Superficial growth is easy; meaningful growth is hard to even define for a computer AI. If
someone could design a system that produces a sequence of art that is meaningful to people
and also significantly evolves over time, that would be truly remarkable. It seems hard to
imagine achieving this without enormous technological advances — it may not be possible
without true AI or social AI in some form.
4.6
Definitions of Art
For guidance about the nature of art, we could have also looked to existing definitions of art.
However, there is no Royal Society that prescribes what is and what is not valid art. Instead,
art is a phenomenon that results from the interplay of cultural institutions and the general
population that we can analyze, and it changes over time. Philosophers have attempted to
devise concise definitions of art that include all existing types (music, dance, painting, etc.)
and styles of art. The Institutional Definition, originated in the 1960s in response to conceptual
art, states, roughly, that art is anything in a style that is broadly accepted as art [12, 15, 42].
A more fine-grained approach is to identify attributes common to many different types of art
[25, 26, 16, 17]. Each of these definitions of art is an attempt to fit the data, and draw a line
between those things that we call art (like theatre) and those that we do not (like spectator
sports). Understandably, these definitions all assume that the artist is always human, without
exploring much whether non-humans can create art, and thus don’t provide much guidance for
this discussion.
23
5
Will an AI ever be an artist?
With this background, I now turn to speculating about the future. As we have seen, authorship
of all current algorithmic art is assigned to the human author behind the algorithms. Will we
ever say that an AI itself created art? Will we ever recognize a piece of software as the author
of a work of art?
Human-Level AI. If we ever develop AI with human-level artificial intelligence and consciousness, by definition, it would be able to create art, since it would have the same capacity
for consciousness, emotions, and social relationships. But, as discussed in Section 2.6, this
scenario is science fiction and we have no idea if this is possible or how it would be achieved.
Making meaningful predictions about a world with “true AI” is impossible [4], because we
have so little idea of how specifically this AI would actually operate. Moreover, this AI would
transform society so much as to make it unrecognizable to us. We may as well speculate about
what kind of artwork is made by aliens from outer-space — if do we ever meet them, we will
have more pressing questions than what kind of music they like.
Hence, the interesting question is whether there could be computer-authored artwork without human-level artificial intelligence or consciousness.
Social AI. If, as I have argued, creating art is a fundamentally social act of expression and
communication, then it follows that AI can be granted authorship when we view the AI as a
social agent, and it is performing some communication or sharing through the art.
What does it mean for us to view an AI as a social agent? We have to view the AI as
deserving of empathy and ethical consideration in some way. However, the AI does not need
human-level intelligence; just as we have social relationships with our pets. But we do expect
that the AI has something to say socially, that suggest an inner consciousness and feeling.
Short of true intelligence (the science fiction scenario), I think that the only way this can
happen is through “shallow AI” agents. People are sometimes “fooled” by shallow AI. The
classic example is Eliza, a simple text-based psychiatrist program developed in 1964, based
on simple pattern-matching and repetition of what the user types [70]. It was meant as a
demonstration of the superficiality of the AIs of the time, but, unexpectedly, many people
attributed human-like emotions to the machine. Since then, there are many anecdotes of people
being fooled by “chatbots” in online settings [18, 22], including the recent plague of Twitter
bots [37]. But, once the veil is lifted, it is clear that these chatbots don’t exhibit real intelligence.
Some software and robots have been designed to have relationships with their owners,
including talking baby dolls, Tamagotchi, and Paro therapeutic baby seals. A related effect is
that people behave toward their computers as if they were social agents in certain ways [53],
even when they don’t believe that they are intelligent. For example, dialogue systems like Siri
and Alexa all use female voices by default, based on many findings that male and female users
both respond better to female voices [29, 53].
Perhaps, for many users, the system doesn’t need to be truly intelligent, it just has to be
perceived as a social agent, like a Siri or Alexa that you can ask to make you an artwork
24
[73]. The day may come in which these agents are so integrated into our daily lives that we
forget that they are carefully-designed software. One can easily imagine the development of
AI that simulates emotion and affection toward the user; it is easy to imagine, for example, a
toy doll that paints pictures for its owner along as one of many behaviors designed to display
companionship and affection.
Non-social AI. It seems possible that non-social algorithms could be successfully promoted
as artists; there have been a few tentative forays in this direction, e.g., [10, 21]. For the reasons
given above, I am skeptical that such methods will be accepted as true artists without some
plausible belief about their underlying social and/or conscious attributes.
Perhaps a curator at a well-known museum would download or otherwise acquire various
artifacts from software “artists,” and list the software systems as the authors. There would
be controversy, and discussion in newspapers and journals. Perhaps other curators and galleries would follow suit. Perhaps people would find enough value in these computer-generated
artworks, while also being convinced that no human could be rightly given credit for their
works. This sort of process has happened for things like abstract expressionism, and not for
chimpanzee art. Could it happen for computer art?
Suitcase words. The term “artist” could come to be used as having multiple meanings, just
as words like “intelligence” and ”learning” have come to mean something different for humans
than they do for algorithms [4]. A software program that, say, automatically stylizes your
photos, could be called an “artist” in the same way that software applications like “calendar”
and “mail” programs have replaced their physical-world namesakes. Unfortunately, the use
of the same word to mean different things in different contexts causes endless confusion, as
discussed in Section 2.6.
Dangers. A continual danger of new AI technology is that human users misunderstand the
nature of the AI [49]. When we call a shallow AI an “artist,” we risk seriously misleading or
lying to people. I believe that, if you convince people that an AI is an artist, then they will also
falsely attribute emotions, feelings, and ethical weight to that AI. If this is true, I would argue
that calling such AIs “artists” is unethical. It leads to all sorts of dangers, including overselling
the competence and abilities of the AI, to misleading people about the nature of art.
It seems likely that some companies will not have any scruples about this. For example,
Hanson Robotics has promoted (in many contexts) a social robot as a truly intelligent being,
even though it is clearly nothing more than a “chatbot with a face” [27] or, in Yann LeCun’s
words, a “Potemkin AI” [69].
A related concern is that we deprive the AI’s designer(s) of authorship credit. At present,
we credit the author of a piece of automatic software with the output of that software. This
usually acknowledges the skill and effort required to engineer and iterate with software so that
it produces good outputs. Artistic credit is important for understanding the real sources of how
something was made.
25
Outside of science fiction, I can see no positive benefit to calling a computer an artist, but I
do see dangers.
Conclusion
I do not believe that any software system in our current understanding could be called an
“artist.” Art is a social activity. I mean this as a warning against misleading oneself and others
about the nature of art. Of course, the ambitious reader could take this as a challenge: I have
laid out some of the serious objections that you must overcome if you wish to create a software
“artist.” I don’t think it can be done anytime soon, but I also know that proving critics wrong
is one of the ways that art and science advance.
One of my main goals in this essay has been to highlight the degree to which technology
contributes to art, rather than being antagonistic. We are lucky to be alive at a time when artists
can explore ever-more-powerful tools. Every time I see an artist create something wonderful
with new technology, I get a little thrill: it feels like a new art form evolving. Danny Rosin’s
Wooden Mirror, Jason Salavon’s The Top Grossing Film of All Time, 1x1, Bob Sabiston’s Snack
and Drink, Michel Gondry’s Like A Rolling Stone, Kutiman’s ThruYOU, Amon Tobin’s Permutation, Ian Bogost’s Cow Clicker, Christian Marclay’s video installations, Íñigo Quilez’s
procedural renderings, and Wesley Allsbrook’s and Goro Fujita’s Virtual Reality paintings are
a few examples of artworks that have affected me this way over the years. Today, through
GitHub and Twitter, there is an extremely fast interplay between machine learning researchers
and artists; it seems like, every day, we see new tinkerers and artists Tweeting their latest creative experiments with RNNs and GANs (e.g., @JanelleCShane, @helena, @christophrhesse,
@quasimondo).
Art maintains its vitality through continual innovation, and technology is one of the main
engines of that innovation. Occasionally, the avant garde has tremendous cultural impact:
electronic music and sampling was once the domain of experimental electronic and musique
concrète pioneers, like Wendy Carlos and Delia Darbyshire. Likewise, at one time, computeranimated films could only be seen at obscure short-film festivals. Today, we are seeing many
intriguing and beguiling experiments with AI techniques, and, as artists’ tools, they will surely
transform the way we think about art in thrilling and unpredictable ways.
Acknowledgements. Thanks to Shira Katz, Craig Kaplan, and Dani Oore for valuable comments on the manuscript. Thanks to everyone who shared discussion and/or encouragement
online, including Aseem Agarwala, Mark Chen, Lyndie Chiou, Michael Cohen, James Landay, Nevena Lazic, Jason Salavon, Adrien Treuille, and many others.
References
[1] AG ÜERA Y A RCAS , B. Art in the Age of Machine Intelligence. Arts 6, 4 (2017).
26
[2] A LEXANDER , R. M. Optima For Animals. Princeton University Press, 1996.
[3] B ERNSTEIN , M. S., L ITTLE , G., M ILLER , R. C., H ARTMANNN , B., ACKERMANN ,
M. S., K ARGER , D. R., C ROWELL , D., AND PANOVICH , K. Soylent: A Word Processor
with a Crowd Inside. In Proc. UIST (2010).
[4] B ROOKS , R. The Seven Deadly Sins of AI Predictions. MIT Technology Review (2017).
[5] C ATMULL , E., AND WALLACE , A. Creativity, Inc: Overcoming the Unseen Forces that
Stand in the Way of True Inspiration. Random House, 2014.
[6] C OHEN , H. The further exploits of AARON, painter. Stanford Humanities Review 4, 2
(1995).
[7] C OHEN , H. ACM SIGGRAPH Awards - Harold Cohen, Distinguished Artist Award for
Lifetime Achievement, 2014. https://youtu.be/_Xbt8lzWxIQ?t=13m20s.
[8] C OHEN , M. F. Imagination amplification. IEEE Comp. Graph. & App. (2000).
[9] C OLTON , S. Creativity Versus the Perception of Creativity in Computational Systems.
In Proc. AAAI spring symposium: creative intelligent systems (2008).
[10] C OLTON , S. The Painting Fool: Stories from Building an Automated Painter. In
Proc. Computers and Creativity (2012).
[11] C OTTER , H. Now in Residence: Walls of Luscious Austerity. The New York Times
(2008).
[12] DANTO , A. The Artworld. The Journal of Philosophy 61, 19 (1964).
[13]
DE LA C ROIX , H., TANSEY, R. G., AND K IRKPATRICK , D. Gardner’s Art Through The
Ages, 9 ed. Harcourt Brace, 1991.
[14] D ESMOND , J. Zoos make money selling paintings made by animals. Are they art? The
Washington Post (2016).
[15] D ICKIE , G. Defining Art. American Philosophical Quarterly 6, 3 (1969).
[16] D UTTON , D. But They Don’t Have Our Concept of Art. In Theories of Art Today,
N. Carroll, Ed. University of Wisconsin Press, 2000.
[17] D UTTON , D. The Art Instinct: Beauty, Pleasure, and Human Evolution. Bloomsbury
Press, 2009.
[18] DVORSKY, G. Computer Science Students Fooled By Artificially Intelligent TA. Gizmodo (2016).
27
[19] DVORSKY, G. This Artificially Intelligent Robot Composes and Performs Its Own Music.
Gizmodo (2017).
[20] E FROS , A. A., AND L EUNG , T. Texture synthesis by non-parametric sampling. In Proc.
International Conference on Computer Vision (1999).
[21] E LGAMMAL , A., L IU , B., E LHOSEINY, M., AND M AZZONE , M. CAN: Creative Adversarial Networks Generating “Art” by Learning About Styles and Deviating from Style
Norms. In Proc. Int. Conf. on Computational Creativity (2017).
[22] E PSTEIN , R. From Russia, with Love. Scientific American Mind (2007).
[23] F I ŠER , J., JAMRI ŠKA , O., L UK Á Č , M., S HECHTMAN , E., A SENTE , P., L U , J., AND
S ÝKORA , D. StyLit: Illumination-guided Example-based Stylization of 3D Renderings.
ACM Trans. on Graphics (SIGGRAPH) 35, 4 (July 2016), 92:1–92:11.
[24] G ATYS , L. A., E CKER , A. S., AND B ETHGE , M. Image Style Transfer Using Convolutional Neural Networks. In Proc. Computer Vision and Pattern Recognition (2016).
[25] G AUT, B. Art as a Cluster Concept. In Theories of Art Today, N. Carroll, Ed. University
of Wisconsin Press, 2000.
[26] G AUT, B. The Cluster Account of Art Defended. The British Journal of Aesthetics
(2005).
[27] G ERSHGORN , D. Inside the mechanical brain of the worlds first robot citizen. Quartz
(2017).
[28] G OODFELLOW, I. J., P OUGET-A BADIE , J., M IRZA , M., X U , B., WARDE -FARLEY, D.,
O ZAIR , S., C OURVILLE , A., AND B ENGIO , Y. Generative Adversarial Nets. In Proc.
Neural Information Processing Systems (2014).
[29] G RIGGS , B. Why computer voices are mostly female. CNN (2011).
[30] H AEBERLI , P. Paint By Numbers: Abstract Image Representations. In Proc. SIGGRAPH
(1990).
[31] H ERTZMANN , A. Painterly Rendering with Curved Brush Strokes of Multiple Sizes. In
Proc. SIGGRAPH (1998).
[32] H ERTZMANN , A. Algorithms for Rendering in Artistic Styles. PhD thesis, New York
University, 2001.
[33] H ERTZMANN , A. A Survey of Stroke-Based Rendering. IEEE Computer Graphics &
Applications 23, 4 (2003).
[34] H ERTZMANN , A. Non-Photorealistic Rendering and the Science of Art. In Proc. NPAR
(2010).
28
[35] H ERTZMANN , A., JACOBS , C. E., O LIVER , N., C URLESS , B.,
Image Analogies. In Proc. SIGGRAPH (2001).
AND
S ALESIN , D. H.
[36] H OCKNEY, D. That’s the way I see It. Chronicle, 1996.
[37] K ESSLER , S. How Twitter Bots Fool You Into Thinking They Are Real People. Fast
Company (2014).
[38] KOBLIN , A. M. The Sheep Market. In Proc. Creativity and Cognition (2009), pp. 451–
452.
[39] KOERDING , K. P., AND W OLPERT, D. M. Bayesian decision theory in sensorimotor
control. Trends in Cognitive Sciences 10, 7 (2006), 319 – 326. Special issue: Probabilistic
models of cognition.
[40] K RAUSS , L. Why Science-Fiction Writers Couldn’t Imagine the Internet. Slate (2017).
[41] L ASSETER , J. Principles of traditional animation applied to 3d computer animation. In
Proc. SIGGRAPH (1987).
[42] L EVINSON , J. Defining art historically. The British Journal of Aesthetics 19, 3 (1979).
[43] L ITWINOWICZ , P. Processing Images and Video for an Impressionist Effect. In Proc.
SIGGRAPH (1997).
[44] M ETZ , C. How A.I. Is Creating Building Blocks to Reshape Music and Art. The New
York Times (2017).
[45] M ILLWARD , D. Monkey selfie case: British photographer settles with animal charity
over royalties dispute. The Telegraph (2017).
[46] M ORDVINTSEV, A., O LAH , C., AND T YKA , M. Inceptionism: Going Deeper into
Neural Networks. Google Research Blog (2015).
[47] N GUYEN , A., YOSINSKI , J., AND C LUNE , J. Understanding Innovation Engines: Automated Creativity and Improved Stochastic Optimization via Deep Learning. Evolutionary
Computation 24, 3 (2016).
[48] N ICHOLAS , G. These stunning A.I. Tools Are About to Change the Art World. Slate
(2017).
[49] O’N EIL , C. Weapons of Math Destruction: How Big Data Increases Inequality and
Threatens Democracy. Crown, 2016.
[50] PAIK , K. To Infinity and Beyond! The story of Pixar Animation Studios. Chronicle Books,
2007.
[51] P EREZ , S. Microsoft’s new drawing bot is an AI artist. TechCruch (2018).
29
[52] P RICE , D. A. The Pixar Touch. Vintage, 2008.
[53] R EEVES , B., AND NASS , C. The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places. CSLI, 1996.
[54] ROSENBLUM , N. A World History of Photography. Abbeville Press, 1984.
[55] ROSIN , P., AND C OLLOMOSSE , J. Image and Video-Based Artistic Stylisation. Springer,
2013.
[56] RUBIN , M. Droidmaker: George Lucas and the Digital Revolution. Triad, 2005.
[57] S CHARF, A. Art and Photography. Penguin, 1968.
[58] S HEIN , E. Computing the Arts. Commun. ACM 60, 4 (2017).
[59] S HELLEY, M. W. Frankenstein; or, the Modern Prometheus. Lackington, Hughes, Harding, Mavor & Jones, 1818.
[60] S IMONITE , T. When it comes to gorillas, Google Photos remains blind. Wired (2018).
[61] S NOW, C. P. The Two Cultures. London: Cambridge University Press, 1959.
[62] S NOWDEN , D. Robert Moog: ’I wouldn’t call this music’ a classic interview to mark a
Google doodle. The Guardian (2012).
[63] S TERLING , B. An Essay on the New Aesthetic. WIRED (2012).
[64] S TERNBERGER , P. S. Between Amateur & Aesthete: The Legimitization of Photography
as Art in America, 1880-1900. University of New Mexico Press, 2001.
[65] S TIGLITZ , J. E. The Price of Inequality. Norton, 2012.
[66] S ZEGEDY, C., Z AREMBA , W., S UTSKEVER , I., B RUNA , J., E RHAN , D., G OODFEL LOW, I., AND F ERGUS , R. Intriguing properties of neural networks. In Proc. Int. Conf.
on Learning Representations (2014).
[67] T ENENBAUM , J. B., K EMP, C., G RIFFITHS , T. L., AND G OODMAN , N. D. How to
Grow a Mind: Statistics, Structure, and Abstraction. Science (2011).
[68] T YKA , M. Exploring the Intersection of Art and Machine Intelligence. Google Research
Blog (2016).
[69] V INCENT, J. Facebooks head of AI really hates Sophia the robot (and with good reason).
The Verge (2018).
[70] W EIZENBAUM , J. ELIZA–a computer program for the study of natural language communication between man and machine. Comm. of the ACM (1966).
30
[71] W ILBER , M. J., FANG , C., J IN , H., H ERTZMANN , A., C OLLOMOSSE , J., AND B E LONGIE , S. BAM! The Behance Artistic Media Dataset for Recognition Beyond Photography. In Proc. International Conference on Computer Vision (2017).
[72] W ILSON , M. AI Is Inventing Languages Humans Can’t Understand. Should We Stop It?
Fast Company (2017).
[73] X U , T., Z HANG , P., H UANG , Q., Z HANG , H., G AN , Z., H UANG , X., AND H E , X. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial
Networks, 2017. arXiv:1711.10485.
31
| 2cs.AI
|
Hypothesis Testing via Affine Detectors
Anatoli Juditsky
∗
Arkadi Nemirovski
†
arXiv:1604.02576v4 [math.ST] 1 Oct 2016
Abstract
In this paper, we further develop the approach, originating in [13], to “computation-friendly”
hypothesis testing via Convex Programming. Most of the existing results on hypothesis testing
aim to quantify in a closed analytic form separation between sets of distributions allowing for
reliable decision in precisely stated observation models. In contrast to this descriptive (and highly
instructive) traditional framework, the approach we promote here can be qualified as operational –
the testing routines and their risks are yielded by an efficient computation. All we know in advance
is that, under favorable circumstances, specified in [13], the risk of such test, whether high or low,
is provably near-optimal under the circumstances. As a compensation for the lack of “explanatory
power,” this approach is applicable to a much wider family of observation schemes and hypotheses
to be tested than those where “closed form descriptive analysis” is possible.
In the present paper our primary emphasis is on computation: we make a step further in
extending the principal tool developed in [13] – testing routines based on affine detectors – to a
large variety of testing problems. The price of this development is the loss of blanket near-optimality
of the proposed procedures (though it is still preserved in the observation schemes studied in [13],
which now become particular cases of the general setting considered here).
1
Introduction
This paper can be considered as an extension of [13] where the following simple observation was the
starting point of numerous developments:
Imagine that we want to decide on two composite hypotheses about the distribution P
of a random observation ω taking values in observation space Ω, i-th hypothesis stating
that P ∈ Pi , where Pi , i = 1, 2, are given families of probability distributions on Ω. Let
φ : Ω → R be a detector, and let the risk of detector φ be defined as the smallest ǫ⋆ such
that
Z
Z
−φ(ω)
eφ(ω) P (dω) ≤ ǫ⋆ ∀P ∈ P2 .
(1)
e
P (dω) ≤ ǫ⋆ ∀P ∈ P1 &
Ω
Ω
K
Then the test T which, given
PK K i.i.d. observations ωt ∼ P ∈ P1 ∪ P2 , t = 1, ..., K,
deduces that P ∈ P1 when t=1 φ(ωt ) ≥ 0, and that P ∈ P2 otherwise, accepts the true
hypothesis with P -probability at least 1 − ǫK
⋆ .
It was shown in [13] that
∗
LJK, Université Grenoble Alpes, B.P. 53, 38041 Grenoble Cedex 9, France, [email protected]
Georgia Institute of Technology, Atlanta, Georgia 30332, USA, [email protected]
The first author was supported by the CNRS-Mastodons project GARGANTUA, and the LabEx PERSYVAL-Lab
(ANR-11-LABX-0025). Research of the second author was supported by NSF grants CMMI-1262063, CCF-1523768.
†
1
1. the detector-based tests are “near optimal” – if the above hypotheses can be decided upon by
a single-observation
test T ∗ with risk δ < 1/2, there exists a detector φ with “comparable” risk
p
ǫ⋆ = 2 δ(1 − δ).
p
Note that while the risk 2 δ(1 − δ) seems to be much larger than δ, especially for small δ, we
can “compensate” for risk deterioration passing from the single-observation test T 1 associated
with the detector φ to the test T K based on the same detector and
pusing K Kobservations. The
K
K
risk of the test T , by the above, is upper-bounded by ǫ⋆ = [2 δ(1 − δ)] and thus is not
worse than the risk δ of the “ideal” single-observation test already for a “quite moderate” value
2
⌋ 1−ln(4(1−δ))/
ln(1/δ) ⌊ of K.
2. There are “good,” in certain precise sense, parametric families of distributions, primarily,
• Gaussian N (µ, Id ) distributions on Ω = Rd ,
• Poisson distributions with parameters µ ∈ Rd+ on Ω = Zd ; the corresponding random
variables have d independent entries, j-th of them being Poisson with parameter µj ,
• Discrete distributions on {1, ..., d}, the parameter µ of a distribution being the vector of
probabilities to take value j = 1, ..., d,
for which the optimal (with the minimal risk, and thus – near-optimal by 1) detectors can be
found efficiently, provided that Pi , i = 1, 2, are convex hypotheses, meaning that they are cut
off the family of distributions in question by restricting the distribution’s parameter µ to reside
in a convex domain. 1
On a closer inspection, the “common denominator” of Gaussian, Poisson and Discrete families of
distributions is that in all these cases the minimal risk detector for a pair of convex hypotheses is
affine,2 and the results of [13] in the case of deciding on a pair of convex hypotheses stemming from
a good family of distributions sum up to the following:
A) the best – with the smallest possible risk – affine detector, same as its risk, can be efficiently
computed;
B) the smallest risk affine detector from A) is the best, in terms of risk, detector available under
the circumstances, so that the associated test is near-optimal.
Note that as far as practical applications of the above approach are concerned, one “can survive”
without B) (near-optimality of the constructed detectors), while A) is a must. In this paper, we focus
on families of distributions obeying A); this class turns out to be incomparably larger than what was
defined as “good” in [13]. In particular, it includes nonparametric families of distributions. Staying
within this much broader class, we still are able to construct in a computationally efficient way the
best affine detectors for a pair of “convex”, in certain precise sense, hypotheses, along with valid upper
bounds on the risks of the detectors. What we, in general, cannot claim anymore, is that the tests
associated with the above detectors are near-optimal. This being said, we believe that investigating
possibilities for building tests and quantifying their performance in a computationally friendly manner
is of value even when we cannot provably guarantee near-optimality of these tests.
1
In retrospect, these results can be seen as a development of the line of research initiated by the pioneering works of
H. Chernoff [9], C. Kraft [16], and L. Le Cam [17], further developed in [1, 2, 3, 4, 7, 8] among many others (see also
references in [15]).
2
affinity of a detector makes sense only when Ω can be naturally identified with a subset of some Rd . This indeed
is the case for Gaussian and Poisson distributions; to make it the case for discrete distributions on {1, ..., d}, it suffices
to encode j ≤ d by j-th basic orth in Rd , thus making Ω the set of basic orths in Rd . With this encoding, every real
valued function on {1, ..., d} becomes affine.
2
The paper is organized as follows. The families of distributions well suited for constructing affine
detectors in a computationally friendly fashion are introduced and investigated in section 2. In particular, we develop a kind of fully algorithmic “calculus” of these families. This calculus demonstrates
that the families of probability distributions covered by our approach are much more common commodity than “good observation schemes” as defined in [13]. In section 3 we explain how to build within
our framework tests for pairs (and larger tuples) of hypotheses and how to quantify performance of
these tests in a computationally efficient fashion. Aside of general results of this type, we work out
in detail the case where the family of distributions giving rise to “convex hypotheses” to be tested is
comprised of sub-Gaussian distributions (section 3.2.3). In section 4 we discuss an application to the
now-classical statistical problem – aggregation of estimators – and show how the results of [12] can be
extended to the general situation considered here. Finally, in section 5 we show how our framework
can be extended in the Gaussian case to include quadratic detectors. To streamline the presentation,
all proofs exceeding few lines are collected in the appendix.
2
Setup
Let us fix observation space Ω = Rd , and let Pj , 1 ≤ j ≤ J, be given families of Borel probability
S
Pj , to
distributions on Ω. Put broadly, our goal is, given a random observation ω ∼ P , where P ∈
j≤J
decide upon the hypotheses Hj : P ∈ Pj , j = 1, ..., J. We intend to address the above goal in the case
when the families Pj are simple – they are comprised of distributions for which moment-generating
functions admit an explicit upper bound.
2.1
Regular and simple probability distributions
Let
• H be a nonempty closed convex set in Ω = Rd symmetric w.r.t. the origin,
• M be a closed convex set in some Rn ,
• Φ(h; µ) : H × M → R be a continuous function convex in h ∈ H and concave in µ ∈ M.
We refer to H, M, Φ(·, ·) satisfying the above restrictions as to regular data. Regular data H, M, Φ(·, ·)
define a family
R = R[H, M, Φ]
of Borel probability distributions P on Ω such that
R
∀h ∈ H ∃µ ∈ M : ln Ω exp{hT ω}P (dω) ≤ Φ(h; µ).
(2)
We say that distributions satisfying (2) are regular, and, given regular data H, M, Φ(·, ·), we refer to
R[H, M, Φ] as to regular family of distributions associated with the data H, M, Φ. The same regular
data H, M, Φ(·, ·) define a smaller family
S = S[H, M, Φ]
of Borel probability distributions P on Ω such that
R
∃µ ∈ M : ∀h ∈ H : ln Ω exp{hT ω}P (dω) ≤ Φ(h; µ).
(3)
We say that distributions satisfying (3) are simple. Given regular data H, M, Φ(·, ·), we refer to
S[H, M, Φ] as to simple family of distributions associated with the data H, M, Φ.
3
Recall that the starting point of our study is a “plausibly good” detector-based test which, given
two families P1 and P2 of distribution with common observation space, and independent observations
ω1 , ..., ωt drawn from a distribution P ∈ P1 ∪ P2 , decides whether P ∈ P1 or P ∈ P2 . Our interest in
regular/simple families of distributions stems from the fact that when the families P1 and P2 are of
this type, building such test reduces to solving a convex-concave game and thus can be carried on in
a computationally efficient manner. We postpone the related construction and analysis to section 3,
and continue with presenting some basic examples of simple and regular families of distributions and
a simple “calculus” of these families.
2.2
2.2.1
Basic examples of simple families of probability distributions
Sub-Gaussian distributions
Let H = Ω = Rd , M be a closed convex subset of the set Gd = {µ = (θ, Θ) : θ ∈ Rd , Θ ∈ Sd+ }, where
Sd+ is cone of positive semidefinite matrices in the space Sd of symmetric d × d matrices, and let
Φ(h; θ, Θ) = θ T h + 21 hT Θh.
In this case, S[H, M, Φ] contains all sub-Gaussian distributions P on Rd with sub-Gaussianity parameters from M, that is, all Borel probability distributions P on Ω admitting upper bound
Eω∼P {exp{hT ω}} ≤ exp{θ T h + 21 hT Θh} ∀h ∈ Rd
(4)
on the moment-generating function, with parameters (θ, Θ) of the bound belonging to M. In particular, S[H, M, Φ] contains all Gaussian distributions N (θ, Θ) with (θ, Θ) ∈ M.
2.2.2
Poisson distributions
Let H = Ω = Rd , let M be a closed convex subset of d-dimensional nonnegative orthant Rd+ , and let
Φ(h = [h1 ; ...; hd ]; µ = [µ1 ; ...; µd ]) =
d
X
i=1
µi [exp{hi } − 1] : H × M → R.
The family S[H, M, Φ] contains all product-type Poisson distributions Poisson[µ] with vectors µ of
parameters belonging to M; here Poisson[µ] is the distribution of random d-dimensional vector with
independent of each other entries, i-th entry being Poisson random variable with parameter µi .
2.2.3
Discrete distributions
Consider a discrete random variable taking values in d-element set {1, 2, ..., d}, and let us think of
such a variable as of random variable taking values ei , i = 1, ..., d, where ei are standard basic orths
in Rd ; probability distribution of such a variable can be identified with a point µ = [µ1 ; ...; µd ] from
the d-dimensional probabilistic simplex
∆d = {ν ∈ Rd+ :
d
X
νi = 1},
i=1
where µi is the probability for the variable to take value ei . With these identifications, setting H = Rd ,
specifying M as a closed convex subset of ∆d and setting
!
d
X
µi exp{hi } ,
Φ(h = [h1 ; ...; hd ]; µ = [µ1 ; ...; µd ]) = ln
i=1
4
the family S[H, M, Φ] contains distributions of all discrete random variables taking values in {1, ..., d}
with probabilities µ1 , ..., µd comprising a vector from M.
2.2.4
Distributions with bounded support
Consider the family P[X] of Borel probability distributions supported on a closed and bounded convex
set X ⊂ Ω = Rd , and let
φX (h) = max hT x
x∈X
be the support function of X. We have the following result (to be refined in section 2.3.5):
Proposition 2.1 For every P ∈ P[X] it holds
Z
1
T
d
(5)
exp{h ω}P (dω) ≤ hT e[P ] + [φX (h) + φX (−h)]2 ,
∀h ∈ R : ln
8
Rd
R
where e[P ] = Rd ωP (dω) is the expectation of P , and the right hand side function in (5) is convex.
As a result, setting
H = Rd , M = X, Φ(h; µ) = hT µ +
1
[φX (h) + φX (−h)]2 ,
8
we get regular data such that P[X] ⊂ S[H, M, Φ].
For proof, see Appendix A.1
2.3
Calculus of regular and simple families of probability distributions
Regular and simple families of probability distributions admit “fully algorithmic” calculus, with the
main calculus rules as follows.
2.3.1
Direct summation
For 1 ≤ ℓ ≤ L, let regular data Hℓ ⊂ Ωℓ = Rdℓ , Mℓ ⊂ Rnℓ , Φℓ (hℓ ; µℓ ) : Hℓ × Mℓ → R be given. Let
us set
Ω = Ω1 × ... × ΩL = Rd , d = d1 + ... + dL ,
H = H1 × ... × HL = {h = [h1 ; ...; hL ] : hℓ ∈ Hℓ , ℓ ≤ L},
M = M1 × ... × ML = {µ = [µ1 ; ...; µL ] : µℓ ∈ Mℓ , ℓ ≤ L} ⊂ Rn , n = n1 + ... + nL ,
P
ℓ ℓ
Φ(h = [h1 ; ...; hL ]; µ = [µ1 ; ...; µL ]) = L
ℓ=1 Φℓ (h ; µ ) : H × M → R.
Then H is a symmetric w.r.t. the origin closed convex set in Ω = Rd , M is a nonempty closed convex
set in Rn , Φ : H × M → R is a continuous convex-concave function, and clearly
• the family R[H, M, Φ] contains all product-type distributions P = P1 × ... × PL on Ω = Ω1 ×
... × ΩL with Pℓ ∈ R[Hℓ , Mℓ , Φℓ ], 1 ≤ ℓ ≤ L;
• the family S[H, M, Φ] contains all product-type distributions P = P1 ×...×PL on Ω = Ω1 ×...×ΩL
with Pℓ ∈ S[Hℓ , Mℓ , Φℓ ], 1 ≤ ℓ ≤ L.
5
2.3.2
IID summation
Let Ω = Rd be an observation space, (H, M, Φ) be regular data on this space, and let λ = {λℓ }K
ℓ=1
be a collection of reals. We can associate with the outlined entities a new data (Hλ , M, Φλ ) on Ω by
setting
L
X
Hλ = {h ∈ Ω : kλk∞ h ∈ H}, Φλ (h; µ) =
Φ(λℓ h; µ) : Hλ × M → R.
ℓ=1
Now, given a probability distribution P on Ω, we can associate withPit and with the above λ a new
probability distribution P λ on Ω as follows: P λ is the distribution of ℓ λℓ ωℓ , where ω1 , ω2 , ..., ωL are
drawn, independently of each other, from P . An immediate observation is that the data (Hλ , M, Φλ )
is regular, and
• whenever a probability distribution P belongs to S[H, M, Φ], the distribution P λ belongs to
S[Hλ , M, Φλ ]. In particular, when ω ∼ P ∈ S[H, M, Φ], then the distribution P L of the sum of
L independent copies of ω belongs to S[H, M, LΦ].
2.3.3
Semi-direct summation
For 1 ≤ ℓ ≤ L, let regular data Hℓ ⊂ Ωℓ = Rdℓ , Mℓ , Φℓ be given. To avoid complications, we assume
that for every ℓ,
• Hℓ = Ω ℓ ,
• Mℓ is bounded.
Let also an ǫ > 0 be given. We assume that ǫ is small, namely, Lǫ < 1.
Let us aggregate the given regular data into a new one by setting
H = Ω := Ω1 × ... × ΩL = Rd , d = d1 + ... + dL , M = M1 × ... × ML ,
and let us define function Φ(h; µ) : Ωd × M → R as follows:
Φ(h = [h1 ; ...; hL ]; µ = [µ1 ; ...; µL ]) = inf λ∈∆ǫ
PL
∆ǫ = {λ ∈ Rd : λℓ ≥ ǫ ∀ℓ &
ℓ=1 λℓ = 1}.
Pd
ℓ
ℓ
ℓ=1 λℓ Φℓ (h /λℓ ; µ ),
(6)
By evident reasons, the infimum in the description of Φ is achieved, and Φ is continuous. In addition,
Φ is convex in h ∈ Rd and concave in µ ∈ M. Postponing for a moment verification, the consequences
are that H = Ω = Rd , M and Φ form a regular data. We claim that
Whenever ω = [ω 1 ; ...; ω L ] is a Borel random variable taking values in Ω = Rd1 × ... × RdL ,
and the marginal distributions Pℓ , 1 ≤ ℓ ≤ L, of ω belong to the families S[Rdℓ , Mℓ , Φℓ ]
for all 1 ≤ ℓ ≤ L, the distribution P of ω belongs to S[Rd , M, Φ].
bℓ ∈ Mℓ such that
Indeed, since Pℓ ∈ S[Rdℓ , Mℓ , Φℓ ], there exists µ
bℓ ) ∀g ∈ Rdℓ .
ln(Eωℓ ∼Pℓ {exp{g T ω ℓ }}) ≤ Φℓ (g; µ
Let us set µ
b = [b
µ1 ; ...; µ
bL ], and let h = [h1 ; ...; hL ] ∈ Ω be given. We can find λ ∈ ∆ǫ such that
Φ(h; µ
b) =
L
X
ℓ=1
λℓ Φℓ (hℓ /λℓ ; µ
bℓ ).
6
Applying Hölder inequality, we get
(
E[ω1 ;...;ωL ]∼P
whence
ln E[ω1 ;...;ωL ]∼P
We see that
)
L
n
oλℓ
X
Y
exp{ [hℓ ]T ω ℓ } ≤
,
Eωℓ ∼Pℓ [hℓ ]T ω ℓ /λℓ
(
ℓ
ℓ=1
)!
L
X
X
ℓ T ℓ
exp{ [h ] ω }
≤
λℓ Φℓ (hℓ /λℓ ; µ
bℓ ) = Φ(h; µ
b).
ln E[ω1 ;...;ωL ]∼P
ℓ
(
ℓ=1
)!
X
exp{ [hℓ ]T ω ℓ }
≤ Φ(h; µ
b) ∀h ∈ H = Rd ,
ℓ
and thus P ∈ S[Rd , Mℓ , Φℓ ], as claimed.
It remains to verify that the function Φ defined by (6) indeed is convex in h ∈ Rd and concave in
µ ∈ M. Concavity in µ is evident. Further, functions λℓ Φℓ (hℓ /λℓ ; µ) (as perspective
of
P transformation
ℓ /λ , µ).
convex functions Φℓ (·; µ)) are jointly convex in λ and hℓ , and so is Ψ(λ, h; µ) = L
λ
Φ
(h
ℓ
ℓ=1 ℓ ℓ
Thus Φ(·; µ), obtained by partial minimization of Ψ in λ, indeed is convex.
2.3.4
Affine image
Let H, M, Φ be regular data, Ω be the embedding space of H, and x 7→ Ax + a be an affine mapping
¯
from Ω to Ω̄ = Rd , and let us set
¯
H̄ = {h̄ ∈ Rd : AT h̄ ∈ H}, M̄ = M, Φ̄(h̄; µ) = Φ(AT h̄; µ) + aT h̄ : H̄ × M̄ → R.
Note that H̄, M̄ and Φ̄ are regular data. It is immediately seen that
Whenever the probability distribution of a random variable ω belongs to R[H, M, Φ] (or
belongs to S[H, M, Φ]), the distribution P̄ [P ] of the random variable ω̄ = Aω + a belongs
to R[H̄, M̄, Φ̄] (respectively, belongs to S[H̄, M̄, Φ̄]).
2.3.5
Incorporating support information
Consider the situation as follows. We are given regular data H ⊂ Ω = Rd , M, Φ and are interested in
the family of distribution P known to belong to S[H, M, Φ]. In addition, we know that all distributions
P from P are supported on a given closed convex set X ⊂ Rd . How could we incorporate this domain
information to pass from the family S[H, M, Φ] containing P to a smaller family of the same type still
containing P ? We are about to give an answer in the simplest case of H = Ω. Specifically, denoting
by φX (·) the support function of X and selecting somehow a closed convex set G ⊂ Rd containing the
origin, let us set
b µ) = inf Φ+ (h, g; µ) := Φ(h − g; µ) + φX (g) ,
Φ(h;
g∈G
where Φ(h; µ) : Rd × M → R is the continuous convex-concave function participating in the original
b is real-valued and continuous on the domain Rd × M (which definitely
regular data. Assuming that Φ
b is
is the case when G is a compact set such that φX is finite and continuous on G), note that Φ
d
b
convex-concave on this domain, so that R , M, Φ is a regular data. We claim that
b contains P, provided the family S[Rd , M, Φ] does so, and the first
The family S[Rd , M, Φ]
of these two families is smaller than the second one.
7
Verification of the claim is immediate. Let P ∈ P, so that for properly selected µ = µP ∈ M and for
all e ∈ Rd it holds
Z
T
exp{e ω}P (dω) ≤ Φ(e; µP ).
ln
Rd
Besides this, for every g ∈ G we have φX (ω) − gT ω ≥ 0 on the support of P , whence for every h ∈ Rd
one has
Z
Z
T
T
T
exp{h ω + φX (g) − g ω}P (dω) ≤ φX (g) + Φ(h − g; µP ).
exp{h ω}P (dω) ≤ ln
ln
Rd
Rd
Since the resulting inequality holds true for all g ∈ G, we get
Z
b µP ) ∀h ∈ Rd ,
exp{hT ω}P (dω) ≤ Φ(h;
ln
Rd
b since P ∈ P is arbitrary, the first part of the claim is justified. The
implying that P ∈ S[Rd , M, Φ];
d
d
b
b ≤ Φ, and the latter is due
inclusion S[R , M, Φ] ⊂ S[R , M, Φ] is readily given by the inequality Φ
b µ) ≤ Φ(h − 0, µ) + φX (0).
to Φ(h,
Illustration: distributions with bounded support revisited. In section 2.2.4, given a convex
compact set X ⊂ Rd with support function φX , we checked that the data H = Rd , M = X,
Φ(h; µ) = hT µ + 18 [φX (h) + φX (−h)]2 are regular and the family S[Rd , M, Φ] contains the family
P[X] of all Borel probability distributions supported on X. Moreover, for every µ ∈ M = X, the
family S[Rd , {µ}, Φ Rd ×{µ} ] contains all supported on X distributions with the expectations e[P ] = µ.
R
T
Note that Φ(h; e[P ]) describes well the behaviour of the logarithm FP (h) = ln Rd eh ω P (dω) of
the moment-generating function of P ∈ P[X] when h is small (indeed, FP (h) = hT e[P ] + O(khk2 ) as
h → 0), and by far overestimates FP (h) when h is large. Utilizing the above construction, we replace
Φ with the real-valued, convex-concave and continuous on Rd × M function
b µ) = inf (h − g)T µ + 1 [φX (h − g) + φX (−h + g)]2 + φX (g) ≤ Φ(h; µ).
Φ(h;
g
8
b ·) still ensures the inclusion P ∈ S[Rd , {e[P ]}, Φ
b d
It is easy to see that Φ(·;
] for every distriR ×{e[P ]}
bution P ∈ P[X] and “reproduces FP (h) reasonably well” for both small and large h. Indeed, since
b e[P ]) ≤ Φ(h; e[P ]), for small h Φ(h;
b e[P ]) reproduces FP (h) even better than Φ(h; e[P ]),
FP (h) ≤ Φ(h;
and we clearly have
1
2
T
b
Φ(h; µ) ≤ (h − h) µ + [φX (h − h) + φX (−h + h)] + φX (h) = φX (h) ∀µ,
8
and φX (h) is a correct description of FP (h) for large h.
3
3.1
Affine detectors and hypothesis testing
Situation
Assume we are given two collections of regular data with common Ω = Rd and H, specifically, the
collections (H, Mχ , Φχ ), χ = 1, 2. We start with a construction of a specific test for a pair of hypotheses
H1 : P ∈ P1 , H2 : P ∈ P2 , where
Pχ = R[H, Mχ , Φχ ], χ = 1, 2.
When building the test, we impose on the regular data in question the following
8
Assumption I: The regular data (H, Mχ , Φχ ), χ = 1, 2, are such that the convex-concave
function
Ψ(h; µ1 , µ2 ) =
1
2
[Φ1 (−h; µ1 ) + Φ2 (h; µ2 )] : H × (M1 × M2 ) → R
(7)
has a saddle point (min in h ∈ H, max in (µ1 , µ2 ) ∈ M1 × M2 ).
We associate with a saddle point (h∗ ; µ∗1 , µ∗2 ) the following entities:
• the risk
ǫ⋆ = exp{Ψ(h∗ ; µ∗1 , µ∗2 )};
(8)
this quantity is uniquely defined by the saddle point value of Ψ and thus is independent of how
we select a saddle point;
• the detector φ∗ (ω) – the affine function of ω ∈ Rd given by
φ∗ (ω) = hT∗ ω + a, a =
1
2
[Φ1 (−h∗ ; µ∗1 ) − Φ2 (h∗ ; µ∗2 )] .
(9)
A simple sufficient condition for existence of a saddle point of (7) is
Condition A:
The sets M1 and M2 are compact, and the function
Φ(h) =
max
µ1 ∈M1 ,µ2 ∈M2
Φ(h; µ1 , µ2 )
is coercive on H, meaning that Φ(hi ) → ∞ along every sequence hi ∈ H with khi k2 → ∞ as i → ∞.
Indeed, under Condition A by Sion-Kakutani Theorem [14] it holds
SadVal[Φ] := inf
max
Φ(h; µ1 , µ2 ) =
sup
inf Φ(h; µ1 , µ2 ),
µ1 ∈M1 ,µ2 ∈M2 h∈H
{z
}
|
{z
}
h∈H µ1 ∈M1 ,µ2 ∈M2
|
so that the optimization problems
Φ(µ1 ,µ2 )
Φ(h)
(P ) :
Opt(P ) = min Φ(h)
(D) :
Opt(D) =
h∈H
max
µ1 ∈M1 ,µ2 ∈M2
Φ(µ1 , µ2 )
have equal optimal values. Under Condition A, problem (P ) clearly is a problem of minimizing a
continuous coercive function over a closed set and as such is solvable; thus, Opt(P ) = Opt(D) is
a real. Problem (D) clearly is the problem of maximizing over a compact set of an upper semicontinuous (since Φ is continuous) function taking real values and, perhaps, value −∞, and not
identically equal to −∞ (since Opt(D) is a real), and thus (D) is solvable. Thus, (P ) and (D) are
solvable with common optimal value, and therefore Φ has a saddle point.
3.2
3.2.1
Pairwise testing regular families of distributions
Main observation
An immediate (and crucial!) observation is as follows:
9
Proposition 3.1 In the situation of section 3.1 and under Assumption I, one has
R
ΩRexp{−φ∗ (ω)}P (dω) ≤ ǫ⋆ ∀P ∈ P1 = R[H, M1 , Φ1 ]
Ω exp{φ∗ (ω)}P (dω) ≤ ǫ⋆ ∀P ∈ P2 = R[H, M2 , Φ2 ].
(10)
Proof. For every µ1 ∈ M1 , we have Φ1 (−h∗ ; µ1 ) ≤ Φ1 (−h∗ ; µ∗1 ), and for every P ∈ P1 , we have
Z
exp{−hT∗ ω}P (dω) ≤ exp{Φ1 (−h∗ ; µ1 )}
Ω
for properly selected µ1 ∈ M1 . Thus,
Z
exp{−hT∗ ω}P (dω) ≤ exp{Φ1 (−h∗ ; µ∗1 )} ∀P ∈ P1 ,
Ω
whence also
Z
Ω
exp{−hT∗ ω − a}P (dω) ≤ exp{ 12 [Φ1 (−h∗ ; µ∗1 ) + Φ2 (h∗ , µ∗2 )]} = ǫ⋆ ∀P ∈ P1 .
Similarly, for every µ2 ∈ M2 , we have Φ2 (h∗ ; µ2 ) ≤ Φ2 (h∗ ; µ∗2 ), and for every P ∈ P2 , we have
Z
exp{hT∗ ω}P (dω) ≤ exp{Φ2 (h∗ ; µ2 )}
Ω
for properly selected µ2 ∈ M2 . Thus,
Z
exp{hT∗ ω}P (dω) ≤ exp{Φ2 (h∗ ; µ∗2 )} ∀P ∈ P2 ,
Ω
whence
Z
Ω
3.2.2
exp{hT∗ ω + a}P (dω) ≤ exp{ 12 [Φ1 (−h∗ ; µ∗1 ) + Φ2 (h∗ , µ∗2 )]} = ǫ⋆ ∀P ∈ P2 .
Testing pairs of hypotheses
Repeated observation.
as follows:
Given Ω = Rd , let random observations ωt ∈ Ω, t = 1, 2, ..., be generated
“In the nature” there exists a random sequence ζt ∈ RN , t = 1, 2, ...,, of driving factors
such that ωt is a deterministic function of ζ t = (ζ1 , ..., ζt ), t = 1, 2, ...;
we refer to this situation as to case of repeated observation ω ∞ .
Let now (H, M, Φ) be a regular data with observation space Ω. We associate with this data
four hypotheses on the stochastic nature of observations ω ∞ = {ω1 , ω2 , ...}. Denoting by P|ζ t−1
the conditional, ζ t−1 being given, distribution of ωt , we say that the (distribution of the) repeated
observation ω ∞ obeys hypothesis
• HR [H, M, Φ], if P|ζ t−1 ∈ R[H, M, Φ] for all t and all ζ t−1 ;
• HS [H, M, Φ], if P|ζ t−1 ∈ S[H, M, Φ] for all t and all ζ t−1 ;
• HRi [H, M, Φ], if P|ζ t−1 is independent of t and ζ t−1 and belongs to R[H, M, Φ];
• HSi [H, M, Φ], if P|ζ t−1 is independent of t and ζ t−1 and belongs to S[H, M, Φ].
10
Note that the last two hypotheses, in contrast to the first two, require from the observations ω1 , ω2 , ...
to be i.i.d. Note also that HR is weaker than HS , HRi is weaker than HSi , and the “non-stationary”
hypotheses HR , HS are weaker than their respective stationary counterparts HRi , HSi .
The tests to be considered in the sequel operate with the initial fragment ω K = (ω1 , ..., ωK ),
of a prescribed length K, of the repeated observation ω ∞ = {ω1 , ω2 , ...}. We call ω K K-repeated
observation and say that (the distribution of) ω K obeys one of the above hypotheses, if ω K is cut off
the repeated observation ω ∞ distributed according to the hypothesis in question. We can think of
HR , HS , HRi and HSi as hypotheses about the distribution of the K-repeated observation ω K .
Pairwise hypothesis testing from repeated observations. Assume we are given two collections
of regular data (H, Mχ , Φχ ), χ = 1, 2, with common observation space Ω = Rd and common H. Given
positive integer K and K-repeated observation ω K = (ω1 , ..., ωK ), we want to decide on the pair of
hypotheses Hχ = HR [H, Mχ , Φχ ], χ = 1, 2, on the distribution of ω K .
Assume that the convex-concave function (7) associated with the pair of regular data in question
has a saddle point (h∗ ; µ∗1 , µ∗2 ), and let φ∗ (·), ǫ⋆ be the induced by this saddle point affine detector and
its risk, see (9), (8). Let us set
K
X
(K) K
φ∗ (ωt ).
φ∗ (ω ) =
t=1
Consider decision rule
T∗K
for hypotheses Hχ , χ = 1, 2, which, given observation ω K ,
(K)
• accepts H1 (and rejects H2 ) if φ∗ (ω K ) > 0;
(K)
• accepts H2 (and rejects H1 ) if φ∗ (ω K ) < 0;
(K)
• in the case of a tie (when φ∗ (ω K ) = 0) the test, say, accepts H1 and rejects H2 .
(K)
In what follows, we refer to T∗K as to test associated with detector φ∗ .
Proposition 3.2 In the situation in question, we have
o
n
(K)
∞ obeys H ;
(a) Eζ K exp{−φ∗ (ω K )} ≤ ǫK
1
⋆ , when ω
o
n
(K) K
when ω ∞ obeys H2 .
(b) Eζ K exp{φ∗ (ω )} ≤ ǫK
⋆ ,
(11)
As a result, the test T∗K accepts exactly one of the hypotheses H1 , H2 , and the risk of this test —
the maximal, over χ = 1, 2, probability not to accept the hypothesis Hχ when it is true (i.e., when the
K-repeated observation ω K obeys the hypothesis HR [H, Mχ , Φχ ]) — does not exceed ǫK
⋆ .
Proof. The fact that the test always accepts exactly one of the hypotheses Hχ , χ = 1, 2, is evident.
Let us denote Eζ t the expectation w.r.t. the distribution of ζ t , and let Eζt+1|ζ t stand for expectation
w.r.t. conditional to ζ t distribution of ζt+1 . Assuming that H1 holds true and invoking the first
inequality in (10), we have
(1)
Eζ1 {exp{−φ∗ (ω1 )}} ≤ ǫ⋆ ,
(t+1)
Eζ t+1 {exp{−φ∗
o
n
(t)
(ω t+1 )} = Eζ t exp{−φ∗ (ω t )}Eζt+1 |ζ t {exp{−φ∗ (ωt+1 )}}
o
n
(t)
≤ ǫ⋆ Eζ t exp{−φ∗ (ω t )} , 1 ≤ t < K
(we have taken into account that ωt+1 is a deterministic function of ζ t+1 and that we are in the case
where the conditional to ζ t distribution of ωt+1 belongs to P1 = R[H, M1 , Φ1 ]), and we arrive at
11
(11.a). This inequality clearly implies that the probability to reject H1 when the hypothesis is true is
(K) K
K
≤ ǫK
⋆ (since φ∗ (ω ) ≤ 0 when T∗ rejects H1 ). Assuming that H2 is true and invoking the second
inequality in (10), similar reasoning shows that (11.b) holds true, so that the probability to reject H2
when the hypothesis is true does not exceed ǫK
⋆ .
3.2.3
Illustration: sub-Gaussian and Gaussian cases
For χ = 1, 2, let Uχ be nonempty closed convex set in Rd , and Uχ be a compact convex subset of the
interior of the positive semidefinite cone Sd+ . We assume that U1 is compact. Setting
Hχ = Ω = Rd , Mχ = Uχ × Uχ , Φχ (h; θ, Θ) = θ T h + 12 hT Θh : Hχ × Mχ → R, χ = 1, 2,
(12)
we get two collections (H, Mχ , Φχ ), χ = 1, 2, of regular data. As we know from section 2.2.1, for
χ = 1, 2, the families of distributions S[Rd , Mχ , Φχ ] contain the families SG[Uχ , Uχ ] of sub-Gaussian
distributions on Rd with sub-Gaussianity parameters (θ, Θ) ∈ Uχ × Uχ (see (4)), as well as families
G[Uχ , Uχ ] of Gaussian distributions on Rd with parameters (θ, Θ) (expectation and covariance matrix)
running through Uχ × Uχ . Besides this, the pair of regular data in question clearly satisfies Condition
A. Consequently, the test T∗K given by the above construction as applied to the collections of regular
data (12) is well defined and allows to decide on hypotheses Hχ = HR [Rd , Uχ , U χ], χ = 1, 2, on
the distribution of the K-repeated observation ω K . The same test can be also used to decide on
stricter hypotheses HχG , χ = 1, 2, stating that the observations ω1 , ..., ωK are i.i.d. and drawn from a
Gaussian distribution P belonging to G[Uχ , Uχ ]. Our goal now is to process in detail the situation in
question and to refine our conclusions on the risk of the test T∗1 when the Gaussian hypotheses HχG
are considered and the situation is symmetric, that is, when U1 = U2 .
Observe, first, that the convex-concave function Ψ from (7) in the situation under consideration
becomes
1
1
(13)
Ψ(h; θ1 , Θ1 , θ2 , Θ2 ) = 12 hT [θ2 − θ1 ] + hT Θ1 h + hT Θ2 h.
4
4
We are interested in solutions to the saddle point problem – find a saddle point of function (13) –
min
h∈Rd
max
θ1 ∈U1 ,θ2 ∈U2
Θ1 ∈U1 ,Θ2 ∈U2
Ψ(h; θ1 , Θ1 , θ2 , Θ2 )
(14)
From the structure of Ψ and compactness of U1 , U1 , U2 , combined with the fact that Uχ , χ = 1, 2,
are comprised of positive definite matrices, it immediately follows that saddle points do exist, and a
saddle point (h∗ ; θ1∗ , Θ∗1 , θ2∗ , Θ∗2 ) satisfies the relations
(a) h∗ = [Θ∗1 + Θ∗2 ]−1 [θ1∗ − θ2∗ ],
(b) hT∗ (θ1 − θ1∗ ) ≥ 0 ∀θ1 ∈ U1 , hT∗ (θ2∗ − θ2 ) ≥ 0 ∀θ2 ∈ U2 ,
(c) hT∗ Θ1 h∗ ≤ hT∗ Θ∗1 h∗ ∀Θ1 ∈ U1 , hT∗ Θ∗2 h∗ ≤ h∗ Θ∗2 h∗ ∀Θ2 ∈ U2 .
(15)
From (15.a) it immediately follows that the affine detector φ∗ (·) and risk ǫ⋆ , as given by (8) and (9),
are
φ∗ (ω) = hT∗ [ω − w∗ ] + 12 hT∗ [Θ∗1 − Θ∗2 ]h∗ , w∗ = 12 [θ1∗ + θ2∗ ];
(16)
ǫ⋆ = exp{− 41 [θ1∗ − θ2∗ ]T [Θ∗1 + Θ∗2 ]−1 [θ1∗ − θ2∗ ]}
1 T
∗
∗
= exp{− 4 h∗ [Θ1 + Θ2 ]h∗ }.
Note that in the symmetric case (where U1 = U2 ), there always exists a saddle point of Ψ with
Θ∗1 = Θ∗2 , and the test T∗1 associated with such saddle point is quite transparent: it is the maximum
likelihood test for two Gaussian distributions, N (θ1∗ , Θ∗ ), N (θ2∗ , Θ∗ ), where Θ∗ is the common value
12
of Θ∗1 and Θ∗2 , and the bound ǫ⋆ for the risk of the test is nothing but the Hellinger affinity of these
two Gaussian distributions, or, equivalently,
∗
∗
(17)
ǫ⋆ = exp − 18 [θ1∗ − θ2∗ ]T Θ−1
∗ [θ1 − θ2 ] .
Applying Proposition 3.2, we arrive at the following result:
Proposition 3.3 In the symmetric sub-Gaussian case (i.e., in the case of (12) with U1 = U2 ), saddle
point problem (13), (14) admits a saddle point of the form (h∗ ; θ1∗ , Θ∗ , θ2∗ , Θ∗ ), and the associated affine
detector and its risk are given by
φ∗ (ω) = hT∗ [ω − w∗ ], w∗ = 21 [θ1∗ + θ2∗ ];
ǫ⋆ = exp{− 18 [θ1∗ − θ2∗ ]T Θ∗−1 [θ1∗ − θ2∗ ]}.
(18)
As a result, when deciding, via ω K , on “sub-Gaussian hypotheses” HS [Rd , Mχ , Mχ ], χ = 1, 2 (in
fact - even on weaker hypotheses HR [Rd , Mχ , Mχ ], χ = 1, 2), the risk of the test T∗K associated with
P
(K)
K
φ∗ (ω K ) := K
t=1 φ∗ (ωt ) is at most ǫ⋆ .
In the symmetric single-observation Gaussian case, that is, when U1 = U2 and we apply the test
T∗ = T∗1 to observation ω ≡ ω1 in order to decide on the hypotheses HχG , χ = 1, 2, the above risk
bound can be improved:
Proposition 3.4 Consider symmetric case U1 = U2 = U , let (h∗ ; θ1∗ ; Θ∗1 , θ2∗ , Θ∗2 ) be “symmetric” –
with Θ∗1 = Θ∗2 = Θ∗ – saddle point of function Ψ given by (13), and let φ∗ be the affine detector given
by (15) and (16):
1
[θ ∗ − θ2∗ ], w∗ = 12 [θ1∗ + θ2∗ ].
φ∗ (ω) = hT∗ [ω − w∗ ], h∗ = Θ−1
2 ∗ 1
Let also
q
q
∗
∗
(19)
δ = hT∗ Θ∗ h∗ = 12 [θ1∗ − θ2∗ ]T Θ−1
∗ [θ1 − θ2 ],
so that
δ2 = hT∗ [θ1∗ − w∗ ] = hT∗ [w∗ − θ2∗ ] and ǫ⋆ = exp{− 21 δ2 }.
(20)
(a) ∀(θ ∈ U1 , Θ ∈ U ) : Probω∼N (θ,Θ) {φ∗ (ω) ≤ α} ≤ Erf(δ − α/δ)
(b) ∀(θ ∈ U2 , Θ ∈ U ) : Probω∼N (θ,Θ) {φ∗ (ω) ≥ −β} ≤ Erf(δ − β/δ),
(21)
Let, further, α ≤
where
δ2 ,
β≤
δ2 .
Then
Z ∞
1
Erf(s) = √
exp{−r 2 /2}dr
2π s
is the normal error function. In particular, when deciding, via a single observation ω, on Gaussian
hypotheses HχG , χ = 1, 2, with HχG stating that ω ∼ N (θ, Θ) with (θ, Θ) ∈ Uχ × U , the risk of the test
T∗1 associated with φ∗ is at most Erf(δ).
Proof. Let us prove (a) (the proof of (b) is completely similar). For θ ∈ U1 , Θ ∈ U we have
Probω∼N (θ,Θ) {φ∗ (ω) ≤ α} = Probω∼N (θ,Θ) {hT∗ [ω − w∗ ] ≤ α}
= Probξ∼N (0,I) {hT∗ [θ + Θ1/2 ξ − w∗ ] ≤ α}
= Probξ∼N (0,I) {[Θ1/2 h∗ ]T ξ ≤ α − hT∗ [θ − w∗ ] } ≤ Probξ∼N (0,I) {[Θ1/2 h∗ ]T ξ ≤ α − δ2 }
| {z }
∗
2
≥hT
∗ [θ1 −w∗ ]=δ
by (15.b),(20)
= Erf([δ2 − α]/kΘ1/2 h∗ k2 )
1/2
≤ Erf([δ2 − α]/kΘ∗ h∗ k2 ) [since δ2 − α ≥ 0 and hT∗ Θh∗ ≤ hT∗ Θ∗ h∗ by (15.c)]
= Erf([δ2 − α]/δ).
The “in particular” part of Proposition is readily given by (21) as applied with α = β = 0.
13
3.3
Testing multiple hypotheses from repeated observations
Consider the situation as follows: we are given
• observation space Ω = Rd and a symmetric w.r.t. the origin closed convex set H;
• J closed convex sets Mj ⊂ Rnj , j = 1, ..., J, along with J convex-concave continuous functions
Φj (h; µj ) : H × Mj → R.
These data give rise to J hypotheses Hj = HR [H, Mj , Φj ] on the distribution of repeated observation
ω ∞ = {ωt ∈ Rd , t ≥ 1}. On the top of it, assume we are given a closeness relation – a subset
C ⊂ {1, ..., J}2 which we assume to contain the diagonal ((j, j) ∈ C for all j ≤ J) and to be symmetric
((i, j) ∈ C if and only if (j, i) ∈ C). In the sequel, we interpret indexes i, j with (i, j) ∈ C (and the
hypotheses Hi , Hj with these indexes) as C-close to each other.
Our goal is, given a positive integer K and K-repeated observation ω K = (ω1 , ..., ωK ), to decide,
“up to closeness C,” on the hypotheses Hj , j = 1, ..., J (which is convenient to be thought of as
hypotheses about the distribution of ω K ).
Let us act as follows3 . Let us make
Assumption II For every pair i, j, 1 ≤ i < j ≤ J, with (i, j) 6∈ C, the convex-concave
function
Ψij (h; µi , µj ) =
1
2
[Φi (−h; µi ) + Φj (h; µj )] : H × (Mi × Mj ) → R
ij
has a saddle point (hij ; µij
i , µj ) on H × (Mi × Mj ) (min in h, max in µi , µj ),
and let us set
ij
ǫij = exp{Ψij (hij ; µij
i , µj )},
φij (ω) = [hij ]T ω h+ aij
aij =
Φi (−hij ; µij
i )−
1
2
i , 1 ≤ i < j ≤ J & (i, j) 6∈ C
Φj (hij ; µij
j )
(cf. (8) and (9)). We further set
(22)
φij (·) ≡ 0, ǫij = 1 ∀(i, j) ∈ C,
and
φij (ω) = −φji (ω), ǫij = ǫji , 1 ≤ j < i ≤ J & (i, j) 6∈ C,
and set
(K)
φij (ω K ) =
K
X
t=1
Now, by construction,
(K)
φij (ωt ), 1 ≤ i, j ≤ J.
(K)
φij (·) = −φji (·), ǫij = ǫji , 1 ≤ i, j ≤ J.
Observe that for every j ≤ J such that the K-repeated observation ω K obeys hypothesis Hj =
HR [H, Mj , Φj ], we have
n
o
(K)
E exp{φij (ω K )} ≤ ǫK
(23)
ij , i = 1, ..., J.
3
The construction we are about to present and the first related result (Proposition 3.5) originate from [13, Section
3]; we reproduce them below to make our exposition self-contained.
14
(we have used (11) along with φij ≡ −φji ).
Furthermore, whenever [αij ]i,j≤J is a skew-symmetric matrix (i.e., αij = −αji ), the shifted detectors
(K)
(K)
φbij (ω K ) = φij (ω K ) + αij
(24)
satisfy the scaled version of (23), specifically, for every j ≤ J such that the K-repeated observation
ω K obeys hypothesis Hj = HR [Rd , Mj , Φj ], we have
n
o
(K)
E exp{φbij (ω K )} ≤ ǫK
(25)
ij exp{αij }, i = 1, 2, ..., J.
The bottom line is as follows:
Proposition 3.5 In the situation in question, given a closeness C, consider the following test TCK
deciding on the hypotheses Hj , 1 ≤ j ≤ J, on the distribution of K-repeated observation ω K : given
skew-symmetric shift matrix [αij ]i,j and observation ω K , TCK accepts all hypotheses Hi such that
(K)
φbij (ω K ) > 0 whenever (i, j) 6∈ C
(∗i )
and reject all hypotheses Hi for which the predicate (∗i ) does not take place. Then
(i) Test TCK accepts some of (perhaps, none of ) hypotheses Hi , i = 1, ..., J, and all accepted
hypotheses, if any, are C-close to each other. Besides, TCK has C-risk at most
X
ǫK
b
ǫ = max
ij exp{−αij },
i
j:(i,j)6∈C
meaning that for every j∗ ≤ J such that the distribution P̄K of ω K obeys the hypothesis Hj∗ (i.e., the
hypothesis Hj∗ is true), the P̄K -probability of the event
”either the true hypothesis HjK∗ is not accepted, or the list of accepted hypotheses contains
a hypothesis which is not C-close to Hj∗ ”
does not exceed b
ǫ.
(ii) The infimum of b
ǫ over all skew-symmetric shifts αij is exactly the spectral norm kE (K) k2,2 of
the symmetric entry-wise nonnegative matrix
E
(K)
=
(K)
Eij
=
ǫK
6 C
ij , (i, j) ∈
0,
(i, j) ∈ C
J
.
(26)
i,j=1
This infimum is attained when the Perron-Frobenius eigenvector g of E (K) can be selected to be positive,
in which case an optimal selection of αij is the selection
αij = ln(gi /gj ), 1 ≤ i, j ≤ J.
(K)
Proof. Given ω K , the test TCK can accept hypotheses Hi and Hj with (i, j) 6∈ C only when φij (ω K ) >
(K)
(K)
(K)
0 and φji (ω K ) > 0, which is impossible due to φij (·) = −φji (·). Thus, TCK can accept Hi and
Hj only when (i, j) ∈ C. Further, let the distribution P̄K of ω K obey hypothesis Hj∗ . Invoking (25)
(K)
(K)
and the relation φbj∗ j (·) = −φbjj∗ (·), for every j ≤ J with (j∗ , j) 6∈ C, the P̄K -probability of the
(K)
(K)
event φb (ω K ) ≤ 0, or, which is the same, of the event φb (ω K ) ≥ 0, is at most ǫK exp{αjj } =
j∗ j
jj∗
15
jj∗
∗
ǫK
j∗ j exp{−αj∗ j }. Using the union bound, the P̄K -probability of the event “Hj∗ is not accepted” is at
most
X
X (K)
ǫK
Ej∗ j exp{−αj∗ j } ≤ b
ǫ.
j∗ j exp{−αj∗ j } =
j
j:(j∗ j)6∈C
By construction of the test, when Hj∗ is accepted and j is not C-close to j∗ , Hj is not accepted
(we have already seen that the test never accepts a pair of hypotheses which are not C-close to each
other). Thus, the C-risk of TCK indeed is at most b
ǫ. Now, E (K) is a symmetric entry-wise nonnegative
matrix, so that its leading eigenvector g can be selected to be nonnegative. When g is positive, setting
αij = ln(gi /gj ), we get for every i
X (K)
X (K)
Eij exp{−αij } =
Eij gj /gi = (E (K) g)i /gi = kE (K) k2,2 ,
j
j
and thus b
ǫ = kE (K) k2,2 . The fact that this is the smallest possible, over skew-symmetric shifts αij ,
value of b
ǫ is proved in [13]. When g is just nonnegative, consider close to E (K) symmetric matrix
b with positive entries E
bij ≥ E (K) ; utilizing the (automatically strictly positive) Perron-Frobenius
E
ij
eigenvector g of this matrix, we, as was just explained, get skew-symmetric shifts αij such that
X
bij exp{−αij } ≤ kEk
b 2,2
E
j
P (K)
for all i; the left hand side in this inequality is ≥ j Eij exp{−αij }, and the right hand side can be
b close enough to E (K) . Thus, we indeed can make b
made arbitrarily close to kE (K) k2,2 by making E
ǫ
(K)
arbitrarily close to kE k2,2 .
3.3.1
Special case: inferring colors
Assume that we are given J collections (H, Mj , Φj ), 1 ≤ j ≤ J, of regular data with common observation space Ω = Rd and common H, and thus have at our disposal J hypotheses Hj = HR [H, Mj , Φj ],
on the distribution of K-repeated observation ω K . Let also the index set {1, ..., J} be partitioned into
L ≥ 2 non-overlapping nonempty subsets I1 , ..., IL ; it is convenient to think about ℓ, 1 ≤ ℓ ≤ L, as the
common color of indexes j ∈ Iℓ , and that the colors of indexes j are inherited by the hypotheses Hj .
The Color Inference (CI) problem we want to solve amounts to decide, given K-repeated observation
ω K obeying one or more of the hypotheses H1 , ..., HJ , on the color of these hypotheses. Note that
it may happen that the distribution of ω K obeys a pair of hypotheses Hi , Hj of different colors. If
it is not the case – that is, no distribution of ω K obeys a pair of hypotheses Hi , Hj of two distinct
colors – we call the CI problem well-posed. In the well-posed case, we can speak about the color of
the distribution of ω K provided this distribution obeys the union, over j = 1, ..., J, of hypotheses Hj ;
this is the color of (any of) the hypotheses Hj obeyed by the distribution of ω K , and the CI problem
is to infer this color given ω K .
In order to process the CI problem via our machinery, let us define closeness C as follows:
(i, j) ∈ C ⇔ i and j are of the same color.
Assuming that the resulting C ensures validity of Assumption II, we can apply the above scheme to
build test TCK . We can then convert this test into a color inference as follows. Given a K-repeated
observation ω K , it may happen that TCK , as applied to ω K , accepts one or more among the hypotheses
Hj . In this case, by item (i) of Proposition 3.5, all accepted hypotheses are C-close to each other (in
other words, are of the same color), and we claim that this is the color of the distribution of K-repeated
16
observation we are dealing with. And if TCK , as applied to ω K , accepts nothing, we claim that the
color we are interested in remains undecided.
Let us analyze the just described color inferring procedure, let it be denoted AK . Observe, first,
that in the situation in question, assuming w.l.o.g. that the sets I1 , I2 , ..., IL are consecutive fragments
in {1, ..., J}, the matrix E (K) given by (26) is naturally partitioned into L × L blocks E pq = (E qp )T ,
(K)
1 ≤ p, q ≤ L, where E pq is comprised of entries Eij with i ∈ Ip , j ∈ Iq . By construction of E (K) , the
diagonal blocks E pp are zero, and off-diagonal blocks are entry-wise positive (since ǫij clearly is positive
for all pairs i, j of different colors). It follows that Perron-Frobenius eigenvectors of E (K) are strictly
positive. This implies that for properly selected shifts αij = −αji , the quantity b
ǫ in Proposition 3.5 is
equal to kE (K) k2,2 ; in what follows we assume that the test TCK utilizes exactly these optimal shifts,
so that we are in the case of b
ǫ = kE (K) k2,2 .
Now, it may happen (“bad case”) that that kE (K) k2,2 ≥ 1; in this case Proposition 3.5 says nothing
meaningful about the quality of the test TCK , and consequently, we cannot say much about the quality
of AK . In contrast to this, we claim that
Lemma 3.1 Assume that b
ǫ := kE (K) k2,2 < 1. Then the CI problem is well posed, and whenever the
distribution P̄K of ω K obeys one of the hypotheses Hj , j = 1, ..., J, AK recovers correctly the color of
P̄K with P̄K -probability at least 1 − b
ǫ.
The proof is immediate. In the good case, all entries in E (K) are of magnitude < 1, whence ǫij < 1
whenever (i, j) 6∈ C, see (26), so that
ǭ := max{ǫij : (i, j) 6∈ C} < 1.
i,j
(27)
It follows that the nonzero entries in E (M) are nonnegative and ≤ ǭM , whence
b
ǫ(M ) := kE (M) k2,2 ≤ J ǭM → 0, as M → ∞.
(28)
In particular, for properly selected M we have b
ǫ(M ) < 1/2. Applying Proposition 3.5 with M in
the role of K, we see that if the distribution P̄K of ω K obeys hypothesis Hj∗ with some j∗ ≤ J
(due the origin of our hypotheses, this is exactly the same as to say that the distribution P̄M of ω M
ǫ(M ) > 1/2 the test TCM accepts hypothesis Hj∗ .
obeys Hj∗ ), then with P̄M -probability at least 1 − b
It follows that if P̄M obeys both Hj ′ and Hj ′′ , then TCM will accept Hj ′ and Hj ′′ simultaneously
with positive P̄M -probability, and since, as we have already explained, TCM never accepts two
hypotheses of different color simultaneously, we conclude that Hj ′ and Hj ′′ are of the same color.
This conclusion holds true whenever the distribution of ω K obeys one or more of the hypotheses
Hj , 1 ≤ j ≤ K, meaning that the CI problem is well posed.
Invoking Proposition 3.5, we conclude that if the distribution P̄K of ω K obeys, for some j∗ ,
ǫ(K); and from
hypothesis Hj∗ , then the P̄K -probability for TCK to accept Hj∗ is at least 1 − b
preceding analysis, whenever TCK accepts Hj∗ such that P̄K obeys Hj∗ , AK correctly infers
color of P̄K , as claimed.
the
the
the
ǫ(K) → 0 as K → ∞, so that the CI
Finally, we remark that when (27) holds, (28) implies that b
problem is well posed, and for every desired risk level ǫ ∈ (0, 1) we can find efficiently observation time
K = K(ǫ) such that b
ǫ(K) ≤ ǫ. As a result, for this K the color inferring procedure AK recovers the
color of the distribution P̄K of ω K (provided this distribution obeys some of the hypotheses H1 , ..., HJ )
with P̄K -probability at least 1 − ǫ.
4
Application: aggregating estimates by testing
Let us consider the situation as follows:
17
• We are given I triples of regular data (H, Mi , Φi ), 1 ≤ i ≤ I, with common H and Ω = Rd
and the parameter sets Mi sharing the common embedding space Rn ; for the sake of simplicity,
assume that Mi are bounded (and thus are nonempty convex compact sets in Rn ) and the
continuous convex-concave functions Φi (h; µ) : H × Mj → R are coercive in H: Φi (ht , µ) → ∞
whenever µ ∈ Mi and sequence {ht ∈ H, t ≥ 1} satisfies kht k2 → ∞, t → ∞.
• We observe a realization of K-repeated observation ω K = (ω1 , ..., ωK ) with i.i.d.
S ωt ’s drawn
S[H, Mi , Φi ].
from unknown probability distribution P̄ known to belong to the family P =
i≤I
Thus, “in the nature” there exists ī ≤ I and µ̄ ∈ Mī such that
ln Eω∼P̄ {exp{hT ω}} ≤ Φī (h, µ̄) ∀h ∈ H.
(29)
we call µ̄ the parameter associated with P̄ , and our goal is to recover from K-repeated observation
ω K the image ḡ = Gµ̄ of µ̄ under a given linear mapping µ 7→ Gµ : Rn → Rm .
Undoubtedly, parameter estimation problem is a fundamental problem of mathematical statistics, and
as such is the subject of a huge literature. In particular, several constructions of estimators based
on testing of convex hypotheses have been studied in connection with signal reconstruction [4, 5, 6]
and linear functionals estimation [10, 11]. Our actual goal to be addressed below is more modest: we
assume that we are given L candidate estimates g1 , ..., gL of ḡ (these estimates could be outputs of
various estimation routines applied to independent observations sampled form P̄ ), and our goal is to
select the best – the k · k2 -closest to ḡ among the estimates g1 , ..., gL . This is the well-known problem
of aggregating estimates, and our goal is to process this aggregation problem via the Color Inference
procedure from section 3.3.1.
4.1
Aggregation procedure
It should be stressed that as stated, the aggregation problem appears to be ill-posed: there could be
several pairs (ī, µ̄) satisfying (29), and the values of Gµ at the µ-components of these pairs could be
different for different pairs, so that ḡ not necessary is well defined. One way to resolve this ambiguity
would be to assume that given P̄ ∈ P, relation (29) uniquely defines µ̄. We, however, prefer another
setting: µ̄ and ī satisfying (29), same as P̄ ∈ P, are “selected by nature” (perhaps, from several
alternatives), and the performance of the aggregating procedure we are about to develop should be
independent of what is the nature’s selection.
When processing the aggregation problem, we assume w.l.o.g. that all points g1 , ..., gL are distinct
from each other. Let us split the space Rm where Gµ takes values into L Voronoi cells
Vℓ = {g ∈ Rm : kg − gℓ k2 ≤ kg − gℓ′ k2 ∀ℓ′ ≤ L}
= {g ∈ Rm : uTℓℓ′ g ≤ vℓℓ′ ∀(ℓ′ ≤ L, ℓ′ 6= ℓ)},
′
′
1 T
uℓℓ′ = kgℓ′ − gℓ k−1
2 [gℓ′ − gℓ ], vℓℓ′ = 2 uℓℓ′ [gℓ′ + gℓ ], 1 ≤ ℓ, ℓ ≤ L, ℓ 6= ℓ .
(30)
Note that Vℓ is comprised of all points g from Rm for which gℓ is (one of) the k · k2 -closest to g among
the points g1 , ..., gL . Let us set
Wℓi = {µ ∈ Mi : Gµ ∈ Vℓ }, 1 ≤ i ≤ I, 1 ≤ ℓ ≤ L,
(31)
so that Wℓi are convex compact sets in Rm . Observe that gℓ can be a solution to the aggregation
problem (that is, the closest to ḡ point among g1 , ..., gL ) only when ḡ = Gµ̄ belongs to Vℓ , that is, only
when µ̄ ∈ Wℓī for some ī, implying that at least one of the sets Wℓ1 , Wℓ2 , ..., WℓI is nonempty. Whether
the latter condition is indeed satisfied for a given ℓ this can be found out efficiently via solving I convex
18
feasibility problems. If the latter condition does not hold for some ℓ (let us call the associated estimate
gℓ redundant), we can eliminate gℓ from the list of estimates to be aggregated without affecting the
solution to the aggregation problem. Then we can redefine the Voronoi cells for the reduced list of
estimates in the role of our original list, check whether this list still contains a redundant estimate,
eliminate the latter, if it exists, and proceed recursively until a list of estimates (which by construction
still contains all solutions to the aggregation problem) with no redundant estimates is built. We lose
nothing when assuming that this “purification” was carried out in advance, so that already the initial
list g1 , ..., gL of estimates does not contain redundant ones. Of course, we lose nothing when assuming
that L ≥ 2. Thus, from now on we assume that L ≥ 2 and for every ℓ ≤ L, at least one of the sets
Wℓi , 1 ≤ i ≤ I, is nonempty.
Note that to solve the aggregation problem to optimality is exactly the same, in terms of the sets
i
Wℓ , as to identify ℓ such that µ̄ ∈ Wℓi for some i. We intend to reduce this task to solving a series of
Color Inference problems. We start with presenting the principal building block of our construction –
Individual Inference procedure.
Individual Inference procedure is parameterized by ℓ ∈ {1, ..., L} and a real δ > 0. Given ℓ and
δ, we initialize the algorithm as follows:
• mark as red all nonempty sets Wℓi along with their elements and the corresponding regular data
(H, Wℓi , Φi H×W i );
ℓ
• look one by one at all sets Wℓi′ with i ≤ I and ℓ′ 6= ℓ, and associate with these sets their chunks
Wℓℓiδ′ = {µ ∈ Wℓi′ : uTℓℓ′ [Gµ] ≥ vℓℓ′ + δ}.
(32)
Note that the resulting sets are convex and compact. Whenever Wℓℓiδ′ is nonempty, we mark blue
this set along with its elements and the corresponding regular data (H, Wℓℓiδ′ , Φi H×W iδ ) .
ℓℓ′
As a result of the above actions, we get a collection of nonempty convex compact subsets Wℓsδ ,
sδ
s = 1, ..., Sℓδ , of Rd and associated regular data Dℓsδ = (H, Wℓsδ , Φsδ
ℓ ); the sets Wℓ , same as their
sδ
sδ
elements and regular data Dℓ , are colored in red and blue. Note that the sets Wℓ of different colors
do not intersect (since their images under the mapping G do not intersect), so that a point µ ∈ Rm
gets at most one color. Note also that our collection definitely contains red components.
δ
sδ
δ
Individual Inference Procedure Aℓδ
K infers the color of a regular data D ∈ Dℓ = {Dℓ , 1 ≤ s ≤ Sℓ }
K
given i.i.d. K-repeated observation ω drawn from a distribution P ∈ S[D]: when the collection Dδℓ
contains both red and blue regular data, Aℓδ
K is exactly Color Inference procedure from section 3.3.1
associated with this collection and our coloring4 ; if no blue regular data is present, Aℓδ
K always infers
that the color is red.
Observe that if the collection Dδℓ of regular data we have built contains no blue data for some value
δ̄ of δ, the same holds true for all δ ≥ δ̄. Let us define the risk b
ǫ(ℓ, δ) of Individual Inference Procedure
δ
with parameters ℓ, δ as follows: when δ is such that Dℓ contains no blue regular data, b
ǫ(ℓ, δ) = 0,
otherwise b
ǫ(ℓ, δ) is as stated in Proposition 3.5. Note that whenever δ > 0, ℓ ≤ L, s ≤ Sℓδ , µ ∈ Wℓsδ
and a probability distribution P satisfies
Eω∼P {exp{hT ω}} ≤ Φs (h, µ) ∀h ∈ H,
4
The procedure is well defined, since by our assumption, all convex-concave functions we need to deal with are
continuous, coercive in the minimization variable, and have closed convex minimization and compact convex maximization
domains, so that the required saddle points do exist.
19
the quantity b
ǫ(ℓ, δ) , by construction, is an upper bound on P -probability of the event “as applied
to observation ω K = (ω1 , ..., ωK ) with ω1 , ..., ωK drawn, independently of each other, from P , Aℓδ
K
does not recover correctly the color of µ.” Observe that b
ǫ(ℓ, δ) = 0 for large enough values of δ (since
for large δ the collection Dδℓ contains no blue data; recall that the parameter sets Mi are bounded).
Besides this, we claim that b
ǫ(ℓ, δ) is nonincreasing in δ > 0.
′′
To support the claim, assume that δ ′ ≥ δ ′′ > 0 are such that Dδℓ contains blue data, and let us
show that ǫ′ := b
ǫ(ℓ, δ ′ ) ≤ ǫ′′ := b
ǫ(ℓ, δ ′′ ). Indeed, recall that ǫ′ and ǫ′′ are k · k2,2 -norms of respective
symmetric entry-wise nonnegative matrices E ′ , E ′′ (see Proposition 3.5). When increasing δ from
δ = δ ′′ to δ = δ ′ , we reduce the associated E-matrix to its submatrix5 and further reduce the
entries in this submatrix6 , and thus reduce the norm of the matrix.
Aggregation procedure
we propose is as follows:
Given tolerance ǫ ∈ (0, 12 ), for every ℓ = 1, ..., L we specify δℓ > 0, the smaller the better,
ℓ
in such a way that b
ǫ(ℓ, δℓ ) ≤ ǫ/L 7 . Given observation ω K , we run the procedures Aℓδ
K ,
ℓδℓ
1 ≤ ℓ ≤ L. Whenever AK returns a color, we assign it to the index ℓ and to the vector
gℓ , so that after all ℓ’s are processed, some gℓ ’s get color “red,” some get color “blue”,
and some do not get color. The aggregation procedure returns, as a solution ĝ(ω K ), a
(whatever) red vector if one was discovered, and returns, say, g1 otherwise.
Proposition 4.1 In the situation and under the assumptions described in the beginning of section 4,
let ω K = (ω1 , ..., ωK ) be K-element i.i.d. sample drawn from a probability distribution P̄ which, taken
along with some ī ≤ I and µ̄ ∈ Mī , satisfies (29). Then the P̄ -probability of the event
kGµ̄ − ĝ(ω K )k2 ≤ min kGµ̄ − gℓ k2 + 2 max δℓ
ℓ≤L
ℓ
(33)
is at least 1 − ǫ.
Simple illustration. Let I = 1, H = Ω = Rd , and let M = M1 be a nonempty convex compact
subset of Rd . Further, suppose that Φ(h; µ) := Φ1 (h; µ) = hT µ + 21 hT Θh : H × M → R, where Θ
is a given positive definite matrix. We are also given a K-element i.i.d. sample ω K drawn from a
sub-Gaussian distribution P with sub-Gaussianity parameters (µ, Θ). Let also Gµ ≡ µ, so that the
aggregation problem we are interested in reads: “given L estimates g1 , ..., gL of the expectation µ of a
sub-Gaussian random vector ω with sub-Gaussianity parameters (µ, Θ), with a known Θ ≻ 0, and Krepeated i.i.d. sample ω K from the distribution of the vector, we want to select gℓ which is k · k2 -closest
to the true expectation µ of ω.” From now on we assume that g1 , ..., gL ∈ M (otherwise, projecting
the estimates onto M, we could provably improve their quality) and that g1 , ..., gL are distinct from
each other.
5
iδ
iδ
′′
The sets Wℓℓ
can become empty
′ shrink as δ grows, thus some of the blue sets Wℓℓ′ which are nonempty at δ = δ
when δ increases to δ ′ .
6
Indeed, by Proposition 3.5, these entries are obtained from saddle point values of some convex-concave functions by
a monotone transformation; it is immediately seen that as δ grows, these functions and the domains of the minimization
argument remain intact, while the domains of the maximization argument shrink, so that the saddle point values cannot
increase.
7
For instance, we could start with δ = δ 0 large enough to ensure that b
ǫ(ℓ, δ 0 ) = 0, and select δℓ as either the last term
i
i
in the progression δ = κ δ0 , i = 0, 1, ..., for some κ ∈ (0, 1), such that b
ǫ(ℓ, δ i ) ≤ ǫ/L, or the first term in this progression
which is “negligibly small” (say, less that 10−6 ), depending on what happens first.
20
In our situation, the sets (30), (31), (32) and functions Φi become
Wℓ = {µ ∈ M : uTℓℓ′ µ ≤ vℓℓ′ , ∀(ℓ′ ≤ L : ℓ′ 6= ℓ)},
g ′ −g
uℓℓ′ = kg′ℓ−gℓℓk2 , vℓℓ′ = 21 [gℓ + gℓ′ ], ℓ 6= ℓ′ ;
ℓ
Wℓℓδ ′ = {µ ∈ M : uTℓℓ′ µ ≥ vℓℓ′ + δ}, 1 ≤ ℓ, ℓ′ ≤ L, ℓ 6= ℓ′ ,
Φ(h; µ) = hT µ + 21 hT Θh,
(we are in the case of I = 1 and thus suppress index i in the notation for W ’s and Φ). Note that
d
Individual Inference Procedure Aℓδ
K deals with exactly one red hypothesis, HSi [R , Wℓ , Φ], and at most
L − 1 blue hypotheses HSi [Rd , Wℓℓδ ′ , Φ] associated with nonempty sets Wℓℓδ ′ and ℓ′ 6= ℓ.
Applying the construction from section 3.3.1, we arrive at the aggregation routine as follows (below
′
ℓ, ℓ vary in {1, ..., L}):
• We set
=
=
√
L L−1
uTℓℓ′ Θuℓℓ′ ;
ǫK
ℓ 6=ℓ
1
′
ℓ′ ;
2 [gℓ + gℓ + δ
i
hℓ uℓℓ′ ] , ℓ 6=
P
K
δℓ
T
′ −
ω
Kw
u
′
t
ℓℓ
T
t=1
2u ′ Θuℓℓ′ ℓℓ
δℓ = max
′
wℓℓ′
(K) K
ψℓℓ′ (ω )
r
ln
ℓℓ
(34)
+ 12 ln(L − 1), ℓ 6= ℓ′ .
(K)
• Given ω K , for 1 ≤ ℓ ≤ L we assign vector gℓ color “red,” if ψℓℓ′ (ω K ) > 0 for all ℓ′ 6= ℓ, otherwise
we do not assign gℓ any color;
• If red vectors were found, we output (any) one of them as solution to the aggregation problem;
if no red vectors are found, we output, say, g1 as solution.
Proposition 4.1 as applied to the situation in question states that whenever ω1 , ω2 , ...ωK are drawn,
independently from each other, from a sub-Gaussian distribution P with parameters µ, Θ, then with
P -probability at least 1 − ǫ the result ĝ(ω K ) of the above aggregation routine satisfies the relation
kµ − ĝ(ω K )k2 ≤ min kµ − gℓ k2 + 2 max δℓ ,
1≤ℓ≤L
ℓ
which essentially recovers the classical ℓ2 oracle inequality (cf. [12, Theorem 4]).
5
5.1
Beyond the scope of affine detectors
Lifted detectors
The tests developed in sections 3.2.2 and 3.3 were based on affine detectors – affine functions φ(ω)
associated with pairs of composite hypotheses H1 : P ∈ P1 , H2 : P ∈ P2 on the probability distribution
P of observation ω ∈ Ω = Rd . Such detectors were built to satisfy the relations
R
R
(35)
Ω exp{−φ∗ (ω)}P (dω) ≤ ǫ⋆ ∀P ∈ P1 & Ω exp{φ∗ (ω)}P (dω) ≤ ǫ⋆ ∀P ∈ P2 ,
with as small ǫ⋆ as possible (cf. (10)), and affinity of φ is of absolutely no importance here: all
constructions in sections 3.2.2, 3.3 were based upon availability of pairwise detectors φ, affine or not,
satisfying, for the respective pairs of composite hypotheses, relations (10) with some known ǫ⋆ . So far,
affinity of detectors was utilized only when building detectors satisfying (35) via the generic scheme
presented in section 3.1.
21
Now note that given a random observation ζ taking values in some Rd along with a deterministic
function Z(ζ) : Rd → RD , we can convert an observation ζ into an observation
ω = (ζ, Z(ζ)) ∈ Rd × RD .
Here ω is a deterministic transformation of ζ which “remembers” ζ, so that to make statistical inferences from observations ζ is exactly the same as to make them from observations ω. However,
detectors which are affine in ω can be nonlinear in ζ: for instance, for Z(ζ) = ζζ T , affine in ω detectors are exactly detectors quadratic in ζ. We see that within the framework of our approach, passing
from ζ to ω allows to consider a wider family of detectors and thus to arrive at a wider family of tests.
The potential bottleneck here is the necessity to bring the “augmented” observations (ζ, Z(ζ)) into
the scope of our setup.
Example: distributions with bounded support. Consider the case where the distribution P
of our observation ζ ∈ Rd belongs to a family P of Borel probability distributions supported on
a given bounded set, for the sake of simplicity – on the unit Euclidean ball B of Rd . Given a
continuous function Z(·) : Rd → RD , our goal is to cover the family P + of distributions P + [P ] of
ω = (ζ, Z(ζ)) induced by distributions P ∈ P of ζ by a family P[H, M, Φ] (or S[H, M, Φ]) associated
with some regular data, thus making the machinery we have developed so far applicable to the family
of distribution P + . Assuming w.l.o.g. that
kZ(z)k2 ≤ 1 ∀(z : kzk2 ≤ 1),
observe that for P ∈ P the distribution P + [P ] is sub-Gaussian with sub-Gaussianity parameters
(θP , Θ = 2Id+D ), where
Z
θP = (ζ, Z(ζ))P (dζ).
B
It follows that
If we can point out a convex compact set U in Rd × RD such that
θP ∈ U ∀P ∈ P,
(36)
then, specifying regular data as
H = Ω := Rd × RD ,
M = U,
Φ(h; µ) = hT µ + hT h : H × M → R,
we ensure that the family S[H, M, Φ] contains the family of probability distributions P + =
{P + [P ] : P ∈ P} on Ω.
How useful is this (by itself pretty crude) observation in the context of our approach depends on how
much information on P can be “captured” by a properly selected convex compact set U satisfying
(36). We are about to consider in more details “quadratic lifting” – the case where Z(ζ) = ζζ T .
5.1.1
Quadratic lifting, Gaussian case
Consider the situation where we are given
• a nonempty bounded set U in Rm ;
22
• a nonempty convex compact subset U of the positive semidefinite cone Sd+ ;
• a matrix Θ∗ ≻ 0 such that Θ∗ Θ for all Θ ∈ U ;
• an affine mapping u 7→ A(u) = A[u; 1] : Rm → Ω = Rd , where A is a given d × (m + 1) matrix.
Now, a pair (u ∈ U, Θ ∈ U ) specifies Gaussian random vector ζ ∼ N (A(u), Θ) and thus specifies a
Borel probability distribution P [u, Θ] of (ζ, ζζ T ). Let Q(U, U ) be the family of probability distributions
on Ω = Rd × Sd stemming in this fashion from Gaussian distributions with parameters from U × U .
Our goal is to cover the family Q(U, U ) by a family of the type S[H, M, Φ], which, as it was already
explained, would allow to use the machinery developed so far in order to decide on pairs of composite
Gaussian hypotheses
H1 : ζ ∼ N (u, Θ) with (u, Θ) ∈ U1 × U1
H2 : ζ ∼ N (u, Θ) with (u, Θ) ∈ U2 × U2
via tests based on detectors which are quadratic in ζ.
It is convenient to represent a linear form on Ω = Rd × Sd as
hT z + 12 Tr(HZ),
where (h, H) ∈ Rd ×Sd is the “vector of coefficients” of the form, and (z, Z) ∈ Rd ×Sd is the argument
of the form.
We denote by b = [0; 0; ...; 0; 1] ∈ Rm+1 the last basic orth of Rm+1 . We assume that for some
δ ≥ 0 it holds
−1/2
− Ik ≤ δ ∀Θ ∈ U ,
(37)
kΘ1/2 Θ∗
−1/2
−1/2
I, whence
where k · k is the spectral norm. Observe that for every Θ ∈ U we have 0 Θ∗ ΘΘ∗
−1/2
1/2
kΘ Θ∗ k ≤ 1, that is, (37)
satisfied with δ = 2. Thus, we can assume w.l.o.g. that
is always
A
δ ∈ [0, 2]. Finally, we set B =
.
bT
A desired “covering” of Q(U, U ) is given by the following
Proposition 5.1 In the notation and under the assumptions of this section, let γ ∈ (0, 1) and a
: Wm+1,m+1 = 1} be given, and let Z
convex compact computationally tractable set Z ⊂ {W ∈ Sm+1
+
contain all matrices Z(u) := [u; 1][u; 1]T with u ∈ U . Denoting by φZ (·) the support function of Z:
φZ (W ) = max Tr(ZW ) : Sm+1 → R,
Z∈Z
let us set
−1
H = Hγ := {(h, H) ∈ Rd × Sd : −γΘ−1
∗ H γΘ∗ }
M = U,
1/2
1/2
1/2
1/2
δ(2−δ)
kΘ∗ HΘ∗ k2F
Φ(h, H; Θ) = − 21 ln Det(I − Θ∗ HΘ∗ ) + 12 Tr([Θ − Θ∗ ]H) + 2(1−γ)
+ΓZ (h, H) : H × U → R,
[k · kF is the Frobenius norm]
T hbT + AT HA + B T [H, h]T [Θ−1 − H]−1 [H, h]B
+
A
ΓZ (h, H) = 21 φZ bhT A
∗
=
1
2 φZ
BT
h
H
hT
(38)
−1 [H, h] B .
+ [H, h]T [Θ−1
∗ − H]
Then H, M, Φ form a regular data, and
Q(U, U ) ⊂ S[H, M, Π].
(39)
Besides this, function Φ(h, H; Θ) is coercive in the convex argument: whenever (hi , Hi ) ∈ H and
k(hi , Hi )k → ∞ as i → ∞, we have Φ(hi , Hi ; Θ) → ∞, i → ∞, for every Θ ∈ U .
23
For proof, see Appendix A.3. Note that
constraint uT Qu + 2q T u + r ≤ 0 which is valid
everyquadratic
Q q
Z ≤ 0 which is valid for all matrices Z(u), u ∈ U , and
on U induces a linear constraint Tr
T
q
r
thus can be incorporated into the description of Z.
Special case. In the situation of Proposition 5.1, let u vary in a convex compact set U . In this case,
the simplest way to define Z such that Z(u) ∈ Z for all u ∈ U is to set
V
u
: W 0, u ∈ U .
Z= W =
T
u
1
Let us compute the function Φ(h, 0; Θ). Setting A = [Ā, a], where a is the last column of A, direct
computation yields
Φ(h, 0; Θ)
=
=
ĀT h
h Ā 2a h + hT Θ∗ h
1
max hT Au + hT Θ∗ h .
u∈U
2
1
φZ
2
T
T
=
V
1
max 2uT ĀT h + 2aT h + hT Θ∗ h :
uT
2 V,u
u
1
0, u ∈ U
Now imagine that we are given two collections (Aχ , Uχ , Uχ ), χ = 1, 2, of the (A, U, U )-data, with the
same number of rows in A1 and A2 and have associated with Uχ -upper bounds Θ∗,χ on matrices
Θ ∈ Uχ . We want to use Proposition 5.1 to build an affine detector capable to decide on the hypotheses
H1 and H2 on the distribution of observation ω, with Hχ stating that this observation is N (Aχ [u; 1], Θ)
with some u ∈ Uχ and Θ ∈ U χ. To this end, we have to solve the convex-concave saddle point problem
SV = min max
h
Θ1 ∈U1 ,
Θ2 ∈U2
1
[Φ1 (−h, 0; Θ1 ) + Φ2 (h, 0; Θ2 )] ,
2
where Φ1 , Φ2 are the functions associated, as explained in Proposition 5.1, with the first, respectively,
with the second collection of the (A, U, U )-data. In view of the above computation, this boils down to
the necessity to solve the convex minimization problem
1 1 T
1
SV = min max
.
h Θ∗,1 h + hT Θ∗,2 h + hT [A2 [u2 ; 1] − A1 [u1 ; 1]]
h u1 ∈U1 , 2 2
2
u ∈U
2
2
An optimal solution h∗ to this problem induces the affine detector
1 1 T
T
T
T
h [Θ∗,1 − Θ∗,2 ]h∗ + max [−h∗ A1 [u1 ; 1]] − max h∗ A2 [u2 ; 1] ,
φ∗ (ω) = h∗ ω + a, a =
u1 ∈U1
u2 ∈U2
2 2 ∗
and the risk of this detector on the pair of families G1 , G2 of Gaussian distributions in question is
exp{SV}.
On the other hand, we could build affine detector for the families G1 , G2 by the machinery from
section 3.2.3, that is, by solving convex-concave saddle point problem
1
1 T
1 T
T
T
SV = min
max
−h A1 [u1 ; 1] + h A2 [u2 , 1] + h Θ1 h + h Θ2 h ;
h u1 ∈U1 ,Θ1 ∈U1 , 2
2
2
u ∈U ,Θ ∈U
2
2
2
2
the risk of the resulting affine detector on G1 , G2 is exp{SV}. Now assume that
(!) Uχ , χ = 1, 2, have -maximal elements, and these elements are selected as Θ∗,χ .
24
ρ
0.5
0.5
0.01
σ1
2
1
1
Table
unrestricted
σ2
H =0 h=0
H and h
2
0.31
0.31
1.00
4
0.24
0.39
0.62
4
0.41
1.00
0.41
1: Risk of quadratic detector
In this case the above computation says that SV and SV are the minimal values of identically equal
to each other functions and thus are equal to each other. Thus, in the case of (!) the machinery of
Proposition 5.1 produces a quadratic detector which can be only better, in terms of risk, than the
affine detector yielded by Proposition 3.3.8
Numerical illustration. To get an impression of the performance of quadratic detectors as compared to affine ones in the case of (!), we present here the results of experiment where U1 = U1ρ = {u ∈
R12 : ui ≥ ρ, 1 ≤ i ≤ 12}, U2 = U2ρ = −U1ρ , A1 = A2 ∈ R8×12 , and Uχ = {Θ∗,χ = σχ2 I8 } are singletons.
The risks of affine, quadratic and “purely quadratic” (with h set to 0) detectors on the pair G1 , G2 of
families of Gaussian distributions, with Gχ = {N (θ, Θ∗χ ) : θ ∈ Aχ Uχρ }, ar given in Table 1.
We see that
• when deciding on families of Gaussian distributions with common covariance matrix and expectations varying in associated with the families convex sets, passing from affine detectors described
by Proposition 3.3 to quadratic detectors, does not affect the risk (first row in the table). This
is a general fact: by the results of [13], in the situation in question affine detectors are optimal
in terms of risk among all possible detectors.
• when deciding on families of Gaussian distributions in the case where distributions from different
families can have close expectations (third row in the table), affine detectors are useless, while
the quadratic ones are not, provided that Θ∗,1 differs from Θ∗,2 . This is how it should be – we
are in the case where the first moments of the distribution of the observation bear no definitive
information on the family this distribution belongs to, which makes affine detectors useless. In
contrast, quadratic detectors are able to utilize information (valuable when Θ∗,1 6= Θ∗,2 ) “stored”
in the second moments of the observation.
• “in general” (second row in the table), both affine and purely quadratic components in a
quadratic detector are useful; suppressing one of them can increase significantly the attainable
risk.
5.1.2
Quadratic lifting: Bounded observations
It is convenient to represent a “quadratically lifted observation” (ζ, ζζ T ) by the matrix
T
ζ
ζζ
Z(ζ) =
∈ Sd+1 .
ζT 1
8
This seems to be tautology – there are more quadratic detectors than affine ones, so that the best risk achievable
with quadratic detectors can be only smaller than the best risk achievable with affine detectors. The point, however,
is that Proposition 5.1 does not guarantee building the best, in terms of its risk, quadratic detector, it deals with
“computationally tractable approximation” of this problem. As a result, the quadratic detector constructed in the latter
proposition can, in principle, be worse than the affine detector yielded by Proposition 3.3.
25
Assume that all distributions from P are supported on the solution set X of a system of quadratic
constraints
fℓ (ζ) := ζ T Aℓ ζ + 2aTℓ ζ + αℓ ≤ 0, 1 ≤ ℓ ≤ L,
P
where Aℓ ∈ Sd are such that ℓ λ̄ℓ Aℓ ≻ 0 for properly selected λ̄ℓ ≥ 0; as a consequence, X is bounded
P
(since f¯(ζ) := L
ℓ=1 λ̄ℓ fℓ (ζ) is a strongly convex quadratic form which is ≤ 0 on X). Setting
Aℓ aℓ
, Qℓ =
, 1 ≤ ℓ ≤ L,
Q0 =
1
aTℓ αℓ
observe that the distribution P + of Z(ζ) induced by a distribution P ∈ P is supported on the closed
convex set
X + = {Z ∈ Sd+1 : Z 0, Tr(Q0 Z) = 1, Tr(Qℓ Z) ≤ 0, ℓ = 1, ..., L}
which is bounded9 . The support function of this set is
Z 0, Zd+1,d+1 = 1
φX + (H) = max Tr(HZ) = max Tr(HZ) :
Tr(Qℓ Z) ≤ 0, 1 ≤ ℓ ≤ L
Z
Z∈X +
Recalling Example 4 in section 2.2 and section 2.3.5, we arrive at the regular data
Tr((h − g)µ) + 18 [φX+ (h − g) + φX+ (g − h)]2 + φX+ (g)
H = Sd+1 , M = X + , Φ(h, µ) = inf
g∈Sd+1
such that
∀P ∈ P : P + ∈ S[Sd+1 , {e[P + ]}, Φ(·, e[P + ])],
where e[P + ] is the expectation of P + , and therefore
P + = {P + : P ∈ P} ⊂ S[Sd+1 , M, Φ].
References
[1] L. Birgé. Approximation dans les spaces métriques et théorie de l’estimation: inégalités de
Cràmer-Chernoff et théorie asymptotique des tests. PhD thesis, Université Paris VII, 1980.
[2] L. Birgé. Vitesses maximales de décroissance des erreurs et tests optimaux associés. Zeitschrift
für Wahrscheinlichkeitstheorie und verwandte Gebiete, 55(3):261–273, 1981.
[3] L. Birgé. Sur un théorème de minimax et son application aux tests. Probab. Math. Stat., 3:259–
282, 1982.
[4] L. Birgé. Approximation dans les espaces métriques et théorie de l’estimation. Zeitschrift für
Wahrscheinlichkeitstheorie und verwandte Gebiete, 65(2):181–237, 1983.
[5] L. Birgé. Model selection via testing: an alternative to (penalized) maximum likelihood estimators. In Annales de l’Institut Henri Poincare (B) Probability and Statistics, volume 42, pages
273–325. Elsevier, 2006.
[6] L. Birgé. Robust tests for model selection. In M. Banerjee, F. Bunea, J. Huang, V. Koltchinskii, ,
and M. Maathuis, editors, From Probability to Statistics and Back: High-Dimensional Models and
Processes – A Festschrift in Honor of Jon A. Wellner, pages 47–64. Institute of Mathematical
Statistics, 2013.
P
+
indeed, for large λ0 > 0, the matrix Q = λ0 Q0 + L
is contained
ℓ=1 λ̄ℓ Qℓ is positive definite, and we conclude that X
d+1
in the bounded set {Z ∈ S
: Z 0, Tr(QZ) ≤ λ0 }.
9
26
[7] M. Burnashev. On the minimax detection of an imperfectly known signal in a white noise background. Theory Probab. Appl., 24:107–119, 1979.
[8] M. Burnashev. Discrimination of hypotheses for gaussian measures and a geometric characterization of the gaussian distribution. Math. Notes, 32:757–761, 1982.
[9] H. Chernoff. A measure of asymptotic efficiency for tests of a hypothesis based on the sum of
observations. The Annals of Mathematical Statistics, pages 493–507, 1952.
[10] D. Donoho and R. Liu. Geometrizing rate of convergence I. Technical report, Tech. Report 137a,
Dept. of Statist., University of California, Berkeley, 1987.
[11] D. L. Donoho and R. C. Liu. Geometrizing rates of convergence, II. The Annals of Statistics,
pages 633–667, 1991.
[12] A. Goldenshluger. A universal procedure for aggregating estimators. The Annals of Statistics,
pages 542–568, 2009.
[13] A. Goldenshluger, A. Juditsky, and A. Nemirovski. Hypothesis testing by convex optimization.
Electronic Journal of Statistics, 9(2):1645–1712, 2015.
[14] J.-B. Hiriart-Urruty and C. Lemarechal. Convex analysis and minimization algorithms i: Fundamentals (grundlehren der mathematischen wissenschaften). 1993.
[15] Y. Ingster and I. A. Suslina. Nonparametric goodness-of-fit testing under Gaussian models, volume
169 of Lecture Notes in Statistics. Springer, 2002.
[16] C. Kraft. Some conditions for consistency and uniform consistency of statistical procedures. Univ.
of California Publ. Statist., 2:493–507, 1955.
[17] L. Le Cam. Convergence of estimates under dimensionality restrictions. The Annals of Statistics,
pages 38–53, 1973.
A
Proofs
A.1
Proof of Proposition 2.1
All we need is to verify (5) and to check that the right hand side function in this relation is convex.
The latter is evident, since φX (h) + φX (−h) ≥ 2φX (0) = 0 and φX (h) + φX (−h) is convex. To verify
(5), let us fix P ∈ P[X] and h ∈ Rd and set
ν = hT e[P ],
so that ν is the expectation of hT ω with ω ∼ P . Note that −φX (−h) ≤ ν ≤ φX (h), so that (5)
definitely holds true when φX (h) + φX (−h) = 0. Now let
η :=
1
2
[φX (h) + φX (−h)] > 0,
and let
a=
1
2
[φX (h) − φX (−h)] , β = (ν − a)/η.
Denoting by Ph the distribution of hT ω induced by the distribution P of ω and noting that this
distribution is supported on [−φX (−h), φX (h)] = [a − η, a + η] and has expectation ν, we get
β ∈ [−1, 1]
27
and
γ :=
for all λ ∈ R. Hence,
Z
T
exp{h ω}P (dω) =
Z
a+η
a−η
[es − λ(s − ν)]Ph (ds)
t
ln(γ) ≤ inf ln
max [e − λ(s − ν)] = a + inf ln max [e − ρ(t − [ν − a])]
ρ
a−η≤s≤a+η
−η≤t≤η
λ
t
t
= a + inf ln max [e − ρ(t − ηβ)] ≤ a + ln max [e − ρ̄(t − ηβ)
ρ
s
−η≤t≤η
−η≤t≤η
with ρ̄ = (2η)−1 (eη − e−η ). The function g(t) = et − ρ̄(t − ηβ) is convex on [−η, η], and
g(−η) = g(η) = cosh(η) + β sinh(η),
which combines with the above computation to yield the relation
ln(γ) ≤ a + ln(cosh(η) + β sinh(η)),
(40)
∀(η > 0, β ∈ [−1, 1]) : βη + 12 η 2 − ln(cosh(η) + β sinh(η)) ≥ 0.
(41)
and all we need to verify is that
Indeed, if (41) holds true (40) implies that
ln(γ) ≤ a + βη + 21 η 2 = ν + 21 η 2 ,
which, recalling what γ, ν and η are, is exactly what we want to prove.
Verification of (41) is as follows. The left hand side in (41) is convex in β for β > − cosh(η)
sinh(η)
containing, due to η > 0, the range of β in (41). Furthermore, the minimum of the left hand side of
cosh(η)
(41) over β > − coth(η) is attained when β = sinh(η)−η
and is equal to
η sinh(η)
r(η) = 21 η 2 + 1 − η coth(η) − ln(sinh(η)/η).
All we need to prove is that the latter quantity is nonnegative whenever η > 0. We have
r ′ (η) = η − coth(η) − η(1 − coth2 (η)) − coth(η) + η −1 = (η coth(η) − 1)2 η −1 ≥ 0,
and since r(+0) = 0, we get r(η) ≥ 0 when η > 0.
A.2
Proof of Proposition 4.1
ℓ
Let Dℓ denote the collection of regular data processed by Aℓδ
K . Let, further, the K-repeated random
K
observation ω , the probability distribution P̄ , index ī ≤ I and vector µ̄ ∈ Mī be as in the premise of
the proposition, let ḡ = Gµ̄, and let ℓ∗ ≤ L be such that kḡ − gℓ∗ k2 ≤ kḡ − gℓ k2 , ℓ = 1, ..., L. Finally,
let E be the event “for every ℓ = 1, ..., L such that P̄ obeys one or more of the hypotheses HSi [D],
ℓ
D ∈ Dℓ , processed by procedure Aℓδ
K , this procedure correctly recovers the color of these hypotheses.”
By construction and due to the union bound, the P̄ K -probability of E is at least 1 − ǫ. It follows that
all we need to verify the claim of the proposition is to show that when ω K ∈ E, relation (33) takes
place. Thus, let us fix ω K ∈ E.
ℓ δ
Observe, first, that ḡ ∈ Vℓ∗ , whence µ̄ ∈ Wℓī∗ . Thus, when running AK∗ ℓ∗ , P̄ K obeys a red one
ℓ δℓ∗
among the hypotheses HSi [D], D ∈ Dℓ∗ processed by AK∗
28
, and since we are in the case of ω K ∈ E,
gℓ∗ gets a color, namely, color “red.” By construction of our aggregation procedure, its output can be
either gℓ∗ – and in this case (33) clearly holds true, or another vector, let it be denoted gℓ+ (ℓ+ 6= ℓ∗ ),
which was also assigned red color. We claim that the vector ḡ = Gµ̄ satisfies the relation
uTℓ+ ℓ∗ ḡ < vℓ+ ℓ∗ + δℓ+ .
(42)
īδℓ
Indeed, otherwise we have µ̄ ∈ Wℓ+ ℓ+∗ , meaning that P̄ K obeys a hypothesis HSi [D] processed when
ℓ+ δℓ+
running AK
(i.e., with D ∈ Dℓ+ ), and this hypothesis is blue. Since we are in the case of E, this
ℓ+ δℓ
implies that the color inferred by AK + is “blue,” which is a desired contradiction.
Now we are nearly done: indeed, gℓ∗ is the k · k2 closest to ḡ point among g1 , ..., gL , implying that
uTℓ+ ℓ∗ ḡ ≥ vℓ+ ℓ∗ .
(43)
Recalling what uℓ∗ ℓ+ and vℓ∗ ℓ+ are, the relations (42) and (43) tell us the following story about the
points ḡ, g∗ := gℓ∗ , g+ := gℓ+ and the hyperplane H = {g ∈ Rm : uℓ∗ ℓ+ g = vℓ∗ ℓ+ }: g∗ and g+ are
symmetric to each other w.r.t. H, and ḡ is at most at the distance δ := δℓ+ of H. An immediate
observation is that in this case
kg+ − ḡk2 ≤ kg∗ − ḡk2 + 2δ,
(44)
and we arrive at (33).
To justify (44) note that by shift and rotation we can reduce the situation to the one when
g∗ , g+ , ḡ belong to the linear span of the first two basic orhts with the first two coordinates of these
three vectors being, respectively, [−r; 0] (for g∗ ), [r; 0] (for g+ ) and [d; h] (for ḡ), with |d| ≤ δ. Hence
kg+ − ḡk2 − kg∗ − ḡk2 =
[(r−d)2 +h2 ]−[(r+d)2 +h2 ]
kg+ −ḡk2 +kg∗ −ḡk2
=
−4rd
kg+ −ḡk2 +kg∗ −ḡk2
≤
4rδ
kg+ −g∗ k2
= 2δ,
as claimed.
A.3
10 .
Proof of Proposition 5.1
For any u ∈ Rm , h ∈ Rd , Θ ∈ Sd+ and H ∈ Sd such that −I ≺ Θ1/2 HΘ1/2 ≺ I we have
Ψ(h, H; u, Θ) := ln Eζ∼N (A(u),Θ) exp{hT ζ + 12 ζ T Hζ}
o
n
= ln Eξ∼N (0,I) exp{hT [A(u) + Θ1/2 ξ] + 21 [A(u) + Θ1/2 ξ]T H[A(u) + Θ1/2 ξ]
= − 21 ln Det(I − Θ1/2 HΘ1/2 )
+hT A(u) + 21 A(u)T HA(u) + 21 [HA(u) + h]T Θ1/2 [I − Θ1/2 HΘ1/2 ]−1 Θ1/2 [HA(u) + h]
= − 21 ln Det(I − Θ1/2 HΘ1/2 ) + 12 [u; 1]T bhT A + AT hbT + AT HA [u; 1]
h
i
(45)
+ 21 [u; 1]T B T [H, h]T Θ1/2 [I − Θ1/2 HΘ1/2 ]−1 Θ1/2 [H, h]B [u; 1]
(because hT A(u) = [u; 1]T bhT A[u; 1] = [u; 1]T AT hbT [u; 1] and HA(u) + h = [H, h]B[u; 1]).
−1
Observe that when (h, H) ∈ Hγ , we have Θ1/2 [I−Θ1/2 HΘ1/2 ]−1 Θ1/2 = [Θ−1 −H]−1 [Θ−1
∗ −H] ,
so that (45) implies that for all u ∈ Rm , Θ ∈ U , and (h, H) ∈ Hγ ,
Ψ(h, H; u, Θ) ≤ − 21 ln Det(I − Θ1/2 HΘ1/2 )
−1
+ 12 [u; 1]T bhT A + AT hbT + AT HA + B T [H, h]T [Θ−1
∗ − H] [H, h]B [u; 1]
{z
}
|
Q[H,h]
=
− 12
ln Det(I − Θ
1/2
HΘ
1/2
)+
1
2 Tr(Q[H, h]Z(u))
≤ − 12 ln Det(I − Θ1/2 HΘ1/2 ) + ΓZ (H, h)
29
(46)
(we have taken into account that Z(u) ∈ Z when u ∈ U (premise of the proposition) and therefore
Tr(Q[H, h]Z(u)) ≤ φZ (Q[H, h])).
20 .
Now let us upper-bound the function
G(h, H; Θ) = − 21 ln Det(I − Θ1/2 HΘ1/2 )
on the domain (h, H) ∈ Hγ , Θ ∈ U . For (h, H) ∈ Hγ and Θ ∈ U fixed we have
−1/2
kΘ1/2 HΘ1/2 k = k[Θ1/2 Θ∗
≤ kΘ
1/2
−1/2 T
1/2
1/2
][Θ∗ HΘ∗ ][Θ1/2 Θ∗
1/2
1/2
−1/2
Θ∗ k2 kΘ∗ HΘ∗ k
≤
] k
1/2
1/2
kΘ∗ HΘ∗ k
=: d(H) ≤ γ
(47)
−1/2
−1/2
(we have used the fact that 0 [Θ1/2 Θ∗ ]T [Θ1/2 Θ∗ ] I due to 0 Θ Θ∗ , whence
−1/2
kΘ1/2 Θ∗ k ≤ 1). Denoting by k · kF the Frobenius norm of a matrix and noting that kABkF ≤
kAkkBkF , computation completely similar to the one in (47) yields
1/2
1/2
kΘ1/2 HΘ1/2 kF ≤ kΘ∗ HΘ∗ kF =: D(H).
(48)
Besides this, setting F (X) = − ln Det(X) : int Sd+ → R and equipping Sd with the Frobenius inner
1/2
1/2
product, we have ∇F (X) = −X −1 , so that with R0 = Θ∗ HΘ∗ , R1 = Θ1/2 HΘ1/2 , and ∆ = R1 −R0 ,
we have for properly selected λ ∈ (0, 1) and Rλ = λR0 + (1 − λ)R1 :
F (I − R1 ) = F (I − R0 − ∆) = F (I − R0 ) + h∇F (I − Rλ ), −∆i = F (I − R0 ) + h(I − Rλ )−1 , ∆i
= F (I − R0 ) + hI, ∆i + h(I − Rλ )−1 − I, ∆i.
We conclude that
F (I − R1 ) ≤ F (I − R0 ) + Tr(∆) + kI − (I − Rλ )−1 kF k∆kF .
(49)
Denoting by µi the eigenvalues of Rλ and noting that kRλ k ≤ max[kR0 k, kR1 k] = d(H) ≤ γ (see
µi
1
= − 1−µ
of I − (I − Rλ )−1 satisfy
(47)), we have |µi | ≤ γ, and therefore eigenvalues νi = 1 − 1−µ
i
i
|νi | ≤ |µi |/(1 − µi ) ≤ |µi |/(1 − γ), whence
kI − (I − Rλ )−1 kF ≤ kRλ kF /(1 − γ).
Noting that kRλ kF ≤ max[kR0 kF , kR1 kF ] ≤ D(H), see (48), we conclude that kI − (I − Rλ )−1 kF ≤
D(H)/(1 − γ), so that (49) yields
F (I − R1 ) ≤ F (I − R0 ) + Tr(∆) + D(H)k∆kF /(1 − γ).
−1/2
Further, by (37) the matrix D = Θ1/2 Θ∗
(50)
− I satisfies kDk ≤ δ, whence
1/2
1/2
1/2
1/2
HΘ = (I + D)R0 (I + D T ) − R0 = DR0 + R0 D T + DR0 D T .
∆=Θ
HΘ
{z
}−Θ
|
| ∗ {z ∗ }
R1
R0
Consequently,
k∆kF ≤ kDR0 kF + kR0 D T kF + kDR0 D T kF ≤ [2kDk + kDk2 ]kR0 kF ≤ δ(2 + δ)kR0 kF = δ(2 + δ)D(H).
This combines with (50) and the relation
1/2
1/2
Tr(∆) = Tr(Θ1/2 HΘ1/2 − Θ∗ HΘ∗ ) = Tr([Θ − Θ∗ ]H)
30
to yield
F (I − R1 ) ≤ F (I − R0 ) + Tr([Θ − Θ∗ ]H) +
and we conclude that for all (h, H) ∈ Hγ and Θ ∈ U ,
δ(2 + δ) 1/2
1/2
kΘ∗ HΘ∗ k2F ,
1−γ
G(h, H; Θ) = 12 F (I − R1 )
δ(2 + δ) 1/2
1/2
1/2
1
b H; Θ) := − 1 ln Det(I − Θ1/2
≤ G(h,
kΘ∗ HΘ∗ k2F .(51)
∗ HΘ∗ ) + 2 Tr([Θ − Θ∗ ]H) +
2
2(1 − γ)
b H; Θ) clearly is convex-concave and continuous on H × M = Hγ × U .
Note that G(h,
30 .
Combining (51), (46), (38) and the origin of Ψ, see (45), we arrive at
∀((u, Θ) ∈ U × U , (h, H) ∈ Hγ = H) : ln Eζ∼N (u,Θ) exp{hT ζ + 12 ζ T Hζ} ≤ Φ(h, H; Θ).
This is all we need, up to verification of the claim that H, M, Φ is regular data, which boils down to
checking that Φ : H × M → R is convex-concave and continuous. The latter check, recalling that
b H; Θ) : H × M indeed is convex-concave and continuous, reduces to verifying that Γ(h, H) is
G(h,
convex and continuous on Hγ . Recalling that Z is nonempty compact set, the function φZ (·) : Sd+1
is continuous, implying the continuity of Γ(h, H) = 12 φZ (Q[H, h]) on H = Hγ (Q[H, h] is defined in
(46)). To prove convexity of Γ, note that Z is contained in Sm+1
, implying that φZ (·) is convex and
+
-monotone. On the other hand, by Schur Complement Lemma, we have
S :=
{(h, H, G) : G
(h, H) ∈ Hγ }
Q[H, h],
G − [bhT A + AT hB + AT HA] B T [H, h]T
0, (h, H) ∈ Hγ ,
=
(h, H, G) :
[H, h]B
Θ−1
∗ −H
implying that S is convex. Since ΦZ (·) is -monotone, we have
{(h, H, τ ) : (h, H) ∈ Hγ , τ ≥ Γ(h, H)} = {(h, H, τ ) : ∃G : G Q[H, h], 2τ ≥ φZ (G), (h, H) ∈ Hγ },
and we see that the epigraph of Γ is convex (since the set S and the epigraph of φZ are so), as claimed.
40 . It remains to prove that Φ is coercive in H, h. Let Θ ∈ U and (hi , Hi ) ∈ Hγ with k(hi , Hi )k → ∞
as i → ∞, and let us prove that Φ(hi , Hi ; Θ) → ∞. Looking at the expression for Φ(hi , Hi ; Θ), it is
immediately seen that all terms in this expression, except for the terms coming from φZ (·), remain
bounded as i grows, so that all we need to verify is that the φZ (·)-term goes to ∞ as i → ∞. Observe
that Hi are uniformly bounded due to (hi , Hi ) ∈ Hγ , implying that khi k2 → ∞ as i → ∞. Denoting
by e the last basic orth of Rd+1 and by b, as before, the last basic orth of Rm+1 , note that, by
construction, B T e = b. Now let W ∈ Z, so that Wm+1,m+1 = 1. Taking into account that the
−1 satisfy αI [Θ−1 − H ]−1 βI for some positive α, β due to H ∈ H ,
matrices [Θ−1
i
i
γ
d
d
∗
∗ − Hi ]
observe that
T
Hi hi
T
−1
−1
−1
+ [Hi , hi ] [Θ∗ − Hi ] [Hi , hi ] = hTi [Θ−1
∗ − Hi ] hi ee + Ri ,
T
hi
|
{z
}
|
{z
}
αi khi k2
2
Qi
where αi ≥ α > 0 and kRi kF ≤ C(1 + khi k2 ). As a result,
φZ (B T Qi B) ≥ Tr(W B T Qi B) = Tr(W B T [αi khi k22 eeT + Ri ]B)
= αi khi k22 Tr(W bbT ) −kBW B T kF kRi kF ≥ αkhi k22 − C(1 + khi k2 )kBW B T kF ,
| {z }
=Wm+1,m+1 =1
and the concluding quantity tends to ∞ as i → ∞ due to khi k2 → ∞, i → ∞.
31
| 10math.ST
|
arXiv:1303.2208v2 [math.AC] 23 Jun 2014
ON THE RELATIVE COHEN-MACAULAY MODULES
MAJID RAHRO ZARGAR
Abstract. Let R be a commutative Noetherian local ring and let a be a proper
ideal of R. In this paper, as a main result, it is shown that if M is a Gorenstein
R-module with c = ht M a, then Hia (M ) = 0 for all i 6= c is completely encoded in
homological properties of Hca (M ), in particular in its Bass numbers. Notice that,
this result provides a generalization of a result of Hellus and Schenzel which has
been proved before, as a main result, in the case where M = R.
1. introduction
Throughout this paper, R is a commutative Noetherian ring, a is a proper ideal
of R and M is an R-module. For a prime ideal p of R, the residue class field Rp /pRp
is denoted by k(p). For each non-negative integer i, let Hia (M ) denotes the i-th
local cohomology module of M with respect to a; see [1] for its definition and basic
results. A Gorenstein module over a local ring R is a maximal Cohen-Macaulay
module of finite injective dimension. This concept was introduced by Sharp in [7]
and studied extensively by him and other authors. In the present paper, we will
use the concept of relative Cohen-Macaulay modules which is a generalization of the
concept of Cohen-Macaulay modules. This kind of modules has been studied in [3]
under the title of cohomologically complete intersections and continued in [5]. In [3]
Hellus and Schenzel, as a main result, showed that if (R, m, k) is a Gorenstein local
ring and a is an ideal of R with ht R a = c and dim R/a = d such that R is relative
Cohen-Macaulay in V(a) \ {m}, then the following statements are equivalent.
(i) Hia (R) = 0 for all i 6= c, i.e R is relative Cohen-Macaulay with respect to a.
(ii) Hd (Hc (R)) ∼
= ER (k) and Hi (Hc (R)) = 0 for all i 6= d.
m
a
m
a
(iii) ExtdR (k, Hca (R)) ∼
= k and ExtiR (k, Hca (R)) = 0 for all i 6= d.
(iv) µi (m, Hca (R)) = δdi .
Moreover, if a satisfies the above conditions, it follows that R̂a ∼
= HomR (Hca (R), Hca (R))
and ExtiR (Hca (R), Hca (R)) = 0 for all i 6= 0, where R̂a denotes the a-adic completion
of R.
2010 Mathematics Subject Classification. Primary: 13D45; Secondary: 13C14.
Key words and phrases. Local cohomology, Gorenstein module, relative Cohen-Macaulay.
This research was in part supported by a grant from IPM (No. 92130024).
1
2
M. R. ZARGAR
As a main result, in Theorem 2.9, we generalize the above result for a Gorenstein
R-module M . Indeed, it is shown that if M is a Gorenstein R-module with c = ht M a
and d = dim M/aM such that M is relative Cohen-Macaulay in SuppR (M/aM ) \
{m}, then the following statements are equivalent.
(i) Hia (M ) = 0 for all i 6= c, i.e M is relative Cohen-Macaulay with respect to a.
(ii) Hdm (Hca (M )) ∼
= ER (k)r(M ) and Him (Hca (M )) = 0 for all i 6= d.
(iii) Extd (k, Hc (M )) ∼
= kr(M ) and Exti (k, Hc (M )) = 0 for all i 6= d.
R
R
a
a
(iv) µi (m, Hca (M )) = r(M )δdi .
Moreover, if a satisfies the above conditions, it follows that HomR (M, M ) ⊗R R̂a ∼
=
i
c
c
c
c
HomR (Ha (M ), Ha (M )) and ExtR (Ha (M ), Ha (M )) = 0 for all i 6= 0, where r(M )
denotes the type of M .
2. Main results
Definition 2.1. We say that a finitely generated R-module M is relative Cohen
Macaulay with respect to a if there is precisely one non-vanishing local cohomology
module of M with respect to a. Clearly this is the case if and only if grade (a, M ) =
cd(a, M ), where cd(a, M ) denotes the cohomological dimension of M with respect
to a, which is the largest integer i for which Hia (M ) 6= 0. Observe that the notion
of relative Cohen-Macaulay module is connected with the notion of cohomologically
complete intersection ideal which has been studied in [3].
Remark 2.2. Let M be a relative Cohen-Macaulay module with respect to a and
let cd(a, M ) = n. Then, in view of [1, Theorems 6.1.4, 4.2.1, 4.3.2], it is easy to
see that Supp Hna (M ) = Supp(M/aM ) and ht M a = grade (a, M ), where ht M a =
inf{ dimRp Mp | p ∈ Supp(M/aM ) }.
Definition 2.3. ( See [7, Theorem 3.11].) A non-zero finitely generated module M
over a local ring (R, m) is said to be a Gorenstein R-module if the following equalities
hold true
depth M = dim M = idR M = dim R.
Definition 2.4. For any prime ideal p of R, the i-th Bass number of M with respect
to p is defined by µi (p, M ) := vdim k(p) ExtiRp (k(p), Mp ). If (R, m) is local and M is
finitely generated of depth t, then the number µt (m, M ) is called type of M and is
denoted by r(M ).
Let M be a non-zero finitely generated module over a local ring (R, m) and let a
be an ideal of R. Let E.R (M ) be a minimal injective resolution for M . It is a wellL
i
known fact that E.R (M )i =
p∈Spec(R) µ (p, M )E R (R/p), where ER (R/p) denotes
the injective hull of R/p. Now, let c = grade (a, M ). Then, Γa (E.R (M )i ) = 0 for all
ON THE RELATIVE COHEN-MACAULAY MODULES
3
i < c. Therefore Hca (M ) = ker(Γa (E.R (M ))c −→ Γa (E.R (M ))c+1 ). This observation
provides an embedding 0 −→ Hca (M )[−c] −→ Γa (E.R (M )) of complexes of R-modules
where Hca (M )[−c] is considered as a complex concentrated in homological degree
zero.
Definition 2.5. ( See [4, Definition 4.1].) The cokernel of the embedding
0 −→ Hca (M )[−c] −→ Γa (E.R (M ))
. (a) and is called the truncation complex. Therefore, there is a
is denoted by CM
short exact sequence
.
(a) −→ 0
0 −→ Hca (M )[−c] −→ Γa (E.R (M )) −→ CM
. (a)) ∼ Hi (M ) for all i > c
of complexes of R-modules. We observe that Hi (CM
= a
i
.
while H (CM (a)) = 0 for all i ≤ c.
The following proposition is of assistance in the proof of the main result.
Proposition 2.6. Assume that M is a Gorenstein module over a local ring R.
Then, with the previous notation the following statements are true.
(i) There is an exact sequence
.
.
0 → Ext0R (CM
(a), M ) → HomR (M, M )⊗R R̂a → ExtcR (Hca (M ), M ) → Ext1R (CM
(a), M ) → 0.
i+1
c
.
∼
(ii) There are isomorphisms Exti+c
R (Ha (M ), M ) = ExtR (CM (a), M ) for all i >
0.
(iii) Suppose that M is relative Cohen-Macaulay with respect to a. Then
∼ HomR (M, M ) ⊗R R̂a and Exti+c (Hc (M ), M ) = 0 for
ExtcR (Hca (M ), M ) =
a
R
all i 6= 0.
Proof. We first notice that, since M is Gorenstein, R is Cohen-Macaulay and M̂ is
a Gorenstein R̂-module. Now, let ωR̂ be a canonical module of the ring R̂. Then, in
view of [2, Exercise 3.3.28], M̂ is isomorphic to a direct sum of finitely many copies
of ωR̂ . Hence, one can use [2, Theorem 3.3.10(c)], [2, Theorem 3.3.4(d)] and the fact
that Exti (M̂ , M̂ ) ∼
= Exti (M, M ) ⊗R R̂ for all i, to see that Exti (M, M ) = 0 for all
R̂
R
R
i > 0 and that HomR (M, M ) is a flat R-module.
(i): Let E . be a minimal injective resolution for M . Then, since E . is a bounded
complex of injective R-modules, by applying the functor HomR (−, E . ) on the exact
sequence of R-complexes in 2.5 we obtain the following short exact sequence of
complexes
.
0 −→ HomR (CM
(a), E . ) −→ HomR (Γa (E . ), E . ) −→ HomR (Hca (M ), E . )[c] −→ 0.
4
M. R. ZARGAR
Now, let x = x1 , . . . , xn be a generating set of the ideal a, and let Čx (R) be the
¯
¯
Čech complex of R with respect to x. Then, in view of [6, Theorem 1.1], there
¯
exists an isomorphism RΓa (M ) ≃ Čx (R) ⊗L
R M in the derived category. Now, since
¯
i
ExtR (M, M ) = 0 for all i > 0, we see that HomR (M, M ) ≃ R HomR (M, M ). Thus,
we get the following isomorphisms
R HomR (RΓa (M ), M ) ≃ R HomR ((Čx (R) ⊗L
R M ), M )
¯
≃ R HomR (Čx (R), R HomR (M, M ))
¯
≃ R HomR (Čx (R), HomR (M, M )),
¯
in the derived category. Now, since HomR (M, M ) is flat, one can use [6, Theorem 1.1] and the above isomorphisms to deduce that Hi (R HomR (RΓa (M ), M ) = 0
for all i 6= 0 and that H0 (HomR (Γa (E . ), E . )) ∼
= HomR (M, M ) ⊗R R̂a . On the
other hand, since Γa ((E . )i ) = 0 for all i < c = grade (a, M ) and HomR (N, X) =
HomR (N, Γa (X)) for any a-torsion R-module N and for all R-modules X, one can
deduce that ExtiR (Hca (M ), M ) = 0 for all i < c. Therefore, with the aid of the above
considerations, the induced long exact cohomology sequence of the above exact sequence of complexes provides the statements (i) and (ii) of the claim.
(iii): Assume that M is relative Cohen-Macaulay with respect to a. Then, one can
. (a) is exact; and so the complex Hom (C . (a), E . )
easily check that the complex CM
R
M
. (a), M ) = Hi (Hom (C . (a), E . )) = 0 for all i.
is also exact. Therefore, ExtiR (CM
R
M
Hence, the assertion follows from (i) and (ii).
The following lemma and definition are needed in the proof of the next result.
Lemma 2.7. ( See [3, Proposition 4.1].) Let n be a non-negative integer and let M
be an arbitrarily R-module over a local ring (R, m). Then the following conditions
are equivalent.
(i) Him (M ) = 0 for all i < n.
(ii) ExtiR (R/m, M ) = 0 for all i < n.
Definition 2.8. Let (R, m) be a local ring and let a be a proper ideal of R. Then, we
say that a non-zero R-module M is relative Cohen-Macaulay in SuppR (M/aM )\{m},
whenever Mp is relative Cohen-Macaulay with respect to aRp and cd(aRp , Mp ) = c
for all p ∈ SuppR (M/aM ) \ {m}, where c = ht M a.
The following theorem extends the main result [3, Theorem 0.1].
Theorem 2.9. Let a be an ideal of a local ring (R, m, k) and M a Gorenstein Rmodule. Suppose that M is relative Cohen-Macaulay in SuppR (M/aM ) \ {m}. Set
c := ht M a and d := dim M/aM . Then the following statements are equivalent.
(i) Hia (M ) = 0 for all i 6= c, i.e M is relative Cohen-Macaulay with respect to a.
ON THE RELATIVE COHEN-MACAULAY MODULES
5
(ii) Hdm (Hca (M )) ∼
= ER (k)r(M ) and Him (Hca (M )) = 0 for all i 6= d.
(iii) Extd (k, Hc (M )) ∼
= kr(M ) and Exti (k, Hc (M )) = 0 for all i 6= d.
(iv)
a
R
c
i
µ (m, Ha (M ))
R
a
= r(M )δdi .
Moreover, if one of the above statements holds, then ExtiR (Hca (M ), Hca (M )) = 0 for
all i 6= 0 and HomR (Hc (M ), Hc (M )) ∼
= HomR (M, M ) ⊗R R̂a , where R̂a denotes the
a
a
a-adic completion of R.
Proof. (i)⇒(ii): First, we can use [5, Proposition 2.8] and the assumption to see
that Him (Hca (M )) ∼
= Hi+c
m (M ) for all i ≥ 0. Now, since M is Cohen-Macaulay of
d
c
dimension n, Hm (Ha (M )) ∼
= Hnm (M ) and Him (Hca (M )) = 0 for all i 6= d. On the
other hand, in view of [5, Theorem 2.5] and the assumption, Hnm (M ) is an injective
R-module. Therefore, one can use [5, Corollay 2.2] to achieve the isomorphism
r(M )
Hn (M ) ∼
.
= ER (k)
m
(ii)⇒(iii): It follows from Lemma 2.7 that ExtiR (k, Hca (M )) = 0 for all i < d.
On the other hand, since Hdm (Hca (M )) is injective, one can use [5, Proposition
∼
2.1] to see that Exti (k, Hc (M )) = 0 for all d < i and that Extd (k, Hc (M )) =
R
R
a
a
HomR (k, Hdm (Hca (M ))) ∼
= kr(M ) , and hence the assertion is done. The implications
(iii)⇔(iv) is clear.
(iii)⇒(ii): First, in view of Lemma 2.7, Him (Hca (M )) = 0 for all i < d. Now,
since dimR Hca (M ) ≤ dimR M/aM = d, one can use the vanishing theorem to see
that Him (Hca (M )) = 0 for all d < i. Therefore, [5, Proposition 2.1] implies that
HomR (R/m, Hd (Hc (M ))) ∼
= kr(M ) , and so it is an Artinian R-module. Hence, by
m
a
[1, Theorem 7.1.2], Hdm (Hca (M )) is Artinian. Thus, one can use [5, Corollary 2.2] to
see that µ1 (m, Hdm (Hca (M ))) = µd+1 (m, Hca (M )) = 0 and µ0 (m, Hdm (Hca (M ))) = r(M ).
Hence Hdm (Hca (M )) is an injective R-module, and so we obtain the isomorphism
Hd (Hc (M )) ∼
= ER (k)r(M ) .
m
a
(ii)⇒(i): Since M is relative Cohen-Macaulay in SuppR (M/aM )\{m}, one can use
[5, Proposition 2.8] to deduce that Hi (Hc (Mp )) ∼
= Hi+c (Mp ) for all prime ideals p
pRp
aRp
pRp
in SuppR (M/aM )\{m} and for all i ∈ Z. On the other hand, since M is Gorenstein,
one can use [5, Theorem 2.5] and [5, Corollay 2.2] to see that Hn (M ) ∼
= ER (k)r(M ) .
m
Therefore, the assertion follows from [4, Theorem 4.4]. Next, one can use Proposition
2.6(iii) and [5, Proposition 2.1] to establish the final assertion.
Next, we provide an example to show that even if, for all p ∈ SuppR (M/aM )\{m},
Mp is relative Cohen-Macaulay with respect to aRp , but with different cohomological
dimensions, then Theorem 2.9 is no longer true.
Example 2.10. Let k be a filed and R = k[[x, y, z]]. Set a := (xy, xz). Then, we
have grade (a, R) = ht a = 1. Also, one can use Mayer-Vietoris sequence to see that
6
M. R. ZARGAR
cd(a, R) = 2 and H1a (R) ∼
= H1(x) (R). On the other hand, it is easy to see that, for
all prime ideals p in V(a) \ {m}, aRp = (x)Rp or aRp = (y, z)Rp ; and hence Rp is
relative Cohen-Macaulay with respect to aRp for all prime ideals p in V(a) \ {m}
with cd(aRp , Rp ) ∈ {1, 2}. Next, since Hi(x) (R) = 0 for all i 6= 1, one can see that
Him (H1(x) (R)) ∼
= Hi+1
m (R) for all i. Hence, by using the fact that R is Gorenstein, we
d
1
∼
have H (H (R)) = ER (k) and Hi (H1 (R)) = 0 for all i 6= d, where d = dim R/a.
m
a
m
a
Next, we single out a certain case of the above theorem for M = ωR , where ωR
denotes a canonical module of a Cohen-Macaulay ring R.
Corollary 2.11. Let (R, m, k) be a local Cohen-Macaulay ring which admits a
canonical R-module ωR and let a be an ideal of R with ht R a = c and dim R/a = d.
Suppose that ωR is relative Cohen-Macaulay in V(a) \ {m}. Then the following
statements are equivalent.
(i) Hia (ωR ) = 0 for all i 6= c.
∼ ER (k) and Hi (Hc (ωR )) = 0 for all i 6= d.
(ii) Hdm (Hca (ωR )) =
m
a
i
(iii) Extd (k, Hc (ωR )) ∼
k
and
Ext
(k,
Hc (ωR )) = 0 for all i 6= d.
=
(iv)
a
R
c
i
µ (m, Ha (ωR ))
R
a
= δdi .
Moreover, if a satisfies the above conditions, then ExtiR (Hca (ωR ), Hca (ωR )) = 0 for all
i 6= 0 and R̂a ∼
= HomR (Hc (ωR ), Hc (ωR )), where R̂a denotes the a-adic completion of
a
a
R.
Proof. We first notice that Supp(ωR /aωR ) = Supp(ωR ) ∩ V(a) = V(a). Therefore,
dim(ωR /aωR ) = dim(R/a) and ht R a = ht ωR a. On the other hand, one can use
[2, Proposition 3.3.11] to see that ωR is a Gorenstein R-module of type 1 and that
HomR (ωR , ωR ) ∼
= R. Hence, the assertion follows from Theorem 2.9.
The following corollary, which is an immediate consequence of Corollary 2.11, has
been proved in [3, Theorem 0.1] as a main result.
Corollary 2.12. Let (R, m, k) be a local Gorenstein ring and let a be an ideal of
R. Set ht R a = c and dim R/a = d. Suppose that R is relative Cohen-Macaulay in
V(a) \ {m}. Then the following statements are equivalent.
(i) Hia (R) = 0 for all i 6= c, i.e R is relative Cohen-Macaulay with respect to a.
(ii) Hd (Hc (R)) ∼
= ER (k) and Hi (Hc (R)) = 0 for i 6= d .
m
a
m
a
(iii) ExtdR (k, Hca (R)) ∼
= k and ExtiR (k, Hca (R)) = 0 for all i 6= d.
(iv) µi (m, Hca (R)) = δdi .
Moreover, if one of the above conditions is satisfied, then R̂a ∼
= HomR (Hca (R), Hca (R))
and ExtiR (Hca (R), Hca (R)) = 0 for all i 6= 0.
ON THE RELATIVE COHEN-MACAULAY MODULES
7
Acknowledgements. I am very grateful to Professor Hossein Zakeri for his kind
comments and assistance in the preparation of this paper. Also, the author is grateful
to the referee for careful reading and for suggesting several improvements of the
manuscript.
References
[1] M. P. Brodmann and R. Y. Sharp, Local cohomology: an algebraic introduction with geometric
applications, Cambridge University Press, Cambridge, 1998.
[2] W. Bruns and J. Herzog, Cohen-Macaulay Rings, Cambridge University Press, Cambridge,
1993.
[3] M. Hellus and P. Schenzel, On cohomologically complete intersections, J. Algebra 320 (10)
(2008) 3733-3748.
[4] M. Hellus and P. Schenzel, Notes on local cohomology and duality, J. Algebra 401 (1) (2014)
48-61.
[5] M. Rahro Zargar and H. Zakeri, On injective and Gorenstein injective dimensions of local
cohomology modules, Algebar. Colloq., to appear.
[6] P. Schenzel, Proregular sequences, local cohomology, and completion, Math. Scand. 92 (2)
(2003) 161-180.
[7] R.Y. Sharp, Gorenstein Modules, Math. Z. 115 (2) (1970) 117-139.
Majid Rahro Zargar, School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box: 19395-5746, Tehran, Iran.
E-mail address: [email protected]
| 0math.AC
|
Distributed Bayesian Filtering using Logarithmic Opinion
Pool for Dynamic Sensor Networks ?
Saptarshi Bandyopadhyay a , Soon-Jo Chung b,1 ,
arXiv:1712.04062v1 [cs.SY] 11 Dec 2017
a
b
Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, USA
Graduate Aerospace Laboratories, California Institute of Technology, Pasadena, CA 91125, USA
Abstract
The discrete-time Distributed Bayesian Filtering (DBF) algorithm is presented for the problem of tracking a target dynamic
model using a time-varying network of heterogeneous sensing agents. In the DBF algorithm, the sensing agents combine their
normalized likelihood functions in a distributed manner using the logarithmic opinion pool and the dynamic average consensus
algorithm. We show that each agent’s estimated likelihood function globally exponentially converges to an error ball centered
on the joint likelihood function of the centralized multi-sensor Bayesian filtering algorithm. We rigorously characterize the
convergence, stability, and robustness properties of the DBF algorithm. Moreover, we provide an explicit bound on the time
step size of the DBF algorithm that depends on the time-scale of the target dynamics, the desired convergence error bound,
and the modeling and communication error bounds. Furthermore, the DBF algorithm for linear-Gaussian models is cast into
a modified form of the Kalman information filter. The performance and robust properties of the DBF algorithm are validated
using numerical simulations.
Key words: Bayesian filtering, distributed estimation, sensor network, data fusion, logarithmic opinion pool.
1
Introduction
A network of time-varying, heterogeneous sensing agents
could use a distributed estimation algorithm to estimate
the states of the target dynamics in a distributed manner. Potential applications include environment and pollution monitoring, analyzing communication and social
networks, and tracking mobile targets in air, land, water,
and space. In this paper, we present a new, discrete-time
distributed estimation algorithm based on the logarithmic opinion pool that guarantees bounded convergence
to the Bayesian-optimal probability distribution of the
states of the target dynamics.
? S. Bandyopadhyay and S.-J. Chung were supported in part
by the AFOSR grant (FA95501210193) and the NSF grant
(IIS-1253758). This research was carried out in part at the
Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and
Space Administration. c 2017 California Institute of Technology. All rights reserved.
Email addresses:
[email protected] (Saptarshi
Bandyopadhyay), [email protected] (Soon-Jo Chung).
1
Corresponding author. Tel.: +1 626 395 6294.
Preprint submitted to Automatica
Discrete-time distributed estimation algorithms can
be broadly classified into three categories based on
their representation of the states of the target dynamics. Algorithms in the first category only estimate the
mean and the covariance matrix of the target’s states
[1,2,3,4,5,6,7]. These algorithms usually deal with linearized target dynamics and measurement models, and
also neglect information captured by the higher-order
moments of the estimated probability distribution of
the target’s states. The second category aims to reach
an agreement across the sensor network over a discrete
set of hypotheses about the states of the target [8,9,10].
Although these algorithms use the entire information
in the estimated probability distribution of the target’s
states, they are only applicable in cases where the target’s states can be represented by a discrete (finite)
set of hypotheses. Therefore, these algorithms are not
suitable for estimation over continuous domains.
The third category of algorithms estimates the posterior probability distribution of the states of the target [11,12,13,14,15,16,17,18]. This category forms the
most general class of distributed estimation algorithms
because these algorithms can be used for estimation
over continuous domains, and can incorporate nonlinear target dynamics, heterogeneous nonlinear measure-
13 December 2017
rithm for a time-varying, heterogeneous sensor network
on a communication graph that is much sparser than a
complete graph so that each agent’s estimate converges
to this Bayesian-optimal posterior probability distribution of the target’s states. Furthermore, we assume that
the time-varying communication network topology is periodically strongly connected and each agent can only
communicate once with its neighboring agents during
each time instant.
ment models, and non-Gaussian uncertainties. These algorithms also use the entire information (i.e., not just
the mean and the covariance matrix) in the estimated
probability distribution of the target’s states. In light
of these advantages, this paper focuses on the development of a distributed estimation algorithm that belongs
to this third category.
In third-category algorithms, the agents exchange their
local probability distributions with their neighboring
agents and combine them using fusion or diffusive coupling rules to estimate the aggregate probability distribution. Schemes for combining probability distributions in a distributed manner, like the Linear Opinion Pool (LinOP) and the Logarithmic Opinion Pool
(LogOP), were first studied in the statistics literature
[19,20,21,22,23]. The LogOP scheme is deemed ideal for
this purpose because of its favorable properties (e.g., externally Bayesian) [22].
In this paper, we present the Distributed Bayesian Filtering (DBF) algorithm to address this open question.
During each time instant, the agents exchange their
normalized likelihood functions with their neighboring
agents only once and then combine them using our fusion rule. Our fusion rule for combining arbitrary probability distributions relies on the LogOP scheme and
the dynamic average consensus algorithm [24,25,26,27].
We show that after finite time instants, the estimated
likelihood function of each agent converges to an error
ball centered on the joint likelihood function of the
centralized multi-sensor Bayesian filtering algorithm.
We also provide an explicit upper bound on the time
step size of the DBF algorithm that depends on the
time-scale of the target dynamics and the convergence
error bound. Moreover, we analyze the effect of communication and modeling errors on the DBF algorithm.
If the target dynamics are linear-Gaussian models, we
show that the DBF algorithm can be simplified to the
modified (Kalman) information filter. Finally, we show
that the distributed estimation algorithms in [14,15] are
special cases of the DBF algorithm.
We now focus on distributed estimation algorithms that
use the LogOP scheme. The first such algorithm is proposed in [11]. In particular, [12] generates informationtheoretically-optimal weights for the LogOP scheme.
Combining probability distributions within the exponential family (i.e., probability distributions that can
be expressed as exponential functions) is discussed in
[13,14]. In the distributed estimation algorithm presented in [16] as well as in our prior work [17,18], the
agents combine their local posterior probability distributions using the consensus algorithm, where the
multiple consensus loops within each time step are executed much faster than the original time steps of the
Bayesian filter. Moreover, [16,17,18] show that each
agent’s estimated probability distribution of the target’s states converges around the pdf that minimizes
the sum of Kullback–Leibler (KL) divergences from all
the posterior probability distributions of the target’s
states. Similar algorithms for combining local likelihood
functions using the consensus algorithm are proposed
in [14,15]. But the number of consensus loops within
each estimator time step grows very fast with the number of agents due to the convergence properties of the
consensus algorithm [24,25,26]. Hence, such algorithms
are not feasible if the time-scale of the target dynamics is comparatively fast. This connection between the
time-scale of the target dynamics and the time step size
of the distributed estimation algorithm has not been
explored in the literature.
Furthermore, [16] analyzed their algorithm using linearGaussian models while [13] focused on probability distributions within the exponential family. In contrast, we
present a rigorous proof technique, which was first introduced in our prior work [17,18], for the LogOP scheme
that is applicable for general probability distributions.
This paper is organized as follows. Section 2 presents
some preliminaries and the problem statement. The LogOP scheme and some general convergence results are
presented in Section 3. The DBF algorithm and its special cases are presented in Section 4. Results of numerical simulations are presented in Section 5 and the paper
is concluded in Section 6.
2
If all the agents are perfectly connected by a complete
communication graph (i.e., each agent could communicate instantaneously with every other agent without any
loss of information in the communication links), then
the agents can exchange their local likelihood functions
and use the centralized multi-sensor Bayesian filtering
algorithm to estimate the Bayesian-optimal posterior
probability distribution of the target’s states. An open
question is how to design a distributed estimation algo-
Preliminaries and Problem Statement
Let N and R represent the sets of natural numbers (positive integers) and real numbers respectively. The state
space of the target’s states X is a closed set in Rnx ,
where nx is the dimension of the states of the target.
Let X be the Borel σ–algebra for X . A probability
space is defined by the three-tuple {X , X , P}, where P
is a complete, σ-additive probability measure on all X .
dP(x)
Let p(x) = dµ(x)
denote the Radon–Nikodým density
2
xk+1 = f k (xk , wk , ∆) , ∀k ∈ N ,
of the probability distribution P(x) with respect to a
measure µ(x). If x ∈ X is continuous and µ(x) is a
Lebesgue measure, p(x) is the probability density function (pdf) [28]. Therefore, the probability of an event
A ∈ X can ´be written as the Lebesgue–Stieltjes integral P(A ) = A p(x) dµ(x). In this paper, we only deal
with the continuous case where the function p(·) represents the pdf and µ(·) is the Lebesgue measure. Let
Φ(X ) represent the set of all pdfs over the state space
X . The L1 distance and the KL divergence between the
pdfs P, Q ∈ Φ(X ) areˆdenoted by:
DL1 (P, Q) =
ˆX
DKL (P||Q) =
X
(1)
where f k : Rnx × Rnw → Rnx is a possibly nonlinear
time-varying function of the state xk , the discretization
time step size ∆, and an independent and identically
distributed (i.i.d.) process noise wk , where nw is the
dimension of the process noise vector.
Consider a network of N heterogeneous sensing agents
simultaneously tracking this target. Let y ik represent the
measurement taken by the ith agent at the k th time instant. The measurement model of the agents is given by:
|P(x) − Q(x)| dµ(x) ,
P(x)
P(x) log Q(x)
dµ(x) .
y ik = hik (xk , v ik ), ∀i ∈ V = {1, . . . , N } , ∀k ∈ N , (2)
where hik : Rnx × Rnvi → Rnyi is a possibly nonlinear
time-varying function of the state xk and an i.i.d. measurement noise v ik , where nyi and nvi are dimensions
of the measurement and measurement noise vectors respectively. Note that the measurements are conditionally independent given the target’s states. We assume
that the target dynamics (1) and measurement models
(2) are known.
In this paper, all the algorithms are presented in discrete
time. Let ∆ be the time step size between any two consecutive time instants. The time index is denoted by a
right subscript. The agent index is denoted by a lowercase right superscript. Frequently used symbols are listed
in the following table.
Table 1: List of frequently used symbols
Symbol
Definition
Ak
Adjacency matrix
b
Periodicity of the communication network
Jki
Lik
LC
k
Inclusive neighbors of the ith agent
N
Number of agents in the network
Ski
Prior pdf
Tki
Uki
Wki
WkC,i
Estimated likelihood function
X
State space
xk
True states of the target
xk|k−1
Predicted states of the target
xk|k
Updated states of the target
y ik
Measurement taken by the ith agent
∆
Time step size
2.1
2.2
Bayesian Filtering Algorithm
Each agent uses the Bayesian filtering algorithm to estimate the pdf of the states of the target [29,28]. Let
xk|k−1 and xk|k represent the predicted and updated
states of the target at the k th time instant. Let the
pdfs Ski = p(xk|k−1 ) ∈ Φ(X ) and Wki = p(xk|k ) =
p(xk|k−1 |y ik ) ∈ Φ(X ) denote the ith agent’s prior and
posterior pdfs of the target’s states at the k th time instant.
Normalized likelihood function
Normalized joint likelihood function
During the prediction step, the prior pdf Ski = p(xk|k−1 )
i
is obtained from the previous posterior pdf Wk−1
=
p(xk−1|k−1 ) using the Chapman–Kolmogorov equation
[28]:
ˆ
i
i
Sk =
p(xk|k−1 |xk−1|k−1 ) Wk−1
dµ(xk−1|k−1 ), (3)
Estimated KL-divergence-minimizing pdf
Posterior pdf
Centralized posterior pdf
X
where the probabilistic model of the state evolution
p(xk|k−1 |xk−1|k−1 ) is obtained from the known target
dynamics model (1). We assume that the prior pdf is
available at the start of the estimation process.
The new measurement y ik is used to compute the posterior pdf Wki = p(xk|k ) = p(xk|k−1 |y ik ) during the update step using the Bayes’ rule [28]:
Wki = ´
Target Dynamics and Measurement Models
X
Let xk represent the true states of the target at the k th
time instant, where xk ∈ X for all k ∈ N. The dynamics
of the target in discrete time is given by:
p(y ik |xk|k−1 ) Ski
.
p(y ik |xk|k−1 ) Ski dµ(xk|k−1 )
(4)
The likelihood function p(y ik |xk|k−1 ) is obtained from
the ith agent’s known measurement model (2). Let the
3
pdf Lik ∈ Φ(X ) represent the normalized likelihood function, i.e.,
Lik = ´
X
p(y ik |xk|k−1 )
.
p(y ik |xk|k−1 ) dµ(xk|k−1 )
Therefore, (4) is equivalent to Wki =
´
X
where the convergence error is given by
DL1 Tki , LC
k ≤ (1 + η)δ , ∀k ≥ κ , ∀i ∈ V , (8)
∀i ∈ V .
(9)
lim DL1 Tki , LC
k ≤ δ,
(5)
k→∞
The DBF algorithm, shown in Fig. 1 and Algorithm 1,
achieves this objective. Note that the agents exchange
their estimated pdfs with their neighboring agents only
once during each time instant before the fusion step.
Lik Ski
.
Lik Ski dµ(xk|k−1 )
If all the sensing agents are hypothetically connected by
a complete graph, then the agents can exchange their
likelihood functions. Each agent can use the centralized multi-sensor Bayesian filtering algorithm to compute the centralized posterior pdf of the target’s states
WkC,i = p(xk|k ) = p(xk|k−1 |y 1k , . . . , y N
k ) ∈ Φ(X ) using
the Bayes’ rule [30]:
i
LC
k Sk
.
C
i
L S dµ(xk|k−1 )
X k k
QN
j
j=1 Lk
C
L k = ´ QN
.
j
j=1 Lk dµ(xk|k−1 )
X
WkC,i = ´
where
Here,
LC
k
Prior Knowledge of
States
Posterior pdf from
previous time instant
Prediction Step
Compute pdf
.
Transmit the
pdf
to neighbors
Exchange pdfs with
neighboring agents
(6)
Receive pdfs
.
from
neighbors
(7)
Fusion Step
Compute pdfs
(7) is the normalized joint likelihood function.
Measurement.
.
Next time instant
Bayesian filtering is optimal because this posterior pdf
WkC,i integrates and uses all the available information
expressed by probabilities [28]. Moreover, an optimal
state estimate with respect to any criterion can be computed from this posterior pdf WkC,i . The minimum meansquare error (MMSE) estimate and the maximum a posteriori (MAP) estimate are given by [31]:
ˆ
M SE
AP
x̂M
=
x WkC,i dµ(x), x̂M
= arg max WkC,i .
k|k
k|k
X
Update Step
Compute pdf
Fig. 1. Flowchart of the DBF algorithm (for the ith agent at
the kth time instant)
2.4
Communication Network Topology
x∈X
The time-varying communication network topology of
the sensor network is denoted by the directed graph
Gk = (V, Ek ). The edge (i, j) ∈ Ek if and only if the ith
agent receives information from the j th agent at the k th
time instant. The inclusive neighbors of the ith agent are
denoted by Jki = {j ∈ V : (i, j) ∈ Ek } ∪ {i}. The matrix Ak ∈ RN ×N represents the adjacency matrix of Gk ,
where Ak [i, j] 6= 0 if and only if j ∈ Jki .
Other potential criteria for optimality, such as maximum likelihood, minimum conditional KL divergence,
and minimum free energy, are discussed in [28,31]. The
main advantage of the original Bayesian filtering formulation is that no approximation is needed during the filtering process; i.e., the complete information about the
dynamics and uncertainties of the model can be incorporated in the filtering algorithm. However, direct implementation of Bayesian filtering (3)–(4) is computationally expensive. Practical implementation of these algorithms, in their most general form, is achieved using particle filtering [29,32] and Bayesian programming [33,34].
2.3
.
Assumption 1 [26,27] The digraph Gk = (V, Ek ) and
its adjacency matrix Ak satisfy the following properties:
(i) There exists some positive integer b ∈ N such that the
directed graph (V, Ek ∪ Ek+1 ∪ . . . ∪ Ek+b−1 ) is strongly
connected for all time instants k ∈ N.
(ii) The matrix Ak is doubly stochastic, i.e., 1T Ak = 1T
and Ak 1 = 1 for all k ∈ N, where 1 = [1, 1, . . . , 1]T .
(iii) The matrix product Ak,k+b−1 is defined as:
Problem Statement
Let the pdf Tki ∈ Φ(X ) denote the estimated joint likelihood function of the ith agent at the k th time instant.
2
) be positive constants.
Let η ∈ (0, 1) and δ ∈ (0, 1+η
Our aim is to design a discrete-time distributed estimation algorithm, over the communication network topology described in Section 2.4, so that each agent’s Tki converges to the normalized joint likelihood function LC
k (7),
Ak,k+b−1 =
k+b−1
Y
!
Aτ
.
(10)
τ =k
There exists a constant γ ∈ (0, 12 ) such that each element
Ak,k+b−1 [i, j] ∈ [γ, 1] ∪ {0} for all i, j ∈ V and k ∈ N.
4
Therefore, the digraph Gk is periodically strongly connected and the matrix Ak is non-degenerate and balanced.
P1
P2
0.1
0.05
Note that if b = 1, then the digraph Gk is strongly connected at all time instants k ∈ N.
LinOP(P1 , P2 )
LogOP(P1 , P2 )
0.1
α1 = 0.5
0.05
0
0
0
10
20
30
0
10
20
x
3
Logarithmic Opinion Pool and Convergence
Results
(a)
0.1
P3
P4
0.08
PkLinOP (x) =
αki Pki (x) ,
0.06
0.04
0.04
0.02
0.02
0
(11)
LinOP(P3 , P4 )
LogOP(P3 , P4 )
0.08
0.06
α1 = 0.5
0
0
N
X
(b)
0.1
Let the pdf Pki ∈ Φ(X ) denote the ith agent’s pdf at the
k th time instant. The LinOP and LogOP schemes for
combining the pdfs Pki are given by [20]:
30
x
10
20
30
40
50
0
10
20
30
x
x
(c)
(d)
40
50
i=1
PkLogOP (x)
αik
i
ΠN
i=1 Pk (x)
=´
,
i
N
i (x̄) αk dµ(x̄)
Π
P
i=1
k
X
Fig. 2. The pdfs in (a) and (c) are combined using LinOP
and LogOP in (b) and (d). Note that the LogOP solution
preserves the modal nature of the original pdfs.
(12)
Definition 1 Under Assumption 2, for any constant
ψ ∈ X , we have Pki (ψ) > 0, ∀i ∈ V and PkLogOP (ψ) > 0.
Using simple algebraic manipulation of (12), we get [23]:
"
# N
X
PkLogOP (x)
LogOP
αki Pki (x) , (13)
=
Pk
(x) := ln
LogOP
Pk
(ψ)
i=1
i
P
(x)
k
where Pki (x) := ln
, ∀i ∈ V .
(14)
Pki (ψ)
PN
where the weights αki are such that i=1 αki = 1 and
the integral in the denominator of (12) is finite. Thus,
the combined pdf obtained using LinOP and LogOP
gives the weighted algebraic and geometric averages of
the individual pdfs respectively. As shown in Fig. 2, the
combined pdf obtained using LogOP typically preserves
the multimodal or unimodal nature of the original individual pdfs [22]. The most compelling reason for using
the LogOP scheme is that it is externally Bayesian; i.e.,
the LogOP combination step commutes with the process
of updating the pdfs by multiplying with a commonly
agreed likelihood pdf Lk ∈ Φ(X ):
Thus, we have represented the LogOP scheme (12)
as a linear equation using these functions Pki (x) and
PkLogOP (x), and removed the effect of the normalizing
constants.
αik
Lk Pki
´
ΠN
i dµ(x̄)
i=1
L
P
k
X
k
.
=
´
αik
LogOP
i
´
L
P
L
P
dµ(x̃)
k
N
k
´
X k k
dµ(x̃)
Πi=1
X
Lk P i dµ(x̄)
Lk PkLogOP
X
We now state some useful convergence results using the
functions in Definition 1. The proofs are given in Appendix.
k
Therefore, the LogOP scheme is ideal for combining pdfs
in distributed estimation algorithms.
Definition 2 (Pointwise Convergence) The pdf Pki converges pointwise to the pdf P ? ∈ Φ(X ), if and only if
limk→∞ Pki (x) = P ? (x) for all x ∈ X .
Due to the multiplicative nature of the LogOP scheme,
each agent has veto power [22]. That is, if Pki (x) = 0
for some x ∈ X and some agent i ∈ V with αki > 0,
then PkLogOP (x) = 0 in the combined pdf irrespective of
the pdfs of the other agents. In order to avoid this veto
condition, we enforce the following assumption which
has been used in the literature.
Lemma 1 If the pdfs P, Q satisfy Assumption 2, then
there exists ψ ∈ X such that P(ψ) = Q(ψ).
i
Lemma 2 If the function
converges pointwise
k (14)
i
h P
?
P
(x)
?
to the function P := ln P ? (ψ) , then the corresponding
pdf Pki also converges pointwise to the pdf P ? .
Assumption 2 [10,22] (Nonzero Probability Property)
In this paper, all pdfs are strictly positive everywhere in
the closed set X .
Definition 3 (Convergence in TV) The measure µPki
is defined as the measure
induced by the pdf Pki on X ,
´
i
where µPki (A ) = A Pk dµ(x) for any event A ∈ X .
Similarly, let µP ? denote the measure induced by the pdf
P ? on X .
In order to analyze the LogOP scheme with general probability distributions that satisfy Assumption 2, we use
the following functions.
5
The TV distance is defined as kµPki − µP ? kTV :=
supA ∈X |µPki (A ) − µP ? (A )|. The measure µPki converges to the measure µP ? in TV, if and only if
k limk→∞ µPki − µP ? kTV = 0.
The necessary conditions
for satisfying (17) are
given by
DKL Lik ||Lik−1 ≤ ∆θL and DKL Lik−1 ||Lik ≤ ∆θL .
We now state the DBF algorithm, whose steps are shown
in Fig. 1. Let the pdf Uki ∈ Φ(X ) denote the estimated
KL-divergence-minimizing pdf of the ith agent at the
k th time instant. The pdf Tki is defined in Section 2.3.
Under Assumptions 1–3, the pseudo-code of the DBF
algorithm is given in Algorithm 1.
Lemma 3 If the pdf Pki converges pointwise to the pdf
P ? , then the measure µPki converges in TV to the measure
µP ? . Moreover, kµPki − µP ? kTV = 12 DL1 Pki , P ? .
Another reason for using the LogOP scheme is that it
minimizes the information lost during the combination
process, where the information loss is measured using
the KL divergence.
Algorithm 1. Distributed Bayesian Filtering Algorithm
1. (ith agent’s steps at k th time instant)
Lemma 4 [16,17] The pdf PkKL ∈ Φ(X ) that globally
minimizes the sum of KL divergences with the pdfs Pki
for all agents is given by:
1
QN
N
i N
X
i=1 Pk
KL
i
Pk = arg min
DKL ρ||Pk = ´ Q
.
1
N
i N dµ(x̄)
ρ∈Φ(X ) i=1
i=1 Pk
X
2.
Compute prior pdf Ski = p(xk|k−1 ) using (3)
3.
Obtain local measurement y ik
4.
Compute normalized likelihood function Lik using (5)
5.
j
Receive pdfs Uk−1
from agents j ∈ Jki
6.
Compute pdfs Uki and Tki as follows
i
L1
−1
i
Λ Lik Lik−1
Uk =
´
−1
Λ Li Lik−1
dµ(x̄)
X Yk
j
Ak [i,j]
where Λ =
(Uk−1 )
,
Note that the pdf PkKL is equivalent to the pdf PkLogOP (12)
obtained using the LogOP scheme with weights αki = N1
for all agents.
Tki =
The proof of Lemma 4 is given in our prior work [17].
Note that the normalized joint likelihood function LC
k
(7) is also given by:
QN
N
j
LKL
j=1 Lk
k
C
Lk = ´ QN
, (15)
=´
N
j
LKL
dµ(x̄)
j=1 Lk dµ(x̄)
X
k
X
QN j N1
j=1 Lk
where LKL
=
.
(16)
k
´ QN j N1
L
dµ(x̄)
j=1
k
X
7.
Tki Ski
.
Tki Ski dµ(x̄)
(20)
The following theorem shows that the DBF algorithm
satisfies the problem statement (8)–(9) in Section 2.3.
Theorem 5 Under Assumptions 1–3, if all the agents
execute the DBF algorithm (Algorithm 1) and the time
step size ∆ for (17) for Algorithm 1 is defined as
∆=
(1 − σm ) log (δ + 1)
√
,
2bN (N − 1) N θL
(21)
then the steady-state convergence error between the pdf
Tki (19) and the pdf LC
k (7) is bounded by δ:
lim
max
DL1 Tki , LC
(22)
k ≤ δ.
k→∞ i∈V={1,...,N }
Furthermore, the convergence error between the pdfs
Tki (19) and LC
k (7) after κ time instants is bounded by
(1 + η)δ:
max
DL1 Tki , LC
k ≤ (1 + η)δ , ∀k ≥ κ , (23)
Assumption 3 For any time step size ∆ > 0, there
exists a time-invariant constant θL > 0 such that for all
agents i ∈ V = {1, . . . , N }:
≤ e∆θL , ∀x ∈ X , ∀k ∈ N .
(18)
(19)
X
In this section, we present the main DBF algorithm, its
convergence and robustness properties, and its application to special cases. We first state an assumption on
the time-varying nature of the pdfs Lik (5) for all agents
that directly link the target dynamics and measurement
models with the time step size of the distributed estimation algorithm.
e−∆θL ≤
,
Compute posterior pdf Wki = p(xk|k ) as follows
Distributed Bayesian Filtering Algorithm
Lik (x)
Lik−1 (x)
if k ≥ 2
,
Wki = p(xk|k ) = ´
We show that the DBF algorithm also estimates the pdf
LKL
k (16) in a distributed manner.
4
j∈Jki
(Uki )N
´
(Uki )N dµ(x̄)
X
if k = 1
i∈V={1,...,N }
(17)
where, κ = 1 if D1 ≤
6
log(δ+1)
3
N2
, otherwise:
b(N − 1)
log
3
κ=
log σm log
+ 1 ,
eN 2 D1
log
δ+1
!
L` (x)
D1 = 2 ln max max 1j
.
`,j∈V x∈X L (x)
1
(1+η)δ+1
δ+1
N
X
(24)
=
1−
4(γ − γ N )
π
sin2
(1 − γ)
2N
j
Ak [i, j]Uk−1
+
N
X
i=1 j=1
i
Lki − Lk−1
,
i=1
N
N
X
X
i
i
Ak [j, i] Uk−1
+
Lki − Lk−1
,
N
X
j=1
i=1
U1i − L1i +
i=1
N
X
Lki = N LkKL .
(30)
i=1
PN
Note that LkKL = N1 i=1 Lki follows from (16). This
shows that if the functions Uki converge towards each
other, then they will converge to the function LkKL . Let
us define the error vector ek as:
h
iT
ek = Uk1 −LkKL , . . . , Uki −LkKL , . . . , UkN −LkKL
.
21
< 1.
N
X
i=1
(25)
2
Here, η ∈ (0, 1) and δ ∈ (0, 1+η
) are positive constants,
and σm = maxk∈N σN −1 (Ak,k+b(N −1)−1 ), where σN −1
denotes the second largest singular value of the matrix,
and σm is upper bounded by:
N X
N
X
i=1
=
σm ≤
Uki =
(26)
The evolution of the error vector ek is given by:
Moreover, b is the periodicity of the communication network topology, γ ∈ (0, 12 ) is the smallest positive element
in Ak,k+b−1 defined in Assumption 1, and θL is defined
in Assumption 3.
The TV error between the measures induced by the pdfs
Tki and LC
k is bounded by:
(1 + η)δ
max kµTki − µLC kTV ≤
,
k
i∈V
2
δ
lim max kµTki − µLC kTV ≤ .
k
k→∞ i∈V
2
∀k ≥ κ ,
where
ek = Ak ek−1 + Ωk,k ,
∀k ≥ 2 ,
1 1
Lk −Lk−1
11T
..
.
Ωk,k = I −
. N
N
N
Lk −Lk−1
The overall evolution of the error vector ek after b ∈ N
time instants is given by:
(27)
ek+b−1 = Ak,k+b−1 ek−1 + Ωk,k+b−1 ,
where Ak,k+b−1 is defined in (10) and for b ≥ 2:
k+b−2
X
(Aτ +1,k+b−1 Ωτ,τ ) + Ωk+b−1,k+b−1 .
τ =k
Proof: Using h Definition
1, we define
i
h C thei functions
LKL
Lk (x)
C
k (x)
LkKL (x) = ln LKL
,
L
(x)
=
ln
, Lki (x) =
C
k
(ψ)
k
h i i
h i i Lk (ψ)
L (x)
U (x)
ln Lik(ψ) , Uki (x) = ln U ki (ψ) , and Tki (x) =
k
h ki i
T (x)
ln T ki (ψ) for all i ∈ V. Since these functions are de-
Note that 1Tek = 0 because
of (30) and 1T Ωk,k+b−1 = 0
T
because 1T I − 11
= 0. Therefore, we investigate
N
the convergence of ek along all directions that are orthogonal to 1T . It follows from Assumption 1 that the
matrix Ak,k+b−1 is irreducible. Therefore, the matrix
Ak,k+b−1 is primitive [35, Lemma 8.5.4, pp. 516] and
λN −1 (Ak,k+b−1 ) < 1, where λN −1 denotes hthe secondi
k
fined for all x ∈ X , we henceforth drop the term (x) for
brevity.
largest eigenvalue of the matrix. Let Vtr = √1N 1, Vs
be the orthonormal matrix of eigenvectors of the symmetric primitive matrix AT1,b A1,b . By spectral decomposition [36], we get:
h
i
1
01×(N −1)
VtrT AT1,b A1,b Vtr = 0(N −1)×1
,
T
T
V A A1,b Vs
Step 1. We first show that the pdf Uki (18) converges to
the pdf LKL
k (16). Equation (18) can be re-written using
these functions as:
(
(32)
(28)
Ωk,k+b−1 =
Uki =
(31)
L1i
PN
if k = 1
, (29)
j
i
i
A
[i,
j]U
+L
−L
if
k≥2
k
k
k−1
j=1
k−1
s
1,b
1 T T
= 1, √1N 1T AT1,b A1,b Vs
N 1 A1,b A1,b 1
01×(N −1) , and VsT AT1,b A1,b 1 √1N = 0(N −1)×1
where
=
are
used. Since the eigenvectors are orthonormal, we have
Vs VsT + N1 11T = I. Left-multiplying (32) with VsT gives:
because Ak [i, j] = 0 if j 6∈ Jki , as defined in Section 2.4.
Since Ak is doubly stochastic, (29) satisfies the conservation property:
7
max |Uki (x) − LkKL (x)| ≤ Ξk ,
VsT ek+b−1 = VsT Ωk,k+b−1
+
=
VsT Ak,k+b−1
VsT Ωk,k+b−1
+
Vs VsT + N1 11 ek−1 ,
VsT Ak,k+b−1 Vs VsT ek−1 .
i∈V
T
(33)
≤ kVsT Ak,k+b−1 Vs k2 kVsT ek−1 k2
,
LkC (x) = N LkKL (x) ,
max |Tki (x) − LkC (x)| ≤ N Ξk ,
Using Lemma 1, we select ψ ∈ X such that Tki (ψ) =
LC
k (ψ). Therefore,
i
T (x)
max ln kC
≤ N Ξk , ∀k ∈ N ,
i∈V
Lk (x)
i
Tk (x)
≤ eN Ξk , ∀k ∈ N .
e−N Ξk ≤ max
i∈V
LC
(x)
k
by (26). Moreover, it follows from Assumption 3 that
√
T
1
N
,...,LkN −Lk−1
]
≤ 2 N ∆θL because |Lki −
[ Lk1 −Lk−1
C
N Ξk
− 1 , ∀k ∈ N .
max Tki (x) − LC
k (x) ≤ Lk (x) e
2
≤ 2∆θL . Therefore, we have:
i∈V
Since x ∈ X can be any point, therefore:
ˆ
i
C
Tki − LC
max DL1 Tk , Lk = max
k dµ(x)
i∈V
i∈V X
ˆ
N Ξk
≤ eN Ξ k − 1
LC
− 1 , ∀k ∈ N .
k dµ(x) = e
√
kVsT Ωk,k+b(N −1)−1 k2 ≤ 2b(N − 1) N ∆θL .
Hence, in the presence of the disturbance term, we get:
kVsT ek+b(N −1)−1 k2 ≤ kVsT Ωk,k+b(N −1)−1 k2
+ σN −1 (Ak,k+b(N −1)−1 )kVsT ek−1 k2
√
σm kVsT ek−1 k2 + 2b(N − 1) N ∆θL .
X
,
Hence the convergence
error between the pdfs is bounded
by eN Ξk − 1 .
(34)
It follows from (23)–(22) that eN Ξk − 1 ≤ (1 + η)δ for
all k ≥ κ and limk→∞ eN Ξk − 1 ≤ δ. The time step
size ∆ (21) is found using the steady-state error term:
√
Using the discrete Gronwall lemma [39, pp. 9] we obtain:
k−1
b(N −1)
b(N −1)
1 − σm
+
1 − σm
kVsT e1 (x)k2
(38)
Thus, the error between the function Tki and the function LkC is bounded by N Ξk . We have:
i
C
Tk (x)
Lk (x)
max ln
≤ N Ξk , ∀k ∈ N .
− ln
i∈V
Tki (ψ)
LC
k (ψ)
N
kVsT e1 k2
k−1
∀k ∈ N .
i∈V
bounded by γ−γ
1−γ , and we have σN −1 (Ak,k+b(N −1)−1 ) ≤
12
N
)
2 π
< 1 [38]. Therefore, σm is given
1 − 4(γ−γ
sin
(1−γ)
2N
Tki (x) = N Uki (x) , ∀i ∈ V .
Therefore, it follows from (36) that:
Since the matrix Ak,k+b−1 is irreducible, the matrix
Ak,k+b(N −1)−1 is a positive matrix because the maximum path length between any two agents is less than or
equal to b(N − 1) [37]. Hence the measure of irreducibility of the matrix ATk,k+b(N −1)−1 Ak,k+b(N −1)−1 is lower
kVsT ek k2 ≤ σm
(37)
Step 2. We now show that the pdf Tki (19) converges to
the pdf LC
k (15). For all x ∈ X , equations (15) and (19)
can be re-written as:
where σmax denotes the largest singular value of the
matrix. Since VsT is orthonormal (i.e., VsT Vs = I)
and also orthogonal to 1T (i.e., VsT 1 = 0) and
the matrix ATk,k+b−1 Ak,k+b−1 is primitive, we have
σmax (Ak,k+b−1 Vs ) = σN −1 (Ak,k+b−1 ) < 1, where σN −1
denotes the second largest singular value of the matrix.
Therefore, the error vector VsT ek is globally exponentially stable in absence of the disturbance term.
≤
b(N −1)
Thus, the error between the function Uki and the function LkKL is bounded by Ξk , which depends on time instant k.
kVsT ek+b−1 k2 = kVsT Ak,k+b−1 Vs VsT ek−1 k2
i
Lk−1
|
√
N ∆θL
σm
N D1 − 2b(N −1)
1−σm
√
2b(N − 1) N ∆θL
+
.
1 − σm
where Ξk =
We first investigate the stability of this system without the disturbance term VsT Ωk,k+b−1 in (33). Let
kVsT ek+b−1 k2 be a candidate Lyapunov function for
this system. Therefore, we get:
≤ σmax (Ak,k+b−1 Vs )kVsT ek−1 k2
= σN −1 (Ak,k+b−1 )kVsT ek−1 k2 ,
√
∀k ∈ N , (36)
k−1
√
2b(N − 1) N ∆θL .
N
e
(35)
√
If
Moreover,
≤ N D1 , where D1 is defined
in (25). Therefore, it follows that for all x ∈ X :
√
N D1 ≤
2b(N −1) N ∆θL
1−σm
√
2b(N −1) N ∆θL
,
1−σm
(39)
then eN Ξk − 1 ≤ (1 + η)δ
for all k ∈ N. Therefore, if D1 ≤
8
−1 = δ.
log(δ+1)
3
N2
, then κ = 1.
Otherwise, for D1 > log(δ+1)
, κ (24) is computed using
3
N2
N Ξk
e
− 1 ≤ (1 + η)δ. The constraint on TV error
follows from Lemma 3. Note that our exponentialconvergence proof is substantially different from the
asymptotic-convergence proof in [27].
4.1
In this section, we study the robustness of the DBF algorithm in the presence of communication and modeling
errors. In order to implement the DBF algorithm, the
j
agents need to communicate their estimated pdfs Uk−1
with their neighboring agents (see line 5 in Algorithm 1).
Remark 1 A key advantage of the DBF algorithm is
that it does not require all the sensors to observe the target. If an agent does not observe the target, then it sets
its normalized likelihood function as the uniform distribution, i.e., Lik (x) = 1. Then this agent’s likelihood function does not influence the joint likelihood function and
the estimated pdfs because of the geometric nature of the
fusion rule.
Moreover, the DBF algorithm avoids double counting because the summation of weights from all paths is a constant due to the weights in the adjacency matrix Ak .
Theorem 5 explicitly bounds the time step size ∆ of the
distributed estimation algorithm with the time-scale of
the target dynamics. But the effectiveness of the DBF
algorithm is predicated on Assumption 3. Moreover, the
upper bound on the time step size ∆max (21) decreases
with increasing number of agents N .
Remark 2 (Communication of pdfs) The information
theoretic approach for communicating pdfs is studied in
[40]. If particle filters are used to implement the Bayesian
filter and combine the pdfs [32,41], then the resampled
particles represent the agent’s estimated pdf. Hence communicating pdfs is equivalent to transmitting these resampled particles. Another approach involves approximating the pdf by a weighted sum of Gaussian pdfs [42,
pp. 213] and then transmitting this approximate distribution. Several techniques for estimating the Gaussian
parameters are discussed in the Gaussian mixture model
literature [41,43,44].
Let the pdf Ûki ∈ Φ(X ) denote the pdf Uki that is corrupted with communication errors. Similarly, let the pdf
L̂ik ∈ Φ(X ) represent the normalized likelihood function
Lik that is corrupted with modeling errors. We first state
the assumptions on these errors and then state the main
result of this section.
The following corollary provides sharper bounds for the
special case of a static, strongly-connected communication network topology.
Corollary 6 If the communication network topology is
time-invariant and strongly-connected, then the time step
size ∆ (21) and κ (24) in Theorem 5 are given by:
(1 − σN −1 (A)) log (δ + 1)
√
,
2N N θL
(1+η)δ+1
log
δ+1
1
3
κ=
log σN −1 (A) log
+ 1,
eN 2 D1
log
δ+1
∆=
Robustness Analysis
Assumption 4 There exists time-invariant constants
εU ≥ 0 and εL ≥ 0 such that for all agents i ∈ V:
(40)
e−εU ≤
Ûki (x)
≤ e εU ,
Uki (x)
∀x ∈ X , ∀k ∈ N ,
(43)
(41)
e−εL ≤
L̂ik (x)
≤ e εL ,
Lik (x)
∀x ∈ X , ∀k ∈ N .
(44)
i
i
ˆi
ˆi
Therefore, |U
h k i− Uki | ≤ 2εU and |Lk −h Lik | ≤i 2εL , where
L̂ (x)
Û (x)
and Lˆi (x) = ln k
.
Uˆi (x) = ln k
where A is the time-invariant adjacency matrix.
k
Ûki (ψ)
k
L̂ik (ψ)
Proof: In this case, (34) is written as:
√
kVsT ek k2 ≤ σN −1 (A)kVsT ek−1 k2 + 2 N ∆θL .
Corollary 7 Under Assumptions 1–4, the time step size
∆ (21) in Theorem 5 is given by:
Using the discrete Gronwall lemma [39, pp. 9] we obtain:
kVsT ek k2
≤ (σN −1 (A))
+
k−1
∆=
kVsT e1 k2
k−1 √
1 − (σN −1 (A))
1 − σN −1 (A)
2 N ∆θL .
(1 − σm ) log (δ + 1) 2εL + εU
√
−
,
θL
2bN (N − 1) N θL
(45)
where εU and εL are defined in Assumption 4.
(42)
Proof: Equation (18) can be written as:
ˆi
if k = 1
L
P1N
i
ˆj
Uk =
,
A
[i,
j]
U
k
j=1,j6=i
k−1
+A [i, i]U i + Lˆi − Lˆi
k
k−1
k
k−1 if k ≥ 2
Hence, we get maxi∈V |Uki (x) − LkKL (x)| ≤ Ξk for all
k ∈ N, where
√
k−1 √
k−1
N −1 (A))
Ξk = (σN −1 (A))
N D1 + 1−(σ
2 N ∆θL .
1−σN −1 (A)
We get ∆ (40) and κ (41) from limk→∞ eN Ξk − 1 ≤ δ
and eN Ξk − 1 ≤ (1 + η)δ for all k ≥ κ respectively.
Substituting the bounds from Assumption 4 gives:
9
(46)
|U1i − L1i | ≤ 2εL ,
Algorithm 2. DBF-Kalman Information Filtering Algorithm
N
X
j
i
|Uki −
Ak [i, j]Uk−1
−Lki +Lk−1
| ≤ 2εU + 4εL .
j=1
1.
(ith agent’s steps at k th time instant)
2.
Compute the prior pdf Ski = N (x̂ik|k−1 , P ik|k−1 )
ẑ ik−1|k−1 = (P ik−1|k−1 )−1 x̂ik−1|k−1 ,
The evolution of the error vector ek is now given by:
ek = Ak ek−1 + Ω̂k,k ,
Z ik−1|k−1 = (P ik−1|k−1 )−1 ,
M ik−1 = (F −1
)T Z ik−1|k−1 F −1
k−1 ,
k−1
−1 i
i
i
i
Z k|k−1 = I − M k−1 M k−1 + Q−1
M k−1 ,
k−1
−1 T i
i
i
i
−1 −1
ẑ k|k−1 = I − M k−1 M k−1 + Qk−1
(F k−1 ) ẑ k−1|k−1 ,
∀k ≥ 2 ,
(47)
√
where kΩ̂k,k k2 ≤ kΩk,k k2 + 2 N (εU + 2εL ) .
Similar to the proof of Theorem 5, we get:
k−1
kVsT ek k2 ≤ σm
k−1
b(N −1)
b(N −1)
+
1 − σm
1 − σm
kVsT e1 k2
√
2b(N − 1) N (∆θL + 2εL + εU ) .
Obtain local measurement y ik
4.
j
Receive pdfs Uk−1
from agents j ∈ Jki
5.
Compute the pdfs Uki and Tki as follows:
k
tik = N uik ,
6.
4.3
In this section, we apply the DBF algorithm to the special case where the target dynamics and measurement
models are given by linear systems with additive Gaussian noise:
(51)
(52)
=
(Z ik|k )−1
,
Z ik|k = Z ik|k−1 + T jk ,
x̂ik|k
= P ik|k ẑ ik|k .
(50)
Special Case: Multiple Consensus Loops within
Each Time Instant
In this section, we show that the proposed DBF algorithm can be easily extended to recursively combine
local likelihood functions using multiple consensus loops
within each time instant so that each agent’s estimated
likelihood function converges to the joint likelihood
function LC
k (15). Then, the resultant DBF algorithm
is equivalent to the Bayesian consensus algorithms in
[14,15]. Note that multiple consensus loops within each
time step significantly reduces the practicality of such
i
i
algorithms. Let the pdfs Uk,ν
∈ Φ(X ) and Tk,ν
∈ Φ(X )
th
denote to the local pdfs of the i agent during the ν th
consensus loop at the k th time instant. Since the pdf Lik
is not updated during the k th time instant, we define
the pdfs Lik,ν = Lik for all ν ∈ N. During the ν th coni
sensus loop, each agent updates its local pdfs Uk,ν
and
i
Tk,ν using the following fusion rule:
i
Lk,1 Q
if ν = 1
j
(Uk,ν−1
)Ak [i,j]
i
j∈J i
, (53)
Uk,ν
=
Q kj
´
if ν ≥ 2
(Uk,ν−1 )Ak [i,j] dµ(x)
X
Special Case: DBF-Kalman Information Filter
∀k ∈ N ,
Compute the posterior pdf Wki = N (x̂ik|k , P ik|k )
P ik|k
It follows from Corollary 7 that in order to generate satisfactory estimates using the DBF algorithm, the bounds
εU , εL should be substantially smaller than δ.
∀k ∈ N , ∀i ∈ V ,
T ik = N U ik ,
ẑ ik|k = ẑ ik|k−1 + tjk ,
We get ∆ (45) and κ (24) from limk→∞ e
−1 ≤δ
and eN Ξk − 1 ≤ (1 + η)δ for all k ≥ κ respectively.
y ik = H ik xk + v ik ,
(49)
3.
N Ξk
xk+1 = F k xk + wk ,
(48)
x̂ik|k−1 = P ik|k−1 ẑ ik|k−1 .
iik = (H ik )T (Rik )−1 y ik ,
I ik = (H ik )T (Rik )−1 H ik ,
(
ii
if k = 1
P
,
uik = 1i
ik − iik−1 + j∈J i Ak [i, j]ujk−1 , if k ≥ 2
k
(
if k = 1
I i1
P
,
U ik =
I ik − I ik−1 + j∈J i Ak [i, j]U jk−1 , if k ≥ 2
Hence, we get maxi∈V |Uki (x) − LkKL (x)| ≤ Ξk for all
k ∈ N, where
k−1
√
b(N −1)
N D1
Ξk = σm
k−1
b(N −1)
√
1 − σm
+
2b(N − 1) N (∆θL + 2εL + εU ) .
1 − σm
4.2
P ik|k−1 = (Z ik|k−1 )−1 ,
where the process noise wk = N (0, Qk ) and the measurement noise v ik = N (0, Rik ) are zero mean multivariate normal distributions.
Therefore, we adopt the information filter-based representation [45,46]. The pseudo-code of the DBFKalman information filtering algorithm for linearGaussian models is given in Algorithm 2. The prior
pdf Ski = N (x̂ik|k−1 , P ik|k−1 ), the posterior pdf
Wki = N (x̂ik|k , P ik|k ), and the estimated pdfs Uki =
N (U ik )−1 uik , (U ik )−1 , Tki = N (T ik )−1 tik , (T ik )−1
are also multivariate normal distributions.
j∈J i
k
i
(Uk,ν
)N
Tki = ´
X
10
i )N dµ(x)
(Uk,ν
.
(54)
5000
4000
Y Axis (m)
Theorem 8 [17,18] Assuming Gk is strongly coni
nected, each agent’s pdf Tk,ν
globally exponentially
C
converges pointwise to Lk (15). After nloop consensus loops, the `2 norm of the error vector ek,ν :=
h
iT
1
N
C
DL1 (Tk,ν
, LC
),
.
.
.
,
D
(T
,
L
)
is bounded by
L
1
k
k,ν
k
√
(nloop −1)
kek,nloop k2 ≤ (σN −1 (Ak ))
2 N.
Benchmark Example
0
0
3
∆2
2
2000
3000
X Axis (m)
Centralized Bayesian Filter
DBF
25
20
15
10
5
0
In this subssection, we compare the performance of
the DBF algorithms with the centralized multi-sensor
Bayesian filtering algorithms using the benchmark example studied in [6,16,47]. The target dynamics is
modeled by a linear (nearly constant velocity) model:
∆3 ∆2
0
0
3
2
1 ∆ 0 0
2
∆
0
0 x + w , where Q = 2 ∆ 0
xk+1 = 00 10 01 ∆
k
k
∆3 ∆2
0
1000
10-1
Time Step ∆ (sec)
10-2
30
Steady-State MSE in Position (m)
Steady-State MSE in Position (m)
30
0
0
4000
5000
Fig. 3. The motion of the target, the position of sensing
agents (5 TOA sensors, 5 DOA sensors, and 40 agents with
no sensors), and their communication network topology are
shown.
Numerical Simulations
0 0 0 1
2000
0
In this section, we demonstrate the properties of the
DBF algorithm using a benchmark example in Section 5.1 and a complex multi-agent estimation and
control task in Section 5.2.
5.1
3000
1000
The proof follows from Theorem 2 and 4 in [17]. Thus, the
distributed estimation algorithm in [14,15] is a special
case of our DBF algorithm.
5
Motion of Target
Agent with no sensor
TOA Sensor
DOA Sensor
Communication Network
Centralized Kalman Filter
DBF-Kalman Information Filter
25
20
15
10
5
0
(a)
10-1
Time Step ∆ (sec)
10-2
(b)
Fig. 4. Variation of steady-state MSE in position with respect
to time step size ∆ is shown for (a) the centralized Bayesian
filtering algorithm and the DBF algorithm in Scenario 1 and
(b) the centralized Kalman filtering algorithm and the DBF
algorithm for linear-Gaussian models in Scenario 2.
2
∆
do not have any sensors. The measurement models for
these sensors are given by:
is the covariance matrix of the process noise wk , ∆ is the
time step size, and the state vector xk denotes the position and velocity components along the coordinate axes,
T
i.e., xk = [xk , ẋk , yk , ẏk ] . As shown in Fig. 3, 50 sensing agents are distributed over the given region and are
able to communicate with their neighboring agents. The
undirected communication network topology is assumed
to be time-invariant. Local-degree weights are used to
compute the doubly stochastic adjacency matrix Ak as:
hik (xk , v ik ) =
(
i
i
i
atan2(x
k − x , yk − y ) + v k,DOA
p
(xk − xi )2 + (yk − y i )2 + v ik,T OA
(55)
for DOA sensor
,
for TOA sensor
where (xi , y i ) denotes the position of the ith agent and
atan2 is the 4-quadrant inverse tangent function. The
DOA sensor’s measurement noise v ik,DOA = N (0, σθ )
has variance σθ = 2◦ and the TOA sensor’s measurement
noise v ik,T OA = N (0, σr ) has variance σr = 10 m.
1
,
∀j ∈ Jki and i 6= j ,
max(di , dj )
X
Ak [i, i] = 1 −
Ak [i, j] ,
Ak [i, j] =
In Scenario 1, each agent executes the DBF algorithm
in Algorithm 1 using particle filters [48] with 104 particles. The comparison between the DBF algorithm and
the centralized Bayesian filtering algorithm for varying
time step sizes (∆) is shown in Fig. 4(a). The same target motion, shown in Fig. 3, is used for all simulations.
We see that the DBF algorithm’s steady-state meansquare-error (MSE) in position converges to that of the
centralized algorithm as the time step size ∆ decreases
(i.e., the steady-state MSE is smaller than 5 m if the
time step size ∆ ≤ 0.05 sec). Note that the MSE of the
j∈V\{i}
where di denotes the degree of the ith agent.
In Scenario 1, five of these agents are equipped with nonlinear position sensors that can measure their distance
to the target using Time of Arrival (TOA) sensors. Another five agents are equipped with Direction of Arrival
(DOA) sensors that can measure the bearing angle between the target and themselves. The remaining agents
11
centralized algorithm does not change much with time
step size because it is constrained by the measurement
noise intensities. This shows that the performance of the
DBF algorithm approaches the performance of the centralized Bayesian filter as the time step size is reduced.
Moreover, Fig. 5 shows that the L1 distances between
the estimated likelihood functions and the joint likelihood function are bounded by δ.
approach to maintain a distance d from their nearest
neighbors, i.e.:
!
i,i
2
(x̂i,j
−
x̂
)
a
d
i,j
i,j
i,i
a rk − i,j ,
AP F (x̂k , x̂k , d) = k i,j k
rk
rk
i,i
where rki,j = kx̂i,j
k − x̂k k2 , and maintain a distance
dCM = 2 cos dπ − π from the estimated center of mass
(2 N)
PN
i,j
1
x̂i,CM
=
j=1 x̂k . In the propagation step of the
k
N
DBF algorithm, the agents use their estimated positions
to estimate the control input applied by other agents.
Therefore, the estimation errors contribute to the process noise in the propagation step. During the fusion step
at k th time instant, the ith agent communicates with the
j th agent if either j ∈ Nki or i ∈ Nkj .
δ (∆ = 0.025 sec)
1
0
0
10
20
30
40
50
60
70
80
90
100
Time (sec)
"
!
DL1 Tki , LC
k
2
δ (∆ = 0.0125 sec)
1
0
0
10
20
30
40
50
60
70
80
90
100
In these simulations, we use a = 0.1, d = 1 m, ∆ = 0.1
sec, and 103 particles to execute each DBF algorithm.
At the start of the estimation process, the particles are
selected from a uniform distribution over the state space
X = [−N, N ] × [−N, N ]. The simulation results for multiple values of N are shown in Fig. 6. Since the agents
only use relative measurements, the orientation of the
final N -sided regular polygon in the global frame is not
fixed. Therefore, we conclude that the N agents successfully estimate their relative positions using the DBF algorithms and achieve the complex desired formations.
Time (sec)
Fig. 5. The trajectories of the L1 distances between the estimated likelihood functions and the joint likelihood function
for the ten sensing agents are shown.
In Scenario 2, the same ten agents (having DOA or
TOA sensors) have linear position sensors hik (xk , v ik ) =
[ 10 00 01 00 ] xk + v ik,lin , with measurement noise v ik,lin =
N (0, Rki ) and covariance matrix Rki = 15I. Here, each
agent executes the DBF-Kalman information filtering
algorithm from Algorithm 2. Fig. 4(b) shows that the
performance of the DBF-Kalman information filtering
algorithm approaches the performance of the centralized Kalman filtering algorithm as the time step size is
reduced.
2
1.5
Agent 1
Agent 2
Agent 3
Agent 4
Polygon
2.5
2
[m]
5.2
3
Agent 1
Agent 2
Agent 3
Polygon
Muti-agent Relative Position Estimation for Formation Control
[m]
"
!
DL1 Tki , LC
k
2
1
1.5
1
0.5
0.5
0
-1
In this subsection, N agents estimate their relative positions using only range measurements, and then reconfigure to a N -sided regular polygon. Specifically, each agent
can only measure the distance to its nearest two neighbors using a TOA sensor, whose measurement model is
described in (55). Each agent simultaneously executes N
DBF algorithms to estimate the relative positions of all
the agents. The ith agent’s dynamics and control inputs
are given by:
-0.5
0
0.5
0
-1.5
1
-1
-0.5
[m]
(a) N = 3
3
2
1
1.5
(b) N = 4
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Polygon
3.5
3
2.5
[m]
[m]
0.5
4
Agent 1
Agent 2
Agent 3
Agent 4
Polygon
2.5
1.5
2
1.5
1
xik+1 = xik + ∆uik ,
X
i,i
i,CM
uik =
AP F (x̂i,j
, x̂i,i
k , x̂k , d) + AP F (x̂k
k , dCM )
0
[m]
1
0.5
0
-1
j∈Nki
0.5
0
1
[m]
(c) N = 4
where Nki denotes the two nearest neighbors of the ith
th
th
agent and x̂i,j
k is the i agent’s estimate of the j agent’s
position, which is obtained using the DBF algorithms.
The agents use the artificial potential field (APF) based
2
0
-2
-1
0
1
2
[m]
(d) N = 5
Fig. 6. The initial position (), the final position ( ), the
trajectories of all the agents, and the final regular polygon
are shown for N = 3, 4, 5 agents.
12
6
Conclusions
[12] N. Ahmed, J. Schoenberg, and M. Campbell, “Fast weighted
exponential product rules for robust general multi-robot data
fusion,” in Robotics: Science and Systems VIII (N. Roy,
P. Newman, and S. Srinivasa, eds.), pp. 9–16, MIT Press,
2013.
In this paper, we presented a novel, discrete-time distributed estimation algorithm, namely the DBF algorithm, that ensures that each agent’s estimated likelihood function converges to an error ball around the
joint likelihood function of the centralized multi-sensor
Bayesian filtering algorithm. We have rigorously proven
the convergence properties of this algorithm. We have
shown an explicit connection between the time step size
of the distributed estimation algorithm and the timescale of the target dynamics. We also presented the DBFKalman information filtering algorithm for the special
case of linear-Gaussian models. The properties of these
algorithms are illustrated using complex numerical examples. We envisage that the novel proof techniques presented in this paper can also be used in other distributed
estimation algorithms which rely on the LogOP scheme.
[13] C. S. R. Fraser, L. F. Bertuccelli, H.-L. Choi, and J. P. How,
“A hyperparameter consensus method for agreement under
uncertainty,” Automatica, vol. 48, no. 2, pp. 374 – 380, 2012.
[14] O. Hlinka,
M. Rupp,
distributed
vol. 60, no.
O. Sluc̆iak, F. Hlawatsch, P. M. Djuric, and
“Likelihood consensus and its application to
particle filtering,” IEEE Trans. Signal Process.,
8, pp. 4334–4349, 2012.
[15] O. Hlinka, F. Hlawatsch, and P. M. Djuric, “Consensusbased distributed particle filtering with distributed proposal
adaptation,” IEEE Trans. Signal Process., vol. 62, no. 12,
pp. 3029–3041, 2014.
[16] G. Battistelli and L. Chisci, “Kullback–Leibler average,
consensus on probability densities, and distributed state
estimation with guaranteed stability,” Automatica, vol. 50,
no. 3, pp. 707–718, 2014.
[17] S. Bandyopadhyay and S.-J. Chung, “Distributed estimation
using Bayesian consensus filtering,” in Proc. Amer. Control
Conf., (Portland, OR), pp. 634–641, June 2014.
References
[18] S. Bandyopadhyay and S.-J. Chung, “Distributed estimation
using Bayesian consensus filtering.” https://arxiv.org/
abs/1403.3117.
[1] J. Speyer, “Computation and transmission requirements for
a decentralized linear-quadratic-Gaussian control problem,”
IEEE Trans. Autom. Control, vol. 24, no. 2, pp. 266–269,
1979.
[19] M. H. DeGroot, “Reaching a consensus,” J. Amer. Statistical
Assoc., vol. 69, no. 345, pp. 688 – 704, 1960.
[2] V. Borkar and P. Varaiya, “Asymptotic agreement in
distributed estimation,” IEEE Trans. Autom. Control,
vol. 27, no. 3, pp. 650 – 655, 1982.
[20] M. Bacharach, “Normal Bayesian dialogues,” J. Amer.
Statistical Assoc., vol. 74, no. 368, pp. 837 – 846, 1979.
[21] S. French, “Consensus of opinion,” European J. Operational
Research, vol. 7, pp. 332 – 340, 1981.
[3] L. Chen, P. O. Arambel, and R. K. Mehra, “Estimation under
unknown correlation: covariance intersection revisited,”
IEEE Trans. Autom. Control, vol. 47, no. 11, pp. 1879–1882,
2002.
[22] C. Genest and J. V. Zidek, “Combining probability
distributions: A critique and an annotated bibliography,”
Statistical Sci., vol. 1, no. 1, pp. 114 – 135, 1986.
[4] M. Kamgarpour and C. Tomlin, “Convergence properties
of a decentralized Kalman filter,” in IEEE Conf. Decision
Control, pp. 3205–3210, IEEE, 2008.
[23] G. L. Gilardoni and M. K. Clayton, “On reaching a consensus
using DeGroot’s iterative pooling,” Ann. Stat., vol. 21, no. 1,
pp. 391 – 401, 1993.
[5] R. Olfati-Saber, “Kalman-consensus filter : Optimality,
stability, and performance,” in IEEE Conf. Decision Control,
(Shanghai, China), pp. 7036–7042, December 2009.
[24] R. Olfati-Saber and R. Murray, “Consensus problems in
networks of agents with switching topology and time-delays,”
IEEE Trans. Autom. Control, vol. 49, no. 9, pp. 1520 – 1533,
2004.
[6] G. Battistelli, L. Chisci, G. Mugnai, A. Farina, and
A. Graziano, “Consensus-based linear and nonlinear
filtering,” IEEE Trans. Autom. Control, vol. 60, no. 5,
pp. 1410–1415, 2015.
[25] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of
groups of mobile autonomous agents using nearest neighbor
rules,” IEEE Trans. Autom. Control, vol. 48, no. 6, pp. 988
– 1001, 2003.
[7] M. Rashedi, J. Liu, and B. Huang, “Communication delays
and data losses in distributed adaptive high-gain ekf,” AIChE
Journal, vol. 62, no. 12, pp. 4321–4333, 2016.
[26] A. Olshevsky and J. N. Tsitsiklis, “Convergence speed in
distributed consensus and averaging,” SIAM Journal on
Control and Optimization, vol. 48, no. 1, pp. 33–55, 2009.
[8] G. Pavlin, P. Oude, M. Maris, J. Nunnink, and T. Hood,
“A multi-agent systems approach to distributed Bayesian
information fusion,” Inform. Fusion, vol. 11, pp. 267–282,
2010.
[27] M. Zhu and S. Martı́nez, “Discrete-time dynamic average
consensus,” Automatica, vol. 46, no. 2, pp. 322–329, 2010.
[28] Z. Chen, “Bayesian filtering: From Kalman filters to particle
filters, and beyond,” Statistics, vol. 182, no. 1, pp. 1–69, 2003.
[9] A. Jadbabaie, P. Molavi, A. Sandroni, and A. TahbazSalehi, “Non-Bayesian social learning,” Games and Economic
Behavior, vol. 76, pp. 210–225, 2012.
[29] J. Pearl, Probabilistic Reasoning in Intelligent Systems:
Networks of Plausible Inference. San Mateo, CA: Morgan
Kaufmann, 1988.
[10] A. Nedić, A. Olshevsky, and C. A. Uribe, “Fast convergence
rates for distributed non-Bayesian learning,” arXiv preprint
arXiv:1508.05161, 2015.
[30] H. Durrant-Whyte and T. C. Henderson, Springer Handbook
of Robotics, ch. Multisensor Data Fusion, pp. 585–610.
Springer, 2008.
[11] T. Bailey, S. Julier, and G. Agamennoni, “On conservative
fusion of information with unknown non-Gaussian
dependence,” in Int. Conf. Information Fusion, (Singapore),
pp. 1876–1883, July 2012.
[31] N. Gordon, B. Ristic, and S. Arulampalam, Beyond the
Kalman Filter: Particle Filters for Tracking Applications.
Artech House, London, 2004.
13
A
[32] M. S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp, “A
tutorial on particle filters for online nonlinear/non-Gaussian
Bayesian tracking,” IEEE Trans. Signal Process., vol. 50,
pp. 174–188, February 2002.
Proof of Lemma 1
If this claim is untrue, then either 0 < P(x) < Q(x)
or
´ 0 < Q(x) < P(x)
´ for all x ∈ X
´ . Hence either
P(x)dµ(x)
=
1
<
Q(x)dµ(x)
or
Q(x)dµ(x) <
X
X
´X
P(x)dµ(x)
=
1,
which
results
in
contradiction
since
´X
Q(x)dµ(x)
=
1.
Hence,
such
a
ψ
∈
X
must
exist.
X
[33] O. Lebeltel, P. Bessiere, J. Diard, and E. Mazer, “Bayesian
robot programming,” Autonomous Robots, vol. 16, pp. 49–
79, January 2004.
[34] M.-H. Chen, “Bayesian computation: From posterior
densities to Bayes factors, marginal likelihoods, and posterior
model probabilities,” in Bayesian Thinking, Modeling and
Computation (D. K. Dey and C. R. Rao, eds.), Handbook of
Statistics, ch. 15, pp. 437 – 457, Amsterdam: Elsevier, 2005.
B
Proof of Lemma 2
Since limk→∞ Pki (x) = P ? (x),
we have
limk→∞ ln Pki (x)−ln Pki (ψ) = ln P ? (x) − ln P ? (ψ).
From Lemma 1, substituting limk→∞ Pki (ψ) = P ? (ψ)
gives limk→∞ Pkj (x) = P ? (x) since logarithm is a
monotonic function.
[35] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge,
England: Cambridge University Press, 1985.
[36] S.-J. Chung, S. Bandyopadhyay, I. Chang, and F. Y.
Hadaegh, “Phase synchronization control of complex
networks of Lagrangian systems on adaptive digraphs,”
Automatica, vol. 49, pp. 1148–1161, May 2013.
[37] S. Bandyopadhyay, S.-J. Chung, and F. Y. Hadaegh,
“Probabilistic and distributed control of a large-scale swarm
of autonomous agents,” IEEE Trans. Robotics, vol. 33,
pp. 1103–1123, Oct. 2017.
C
Proof of Lemma 3
It follows from Scheffé’s theorem [49, pp. 84] that if the
pdfs converge pointwise, then their induced measures
converge in TV. The relationship between TV error and
L1 distance follows from [50, pp. 48].
[38] M. Fiedler, “Bounds for eigenvalues of doubly stochastic
matrices,” Linear Algebra and Its Applications, vol. 5, no. 3,
pp. 299–310, 1972.
[39] A. Stuart and A. R. Humphries, Dynamical Systems
and Numerical Analysis. Cambridge, England: Cambridge
University Press, 1998.
[40] G. Kramer and S. A. Savari, “Communicating probability
distributions,” IEEE Trans. Inf. Theory, vol. 53, pp. 518–
525, February 2007.
[41] J. H. Kotecha and P. M. Djuric, “Gaussian sum particle
filtering,” IEEE Trans. Signal Process., vol. 51, pp. 2602–
2612, Oct. 2003.
[42] B. D. O. Anderson and J. B. Moore, Optimal Filtering.
Mineola, New York: Dover Publications, 2005.
[43] G. J. McLachlan and K. E. Basford, Mixture Models:
Inference and Applications to Clustering. New York, N.Y.:
M. Dekker, 1988.
[44] D. A. Reynolds, “Gaussian mixture models,” Encyclopedia
of Biometric Recognition, February 2008.
[45] A. G. O. Mutambara, Decentralized Estimation and Control
for Multisensor Systems. CRC press, 1998.
[46] H. Fourati, Multisensor Data Fusion: From Algorithms
and Architectural Design to Applications. Series: Devices,
Circuits, and Systems, CRC Press, Taylor & Francis Group
LLC, 2015.
[47] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation
with Applications to Tracking and Navigation: Theory,
Algorithms and Software. John Wiley & Sons, 2004.
[48] D. A. A. Marı́n, “Particle filter tutorial.” (http://www.
mathworks.com/matlabcentral/fileexchange/
35468-particle-filter-tutorial)
[retrieved
on
13
December 2017].
[49] R. Durrett, Probability: Theory and Examples. Thomson
Brooks, 2005.
[50] D. A. Levin, Y. Peres, and E. L. Wilmer, Markov Chains
and Mixing Times. American Mathematical Soc., 2009.
14
| 7cs.IT
|
arXiv:cs/0609092v1 [cs.PL] 16 Sep 2006
Analysis of Equality Relationships
for Imperative Programs
Pavel Emelianov∗
Institute of Informatics Systems,
6 avenue Lavrentiev
630090 Novosibirsk, Russia
[email protected]
March 1, 2018
Abstract
In this article, we discuss a flow–sensitive analysis of equality relationships
for imperative programs. We describe its semantic domains, general purpose operations over abstract computational states (term evaluation and identification,
semantic completion, widening operator, etc.) and semantic transformers corresponding to program constructs. We summarize our experiences from the last
few years concerning this analysis and give attention to applications of analysis of
automatically generated code. Among other illustrating examples, we consider a
program for which the analysis diverges without a widening operator and results
of analyzing residual programs produced by some automatic partial evaluator. An
example of analysis of a program generated by this evaluator is given.
Keywords: abstract interpretation, value numbering, equality relationships for
program terms, formal grammars, semantic transformers, widening operator, automatically generated programs.
Introduction
Semantic analysis is a powerful technique for building effective and reliable programming systems. In [17, 18, 16] we presented a new kind of a semantic flow–sensitive
analysis designed in the framework of abstract interpretation [9, 10, 11]. This analysis which determines an approximation of sets of invariant term equalities t1 = t2 was
called the analysis of equality relationships for program terms (hereinafter referred to as
ERA).
Most traditional static analyses of imperative programs are interested in finding the
(in)equalities of a specific kind (so–called value analyses; only they are discussed here)
∗ This work was partly done when the author was in Laboratoire d’informatique, Ecole polytechnique
(Palaiseau, France) and Ecole normale supérieure d’ingenieur (Bourges, France).
1
describing regular approximations (i.e. they have simple mathematical descriptions and
machine representations) of sets of values: convex polyhedrons/octahedrons/octagons
[14, 27, 36, 8], affine [32, 25] and congruent [24, 34] hyper-planes, their non-relational
counterparts [9, 43, 23, 35] as well, etc. They are carefully designed to be reasonable
(i.e. they express non-trivial semantic properties) and effectively computed (i.e. there
are polynomial algorithms to handle them1 ) but “regular” nature does not allow them
to treat well programs with irregular control/data-flows. Hence of special interest is
investigations of approximations based on sets of terms which can have potentially
arbitrary nature, i.e. they could be powerful (due to their irregularity) but effectively
computed. One well known example is the set–based analysis [29, 12].
In our case, terms represent all expressions computed in programs. This enables the
analysis to take into account different aspects of program behavior in a unified way. A
such unified treatment of all semantic information allows the analysis to improve its
accuracy. This does not mean that ERA is a generalization of all other value analyses
(except the constant propagation one), because they use different approaches (semantic
domains and transformers) to extract effectively and precisely the limited classes of
semantic properties. In general, the results of the analyses are not comparable.
ERA provides interesting possibilities for gathering and propagating different invariant information about programs in a unified way. This information can be used
both for verification and optimization purposes. The second is especially interesting for
automatically generated programs: residual, i.e. obtained in the process of the partial
evaluation, and synthesized from high–level specifications. Due to nature of automatic
generation processes, such programs have specific control flows (for example, hierarchy
of nested conditional statements with specific conditions; in the case of residual programs this hierarchy is more deep as “degree” of the partial evaluation increases) that
can be successfully optimized on base of gathered invariant information.
Besides the peculiarity of ERA mentioned above, let us discuss some common properties of the semantic analyses. Such taxonomic properties of the analysis algorithms
as the attribute (in)dependence, context (in)sensitivity, flow (in)sensitivity, scalability
and some other properties are well known. However, it is the author’s opinion that a
notion of “interpretability of a semantic analysis” has not been considered adequately
yet. Here the interpretability of analysis means how extensively the properties of primitive operations of the language (arithmetical, logical, etc.) and type information are
allowed for analyzing and can be handled when the analysis works.
One extreme point of view on the interpretability is an approach accepted in the
“pure” program scheme theory where no interpretations of functional symbols or type
information are allowed2 . Unfortunately, the results obtained under this approach are
not reasonably strong. Nevertheless, it must be underscored that ERA dates back to
V. Sabelfeld’s works in the program scheme theory [39, 40]. Another extreme leads to
the complete description of the program behavior that is also not workable. Obviously,
it is closely allied to its flow sensitivity (ignoring some part of semantic information
1 An example of this sort is an approximation of value sets by conic shapes that has only one “computational disadvantage”: semantic transformers involve algorithms known to be N P -hard. Proposed
in [4] many years ago it did not gain ground.
2 More precisely, there exist some works in the program scheme theory where some semantic interpretations of functional symbols (like commutativity of superposition f ◦ g ≡ g ◦ f , etc.) are considered.
2
does not allow us to treat precisely some control flow constructs of analyzed programs)
and its scalability (attempts to take into account large quantity of semantic properties,
for example, using some theorem prover which is invoked while an semantic analyzer
works and deduces new properties, can lead to combinatorial explosion for abstract
computational states).
It is possible that the interpretability has not been highlighted enough, because most
of analysis algorithms take into account the limited classes of primitive operations and
type information and they cannot be enriched in some natural way. For example, an
interval analysis is not able to incorporate congruence properties in some natural way,
etc3 . Essentially another case is ERA where we have a choice to handle expressiveness
of the analysis. We intend to illustrate the notion of “interpretability of analysis”, its
importance and usefulness on this example of analysis.
Among the analyses closely related to ours we would like to point out the following.
A semantic analysis for detection of equalities between program variables (and simple
relationships among them) was described in [2]. It makes a list of sets of variables
discovered to be equal by using the Hopcroft’s partitioning algorithm for finite–state
automata. This algorithm being quite efficient is not however precise enough. Further
value-numbering techniques were developed in [38, 21, 26]. These algorithms demonstrate that adequacy of value numbering is resource-consuming. For example, in the
last case time complexity of the algorithm is O(k 3 jN ) where k is a number of program
variables, j is a number join-points, and N is program size.
Another important example is the set-based analysis [29] mentioned above. Here
approximating sets of terms are found with resolving some system of set-theoretical
equations. Formal grammars were used for an analysis of recursive data structures of
functional languages (see, for example, [31]). Formal languages were applied to coding
of memory access paths in [15, 42] and values of program variables in the set-based
analysis. [12] established common foundations connecting and generalizing different
approaches using formal languages to represent semantic properties. Of course, we
should mention techniques from the automatic proof theory and the term rewriting
theory which can be widely applied both at the analysis stage to improve its accuracy
and the post-processing stage to present its results to the user.
This article is organized as follows: In Sections 1.1 and 1.2 we describe the semantic properties, concrete and abstract, respectively, which are considered in ERA.
In Section 1.3 we discuss some basic operations over the semantic properties used to
define the semantic transformers which are next presented in Section 2. In Section
3 we consider a widening operator and the complexity of ERA is discussed. Finally,
Section 4 describes processing of ERA invariants and presents some results of our experiments with ERA. In Appendix an example of analysis of some residual program
is considered.
3 Of course, it is possible to use sophisticated approaches for combinations of analyses but it introduces complicated problems under implementations and it is not an enrichment of original ones.
3
1
Properties of interest
1.1
Concrete properties
A usual choice for the description of the operational semantics is a specification of some
transition relationship on the pairs <control point, state of program memory> (see, for
example, [30, 22]) where the states of program memory are described by mapping the
cells of memory into a universe of values. Here variables (groups of cells) and their values
(constants) are in asymmetric roles. Another example of “asymmetry”: manipulations
over the structured objects of programs (arrays, records etc.) are not so transparent
as over the primary ones. To describe the operational semantics for ERA, we used
another approach. All objects of a program are considered to be “identical” in the
following meaning.
Let CV be a set of 0–ary symbols representing variables and constants. The last ones
may be of the following kinds: scalars, compositions over scalars (i.e., constant arrays,
records, etc.), names of record fields, and indefiniteness. Let FP be a set of n–ary
(functional) symbols which represent primitive operations of programming languages:
arithmetic, logic, type casting, and all the kinds of memory addressing, as well. Let TRS
be a set of well-formed terms over CV and FP, hereinafter referred to as program terms.
They represent expressions computed during execution of a program. So, as a state
of program memory we take a reflexive, symmetrical, and transitive relationship (i.e.,
equivalence relationship) over TRS. The relationship defines some set of term equalities
which we use to describe the operational semantics and call a computation state.
Suppose that the following code
var x,i,j : integer; a : array [1..3] of integer={1,2,3};
...
i := 3;
j := i-1;
if odd(x) then
i := i mod j;
j := 1;
else
j := a[i];
a[i] := a[1];
a[1] := j
end
...
Example 1
is executed at least twice for the different parities of the variable x. In Table 1 static
semantics for five control points is given. We present a minimum4 subset of term
equalities concerning dynamic behavior of the piece. We shall use the property π (see
Table 1) to illustrate our further reasoning.
4A
set of equalities can be completed with any number of consistent equalities.
4
then-branch
else-branch
(
a[1]=1,a[2]=j=2,a[3]=i=3,
entry
a= 1 2 3 ,ODD(x)=TRUE
exit
(
a[1]=i=j=1,a[2]=2,a[3]=3,
a= 1 2 3 ,ODD(x)=TRUE
(
a[1]=1,a[2]=j=2,a[3]=i=3,
a= 1 2 3 ,ODD(x)=FALSE
)
(
a[1]=i=j=3,a[2]=2,a[3]=1,
a= 3 2 1 ,ODD(x)=FALSE
)
)
)
exit of if–statement
π=
) (
)
(
a[1]=i=j=1,a[2]=2,a[3]=3,
a[1]=i=j=3,a[2]=2,a[3]=1,
a=
1
2
3
,ODD(x)=TRUE
3
2
1
,ODD(x)=FALSE
a=
,
Table 1: Description of collecting semantics for Example 1 (here the constant
c1 c2 c3 represents constant arrays).
5
Equality
relationships
a = h(y)
x = f (a, y)
x = f (h(y), y)
f (a, y) = f (h(y), y)
Grammar
representation
S → A1 = A1 |A2 = A2 |A3 = A3
A1 → x|f (A2 , A3 )
A2 → a|h(A3 )
A3 → y
Functional net
representation
✏
✓
x
f
✒ ✓❏ ✑
✓
❏
✓
❏
❏
❫
✓
✏✗✔
✓ ✴
✲
y
a
h
✒
✑✖✕
Figure 1: Semantic properties and their representations. The set of equalities does not
contain trivial equalities like x = x and equalities given by the symmetry of the equality
relationship.
Formally it is described as follows. Let EQS be a set of all equalities of the terms
from TRS, i.e., EQS = {t1 = t2 | t1 , t2 ∈ TRS}. A set S ∈ ℘(EQS) is a computation
state interpreted in the following way. For each equality ti = tj ∈ S values of the
expressions represented by ti and tj must be equal at this point for some execution
trace. We take the set ℘(℘(EQS)) as a set for a concrete semantic domain describing
the collecting semantics of ERA. So, an element of the concrete semantic domain is a
set of computation states. For a particular point in a particular program it is a set of
computation states each of them corresponds to some execution trace in the program
that reaches this point.
Properties considered in ERA are presented by means of context–free grammars
G = (N, T, P, S) where N is a finite set of nonterminals denoted by capital letters,
S ∈ N is the initial symbol of the grammar G, T = CV ∪ FP ∪ { (, ), =, } is a finite
set of terminal symbols, and P is a finite set of grammar rules. We do not give their
precise description because we shall use quite simple machinery from formal languages
theory that is not an object of considerations itself and serves for demonstrations. We
could use functional nets language as well but it is rather machine–oriented and is
not widely used. We expect that all these descriptive ways become apparent from the
examples on Figure 1 and further ones.
In that way, a state of computation is represented by a language L(G) generated by
some grammar G of the described form.
If for A ∈ N
:
+
+
A⇒
t1 ∧ A⇒
t2 , i.e. t1 = t2 ∈ L(G), we say that the nonterminal A and the lanG
G
guage L(G) know the terms t1 , t2 . Obviously, such a grammar representation has
some superfluous “syntactic sugaring”. We can use S → A rules only and say about
A–nonterminals as classes of equal values.
Evidently we may suppose that the set of rules P does not contain rules having
identical right parts. It is convenient to consider the grammars which do not contain
useless and redundant nonterminals and rules. A rule is useless if it produces only
one term (the language knows only a trivial equality like t = t) and this term is not an
argument of other terms. A nonterminal is useless if it does not participate in derivations
,
6
of sentential forms [1]. Such grammars can arise as result of operations on grammars.
If nonterminals and rules are useless or redundant it is possible to remove them (see
Lemma 2). This operation called state reduction consist in detecting and removing
a set of useless/redundant nonterminals/rules. They are revealed with the well–known
incremental markup algorithms (see, for example, [1]).
For illustrating purposes we shall use special functional nets for these grammars.
Here nonterminals are represented by ovals containing 0–ary and functional symbols
from right parts of rules. Arcs from functional symbols to ovals represent argument
dependencies ordered from left to right (see Figure 1).
1.2
Abstract properties
It is an interesting peculiarity of ERA that abstract (i.e. approximate) properties have
the same nature as the computation states of the operational semantics. Formally this
approximation is defined as follows.
1.2.1
Functions of abstraction and concretization
Given a concrete property π ∈ ℘(℘(EQS)) and an abstract property π̃ ∈ ℘(EQS), the
abstraction function α : ℘(℘(EQS)) → ℘(EQS) and the concretization one γ : ℘(EQS) →
℘(℘(EQS)) are defined in the following way
if π = ∅,
EQS,
\
α(π) =
and
γ(π̃) = ∪{ π | α(π) ⊇ π̃ }
S
otherwise
S∈π
where ∪ is the set–theoretical union on ℘(℘(EQS)) and ⊇ and ∩ are the set–theoretical
inclusion and intersection of the languages (i.e. on ℘(EQS)) respectively. We take the
empty language as infimum ⊥′ (there are no computed expressions) of the semi-lattice
of abstract semantic properties. The supremum ⊤′ (an inaccessible computation state)
is the language containing all possible equalities of program terms. Also ⊆ is the set–
theoretical inclusion on ℘(℘(EQS))
Lemma 1 The abstraction function α is monotonic.
Proof The function α is monotonic iff ∀π1 , π2 ∈ ℘(℘(EQS)) : π1 ⊆π2 ⇒
α(π1 )⊇α(π2 ). Because π1 ⊆π2 , then π2 = π1 ∪(π2 \ π1 ). So, we have α(π1 )⊇
α(π1 ∪(π2 \ π1 )) = α(π1 ) ∩ α(π2 \ π1 ) = α(π1 ) ∩ (∩S∈π2 \π1 S). Q.E.D.
γ(π̃) is the most imprecise element of ℘(℘(EQS)) that can be soundly approximated
by π̃ ∈ ℘(EQS). For the example of Table 1, the best approximation of the concrete
property π is
α(π) =
a[1] = i = j, a[2] = 2 .
7
(∗)
✎
☞
g
✍
✌
✡ ✡ ❙
L1
✡ ✡
❙
✡ ✡
❙
✢ ☞
✢ ✡
✎ ✡
✎ ✇
❙
f
✍
z
f
✂
✌
✍
x
a
✎
✁✻
g
✎
✍
f
✂
☞
✍
✌
✡ ✡ ❙
L2
✡ ✡
❙
✡ ✡
❙
✢ ☞
✢ ✡
❙
☞ ✎ ✡
✎ ✇
y
x
a
f ✛ f
✌ ✍
✌
✍
✂
✁✻
y
☞
✍
✌
✡ ✡ ❙
✡ ✡
❙
✡ ✡
❙
✢ ☞ ✎✇
✡
✢ ✡
✎
❙
f
g
f
✌ ✍
a
✎
✍
❈
✌ ✍
✁✻
✌
❈
❈
☞ ✎ ❈❲
y
☞
z
f
x
☞
z
☞
✌
L1 ∩L2
✌
Figure 2: Intersection of computation states.
1.2.2
Intersection of ERA–languages
Finding of the intersection of context–free languages is an undecidable problem in the
general case. In our case, for the languages of term equalities, an algorithm exists (see
an example at Fig. 2). It is similar to constructing a Cartesian product of automata.
Algorithm. Intersection of two languages of term equalities.
Input: grammars G1 = (N1 , T, P1 , S1 ) and G2 = (N2 , T, P2 , S2 ).
Output: grammar G = (N, T, P, S) such that L(G) = L(G1 ) ∩ L(G2 ).
Description:
1. Let N = {hN1 , N2 i|N1 ∈ N1 & N2 ∈ N2 } = N1 × N2 .
2. The set of rules P is defined as follows:
• The rule hN1 , N2 i → t is introduced if and only if t ∈ CV & N1 → t ∈
P1 & N 2 → t ∈ P2 .
• The rule hN1 , N2 i → t(hN11 , N21 i, . . . , hN1k , N2k i) is introduced if and only if
t ∈ FP & N1 → t(N11 , . . . , N1k ) ∈ P1 & N2 → t(N21 , . . . , N2k ) ∈ P2 .
3. Add rules S → hN1 , N2 i = hN1 , N2 i for the initial nonterminal S of G and for all
N1 ∈ N1 \ {S1 }, N2 ∈ N2 \ {S2 }.
4. Apply state reduction.
The described algorithm of intersection is very “naı̈ve” and impractical. To improve
it we should choose a more efficient strategy for generating functional symbols. To do
this we first do a topological sorting of the functional symbols appearing in the right
8
parts of rules; intersect the 0–ary symbol sets; and then generate the next functional
symbol in conformity with its topological order if and only if all arguments of this
symbol already exist in the new grammar. For practical cases this intersection can
be done in an linear average time with respect to grammar size and linear space. It
demands quadratic time in the worst case.
1.3
Operations over semantic properties
Now we shall discuss some basic operations over semantic properties
{| }| : ℘(EQS) → ℘(EQS) used to define the semantic transformers of ERA. For all
{| }| the notation L{| }| {| }| means (L{| }| ){| }| .
1.3.1
Removing terms
Operations over abstract computation states use certain common transformation of the
sets of term equalities which consists in removing some subset L′ . The following
statement holds.
Lemma 2 Removing any subset of term equalities preserves correctness of an approximation.
Proof It easy to see that
∪{ π | α(π) ⊇ π̃ } = γ(π̃)
⊆
γ(π̃ \ L′ ) = ∪{ π | α(π)
\ ⊇ (π̃ \ L) } =
∪{ π |
(L∪L′ ) ⊇ π̃ },
L∈π
which states that removing term equalities makes the approximation more rough but it
does preserve its correctness. Q.E.D.
For (∗), for example, γ({a[1] = i = j, a[2] = 2})⊆γ({a[1] = i}).
We shall write L{| ↓ t|} and L{| ↓ T }| for single term and term set removing followed
by the state reduction operation defined above.
1.3.2
Term evaluation
We define the abstract semantics for an evaluation of a term in an abstract computation state. The result of the term evaluation is a state knowing the evaluated
term.
Term evaluation L{|t|}.
1. If t = t ∈ L, then L{|t|} = L.
2. Otherwise, if t ∈ CV, then add the new rules S → A = A and A → t to the
grammar G, where A is a nonterminal which does not exist in G.
9
✗
✔
f
✗
x
h
✖
✗ ✌✂
y
✖
✂
✔
✂
✂
✕
a
✔
✕
✖
✑ ◗ ✕
◗
✑
◗
✑
◗
✑
◗
✑
◗
✑
✗✰ ✔
✗
✔
✗
s
◗
g
x
h
f
✔
✖ ✕
✖
✕
✖ ✕
✂ ✻
✻
✍
✌
✍
✌
✂
✂
✗❄ ✔
✔
✗ ✂✌
✍ ✲ y
a
b
✖
✕
✖ ✕
L{|f (g(a, x), f (h(y), b))|}
L
Figure 3: Term evaluation.
3. Otherwise, if t = f (t1 , . . . , tn ) where f ∈ FP is a functional n–ary symbol and the
+
+
tn ),
t1 , . . . , An ⇒
sub-terms t1 , . . . , tn have been calculated (i.e. there exist derivations A1 ⇒
G
G
then add the new rules S → A = A and A → f (A1 , . . . , An ) to the grammar G,
where A is a nonterminal which still does not exist in G.
To improve an accuracy of analysis we can take into account, for example, the commutativity of primitive operations. If L knows f (t1 , t2 ) and f is commutative then
L{|f (t2 , t1 )|} = L.
1.3.3
Identification of terms
When the standard semantics defines that during program execution values of computed
expressions are equal, we can incorporate this information in the computation state (but
it can also be left out). Identification of terms transforms the state into a new one
incorporating this information. For example, we know that a value of a term representing
the conditional expression of an IF-statement coincides with 0-ary terms representing
the constants TRUE or FALSE when, respectively, THEN-branch or ELSE-branch
is being executed. So, identification of terms along with semantic completion considered
below provides powerful facilities to take into account real control flow in programs.
Identification of terms L{|t1 ≡ t2 }| .
1. If t1 = t2 ∈ L then L{|t1 ≡ t2 }| = L.
+
+
t2 . We replace the nonterminal A2 by the nonterminal A1
t1 and A2 ⇒
2. Let A1 ⇒
G
G
in all rules of P. If rules with an identical right side B1 → w, . . . , Bk → w
have appeared, then a certain nonterminal from the left sides of the rules (for
example B1 ) must be taken and all nonterminals B2 , . . . , Bk in the grammar must
be replaced by it.
10
✛
g
a
✚
✓ ❅
❅
✓
✴
✓
❘
❅
✛
✛✘
✘
✛
g
a
✚
❙
❙
❄ ❙✇
✛
✙
✘
✘
✙
✘
☛✲ g
f
✚
✙
✍
✌
❄ ❄ ✘
✛
g
f ✛
f
✚✙
✚
✙
✂
❇
✂
❇
❇◆ ✂✌
❄
✛✘
✛✘
x
b
✚✙✚✙
x
✚
b
L{|x ≡ b|}
L
✙
Figure 4: Identification of values of terms.
3. Repeat step 2 until stabilization. If after that we have a state L′ containing
inconsistent term equalities5 then the result is ⊤′ else it is a reduction of L′ .
An example of identification is given in Figure 4.
Lemma 3 Identification of values of terms is a correct transformation and the resulting
state is unique.
Proof
Let π = { Li | Li ∈ ℘(EQS) } be a concrete semantic property which holds before
identification of terms t1 and t2 . If the values of t1 and t2 are equal in the concrete semantics ∀Li : t1 = t2 ∈ Li , then they are equal in the abstract semantics t1 = t2 ∈ α(π),
too. If their values are not equal, then identification gives us an inconsistent computation state which obviously includes L{|t1 ≡ t2 }| for all t1 and t2 . So, this transformation
is correct.
Identification is done in finite steps because the size of grammar decreases at each
step. Uniqueness of the resulting state is explained by the following observation. If we
have two pairs of terms which are candidates for identification, then identification of one
of them does not close a possibility of it for another, because we remove a duplication
of the functional symbols only. In fact, after identification of a pair of terms we obtain
a new state, including the source one, and thus other existing identification possibilities
remain. So, the order of “merging” of term pairs is not important for the resulting
state. Q.E.D.
1.3.4
Semantic completion
We have not yet considered any interpretation of constants and functional symbols.
We could continue developing ERA in the same way. As a result we shall obtain a
5 There exists a wide spectrum of inconsistency conditions. The simplest of them is an equality of
two different constants (see Section 1.3.4 for the further discussion).
11
✗
x
✖
✔
✗
y
∗
✕ ✖ ✚
✻
✚
✡
✠
✚
✚
✚
✚
✗❄ ✔
✗❄
✚
❂
✚
z
0
✖
✕
✖
+
✔
✕
✔
✕
L
✓✏
❄
✔
✗
✛✟
x
z
+
✖❇
✕
❇
✒✑
❇
❇
✗ ◆❇
✔
y
∗
0
✖
✕
✻
✍ ✌
✒ ✑
Co(L)
Figure 5: Semantic completion.
noninterpretational version of the analysis likewise analysis algorithms in the program
scheme theory. However, it is natural to use semantics of primitive operations of the
programming language of interest in order to achieve better accuracy.
ERA provides us with possibilities of taking into account properties of language
constructs, and, what is especially important, we can easily handle complexity of these
manipulations. In fact, inclusion of these properties corresponds to carrying out some
finite part of completion of the computation states by consistent equalities. This manipulation is called semantic completion (about “conjunctive/disjunctive completion”
see [11]).
As a basic version of semantic completion Co we take computations over constant
and equal arguments. When we detect that some term t has some specific value v then
Co(L)=L{|t ≡ v|}. Also, it is possible to apply identification involving dependencies
among result of an operation and its arguments: if (t1 AND t2 ) = T RUE ∈ L then
Co(L)=L{|t1 ≡ T RUE|}{|t2 ≡ T RUE|} etc. This identification process is iterative because new possibilities for identification can appear at the next steps. It is conceivable
that in doing so we shall detect inconsistency of a computation state. In this case result
of semantic completion is ⊤′ .
This version can be extended by some intelligent theorem prover inferring new reasonable equalities and checking inconsistency of computation states. Such combining of
analysis with proofs offers powerful facilities to the analyzer (see [28, 13]) and, as it was
mentioned in the previous works, this prover is reusable for consequent (semi-)automatic
processing of results of the analysis. “Size” of used completion can be tuned by options
of interpretability of the analyzer.
Some arithmetical errors (such as division by zero, out of type range etc.) will appear
during the semantic completion. In this case the analyzer tells us about the error and
sets the current computation state to ⊤′ . Notice that for the languages where incomplete
Boolean evaluation is admissible semantic completion over Boolean expressions should
be carefully designed especially in presence of pointers.
An example of semantic completion is presented in Figure 5. Turning back to the
identification example at Figure 4, we consider the following interpretation of constants
12
and functional symbols: g is the exclusive disjunction, f is the negation, a is the constant
TRUE and b is the constant FALSE. It is easy to see that in this case application of
semantic completion gives us L{|x ≡ b|} = ⊤′ .
In our analyzer we implemented an interpretational version of ERA which uses
semantic completion Co(L). Under this approach, the definitions of the basic transformations mentioned above are changed to the following:
′
L{|t|} = Co(L{|t|}),
′
L{|t1 ≡ t2 }| = Co(L{|t1 ≡ t2 }| ).
In short, we shall omit this “interpretability” prime.
2
Semantic transformers
In this section we describe semantic transformers [[]] : ℘(EQS) → ℘(EQS) corresponding
to common statements existing in imperative programming languages. If for a statement
S an input computation state is L then L[[S]] is its output computation state. For all
[[]] the notation L[[]][[]] means (L[[]])[[]].
2.1
Assignment statement
Among all program terms considered in ERA we can pick out access program terms
including array elm, record fld, and pointer val referencing and playing an important
role in determination of effect of an assignment statement. As in [15, 42] our abstraction
of program memory manipulations is storeless and based on notion of memory access
paths represented by access program terms. For example, for an address expression
bar[i][j]ˆ.foo the access term is fld(val(elm(elm(bar, i), j)), f oo).
We shall assume that no operations other than memory addressing (for example comparisons) are allowed for structured variables such as arrays and records. So, for the previous
example
neither
a
and
elm(a, i)
(a[i])
nor
val(elm(elm(a, i), j)) (a[i][j]ˆ) can appear as arguments for operations other than elm
and fld respectively. This limitation allows us to simplify definition of our assignment
statement abstraction6 .
To preserve safety of the analysis we have to take into account memory aliasing
appearing in programs. Two access terms are alias if they address the same memory
location. In the general case ERA is inadequate itself to handle precisely all kinds
of aliasing and we should use other analyses. Next it is assumed that for each access
term ta we know a set A(ta ) of access terms covering a set of aliases for ta (may–alias
information about ta ). Let
+
A(ta , L(G)) = (A(ta ) ∩ CV) ∪ {f ∈ FP | A⇒
f (. . .) ∧ f (. . .) ∈ A(ta )}
G
6 Otherwise we have either to accept that each assignment to some structured variables destroys
all equalities involving other components of it or to implement some strategy (for example copying)
preserving useful and safe access terms.
13
(“roots” of memory access terms in A(ta )). Notice that flow–insensitive approximations
of alias information may cause conservative results of ERA. Therefore ERA and alias
analyses used in its implementation should have the same sensitivity to the control flow.
Assignment statement L[[v := exp]].
1. In the state L evaluate exp using evaluation transformer formally defined earlier
in Section 1.3.2. Let L′ be a result of the evaluation and let E be a nonterminal
such that E⇒exp.
2. Perform L′ {| ↓ A(v, L′ )|}. Do not remove E.
3. Add the term v so that E⇒v.
Unfortunately, in some cases this abstraction of the assignment statement fails
as before. For example, this assignment transformer corresponding to x:=x+1 and
being applied to the state L = {(x > 0) = T RUE} gives only the trivial identity
L[[x := x + 1]] = {x = x}. To improve accuracy of the analysis in these cases we can
consider “artificial” variables associated with scalar variables of the program which will
store previous values of the original ones. Under this approach between first and second
steps of the assignment statement effect definition we should insert the step
Let A → v and B → v ′ where v ′ is associated with v. Remove the second rule and
add A → v ′ if it is needed.
...
Under this approach we shall have L[[x := x + 1]] = {(x′ > 0) = T RUE, x = x′ + 1}
from where we can deduce that x > 0 also.
2.2
Other transformers
• Program.
Given a program
PROGRAM;
VAR x : T;
BEGIN
S
END.
(∗ variables ∗)
(∗ statements ∗)
we can define the following transformer corresponding to it
⊥′ [[P ROGRAM ]] = ⊥′ [[x := ω]][[S]]
where ω represents the indefinite value. Notice that ω is not a constant.
• Empty statement .
L[[ ]] = L
• Sequence of statements.
L[[S1 ; S2 ]] = L[[S1 ]][[S2 ]]
14
• Read statement .
L[[READ(x)]] = L{|x|}{| ↓ A(x)|}
Notice that if for read statement as well as for other statements some set of user’s
pre– or post–assertions represented in the form of equalities of program terms is
supplied then the analyzer can take them into consideration to check consistency
and to include in the current computation state.
• Write statement .
L[[WRIT E(x)]] = L{|x|}
• Conditional statement .
IF p THEN St
ELSE
Sf
END.
If L′ = L{|p|} then
L[[IF ]] = L′ {|p ≡ TRUE|}[[St ]]∩L′ {|p ≡ FALSE|}[[Sf ]].
• Cycle statement .
CYCLE
S
(∗ body of cycle ∗)
END
where S is a composed statement that possibly contains occurrences of exit-ofcycle statements EXITk . When the sequence
L0 = L, Ln = Ln−1 [[S]] for n > 0
becomes stabilize
L[[CY CLE]] = ∩k Ek
where Ek is a stationary entry state for EXITk . If this process does not become
stabilize then some widening operator should be used (see Section 3).
• Halt, exit and return statements.
L[[EXIT ]] = L[[RET U RN ]] = ⊤′
• Call of function. We assume that return of results of function calls is implemented
as an assignment to variables having the same names as invoked functions (connection with their call sites should be taken into account). The function bodies
may contain RETURN statements as well.
FUNCTION F(x: T1 ) : T;
VAR y: T2 ;
BEGIN
S
END.
(∗ local variables ∗)
(∗ statements ∗)
L[[F U N CT ION ]] = L[[x := e; y := ω; F := ω]][[S]]∩R
15
✓
x
✒
L
y
✓
✏
x
✒
✏
✑
❅
❅
✗✔
✗✔
✠
❘
❅
✑
y
+
✖✕
F
1
✖✕
L[[x := F (x)]]
Figure 6: Function call for F (a) = a + 1.
where e is a factual parameter of the function (see the note for READ statement)
and R = ∩Rk is intersection of stationary entry states for return statements in S
if they exist and R = ⊤′ otherwise. In the term being evaluated the result of a
function call is represented by F .
3
Widening operator and convergence of the analysis
Our abstract semantic domain does not satisfy the (descending) chain condition and
therefore it requires some widening operator [9, 10, 11]. To guarantee the convergence
of the abstract interpretation we should use a dual widening operator:
˜ & y ⊇ x▽y,
˜
• ∀x, y ∈ ℘(EQS) ⇒ x ⊇ x▽y
• for all decreasing chains x0 ⊇x1 ⊇ . . ., the decreasing chains defined by y0 = x0 , . . . , yi+1 =
˜ i+1 , . . . are not strictly decreasing.
yi ▽x
The iteration sequence with widening is convergent and its limit is a sound approximation of the fixpoint.
3.1
Widening operator for ERA–grammars
Infinite chains can appear because corresponding languages have common infinite subsets generated by cyclic derivations in grammars. The source of that in ERA is term
identification. We can avoid this problem by imposing the constraint that grammars
must be acyclic. Within the semilattice ℘(EQS)(⊇, EQS, ∩), the subsemilattice of finite
languages7 generated by such grammars satisfies the chain condition, but such languages
are not expressive enough. Our solution is the following. Grammars are not originally
restricted but if in course of abstract interpretation the grammar size becomes greater
than some parameter, then “harmful” cycles must be destroyed. To this end we remove
grammar rules participating in cyclic derivations. Correctness of this approximation of
intersection follows from Lemma 2.
7 Notice that sets of term equalities of a special form corresponding to these languages were used
by V. Sabelfeld to develop effective algorithms of recognizing equivalence for some classes of program
schemata [39, 40].
16
Detecting such rules is no simpler than the “minimum-feedback-arc/vertex-set”
problem (MFAS or MFVS) if we consider the grammars as directed graphs. These
sets are the smallest sets of arcs or vertices, respectively, whose removal makes a graph
acyclic. We suppose that the “feedback vertices” choice is more natural for our purposes.
In the general case this problem is NP –hard, but there are approximate algorithms that
solve this problem in polynomial [41] or even linear [37] time. Consideration of weighted
digraphs makes it possible to distinguish grammar rules with respect to their worth for
accuracy of the analysis algorithm. However, perspectives of this approach are not clear
now for complexity/precision reasons of such algorithms. For example, [19] proposes
an algorithm for weighted FVS-problem requiring O(n2 µ(n) log2 n) time where µ(n) is
complexity of matrix multiplication.
✏
✏
✓
✓
✟
x
a ✛
x
a
f
f
✒
✑
✒
✑
✓ ❙
❙
✓
❙
❙
✓
❙
❙
✓
❙
❙
❙
✓
✇
✇
✓ ✠
✓ ✴
✏
✓ ❙
✏
✏
✓
✲
✲
g
y
g
y
r
r
f
h
h
✒
✑
✒
✑
✒
✑
✒
❅
✓
❙
✻
✻
✡
✠✡
✠
✓
❙❅
✓
❙ ❅
✓
❙ ❅
❅
✓
❙✓
❫ ✇
❘
❅
✓
☛ ✏
✏
✴
✓
✇
❙
z
z
h
✒
✑
✒
✑
✚
✙
L(G)
L(G)\fvs
f
g
r
f
h
f
g
r
h
Acyclic subgraph of the cyclomatic graph
of the grammar
h
Cyclomatic graph for the grammar G
Figure 7: FVS–transformation.
17
✏
✑
Our widening operator for the analysis of equality relationships is defined in the
following way. A vertex set of the cyclomatic graph8 of the grammar G is a set of
functional symbols existing in G. An arc (f, g) belongs to its arc set if G contains rules
A → f (. . . , B, . . .) and B → g(. . .). A transformation of a cyclomatic graph which
involves detecting some FVS (it can be an upper approximation of a minimal feedback
set) and removing all vertices from the FVS is said to be an FVS–transformation
(an example is shown in Figure 7). Let L\FVS be a language obtained from L by
FVS–transformation applied to the grammar generating L. We define
L(G1 )\FVS ∩ L(G2 ) if |G2 | ≥ |G1 | > d,
∼
L(G1 ) ∩ L(G2 )\FVS if |G1 | > |G2 | > d,
L(G1 )▽L(G
2) =
L(G1 ) ∩ L(G2 )
otherwise,
where d is a user-defined parameter. It is reasonable to choose this parameter, depending
on number of variables of analyzed programs, as a linear function with a small factor of
proportionality. Notice that in this case the lengths of appearing chains linearly depend
on number of variables living simultaneously.
3.2
Divergence of the analysis
Is the widening operator, being rather complex, really needed for the analysis of equality
relationships? Are there programs which, being analyzed, generate infinite chains of
semantic properties? It should be mentioned that constructing such program examples
has been a problem for a long time. In [18] we stated our belief that their existence seems
hardly probable. These attempts failed, because they concentrated on constructing an
example with completely non-interpretable functional symbols, i.e. in the frame of the
“pure” theory of program schemata.
As already noticed, we can widely vary the interpretability of the analysis algorithm.
In order to construct an example, it will suffice to use the following rule of completion:
if (t1 = t2 ) = T RUE ∈ L then Co(L) = L{|t1 ≡ t2 }| .
We consider the following example:
...
x:=f(y);
if f(x)=f(y) then
while y=f(g(y)) do
y:=g(y)
end
end
...
program scheme
...
x:=sign(y);
if sign(x)=sign(y) then
while y=sign(abs(y)) do
y:=abs(y)
end
end
...
“real-world” program
The properties computed at the body entry belong to an infinite decreasing chain.
On Figure 8 a state Le describes properties valid before cycle execution; states L1
8 A graph represents cyclic derivations in the grammar. The author did not find another appropriate
name for this object in the Computer Science and Discrete Mathematics literature.
18
✓✏
❄✔
✗
x f f
✖
✕
✒✑
❄
✗✔
y
✖✕
Le
✓✏
❄✔
✗
x f f
✖
✕
✗❄
y f
✖
✒✑
✔
✛✟
✕
❄
✗✔
g
✖✕
L1
✒✑
✓✏
❄✔
✗
x f f
✖
✕
✒✑
❄
✗✔
☛✲ f
✖✕
✗❄
✔
∗
✛
✟
y
g
f
✖
✕
✒✑
❄
✗✔
g
✖✕
✒
✑
L2
Figure 8: The case of divergence of the analysis.
and L2 describe properties at the entry of the cycle body for first and second iteration,
respectively. It is easy to see that L1 ∩L2 coincides with L2 except in the equality
relationships containing terms generated by g ∗ . Therefore every time we obtain the
next state, the functional element g ∗ is absent and the sub-net placed in the dashed box
will repeat more and more times.
To obtain a “real-world” program from this program scheme, we interpret the functional symbols in the following manner: f ≡ sign and g ≡ abs. We would like to point
out the following interesting property. Execution of this piece of code (i.e., its behavior
determined with the standard semantics) diverges only for two values of y: 0 and 1.
At the same time the analysis algorithm (i.e. execution of the piece of code under our
nonstandard semantics) is always divergent on condition that a widening operator is
not used and the assumption on interpretation mentioned above holds.
Is this program actually real-world? The reader should decide that by himself but
we would like to underline the following9 . On the one hand, the interpretability of the
analysis algorithm can be varied in wide ranges and, on the other, we are not able to
prove formally the impossibility of such behavior of the analyzer under the considered
interpretation. So, we can choose either a lean analysis using acyclic grammars only or
another one using arbitrary grammars and some widening operator.
9 The penetrating reader may notice that this program to be considered as human written is really
stupid (it can be slightly intellectualized if we interpret g as “add 1”). But for automatic generators of
programs such a code does not seem improbable.
19
3.3
Complexity of the analysis
In [18] we pointed out the following upper bound on time for the algorithm of ERA:
O(nmG2max ), where n is program size, m is maximum of number of program variables
existing at a time, and Gmax is maximum of sizes of grammars appearing in course
of the analysis. Due to our construction of widening operator (namely, choice of the
parameter d linearly depended on m) we can assume that Gmax = O(m).
This bound can be deduced with help of Theorem 6 in [5]. The theorem states
that for the recursive strategy of chaotic iteration maximum complexity is
X
h·
δ(c)
≤ h · |C| · |W | ,
(∗∗)
c∈C
where h is maximum length of increasing chains built by a widening operator, C is set
of control program points, δ(c) is depth of the control point c in hierarchy of nested
strongly connected components of the control flow graph containing c, and W is set of
vertices where a widening operator is applied during the analysis.
For well-structured programs we can assume that maximum depth of nested loops
does not depend on program size and is bounded by some constant. By (∗∗) we conclude
that number of algorithm steps does not exceed O(nm). Since time complexity of
all operations used in the analysis are estimated by G2max we obtain O(nmG2max ) (or
O(nm3 )) upper bound. Notice that to improve the results of the analysis it is possible
to use rich semantic completion and more precise FVS–algorithms that have more than
quadratic time complexity.
However, experimental results show that an approximation of a fixed point for the
heads of cycle bodies is usually attained after at most two iterations and time complexity
of the analysis is proportional to nGmax . Also, the user can turn off checking a threshold
after which widening is started. In this case, he (consciously) admits some chance that
the analysis diverges but we believe that this chance is not too big.
It is easy to see that the space complexity of the equality relationship analysis is
O(nGmax ) and it is essentially depended on the number of variables. We estimate
the actual space requirements as 1.5–2.0 Mb per 1000 program lines for middle-size
programs.
4
4.1
Processing of invariants and experimental results
Usage of ERA–invariants
ERA produces some set of invariants involving program terms which can be useful
at different steps of program development and processing: debugging, verification (for
that the invariants are interesting themselves), specialization, and optimization. The
automatic prover mentioned above can be used at step of post-processing results of the
analysis.
We notice that the analyzer can tell the user useful information both at the stage
of analyzing (this means that it is possible that there exist execution traces where such
computational states appear; we mark with + properties which can be detected at this
20
stage) and at the stage of processing of results of the analysis (these properties hold for
each execution trace leading to this program point). We shall briefly list some program
properties that can be extracted from computation states Lin and Lout being for a
statement S the input and output states respectively.
+ Variable x has an indefinite value if Lin contains x = ω.
+ Error in evaluation of an expression: division by zero, out of type ranges, nil–pointer
dereferencing, etc.
− S is inaccessible if Lin = ⊤′ . This information can correspond to different properties
of program execution: potentially infinite cycles and recursive calls, dead branches
of conditional statements, useless procedural definitions, etc.
− Assignment statement v:=exp is redundant if Lout contains v = v ′ where v ′ is the
variable associated with v.
− Unused definitions (constants, variables, types).
− Constant propagation. Notice that ERA can detect that an expression is constant
not only when constants for all variables in this expression are known.
− More general: for some expression there exists an expression that is equal to the
original one and calculated more efficiently (with respect to an given criterion
time/space and target computer architecture).
Obviously, this list is not complete and there are many other properties which can
be extracted from the invariants. For example, we can consider systems of equations/inequalities contained in the gathered invariants and try to solve them to derive
more precise ranges for values of expressions or the inconsistency of this computational
state.
Apart from the automatic mode when invariants are processed automatically, we
provide an interactive mode to visualize results of our analysis in a hypertext system.
HyperCode presented in [7, 3] is an open tunable system for visualization of properties
of syntactic structures. There are two cases: visualization of all properties detected in
the automatic mode and the user-driven processing and visualization of properties.
The experiments show that not all program properties of interest can be automatically extracted out of the computed invariants. It is not judicious to consider many
particular cases and to hardly embed them into the system. Instead, the system facilitates the specification of the user request with some friendly interface. He chooses a
program point and an expression and obtains those and only those equality relationships, valid at this point, where this expression occurs as a super- or sub-term.
4.2
Program examples
An example of a program is presented below. The properties detected by the analyzer
are indicated in comments.
21
program
KMP
Lambert
Automaton
IntF ib
Ackerman
length (lines)
M2Mix
167
361
37
87
64
ERA
133
326
35
77
62
average
improv.
20.35%
9.7%
5.4%
11.5%
3.1%
10.01%
size (bytes)
M2Mix
2996
6036
969
1647
1384
ERA
2205
2564
926
1432
1322
average
improv.
26.4%
57.5%
4.5%
13.05%
4.5%
21.19%
Table 2: Comparison ERA and M2Mix.
var x,y,z: integer;
procedure P(a,b: integer):integer;
begin
(*parameters are always equal*)
return a+b
(*expression can be simplified: 2*a*)
end P;
begin
Read(x);
while x≤0 do
Read(x);
x := x+1;
z := x+z;
(*variable z might be uninitialized*)
y := x+1;
if x=0 then
z := y;
(*r-value can be simplified: z:=1*)
else
z := x+1;
(*r-value can be simplified: z:=y*)
x := y;
end;
Write( P(y,z) )
(*call can be transformed: Write(2*y)*)
end;
x := z div (y - z); (*arithmetical error*)
Write(x)
(*inaccessible point*)
end.
On basis of the analysis, this program can be transformed into the following:
var x:integer;
begin
Read(x);
while x≤0 do
Read(x); x := x+2; Write(2*x)
end
ERROR EXCEPTION;
end.
In Table 2 we present some results of optimization based on our analysis of residual programs generated by M2Mix specializer [6, 33]. To compile these examples, we
22
used XDS Modula/Oberon compiler v.2.30 [20]. The following programs have been
investigated:
• KMP — the “naı̈ve” matching algorithm specialized with respect to some pattern;
the residual program is comparable to Knuth, Morris, and Pratt’s algorithm in
efficiency. (see also Appendix).
• Lambert — a program drawing Lambert’s figure and specialized with respect to
number of points.
• Automaton — an interpreter of a deterministic finite-state automaton specialized
with respect to some language.
• IntF ib — an interpreter of MixLan [33] specialized with respect to a program
computing Fibonacci numbers.
• Ackerman — a program computing some values of Akcerman’s function and
specialized with respect to its first argument.
Let us comment briefly on the obtained results. Reducing length of a program can
be considered as reducing number of operators and declarations. In these examples the
optimizing effect was typically attained by the removal of redundant assignments, dead
operators, unused variables and the reduction of operator strength. The only exception
is KMP program characterized by high degree of polyvariance (roughly speaking it
means presence of deep–nested conditional statements) and an active usage of array
references. Here some IF–statements with constant conditions and redundant range
checks were eliminated. Notice that the last optimizing transformation is very important
for Modula–like languages where such checks are defined by the language standard. Such
notable optimizing effect for the Lambert program is explained by deep reduction of
power of floating-point operations which cannot be achieved by optimizing techniques
now used in compilers. Since Automaton and Ackerman programs are quite small,
their optimization gives conservative results. However, they would be better for the
Ackerman program if the implementation of ERA were context-sensitive. Substantial
speed-up of these optimized programs was not obtained (it was less than 2%) and this
is not surprising since the great bulk of specializers take it as a criterion of optimality.
These experiments show that an average reduction of size of residual programs is 20–
25%. Because the case of KMP program seems to be the most realistic10 , we suppose
that such improvement can be achieved in practice for real–world programs and it will
be increased for large residual programs with a high degree of polyvariance and active
usage of arrays and float-point arithmetics. It is the author opinion that the analysis
of automatically generated (from high–level specifications as well) programs is the most
promising direction of its application, especially in the context-sensitive implementation
of ERA.
10 Unfortunately our experiments were not exhaustive enough since the partial evaluation is not
involved yet in real technological process of the software development and hence finding large resudial
programs is a hard problem.
23
Acknowledgement
The author wishes to thank M.A. Bulyonkov, P. Cousot, R. Cousot, and V.K. Sabelfeld
for support, useful discussions and remarks.
References
[1] A. Aho and J. Ullman. The theory of parsing, translation and compilation, volume 1. Prentice-Hall Inc., 1972.
[2] B. Alpern, M.N. Wegman, and F.K. Zadeck. Detecting equality of variables in
programs. In Proc. of the 15th Annual ACM Symposium Principles of Programming
Languages, pages 1–11. ACM Press, 1988.
[3] D. Baburin, M. Bulyonkov, P. Emelianov, and N. Filatkina. Visualization facilities
in program reengineering. Programmirovanie, 27(2):21–33, 2001.
[4] L. Berman and G. Markowsky. Linear and non-linear approximate invariants. Research Report RC7241, T.J. Watson Research Center of IBM, Yorktown, NY, 1976.
[5] F. Bourdoncle. Efficient chaotic iteration strategies with widenings. In Proc. of the
International Conference on Formal Methods in Programming and Their Applications, volume 735 of Lecture Notes in Computer Science, pages 129–141. SpringerVerlag, 1993.
[6] M. Bulyonkov and D. Kochetov. Practical aspects of specialization of Algol-like
programs. In Proc. of the International Seminar on Partial Evaluation, volume
1110 of Lecture Notes in Computer Science, pages 17–32. Springer-Verlag, 1996.
[7] M. Bulyonkov and D. Kochetov. Visualization of program properties. Research
Report 51, Institute of Informatics Systems, Novosibirsk, Russia, 1998.
[8] R. Clarisó and J. Cortadella. The octahedron abstract domain. In Proc. of the
11th International Static Analysis Symposium, volume 3148 of Lecture Notes in
Computer Science, pages 312–327. Springer–Verlag, 2004.
[9] P. Cousot and R. Cousot. Abstract interpretation: A unified lattice model for static
analysis of programs by construction or approximation of fixpoints. In Rec. of the
4th ACM Symposium on Principles of Programming Languages, pages 238–252.
ACM Press, 1977.
[10] P. Cousot and R. Cousot. Abstract interpretation and application to logic program.
Journal of Logic Programming, 13(2–3):103–180, 1992.
[11] P. Cousot and R. Cousot. Abstract interpretation frameworks. Journal of Logic
and Computation, 2(4):511–547, 1992.
24
[12] P. Cousot and R. Cousot. Formal languages, grammar and set–constraint–based
program analysis by abstract interpretation. In Rec. of the Conference on Functional Programming Languages and Computer Architecture, pages 170–181. ACM
Press, 1995.
[13] P. Cousot and R. Cousot. Software analysis and model checking. In Proceedings of
the 14th International Conference on Computer Aided Verification, volume 2404 of
Lecture Notes in Computer Science, pages 37–56. Springer-Verlag, 2002.
[14] P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In Records of the 5th annual ACM Symposium on Principles
of Programming Languages, pages 84–96. ACM, ACM Press, January 1978.
[15] A. Deutsch. Interprocedural may–alias analysis for pointers: beyond k-limiting.
SIGPLAN Notices, 29(6):230–241, 1994. Proc. of the ACM SIGPLAN’94 Conference on Program Language Design and Implementation.
[16] P.G. Emelianov and D.E. Baburin. Semantic analyzer of Modula-programs. In
Proc. of the 4th International Static Analysis Symposium, volume 1302 of Lecture
Notes in Computer Science, pages 361–363. Springer–Verlag, 1997.
[17] P.G. Emelianov and V.K. Sabelfeld. Analyzer of semantic properties of Modulaprogramms. In Software intellectualization and quality, pages 100–107. Institute of
Informatics Systems, Novosibirsk, Russia, 1994.
[18] P.G. Emelianov. Analysis of the equality relation for the program terms. In Proc. of
the Third International Static Analysis Symposium, volume 1145 of Lecture Notes
in Computer Science, pages 174–188. Springer–Verlag, 1996.
[19] G. Even, J. Naor, B. Schieber, and M. Sudan. Approximating minimum feedback sets and multi-cuts. In Proc. of the 4th International Conference on Integer
Programming and Combinatorial Optimization, volume 920 of Lecture Notes in
Computer Science, pages 14–28. Springer–Verlag, 1995.
[20] Excelsior LLC. Native xds-x86 modula-2/oberon-2 development toolset.
[21] K. Gargi. A sparse algorithm for predicated global value numbering. In Proceedings
of the ACM SIGPLAN 2002 Conference on Programming Language Design and
Implementation, pages 45–56, 2002.
[22] A.D. Gordon and A.M. Pitts, editors. High Order Operational Techniques in Semantics. Publications of Newton Institute. Cambridge University Press, 1998.
[23] Ph. Granger. Static analysis of arithmetical congruences. International Journal of
Computer Mathematics, 30:165–199, 1989.
[24] Ph. Granger. Static analysis of linear congruence equalities among variables of a
program. In Proceedings of the International Joint Conference on Theory and Practice of Software Developement, volume 493 of Lecture Notes in Computer Science,
pages 169–192. Springer–Verlag, 1991.
25
[25] S. Gulwani and G.C. Necula. Discovering affine equalities using random interpretation. In Proc. of the ACM SIGPLAN-SIGACT 2003 Principles of Programming
Languages, pages 74–84, 2003.
[26] S. Gulwani and G.C. Necula. A polynomial-time algorithm for global value numbering. In Proc. of the 11th International Static Analysis Symposium, volume 3148
of Lecture Notes in Computer Science, pages 212–227. Springer–Verlag, 2004.
[27] N. Halbwachs, Y.-E. Proy, and P. Roumanoff. Verification of real-time systems
using linear relation analysis. Formal Methods in System Design, 11(2):157–185,
1997.
[28] Nevin Heintze, Joxan Jaffar, and Răzvan Voicu. A framework for combining analysis and verification. In Records of the 27th Annual ACM Symposium on Principles
of Programming Languages, pages 26–39, 2000.
[29] N. Heintze and J. Jaffar. Set constraint and set–based program analysis. In A. Borning, editor, Principles and Practice of Constraint Programming, volume 874 of
Lecture Notes in Computer Science, pages 281–298. Springer–Verlag, 1994.
[30] N.D. Jones and S.S. Muchnick, editors. Program Flow Analysis: Theory and Applications. Prentice-Hall, 1981.
[31] N.D. Jones. Flow analysis of lazy higher–order functional programs. In S. Abramsky
and C. Hankin, editors, Abstract Interpretation of Declarative Languages, pages
103–122. Ellis Horwood, 1987.
[32] M. Karr. Affine relationships among variables of a program. Acta Informatica,
6:133–151, 1976.
[33] D.V. Kochetov. Effective specialization of Algol–like programs. Ph.D. Thesis, Institute of Informatics Systems, Novosibirsk, Russia, 1995.
[34] F. Masdupuis. Array operations abstraction using semantic analysis of trapezoid
congruences. In Proceedings of the International Conference on Supercomputing,
pages 226–235. ACM Press, 1992.
[35] F. Masdupuis. Semantic analysis of interval congruences. In Proceedings of the
International Conference Formal Methods in Programming and Their Applications,
volume 735 of Lecture Notes in Computer Science, pages 142–155. Springer–Verlag,
1993.
[36] A. Miné. A new numerical abstract domain based on difference-bound matrices. In
Proceedings of the Second Symposium on Programs as Data Objects, volume 2053,
pages 155–169, 2001.
[37] B.K. Rosen. Robust linear algorithms for cutsets. Journal Algorithms, 3:205–217,
1982.
26
[38] O. Rüthing, J. Knoop, and B. Steffen. Detecting equalities of variables: Combining efficiency with precision. In Proc. of the 6th International Static Analysis
Symposium, volume 1694 of Lecture Notes in Computer Science, pages 232–247.
Springer–Verlag, 1999.
[39] V. Sabelfeld. Polynomial upper bound for the complexity of the logic-termal equivalence decision. Doklady Akademii Nauk, Matematika, 249(4):793–796, 1979.
[40] V. Sabelfeld. The logic–termal equivalence is polynomial–time decidable. Information Processing Letters, 10(2):102–112, 1980.
[41] E. Speckenmeyer. On feedback problems in digraphs. In Proceedings of the 15th
International Workshop on Graph–Theoretic Concepts in Computer Science, volume 411 of Lecture Notes in Computer Science, pages 218–231. Springer–Verlag,
1990.
[42] A. Venet. Automatic analysis of pointer aliasing for untyped programs. Science of
Computer Programming, 35(2-3):223–248, 1999.
[43] M. Wegman and F.K. Zadeck. Constant propagation with conditional branches.
ACM Transactions on Programming Language and Systems, 13(2):181–210, 1991.
27
Appendix. Analysis of KMP
The appendix presents results of application of ERA to KMP program generated by
the specializer M2Mix. This program is a specialization of a program implementing
the naı̈ve pattern matching Match(p,str) with respect to the pattern p=”ababb”.
The invariants are written as comments at program points where they hold. We can
conclude that:
• The target string necessarily ends with ”#” and the variable ls is equal to the
string length (line 10).
• Every time when some element of str (lines 20, 31, 42, 53, 58, 67, 72, 78, 86, 108,
113, 129, 140, 151, 162) is used in second LOOP, the value of its index expression
does not exceed the value of the variable ls. The same is true for the value of a
variable before the increment statements INC(s) (lines 23, 34, 45, 62, 81, 89, 94,
98, 103, 116, 120, 132, 143, 154, 165). Therefore, it suffices to check that a value
of ls is not beyond the ranges determined by the type TYPE354a04 during
input of the target string (line 8). So, in the second cycle all range checks can be
eliminated.
• The assignment cfg counter:=0 is redundant (line 24).
• Conditions str[s+2]=’a’ and str[s]=’a’ are always false (lines 78 and 86, respectively) because two different constants are equal. So, the code of THEN–branches
is dead.
• The conditions at the lines 63, 68, 74, 82, 104, 109 are false, too. However,
automatic detection of these properties are not as easily as the previous.
Using this semantic information, it is possible to build a new program functionally
equivalent to Match(”ababb”,str). In text of the program given below the underlined
code can be eliminated.
MODULE Match;
FROM FIO IMPORT File,Open,ReadChar,WriteInt,stdout;
VAR cfg counter : CARDINAL; str file : File;
TYPE TYPE354a04 = [0..20];
TYPE TYPE355004 = ARRAY TYPE354a04 OF CHAR;
VAR str : TYPE355004; ls,s : TYPE354a04;
BEGIN
1:
str file := Open(”target.dat”);
2:
ls := 0;
3:
LOOP
4:
str[ls] := ReadChar(str file);
5:
IF (str[ls]=’#’) THEN
(∗str[ls] =′#′ ∗)
6:
EXIT
7:
ELSE
(∗str[ls] 6=′#′ ∗)
8:
INC(ls)
9:
END
10: END;
(∗str[ls] =′#′ ∗)
11: s := 0;
cfg counter := 0;
(∗s = cf g counter = 0∗)
12:
28
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:
55:
56:
57:
58:
59:
60:
61:
62:
63:
64:
65:
66:
67:
68:
69:
70:
71:
LOOP
CASE cfg counter OF
|0:
IF ((s+0)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+0)]=’a’) THEN
cfg counter := 1
ELSE
INC(s);
cfg counter := 0
END
|1:
IF ((s+1)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+1)]=’b’) THEN
cfg counter := 2
ELSE
INC(s);
cfg counter := 0
END
|2:
IF ((s+2)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+2)]=’a’) THEN
cfg counter := 3
ELSE
INC(s);
cfg counter := 4
END
|3:
IF ((s+3)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+3)]=’b’) THEN
IF ((s+4)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+4)]=’b’) THEN
WriteInt(stdout,s,0);
EXIT
ELSE
INC(s);
IF ((s+0)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+0)]=’a’) THEN
IF ((s+1)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
29
(∗ cf g counter = 0∗)
(∗s ≥ ls∗)
(∗0 < ls, s < ls∗)
(∗str[s] =′a′ , 0 < ls, s < ls∗)
(∗0 < ls, s < ls, str[0] 6=′a′ ∗)
(∗ cf g counter = 0∗)
(∗0 < ls, s < ls∗)
(∗ cf g counter = 1∗)
(∗s + 1 ≥ ls∗)
(∗s + 1 < ls∗)
(∗str[s + 1] =′b′ ∗)
(∗str[s + 1] 6=′b′ ∗)
(∗str[s] 6=′b′ , s < ls∗)
(∗ cf g counter = 2∗)
(∗s + 2 ≥ ls∗)
(∗s + 2 < ls∗)
(∗str[s + 2] =′a′ ∗)
(∗str[s + 2] =
6 ′a′ ∗)
(∗str[s + 1] =
6 ′a′ , s + 1 < ls∗)
(∗ cf g counter = 3∗)
(∗s + 3 ≥ ls∗)
(∗s + 3 < ls∗)
(∗str[s + 3] =′b′ ∗)
(∗s + 4 ≥ ls, s + 3 < ls∗)
(∗s + 4 < ls∗)
(∗str[s + 3] = str[s + 4] =′b′ ∗)
(∗str[s + 3] =′b′ , str[s + 4] =
6 ′b′ , s + 4 < ls∗)
(∗str[s + 2] =′b′ , str[s + 3] =
6 ′b′ , s + 3 < ls∗)
(∗s ≥ ls, s + 3 < ls∗)
(∗str[s + 2] =′b′ , str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗str[s] =′a′ , str[s + 2] =′b′ , str[s + 3] 6=′b′ ,
s + 3 < ls∗)
(∗s + 1 ≥ ls, s + 3 < ls∗)
(∗the same as at line 67∗)
72:
IF (str[(s+1)]=’b’) THEN
73:
74:
75:
76:
77:
78:
79:
80:
IF ((s+2)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+2)]=’a’) THEN
cfg counter := 3
ELSE
81:
INC(s);
82:
IF ((s+0)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
83:
84:
85:
IF (str[(s+0)]=’a’) THEN
cfg counter:=14
ELSE
86:
87:
88:
89:
INC(s);
90:
cfg counter:=4
END
END
91:
92:
93:
94:
95:
96:
97:
98:
99:
100:
101:
102:
103:
104:
105:
106:
107:
108:
109:
110:
111:
112:
ELSE
INC(s);
cfg counter := 12
END
ELSE
INC(s);
cfg counter := 12
END
END
ELSE
INC(s);
IF ((s+0)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
IF (str[(s+0)]=’a’) THEN
IF ((s+1)≥ls) THEN
WriteInt(stdout,(-1),0);
EXIT
END;
30
(∗str[s] =′a′ , str[s + 1] =′b′ , str[s + 2] =′b′ ,
str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗s + 2 ≥ ls, s + 3 < ls∗)
(∗str[s] =′a′ , str[s + 1] =′b′ , str[s + 2] =′b′ ,
str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗inaccessible point∗)
(∗str[s] =′a′ , str[s + 1] =′b′ , str[s + 2] =′b′ ,
str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗str[s − 1] =′a′ , str[s] =′b′ , str[s + 1] =′b′ ,
str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗s ≥ ls, s + 2 < ls∗)
(∗str[s − 1] =′a′ , str[s] =′b′ , str[s + 1] =′b′ ,
str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗inaccessible point∗)
(∗str[s − 1] =′a′ , str[s] =′b′ , str[s + 1] =′b′ ,
str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗str[s − 2] =′a′ , str[s − 1] =′b′ , str[s] =′b′ ,
str[s + 1] 6=′b′ , s + 1 < ls∗)
(∗str[s − 2] =′a′ , str[s − 1] =′b′ , str[s] =′b′ ,
str[s + 1] 6=′b′ , s + 1 < ls∗)
(∗str[s] =′a′ , str[s + 1] 6=′b′ , str[s + 2] =′b′ ,
str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗str[s − 1] =′a′ , str[s] 6=′b′ , str[s + 1] =′b′ ,
str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗str[s] =′a′ , str[s + 2] =′b′ , str[s + 3] 6=′b′ ,
s + 3 < ls∗)
(∗str[s − 1] =′a′ , str[s + 1] =′b′ , str[s + 2] 6=′b′ ,
s + 2 < ls∗)
(∗str[s + 3] 6=′b′ , s + 3 < ls∗)
(∗str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗s ≥ ls, s + 2 < ls∗)
(∗str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗str[s] =′a′ , str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗s + 1 ≥ ls, s + 2 < ls∗)
(∗str[s] =′a′ , str[s + 2] 6=′b′ , s + 2 < ls∗)
113:
IF (str[(s+1)]=’b’) THEN
cfg counter := 2
ELSE
116:
INC(s);
cfg counter := 10
117:
118:
END
119:
ELSE
120:
INC(s);
cfg counter := 10
121:
122:
END
123:
END
124:
|4:
125:
IF ((s+0)≥ls) THEN
126:
WriteInt(stdout,(-1),0);
127:
EXIT
128:
END;
129:
IF (str[(s+0)]=’a’) THEN
cfg counter := 1
130:
131:
ELSE
132:
INC(s);
133:
cfg counter := 0
134:
END
135:
| 10:
136:
IF ((s+0)≥ls) THEN
137:
WriteInt(stdout,(-1),0);
138:
EXIT
139:
END;
140:
IF (str[(s+0)]=’a’) THEN
cfg counter := 1
141:
142:
ELSE
143:
INC(s);
144:
cfg counter := 0
145:
END
146:
| 12:
147:
IF ((s+0)≥ls) THEN
148:
WriteInt(stdout,(-1),0);
149:
EXIT
150:
END;
151:
IF (str[(s+0)]=’a’) THEN
cfg counter := 14
152:
153:
ELSE
154:
INC(s);
cfg counter := 4
155:
156:
END
157:
| 14:
158:
IF ((s+1)≥ls) THEN
159:
WriteInt(stdout,(-1),0);
160:
EXIT
161:
END;
162:
IF (str[(s+1)]=’b’) THEN
163:
cfg counter := 2
164:
ELSE
165:
INC(s);
cfg counter := 4
166:
167:
END
168:
END
END
END Match.
(∗str[s] =′a′ , str[s + 1] =′b′ , str[s + 2] 6=′b′ ,
s + 2 < ls∗)
114:
(∗str[s] =′a′ , str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗str[s − 1] =′a′ , str[s + 1] 6=′b′ , s + 1 < ls∗)
115:
(∗str[s + 2] 6=′b′ , s + 2 < ls∗)
(∗str[s + 1] 6=′b′ , s + 1 < ls∗)
(∗ cf g counter = 4∗)
(∗s ≥ ls∗)
(∗s < ls∗)
(∗str[s] =′a′ , s < ls∗)
(∗str[s] 6=′a′ , s < ls∗)
(∗str[s − 1] 6=′a′ , s ≤ ls∗)
(∗ cf g counter = 10∗)
(∗s ≥ ls∗)
(∗s < ls∗)
(∗str[s] =′a′ , s < ls∗)
(∗str[s] 6=′a′ , s < ls∗)
(∗str[s − 1] 6=′a′ , s ≤ ls∗)
(∗ cf g counter = 12∗)
(∗s ≥ ls∗)
(∗s < ls∗)
(∗str[s] =′a′ , s < ls∗)
(∗str[s] 6=′a′ , s < ls∗)
(∗str[s − 1] 6=′a′ , s ≤ ls∗)
(∗ cf g counter = 14∗)
(∗s + 1 ≥ ls∗)
(∗s + 1 < ls∗)
(∗str[s + 1] =′b′ , s + 1 < ls∗)
(∗str[s + 1] 6=′b′ , s + 1 < ls∗)
(∗str[s] 6=′b′ , s < ls∗)
31
| 2cs.AI
|
arXiv:1605.05448v1 [cs.NE] 18 May 2016
The Bees Algorithm for the
Vehicle Routing Problem
Aish Fenton
Department of Computer Science
University of Auckland
Auckland, New Zealand
Colophon:
Typeset in LATEX using typefaces Computer Modern.
Title-page image is courtesy of [1].
Preface
This MSc. thesis has been prepared by Aish Fenton at the University of Auckland,
Department of Computer Science. It has been supervised by Dr. Michael Dinneen.
The work undertaken in this thesis has grown out of a research project sponsored
by New Zealand Trade and Enterprise (NZTE) for the company vWorkApp Inc. to
research vehicle route optimisation for use within their software product.
Acknowledgements
I’d like to thank my partner in crime, Anna Jobsis, for her encouragement, cajoling,
threatening, bribing, doing my dishes, proof reading, guilt tripping, grammar policing,
comforting. . . doing whatever it took to help me get it done. Also, I’d like to thank
my mum and dad for their encouragement and for always making me feel like higher
education was within my reach.
Thanks to Steve Taylor and Steve Harding for holding down the fort at work while
I disappeared to do this thesis. And likewise thanks to my team, Jono, Rash, Elena,
Bob, Marcus, Yuri, and Robin for being on the ball (as always) despite my absence.
Thanks to Brendon Petrich and vWorkApp Inc. for providing me with time off work
and being supportive of me undertaking this.
And lastly, I especially owe Dr. Michael Dinneen, my supervisor, a big thank you for
persevering with me even though he must have doubted that I’d ever finish.
iii
Abstract
In this thesis we present a new algorithm for the Vehicle Routing Problem called the Enhanced Bees Algorithm. It is adapted
from a fairly recent algorithm, the Bees Algorithm, which was
developed for continuous optimisation problems. We show that
the results obtained by the Enhanced Bees Algorithm are competitive with the best meta-heuristics available for the Vehicle
Routing Problem—it is able to achieve results that are within
0.5% of the optimal solution on a commonly used set of test
instances. We show that the algorithm has good runtime performance, producing results within 2% of the optimal solution
within 60 seconds, making it suitable for use within real world
dispatch scenarios. Additionally, we provide a short history of
well known results from the literature along with a detailed description of the foundational methods developed to solve the
Vehicle Routing Problem.
Contents
1 Introduction
1.1
Content Outline
5
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background
2.1
7
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1
6
8
TSP Introduction and History . . . . . . . . . . . . . . . . . .
10
2.2
Exact Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.3
Classic Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
2.3.1
Constructive Heuristics . . . . . . . . . . . . . . . . . . . . . .
13
2.3.2
Two-phase Heuristics
. . . . . . . . . . . . . . . . . . . . . . .
15
2.3.3
Iterative Improvement Heuristics . . . . . . . . . . . . . . . . .
17
Meta-heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.4.1
Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . .
19
2.4.2
Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . .
21
2.4.3
Tabu Search
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
2.4.4
Large Neighbourhood Search . . . . . . . . . . . . . . . . . . .
25
Swarm Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
27
2.4
2.5
1
2.5.1
Ant Colony Optimisation . . . . . . . . . . . . . . . . . . . . .
28
2.5.2
Bees Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
3 Problem Definition
35
3.1
Capacitated Vehicle Routing Problem . . . . . . . . . . . . . . . . . .
35
3.2
Variants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
3.2.1
Multiple Depot Vehicle Routing Problem . . . . . . . . . . . .
37
3.2.2
Vehicle Routing with Time Windows . . . . . . . . . . . . . . .
37
3.2.3
Pickup and Delivery Problem . . . . . . . . . . . . . . . . . . .
38
4 Algorithm
39
4.1
Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
4.2
Problem Representation . . . . . . . . . . . . . . . . . . . . . . . . . .
40
4.3
Enhanced Bees Algorithm . . . . . . . . . . . . . . . . . . . . . . . . .
41
4.3.1
Bee Movement . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
4.3.2
Search Space Coverage . . . . . . . . . . . . . . . . . . . . . . .
44
Search Neighbourhood . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
4.4.1
Destroy Heuristic . . . . . . . . . . . . . . . . . . . . . . . . . .
45
4.4.2
Repair Heuristic . . . . . . . . . . . . . . . . . . . . . . . . . .
45
4.4.3
Neighbourhood Extent . . . . . . . . . . . . . . . . . . . . . . .
46
4.4
5 Results
49
5.1
Enhanced Bees Algorithm . . . . . . . . . . . . . . . . . . . . . . . . .
49
5.2
Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
3
5.3
5.2.1
Bees Algorithm versus Enhanced Bees Algorithm . . . . . . . .
52
5.2.2
Large Neighbourhood Search . . . . . . . . . . . . . . . . . . .
53
5.2.3
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
6 Conclusion
63
4
Chapter 1
Introduction
In this thesis we present a new algorithm to solve the Vehicle Routing Problem. The
Vehicle Routing Problem describes the problem of assigning and ordering geographically distributed work to a pool of resources. The aim is minimise the travel cost
required to complete the work, while meeting any specified constraints, such as a
maximum shift duration. Often the context used is that of a fleet of vehicles delivering goods to their customers, although the problem can equally be applied across many
different industries and scenarios, and has been applied to non-logistics applications,
such as microchip layout.
Interest in the Vehicle Routing Problem has increased over the last two decades as the
cost of transporting and delivering goods has become a key factor of most developed
economies. Even a decrease of a few percent on transportation costs can offer a
savings of billions for an economy. In the context of New Zealand (our home country)
virtually every product grown, made, or used is carried on a truck at least once during
its lifetime [33]. The success of New Zealand’s export industries are inextricably linked
to the reliability and cost effectiveness of our road transport. Moreover, a 1% growth
in national output requires a 1.5% increase in transport services [33].
The Vehicle Routing Problem offers real benefits to transport and logistics companies. Optimisation of the planning and distribution process, such as modelled by the
Vehicle Routing Problem, can offer savings of anywhere between 5% to 20% of the
transportation costs [52]. Accordingly, the Vehicle Routing Problem has been the
focus of intense research since its first formal introduction in the fifties. The Vehicle
Routing Problem is one of the most studied of all combinatorial optimisation problems, with hundreds of papers covering it and the family of related problems since its
introduction fifty years ago.
The challenge of the Vehicle Routing Problem is that it combines two (or more, in
the case of some of its variants) combinatorially hard problems that are themselves
5
6
CHAPTER 1. INTRODUCTION
known to be N P-hard. Its membership in the family of N P-hard problems makes it
very unlikely that an algorithm exists, with reasonable runtime performance, that is
able to solve the problem exactly. Therefore heuristic approaches must be developed
to solve all but the smallest sized problems.
Many methods have been suggested for solving the Vehicle Routing Problem. In this
thesis we develop a new meta-heuristic algorithm that we call the Enhanced Bees
Algorithm. We adapt this from another fairly recent algorithm, named the Bees
Algorithm, that was developed for solving continuous optimisation problems.
We show that the results obtained by the Enhanced Bees Algorithm are competitive
with the best modern meta-heuristics available for the Vehicle Routing Problem. Additionally, the algorithm has good runtime performance, producing results within 2%
of the optimal solution within 60 seconds. This makes the Enhanced Bees Algorithm
suitable for use within real world dispatch scenarios where often the dispatch process
is dynamic and hence it is impractical for a dispatcher to wait minutes (or hours) for
an optimal solution1 .
1.1
Content Outline
We start in Chapter 2 by providing a short history of the Vehicle Routing Problem, as
well as providing background material necessary for understanding the Enhanced Bees
Algorithm. In particular, we review the classic methods that have been brought to
bear on the Vehicle Routing Problem along with the influential results achieved in the
literature. In Chapter 3 we provide a formal definition of the Vehicle Routing Problem
and briefly describe the variant problems that have been developed in the literature.
In Chapter 4 we provide a detailed description of the Enhanced Bees Algorithm and
its operation, along with a review of the objectives that the algorithm is designed to
meet, and a description of how the algorithm internally represents the Vehicle Routing
Problem. In Chapter 5 we provide a detailed breakdown of the results obtained by
the Enhanced Bees Algorithm. The algorithm is tested against the well known set of
test instances due to Christofides, Mingozzi and Toth [11] and is contrasted with some
well known results from the literature. Finally, in Chapter 6 we provide a summary of
the results achieved by the Enhanced Bees Algorithm in context of the other methods
available for solving the Vehicle Routing Problem from the literature. Additionally,
we offer our thoughts on future directions and areas that warrant further research.
1
The Enhanced Bees Algorithm was developed as part of a New Zealand Trade and Enterprise Research and Development grant for use within the dispatch software product, vWorkApp. Hence more
consideration has been given to its runtime performance than is typically afforded in the literature.
Chapter 2
Background
This chapter provides a short history and background material on the Vehicle Routing
Problem. In particular we review the solution methods that have been brought to
bear on the Vehicle Routing Problem and some of the classic results reported in the
literature.
This chapter is laid out as follows. We start in Section 2.1 by informally defining what
the Vehicle Routing Problem is and by providing a timeline of the major milestones
in its research. We also review a closely related problem, the Traveling Salesman
Problem, which is a cornerstone of the Vehicle Routing Problem. We then review in
Section 2.2 the Exact Methods that have been developed to solve the Vehicle Routing
Problem. These are distinguished from the other methods we review in that they
provide exact solutions, where the globally best answer is produced. We follow this
in Section 2.3 by reviewing the classic Heuristics methods that have been developed
for the Vehicle Routing Problem. These methods are not guaranteed to find the
globally best answer, but rather aim to produce close to optimal solutions using algorithms with fast running times that are able to scale to large problem instances. In
Section 2.4 we review Meta-heuristic methods that have been adapted for the Vehicle Routing Problem. These methods provide some of the most competitive results
available for solving the Vehicle Routing Problem and are considered state-of-the-art
currently. Lastly, in Section 2.5 we review a modern family of meta-heuristics called
Swarm Intelligence that has been inspired by the problem solving abilities exhibited
by some groups of animals and natural processes. These last methods have become
a popular area of research recently and are starting to produce competitive results to
many problems. This thesis uses a Swarm Intelligence method for solving the Vehicle
Routing Problem.
7
8
CHAPTER 2. BACKGROUND
2.1
Overview
The Vehicle Routing Problem (commonly abbreviated to VRP) describes the problem
of assigning and ordering work for a finite number of resources, such that the cost of
undertaking that work is minimised. Often the context used is that of a fleet of vehicles
delivering goods to a set of customers, although the problem can equally be applied
across many different industries and scenarios (including non-logistics scenarios, such
as microchip layout). The aim is to split the deliveries between the vehicles and to
specify an order in which each vehicle undertakes its work, such that the distance
travelled by the vehicles is minimised and any pre-stated constraints are met. In the
classic version of the VRP the constraints that must be met are:
1. Each vehicle must start and end its route at the depot.
2. All goods must be delivered.
3. The goods can only be dropped off a single time and by a single vehicle.
4. Each good requires a specified amount of capacity. However, each vehicle has a
finite amount of capacity that cannot be exceeded. This adds to the complexity
of the problem as it necessarily influences the selection of deliveries assigned to
each vehicle.
Figure 2.1: An example of customers being assigned to three different
vehicle routes. The depot is the black dot in the centre.
More formally, the VRP can be represented as a graph, (V, E). The vertices of the
graph, V , represent all locations that can be visited; this includes each customer
location and the location of the depot. For convenience let v d denote the vertex that
represents the depot. We denote the set of customers as C = {1, 2, . . . , n}. Next
let the set of edges, E, correspond to the valid connections between customers and
connections to the depot—typically for the VRP all connections are possible. Each
edge, (i, j) ∈ E, has a corresponding cost cij . This cost is typically the travel distance
between the two locations.
2.1. OVERVIEW
9
A solution to a given VRP instance can be represented as a family of routes, denoted
by S. Each route itself is a sequence of customer visits that are performed by a single
vehicle, denoted by R = [v1 , v2 , . . . , vk ] such that vi ∈ V , and v1 , vk = v d . Each
customer has a demand di , i ∈ C, and q is the maximum demand permissible for any
route (i.e. its maximum capacity). The cost of the solution, and the value we aim to
minimise, is given by the following formula:
X X
cvi ,vi+1
R∈S vi ∈R
We can now formalise the VRP constraints as follows:
[
=V
(2.1)
R∈S
(Ri − v d ) ∩ Rj = ∅
(2.2)
d
vi = vj
v0 , vk ∈ Ri = v
X
dv < q
∀Ri , Rj ∈ S
d
∀Ri ∈ S, ¬∃vi , vj ∈ (Ri − v )
(2.3)
∀Ri ∈ S
(2.4)
∀Ri ∈ S
(2.5)
v∈Ri
Equation (2.1) specifies that all customers are included in at least one route. Equations (2.2) and (2.3) ensure that each customer is only visited once, across all routes.
Equation (2.4) ensures that each route starts and ends at the depot. Lastly, Equation (2.5) ensures that each route doesn’t exceed its capacity.
This version of the problem has come to be known as the Capacitated Vehicle Routing
Problem (often appreciated to CVRP in the literature). See Chapter 3 for an alternative formation, which states the problem as an Integer Linear Programming problem,
as is more standard in the VRP literature1 .
VRP was first formally introduced in 1959 by Dantzig and Ramser in their paper,
the Truck Scheduling Problem [14]. The VRP has remained an important problem in
logistics and transport, and is one of the most studied of all combinatorial optimisation
problems. Hundreds of papers have been written on it over the intervening fifty
years. From the large number of implementations in use today it is clear that the
VRP has real benefits to offer transport and logistics companies. Anywhere from
1
We believe that the formation provided in this chapter is simpler and more precise for understanding the algorithmic methods described in this chapter. However, we do provide a more standard
formation in Chapter 3.
10
CHAPTER 2. BACKGROUND
5% to 20% savings have been reported where a vehicle routing procedure has been
implemented [52].
From the VRP comes a family of related problems. These problems model other constraints that are encountered in real world applications of the VRP. Classic problems
include: VRP with Time Windows (VRPTW), which introduces a time window constraint against each customer that the vehicle must arrive within; VRP with Multiple
Depots (MDVRP), where the vehicles are dispatched from multiple starting points;
and the Pickup and Delivery Problem (PDP), where goods are both picked up and
delivered during the course of the route (such as a courier would do).
2.1.1
TSP Introduction and History
The VRP is a combination of two problems that are combinatorial hard in themselves:
the Traveling Salesman Problem (more precisely the Multiple Traveling Salesman
Problem), and the Bin Packing Problem.
The Traveling Salesman Problem (TSP) can informally be defined as follows. Given
n points on a map, provide a route through each of the n points such that each point
is only used once and the total distance travelled is minimised. The problem’s name,
the Traveling Salesman, comes from the classic real world example of the problem. A
salesman is sent on a trip to visit n cities. They must select the order in which to
visit the cities, such that they travel the least amount of distance.
Although the problem sounds like it might be easily solvable, it is in fact N P-hard.
The best known exact algorithms for solving the TSP still require a running time
of O(2n ). Karp’s famous paper, Reducibility Among Combinatorial Problems [26], in
1972 showed that the Hamiltonian Circuit problem is N P-complete. This implied
the NP-hardness of TSP, and thus supplied the mathematical explanation for the
apparent difficulty of finding optimal traveling salesman tours.
TSP has a history reaching back many years. It is itself related to another classic graph
theory problem, the Hamiltonian circuit. Hamiltonian circuits have been studied since
1856 by both Hamilton [24] and Kirkman [27]. Whereas the TSP has been informally
discussed for many years [46], it didn’t become actively studied until after 1928, where
Menger, Whitney, Flood and Robinson produced much of the early results in the field.
Robinson’s RAND report [43] is probably the first article to call the problem by the
name it has since become known as, the Traveling Salesman Problem.
2.1. OVERVIEW
11
From Robinson’s RAND report:
The purpose of this note is to give a method for solving a problem related
to the traveling salesman problem. One formulation is to find the shortest route for a salesman starting from Washington, visiting all the state
capitals and then returning to Washington. More generally, to find the
shortest closed curve containing n given points in the plane.
Figure 2.2: Shown is an example of a 16 city TSP tour. This tour is one
of 20,922,789,888,000 possible tours for these 16 cities.
An early result was provided by Dantzig, Fulkerson, and Johnson [13]. Their paper
gave an exact method for solving a 49 city problem, a large number of cities for the
time. Their algorithm used the cutting plane method to provide an exact solution.
This approach has been the inspiration for many subsequent approaches, and is still
the bedrock of algorithms that attempt to provide an exact solution.
A generalisation of the TSP is Multiple Traveling Salesman Problem (MTSP), where
multiple tours are constructed (i.e. multiple salesman can be used to visit the cities).
The pure MTSP can trivially be turned into a TSP by constructing a graph G with
n − 1 additional copies of the starting vertex and by forbidding travel directly between
the n starting vertices. However, the pure formulation of MTSP places no additional
constraints on how the routes are constructed. Real life applications of the MTSP
typically require additional constraints, such as limiting the size or duration of each
route (i.e. one salesman shouldn’t be working a 12 hour shift, while another has no
work).
MTSP leads us naturally into the family of problems given by the VRP. VRP, and
its family of related problems, can be understood as being a generalisation of MTSP
that incorporates additional constraints. Some of these constraints, such as capacity
limits, introduce additional dimensions to the problem that are in themselves hard
combinatorial problems.
12
2.2
CHAPTER 2. BACKGROUND
Exact Methods
The first efforts at providing a solution to the VRP were concerned with exact methods. These started by sharing many of the techniques brought to bear on TSP. We
follow Laporte and Nobert’s survey [29] and classify exact methods for the VRP into
three families: Direct Tree Search methods, Dynamic Programming, and Integer Linear Programming.
The first classic Direct Tree Search results are due to Christolds and Ellison. Their
1969 paper provided the first branch and bound algorithm for exactly solving the
VRP [10]. Unfortunately its time and memory requirements were large enough that it
was only able to solve problems of up to 13 customers. This result was later improved
upon by Christolds in 1976 by using a different branch model. This improvement
allowed him to solve for up to 31 customers.
Christofides, Mingozzi, and Toth [11], provide a lower bound method that is sufficiently
quick (in terms of runtime performance) to be used as a lower bound for excluding
nodes from the search tree. Using this lower bound they were able to provide solutions
for problems containing up to 25 customers. Laporte, Mercure and Nobert [28] used
MTSP as a relaxation of the VRP within a branch and bound framework to provide
solutions for more realistically sized problems, containing up to 250 customers.
A Dynamic Programming approach was first applied to the VRP by Eilon, WatsonGandy and Christofides [16]. Their approach allowed them to solve exactly for problems of 10 to 25 customers. Since then, Christofides has made improvements to this
algorithm to solve exactly for problems up to fifty customers.
A Set Partitioning method was given by Balinski, and Quandt in 1964 [5] to produce
exact VRP solutions. However, the problem sets they used were very small, only
containing between 5 to 15 customers; and even then they were not able to produce
solutions for some of the problems. However, taking their approach as a starting
point, many authors have been able to produce more powerful methods. Rao and
Zionts [40], Foster and Ryan [18], and Desrochers, Desrosiers and Solomon [15] have
all extended the basic set partitioning algorithm using the Column Generation method
from Integer Programming. These later papers have produced some of the best exact
results.
Notwithstanding the preceding discussion, exact methods have been of more use in
advancing the theoretical understanding of the VRP than they have been in providing
solutions to real life routing problems. This can mostly be attributed to the fact that
real life VRP instances often involve at least tens of customers (and often hundreds),
and involve richer constraints than are modelled in the classic VRP.
2.3. CLASSIC HEURISTICS
2.3
13
Classic Heuristics
In this section we review the classic heuristic methods that have been developed for the
VRP. These methods are not guaranteed to find the globally best answer, but rather
aim to produce close to optimal solutions using algorithms with fast running times
that are able to scale to large problem instances. Classic heuristics for the VRP can
be classified into three families: constructive heuristics; two-phase heuristics, which
can again be divided into two subfamilies, cluster first and then route, and route first
and then cluster; and finally improvement methods.
2.3.1
Constructive Heuristics
We start by looking at the Constructive Heuristics. Constructive heuristics build a
solution from the ground up. They typically provide a recipe for building each route,
such that the total cost of all routes is minimised.
A trivial but intuitive constructive heuristic is the Nearest Neighbour method. In this
method routes are built up sequentially. At each step the customer nearest to the
last routed customer is chosen. This continues until the route reaches its maximum
capacity, at which point a new route is started. In practice the Nearest Neighbour
algorithm tends to provide poor results and is rarely used.
j
i
Figure 2.3: Shown is an example of the Nearest Neighbour method being
applied. A partially constructed route selects customer j to add, as its closest
to the last added customer, i.
An early and influential result was given by Clarke and Wright in their 1964 paper [12].
In their paper they present a heuristic extending Dantzig and Ramser’s earlier work,
which has since become known as the Clarke Wright Savings heuristic. The heuristic
is based on the simple premise of iteratively combining routes in order of those pairs
that provide the largest saving.
14
CHAPTER 2. BACKGROUND
j
i
R'
Figure 2.4: Clark Wright Savings Algorithm. Customers i, j are selected
as candidates to merge. The merge results in a new route R0 .
The algorithm works as follows:
Algorithm 1: Clark Write Savings Algorithm
initialiseRoutes()
M = savingsMatrix(V )
L = sortBySavings(SM )
for lij ← L do
Ri , Rj = findRoutes(lij )
if feasibleMerge(Ri , Rj ) then
combineRoute(Ri , Rj )
end
end
The algorithm starts by initialising a candidate solution. For this it creates a route
R = [v d , vi , v d ] for all v ∈ V . It then calculates a matrix M that contains the savings
sij = ci0 + cj0 − cij for all edges (i, j) ∈ E. It then produces a list, L, that enumerates
each cell i, j of the matrix in descending order of the savings. For each entry in the
list, lij ∈ L, it selects the two routes, Ri , Rj , that contain customers i, j ∈ V and tests
to see if the two routes can be merged. A merge is permissible if and only if:
1. Ri 6= Rj .
2. i, j are the first or last vertices (excluding the depot v d ) of their respective routes.
3. The combined demand of the two routes doesn’t exceed the maximum allowed,
q.
The heuristic comes in two flavours, sequential and parallel. The sequential version
adds the additional constraint that only one route can be constructed at a time. In this
case one of the two routes considered, Ri , Rj , must be the route under construction. If
2.3. CLASSIC HEURISTICS
15
neither of the routes are the route under construction then the list item is ignored and
processing continues down the list. If the merge is permissible then we merge routes
Ri , Rj such that R0 = [v0 , . . . , i, j, . . . , vk ]. In the parallel version, once the entire list
of savings has been enumerated then the resulting solution is returned as the answer.
In the sequential version the for loop is repeated until no feasible merges remain.
The Clark Write Savings heuristic has been used to solve problems of up to 1000
customers with results often within 10% of the optimal solution using only a 180
seconds of runtime [52]. The parallel version of the Clark Write Savings Algorithm
outperforms the sequential version in most cases [30] and is typically the one employed.
The heuristic has proven to be surprisingly adaptable and has been extended to deal
with more specialised vehicle routing problems where additional objectives and constraints must be factored in. Its flexibility is a result of its algebraic treatment of the
problem [30]. Unlike many other VRP heuristics that exploit the problem’s spatial
properties (such as many of the two-phase heuristics, see Section 2.3.2), the savings
formula can easily be adapted to take into consideration other objectives. An example
of this is Solomon’s equally ubiquitous algorithm [48] which extends the Clark Wright
Savings algorithm to cater for time constraints.
This classic algorithm has been extended by Gaskell [19], Yellow [55] and Paessens [37],
who have suggested alternatives to the savings formulas used by Clarke and Wright.
These approaches typically introduce additional parameters to guide the algorithm
towards selecting routes with geometric properties that are likely to produce better
combinations. Altinkemer and Gavish provide an interesting variation on the basic
savings heuristic [4]. They use a matching algorithm to combine multiple routes in
each step. To do this they construct a graph such that each vertex represents a route,
each edge represents a feasible saving, and the edges’ weights represent the savings
that can be realised by the merge of the two routes. The algorithm proceeds by solving
a maximum cost weighted matching of the graph.
2.3.2
Two-phase Heuristics
We next look at two-phase heuristics. We start by looking at the cluster first, route
second subfamily. One of the foundational algorithms for this method is given to us
by Gillett and Miller who provided a new approach called the Sweep Algorithm in
their 1974 paper [22]. This popularised the two-phase approach, although a similar
method was suggested earlier by Wren in his 1971 book, and subsequently in Wren
and Holliday’s 1972 paper [54]. In this approach, an initial clustering phase is used
to cluster the customers into a base set of routes. From here the routes are treated as
separate TSP instances and optimised accordingly. The two-phase approach typically
doesn’t prescribe a method for how the TSP is solved and assumes that already developed TSP methods can be used. The classic Sweep algorithm uses a simple geometric
method to cluster the customers. Routes are built by sweeping a ray, centered at
16
CHAPTER 2. BACKGROUND
the depot, clockwise around the space enclosing the problem’s locations. The Sweep
method is surprisingly effective and has been shown to solve several benchmark VRP
problems to within 2% to 9% of the best known solutions [52].
i
Figure 2.5: This diagram shows an example of the Sweep process being run.
The ray is swept clockwise around the geographic area. In this example one
route has already been formed, and a second is about to start at customer i.
Fisher and Jaikumars’s 1981 paper [17] builds upon the two-phase approach by providing a more sophisticated clustering method. They solve a General Assignment
Problem to form the clusters instead. A limitation of their method is that the amount
of vehicle routes must be fixed up front. Their method often produces results that are
1% to 2% better than results produced by the classic Sweep algorithm [52].
Christofides, Mingozzi, and Toth expanded upon this approach in [11] and proposed a
method that uses a truncated branch and bound technique (similar to Christofides’s
Exact method). At each step it builds a collection of candidate routes for a particular
customer, i. It then evaluates each route by solving it as a TSP, from which it then
selects the shortest TSP as the route.
The Petal algorithm is a natural extension to the Sweep algorithm. It was first proposed by Balinski and Quandt [5] and then extended by Foster and Ryan [18]. The
basic process is to produce a collection of overlapping candidate routes (called petals)
and then to solve a set partitioning problem to produce a feasible solution. As with
other two-phase approaches it is assumed that the order of the customers within each
route is solved using an existing TSP heuristic. The petal method has produced competitive results for small solutions, but quickly becomes impractical where the set of
candidate routes that must be considered is large.
Lastly, there are route first, cluster second methods. The basic premise of these
techniques are to first construct a ‘grand’ TSP tour such that all customers are visited.
The second phase is then concerned with splitting this tour into feasible routes. Route
first, cluster second methods are generally thought to be less competitive than other
methods [30], although interestingly, Haimovich and Rinnooy Kan have shown that if
all customers have unit demand then a simple shortest path algorithm (which can be
solved in polynomial time) can be used to produce a solution from a TSP tour that
2.3. CLASSIC HEURISTICS
17
is asymptotically optimal [23].
2.3.3
Iterative Improvement Heuristics
Iterative Improvement methods follow an approach where an initial candidate solution is iteratively improved by applying an operation that improves the candidate
solution, typically in a small way, many thousands of times. The operations employed
are typically simple and only change a small part of the candidate solution, such as
the position of a single customer or edge within the solution. The set of solutions that
are obtainable from the current candidate solution, S, by applying an operator Op is
known as S’s neighbourhood. Typically, with Iterative Improvement heuristics, a new
solution S 0 is selected by exhaustively searching the entire neighbourhood of S for the
best improvement possible. If no improvement can be found then the heuristic terminates. The initial candidate solution (i.e. the starting point of the algorithm) can be
randomly selected or can be produced using another heuristic. Constructive Heuristics are typically used for initially seeding an improvement heuristic, see Section 2.3.1
for more information on these.
Probably one of the best known improvement operators is 2-Opt. The 2-Opt operator takes two edges (i, j), (k, l) ∈ T , where T are the edges traversed by a particular
route R = [v1 , . . . , vi , vj , . . . , vk , vl , . . . , vn ], and removes these from the candidate solution. This splits the route into two disconnected components, D1 = [vj , . . . , vk ], D2 =
[v1 , . . . , vi , vl , . . . , vn ]. A new candidate solution is produced by reconnecting D1 to
D2 using the same vertices i, j, l, k but with alternate edges, such that (i, j), (j, l) ∈ T .
j
j
l
l
k
k
i
R
i
R'
Figure 2.6: This diagram shows 2-Opt being applied to a candidate solution
R and producing a new solution R0 . In this example edges (i, j), (k, l) are
exchanged with edges (i, k), (j, l).
The rationale behind 2-Opt is that, due to the triangle inequality, edges that cross
themselves are unlikely to be optimal. 2-Opt aims to detangle a route.
18
CHAPTER 2. BACKGROUND
There are a number of other operations suggested in the literature. Christofides and
Eilon give one of the earliest iterative improvement methods in their paper [10]. In the
paper they make a simple change to 2-Opt to increase the amount of edges removed
from two to three—the operation fittingly being called 3-Opt. They found that their
heuristic produced superior results than 2-Opt.
In general, operations such as 3-Opt, that remove edges and then search for a more
optimal recombination of components take O(ny ) where y is the number of edges
removed. A profitable strain of research has focused on producing operations that
reduce the amount of recombinations that must be searched. Or presents an operation
that has since come to be known as Or-Opt [35]. Or-Opt is a restricted 3-Opt. It
searches for a relocation of all sets of 3 consecutive vertices (which Or calls chains),
such that an improvement is made. If an improvement cannot be made then it tries
again with chains of 2 consecutive vertices, and so on. Or-Opt has been shown to
produce similar results to that of 3-Opt, but with a running time of O(n2 ). More
recently Renaud, Boctor, and Laptorte [42] have presented a restricted version of 4Opt, called 4-Opt*, that operates in a similar vein to Or-Opt. 4-Opt* has a running
time of O(wn2 ) where w denotes the number of edges spanned by 4-Opt* when building
a chain.
Iterative improvement heuristics are often used in combination with other heuristics.
In this case they are run on the candidate solution after the initial heuristic has
completed. However, if used in this way there is often a fine balance between producing
an operation that improves a solution, and one that is sufficiently destructive enough
to escape a local minimum. Interest in Iterative Improvement heuristics has grown as
the operations developed for them, such as Or-Opt, are directly applicable to more
modern heuristics, such as the family known as meta-heuristics presented in the next
section.
2.4
Meta-heuristics
Meta-heuristics are a broad collection of methods that make few or no assumptions
about the type of problem being solved. They provide a framework that allows for
individual problems to be modelled and ‘plugged in’ to the meta-heuristic. Typically, meta-heuristics take an approach where a candidate solution (or solutions) is
initially produced and then is iteratively refined towards the optimal solution. Intuitively meta-heuristics can be thought of searching a problem’s search space. Each
iteration searches the neighbourhood of the current candidate solution(s) looking for
new candidate solutions that move closer to the global optimum.
A limitation of meta-heuristics is that they are not guaranteed to find an optimal
solution (or even a good one!). Moreover, the theoretical underpinnings of what
makes one meta-heuristic more effective than another are still poorly understood.
2.4. META-HEURISTICS
19
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
10
10
5
5
0
0
-5
-5
-10
-10
Figure 2.7: This diagram shows an example of a search space that a metaheuristic moves through. In this example the peak at the centre of the figure
is the globally best answer, but there are also hills and valleys that the metaheuristic may become caught in. These are called local minima and maxima.
Meta-heuristics within the literature tend to be tuned for specific problems and then
validated empirically.
There have been a number of meta-heuristics produced for the VRP in recent years
and many of the most competitive results produced in the last ten years are due to
them. We next review some of the more well known meta-heuristic results for the
VRP.
2.4.1
Simulated Annealing
Simulated Annealing is inspired by the annealing process used in metallurgy. The
algorithm starts with a candidate solution (which can be randomly selected) and then
moves to nearby solutions with a probability dependent on the quality of the solution
and a global parameter T , which is reduced over the course of the algorithm. In classic
implementations the following formula is used to control the probability of a move:
e−
f (s0 )−f (s)
T
Where f (s) and f (s0 ) represent the solution quality of the current solution, and the
new solution respectively. By analogy with the metallurgy process, T represents the
current temperature of the solution. Initially T is set to a high value. This lets the
algorithm free itself from any local optima that it may be caught in. It is then cooled
over the course of the algorithm forcing the search to converge on a solution.
20
CHAPTER 2. BACKGROUND
One of the first Simulated Annealing results for the VRP was given by Robuste,
Daganzo and Souleyrette [44]. They define the search neighbourhood as being all
solutions that can be obtained from the current solution by applying one of two
operations: relocating part of a route to another position within the same route,
or exchanging customers between routes. They tested their solution on some large
real world instances of up to 500 customers. They reported some success with their
approach, but as their test cases were unique, no direct comparison is possible.
Osman has given the best known Simulated Annealing results for the VRP [36]. His
algorithm expands upon many areas of the basic Simulated Annealing approach. The
method starts by using the Clark and Wright algorithm to produce an initial position.
It defines its neighbourhood as being all candidate solutions that can be reached by
applying an operator he names the λ-interchange operation.
λ-interchange works by selecting two sequences (i.e. chains) of customers Cp , Cq from
two routes, Rp and Rq , such that |Cp |, |Cq | < λ (note that the chains are not necessarily
of the same length). The customers within each chain are then exchanged with each
other in turn, until an exchange produces an infeasible solution. As the neighbourhood
produced by λ-interchange is typically quite large, Osman restricts λ to being less than
2 and suggests that the first move that provides an improvement is used rather than
exhaustively searching the entire neighbourhood.
Rq
Rp
Rq
Rp
Figure 2.8: This diagram shows an example of λ-interchange being applied
to a candidate solution. Two sequences of customers Cp and Cq are selected
from routes Rp , Rq respectively. Customers from Cp are then swapped with
Cq where feasible.
Osman also uses a sophisticated cooling schedule. His main change being that the
temperature is cooled only while improvements are found. If no improvement is found
he then resets the temperature using Ti = max( T2r , Tb ), where Tr is the reset temperature, and Tb is temperature of the best solution found so far.
Although Simulated Annealing has produced some good results, and in many cases
outperforms classic heuristics (compare [30] with [21]), it is not competitive with the
Tabu Search methods discussed in Section 2.4.3.
2.4. META-HEURISTICS
2.4.2
21
Genetic Algorithms
Genetic Algorithms were first proposed in [25]. They have since been applied to many
problem domains and are particularly well suited to applications that must work
across a number of different domains. In fact they were the first evolutionary-inspired
algorithm to be applied to combinatorial problems [39]. The basic operation of a
Genetic Algorithm is as follows:
Algorithm 2: Simple Genetic Algorithm
Generate the initial population
while termination condition not met do
Evaluate the fitness of each individual
Select the fittest pairs
Mate pairs and produce next generation
Mutate (optional)
end
In a classic Genetic Algorithm each candidate solution is encoded as a binary string
(i.e. chromosome). Each individual (i.e. candidate solution) is initially created randomly and used to seed the population. A technique often employed in the literature
is to initially ‘bootstrap’ the population by making use of another heuristic to produce the initial population. However, special care must be taken with this approach
to ensure that diversity is maintained across the population, as you risk premature
convergence by not introducing enough diversity in the initial population.
Next, the fittest individuals are selected from the population and are mated in order
to produce the next generation. The mating process uses a special operator called
a crossover operator that takes two parents and produces offspring from these by
combining parts of each parent. Optionally, a mutation operation is also applied, that
introduces a change that doesn’t exist in either parent. The classic crossover operation
takes two individuals encoded as binary strings and splits these at one or two points
along the length of the string. The strings are then recombined to form a new binary
string, which in turn encodes a new candidate solution. The entire process is continued
until a termination condition is met (often a predetermined running time), or until
the population has converged on a single solution.
Special consideration needs to be given to how problems are encoded and to how
the crossover and mutation operators work when using Genetic Algorithms to solve
discrete optimisation problems, such as the VRP. For example, the classic crossover
operation, which works on binary strings, would not work well on a TSP tour. When
two components of two tours are combined in this way they are likely to contain
duplicates. Therefore, it is more common for the VRP (and the TSP) to use a direct
representation and to use specially designed crossover operators. In this instance the
VRP is represented as a set of sequences, each holding an ordered list of customers.
22
CHAPTER 2. BACKGROUND
The crossover operators are then designed so that they take into consideration the
constraints of the VRP.
Two crossover operators commonly used with combinatorial problems are the Order
Crossover (OX) and the Edge Assembly Crossover (EAX). OX [34] operates by
selecting two cut points within each route. The substring between the two cut points
is copied from the second parent directly into the offspring. Likewise, the string
outside the cut points is copied from the first parent into the offspring, but with any
duplicates removed. This potentially leaves a partial solution, where not all customers
have been routed. The partial solutions is then repaired by inserting any unrouted
customers into the child in the same order that they appeared in the second parent.
3 2 1 4 5 6 7
3 2 6 7 1 4 5
2 5 6 7 1 3 4
Figure 2.9: This diagram shows the OX crossover operator being applied to
two tours from a TSP. A child is produced by taking customers at position
3, 4, 5 from the second parent and injecting these into the same position
in the first parent, removing any duplicates. This leaves customers 4, 5
unrouted, so they are reinserted back into the child in the order that they
appear in the first parent.
Another common crossover operator is EAX. EAX was originally designed for the
TSP but has been adapted to the VRP by [32]. EAX operates using the following
process:
1. Combine the two candidate solutions into a single graph by merging each solution’s edge sets.
2. Create a partition set of the graph’s cycles by alternately selecting an edge from
each graph.
3. Randomly select a subset of the cycles.
4. Generate a (incomplete) child by taking one of the parents and removing all
edges from the selected subset of cycles, then add back in the edges from the
parent that wasn’t chosen.
5. Not all cycles in the child are connected to the route. Repair them by iteratively
merging the disconnected cycles to the connected cycles.
An alternative and interesting approach found in the literature is to instead encode
a set of operations and parameters that are fed to another heuristic, that in turn
2.4. META-HEURISTICS
23
P1
P2
Figure 2.10: This diagram shows an example of of EAX being applied
on two parent solutions, P 1 and P 2. The parents are first merged together.
Then a new graph is created by selecting alternate edges from each parent
P 1, P 2. A subset of cycles are then taken and applied to P 1, such that
any edges from P 1 are removed. The child solution produced is infeasible
(it contains broken routes). These would need to be repaired.
produces a candidate solution. A well known example of this approach was suggested
in [7] which encoded an ordering of the customers. The ordering is then fed into an
insertion heuristic to produce the actual candidate solutions.
An influential result that uses Genetic Algorithms to solve VRPTW is given in [50]
with their GIDEON algorithm. GIDEON uses an approach inspired by the Sweep
method (an overview of the Sweep method is provided in Section 2.3.2). It builds
routes by sweeping a ray, centered at the depot, clockwise around the geographic space
enclosing the customer’s locations. Customers are collected into candidate routes
based on a set of parameters that are refined by the Genetic Algorithm. GIDEON
uses the Genetic Algorithm to evolve the parameters used by the algorithm, rather
than to operate on the problem directly. Finally, GIDEON uses a local search method
to optimise customers within each route, making use of the λ-interchange operator (A
description of this operator is provided in Section 2.4.1).
Generally speaking, Genetic Algorithms are not as competitive as other meta-heuristics
24
CHAPTER 2. BACKGROUND
at solving the VRP. However, more recently there have been two very promising applications of Genetic Algorithms being used to solve the VRP. Nagata [32] has adapted
the EAX operator for use with the VRP. And Berger and Barkaoui have presented a
Hybrid Genetic Algorithm called HGA-VRP in [6]. HGA-VRP adapts a construction
heuristic for use as a crossover operator. The basic premise is to select a set of routes
from each parent that are located close to one another. Customers are then removed
from one parent and inserted into the second using an operation inspired by Solomon’s
construction heuristic for VRPTW [48].
Both methods have reached the best known solution for a number of the classic VRP
benchmark instances by Christofides, Mingozzi and Toth [11] and are competitive with
the best Tabu Search methods.
2.4.3
Tabu Search
Tabu Search follows the general approach shared by many meta-heuristics; it iteratively improves a candidate solution by searching for improvements within the current
solution’s neighbourhood. Tabu search starts with a candidate solution, which may be
generated randomly or by using another heuristic. Unlike Simulated Annealing, the
best improvement within the current neighbourhood is always taken as the next move.
This introduces the problem of cycling between candidate solutions. To overcome this
Tabu Search introduces a list of solutions that have already been investigated and are
forbidden as next moves (hence its name).
The first instance of Tabu Search being used for VRP is by Willard [53]. Willard’s
approach made use of the fact that VRP instances can be transformed into MTSP
instances and solved. The algorithm uses a combination of simple vertex exchange
and relocation operations. Although opening the door for further research, its results
weren’t competitive with the best classic heuristics.
Osman gives a more competitive use of Tabu Search in [36]. As with his Simulated
Annealing method he makes use of the λ-interchange operation to define the search
neighbourhood. Osman provides two alternative methods to control how much of
the neighbourhood is searched for selecting the next move: Best-Improvement (BI)
and First-Improvement (FI). Best-Improvement searches the entire neighbourhood
and selects the move that is the most optimal. First-Improvement searches only until
a move is found that is more optimal than the current position. Osman’s heuristic
produced competitive results that outperformed many other heuristics. However, it
has since been refined and improved upon by newer Tabu Search methods.
Toth and Vigo introduced the concept of Granular Tabu Search (GTS) [51]. Their
method makes use of a process that removes moves from the neighbourhood that are
unlikely to produce good results. They reintroduce these moves back into the process
if the algorithm is stuck in a local minimum. Their idea follows from an existing idea
2.4. META-HEURISTICS
25
known as Candidate Lists. Toth and Vigo’s method has produced many competitive
results.
Taillard has provided one of the most successful methods for solving the VRP in his
Tabu Search method in [49]. Talliard’s Tabu Search uses Or’s λ-interchange as its
neighbourhood structure. It borrows two novel concepts from [20]: the use of a more
sophisticated tabu mechanism, where the duration (or number of iterations) that an
item is tabu for is chosen randomly; and a diversification strategy, where vertices that
are frequently moved without giving an improvement are penalised. A novel aspect
of Taillard’s algorithm is its decomposition of the problem into sub-problems. Each
problem is split into regions using a simple segmentation of the region centred about
the depot (Taillard also provides an alternative approach for those problems where
the customers are not evenly distributed around the depot). From here each subproblem is solved individually, with customers being exchanged between neighbouring
segments periodically. Taillard observes that exchanging customers beyond geographically neighbouring segments is unlikely to produce an improvement, so these moves
are safely ignored. Taillard’s method has produced some of the currently best known
results for the standard Christofides, Mingozzi and Toth problem sets [11].
2.4.4
Large Neighbourhood Search
Large Neighbourhood Search (commonly abbreviated to LNS) was recently proposed
as a heuristic by Shaw [47]. Large Neighbourhood Search is a type of heuristic belonging to the family of heuristics known as Very Large Scale Neighbourhood search
(VLSN)2 . Very Large Scale Neighbourhood search is based on a simple premise; rather
than searching within a neighbourhood of solutions that can be obtained from a single (and typically quite granular) operation, such as 2-opt, it might be profitable
to consider a much broader neighbourhood—a neighbourhood of candidate solutions
that are obtained from applying many simultaneous changes to a candidate solution.
What distinguishes these heuristics from others is that the neighbourhoods under consideration are typically exponentially large, often rendering them infeasible to search.
Therefore much attention is given to providing methods that can successfully traverse
these neighbourhoods.
2
LNS is somewhat confusingly named given that it a type of VLSN, and not a competing approach.
26
CHAPTER 2. BACKGROUND
Large Neighbourhood Search uses a Destroy and Repair metaphor for how it searches
within its neighbourhood. Its basic operation is as follows:
Algorithm 3: Large Neighbourhood Search
x = an initial solution
while termination condition not met do
xt = x
destroy(xt )
repair(xt )
if xt better than current solution then
x = xt
end
end
Result: x
It starts by selecting a starting position. This can be done randomly or by using
another heuristic. Then for each iteration of the algorithm a new position is generated
by destroying part of the candidate solution and then by repairing it. If the new
solution is better than the current solution, then this is selected as the new position.
This continues until the termination conditions are met. Large Neighbourhood Search
can be seen as being a type of Very Large Scale Neighbourhood search because at
each iteration the number of neighbouring solutions is exponentially large, based on
the number of items removed (i.e. destroyed).
Obviously the key components of this approach are the functions used to destroy and
repair the solution. Care must be given to how these functions are constructed. They
must pinpoint an improving solution from a very large neighbourhood of candidates,
while also providing enough degrees of freedom to escape a local optimum.
Empirical evidence in the literature shows that even surprisingly simple destroy and
repair functions can be effective [47] [45]. In applications of Large Neighbourhood
Search for VRP a pair of simple operations are commonly used (often alongside more
complex ones too) for the destroy and repair functions. Specifically, the solution is
destroyed by randomly selecting and removing n customers. It is then repaired by
finding the least cost reinsertion points back into the solution of the n customers.
Shaw applied Large Neighbourhood Search to VRP in his original paper introducing
the method [47]. In this he introduced a novel approach for his destroy and repair
functions. The destroy function removes a set of ‘related’ customers. He defines a
related customer to be any two customers that share a similar geographic location,
that are sequentially routed, or that share a number of similar constraints (such as
overlapping time windows if time constraints are used). The idea of removing related
customers, over simply removing random customers, is that related customers are
more likely to be profitably exchanged—or stated another way, unrelated customers
are more likely to be reinserted back in their original positions. Shaw’s repair func-
2.5. SWARM INTELLIGENCE
27
tion makes use of a simple branch and bound method that finds the minimum cost
reinsertion points within the partial solution. His results were immediately impressive
and reached many of the best known solutions on the Christofides, Mingozzi and Toth
problems [11].
More recently Ropke proposed an extension to the basic Large Neighbourhood Search
process in [45]. His method adds the concept of using a collection of destroy and repair
functions, rather than using a single pair. Which function to use is selected at each
iteration based on its previous performance. In this way the algorithm adapts itself
to use the most effective function to search the neighbourhood.
Ropke makes use of several destroy functions. He uses a simple random removal
heuristic, Shaw’s removal heuristic, and a worst removal heuristic, which removes the
most costly customers (in terms of that customer’s contribution to the route’s overall
cost). Likewise, he makes use of several different insertion functions. These include
a simple greedy insertion heuristic, and a novel insertion method he calls the ‘regret
heuristic’. Informally, the regret heuristic reinserts those customers first who are
most impacted (in terms of increased cost) by not being inserted into their optimum
positions. Specifically, let U be the set of customers to be reinserted and let xik be a
variable that gives the k’th lowest cost for inserting customer i ∈ U into the partial
solution. Now let c∗i = xi2 − xi1 , in other words the cost difference between inserting
customer i into its second best position and its first. Now in each iteration of the
repair function choose a customer that maximises:
max c∗i
i∈U
Ropke presents a series of results that show that his Large Neighbourhood Search is
very competitive for solving the VRP and its related problems (i.e. VRPTW, PDPTW,
and DARP). Considering that Large Neighbourhood Search was only proposed in
1998, it has been very successful. In a short space of time it has attracted a large
amount of research and has produced some of the most competitive results for solving
the VRP.
2.5
Swarm Intelligence
A recent area of research is in producing heuristics that mimic certain aspects of swarm
behaviour. Probably the most well known heuristics in this family are Particle Swarm
Optimisation (PSO) and Ant Colony Optimisation (ACO). Real life swarm intelligence
is interesting to combinatorial optimisation researchers as it demonstrates a form of
emergent intelligence, where individual members with limited reasoning capability and
simple behaviours, are still able to arrive at optimal solutions to complex resource
allocation problems.
28
CHAPTER 2. BACKGROUND
In the context of combinatorial optimisation, these behaviours can be mimicked and
exploited. Algorithms that make use of this approach produce their solutions by
simulating behaviour across a number of agents, who in themselves, typically only
perform rudimentary operations. A feature of this class of algorithms is the ease with
which they can be parallelised, making them more easily adaptable to large scale
problems.
Swarm Intelligence algorithms have been employed to solve a number of problems. We
look at two examples here, Ant Colony Optimisation and the Bees Algorithm, which
this thesis makes use of.
2.5.1
Ant Colony Optimisation
Ant Colony Optimisation is inspired by how ants forage for food and communicate
promising sites back to their colony. Real life ants initially forage for food randomly.
Once they find a food source they return to the colony and in the process lay down
a pheromone trail. Other ants that then stumble upon the pheromone trail follow it
with a probability dependent on how strong (and therefore how old) the pheromone
trail is. If they do follow it and find food, they then return to the colony, thus also
strengthening the pheromone trail. The strength of the pheromone trail reduces over
time meaning that younger and shorter pheromone trails, that do not take as long to
traverse, attract more ants.
Figure 2.11: This diagram depicts how ants make use of pheromone trails
to optimise their exploitation of local food sources.
Ant Colony Optimisation mimics this behaviour on a graph by simulating ants marching along a graph that represents the problem being solved. The basic operation of
2.5. SWARM INTELLIGENCE
29
the algorithm is as follows:
Algorithm 4: Ant Colony Optimisation
Data: A graph representing the problem
while termination condition not met do
positionAnts()
while solution being built do
marchAnts()
end
updatePheromones()
end
At each iteration of the algorithm the ants are positioned randomly within the graph.
The ants are then stochastically marched through the graph until they have completed
a candidate solution (in the case of a TSP this would be a tour of all vertices). At each
stage of the march each ant selects their next edge based on the following probability
formula:
pkij = P
β
[τijα ][ηij
]
β
α
l∈N k [τil ][ηil ]
Where pkij is the probability that ant k will traverse edge (i,j), N k is the set of all edges
that haven’t been traversed by ant k yet, τ is the amount of pheromone that has been
deposited at an edge, η is the desirability of an edge (based on a priori knowledge
specific to the problem), and α and β are global parameters that control how much
influence each term has.
Once the march is complete and a set of candidate solutions have been constructed
(by each ant, k), pheromone is deposited on each edge using the following equation:
τij = (1 − ρ)τij +
m
X
∆τijk
k=1
Where 0 < ρ ≤ 1 is the pheromone persistence, and ∆τijk is a function that gives the
amount of pheromone deposited by ant k. The function is defined as:
∆τijk =
1/C k
0
if edge (i, j) is visited by ant k
otherwise
30
CHAPTER 2. BACKGROUND
Where C k represents the total distance travelled through the graph by ant k. This
ensures that shorter paths result in more pheromone being deposited.
As an example of Ant Colony Optimisation’s use in combinatorial problems, we show
how it can be applied to the TSP. We build a weighted graph with i ∈ V representing
each city to be visited and (i, j) ∈ E and wij representing the cost of travel between
each city. Then at each step of the iteration we ensure that the following constraints
are met:
• Each city is visited at most once.
• We set ηij to be equal to wij .
When the Ant Colony Optimiser starts, it positions each ant at a randomly selected
vertex (i.e. city) within the graph. Each step of an ant’s march then builds a tour
through the cities. Once an ant has completed a tour it serves as a candidate solution
for the TSP. Initially the solutions will be of low quality, so we use the length of the
tours to ensure that more pheromone is deposited on the shorter tours. At the end
of n iterations the ants will have converged on a near optimal solution (but like all
meta-heuristics there’s no guarantee that this will be the global optimum).
Figure 2.12: Shown is an example of ACO being used to solve a TSP.
Initially the ants explore the entire graph. At the end of each iteration the
more optimal tours will have more pheromone deposited on them, meaning in the next iteration the ants are more likely to pick these edges when
constructing their tour. Eventually the ants converge on a solution.
Ant Colony Optimisation has been applied to VRP by Bullnheimer, Hartl, and Strauss
in [8] [9]. They adapted the straightforward implementation used for the TSP, detailed
in the preceding discussion, by forcing the ant to create a new route each time it
exceeds the capacity or maximum distance constraint. They also use a modified edge
selection rule that takes into account the vehicle’s capacity and its proximity to the
depot. Their updated rule is given by:
2.5. SWARM INTELLIGENCE
pkij
=P
31
β
[τijα ][ηij
][sij ][κij ]
β
α
l∈N k [τil ][ηil ][sil ][κil ]
Where s represents the proximity of customers i, j to the depot, and κ = (Qi +q j )/Q—
Q, giving the maximum capacity, Qi giving the capacity already used on the vehicle,
and q j is the additional load to be added. κ influences the ants to take advantage of
the available capacity.
Bullnheimer et al.’s implementation of Ant Colony Optimisation for VRP produces
good quality solutions for the Christofides, Mingozzi and Toth problems [11], but is
not competitive with the best modern meta-heuristics.
More recently Reimann, Stummer, and Doerner have presented a more competitive
implementation of Ant Colony Optimisation for VRP [41]. Their implementation
operates on a graph where (i, j) ∈ E represent the savings of combining two routes,
as given by the classic Clark and Wright Savings heuristic (see Section 2.3.1 for more
on this heuristic). Each ant selects an ordering of how the merges are applied. This
implementation is reported to be competitive with the best meta-heuristics [39].
2.5.2
Bees Algorithm
Over the last decade, and inspired by the success of Ant Colony Optimisation, there
have been a number of algorithms proposed that aim exploit the collective behaviour
of bees. This includes: Bee Colony Optimisation, which has been applied to many
combinatorial problems, Marriage in Honey Bees Optimization (MBO) that has been
used to solve propositional satisfiability problems, BeeHive that has been used for
timetabling problems, the Virtual Bee Algorithm (VBA) that has been used for function optimisation problems, Honey-bee Mating Optimisation (HBMO) that has been
used for cluster analysis, and finally, the Bees Algorithm that is the focus of this thesis.
See [31] for a bibliography and high level overview on many of these algorithms.
The Bees Algorithm was first proposed in [38]. It is inspired by the foraging behaviour
of honey bees. Bee colonies must search a large geographic area around their hive in
order to find sites with enough pollen to sustain a hive. It is essential that the colony
makes the right choices in which sites are exploited and how much resource is expended
on a particular site. They achieve this by sending scout bees out in all directions from
the hive. Once a scout bee has found a promising site it then returns to the hive and
recruits hive mates to forage at the site too. The bee does this by performing a waggle
dance. The dance communicates the location and quality of the site (i.e. fitness). Over
time, as more bees successfully forage at the site, more are recruited to exploit the
site—in this aspect bee behaviour shares some similarities with ant foraging behaviour.
32
CHAPTER 2. BACKGROUND
Figure 2.13: Shown is the waggle dance performed by a honey bee (image
courtesy of [3]). The direction moved by the bee indicates the angle that
the other bees must fly relative to the sun to find the food source. And the
duration of the dance indicates its distance.
Informally the algorithm can be described as follows. Bees are initially sent out to
random locations. The fitness of each site is then calculated. A proportion of the
bees are reassigned to those sites that had the highest fitness values. Here each bee
searches the local neighbourhood of the site looking to improve the site’s fitness. The
remainder of the bees are sent out scouting for new sites, or in other words, they are
set to a new random position. This process repeats until one of the sites reaches a
satisfactory level of fitness—or a predetermined termination condition is met.
More formally, the algorithm operates as follows:
Algorithm 5: Bees Algorithm
B = {b1 , b2 , . . . , bn }
setToRandomPosition(B)
while termination condition not met do
sortByFitness(B)
E = {b1 , b2 , . . . , be }
R = {be+1 , be+2 , . . . , bm }
searchNeighbourhood(E ∪ {c1 , . . . , cnep })
searchNeighbourhood(R ∪ {d1 , . . . , dnsp })
setToRandomPosition(B − (E ∪ R))
end
B is the set of bees that are used to explore the search space. Initially the bees are
set to random positions. Function sortByF itness sorts the bees in order of maximum
fitness. It then proceeds by taking the m most promising sites found by the bees. It
does this by partitioning these into two sets, E, N ⊂ B. E is the first e best sites,
and represents the so called elite bees. N is the m − e next most promising sites. The
searchN eighbourhood function explores the neighbourhood around a provided set of
bees. Each site in E and N is explored. nep bees are recruited for the search of each
2.5. SWARM INTELLIGENCE
33
b ∈ E, and nsp are recruited for the search of each b ∈ N . In practice this means
that nep and nsp number of positions are explored within the neighbourhoods of E
and N ’s sites, respectively. These moves are typically made stochastically, but it is
possible for a deterministic approach to be used too. The remaining n − m bees (in
other words, those not in E and N ) are set to random positions. This is repeated
until the termination condition is met, which may be a running time threshold or a
predetermined fitness level.
The advantage promised by the Bees Algorithm over other meta-heuristics is its ability
to escape local optima and its ability to navigate search topologies with rough terrain
(such as in Figure 2.14). It achieves this by scouting the search space for the most
promising sites, and then by committing more resources to the exploration of those
sites that produce better results.
2.5
2
1.5
1
0.5
0
10
10
5
5
0
0
-5
-5
-10 -10
Figure 2.14: Shown is a search space with many valleys and hills. These
search spaces provide a challenge to meta-heuristic approaches as there are
many local minima and maxima to get caught in. The Bees Algorithm ameliorates this by searching in many different areas simultaneously.
The Bees Algorithm has been applied to manufacturing cell formation, training neural networks for pattern recognition, scheduling jobs for a production machine, data
clustering, and many others areas. See [2] for more examples and a comprehensive
bibliography. However, to the best of our knowledge, the Bees Algorithm hasn’t been
adapted for the Vehicle Routing Problem until now.
34
CHAPTER 2. BACKGROUND
Chapter 3
Problem Definition
In this chapter we provide a formal definition of the VRP and briefly describe the
variant problems that have arisen in the literature. The Capacitated Vehicle Routing
Problem (CVRP) is the more correct name for the VRP that distinguishes it from its
variants. We start in Section 3.1 by providing a formal definition of the CVRP. We
formulate it as an integer linear programming problem, as has become standard in the
VRP literature. And follow this in Section 3.2 by an overview of the VRP variants
that are commonly used.
3.1
Capacitated Vehicle Routing Problem
We formulate the CVRP here as an integer linear programming problem. Although it
is possible to solve the CVRP using an integer programming solver, this is uncommon
in practice as the best solvers are still only able to solve for small problem sizes. We
provide this formulation as it has become the lingua franca of combinatorial problems.
We start the formation by specifying the variables used within it. We represent the
CVRP on a weighted graph, G = (V, E). The vertices of the graph V represent all
locations that can be visited, this includes each customer location and the location
of the depot. For convenience we let v d denote the vertex that represents the depot,
and we denote the set of customers as C = {1, 2, . . . , n}. Thus the set of vertices
is given by V = v d ∪ C. We now let the set of edges, E, correspond to the valid
connections between customers and connections to the depot. For the CVRP all
connections are possible, in other words, we set G to be a clique. Each edge (i, j) ∈ E
has a correspondingpcost cij . We let the cost be the euclidian distance between the
two locations cij = (xj − xi )2 + (yj − yi )2 . Where xi and yi for i ∈ V represent the
coordinates of the customer’s location.
35
36
CHAPTER 3. PROBLEM DEFINITION
We use K to denote the set of vehicles that are used to visit customers, such that
|K| = m and m is the maximum number of vehicles allowed. We define q and t to be
the maximum capacity and the maximum work duration, respectively, allowable for a
vehicle. The demand (i.e. required capacity) for each customer is denoted by di , i ∈ C.
Likewise, we denote the service time required by each customer as ti , i ∈ C. We then
use the decision variable Xijk to denote if a particular edge (i, j) ∈ E is traversed by
vehicle k ∈ K, in other words, k travels between customers i, j ∈ C. Where this is
true we let Xijk = 1, and Xijk = 0 where it is not. We use ui , i ∈ C as a sequencing
variable that gives the position of customer i within the route of the vehicle that visits
it.
We are now able to define the problem as follows:
Minimise:
X X
cij Xijk
(3.1)
k∈K (ij)∈E
Subject to:
XX
Xijk = 1
∀i ∈ C
(3.2)
∀k ∈ K
(3.3)
∀k ∈ K
(3.4)
Xvkd j = 1
∀k ∈ K
(3.5)
k
Xjv
d = 1
∀k ∈ K
(3.6)
∀c ∈ C and ∀k ∈ K
(3.7)
ui − uj + |V |Xijk ≤ |V | − 1
∀(i, j) ∈ E − v d and ∀k ∈ K
(3.8)
Xijk
∀(i, j) ∈ E and ∀k ∈ K
(3.9)
k∈K j∈V
X
di
i∈C
X
i∈C
X
X
Xijk ≤ q
j∈C
ti
X
Xijk +
j∈C
X
cij Xijk ≤ t
(ij)∈E
j∈V
X
j∈V
X
i∈V
k
Xic
−
X
k
Xcj
=0
j∈V
∈ {0, 1}
The objective function (3.1) minimises the costs cij . Constraint (3.2) ensures that
each customer can only be serviced by a single vehicle. Constraint (3.3) enforces
the capacity constraint; each vehicle cannot exceed its maximum vehicle capacity q.
Likewise Constraint (3.4) enforces the vehicle’s work duration constraint. A vehicle’s
work duration is the sum of its service times (ti , i ∈ C where customer i is visited by
the vehicle) and its travel time. By convention the travel time is taken to be equal to
3.2. VARIANTS
37
the distance traversed by the vehicle, which in turn is equal to the costs, cij , of the
edges it traverses. Constraints (3.5) and (3.6) ensure that each vehicle starts at the
depot and finishes at the depot, and that they do this exactly once. Constraint (3.7)
and Constraint (3.8) are flow constraints that ensure that the number of vehicles
entering a customer is equal to the number of vehicles leaving, and that sub-tours are
eliminated. Lastly, Constraint (3.9) ensures the integrality conditions.
Constraint (3.4), which enforces a maximum vehicle work duration, t, is often left out
of the traditional CVRP formation but is included here as it is present in the problem
instances we use for benchmarks in Chapter 5.
3.2
Variants
In this section we provide an overview of the common variations of the VRP that are
used. These variations have arisen from real world vehicle routing scenarios, where
the constraints are often more involved than is modelled in the CVRP.
3.2.1
Multiple Depot Vehicle Routing Problem
A simple extension to the CVRP is to allow each vehicle to start from a different
depot. Part of the problem now becomes assigning customers to depots, which in
itself is a hard combinatorial problem. The CVRP formation can easily be relaxed to
allow this. There are two variations of the problem. One constrains each vehicle to
finish at the same depot that it starts from. The other allows vehicles to start and
finish at any depot, as long as the same number of vehicles return to the depot as left
from it.
3.2.2
Vehicle Routing with Time Windows
The Vehicle Routing Problem with Time Windows (VRPTW) adds the additional
constraint to the classic VRP that each customer must be visited within a time window
specified by the customer. More formally, for VRPTW each customer i ∈ C also has a
corresponding time window [ai , bi ] in which the goods must be delivered. The vehicle
is permitted to arrive before the start time, ai . However, in this case the vehicle must
wait until time ai adding to the time it takes to complete the route. However, it is
not permitted for the job to start after time bi .
An additional constraint is added to the formation of CVRP to ensure that time
window constraints are met: ai ≤ Sik ≤ bi where the decision variable Sik provides the
time that each vehicle k ∈ K arrives at customer i ∈ V .
38
CHAPTER 3. PROBLEM DEFINITION
3.2.3
Pickup and Delivery Problem
The Pickup and Delivery Problem (PDP) generalises the VRP. In this problem goods
are both picked up and delivered by the vehicle along its route. The vehicle’s work
now comes in two flavours: pickup jobs, P = {p1 , p2 , . . . , pk }, and delivery jobs,
D = {d1 , d2 , . . . , dl }, such that C = P ∪ D. Additional constraints are added to the
CVRP formation to ensure that:
1. Pickup and deliver jobs are completed by the same vehicle, that is pi ∈ Rk ⇒
di ∈ Rk where Rk represents a sequence of jobs undertaken by a vehicle k.
2. The pickup job, pi , appears before its corresponding delivery job, di , in the
sequence of jobs undertaken by a vehicle.
3. The vehicles capacity is not exceeded as goods are loaded and unloaded from
it. This requires the use of an intermediate variable, yik , i ∈ V, k ∈ K that
represents the load of vehiclePk at customer i. It adds constraints: y0k = 0,
Xijk = 1 ⇒ yjk = yik + di , and i∈V yik ≤ q for all k ∈ K, to enforce this.
There is also a variation on PDP that adds time windows, called PDPTW. In this case
the extra constraints from the VRPTW problem are merged with those given here.
PDP is a much harder problem computationally than CVRP, as its extra constraints
add new dimensions to the problem. Because of its complexity PDP has only been
actively researched in the last decade.
Chapter 4
Algorithm
This chapter provides a detailed description of the Enhanced Bees Algorithm, the
algorithm developed for this thesis, and its operation. We start by reviewing the
objectives that the algorithm was designed to meet in Section 4.1. In Section 4.2 we
provide a description of how the algorithm internally represents the VRP problem
and its candidate solutions. Next in Section 4.3 we provide a detailed description of
the operation of the algorithm. Finally, in Section 4.4 we describe the neighbourhood
structures that are used by the algorithm to define its search space.
4.1
Objectives
The Enhanced Bees Algorithm was built for use in a commercial setting. It was
developed as part of a New Zealand Trade and Enterprise grant for the company
vWorkApp Inc.’s scheduling and dispatch software. Accordingly, different objectives
were aimed for with its design (such as runtime performance) than are typically sought
in the VRP literature. The algorithm’s objectives, in order of priority, are as follows:
1. Ensure that all constraints are met. Specifically that the route’s maximum
duration is observed.
2. Have a good runtime performance. It is more desirable for the algorithm to
produce a reasonable quality result quickly (within 60 seconds), than for it to
produce a better result but require a longer processing time. Specifically if the
algorithm could reach 5% of the optimum value within 60 seconds then this
would be sufficient.
3. Produce good quality results. Notwithstanding objective 2, the results produced
must be close to the global optimum.
39
40
CHAPTER 4. ALGORITHM
4. Have a design that lends itself to parallelisation and is able to make use of the
additional processing cores available within modern hardware.
4.2
Problem Representation
The Enhanced Bees Algorithm represents the problem in a direct and straightforward
manner. It directly manipulates a candidate solution S, where S is a set of routes
R ∈ S, and each route contains an ordered sequence of customers vi ∈ R starting and
ending at the depot vertex v d .
=
R1 =
1 5 3
R2 =
2 4 7 11
R3 =
9 8 10 6
R1
3
1
5
Figure 4.1: Shown is an example of a simple VRP candidate solution as
represented internally by the Enhanced Bees Algorithm.
More general representations are sometimes used for meta-heuristics, as is commonly
seen with Genetic Algorithms, as they allow the algorithm to be easily adapted to other
combinatorial problems. However, this often comes at a cost of added complexity and
inferior results1 . This algorithm was designed specifically for solving instances of the
VRP so a direct representation was chosen.
The algorithm makes use of a fitness concept, common to many meta-heuristics, to
describe the cost of the solution. The fitness function f () includes terms for the distance (i.e. cost) of the solution and penalties for breaking the capacity and maximum
route duration constraints. The Enhanced Bees Algorithm uses penalties to encourage
feasible solutions to be produced. Rather than outright barring infeasible solutions,
the fitness function allows the algorithm some wriggle room to traverse through these
on its way towards a feasible solution.
1
This occurs because the operators that act on the problem representation can no longer exploit
information that is specific to the problem domain and must rely on general purpose operations
instead.
4.3. ENHANCED BEES ALGORITHM
41
Specifically f () is defined as follows:
X
c(R) =
ci,i+1
(4.1)
i∈R
!
d(R) = max
X
di − q, 0
(4.2)
i∈R
!
X
ti + c(R) − t, 0
(4.3)
(αc(R) + βd(R) + γc(R))
(4.4)
t(R) = max
i∈R
f (S) =
X
R∈S
Function c(R) calculates the cost (i.e distance) of a given route, and function d(R)
calculates how overcapacity the given route is. We define overcapacity to be how
much larger the sum of the route’s demands, di , i ∈ R, are than the stated maximum
allowable capacity q. Likewise function t(R) calculates the overtime of a given route.
A route’s duration is calculated as being the sum of its customer’s service times,
ti , i ∈ R, and its travel time. By convention the travel time is equal to the distance of
the route. Function t(R) then returns how much over the maximum allowable route
duration, t, the duration is. Lastly, the fitness function f () is the weighted sum of
these three terms. Parameters α, β, and γ are used used to control how much influence
each term has on determining the candidate solution’s fitness.
For the purposes of benchmarking our algorithm (see Chapter 5) we use a travel cost
that is equal to the 2D Euclidian distance2 between the two points. For real life
problems we have found that using a manhattan distance3 often provides superior
results. This is presumably due to the manhattan distance better modelling the road
system we tested on (Auckland, New Zealand), which although not a strict grid, is
still closer to this than the Euclidian distance models.
4.3
Enhanced Bees Algorithm
Our Algorithm is based on the Bees Algorithm (see Section 2.5.2 for an overview of
the standard Bees Algorithm). The Enhanced Bees Algorithm makes some changes
to adapt the Bees Algorithm to the VRP domain. An interesting aspect of the Bees
Algorithm is that it covers a broad search area, minimising the risk of being stuck in
a local optimum. It achieves this by randomly probing (or in the Bees Algorithm parlance, ‘scouting’) many areas of the search space through its entire run. However, this
approach is not well suited to hard combinatorial problems, where a newly constructed
p
Specifically, we use: cij = (xj − xi )2 + (yj − yi )2
3
Specifically, we use: cij = (xj − xi ) + (yj − yi )
2
42
CHAPTER 4. ALGORITHM
solution, let alone a randomly generated one, is often far from optimal (for instance,
the Clark Wright Savings heuristic still produces solutions that are up to 15% from
the global optimum and would require many operations to get close to optimal). We
have adapted the Bees Algorithm such that many of its unique characteristics, like
its relative robustness, are maintained while working well with hard combinatorial
problems, such as the VRP.
The Enhanced Bees Algorithm can be summarised, at a high level, as follows:
Algorithm 6: Enhanced Bees Algorithm
S = seedSites()
while termination condition not met do
for si ∈ S do
explore(si , d)
if i < λ then
removeWorstSite
end
end
end
The algorithm maintains a collection of sites S, and each site si ∈ S maintains a
collection of bees, Bi . Each bee is a proxy to the problem domain that we are trying
to solve. In our case this is the VRP problem representation covered in Section 4.2.
Initially each site is seeded, such that each site, si ∈ S, contains a collection of bees
Bi , and each bee has a corresponding VRP candidate solution, S. Each candidate
solution is initialised by seeding each route with a randomly chosen customer, which is
then filled out using the insertion heuristic outlined in Section 4.4.2. Each site is then
in turn improved upon. This is achieved by iteratively exploring the neighbourhood of
each site. The process used to explore each site is where the majority of the algorithm’s
processing takes place and where the interesting aspects of the algorithm come into
play. The exploration process is covered in detail in Sections 4.3.1, 4.3.2, and 4.4.
The number of sites explored is reduced over the run of the algorithm. This borrows
from the idea of a cooling schedule used in Simulated Annealing. Sites are reduced
using the formula:
S = S − sw
if i mod λ ≡ 0
(4.5)
Where sw represents the worst site, in terms of fitness, i represents the current iteration
of the algorithm, and λ represents the period of iterations with which the number of
sites are reduced. Once the algorithm is complete the solution, S, with the best
4.3. ENHANCED BEES ALGORITHM
43
overall fitness is returned as the answer. In the next section we review in more detail
each aspect of the algorithm.
4.3.1
Bee Movement
Bees are moved around the search space to look for improvements to the collection of
candidate solutions being maintained. Each bee represents a candidate solution, S,
so a valid bee move is any new candidate solution S0 that can be reached within the
neighbourhood of S (see Section 4.4 for the operations under which the neighbourhood
is defined).
Neighbourhood
Site1
Best site position
Last x best positions
Sitek
Sitei
Figure 4.2: Three sites are shown along with their neighbourhoods. Site
i shows in more detail that each site maintains a list of the last most
promising positions for exploration.
A feature of the Enhanced Bees Algorithm is that two Bees cannot occupy the same
position. The algorithm maintains a register of the current positions occupied by each
bee. We use the current fitness, f (S), as a quick and simple representation of a bee’s
current position4 . If a bee tries to occupy the same position as another bee (i.e. they
share the same candidate solution) then the bee trying to occupy that position is
forced to explore the neighbourhood again and find another position.
4
This obviously will not work in circumstances where there is a reasonable likelihood of two candidate solutions, Si and Sj having f (Si ) = f (Sj ). While this is not the case with the problem
instances we have used in this thesis, this may need to be modified if the algorithm is to be used on
more general problem instances.
44
CHAPTER 4. ALGORITHM
Enforcing the constraint that each bee must occupy a unique position has two benefits: it forces diversification between the bees and sites, hence encouraging a greater
proportion of the search space to be explored; and it increases the chance of a local
optimum being escaped, as a bee ensnared in the local optimum now forces the remainder of the hive to explore alternative positions. This feature has a similar intent
and effect to the tabu lists used by Tabu Search.
Another feature of the Enhanced Bees Algorithm is the role that sites play in concentrating exploration on certain areas of the search space. Each site maintains a
list, M = [S1 , . . . , S ], of the last best positions. Each Si ∈ M is then taken as a
launching point for a site’s bees to explore. θ bees are recruited for the exploration of
each Si ∈ M . Once all positions in M have been explored then the best positions
are again taken and used as the launching points for the site’s next round of exploration. This exploration method has two purposes: firstly, it allows for a simple type
of branching, as of the most promising positions that were traversed through on the
way to the current position are also explored; secondly, it prevents cycling between
promising solutions that are in close vicinity to each other.
Conversely, sites do not interact with each other, as each site maintains its own unique
list of promising positions. The constraint that no two bees can occupy the same
position ensures that each site covers a non overlapping area of the search space. In
practice we have found that this is sufficient to encourage sites to diverge and explore
distinct areas of the search space.
4.3.2
Search Space Coverage
As mentioned, one of the unique aspects of the Bees Algorithm is its ability to produce
robust results through probing a large area of the search space. However, this doesn’t
work well with hard combinatorial problems, where it cannot be ascertained quickly
if an area in the search space shows promise or not.
To overcome this limitation we instead use an approach inspired by Simulated Annealing’s use of a cooling schedule. Bees are initially divided equally between each site
si ∈ S, ensuring that each site is explored equally. Then every λ period of iterations
we reduce the number of sites maintained, such that S = S − sw , where sw is the site
with lowest fitness. We measure each site’s fitness from the fitness of its best position
found to date.
This process continues until a single site remains. We show experimentally in Chapter 5 that this process improves the robustness of the algorithm and produces better
results overall than the standard Bees Algorithm.
4.4. SEARCH NEIGHBOURHOOD
4.4
45
Search Neighbourhood
As already discussed, each bee seeks to improve upon its current fitness by exploring
the local neighbourhood of the solution it represents. The Enhanced Bees Algorithm
does this by applying a Large Neighbourhood Search (LNS) operator to its candidate
solution S. The LNS operator differs from the more common VRP operators in that
a single operation applies many changes to the candidate solution S. This widens
the neighbourhood of S to encompass exponentially many candidate solutions. LNS
navigates through the vast space it spans by selecting only those changes that have a
high likelihood of improving the solution.
The LNS operation is comprised of two-phases: a destroy phase, and a repair phase.
LNS’s destroy phase, when used for the VRP, typically involves removing a proportion
of the customers from the solution. The Enhanced Bees Algorithm’s destroy phase
uses two heuristics along those lines: a somewhat intelligent heuristic that attempts to
remove those customers that are more likely to be able to be recombined in a profitable
way; and a simple random selection. These heuristics are covered in more detail in
Section 4.4.1. The second phase of LNS is used to repair the partial solution. The
Enhanced Bees Algorithm uses a simple insertion heuristic that inserts the customers
into those locations that have the lowest insertion cost. This heuristic is covered more
formally in Section 4.4.2.
4.4.1
Destroy Heuristic
The Enhanced Bees Algorithm employs two destroy heuristics. The first simply selects
l customers randomly from a solution S and removes these from their routes. The
second is slightly more complicated and is due to Shaw [47]. Shaw’s removal heuristic
stochastically selects customers such that there is a higher likelihood of customers
that are related to one another being removed. For our purposes we define related to
mean that for any two customers, vi , vj ∈ V , then either vi , vj are geographically close
to one another (i.e. cij is small), or they share an adjacent position within the same
route, R = [. . . , vi , vj , . . .].
The rationale to removing related customers is that they are the most likely to be
profitably exchanged with one another. Conversely, unrelated customers are more
likely to be reinserted back into the same positions that they were removed from.
4.4.2
Repair Heuristic
The repair heuristic used by the Enhanced Bees Algorithm randomly selects one of
the removed customers vj , and calculates a cost for reinserting vj between each pair
46
CHAPTER 4. ALGORITHM
of jobs vi , vk ∈ Ri for all Ri ∈ S (actually not all reinsertion positions are considered,
see Section 4.4.3 for a description of which positions are considered). The reinsertion
cost is calculated as follows:
c∗ = cij + cjk − cik
∗
0
(4.6)
0
cost = c + d(R ) − d(R) + t(R ) − t(R)
(4.7)
Where c∗ calculates the cost difference in terms of travel distance. R and R0 are
defined as the route before and after the customer is inserted, respectively. And
functions d(R) and t(R) are all defined as they are in Section 4.2. The final cost is
the sum of the added travel distance and the two extra penalties, if the route is now
overcapacity or over its maximum duration. The algorithm selects the position with
the lowest insertion cost to reinsert the customer. This is repeated until all customers
are reinserted into the solution.
The reason that customers are reinserted in a random order is that it adds a beneficial
amount of noise to the heuristic. This ensures that a healthy diversity of solutions are
generated from the heuristic.
4.4.3
Neighbourhood Extent
We use two techniques to adjust the extent of the neighbourhood being searched. The
first technique that we use allows the algorithm some flexibility in selecting infeasible
solutions. As can be seen from our formulation of the candidate solutions’ fitness values (see Section 4.2) violations of the problem’s capacity and duration constraints are
penalised rather than forbidden. This allows the bees to navigate through infeasible
solutions, where other aspects of that solution are sufficiently attractive enough to
outweigh the penalties. However, only feasible solutions are allowed to be counted as
final solutions returned by the algorithm.
The second technique that we use is to adjust the number of insertion positions considered as part of the repair heuristic. The number of insertion positions considered
starts with both sides of the three closest customers and increases as the site ages.
More formally, let vi ∈ V be the customer that is being inserted. An ordered sequence
of candidate insertion points, Lvi = [v1 , . . . , vn ] such that vj ∈ V − vi , is kept that
lists customers in increasing geographic distance from vi , that is cij ∀j ∈ V − i. The
LNS repair operator tests µ positions from Lvi to find the cheapest insertion point.
The repair operator tests both possible insertion points represented by vj ∈ Lvi , that
is, it tests both the insertion cost of inserting vi immediately before and after vj in
the route R that contains vj .
4.4. SEARCH NEIGHBOURHOOD
47
For each site si ∈ S we also maintain a counter ai , i ∈ S that denotes the age of the
site. A site’s age is incremented for each iteration that a site doesn’t improve upon its
currently best known solution (as defined by the solution’s fitness, f (S)). Whenever a
site improves upon its best known solution then the counter is reset, such that ai = 0.
initial neighbourhood
final neighbourhood
Figure 4.3: The diagram shows a vertex being considered for reinsertion
back into a solution. It starts by only considering insertion points that are
geographically close to its position. The search area is widened over the
course of the algorithm to take into consideration a larger set of insertion
points.
We then use the following formula to increase how much of Lvi is considered as the
site ages.
µ = |Lvi | min
a
i
k
,1
Where k is a constant that controls the rate at which the search area is expanded.
As this process extends the number of insertion positions that are considered by the
repair heuristic, this also serves to extend the neighbourhood of solutions surrounding
a candidate solution S. In this way the algorithm also dynamically extends the size
of a neighbourhood surrounding a site, si ∈ S, if si becomes stuck in a local optimum.
48
CHAPTER 4. ALGORITHM
Chapter 5
Results
In this chapter we provide a detailed breakdown of the results obtained by the Enhanced Bees Algorithm. The algorithm is tested against the well known set of test
instances from Christofides, Mingozzi and Toth [11]. We start in Section 5.1 by presenting the results obtained by running the algorithm in its two standard configurations: the first configuration is optimised to produce the best overall results, regardless
of the runtime performance; the second configuration is optimised to produce the best
results possible within a 60 second runtime threshold. We follow this in Section 5.2
by providing contrasting results on the same problem instances instead using a standard Bees Algorithm and a LNS local search. Here we aim to demonstrate that the
enhancements suggested in this thesis do in fact improve the solution quality. Finally,
we end in Section 5.3 by comparing and ranking how the Enhanced Bees Algorithm
performs compared to the results in the literature.
5.1
Enhanced Bees Algorithm
The results depicted in Figures 5.1, 5.2, 5.3 and 5.4 show the algorithm’s performance
on two configurations. The first configuration is optimised to produce the best overall
results, that is, the minimum travel distance that meets all capacity and duration
constraints. No consideration is made for the algorithm’s runtime performance in this
configuration. This configuration is denoted as Best in the following diagrams and
tables. The second configuration is optimised to produce the best results possible
within a 60 second runtime window. This configuration is denoted as Fast in the
following diagrams and tables.
The results depicted in Figures 5.1, 5.2, 5.3 and 5.4 are for the standard 14 VRP
problem instances due to Christofides, Mingozzi and Toth [11] and were all obtained
49
50
CHAPTER 5. RESULTS
1
Percentage of Best Known
0.8
0.6
0.4
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.2
0
0
200
400
600
800
Time (secs)
1000
1200
1400
1600
Figure 5.1: Best. Shown are the results obtained on the standard 14
Christofides, Mingozzi and Toth problem instances when the algorithm is
optimised for producing the best overall results. The algorithm was left to
run for 30 minutes. The lefthand axis shows the relative percentage compared to the best known result for the same problem instance. The bottom
axis shows the elapsed runtime in seconds. Infeasible solutions are shown
as being 0% of the best known result. The average result obtained across all
problem instances is shown in red.
on a MacBook Pro 2.8 GHz Intel Core 2 Duo. The best result for each problem
instance was selected from 10 runs of the algorithm.
In Figures 5.1 and 5.2 (i.e. the Best configuration) the algorithm was set to start with
|S| = 100 (i.e. 100 sites), and reduce this number each 50 iterations (λ = 50) by 1%
until only |S| = 3. The number of promising solutions remembered by each site was
set to |M | = 5. The algorithm was left to run for 30 minutes on each problem before
being terminated. Infeasible solutions (i.e. solutions over their capacity or duration
constraints) were allowed to be traversed through, but were scored as being 0% of the
best known solution.
Conversely, in Figures 5.3 and 5.4 (i.e. the Fast configuration) the algorithm was set to
start with |S| = 25 (i.e. 25 sites), and reduce this number each iteration (λ = 1) by 1%
until only |S| = 1. The number of promising solutions remembered by each site was
set to |M | = 5. The algorithm was left to run for 60 seconds on each problem before
5.2. EXPERIMENTS
51
1
0.98
0.96
Percentage of Best Known
0.94
0.92
0.9
0.88
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.86
0.84
0.82
0.8
0
200
400
600
800
Time (secs)
1000
1200
1400
1600
Figure 5.2: Best—Blow up. Shown are the same results as are depicted
in Figure 5.1, but with the area between 0.8 and 1.0 of the left axis blown
up.
being terminated. The LNS improvement heuristic was set to destroy between 0% and
80% (with a mean of 40%) of the solution at each step. The repair operator initially
only considers the first 3 closest customers as reinsertion points, but it increases this
to 50% of the closest customers as the site ages. Infeasible solutions (i.e. solutions
over duration or service time constraints) were allowed to be traversed through, but
were scored as being 0% of the best known solution.
Table 5.1 provides a tabular summary of the results covered in this section.
5.2
Experiments
In this section we review the results obtained by implementing the standard Bees
Algorithm and a LNS local search. The aim of these experiments is to prove that the
algorithmic enhancements suggested in this thesis do in fact produce better results
than would have been obtained if we had used a standard Bees Algorithm. We also
demonstrate that the combination of the Bees Algorithm with the LNS local search
produces better results than if either algorithm were used separately.
52
CHAPTER 5. RESULTS
1
Percentage of Best Known
0.8
0.6
0.4
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.2
0
0
10
20
30
Time (secs)
40
50
60
Figure 5.3: Fast. Shown are the results obtained on the standard 14
Christofides, Mingozzi and Toth problem instances when the algorithm is
optimised for producing the best results within a 60 second runtime limit.
The lefthand axis shows the relative percentage compared to the best known
result for the same problem instance. The bottom axis shows the elapsed
runtime in seconds. Infeasible solutions are shown as being 0% of the best
known result.
5.2.1
Bees Algorithm versus Enhanced Bees Algorithm
We start in Figures 5.5 and 5.6 by showing the results obtained by using the standard
Bees Algorithm as described by Pham et al. in [38]. The same problem instances as
Section 5.1 are used, so that the results can be compared directly.
The results depicted in Figures 5.5 and 5.6 were obtained on a MacBook Pro 2.8
GHz Intel Core 2 Duo. The best result for each instance was selected from 10 runs
of the algorithm. The algorithm was configured with the following parameters. It
used 25 sites, and selected the best 6 sites as being elite. Each elite site had 3 bees
recruited for the search. Another 6 sites were selected as being non-elite, and had
2 bees recruited for the search. The bees from the remaining 13 sites were left to
search randomly. The algorithm was left to run for 60 seconds on each problem before
being terminated. Infeasible solutions (i.e. solutions over their capacity or duration
constraints) were allowed to be traversed through, but were scored as being 0% of the
5.2. EXPERIMENTS
53
1
0.98
0.96
Percentage of Best Known
0.94
0.92
0.9
0.88
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.86
0.84
0.82
0.8
0
10
20
30
Time (secs)
40
50
60
Figure 5.4: Fast—Blow up. Shown are the same results as are depicted in
Figure 5.3, but with the area between 0.8 and 1.0 of the left axis blown up.
best known solution. A λ-interchange (with λ = 2) improvement heuristic was used
for the improvement phase of each bee (see Chapter 2 for an overview on how this
heuristic works).
5.2.2
Large Neighbourhood Search
Next we show in Figures 5.7 and 5.8 the results obtained by using a standalone LNS
search embedded within a hill climb meta-heuristic. The LNS search used in this
section is the same one that is employed by the Enhanced Bees Algorithm. It should
be noted that there are more sophisticated LNS algorithms available than the comparatively simple one used here. And that these would most probably produce better
results than the LNS results presented here. However, we believe one of the attractive
features of the Enhanced Bees Algorithm is that it uses a fairly simple local search
procedure. Moreover, our aim in this experiment was to demonstrate that the limitations of our simple local method are offset by it being embedded within a Bees
Algorithm.
The results depicted in Figures 5.7 and 5.8 were obtained on a MacBook Pro 2.8 GHz
Intel Core 2 Duo. The best result for each problem instance was selected from 10 runs
54
CHAPTER 5. RESULTS
Table 5.1: Enhanced Bees Algorithm. Given are the results obtained by running the Enhanced Bees Algorithm using two different
configurations on the standard 14 Christofides, Mingozzi and Toth
problem problem instances. The fast configuration is optimised to
produce results within a 60 second runtime limit. Conversely, the
best configuration is optimised for producing the best results within
the much longer period of 30 minutes.
Instance
Fast1
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
1
2
3
Best Known3
Results
524.61
835.77
826.14
1057.40
1360.85
555.43
913.37
865.94
1196.32
1490.49
1080.20
819.56
1555.30
866.37
100.00%
99.94%
100.00%
97.26%
94.90%
100.00%
99.60%
100.00%
97.18%
93.65%
96.47%
100.00%
99.09%
100.00%
Best2
524.61
835.26
826.14
1036.12
1327.48
555.43
909.68
865.94
1169.24
1428.54
1048.24
819.56
1545.19
866.37
98.44%
100.00%
100.00%
100.00%
99.26%
97.29%
100.00%
100.00%
100.00%
99.43%
97.71%
99.42%
100.00%
99.74%
100.00%
524.61
835.26
826.14
1028.42
1291.45
555.43
909.68
865.94
1162.55
1395.85
1042.11
819.56
1541.14
866.37
99.49%
Run for 60 seconds.
Run for 30 minutes.
As reported by Gendreau, Laporte, and Potvin in [21].
of the algorithm. The algorithm was initialised to a starting position generated by
a simple insertion heuristic. The LNS heuristic was set to destroy between 0% and
80% (with a mean of 40%) of the solution at each step. The repair enumerated all
customers when deciding the best reinsertion point. Infeasible solutions (i.e. solutions
over their capacity or duration constraints) were allowed to be traversed through, but
were scored as being 0% of the best known solution.
5.2.3
Summary
Table 5.2 provides a summary of the results obtained by the Bees Algorithm and LNS
experiments, alongside the results obtained by our Enhanced Bees Algorithm. The
Bees Algorithm is the worst of the three. This is not surprising given that the Bees
Algorithm was devised to solve continuous problems, rather than discrete problems.
5.2. EXPERIMENTS
55
1
Percentage of Best Known
0.8
0.6
0.4
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.2
0
0
10
20
30
Time (secs)
40
50
60
Figure 5.5: Standard Bees Algorithm. Shown are the results obtained on
the standard 14 Christofides, Mingozzi and Toth instances when the algorithm is optimised for producing the best results within a 60 second runtime
limit. The lefthand axis shows the relative percentage compared to the best
known result for the same problem instance. The bottom axis shows the
elapsed runtime in seconds. Infeasible solutions are shown as being 0% of
the best known result. The average result obtained across all problem instances is shown in red.
Where the Bees Algorithm has been used for discrete problems in the literature it has
been adapted to incorporate more sophisticated local search techniques, much as the
Enhanced Bees Algorithm has been here. Also of note is that two of the problem instances didn’t produce feasible solutions at all within the 60 second runtime threshold.
We believe that given a longer running time the algorithm would most probably have
found a feasible solution. However, one of the objectives of the Enhanced Bees Algorithm is to produce robust results reliably; in this count the standard Bees Algorithm
is not competitive.
The LNS improvement heuristic produced much stronger results. This shows that the
LNS improvement plays an important part in the results obtained by the Enhanced
Bees Algorithm. The LNS heuristic is a fairly new heuristic (in the VRP research
at least); nevertheless it has produced some of the most competitive results. This is
borne out by the results obtained here and in Section 5.1. However, the LNS local
search did fail to find a feasible solution for one of the problem instances. Again this
56
CHAPTER 5. RESULTS
1
0.98
0.96
Percentage of Best Known
0.94
0.92
0.9
0.88
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.86
0.84
0.82
0.8
0
10
20
30
Time (secs)
40
50
60
Figure 5.6: Standard Bees Algorithm—Blow up. Shown are the same
results as are depicted in Figure 5.5, but with the section between 0.8 and
1.0 of the left axis blown up. Note that the average is not shown on the blow
up because it is below 80% percent.
is probably due to the limited runtime permitted. If this problem instance is removed
from the results, then the LNS search’s average result becomes 96.39%, getting us
much closer to the results obtained by the Enhanced Bees Algorithm.
5.3
Comparison
Lastly, in Table 5.3 and Figure 5.9 we provide a comparison of the Enhanced Bees
Algorithm along with other well known results from the literature. As can be seen
from Table 5.3 and Figure 5.9 some of the best results known are due to Taillard’s
Tabu Search heuristic. He reaches 12 of the best known solutions from the set of
14 problems. The Enhanced Bees Algorithm, by comparison, finds 8 of the 14 best
known solutions. However, the Enhanced Bees Algorithm is still very competitive.
The runtime duration required to find a best known solution is smaller than many
of the other meta-heuristics (although a direct comparison is hard to make as many
of the reported results were run on significantly older hardware). Additionally, the
solutions produced by the Enhanced Bees Algorithm are within 0.5% of the best known
5.3. COMPARISON
57
1
Percentage of Best Known
0.8
0.6
0.4
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.2
0
0
10
20
30
Time (secs)
40
50
60
Figure 5.7: LNS algorithm. Shown are the results obtained on the standard 14 Christofides, Mingozzi and Toth instances when the algorithm is
optimised for producing the best results within a 60 second runtime limit.
The lefthand axis shows the relative percentage compared to the best known
result for the same problem instance. The bottom axis shows the elapsed
runtime in seconds. Infeasible solutions are shown as being 0% of the best
known result. The average result obtained across all problem instances is
shown in red.
solutions on average, meaning that the algorithm is very competitive with the best
meta-heuristics available for the VRP.
58
CHAPTER 5. RESULTS
1
0.98
0.96
Percentage of Best Known
0.94
0.92
0.9
0.88
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Average
0.86
0.84
0.82
0.8
0
10
20
30
Time (secs)
40
50
Figure 5.8: LNS algorithm—Blow up. Shown are the same results as are
depicted in Figure 5.7, but with the area between 0.8 and 1.0 of the left axis
blown up.
60
5.3. COMPARISON
59
Table 5.2: Experiments. Given are results that contrast the standard
Bees Algorithm, a LNS local search, and the Enhanced Bees Algorithm
with one another. Each algorithm was run for 60 seconds and used the
standard 14 Christofides, Mingozzi and Toth problem instances.
Instance
Bees Algorithm1
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
557.17
925.97
873.75
1203.47
1520.89
565.21
1153.57
941.34
1356.43
2153.48
1091.42
825.38
1662.70
903.25
Average
1
2
3
94.16%
90.20%
94.55%
85.45%
84.91%
98.27%
00.00%
91.99%
85.71%
00.00%
95.48%
99.30%
92.69%
95.92%
79.18%
LNS2
543.16
854.35
866.13
1061.75
1360.85
560.24
1215.12
890.93
1196.31
1515.32
1168.91
826.14
1555.29
879.69
96.58%
97.77%
95.38%
96.86%
94.90%
99.14%
00.00%
97.19%
97.18%
92.12%
89.15%
99.20%
99.09%
98.49%
89.50%
Run for 60 seconds and configured as described in Section 5.2.1.
Run for 60 seconds and configured as described in Section 5.2.2.
Run for 60 seconds and configured as described in Section 5.1.
Enhanced Bees3
524.61
835.77
826.14
1057.40
1360.85
555.43
913.37
865.94
1196.32
1490.49
1080.20
819.56
1555.30
866.37
100.00%
99.94%
100.00%
97.26%
94.90%
100.00%
99.60%
100.00%
97.18%
93.65%
96.47%
100.00%
99.09%
100.00%
98.44%
CHAPTER 5. RESULTS
60
1
2
3
4
5
6
7
8
9
584.64
900.26
886.83
1133.43
1395.74
618.40
975.46
973.94
1287.64
1538.66
1071.07
833.51
1596.72
875.75
CW1
532
874
851
1079
1389
560
933
888
1230
1518
1266
937
1776
949
Sweep2
524
857
833
1014
1420
560
916
885
1230
1518
824
876
GenAsgn3
578.56
888.04
878.70
1128.24
1386.84
616.66
974.79
968.73
1284.64
1538.66
1049.43
824.42
1587.93
868.50
3-Opt4
528 (167)
838.62 (6434)
829.18 (9334)
1058 (5012)
1378 (1291)
555.43 (3410)
909.68 (626)
866.75 (957)
1164.12 (84301)
1417.85 (5708)
1176 (315)
826 (632)
1545.98 (7622)
890 (305)
SA5
524.61 (360)
835.26 (3228)
826.14 (1104)
1028.42 (3528)
1298.79 (5454)
555.43 (810)
909.68 (3276)
865.94 (1536)
1162.55 (4260)
1397.94 (5988)
1042.11 (1332)
819.56 (960)
1541.14 (3552)
866.37 (3942)
TS6
524.61 (6)
844.31 (78)
832.32 (228)
1061.55 (1104)
1343.46 (5256)
560.24 (6)
916.21 (102)
866.74 (288)
1195.99 (1650)
1451.65 (4908)
1065.21 (552)
819.56 (300)
1559.92 (660)
867.07 (348)
ACO7
524.61 (5)
835.26 (225)
826.14 (50)
1036.12 (1215)
1327.48 (785)
555.43 (5)
909.68 (175)
865.94 (20)
1169.24 (1610)
1428.54 (1540)
1048.24 (960)
819.56 (35)
1545.19 (1500)
866.37 (40)
EBA8
524.61
835.26
826.14
1028.42
1291.45
555.43
909.68
865.94
1162.55
1395.85
1042.11
819.56
1541.14
866.37
Best Known9
Table 5.3: Results Comparison. Given is a comparison of the Enhanced Bees Algorithm alongside other well known results from the
literature. The results are for the standard 14 Christofides, Mingozzi and Toth problem instances commonly used in the VRP literature.
Running times are given in parenthesis where known.
Instance
P01E51K05
P02E76K10
P03E101K08
P04E151K12
P05E200K17
P06D51K06
P07D76K11
P08D101K09
P09D151K14
P10D200K18
P11E121K07
P12E101K10
P13D121K11
P14D101K11
Clark Write’s Savings (Parallel) algorithm. Implemented by Laporte and Semet [30].
Sweep Algorithm due to Gillett and Miller [22]. Implemented by Christofides, Mingozzi and Toth [11]. Reported in [30].
Generalised Assignment due to Fisher and Jaikumar [17]. Reported in [30].
3-Opt local search applied after Clark Write’s Savings (Parallel) algorithm. First improvement taken. Implemented by Laporte and Semet [30].
Simulated Annealing due to Osman [36]. Runtime duration is given in parentheses and is reported in seconds on a VAX 86000.
Tabu Search due to Taillard [49]. Runtime duration is given in parentheses and is reported in seconds on a Sillicon Graphics Workstation, 36Mhz.
Ant Colony Optimisation Bullnheimer, Hartl, Strauss [9]. Bullnheimer et al. provided two papers on Ant Colony Optimisation for VRP, the better of the two is
used. Runtime duration is given in parentheses and is reported in seconds on a Pentium 100.
Enhanced Bees Algorithm. Results are shown from the best configuration in Section 5.1. Runtime duration is given in parentheses and is reported in seconds on
a MacBook Pro 2.8 GHz Intel Core 2 Duo.
Best known results as reported by Gendreau, Laporte, and Potvin in [21].
5.3. COMPARISON
61
1
TS
0 . 99
ACO
SA
0 . 98
Gen Assign
Percent of best known solution
0 . 97
0 . 96
0 . 95
0 . 94
3-Opt
Sweep
CW
0 . 93
0 . 92
0 . 91
EBA Average
0.9
0
2 00
4 00
6 00
8 00
1 000
1 200
1 400
1 600
Time (seconds)
Figure 5.9: Results Comparison. Shown are the average results across all
problem instances obtained by the Enhanced Bees Algorithm (shown in red)
mapped against the results listed in Table 5.3.
62
CHAPTER 5. RESULTS
Chapter 6
Conclusion
In this thesis we have described a new meta-heuristic for the VRP called the Enhanced
Bees Algorithm. The results obtained are competitive with the best meta-heuristics
available for the Vehicle Routing Problem. Additionally, the algorithm has good
runtime performance, producing results within 2% of the optimal solution within 60
seconds.
We took the Bees Algorithm as our starting point. The Bees Algorithm was originally developed for solving continuous optimisation problems, so part of the work
undertaken for this thesis was to adapt it for use on the VRP (although, it could be
argued that the Enhanced Bees Algorithm is more ‘inspired by’ than ‘adapted from’).
The approach developed for the Enhanced Bees Algorithm could equally be applied to
other combinatorial optimisation problems, such as the Traveling Salesman Problem,
the Job Shop Scheduling Problem, or the Cutting Stock Problem.
We showed empirically that the quality of the solutions obtained by the Enhanced
Bees Algorithm are competitive with the best modern meta-heuristics available for
the VRP. Additionally, we showed that the algorithm has good runtime performance,
producing results within 2% of the optimal solution within 60 seconds. The runtime
performance of the algorithm makes it suitable for use within real world dispatch
scenarios, where often the dispatch process is fluid and hence it is impractical for the
optimisation to take minutes (or hours). In these environments, it is acceptable to
trade a fraction of a percent off the solution quality for quicker runtime performance.
We also gave results that demonstrated that the algorithmic enhancements suggested
in this thesis did in fact produce better results than would have been obtained by
using a standard Bees Algorithm. We also demonstrated that the combination of the
Bees Algorithm with a LNS local search heuristic produced better results than if the
algorithms had been used separately.
63
64
CHAPTER 6. CONCLUSION
Additionally, we provided a comprehensive survey of the VRP literature. In this we
provided a short history of the results that have been foundational to VRP research, as
well as providing in depth material and descriptions for some of the classic algorithms
developed for the VRP.
There are a number of areas where the research undertaken in this thesis could be
continued. These include:
• Introduce a mating process to generate new sites. An interesting extension to the
Enhanced Bees Algorithm would be to incorporate a crossover operation for generating new sites. In our version of the Enhanced Bees Algorithm unpromising
sites are simply culled off. An alternative approach, borrowing a concept from
Genetic Algorithms and Tabu Search (in this case, Taillard’s Adaptive Memory), would be to replace the site with a recombination of two other successful
sites. The advantage to this approach is that it would open a new area of the
search space to exploration; an area that should be promising as it combines
components from two already successful sites. The crossover process could be
as simple as using an existing VRP crossover operator, such as OX, on the two
fittest solutions from each site. Although much more sophisticated crossover
operations are imaginable.
• Extend to other combinatorial problems. It should be possible to follow a similar
approach that we have taken in adapting the Bees Algorithm to the VRP and
apply this to other combinatorial problems. The Bees Algorithm is especially
strong at providing robust solutions where the search space contains many local
optima. We imagine that there are many other combinatorial problems where
this would be an advantage. An obvious starting point would be to apply it to
the Job Shop Scheduling Problem, which shares many characteristics with the
VRP.
• Include more real world constraints. Although the literature is filled with variations of the VRP that add additional constraints (i.e. VRPTW, PDP, etc), the
focus has been on tackling the computationally hard constraints, such as time
windows, multiple deploys, etc. An area that is not often addressed in the literature is how to deal with soft constraints, such as, on the day disruptions, differing
skill sets across a fleet, or factoring in a dispatcher’s assignment preferences. Unfortunately, these constraints are a barrier to vehicle route optimisation being
adopted in many logistics companies.
An interesting line of research would be to extend the optimisation methods
developed for the VRP to include feedback from a dispatcher—we envision an
interactive process where the dispatcher can feedback into the optimisation process the soft constraints that are not modelled within the algorithm. It would
also be very interesting to see if the supervised learning methods from artificial
intelligence, such as the Naive Bayes classifier, could be incorporated into the
optimisation process.
Bibliography
[1] Bee image. http://etc.usf.edu/clipart/27800/27890/bee 27890.htm.
[2] The bees algorithm website. http://www.bees-algorithm.com.
[3] Biodidac,
a
bank
of
digital
http://biodidac.bio.uottawa.ca/.
resources
for
teaching
biology.
[4] K. Altinkemer and B. Gavish. Parallel savings based heuristic for the delivery
problem. Operations Research, 39:456–469, 1991.
[5] M.L. Balinski and R.E. Quandt. On an integer program for a delivery problem.
Operations Research, 12(2):300–304, 1964.
[6] Jean Berger and Mohamed Barkaoui. A hybrid genetic algorithm for the capacitated vehicle routing problem. In Genetic and Evolutionary Computation—
GECCO 2003, volume 2723 of Lecture Notes in Computer Science, pages 198–198.
2003.
[7] J.L. Blanton and R.L. Wainwright. Multiple vehicle routing with time and capacity constraints using genetic algorithms. Proceedings of the 5th International
Conference on Genetic Algorithms, pages 451–459, 1993.
[8] B. Bullnheimer, R. F. Hartl, and C. Strauss. Applying the ant system to the
vehicle routing problem. Meta-heuristics: Advances and trends in local search
paradigms for optimization, pages 285–296, 1999.
[9] B. Bullnheimer, R. F. Hartl, and C. Strauss. An improved ant system algorithm
for the vehicle routing problem. Annals of Operations Research, 89:319328, 1999.
[10] N. Christofides and S. Eilon. An algorithm for the vehicle dispatching problem.
ORQ, 20:309–318, 1969.
[11] N. Christofides, A. Mingozzi, and P. Toth. Exact algorithms for the vehicle routing problem, based on spanning tree and shortest path relaxations. Mathematical
Programming, 20:255–282, 1981a.
[12] G. Clark and J.W. Wright. Scheduling of vehicles from a central depot to a
number of delivery points. Operations Research, 12:568–581, 1964.
65
66
BIBLIOGRAPHY
[13] Dantzig, Fulkerson, and Johnson. Solution of a large-scale traveling salesman
problem. Operations Research, (2):393410, 1954.
[14] G. B. Dantzig and J. H. Ramser. The truck dispatching problem. Management
Science, 6(1):80–91, Oct, 1959.
[15] Martin Desrochers, Jacques Desrosiers, and Marius Solomon. A new optimization algorithm for the vehicle routing problem with time windows. Operations
Research, 40:342–354, March 1992.
[16] S. Eilon, C.D. Watson-Gandy, and N. Christofides. Distribution management:
Mathematical modelling and practical analysis. Griffin (London), 1971.
[17] M.L. Fisher and R. Jaikumar. A generalized assignment heuristic for solving the
vrp. Networks, 11:109–124, 1981.
[18] B. A. Foster and D.M. Ryan. An integer programming approach to the vehicle
scheduling problem. Operations Research, 27:367384, 1976.
[19] T. Gaskell. Bases for vehicle fleet scheduling. Operational Research Quarterly,
18:281–295, 1967.
[20] M. Gendreau, A. Hertz, and G. Laporte. A tabu search heuristic for the vehicle
routing problem. Management Science, 40:1276–1290, 1994.
[21] M Gendreau, G Laporte, and J Potvin. Metaheuristics for the vehicle routing
problem. Technical report, Les Cahiers du Gerad, 1998, revised 1999.
[22] B.E. Gillett and L.R. Miller. A heuristic algorithm for the vehicle dispatch problem. Operations Research, 22:340–349, 1974.
[23] M Haimovich and A H G R Kan. Bounds and heuristics for capacitated routing
problems. Mathematics of Operations Research, 10:527–542, 1985.
[24] W.R. Hamilton. Memorandum respecting a new system of roots of unity (the
icosian calculus). Philosophical Magazine, 12, 1856.
[25] J. H. Holland. Adaptation in natural and artificial systems. The University of
Michigan Press, 1975.
[26] R. Karp. Reducibility among combinatorial problems. In R. Miller and
J. Thatcher, editors, Complexity of Computer Computations, pages 85–103.
Plenum Press, 1972.
[27] T.P. Kirkman. On the representation of polyhedra. Philosophical Transactions
of the Royal Society of London Series A, 146:413–418, 1856.
[28] G. Laporte, H. Mercure, and Y. Nobert. An exact algorithm for the asymmetrical
capacitated vehicle routing problem. Networks, 16:33–46, 1986.
[29] G. Laporte and Y. Nobert. Surveys in Combinatorial Optimization, chapter Exact
algorithms for the vehicle routing problem. 1987.
BIBLIOGRAPHY
67
[30] G. Laporte and F. Semet. Classical heuristics for the vehicle routing problem.
Technical report, Les Cahiers du Gerad, 1999.
[31] Chee Peng Lim, Lakhmi C. Jain, and Satchidananda Dehuri, editors. Innovations in Swarm Intelligence, volume 248 of Studies in Computational Intelligence.
Springer, 2009.
[32] Y. Nagata. Edge assembly crossover for the capacitated vehicle routing problem.
EvoCOP, LNCS, 4446:142–153, 2007.
[33] Road Transport Forum NZ. Road transport forum nz—transport facts. Website.
http://www.rtfnz.co.nz/cm-transport-facts.php.
[34] I.M Oliver, D.J. Smith, and J.R.C Holland. A study of permutation crossover
operators on the traveling salesman problem. Grefenstette—Proceedings of the
Second International Conference on Genetic Algorithms and Their Applications,
pages 224–230, 1987.
[35] I Or. Traveling salesman-type combinatorial problems and their relation to the
logistics of regional blood banking. PhD thesis, Evanston, IL: Northwestern University, 1976.
[36] I. H. Osman. Metastrategy simulated annealing and tabu search algorithm for
the vehicle routing problem. Annals of Operations Research, 41:421–451, 1993.
[37] H. Paessens. The savings algorithm for the vehicle routing problem. European
Journal of Operational Research, 34:336–344, 1988.
[38] Pham, Ghanbarzadeh, Koc, S. Otri, S. Rahim, and M. Zaidi. The bees algorithm.
Technical report, 2005.
[39] Jean-Yves Potvin. A review of bio-inspired algorithms for vehicle routing. In
Bio-inspired Algorithms for the Vehicle Routing Problem, pages 1–34. 2009.
[40] M. R. Rao and S. Zionts. Allocation of transportation units to alternative tripsa
column generation scheme with out-of-kilter subproblems. Operations Research,
16:52–63, 1968.
[41] M. Reimann, M. Stummer, and K. Doerner. A savings-based ant system for the
vehicle routing problem. Proceedings of the Genetic and Evolutionary Computation Conference, pages 1317–1325, 2002.
[42] J Renaud, F F Boctor, and G Laporte. A fast composite heuristic for the symmetric traveling salesman problem. INFORMS Journal on Computing, 8:134–143,
1996.
[43] J. Robinson. On the hamiltonian game (a traveling salesman problem). Research
Memorandum RM-303, 1949.
[44] F. Robuste, C.F. Daganzo, and R. Souleyrette II. Implementing vehicle routing
models. Transportation Research, 24B:263–286, 1990.
68
BIBLIOGRAPHY
[45] Stefan Ropke. Heuristic and exact algorithms for vehicle routing problems. PhD
thesis, Department of Computer Science, University of Copenhagen (DIKU),
2005.
[46] Alexander Schrijver. On the history of combinatorial optimization (till 1960). In
Operations Research and Management. Elsevier, 2005.
[47] P. Shaw. Using constraint programming and local search methods to solve vehicle
routing problems. pages 417–431. 1998.
[48] M. M. Solomon. Algorithms for the vehicle routing and scheduling problems with
time window constraints. Oper. Res., 35(2):254–265, 1987.
[49] E. D. Taillard. Parallel iterative search methods for vehicle routing problems.
Networks, 23:661–673, 1993.
[50] S.R. Thangiah, K.E. Nygard, and P.L. Juell. Gideon: A genetic algorithm system
for vehicle routing with time windows. Proceedings of 7th IEEE Conference on
Artificial Intelligence Applications, pages 322–328, 1991.
[51] P. Toth and D. Vigo. The granular tabu search (and its application to the vehicle
routing problem). Technical report, 1998.
[52] Paolo Toth and Daniele Vigo, editors. The vehicle routing problem. Society for
Industrial and Applied Mathematics, Philadelphia, PA, USA, 2001.
[53] J. A. G. Willard. Vehicle routing using r-optimal tabu search. Master’s thesis,
The Management School, Imperial College, London, 1989.
[54] A. Wren and A. Holliday. Computer scheduling of vehicles from one or more
depots to a number of delivery points. Operations Research Quarterly, pages
333–344, 1972.
[55] P. Yellow. A computational modification to the savings method of vehicle scheduling. Operational Research Quarterly, 21:281–283, 1970.
| 9cs.NE
|
An Extreme Learning Machine Approach to Predicting
Near Chaotic HCCI Combustion Phasing in Real-TimeI
Adam Vaughana,*, Stanislav V. Bohaca
a Dept.
of Mechanical Engineering, University of Michigan, Ann Arbor, MI 48109, USA
arXiv:1310.3567v3 [cs.LG] 5 May 2015
Abstract
Fuel efficient Homogeneous Charge Compression Ignition (HCCI) engine combustion timing predictions must contend
with non-linear chemistry, non-linear physics, period doubling bifurcation(s), turbulent mixing, model parameters that
can drift day-to-day, and air-fuel mixture state information that cannot typically be resolved on a cycle-to-cycle basis,
especially during transients. In previous work, an abstract cycle-to-cycle mapping function coupled with 𝜖-Support
Vector Regression was shown to predict experimentally observed cycle-to-cycle combustion timing over a wide range of
engine conditions, despite some of the aforementioned difficulties. The main limitation of the previous approach was
that a partially acausual randomly sampled training dataset was used to train proof of concept offline predictions. The
objective of this paper is to address this limitation by proposing a new online adaptive Extreme Learning Machine (ELM)
extension named Weighted Ring-ELM. This extension enables fully causal combustion timing predictions at randomly
chosen engine set points, and is shown to achieve results that are as good as or better than the previous offline method.
The broader objective of this approach is to enable a new class of real-time model predictive control strategies for high
variability HCCI and, ultimately, to bring HCCI’s low engine-out NOx and reduced CO2 emissions to production engines.
Keywords: non-linear, non-stationary, time series, chaos theory, dynamical system, adaptive extreme learning machine
1. Introduction
Since the 1800s, gasoline engines have largely been operated by (1) controlling power output with a throttle that
restricts airflow, (2) using a simple spark to control burn
timing, and (3) operating close to fuel-air stoichiometry
for reliable spark ignition and so catalysts can reduce NOx ,
HC, and CO emissions. The throttle hurts fuel efficiency
with pumping losses (especially at low-load), and the stoichiometric mixtures used are thermodynamically less fuel
efficient than mixtures diluted with air or exhaust gases.
With the broad availability of enabling technologies
(e.g. variable valve timing), a relatively new type of combustion called Homogeneous Charge Compression Ignition
(HCCI) has received increased research interest over the
past decade. HCCI uses autoignition to burn lean (excess
air) mixtures and can produce ultra-low NOx quantities
that do not require expensive catalyst aftertreatment. Instead of a spark, combustion timing is controlled by the
thermodynamic trajectory of the mixture and complex
chemical kinetics. With both ultra-low NOx production
and freedom from the stoichiometric shackles of spark igI There
are no fundamental changes in this updated version of the
original October, 2013 pre-print paper [1]. This version includes algebraic simplifications, minor corrections, and improved body text.
* Corresponding author
Email addresses: [email protected] (Adam Vaughan),
[email protected] (Stanislav V. Bohac)
Preprint submitted to Elsevier
nition, HCCI achieves greater fuel efficiency through thermodynamically ideal lean mixtures and unthrottled operation. This improved fuel economy, has real-world relevance
to near-term sustainability, national oil independence, and
greenhouse gas initiatives that seek to curb petroleum usage.
The primary challenge of HCCI autoignition is to ensure that the burn timing is synchronized against the motion of the piston. This is important for efficient extraction
of mechanical work from the fuel-air mixture and to avoid
unsafe, noisy combustion or unstable (near chaotic) combustion oscillations. This synchronization is so important
that combustion researchers do not use normal units of
time. Instead, they use the angle of the crank, which (1)
describes the position of the piston, and (2) represents time
because each crank angle takes a certain amount of time at
a fixed crank rotation speed. The thermodynamic trajectory of the air-fuel mixture is driven by the piston varying
the cylinder volume as function of crank angle (Fig. 1).
These angles are measured relative to when the piston is
at the top of the cylinder, or Top Dead Center (TDC). In a
four-stroke engine, TDC occurs twice per cycle. In different regions, the piston may be compressing or expanding
the mixture, or, if a valve is open, moving the mixture into
or out of the intake or exhaust manifolds.
Highlighted on the cylinder volume curve are two regions, one for when the exhaust valve is open and the
other for when the intake valve is open. Note that the two
May 7, 2015
Cylinder Pressure (P)
Cylinder Volume
to as high Cyclic Variability (CV) and it severely constrains the available load limits of HCCI.
CA50n+1
function( CA90n , …) = CA50n+1
CA90n
Mean pressures
from air and
residual gases
1.1. Motivation and goals
A primary constraint for HCCI is the need to keep combustion timing between the ringing and combustion stability limits [12]. At the ringing limit, excessive pressure rise
rates are encountered, and at the stability limit, high CV
in combustion timing is observed [12]. Since these limits
play a key role in constraining HCCI’s usable operating
range, it is desirable to explore new methods to predict the
behavior at and beyond these constraints. In particular,
the ability to predict and correct for high CV might enable
the use of late phased combustion to mitigate the excessive
pressure rise rates that currently constrain HCCI’s highload operation [13], while also potentially addressing the
high CV experienced at low-load. Towards the end goal
of expanding the HCCI load envelope, this paper builds
on previous work [6] by describing a new online adaptive
machine learning method that enables fully causal cycleto-cycle combustion timing predictions across randomly
chosen engine set point transients that include both the
stable and near chaotic bifurcation behavior.
Start of fuel injection timing
Quantity of fuel
PNVO
330º to 358º
PEVO
PIVC
-50º to -30º
55º to 75º
48 milliseconds at 2,500 RPM
Exhaust partially out
-180
0
TDC
NVO
180
360 TDC 540
Crank angle [°CA]
Air in
720
TDC 900
Figure 1: A schematic of key engine cycle variables.
valve events are separated by a number of crank angle degrees, termed Negative Valve Overlap or NVO. Unlike conventional engines, NVO prevents some of the hot exhaust
gases from leaving the cylinder (typically 20-60% [2, 3]).
This stores “residual gases” for the next cycle, offering
a practical way to raise the mixture temperature to ensure HCCI autoignition [4]. By changing the amount of
NVO, one can affect the mixture temperature and dilution and ultimately control the chemical kinetics behind
combustion timing. Temperature and dilution work in opposite directions, but typically temperature dominates [5].
NVO is not instantly adjustable with common variable
valve timing systems, and the reader is cautioned that
many researchers publish results with fully variable (lift
and timing) electric or hydraulic valve actuation systems
that are expensive to implement in production engines.
The use of NVO residual gases introduces strong cycleto-cycle coupling on top of the already non-linear chemistry and physics that occur throughout a complete engine
cycle [6]. Further compounding the issues with residual
gases is that neither the airflow to the cylinder(s) nor the
quantity of residual gases in the cylinder can be accurately
resolved before a burn happens on a cycle-to-cycle (not
mean value) basis with commonly available sensors, especially during transients. Beyond residual gas influences,
there are also complex secondary influences on combustion
behavior such as turbulent mixing, manifold resonance effects, combustion deposits, different varieties of fuel and
even ambient temperature variations [7, 8].
While HCCI is already a significant challenge given the
above complexity, the combustion mode also exhibits a period doubling bifurcation cascade to chaos [6, 9, 10], similar
to what is seen in high residual spark ignition engines [11].
When nearly chaotic, HCCI is still deterministic, but becomes oscillatory and very sensitive to parameter variations (e.g. residual gas fraction fluctuations [9, 10]). This
oscillatory “stability limit” behavior is commonly referred
1.2. Experimental Observations
In the authors’ previous publication [6], an abstract
mapping function for engine combustion was created
within the framework of a discrete dynamical system:
𝑛𝑒𝑥𝑡 𝑐𝑜𝑚𝑏𝑢𝑠𝑡𝑖𝑜𝑛 =
𝑓 𝑢𝑛𝑐𝑡𝑖𝑜𝑛( 𝑝𝑟𝑒𝑣𝑖𝑜𝑢𝑠 𝑐𝑜𝑚𝑏𝑢𝑠𝑡𝑖𝑜𝑛, 𝑝𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟𝑠)
(1)
This simple abstraction is intended to convey a conceptual
understanding of the experimental cycle-to-cycle behavior
seen in Fig. 2’s return maps. These return maps show
the experimentally observed combustion timing for a given
cycle 𝑛 along the abscissa and the next cycle 𝑛 + 1 along
the ordinate under random engine actuator set points [6].
The value CA90 is the time in Crank Angle Degrees (∘ CA)
where 90% of the fuel’s net heat release is achieved, and
thus measures the timing of the end of the burn in relation
to piston’s position as the crank rotates.1
The reader should note that there is structure to the
cycle-to-cycle behavior despite the random actuator set
points used to generate Fig. 2. This structure shows a deterministic transition to oscillatory high CV behavior as
combustion moves towards later combustion timing that
can be viewed as at least a single period doubling bifurcation with sensitive dependence on the engine set point [6,
9, 10]. While mathematically interesting, this oscillatory
high CV structure undesirably constrains practical HCCI
engine operation. A more thorough description of these
data is provided in [6].
1 While
it is not shown here, there is similar structure in the CA10
and CA50 percent burn metrics, although less pronounced (especially in CA10, see [6]).
2
−4
CA50
CA10
15
0
−15
−15
45
0
15
30
CA10 n [°ATDC]
45
(a) Cyl. 1
High CV
−5
−15
−15
−3
45
−4
next =
function(previous)
0
15
30
CA90 n [°ATDC]
45
(c) Cyl. 3
CA90 n+1
CA90 n+1
45
(b) Cyl. 2
−5
“predictive” in the sense that when injected with random
(i.e. unpredictable) noise could it generate a qualitative
return map shape similar to what is seen experimentally.
Time series predictions were not shown, only a cloud of
possible combustion timings ranging from stable to oscillatory. This fact was highlighted in an extension of the
work [18] that again used random residual noise because
“the actual time series of disturbances to the experiments
are unknown.” That said, these models are useful for
showing that a period doubling cascade to chaos driven
by residual gas fraction can explain the observed high CV
behavior.
In the context of the above, machine learning provides
a computationally efficient way to capture complex combustion patterns while simultaneously avoiding explicit
knowledge of the underlying mixture state and composition (provided an appropriate abstract mapping function
is chosen). While there are clearly benefits to this machine
learning approach, a key issue is that machine learning
is data driven, and relatively large quantities of data are
needed to adequately cover large dimensional spaces. As
shown conceptually in Fig. 3, these high dimensional data
might be viewed as a “porcupine” [19]. Each engine operating condition might be viewed as a “quill” of Eq. 2’s
six-dimensional “porcupine,” and machine learning algorithms know nothing about the ideal gas law or chemical
kinetics, so their ability to extrapolate between “quills”
is limited, especially when provided sparse data. Previous work [6] used a random sampling of cycle time series
for training to ensure the data driven model had data to
fit the “quills,” and then assessed the model’s ability to
predict on the remaining (randomly chosen) cycles. Thus,
the training dataset was partially acausual and that the
model itself wasn’t shown to adapt to new conditions.
−3
15
−5
−15
−15
−3
45
−4
0
15
30
CA90 n [°ATDC]
45
(d) Cyl. 4
−5
−3
30
15
−4
0
CA90 n+1
CA90 n+1
45
0
30
−15
−15
0
15
30
CA50 n [°ATDC]
30
15
−15
−15
−4
0
30
0
15
15
−4
0
0
15
30
CA90 n [°ATDC]
45
−5
−15
−15
0
15
30
CA90 n [°ATDC]
45
−5
Figure 2: Return map probability histograms of CA90 generated from 129,964 cycles and 2,221 random engine set
points. Outliers are omitted and total only ∼3% of the
data. The colormap is 𝑙𝑜𝑔10 to show order of magnitude
differences. [6]
1.3. Modeling Approaches
In [6], a skeletal functional form for the abstract mapping function was built out using measurable quantities,
thermodynamics, and known correlations. Then, unlike
the physics-based approaches that are usually discussed in
the engine literature, the machine learning technique of
𝜖-Support Vector Regression (𝜖-SVR) was combined with
the abstract functional form to provide quantitative predictions. The primary motivation for this machine learning approach was not that existing chemical kinetics with
Computational Fluid Dynamics (CFD) cannot capture engine behavior (see [14] and gasoline mechanism validation [15]) but that the methods are:
Six-dimensional
offline mapping
function fit with
original β0
∙ Too computationally intensive for real-time predictions.
∙ Subject to experimental uncertainties in the cycleto-cycle (not mean value) mixture state and composition.
Online adaptation to move between the
“quills” and also adjust for parameter
variation with updated β1
Figure 3: High dimensional data might be viewed conceptually as a “porcupine” [19]. The primary goal of this
paper is to design an online adaptive algorithm to fit new
data between the “quills.”
As a point of reference, the simulation time of a 2,500
RPM, 48 millisecond engine cycle is measured in ∼day(s)
for a single core of a modern computer.
At the other computational complexity extreme, loworder approximation models of HCCI for control have been
developed since at least the early 2000s [16], based on spark
ignition engine knock models developed in the 1950s [17].
Recently, efforts have been made to extend this type of
model to the high CV regions of HCCI by injecting random residual gas fraction noise to capture uncertainties in
the mixture state and composition [10]. This model was
tuned for a limited set of steady-state conditions and only
1.4. Contribution
The primary contribution of this work is the development of a new online learning method to provide real-time
adaptive, fully causal predictions of near chaotic HCCI
combustion combustion timing. This method, called
Weighted Ring - Extreme Learning Machine (WR-ELM),
3
enables robust online updates to an Extreme Learning
Machine (ELM) model that is trained to fit the “quills”
of offline data.
WR-ELM is developed in this work as a weighted least
squares extension to Online Sequential - ELM [20]. While
developed independently, a similar derivation is available
in [22]. The difference between this work and the classification application in [22] is the use of a ring buffer data
structure “chunk” for online updates to an offline trained
regression model.
The data in the WR-ELM ring buffer can be weighted
more heavily than the data originally used to fit the offline
model. This allows emphasis to be placed on recent measurements that might be between the “quills” of Fig. 3’s
offline trained model or the result of day-to-day engine
parameter variation. Thus, this approach allows one to
prescribe a partitioned balance between the offline model
fit against the need to adapt to the most recent conditions. It also explicitly avoids over adaptation to the local
conditions (that could compromise global generality) by
“forgetting” old ring buffer data that eventually exit the
buffer. Fig. 4 gives a schematic representation of this approach.
2. Methods
2.1. Mapping Function Modifications
In previous work [6], engine combustion was abstracted
to the following mapping function:
𝐶𝐴50𝑛+1 =
𝑓 𝑢𝑛𝑐𝑡𝑖𝑜𝑛( 𝐶𝐴90𝑛 , 𝑇 𝐼, 𝑆𝑂𝐼, 𝑃𝐼𝑉 𝐶 , 𝑃𝐸𝑉 𝑂 , 𝑃𝑁 𝑉 𝑂 )
(2)
where 𝑛 is the cycle iteration index, CA50 is the time in
∘
CA where 50% of net heat release has occurred, CA90
is the time in ∘ CA when 90% of net heat release has
occured, 𝑇 𝐼 is the injection pulse width in milliseconds,
𝑆𝑂𝐼 is Start of Injection in ∘ CA Before Top Dead Center
(∘ BTDC), and the pressure variables measurements are
mean pressures during specific regions of the combustion
cycle (see Fig. 1).2 Details of the simplified net heat release
algorithm are available in [6]. Fuel rail pressure is constant; however, the reader should note that the pressure
drop to cylinder pressure during NVO injections varies
with each transient step and during high CV regions. The
cylinder pressure variables 𝑃𝐼𝑉 𝐶 , 𝑃𝐸𝑉 𝑂 , and 𝑃𝑁 𝑉 𝑂 were
chosen to capture cycle-to-cycle residual coupling and air
flow without the difficulties of explicitly modeling those
quantities. To meet real-time engine controller timing requirements, 𝑃𝐼𝑉 𝐶 and 𝑃𝑁 𝑉 𝑂 have been modified from [6].
𝑃𝐼𝑉 𝐶 has been moved to the previous cycle, and the range
of 𝑃𝑁 𝑉 𝑂 ’s mean has been shortened. 𝑃𝐼𝑉 𝐶 has also been
moved closer to TDC to take advantage of the inherent
signal amplification provided by the compression process.
The subscripts IVC, EVO, and NVO refer to the general
timing regions Intake Valve Close, Exhaust Valve Open,
and Negative Valve Overlap, respectively.
You want to predict the future output CA50n+2
… and you have the input vector xn+1
Cycle timeline:
n-8
n-7
forget
n-6
n-5
n-4
n-3
n-2
n-1
n
n+1
n+2
Weighted ring buffer of previous input-output data pairs
Figure 4: A schematic of overview WR-ELM.
2.2. WR-ELM Overview
Other differences from [20, 22] are that the derivation
below lacks a bias vector 𝑏, uses the Gaussian distribution for a, drops the unnecessary logistic function exponential negative, and uses a Padé approximant for 𝑒𝑥𝑝(𝑥).
It was found empirically that the computation of the bias
𝑏 addition step could be removed with no loss of fitting
performance if a’s elements were drawn from the Gaussian distribution 𝒩 (0, 1). ELM theory only requires the
distribution to be continuous [21], although the ability to
remove the bias is likely problem specific.
The primary benefits of an ELM approach over the
𝜖-SVR method used in [6] are:
2.3. WR-ELM Core Algorithm
The basic goal of an Extreme Learning Machine (ELM)
is to solve for the output layer weight vector 𝛽 that scales
the transformed input H to output T:
∙ An ELM is easily adapted to online adaptation [20].
∙ An ELM provides good model generalization when
the data are noisy [21].
H𝛽 = T
∙ An ELM is extremely computationally efficient [20,
21].
(3)
where H is the hidden layer output matrix of a given input
matrix and T is the target vector.
̃︀ neurons
For a set of 𝑛 input-output data pairs and 𝑁
at the 𝑛th cycle timestep, these variables are given by
⎡
⎤
𝐺(a1 , x1 ) . . . 𝐺(a𝑁̃︀ , x1 )
⎢
⎥
..
..
..
H(a, x) = ⎣
(4)
⎦
.
.
.
2 Since
CA90 is a stronger indicator of the oscillatory behavior seen
in Fig. 2, it is used as the model input. CA50 is more commonly
encountered in the engine literature and is thus used as the model
output. The two are related quantities, and in terms of model fit
statistics there was no significant benefit of using one over the other.
That said, a few isolated instances were observed where CA90 did
a better job predicting large oscillations.
𝐺(a1 , x𝑛 ) . . .
4
𝐺(a𝑁̃︀ , x𝑛 )
̃︀
𝑛×𝑁
where 𝐺(a𝑖 , x) is the neuron activation function, chosen
to be a commonly used logistic function, but without the
unnecessary negative:
𝐺(a𝑖 , x) =
1
.
1 + 𝑒𝑥𝑝(x · a𝑖 )
The solution can then be split between an offline and
online “chunk” of recent input-output data pairs to avoid
both the computational and storage burden of using offline
data directly. To do this, the matrices are partitioned with
subscript 0 and 1 denoting the offline and online updated
components, respectively:
[︂ ]︂
[︂
]︂
H0
W0
0
H=
W=
H1 𝑛×𝑁̃︀
0
W1 𝑛×𝑁
[︂ ]︂
(13)
T0
T=
.
T1 𝑛×1
(5)
Using a random input weight vector a𝑖 that is composed
of random variable (r.v.) samples from a Gaussian distribution for each of the 𝑧 input variables gives
⎤
⎡
𝑟.𝑣. ∼ 𝒩 (0, 1)
⎥
⎢
..
.
(6)
a𝑖 = ⎣
⎦
.
𝑟.𝑣.
∼
𝒩 (0, 1)
Then, following a similar derivation in [20] for recursive
least squares but adding the weight matrix, the inversion
portion H⊤ WH of the weighted normal equations Eq. 12
can be re-written in terms of K0 and K1 :
𝑧×1
̃︀
The use of a random a𝑖 that is re-sampled for each of the 𝑁
individual neurons during initialization is the main difference of an Extreme Learning Machine versus conventional
neural networks that iteratively train each a𝑖 [21]. These
a𝑖 vectors can then be collected into a single input weight
matrix a, which is held fixed across all 𝑛 input row vectors
x
x=
(7)
[𝐶𝐴90𝑛 𝑇 𝐼 𝑆𝑂𝐼 𝑃𝐼𝑉 𝐶 𝑃𝐸𝑉 𝑂 𝑃𝑁 𝑉 𝑂 ]𝑛×𝑧
K1 =
= H⊤ WH
[︂ ]︂⊤ [︂
]︂ [︂ ]︂
H0
W0
0
H0
=
H1
0
W1 H1
[︂
]︂ [︂ ]︂
[︀
]︀ W0
0
H0
⊤
= H⊤
H
0
1
0
W 1 H1
[︂ ]︂
[︀
]︀ H0
⊤
= H⊤
W
H
W
0
1
0
1
H1
and 𝑛 output values
[︀
]︀
T = 𝐶𝐴50𝑛+1 𝑛×1 .
(8)
⊤
= H⊤
0 W0 H0 + H1 W1 H1
= K0 + H⊤
1 W1 H1 .
While the above logistic works well, one modification
improves the computational efficiency on processors without a dedicated 𝑒𝑥𝑝(𝑥) instruction, such as the Raspberry
R
Pi○
. The modification is to replace the exponential with
the following Padé approximant:
2
𝑒𝑥𝑝(𝑦) ≈ 𝑝(𝑦) =
The non-inverted portion of the normal equations can similarly be re-written using existing relations:
H⊤ WT =
[︂ ]︂⊤ [︂
]︂ [︂ ]︂
H0
W0
0
T0
=
H1
0
W1 T1
[︂
]︂ [︂ ]︂
[︀ ⊤
]︀ W0
0
T0
⊤
= H0 H1
0
W1 T1
[︂ ]︂
[︀ ⊤
]︀ T0
⊤
= H0 W0 H1 W1
T1
3
120 + 60𝑦 + 12𝑦 + 𝑦
,
120 − 60𝑦 + 12𝑦 2 − 𝑦 3
(9)
which has the following simple logistic relations:
1
1
(120 + 12𝑦 2 ) − 60𝑦 − 𝑦 3
≈
=
(10)
1 + 𝑒𝑥𝑝(𝑦)
1 + 𝑝(𝑦)
2 · (120 + 12𝑦 2 )
The small number of floating point operations used, reused
intermediate terms, known boundedness of the normalized inputs, and known a weights make this approximant
work well in this application. No significant degradation
in model performance was found, and as such it is used in
all implementations described hereafter.
The normal equations can then be used to solve for the
least squares solution 𝛽 of Eq. 3 with
𝛽 = (H⊤ H)−1 H⊤ T .
=
=
=
(15)
⊤
H⊤
0 W0 T0 + H1 W1 T1
⊤
⊤
K0 K−1
0 H0 W0 T0 + H1 W1 T1
K0 𝛽 0 + H⊤
1 W1 T1
(︁
)︁
⊤
= K1 − H⊤
1 W1 H1 𝛽 0 + H1 W1 T1
⊤
= K1 𝛽 0 − H⊤
1 W1 H1 𝛽 0 + H1 W1 T1 .
Substituting Eq. 15 into the full online solution
(11)
𝛽1 =
)︁
(︀
)︀ (︁ ⊤
= K−1
H
WT
1
To extend this to a weighted least squares solution, one
can incorporate a diagonal weight matrix W to the normal
equations [23]:
𝛽 = (H⊤ WH)−1 H⊤ WT .
(14)
⊤
−1 ⊤
= 𝛽 0 − K−1
1 H1 W1 H1 𝛽 0 + K1 H1 W1 T1
⊤
= 𝛽 0 + K−1
1 H1 W1 (T1 − H1 𝛽 0 )
(12)
5
(16)
A = P0 H ⊤
1 , B = H1 A
(︀ −1
)︀−1
𝛽 1 = 𝛽 0 + A W1 + B
(T1 − H1 𝛽 0 )
yields the online solution without the need for the offline
̃︀ × 𝑁
̃︀
dataset. To trade the computational burden of the 𝑁
sized K inverse for an inverse that scales with a smaller
sized ring buffer, one can let P = K−1
(︁
)︁−1
⊤
P0 = K−1
(17)
0 = H0 W0 H0
)︁−1
(︁
⊤
−1
(18)
P1 = K−1
1 = P0 + H1 W1 H1
ONLINE PREDICTIONS
CA50𝑛+2 = T𝑛+1 = H(a, x𝑛+1 )𝛽 1
P1 =
⊤
−1
P0 − P0 H⊤
1 W1 + H1 P0 H1
)︁−1
(19)
H1 P0 .
Unlike OS-ELM and WOS-ELM [20, 22], additional
simplification is possible because the WR-ELM algorithm
does not propagate P1 in time. To begin, append the
⊤
H⊤
1 W1 portion of Eq. 16 to Eq. 19 and distribute H1 to
give:
P1 H ⊤
1 W1 =
[︂
)︁−1
(︁
−1
⊤
⊤
W
+
H
P
H
−
P
H
P0 H ⊤
1
0
0
1
1
1
1
]︂
·H1 P0 H⊤
1 W1 .
2.4. Usage procedure
1. Scale x and T columns between zero and unity for
each variable. For the combustion implementation,
column variable values below the 0.1% and above
the 99.9% percentile were saturated at the respective
percentile value, and then normalized between zero
and unity between these percentile based saturation
limits. This was done to both adequately represent
the distribution tails and avoid scaling issues.
2. The random non-linear transformation that enables
the low computational complexity of the WR-ELM
algorithm may result in ill-conditioned matrices. All
numerical implementations should use double precision. Additionally, one should consider using Singular Value Decomposition for ill-conditioned matrix
inversions.
3. Using the 𝒩 (0, 1) Gaussian distribution, initial̃︀ ELM input weights a and hold
ize the 𝑧 × 𝑁
them fixed for all training / predictions. For the
combustion implementation, this was done with
R
MATLAB○
’s built-in 𝑟𝑎𝑛𝑑𝑛() function and the
Mersenne Twister pseudo random number generator
̃︀ of 64 was used based
with seed 7,898,198. An 𝑁
on initial trials, and each cylinder’s individually
computed WR-ELM model used an identical input
weight matrix a.
4. Build H0 (a, x0 ) from previously acquired samples
that cover a wide range of conditions with Eq. 4
using an input matrix x0 and output target vector
T0 (the formats of these are given in Eqs. 7 and 8,
respectively). For the combustion implementation,
the initial training data were ∼40 minutes of random
engine set points covering 53,884 cycles and 1,179
random engine set points at a single engine speed;
however, it appears that only ∼20 minutes of data
may be sufficient. Pruning the training data to only
(20)
Eq. 20 can then be simplified with the substitutions A =
P0 H⊤
1 , B = H1 A and then distributing W1 to provide:
P1 H ⊤
1 W1 =
[︁
]︁
(︀ −1
)︀−1
A W1 − W1 + B
BW1 .
(21)
Transforming Eq. 21 with the identity (X + Y)−1 Y =
X−1 (X−1 + Y−1 )−1 [24] gives:
P1 H ⊤
1 W1 =
[︁
]︁
(︀
)︀−1
A W1 − W1 W1 + B−1
W1 .
(22)
Eq. 22 is then in a form where the identity X − X(X +
Y)−1 X = (X−1 + Y−1 )−1 [24] can be applied to yield a
substantially simpler form with a ring buffer sized inverse:
(︀ −1
)︀−1
P1 H ⊤
.
1 W1 = A W1 + B
(23)
−1 ⊤
Finally, noting that P1 H⊤
1 W1 = K1 H1 W1 , one can
then substitute Eq. 23 into Eq. 16 and arrive at the following algorithm summary:
OFFLINE TRAINING
[︂(︁
)︁−1 ]︂
⊤
P0 = H0 W0 H0
̃︀ ×𝑁
̃︀
𝑁
𝛽0 =
[︁
P0 H⊤
0 W0 T0
(26)
The reader should note that only P0 and 𝛽 0 are needed
for online adaptation, and the size of these matrices scales
̃︀ . None of
only with an increasing number of neurons 𝑁
the original offline data are needed. Additionally,
note that Eq. 26 is simply the reverse of Eq. 3 with the
most recent x𝑛+1 cycle vector and 𝛽 1 updated from the
weighted ring buffer. Finally, it should mentioned that the
resulting update law Eq. 25 is structurally similar that of
the steady-state Kalman filter [23], which also uses recursive least squares. Future work should look at applying
Kalman filtering algorithm improvements (e.g. square root
filtering) to WR-ELM.
and use the matrix inversion lemma on Eq. 18 to yield:
(︁
(25)
(24)
]︁
̃︀ ×1
𝑁
ONLINE ADAPTATION
6
include ∼6 cycles before and ∼9 cycles after a transient set point step provided a small model fitting
performance improvement.
5. Specify a weight matrix W0 for offline measurements. For the combustion implementation, a simple scalar value W0 = 3.5 × 10−3 was chosen using
a design of experiments. While this weight works
well as a proof of concept, future work should more
rigorously determine the weight(s), perhaps with optimization techniques. Note that W0 allows weighting to be applied offline and that a small offline
weighting is equivalent to a large online weighting.
6. Solve for the offline solution P0 and 𝛽 0 using Eqs. 24
and hold these values constant for all future predictions.
7. Populate a ring buffer of size 𝑟 with recently completed input-output pairs using:
⎤
⎡
x𝑛−𝑟+1
input
⎥
⎢
= x1 = ⎣ ... ⎦
ring buffer
x𝑛
𝑟×𝑧
(27)
⎡
⎤
𝐶𝐴50𝑛−𝑟+2
output
⎢
⎥
..
= T1 = ⎣
.
⎦
.
ring buffer
𝐶𝐴50𝑛+1 𝑟×1
was solved at an average rate of 1.1 𝜇𝑠 per combustion
R
cycle per cylinder on an Intel○
i7 860 2.8 GHz desktop
R
computer running Gentoo GNU/Linux○
. The online predictions from Eqs. 25, 27, and 26 were recast into a 𝑝𝑎𝑟𝑓 𝑜𝑟
loop that automatically parallelized the code across four
worker threads to provide predictions at an average rate
of 66 𝜇𝑠 per combustion cycle per cylinder. This level of
performance is more than adequate for real-time.
Although algorithm development is the main focus
of this paper, a real-time implementation of the WRELM algorithm has been built using custom, 18-bit RaspR
berry Pi○
data acquisition hardware. A two-minute video
demonstrating both predictions and control is available
at [25]. The software for this system is comprised of:
Then execute the WR-ELM update algorithm between combustion cycle 𝑛 + 1 and 𝑛 + 2 as shown
in Fig. 4. For the combustion implementation, 𝑟
was taken to be 8 cycles after tuning with existing
datasets. If desired, 𝑟 can vary cycle-to-cycle.
As with the offline data, build H1 (a, x1 ) with Eq. 4
using an input matrix x1 and output target vector
T1 . Specify a weight matrix W1 . For the combustion implementation the identity matrix (W1 =
I) was chosen since weighting was already applied
to the offline data in step 5. Gradually increased
weighting on the most recent time steps in the ring
buffer was explored; however, it did not net a significant improvement to model fitting performance
over a simple scalar value on offline data. Although
not explored in the current implementation, W1 can
vary cycle-to-cycle.
Solve for the updated 𝛽 1 solution using Eqs. 25.
After cycle 𝑛 + 1’s input vector x𝑛+1 is fully populated, transform vector into H𝑛+1 using Eq. 4 and
solve for a predicted target value T𝑛+1 or CA50𝑛+2
using Eq. 26.
Repeat steps 7-10 for each new time step, caching results (e.g. hidden layer outputs) from previous time
steps to reduce computational requirements.
R
∙ A multi-threaded Linux○
user space application
that runs heat release calculations and WR-ELM.
This software leverages the Eigen C++ matrix library and a custom assembly code matrix multiply3
for the Raspberry Pi 1’s VFPv2 double precision
floating point unit. Asynchronous fasync notification is used for specific crank angle events published
from the FIQ code to synchronize the user space
software’s execution against the crank’s rotation.
8.
9.
10.
11.
R
kernel with mi∙ A PREEMPT RT patched Linux○
nor patches to completely disable Fast Interrupt reQuest (FIQ) usage by the USB driver.
∙ ARM assembly code for high-speed pressure data acquisition using the FIQ (up to ∼240 kilosamples per
second, total all cylinder channels).
R
∙ A C code Linux○
kernel module that contains the
FIQ assembly code with page-based memory allocation and standard mmap, ioctl, and fasync hooks to
R
user space.
Linux○
∙ A low-priority second user space thread that uses
standard WebSockets from the libwebsockets C code
library to stream processed data to a web-based user
interface coded in JavaScript with the d3.js library.
R
∙ A minimal Raspbian GNU/Linux○
distribution [26].
After the adaptation routine is run, it is possible to perform 11 model predictive control predictions within a worst
case task context switch and calculation latency window
of ∼300 𝜇𝑠. This level of real-time performance is needed
to ensure control authority with the 𝑆𝑂𝐼 actuator immediately after 𝑃𝑁 𝑉 𝑂 is measured.
2.6. Experimental Setup
Table 1 provides a summary of the experimental setup
and conditions visited. In-cylinder pressure was acquired
2.5. Real-Time Implementation
R
A collection of unoptimized MATLAB○
software routines was developed using the techniques described in the
previous sections. The offline solution provided by Eqs. 24
3 This
code was benchmarked to be 29% faster than OpenBLAS’s
VFPv2 GEneric Matrix Multiply (GEMM) and 126% faster than
Eigen C++.
7
on a 1.0 ∘ CA basis and pegged thermodynamically for each
cycle after IVC using a polytropic exponent of 1.35. This
exponent was chosen to most closely match the pegging
results achieved using the single intake runner high speed
pressure sensor on cylinder 1. For the purpose of computing cycle-to-cycle net Indicated Mean Effective Pressure
(IMEP), a cycle was defined as starting at 360 ∘ BTDC
firing and ending at 359 ∘ ATDC firing. The reference cam
lift for timing and duration in Table 1 is 0.5 mm. The airfuel ratio range indicated in Table 1 was measured postturbine, and represents a mixture from all four cylinders.
Fuel mass per cycle was estimated using the fuel’s lower
heating value, assuming that the gross heat release was
20% greater than the net heat release, and that the combustion efficiency was 100%.
with 𝑆𝑂𝐼 fixed is only used as part of Fig. 2, and not during the model training and testing presented here. Total
cycle counts are reported after outliers are removed. The
outlier criteria (detailed in [6]) are intended to remove misfires and partial burns. These criteria are fairly permissive
and remove only ∼3% of the data.
The offline solution was trained using ∼40 minutes of
test cell time covering 53,884 cycles and 1,179 random engine set points at 2,500 rpm (two random subsequences).
These subsequences are comprised of random, transient
set point steps occurring approximately every 0.5 - 10 seconds with occasional misfires covering the nominal variable
ranges given in Table 1. The training data were pruned to
only include 6 cycles before and 9 cycles after a transient
set point step for a small model fitting performance improvement. The online solution was run with a separate
random subsequence and fed unseen cycles one-by-one,
similar to what would be experienced in a real-time implementation. This online dataset is comprised of 25,323 consecutive cycles with 521 random engine set points. Longer
online sequences were also tested, and achieved similar results.
2.7. Dataset Description
The full collection of 129,964 cycles is comprised of five
∼20 minute random test subsequences. Each random subsequence covers the same nominal ranges listed in Table 1;
however, one subsequence holds 𝑆𝑂𝐼 fixed. The sequence
Table 1: Experimental setup and test conditions
3. Results and Discussion
Engine
The fitting performance on a 25,323 cycle dataset (excluding outliers as defined in [6]) is shown in Table 2 and
in Figs. 5, 6, and 7. The minimum coefficient of determination (𝑅2 ) given in Table 2 shows that at least 80% of the
cycle-to-cycle variance can be explained by the model as
it is currently defined for a dataset with random transient
steps occurring approximately every 0.5 - 10 seconds and
occasional misfires. This is better than the 76% achieved
with 𝜖-Support Vector Regression (𝜖-SVR) on the same
dataset in [6]. However, this is not a 1:1 comparison because WR-ELM is fully predicting the entire 25,323 cycle dataset, whereas 𝜖-SVR’s training strategy ensured the
data-driven model had partially seen the operating points
it was trying to predict. Steady-state Root Mean Squared
Error (RMSE) in Table 2 was assessed at a single set point
with a mean CA50 of 3.9 ∘ ATDC and a net Indicated Mean
Effective Pressure (IMEP) of 2.8 bar before the transient
sequence started.
Make / model
GM / LNF Ecotec
Cylinder layout
in-line 4
Overall displacement
2.0 L
Bore / stroke
86 / 86 mm
Geometric compression ratioa
11.2 : 1
Cam lifta
3.5 mm
Cam durationa
93 ∘ CA
Cam phaser type
hydraulic
Fuel injector type
direct, side mounted, wall guided
Fuel
Designation
Haltermann HF0437, EPA Tier II EEE
Description
U.S. Federal Emission Cert. Gasoline
Research Octane Number
97.0
Motor Octane Number
88.1
ASTM D240 heating value
42.8 MJ / kg
Aromatic / olefin / saturate fractions
28 / 1 / 71 % volume
Test conditions
Throttle position
Turbocharger
Supercharger
Residual retention strategy
IVO set point rangeb
EVC set point rangeb
𝑆𝑂𝐼 set point rangeb
𝑇 𝐼 set point rangeb
Net IMEP values visitedb
Air-fuel ratios visitedb
Estimated fuel per cycleb
Intake runner temps., all cyls.
Fuel injection pressure
Coolant temperature
Engine speed
wide open
wastegate open
bypassed
negative valve overlap
78.6 / 128 ∘ ATDC
-118 / -83.0 ∘ ATDC
272 / 378 ∘ BTDC
0.582 / 1.01 ms
1.85 / 3.62 bar
0.90 / 1.6
6 / 11 mg
𝜇 = 52.6 ∘ C, 𝜎 = 1.6 ∘ C
𝜇 = 70.0 bar, 𝜎 = 0.85 bar
𝜇 = 89.5 ∘ C, 𝜎 = 3.4 ∘ C
𝜇 = 2, 500 RPM, 𝜎 = 6 RPM
Table 2: WR-ELM model of Eq. 2 error statistics
Cylinder
#
1
2
3
4
c
a
b
Overall𝑐
𝑅2
0.81
0.81
0.80
0.83
Overall
RMSE [∘ CA]
1.85
2.06
2.17
1.64
Steady-State
RMSE [∘ CA]
0.84
0.97
0.97
0.86
25,323 consecutive cycles with random transient steps
occurring approx. every 0.5 - 10 sec. and occasional misfires.
Modified from stock engine
First to 99th percentile
8
µ = 0.3
σ = 1.8
1500
0
−8 −6 −4 −2 0 2 4 6 8
CA50 model error [deg]
(c) Cyl. 3
4500
µ = 0.3
σ = 2.2
3000
1500
0
−8 −6 −4 −2 0 2 4 6 8
CA50 model error [deg]
4500
Frequency [# of cycles]
3000
(a) Cyl. 1
Frequency [# of cycles]
Frequency [# of cycles]
Frequency [# of cycles]
4500
3000
(b) Cyl. 2
under predicted and that the positive bias is largely from
the midrange values of CA50. Fig. 7a-d shows the cycleto-cycle time series predictions, which can be computed as
early as 358 ∘ BTDC firing. Missing segments are the outliers described earlier. Fig. 7e-h provide qualitative insight
into the model’s 𝛽 1 weights under online adaptation. The
neurons are sorted by the 2-norm of their respective input weight vector a𝑖 . The same non-linear transformation
specified by a is used for each cylinder, and any cylinderto-cylinder differences in the cycle-to-cycle 𝛽 1 are due to
different characteristics of each cylinder. Fig. 7i shows the
IMEP and Fig. 7j-l shows the random engine actuator inputs that are driving each engine set point transient.
The model predictions of Fig. 7a-d generally show good
agreement; however, there are occasional tracking errors.
It is unclear what the source of these tracking errors is
(e.g. is it a fundamental model limitation, the need for
more inputs,4 the influence of other misfiring cylinders going through a harsh reignite, the need for more weight tuning or offline training data, or perhaps something else?).
Future work will try to answer these questions. Overall,
however, the authors believe the level of fit shown in Table 2 and in Figs. 5, 6, and 7 is very good considering that
the dataset includes both transients and operating points
with high CV, right up to complete engine misfire.
µ = 0.4
σ = 2.0
1500
0
−8 −6 −4 −2 0 2 4 6 8
CA50 model error [deg]
(d) Cyl. 4
4500
µ = 0.2
σ = 1.6
3000
1500
0
−8 −6 −4 −2 0 2 4 6 8
CA50 model error [deg]
Figure 5: Error histograms for WR-ELM model of Eq. 2
across 25,323 consecutive cycles with random transient
steps occurring approx. every 0.5 - 10 seconds and occasional misfires.
(a) Cyl. 1 CA50
20
10
0
−10
30
Model predicted
Model predicted
30
20
10
0
−10
−20
−20 −10 0 10 20 30
Measured [°ATDC]
20
4. Summary and Conclusion
10
This work presents a new online adaptation algorithm
named Weighted Ring - Extreme Learning Machine. The
approach uses a weighted ring buffer data structure of recent measurements to recursively update an offline trained
Extreme Learning Machine solution. It is shown that WRELM can be used to approximate the combustion mapping
function developed in [6] and provide reasonably accurate,
causal predictions of near chaotic combustion behavior. In
the combustion application only in-cylinder pressure and
crank encoder sensors are needed for predictions, and these
predictions can be computed as early as 358 ∘ BTDC firing. The algorithm is fast, and has been implemented in
R
real-time on the low-cost Raspberry Pi○
platform (a twominute video demonstrating this is available at [25]).
Future work will explore optimal selection of weight(s)
and try to better understand the situations that lead to
the occasional model tracking errors. Finally, the broader
objective of this new modeling approach is to enable a new
class of cycle-to-cycle model predictive control strategies
that could potentially bring HCCI’s low engine-out NOx
and reduced CO2 emissions (higher fuel efficiency) to production gasoline engines.
0
−10
−20
−20 −10 0 10 20 30
Measured [°ATDC]
(d) Cyl. 4 CA50
30
Model predicted
Model predicted
−20
−20 −10 0 10 20 30
Measured [°ATDC]
(c) Cyl. 3 CA50
30
(b) Cyl. 2 CA50
20
10
0
−10
−20
−20 −10 0 10 20 30
Measured [°ATDC]
Figure 6: Predicted versus measured WR-ELM model of
Eq. 2 across 25,323 consecutive cycles with random transient steps occurring approximately every 0.5 - 10 seconds
and occasional misfires. Late combustion timing is under
predicted, but almost all prediction outliers capture the
correct directionality.
Fig. 5 shows the distribution of model errors. It is
clear that there is a slight positive bias to the predictions.
Fig. 6 provides insight into the tails of Fig. 5 and shows
that model errors still generally capture the correct directionality. Fig. 6 also shows that late combustion timing is
4 Adding
additional inputs is not necessarily practical computationally, experimentally, or even advisable given Occam’s razor.
9
−10
The model can track
near chaotic, high CV (c) Cyl. 3
Model predicted
5
Sorted Neuron #
10
Measured
20
misfire region...
−10
−25
35
The model tracks harsh
Each cyl. is different
transients up to misfire (d) Cyl. 4 despite same set-points
Measured
Model predicted
20
5
−10
−25
11,375
11,420
11,465
Cycle #
11,510
11,555
Cyl. 1
2
3
4
3
4
3
4
4.5
3
1.5
0
−25
(f) WR−ELM β Weight for Cyl. 2
64
55
46
37
28
19
10
1
25
(j) Injection Pulse Width
1.1
Cyl. 1
2
0
TI [ms]
0.95
0.8
0.65
(g) WR−ELM β Weight for Cyl. 3
64
55
46
37
28
19
10
1
64
55
46
37
28
19
10
1
IMEP [bar]
0
(i) Indicated Mean Effective Pressure
6
25
−25
0.5
25
380
0
(h) WR−ELM β Weight for Cyl. 4
SOI [°BTDC]
Sorted Neuron #
5
Sorted Neuron #
−25
The model often tracks
Tracking errors
CA50 cycle-to-cycle (b) Cyl. 2 occur occasionally
35
Measured
Model predicted
20
Sorted Neuron #
CA50 [°ATDC]
CA50 [°ATDC]
−10
35
CA50 [°ATDC]
Model predicted
5
−25
CA50 [°ATDC]
Measured
20
(e) WR−ELM β Weight for Cyl. 1
64
55
46
37
28
19
10
1
(k) Start of Injection
Cyl. 1
320
290
260
25
150
(l) Approx. EVC [ °BTDC] and IVO [ °ATDC]
EVC Cyl. 1
2
130
0
2
350
−25
[°CA]
(a) Cyl. 1
35
IVO Cyl. 1
3
4
110
90
11,375
11,420
11,465
Cycle #
11,510
11,555
−25
70
11,375
11,420
11,465
Cycle #
11,510
11,555
Figure 7: The WR-ELM CA50 model of Eq. 2 can track CA50 through transients every 0.5 - 10 seconds, operating points with near chaotic high CV, and at
steady-state during a particularly harsh region of the 25,323 cycle dataset that includes misfires. The colormaps (linearly scaled) provide qualitative insight
into the level of cycle-to-cycle adaptation and into cylinder-to-cylinder model differences.
Acknowledgments
[10] E. Hellström, A. G. Stefanopoulou, L. Jiang, Cyclic Variability
and Dynamical Instabilities in Autoignition Engines with High
Residuals, IEEE Transactions on Control Systems Technology
21 (5) (2013) 1527–1536.
[11] J. C. Kantor, A Dynamical Instability of Spark Ignited
Engines, Science 224 (4654) (1984) 1233–1235.
[12] L. Manofsky, J. Vavra, D. Assanis, A. Babajimopoulos,
Bridging the Gap between HCCI and SI: Spark-Assisted
Compression Ignition SAE Paper 2011-01-1179, 2011.
[13] M. Sjöberg, J. E. Dec, A. Babajimopoulos, D. Assanis,
Comparing Enhanced Natural Thermal Stratification Against
Retarded Combustion Phasing for Smoothing of HCCI
Heat-Release Rates SAE Paper 2004-01-2994, 2004.
[14] M. Mehl, W. J. Pitz, M. Sjöberg, J. E. Dec, Detailed Kinetic
Modeling of Low-Temperature Heat Release for PRF Fuels in
an HCCI Engine SAE Paper 2009-01-1806, 2009.
[15] Experimental and surrogate modeling study of gasoline
ignition in a rapid compression machine, Combustion and
Flame 159 (10) (2012) 3066 – 3078.
[16] G. M. Shaver, Physics-Based Modeling and Control of
Residual-Affected HCCI Engines using Variable Valve
Actuation, Ph.D. thesis, Stanford University, 2005.
[17] J. Livengood, P. Wu, Correlation of autoignition phenomena in
internal combustion engines and rapid compression machines,
Symposium International on Combustion 5 (1) (1955) 347–356.
[18] S. Jade, J. Larimore, E. Hellström, A. Stefanopoulou,
L. Jiang, Controlled Load and Speed Transitions in a
Multicylinder Recompression HCCI Engine, Control Systems
Technology, IEEE Transactions on PP (99) (2014) 1–1.
[19] V. Cherkassky, F. Mulier, Learning from Data: Concepts,
Theory, and Methods, Wiley, 2007.
[20] N.-Y. Liang, G.-B. Huang, P. Saratchandran,
N. Sundararajan, A fast and accurate online sequential
learning algorithm for feedforward networks, Neural Networks,
IEEE Transactions on 17 (6) (2006) 1411–1423.
[21] G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning
machine: Theory and applications, Neurocomputing 70 (1-3)
(2006) 489 – 501.
[22] B. Mirza, Z. Lin, K.-A. Toh, Weighted Online Sequential
Extreme Learning Machine for Class Imbalance Learning,
Neural Processing Letters (2013) 1–22.
[23] D. Simon, Optimal State Estimation: Kalman, H Infinity, and
Nonlinear Approaches, Wiley-Interscience, 2006.
[24] John Halleck’s Matrix Identities (also Ring Identities), URL:
http://www.cc.utah.edu/∼nahaj/math/matrix-identities.html,
[Online; accessed 12/31/2014], 2014.
[25] Version 0.1 Raspberry Pi Engine Control with Real-Time
Adaptive Extreme Learning Machine, URL:
https://www.youtube.com/watch?v=qQG7ocnE3EA, [Online;
accessed 1/8/2015], 2015.
[26] Minimal Raspbian unattended netinstaller, URL:
https://github.com/debian-pi/raspbian-ua-netinst, [Online;
accessed 9/5/2014], 2014.
This material is based upon work supported by the Department of Energy [National Energy Technology Laboratory] under Award Number(s) DE-EE0003533. This work
is performed as a part of the ACCESS project consortium
(Robert Bosch LLC, AVL Inc., Emitec Inc., Stanford University, University of Michigan) under the direction of PI
Hakan Yilmaz and Co-PI Oliver Miersch-Wiemers, Robert
Bosch LLC.
The authors thank Vijay Janakiraman for providing
the raw data analyzed in this paper. The authors also
thank Jeff Sterniak for both his test cell and project support. A. Vaughan thanks Nisar Ahmed for many helpful discussions, and his advisors S. V. Bohac and Claus
Borgnakke for the freedom to explore the interesting topic
of near chaotic combustion.
Conflict of Interest
The University of Michigan has filed a provisional
patent on the work described in this publication. A.
Vaughan is named as the inventor, which includes royalty
rights. S. V. Bohac declares no conflict of interest.
References
[1] A. Vaughan, S. V. Bohac, An Extreme Learning Machine
Approach to Predicting Near Chaotic HCCI Combustion
Phasing in Real-Time, arXiv pre-print (1310.3567), 2013.
[2] A. Babajimopoulos, V. Challa, G. Lavoie, D. Assanis,
Model-Based Assessment of Two Variable CAM Timing
Strategies for HCCI Engines: Recompression vs. Rebreathing,
in: Proceedings of the ASME Internal Combustion Engine
Division 2009 Spring Technical Conference, ICES2009-76103,
2009.
[3] E. A. O. Soto, J. Vavra, A. Babajimopoulos, Assessment of
Residual Mass Estimation Methods for Cylinder Pressure Heat
Release Analysis of HCCI Engines With Negative Valve
Overlap, Journal of Engineering for Gas Turbines and Power
134 (8) (2012) 082802.
[4] L. M. Olesky, J. Vavra, D. Assanis, A. Babajimopoulos,
Effects of Charge Preheating Methods on the Combustion
Phasing Limitations of an HCCI Engine With Negative Valve
Overlap, Journal of Engineering for Gas Turbines and Power
134 (11) 112801.
[5] S. Saxena, I. D. Bedoya, Fundamental phenomena affecting
low temperature combustion and HCCI engines, high load
limits and strategies for extending these limits, Progress in
Energy and Combustion Science 39 (5) (2013) 457 – 488.
[6] A. Vaughan, S. V. Bohac, A Cycle-to-Cycle Method to Predict
HCCI Combustion Phasing, in: Proceedings of the ASME
Internal Combustion Engine Division 2013 Fall Technical
Conference, ICEF2013-19203, 2013.
[7] J. S. Lacey, The effects of advanced fuels and additives on
homogeneous charge compression ignition combustion and
deposit formation, Ph.D. thesis, University of Michigan, 2012.
[8] A. Vaughan, G. Delagrammatikas, A High Performance,
Continuously Variable Engine Intake Manifold SAE Paper
2011-01-0420, 2011.
[9] Understanding the transition between conventional
spark-ignited combustion and HCCI in a gasoline engine,
Proceedings of the Combustion Institute 31 (2) (2007) 2887 –
2894.
DISCLAIMER: This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United
States Government nor any agency thereof, nor any of their employees,
makes any warranty, express or implied, or assumes any legal liability or
responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its
use would not infringe privately owned rights. Reference herein to any
specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States
Government or any agency thereof. The views and opinions of authors
expressed herein do not necessarily state or reflect those of the United
States Government or any agency thereof.
11
| 5cs.CE
|
arXiv:1611.02851v2 [math.PR] 24 Jan 2018
Vol. 0 (0000)
Regularity Properties and Simulations
of Gaussian Random Fields on the
Sphere cross Time
Jorge Clarke De la Cerda
∗
Departamento de Matemática, Universidad Técnica Federico Santa Marı́a.
Avda. España 1680, Valparaı́so, Chile.
CEREMADE, UMR CNRS 7534 Université Paris-Dauphine, PSL Research university,
Place du Maréchal de Lattre de Tassigny 75016 Paris, FRANCE.
e-mail: [email protected]
Alfredo Alegrı́a
†
Newcastle University, School of Mathematics and Statistics.
e-mail: [email protected]
and
Emilio Porcu‡
Newcastle University, School of Mathematics and Statistics.
Chair of Spatial Analytics Methods Center.
Departamento de Matemática, Universidad Técnica Federico Santa Marı́a.
Avda. España 1680, Valparaı́so, Chile.
e-mail: [email protected]
url: eporcu.mat.utfsm.cl
Abstract: We study the regularity properties of Gaussian fields defined
over spheres cross time. In particular, we consider two alternative spectral
decompositions for a Gaussian field on Sd × R. For each decomposition, we
establish regularity properties through Sobolev and interpolation spaces.
We then propose a simulation method and study its level of accuracy in
the L2 sense. The method turns to be both fast and efficient.
MSC 2010 subject classifications: Primary 60G60, 60G17, 41A25; secondary 60G15, 33C55, 46E35, 33C45.
Keywords and phrases: Gaussian random fields, Global data, Big data,
Space-time covariance, Karhunen-Loève expansion, Spherical harmonics functions, Schoenberg’s functions.
∗ supported
by Proyecto FONDECYT Post-Doctorado No 3150506.
by Beca CONICYT-PCHA / Doctorado Nacional / 2016-21160371.
‡ supported by Proyecto FONDECYT Regular No 1170290.
† supported
0
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
1
1. Introduction
Spatio-temporal variability is of major importance in many fields, in particular for anthropogenic and natural processes, such as earthquakes, geographic
evolution of diseases, income distributions, mortality fields, atmospheric pollutant concentrations, hydrological basin characterization and precipitation fields,
among others. For many natural phenomena involving, for instance, climate
change and atmospheric variables, many branches of applied sciences have been
increasingly interested in the analysis of data distributed over the whole sphere
representing planet Earth and evolving through time. Hence the need for random fields models where the spatial location is continuously indexed through
the sphere, and where time can be either continuous or discrete. It is common
to consider the observations as a partial realization of a spatio-temporal random field which is usually considered to be Gaussian (see [6, 7, 10]). Thus, the
dependence structure in space-time is governed by the covariance of the spatiotemporal Gaussian field, and we refer the reader to [13, 24, 27] for significant
contributions in this direction.
Specifically, let d be a positive integer, and let Sd = {x ∈ Rd+1 , kxk = 1} be
the d-dimensional unit sphere in the Euclidean space Rd+1 , where k · k denotes
the Euclidean norm of x ∈ Rd+1 . We denote Z = {Z(x, t), (x, t) ∈ Sd × R} a
Gaussian field on Sd × R.
The tour de force in [17] characterizes isotropic Gaussian random fields on the
sphere Sd through Karhunen–Loève expansions with respect to the spherical
harmonics functions and the angular power spectrum. They show that the
smoothness of the covariance is connected to the decay of the angular power
spectrum and then discuss the relation to sample Hölder continuity and sample
differentiability of the random fields.
The present paper extends part of the work of [17] to space-time. Such extension is non-trivial and depends on two alternative spectral decompositions of a
Gaussian field on spheres cross time. In particular, we propose either Hermite or
classical Karhunen-Loève expansions, and show how regularity properties can
evolve dynamically over time. The crux of our arguments rely on recent advances
on the characterization of covariance functions associated to Gaussian fields on
spheres cross time (see [3] and [21]). Notably, the Berg-Porcu representation
in terms of Schoenberg functions inspires the proposal of alternative spectral
decompositions for the temporal part, which become then crucial to establish
the regularity properties of the associated Gaussian field.
The second part of the paper is devoted to simulation methods which should be
computationally fast while keeping a reasonable level of accuracy, resulting in a
notable step forward. Efficient simulation methods for random fields defined on
the sphere cross time are, currently, almost unexplored. Cholesky decomposition
is an appealing alternative since it is an exact method. However, the method
has an order of computation of O(N 3 ), with N denoting the sample size. This
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
2
makes its implementation computationally challenging for large scale problems
(the so called Big “n”problem). It is therefore mandatory to investigate efficient
simulation methods. Here, we propose a simulation method based on a suitable
truncation of the proposed double spectral decompositions. We establish its
accuracy in the L2 sense and illustrate how the model keeps a reasonable level
of accuracy while being considerably fast, even when the number of spatiotemporal locations is very high.
The remainder of the article is as follows. Section 2 provides the basic material
for a self-contained exposition. The expansions for the kernel covariances and the
random field are presented in Section 3. Section 4 presents the regularity results
for the kernel covariance functions in terms of weighted Sobolev spaces and
weighted bi-sequence spaces. In Section 5 our simulation method is developed
and its accuracy is studied. Also, we provide some numerical experiments for
illustrative purposes. In Appendix 7.1 we also provide a rather general version
of the Karhunen-Loève theorem.
The manuscript is intended for complex-valued random fields over Sd × R with
d ∈ N, except for Section 5 where the simulations considered are for real-valued
random fields over S2 × R.
2. Preliminaries
This section is largely expository and devoted to the illustration of the framework and notations that will be of major use throughout the manuscript. All
the tools presented in this section are valid in Sd , for any d ∈ N. Some particular
references to the case d = 2 are exposed, as they will be of use in Section 5.
2.1. Spherical Harmonics Functions and Gegenbauer Polynomials
Spherical harmonics are restrictions to the unit sphere Sd of real harmonics
polynomials in Rd+1 . They are also the eigenfunctions of the Laplace-Beltrami
operator on Sd . A deeper overview on spherical harmonics along with the properties listed in this subsection can be found in [9].
For d ∈ N, let L2 (Sd , σd ; C) = L2 (Sd ) denote the space of complex-valued square
integrable functions over Sd , where σd denotes the surface area measure, and
kσd k denotes the surface area of Sd ,
Z
kσd k :=
dσd =
Sd
with Γ being the Gamma function.
2π (d+1)/2
,
Γ((d + 1)/2)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
3
For j = 0, 1, . . . , let Hjd denote the linear space of spherical harmonics of degree
j over Sd . For different degrees, spherical harmonics are orthogonal with respect
to the inner product of L2 (Sd ). If Yj,m,d ∈ Hjd and Yj 0 ,m,d ∈ Hjd0 , then
Z
hYj,m,d , Yj 0 ,m,d iL2 (Sd ) :=
Yj,m,d (x)Yj 0 ,m,d (x) dσd (x)
Sd
= δj,j 0 · kσd k;
j, j 0 ∈ N, m = 0, . . . , dim(Hjd ),
where δj,j 0 is the Kronecker delta function, being identically equal to one if
j = j 0 , and zero otherwise.
Corollary 1.1.4 in [9] shows that
dim(Hjd ) =
(2j + d − 1)(j + d − 2)!
,
j!(d − 1)!
j ≥ 1,
dim H0d = 1.
(2.1)
Let Yj,m,d : m = 1, . . . , dim(Hjd ) be any orthonormal basis of Hjd . Then, the
family Sd := Yj,m,d : j ∈ N0 , m = 1, . . . , dim(Hjd ) constitutes an orthonormal
basis of L2 (Sd ) and Theorem 2.42 of [19] shows that
L2 (Sd ) =
+∞
M
Hjd .
j=0
Besides, the addition formula for spherical harmonics states
dim(Hd
j)
X
Yj,m,d (x)Yj,m,d (y) =
m=1
(2j + d − 1) (d−1)/2
Cj
(hx, yiRd+1 ),
(d − 1)
j ∈ N, (2.2)
where h·, ·iRd+1 denotes the inner product in Rd+1 . Here Cjr (·) are the Gegenbauer (or ultraspherical) polynomials of degree j and order r, defined by
Cjr (x) =
(2r)j
(r−1/2),(r−1/2)
P
(x),
(r + 1/2)j j
r > −1/2, x ∈ [−1, 1],
(α,β)
where Pj
(·) denotes the Jacobi polynomial of parameters α, β > −1 and
degree j, and (·)j denotes the Pochhammer symbol (rising factorial) defined by
(r)j := r(r + 1) · . . . · (r + j − 1),
r ∈ R, j ∈ N0 ,
where (r)j = Γ(r + j)/Γ(r), provided r is not a negative integer.
Gegenbauer polynomials constitute a basis of L2 (−1, 1) and satisfy the orthogonality relation (section 3.15 in [2])
Z 1
π21−2r Γ(j + 2r)
Cjr (x)Cjr0 (x)(1 − x2 )r−1/2 dx =
δj,j 0 .
(2.3)
j!(j + r)Γ(r)2
−1
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
4
By Stirling’s inequalities, for n ∈ N fixed, there exists constants c1 (n) and c2 (n)
such that
π21−n Γ(j + n)
≤ c2 (n)j n−2 .
c1 (n)j n−2 ≤
j!(j + n/2)Γ(n/2)2
Hence, assuming 2r ∈ N, relation (2.3) becomes
Z
1
Cjr (x)Cjr0 (x)(1 − x2 )r−1/2 dx ' j 2r−2 δj,j 0 .
(2.4)
−1
Besides, for n ≤ j we observe that (see section 10.9 in [2])
∂n r
r+n
r+n
C (x) = 2n (r)n Cj−n
(x) ' Cj−n
(x), x ∈ [−1, 1]
∂xn j
(2.5)
In what follows cj (d, x) denotes the standardized Gegenbauer polynomial, being
identically equal to one for x = 1 and r = (d − 1)/2, that is
(d−1)/2
cj (d, x) =
Cj
(x)
(d−1)/2
Cj
(1)
=
j!
(d−1)/2
C
(x),
(d − 1)j j
x ∈ [−1, 1].
(2.6)
It is straightforward to see that
dim(Hjd )
(d−1)/2
Cj
(1)
=
2j + d − 1
' j.
d−1
(2.7)
1/2
Remark 2.1. When d = 2, i.e., the S2 case, Cj (·) = Pj (·), where Pj (·) is the
Legendre polynomial of degree j [9].
See [9] and [19] for a deeper overview of spherical harmonics, and [2] for a
detailed description of Gegenbauer, Jacobi and Legendre polynomials.
2.2. Isotropic Stationary Gaussian Random Fields on the Sphere
cross Time
Let (Ω, F, P) be a complete probability space. Consider Sd and Sd × R as manifolds contained in Rd+1 and in Rd+2 respectively.
Definition 2.1. The geodesic distance θ on Sd (“great circle” or “spherical”
distance) is defined by
θ(x, y) = arccos(hx, yiRd+1 ),
x, y ∈ Sd .
The geodesic distance ρ on Sd × R is defined through
q
2
ρ((x, s), (y, t)) = θ (x, y) + (t − s)2 , (x, s), (y, t) ∈ Sd × R.
(2.8)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
5
Throughout, unless it is explicitly presented in a different way, we write Sd
instead of (Sd , σd ), σd being the surface area measure, which is equivalent to
other uniformly distributed measures on Sd , such as the Haar measure or the
Lebesgue spherical measure [8]. Analogously, we write R instead of (R, µ), for µ
being the Lebesgue measure.
Definition 2.2. A F ⊗ B(Sd × R)-measurable mapping, Z : Ω × Sd × R 7→ C,
is called a complex-valued random field on Sd × R.
A complex-valued random field Z = Z(x, t) : x ∈ Sd , t ∈ R is called Gaussian if for all k ∈ N, (x1 , t1 ), . . . , (xk , tk ) ∈ Sd × R, the random vector
>
(ReZ(x1 , t1 ), . . . , ReZ(xk , tk ), ImZ(x1 , t1 ), . . . , ImZ(xk , tk )) ,
is multivariate Gaussian distributed. Here, > denotes the transpose operator.
A function h : Sd × R × Sd × R −→ C is positive definite if
n X
n
X
ci cj h(xi , ti , xj , tj ) ≥ 0,
(2.9)
i=1 j=1
n
for all finite systems of pairwise distinct points {(xk , tk )}k=1 ⊂ Sd × R and
constants c1 , . . . , cn ∈ C. A positive definite function h is strictly positive definite
if the inequality (2.9) is strict, unless c1 = · · · = cn = 0.
We call a function h : Sd × R × Sd × R −→ C spatially isotropic and temporally
stationary if there exists a function ψ : [−1, 1] × R −→ C such that
h(x, t, y, s) = ψ (cos θ(x, y), t − s) ,
(x, t), (y, s) ∈ Sd × R.
(2.10)
Hence, a spatially isotropic and temporally stationary function depends on its
arguments via the great circle distance θ(x, y) and the time lag, or equivalently,
via the inner product hx, yiRd+1 and the time
lag.
Definition 2.3. We call a random field Z = Z(x, t) : x ∈ Sd , t ∈ R 2-weakly
isotropic stationary if E [Z(x, t)] is constant for all (x, t) ∈ Sd × R, and if the
covariance h is a spatially isotropic and temporally stationary function on (Sd ×
R)2 . The associated function ψ in (2.10) is called the covariance kernel or simply
kernel.
Remark 2.2. A Gaussian random field (GRF) which is 2-weakly isotropic stationary on Sd ×R, is in fact isotropic in the spatial variable and stationary in the
time variable (see [18]), hence, it has an invariant distribution under rotations
on the spatial variable, and under translations on the temporal variable.
Throughout the manuscript we work with zero-mean random fields, with no loss
of generality.
2.3. Kernel Covariance Functions on the Sphere cross Time
In his seminal paper, [23] characterized the class of continuous functions f :
[−1, 1] 7→ R such that f (cos θ) is positive definite over the product space
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
6
Sd × Sd , with θ defined in equation (2.8). Recently, [3] extended Schoenberg’s
characterization by considering the product space Sd × G, with G being a locally compact group. They defined the class P(Sd , G) of continuous functions
ψ : [−1, 1] × G 7→ C such that ψ(cos θ, u−1 · v) is positive definite on (Sd × G)2 .
In particular, the case G = R offers a characterization of spatio-temporal covariance functions of centred 2-weakly isotropic stationary random fields over
the sphere cross time.
Let P(R) denote the set of continuous and positive definite functions on R.
For d = 1, 2, . . . , we consider the class P(Sd , R) of continuous functions ψ :
[−1, 1]×R 7→ C such that the associated spatially isotropic temporally stationary
function h in (2.10) is positive definite. The following result, rephrased from [3],
allows to identify the class P(Sd , R) with the covariances functions of 2-weakly
isotropic stationary random fields on Sd × R.
Theorem 2.1. Let d ∈ N and let ψ : [−1, 1] × R 7→ C be a continuous mapping.
Then, ψ ∈ P(Sd , R) if and only if there exists a sequence {ϕj,d }j∈N ∈ P(R) with
∞
X
ϕj,d (0) < +∞ such that
j=0
ψ (cos θ, t) =
∞
X
ϕj,d (t)cj (d, cos θ) dim(Hjd ),
θ ∈ [0, π], t ∈ R,
(2.11)
j=0
where
ϕj,d (t) =
kσd k
kσd+1 k
Z
1
ψ(x, t)cj (d, x)(1 − x2 )d/2−1 dx,
(2.12)
−1
are called Schoenberg’s functions, dim(Hjd ) is given by (2.6), and cj (·, ·) by (2.1).
For (θ, t) ∈ [0, π] × R, the series in (2.11) is uniformly convergent.
Remark 2.3. Some comments are in order:
• For ε1 = (1, 0, . . . , 0) a unit vector in Rd+1 , we may consider the mapping
ψε1 : Sd × R 7→ C given by
ψε1 (y, t) = ψ(hε1 , yiRd+1 , t),
(y, t) ∈ Sd × R.
Then, arguments at page 13 of [3], show that
Z
ϕj,d (t) =
ψ(hε1 , yiRd+1 , t)cj (d, hε1 , yiRd+1 ) dσd (y)
Sd
=
hψε1 (·, t), cj (d, hε1 , ·iRd+1 )iL2 (Sd ) .
Hence, for each t ∈ R, the Schoenberg’s functions ϕj,d (t) given by (2.12)
may be understood as the orthogonal projection of ψε1 (·, t) onto Hjd .
• We note that, in comparison with the representation of covariance functions for 2-weakly isotropic random fields on the sphere, representation
(3.1) does not consider Schoenberg coefficients ϕj,d but Schoenberg functions ϕj,d (·), which will play a fundamental role subsequently.
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
7
3. Expansions for Isotropic Stationary GRFs and Kernel Covariance
Functions
According to Theorem 2.1, the kernel ψ of an isotropic stationary GRF Z on
Sd × R admits the following representation:
ψ(x, u) =
∞
X
ϕj,d (u)cj (d, x) dim(Hjd ),
x ∈ [−1, 1], u ∈ R,
(3.1)
j=0
where {ϕj,d }j∈N is a sequence of functions in P(R), such that the series is
uniformly convergent.
Expression (3.1) allows to consider different expansions for the kernel. Before
introducing these representations, we present the expansion for the random field
which motivates the simulation methodology.
3.1. Karhunen-Loève Expansions for Isotropic Stationary GRFs on
the Sphere cross Time
In [15], the following Karhunen-Loève representation for an isotropic stationary
GRF Z over S2 × R is proposed,
Z(x, t) =
j
∞ X
X
Xj,m (t)Yj,m (x),
(x, t) ∈ S2 × R,
(3.2)
j=0 m=−j
where {Xj,m (t)}j,m is a sequence of one-dimensional complex-valued mutually
independent stochastic processes. The set of all Xj,m (t) forms a denumerable
infinite dimensional stochastic process which completely defines the process on
the sphere, and Yj,m (x) are the elements of an orthonormal basis of Hj2 .
Representations (3.1) for the spatio-temporal covariance and (3.2) for the random field, allow us to introduce the following family of GRFs on Sd × R for any
d ∈ N.
Definition 3.1. Let Z be a random field on Sd × R defined, in the mean square
sense, as
dim(Hd )
Z(x, t) =
∞
X
Xj
j=0
Xj,m,d (t)Yj,m,d (x).
(3.3)
m=0
Here, for each j, m and d, {Xj,m,d (t), t ∈ R} is a complex-valued zero-mean
stationary Gaussian process such that
cov{Xj,m,d (t), Xj 0 ,m0 ,d (s)} := E{Xj,m,d (t) Xj 0 ,m0 ,d (s)}
= ϕj,d (t − s)δj,j 0 δm,m0 ,
where {ϕj,d }j∈N represents the Schoenberg’s functions associated to the mapping ψ in Equation (3.1).
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
8
Remark 3.1. Examples of random field satisfying Definition (3.1) are found in
the Appendix of [21].
Proposition 3.1. Let Z be a random field as in Definition 3.1. Then, Z is an
isotropic stationary GRF, with zero-mean and covariance function given by ψ
as in (3.1).
Proof. The proof follows straight by using the properties of the process Xj,m,d ,
and the addition formula in Equation (2.2).
For each j, m, d, the process {Xj,m,d (t), t ∈ R} has the following KarhunenLoève expansion (see Appendix 7.1),
Xj,m,d (t) =
∞
X
λj,k,d ζk (t),
t ∈ R,
(3.4)
k=0
where for each j, m, d ∈ N, {λj,k,m,d }k is a sequence of independent complexvalued random variables defined by
Z
λj,k,m,d :=
Xj,m,d (t)ζk (t)dµ(t).
R
Also, λj,k,m,d ∼ N (0, aj,k,d ), where {aj,k,d }k and {ζk }k are the eigenvalues and
eigenfunctions (respectively) of the integral operator Kϕ : L2 (R) 7→ L2 (R) associated to ϕj,d , defined by
Z
Kϕ (f )(t) :=
ϕj,d (t − s)δj,j 0 δm,m0 f (s)dµ(s), f ∈ L2 (R)
R
Remark 3.2. By (3.4), an alternative way to write (3.3) is:
dim(Hd )
Z(x, t) =
∞
∞
X
Xj X
j=0
m=1
λj,k,m,d ζk (t)Yj,m,d (x),
(x, t) ∈ Sd × R.
(3.5)
k=0
Expressions (3.3) or (3.5) represent a way to construct isotropic stationary GRFs
on the sphere cross time, and suggest a spectral simulation method. However,
it is not yet proved that any isotropic stationary GRF on the sphere cross time
can be written in this way.
3.2. Double Karhunen-Loève Expansion of Kernel Covariance
Functions
By stationarity of the process X and Karhunen-Loève theorem we have,
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
9
ϕj,d (u)δj,j 0 δm,m0 = cov{Xj,m,d (u), Xj 0 ,m0 ,d (0)}
∞
X
=
aj,k,d ζk (u)ζk (0), u ∈ R.
k=0
Hence, a slight abuse of notation allows to reformulate the last expression to
ϕj,d (u) =
∞
X
aj,k,d ζk (u),
u ∈ R.
k=0
Therefore, the kernel covariance function ψ in (3.1) also admits the expansion
ψ (x, t) =
∞
X
dim(Hjd )
j=0
∞
X
aj,k,d ζk (t)cj (d, x) ,
(x, t) ∈ [−1, 1] × R.
(3.6)
k=0
Following [17], we call the series {aj,k,d }j,k∈N ⊂ R the spatio-temporal angular
power spectrum.
P∞
Theorem 2.1 implies that { k=0 aj,k,d ζk (·)}j∈N is a sequence of continuous and
positive definite functions. Further,
∞ X
∞
X
aj,k,d ζk (0) < +∞.
j=0 k=0
3.3. Hermite Expansion of Kernel Covariance Functions
It is well known that any ϕ ∈ P(R) satisfies |ϕ(t)| ≤ ϕ(0) (see [22]). Therefore,
ϕ ∈ P(R), ensures that ϕ ∈ L2 (R, ν) for any finite measure ν, in particular, any
Gaussian measure.
Let ν be the standard Gaussian measure. As the Schoenberg’s functions {ϕj,d }j∈N
associated to ψ in Equation (3.1) belong to the class P(R) ⊂ L2 (R, ν), they can
be expanded in terms of normalized Hermite polynomials. For each j, d ∈ N
there exist constants {αj,k,d }k ∈ C such that
ϕj,d (u) =
∞
X
αj,k,d Hk (u),
u ∈ R.
k=0
The series converges in L2 (R, ν), and each Hk is a normalized Hermite polynomial of degree k given by
Hk (u) =
k
−u2
u2 d
(−1)k
2 ,
√
e2
e
duk
(k! 2π)1/2
k = 0, 1, 2, . . .
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
10
Consequently, the kernel ψ in (3.1) can be reformulated as
ψ(x, t) =
∞
X
j=0
dim(Hjd )
∞
X
αj,k,d Hk (t)cj (d, x),
(x, t) ∈ [−1, 1] × R.
(3.7)
k=0
We call the series {αk,j,d }j,k∈N ⊂ C the spatio-temporal Hermite power spectrum.
4. Regularity Properties
This section is devoted to study the behaviour of the kernel covariance functions
associated to an isotropic stationary GRF Z. It will be shown that the regularity
of such kernels is closely related to the decay of the Hermite power spectrum or
the angular power spectrum. Moreover, the latter characterizes also the (K −J)term truncation of a GRF Z as in Equation (3.5). We recall that we have
introduced two expansions for the kernel covariance function of an isotropic
stationary GRF Z:
i.- The Spatio-temporal power spectrum, by using a double Kaurhunen-Loève
expansion according to formula (3.6), valid for the kernel covariancefunction of any isotropic stationary GRF as in (3.5) over Sd × R, σd ⊗ µ with
µ the Lebesgue measure.
ii.- The Hermite power spectrum, by using Hermite polynomials according
to formula (3.7), valid for the kernel covariance
function of any isotropic
stationary GRF Z over Sd × R, σd ⊗ ν with ν the standard Gaussian
measure.
We recall that, for all x, y ∈ Sd and t, s ∈ R,
cov (Z(x, t), Z(y, s)) = ψ(cos θ(x, y), t − s) = ψ(hx, yiRd+1 , t − s).
Remark 4.1. Considering the relations between Gegenbauer and Legendre polynomials, in the case d = 2, the kernel (3.1) turns out to be
∞
X
ϕj,2 (t − s)(2j + 1)Pj (hx, yiR3 ) = ψ(hx, yiR3 , t − s),
x, y ∈ S2 , t, s ∈ R.
j=0
In what follows we present the regularity analysis of the kernel in terms of the
behaviour of the two proposed expansions (3.6) and (3.7). We address first the
relation with the Hermite power spectrum (3.7).
4.1. Regularity analysis for the Hermite expansion
In this part of the manuscript we consider the measure space (R, ν) with ν the
standard Gaussian measure.
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
11
For n, m ∈ N let the spaces H n (−1, 1) ⊂ L2 (−1, 1) and H m (R) ⊂ L2 (R) be
the standard Sobolev spaces. We extend the proposal in [17] and consider the
function spaces W n,m := W n,m ((−1, 1)×R) as the closure of H n (−1, 1)×H m (R)
with respect to the weighted norm k · kW n,m ((−1,1)×R) given by
kf k2W n,m :=
n X
m
X
|f |2W j,k ,
j=0 k=0
where for j, k ∈ N
|f |2W j,k
Z Z
2
1
∂k ∂j
f (x, t) (1 − x2 )d/2−1+j dx dν(t).
∂tk ∂xj
:=
−1
R
Note that (W n,n , n ∈ N) is a decreasing scale of separable Hilbert spaces, i.e.
L2 ((−1, 1) × R) ∼
= L2 (−1, 1) ⊗ L2 (R) = W 0,0 ⊃ W 1,1 ⊃ . . . ⊃ W n,n ⊃ . . .
We abuse of notation by writing L2 instead of L2 ((−1, 1) × R), and we consider
the canonical partial order relation in N2 : (n, m) ≤ (n0 , m0 ) if and only if [n ≤
n0 and m ≤ m0 ].
By Theorem 5.2 in [1], the norm of W n,n is equivalent to the first and the last
element of the sum, i.e.
kf k2W n,n ' kf k2L2 + |f |2W n,n , f ∈ W n,n .
We now derive another equivalent norm of W n,n in terms of summability of
the spectrum. We first observe that, as the normalized Hermite polynomials
{Hk }k∈N constitute a orthonormal basis of L2 (R), and that for any fixed r >
−1/2 the Gengenbauer polynomials {Cjr }j∈N a basis for L2 (−1, 1), it is apparent
that {Hk ·Cjr }j,k∈N is a basis for L2 ((−1, 1)×R) ∼
= L2 (−1, 1)⊗L2 (R). Therefore,
2
any f ∈ L can be expanded in the series
f (x, t) =
∞
X
bk,j Hk (t)Cjr (x),
(x, t) ∈ [−1, 1] × R,
k,j=0
with
Z Z
1
bk,j :=
R
Putting αj,k :=
f (x, t) =
2j+d−1
d−1 bk,j ,
∞
X
j=0
f (x, t)Hk (t)Cjr (x)dxdν(t),
k, j ∈ N0 .
−1
we get
dim(Hjd )
∞
X
αj,k Hk (t)cj (d, x),
(x, t) ∈ [−1, 1] × R,
k=0
that is, f can be written as a covariance kernel ψ of the type (3.7).
(4.1)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
12
This allows to tackle the problem in a different way. Instead of using spectral techniques, regularity of kernels might be shown through an isomorphism
between the spaces W n,m ((−1, 1) × R) and the weighted square summable bisequence spaces
(kj)m,n := `2 k m/2 · j (d−1)/2+n ; k, j ∈ N ,
(4.2)
where k m/2 · j (d−1)/2+n ; k, j ∈ N denotes the sequence of weights.
From now on, for the sake of simplicity, we only consider weighted Sobolev
spaces W n := W n,n , and the weighted square summable bi-sequence spaces
(kj)n := (kj)n,n , obtained as a special case of Equation (4.2) when n = m.
In order to extend the isomorphism to spaces W η with η being not an integer,
following [26] we now introduce the interpolation spaces W η := W η ((−1, 1)×R)
for n < η < n + 1, defined through:
W η := W n , W n+1 η−n,2 ,
equipped with the norm kf kW η given by
Z ∞
dζ
kf k2W η =
ζ −2(η−n) |J(r, f )|2 ,
ζ
0
where the functional J is defined by
J(r, f ) =
inf
f =v+w
v∈W n , w∈W n+1
(kvkW n + ζ · kwkW n+1 ) ,
ζ > 0.
The definition of the interpolation spaces (kj)η for η non-integer is carried out
in analogous way.
The interpolation property (see section 2.4.1 in [25]), implies that, if the spaces
W n and (kj)n are isomorphic for all n ∈ N, then they are isomorphic for all
η ∈ R+ .
Theorem 4.1. Let f ∈ L2 having expansion (4.1) and η ∈ R+ be given. Then,
f ∈ W η if and only if
∞
X
|bk,j |2 k η · j d−1+2η
< ∞,
k,j=0
i.e.
kf k2W η '
∞
X
k,j=0
η
is an equivalent norm in W .
|bk,j |2 k η · j d−1+2η
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
13
For ψ ∈ W n , n ∈ N, the equivalence is reduced to: αk,j k n/2 j (d−1)/2+n , k, j ∈ N
∂n ∂n
2 (d−2)/4+n/2
is in `2 (N2 ) if and only if ∂t
is in L2 ((−1, 1) × R),
n ∂xn ψ(x, t)(1−x )
where {αk,j }k,j∈N is the space-time Hermite power-spectrum. Rephrased,
∞
X
|αk,j |2 k n · j d−1+2n < +∞
k,j=n
if and only if
Z Z
R
1
−1
2
∂n ∂n
ψ(x, t) (1 − x2 )d/2−1+n dx dν(t) < +∞.
∂tn ∂xn
Proof. Assume first that the claim is already proved for η ∈ N, i.e., W n is
isomorphic to the weighted bi-sequence space (kj)n for all n ∈ N.
Fix n ∈ N, let n < η < n + 1 and set κ := η − n. By the interpolation theorem
of Stein-Weiss (see Theorem 5.4.1 in [4]), the weights of (kj)η are given by
1−κ n+1 d−1+2(n+1) κ
= k η j d−1+2η .
k n j d−1+2n
· k
j
Now, we prove the isomorphism between W n and (kj)n for n ∈ N, which is
equivalent to prove the second formulation of the theorem. We have that,
2
1
∂n ∂n
ψ(x, t) (1 − x2 )d/2−1+n dx dν(t)
n
n
R −1 ∂t ∂x
Z Z 1 X
∞
∞
2
X
∂n
∂n
αk,j n Hk (t) n cj (d, x) (1 − x2 )d/2−1+n dx dν(t)
=
dim(Hjd )
∂t
∂x
R −1 j=0
k=0
Z
∞
X
∂n
∂n
αk,j αk0 ,j 0
=
H
(t)
Hk0 (t) dν(t)
k
n
∂tn
R ∂t
k,k0 ,j,j 0 =0
Z 1 n
∂
∂n
·
cj (d, x) dim(Hjd ) n cj 0 (d, x) dim(Hjd0 )(1 − x2 )d/2−1+n dx
n
∂x
−1 ∂x
∞
X
(4.3)
=
αk,j αk0 ,j 0 · Ik,k0 · Iej,j 0 ,
Z Z
k,k0 ,j,j 0 =0
where
Z
Ik,k0 =
R
∂n
∂n
H
(t)
Hk0 (t) dν(t)
k
∂tn
∂tn
and
Z
1
Iej,j 0 =
−1
n
∂n
d ∂
c
(d,
x)
dim(H
)
cj 0 (d, x) dim(Hjd0 )(1 − x2 )d/2−1+n dx.
j
j
∂xn
∂xn
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
14
Standard properties of normalized Hermite polynomials show that Ik,k0 = 0 if
k < n, and
k!
δk,k0 ; k ≥ n.
Ik,k0 =
(k − n)!
On one hand, Stirling inequality
√
2π k k+1/2 e−k ≤ k! ≤ k k+1/2 e−k+1 ,
implies that
√
2π e−(n+1)
k k+1/2
k!
e−n+1
k k+1/2
√
≤
≤
.
(k − n)!
(k − n)k−n+1/2
2π (k − n)k−n+1/2
On the other hand,
kk
kn
=
,
k−n
(k − n)
(1 − n/k)k(1−n/k)
where for k ≥ n,
1
−−−−→ e−n .
(1 − n/k)k−n k→∞
Hence, for k ≥ n, there exists constants c1 (n) and c2 (n) such that
c1 (n)k n ≤
k!
≤ c2 (n)k n .
(k − n)!
Therefore,
Ik,k0 ' k n δk,k0 , k ≥ n.
(4.4)
For Ij,j 0 (x), expressions (2.4), (2.5), (2.6) and (2.7) allow to conclude that
Ij,j 0 (x) = 0 if j < n, and
Ij,j 0 (x) ' j d−1+2n δj,j 0 , j ≥ n.
Therefore, from (4.3), (4.4) and (4.5) we deduce that
Z Z
R
'
1
−1
∞
X
2
∂n ∂n
ψ(x, t) (1 − x2 )d/2−1+n dx dν(t)
∂tn ∂xn
|αk,j |2 k n j d−1+2n ,
k,j=n
which concludes the proof.
(4.5)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
15
4.2. Regularity analysis for the double Karhunen-Loève expansion
With the exception of minor details on the definitions of the spaces, this part
of the manuscript follows similarly to Section 4.1. On the other hand, we now
consider the measure space (R, µ) with µ the Lebesgue measure.
For n ∈ N consider the function spaces VTn := VTn ((−1, 1) × R) as the closure of
H n (−1, 1) × L2 (R) with respect to the weighted norms k · kVTn given by
kf k2V n ((−1,1)×R) :=
n
X
T
|f |2V j ((−1,1)×R) ,
j=0
T
where for j ∈ N
|f |2V j :=
T
(VTn )n∈N
Z Z
R
2
1
∂j
f (x, t) (1 − x2 )d/2−1+j dx dµ(t).
∂xj
−1
is a decreasing sequence of separable Hilbert spaces, and
kf k2VTn ' kf k2L2 + |f |2VTn ,
f ∈ VTn .
Now we look at the weighted square summable bi-sequence spaces
(kj)T,n := `2 j (d−1)/2+n ; k, j ∈ N .
As in Section 4.1, we consider the interpolation spaces VTη for n < η < n+1. The
proof of next result follows exactly the same lines as the Theorem 4.1, hence it
is omitted.
Theorem 4.2. Let f ∈ L2 and η ∈ R+ be given. Then, f ∈ VTη if and only if
∞
X
b2k,j j d−1+2η < ∞,
k,j=0
i.e.
kf k2V η
T
'
∞
X
b2k,j j d−1+2η
k,j=0
is an equivalent norm in
VTη .
For ψ ∈ VTn , n ∈ N, the equivalence is reduced to: ak,j j (d−1)/2+n , k, j ∈ N
∂n
2 (d−2)/4+n/2
is in `2 (N2 ) if and only if ∂x
is in L2 ((−1, 1) × R),
n ψ(x, t)(1 − x )
where {ak,j }k,j∈N is the space-time angular power-spectrum. Rephrased,
∞ X
∞
X
j=n k=0
a2k,j j d−1+2n < +∞
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
16
if and only if
Z Z
R
1
−1
2
∂n
ψ(x, t) (1 − x2 )d/2−1+n dx dν(t) < +∞.
∂xn
Remark 4.2. By taking into account the normalizing constants, all the previous
results encompasses the results in Section 3 of [17].
5. Spatio-Temporal Spectral Simulation
We now study a spectral simulation method for random fields on S2 × [0, T ],
where T denotes the time horizon. For a neater exposition, along this section,
we omit the subscripts associated to the spatial dimension d = 2.
We must first introduce some notation. Let j ∈ N0 and m ∈ {0, . . . , j}. For
x ∈ [−1, 1], the associated Legendre polynomials Pj,m are defined through
dm
(Pj (x)),
dxm
m+1/2
= (2m − 1)!!(−1)m (1 − x2 )m/2 Cj−m (x).
Pj,m (x) = (−1)m (1 − x2 )m/2
The spherical harmonic basis functions, Yj,m : S2 7→ C, are defined by
s
2j + 1 (j − m)!
Yj,m (x) =
Pj,m (cos β1 ) exp(imβ2 ), j ∈ N0 , m ∈ {0, . . . , j}
4π (j + m)!
Yj,m (x) = (−1)m Yj,−m (x),
j ∈ N, m ∈ {−j, . . . , −1}.
where (β1 , β2 ) ∈ [0, π] × [0, 2π) represents the spherical coordinates of x ∈ S2 .
On the other hand, let {Xj,m (t), j ∈ N0 , m ∈ {−j, . . . , j}} be a collection of
stochastic processes. Thus, we consider the space-time random field
Z(x, t) =
j
∞ X
X
Xj,m (t)Yj,m (x),
x ∈ S2 , t ∈ [0, T ].
(5.1)
j=0 m=−j
In order to obtain a real-valued field, we must impose some conditions on the
stochastic processes {Xj,m (t), j ∈ N0 , m ∈ {0, . . . , j}}. Throughout, we assume that {Xj,m (t), j ∈ N0 , m ∈ {0, . . . , j}} are mutually independent, with
Im(Xj,0 (t)) being identically equal to zero, and for j ∈ N and m ∈ {−j, . . . , −1},
Xj,m (t) = (−1)m Xj,−m (t).
(5.2)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
17
Note that, using standard algebra of complex numbers, coupled with condition
(5.2), Equation (5.1) can be written as
Z(x, t) =
∞
X
Xj,0 (t)Yj,0 (x)
j=0
+
j
X
!
Xj,m (t)Yj,m (x) + (−1) Xj,m (t)(−1) Yj,m (x)
m
m
m=1
=
∞
X
(5.3)
Xj,0 (t)Yj,0 (x)
j=0
+2
j
X
!
Re(Xj,m (t))Re(Yj,m (x)) − Im(Xj,m (t))Im(Yj,m (x))
.
m=1
We consider Re(Xj,m (t)) and Im(Xj,m (t)) as independent processes with the
following Fourier expansions
Re(Xj,m (t))
Im(Xj,m (t))
=
=
A1j,0,m
A2j,0,m
+
+
∞
X
k=1
∞
X
A1j,k,m
A2j,k,m
cos
cos
k=1
πkt
2T
πkt
2T
+
1
Bj,k,m
+
2
Bj,k,m
sin
sin
πkt
2T
πkt
2T
,
.
q
Here, {Aqj,k,m } and {Bj,k,m
}, for q = 1, 2, are sequences of independent centred
real-valued Gaussian random variables, such that
1
var(A1j,k,0 ) = var(Bj,k,0
) = aj,k
and
q
var(Aqj,k,m ) = var(Bj,k,m
) = aj,k /2,
for m 6= 0,
where {aj,k }k,j∈N is a summable bi-sequence of non-negative coefficients. A
direct calculation shows that the covariance function of Z(x, t) is spatially
isotropic and temporally stationary. More precisely, we have that
cov{Z(x, t), Z(y, s)} =
∞ X
∞
X
(2j + 1)
j=0 k=0
4π
aj,k cos
πk(t − s)
2T
cj (hx, yiR3 ), x, y ∈ S2 .
Finally, given two positive integers J and K, we truncate expression (5.3) in the
index j and k, respectively. Thus, we simulate a space-time Gaussian random
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
18
field on S2 × [0, T ] using the explicit approximation
b t) =
Z(x,
J
X
A1j,0,0 Pej,0 (cos β1 )
j=0
j
X
+2
1
2
e
Pj,m (cos β1 ) Aj,0,m cos(mβ2 ) − Aj,0,m sin(mβ2 )
m=1
K
X
πkt
πkt
1
(5.4)
+ Bj,k,0
sin
2T
2T
k=1
j
K
X
X
πkt
πkt
1
+ Bj,k,m
sin
+2
Pej,m (cos β1 ) cos(mβ2 )
A1j,k,m cos
2T
2T
m=1
k=1
!
j
K
X
X
πkt
πkt
2
2
Aj,k,m cos
,
−2
Pej,m (cos β1 ) sin(mβ2 )
+ Bj,k,m sin
2T
2T
m=1
+Pej,0 (cos β1 )
A1j,k,0 cos
k=1
where Pej,m (·) =
p
(2j + 1)(j − m)!/(4π(j + m))!Pj,m (·).
We assess the L2 (Ω × S2 × [0, T ]) error associated to the truncated expansion
b t) given in Equation (5.4), in terms of the positive integers J and K. We
Z(x,
follow the scheme used by [17] in the spatial context and extend their result to
the space-time case. Next, we state the main result of this section.
Theorem 5.1. Let ν1 , ν2 ≥ 2. Suppose that there exist positive constants Ci ,
for i = 1, 2, 3, and positive integers j0 and k0 , such that aj,0 ≤ C1 j −ν1 , a0,k ≤
C2 k −ν2 and aj,k ≤ C3 j −ν1 k −ν2 , for all j ≥ j0 and k ≥ k0 . Then, the following
inequality holds
b t)k2 2
e −(ν1 −2) + C
e2 JK −(ν2 −1) ,
kZ(x, t) − Z(x,
L (Ω×S2 ×[0,T ]) ≤ C1 J
(5.5)
e1 and C
e2 .
for some positive constants C
b t) in Equation (5.4) into two independent
Proof. We decompose Z(x, t) − Z(x,
terms,
b t) = T1 (x) + T2 (x, t),
Z(x, t) − Z(x,
where
X
∞
T1 (x) =
A1j,0,0 Pej,0 (cos β1 )
j=J+1
+2
j
X
Pej,m (cos β1 ) A1j,0,m cos(mβ2 ) − A2j,0,m sin(mβ2 )
m=1
and
T2 (x, t) =
∞ X
∞
X
j=J+1 k=1
∆j,k (x, t) +
J
∞
X
X
j=0 k=K+1
∆j,k (x, t),
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
19
with ∆j,k (x, t) defined as
πkt
πkt
1
1
e
∆j,k (x, t) = Pj,0 (cos β1 ) Aj,k,0 cos
+ Bj,k,0 sin
2T
2T
j
X
πkt
πkt
1
1
e
+2
Pj,m (cos β1 ) cos(mβ2 ) Aj,k,m cos
+ Bj,k,m sin
2T
2T
m=1
j
X
πkt
πkt
2
−2
Pej,m (cos β1 ) sin(mβ2 ) A2j,k,m cos
+ Bj,k,m
sin
.
2T
2T
m=1
For the second term, we have used the identity
∞ X
∞
X
j=0 k=1
ξj,k −
J X
K
X
∞ X
∞
X
ξj,k =
j=0 k=1
ξj,k +
j=J+1 k=1
J
∞
X
X
ξj,k ,
j=0 k=K+1
which is satisfied for any summable bi-sequence {ξj,k }j,k∈N .
The independence of T1 (x) and T2 (x, t) implies that
2
2
b t)k2 2
kZ(x, t) − Z(x,
L (Ω×S2 ×[0,T ]) = kT1 (x)kL2 (Ω×S2 ) + kT2 (x, t)kL2 (Ω×S2 ×[0,T ]) .
In [17] it is shown that there exists a positive constant L1 , depending on C1 ,
such that
kT1 (x)k2L2 (Ω×S2 ) ≤ L1 J −(ν1 −2) .
On the other hand, since
k cos(πkt/(2T ))k2L2 ([0,T ]) + k sin(πkt/(2T ))k2L2 ([0,T ]) = T,
kPej,0 (cos β1 )k2L2 (S2 ) = 1
and
kPej,m (cos β1 ) cos(mβ2 )k2L2 (S2 ) + kPej,m (sin β1 ) cos(mβ2 )k2L2 (S2 ) = 1,
we have that k∆j,k (x, t)k2L2 (Ω×S2 ×[0,T ]) = T (2j + 1)aj,k . Therefore,
X
∞
∞
X
2
kT2 (x, t)kL2 (Ω×S2 ×[0,T ]) ≤ T
(2j + 1)
aj,k
j=J+1
k=1
∞
J
X
X
+
a0,k +
(2j + 1)aj,k
j=1
k=K+1
X
X
∞
∞
≤T
C3
k −ν2
(2j + 1)j −ν1
j=J+1
k=1
X
J
∞
X
+ C2 + C3
(2j + 1)j −ν1
k −ν2
j=1
≤ L2 J
−(ν1 −2)
+ L3 JK
k=K+1
−(ν2 −1)
,
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
20
where L2 and L3 are positive constants depending on C2 and C3 . In particular,
the last inequality follows from the integral bounds of the corresponding series
(see [17]):
∞
X
−ν1
≤
k −ν2
≤
(2j + 1)j
j=J+1
∞
X
k=K+1
2
1
+
ν1 − 2 ν1 − 1
J −(ν1 −2)
1
K −(ν2 −1) .
ν2 − 1
The proof is completed.
Simple examples can be generated from the following space-time angular power
spectrum
1
aj,k =
,
(5.6)
1 + (1 + j)ν1 (1 + k)ν2
with νi > 2, for i = 1, 2. We illustrate space-time realizations on S2 × {1, 2},
over 24000 spatial locations, with coefficients (5.6), in two cases:
(a) ν1 = 3 and ν2 = 5, and
(b) ν1 = ν2 = 5.
Figures 1 and 2 show the corresponding realizations for Scenarios (a) and (b),
respectively. For each case, we truncate the series using K = J = 50. Note that
the parameter ν1 is the responsible of the spatial scale and smoothness of the
realization. In [17], some realizations are illustrated using a similar spectrum,
in a merely spatial context.
We now compare the empirical and theoretical convergence rates for the cases
(a) and (b) described above. In our experiment, we consider J = K and study
the (Log) error in terms of (Log) J, taking as the exact solution J = 50. Note
that, under this choice, the bound (5.5) implies that the order of convergence is
min{(ν1 − 2)/2, (ν2 − 2)/2}. Following [17], instead of calculating the L2 -error,
we take the maximum error over all the points on the space-time grid. The
empirical errors are calculated on the basis of 100 independent samples. Our
studies reflect the theoretical results (see Figure 3).
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
21
Figure 1. Space-time realization on S2 × {1, 2}, with spectrum (5.6), with ν1 = 3 and ν2 = 5.
Figure 2. Space-time realization on S2 × {1, 2}, with spectrum (5.6), with ν1 = ν2 = 5.
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
22
Figure 3. Empirical versus theoretical (Log) L2 -error of the simulation method, in terms of
(Log) J, with K = J. We consider the spectrum (5.6) in two cases: (a) ν1 = 3 and ν2 = 5,
and (b) ν1 = ν2 = 5.
6. Conclusions and discussion
The present work has provided a deep look at the regularity properties of Gaussian fields evolving temporally over spheres. We hope this effort will put the
basis for facing important challenges related to space-time processes. There are
in fact many open problems related to mathematical modeling as well as to
statistical inference and to optimal prediction. A list of open problems is included in the recent survey [20]. Amongst them, our paper is certainly related
to Problem 1, that is to the construction of non-stationary processes on spheres
cross time. Our work could also put the basis to solve Problem 2, related to
the construction of multivariate space-time processes. It might be interesting to
extend the study of regularity properties to the vector valued case. This would
imply the use of a pretty different machinery. Problem 10 is closely related to
our approach, because regularity properties are crucial to study Gaussian fields
under infill asymptotics.
On the other hand, a question arise naturally: is it possible to make inference
with a representation like (3.5) ? Or with its respective spectral decomposition?
The answer is, a priori, no. Establishing a clear relation between the parameters
of any random field and its spectrum is not an obvious task, in fact, up today
the only familiar stochastic process with known spectrum is Brownian motion,
and his closer generalization, the fractional Brownian motion, doesn’t have yet
a known spectrum. However, under relatively weak hypotheses, the covariance
kernel of a GRF turns out to be a Mercer kernel. This opens an alternative to the
negative answer previously mentioned, by considering the eigenvalue problem
associated to the integral operator induced by the kernel.
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
23
7. Appendix
7.1. Karhunen-Loève Theorem
Recent results on functional analysis (see [11] and [12]) allow to construct
Mercer’s kernels in more general contexts. A proper interpretation of these results allows to generalize the classic Karhunen-Loève theorem in a very neat
way. We first introduce the framework and basics notations from [11].
Let S be a nonempty set and K a positive definite kernel on S, i.e., a function
K : S × S −→ C satisfying the inequality
n
X
ci cj K(xi , xj ) ≥ 0,
i,j
whenever n ≥ 1, {x1 , x2 , . . . , xn } is a subset of S and {c1 , c2 , . . . , cn } is a subset
of C. The set of all positive definite kernels over S is denoted by PD(S).
If S is endowed with a measure υ, denote by L2 PD(S, υ) the class of kernels
such that the associated integral operator
Z
K(f )(x) :=
K(x, y)f (y)dυ(y), f ∈ L2 (S, υ), x ∈ S,
S
is positive, that is, when the following conditions holds
Z Z
K(x, y)f (y)dυ(y) f (x)dυ(x) ≥ 0, f ∈ L2 (S, υ),
S
S
i.e.,
hK(f ), f iL2 (S,υ) ≥ 0,
f ∈ L2 (S, υ).
Finally we define what is a Mercer’s kernel according to [12]: A continuous kernel
K on S is a Mercer’s kernel when it possesses a series representation of the form
K(x, y) =
∞
X
aj (K)ζj (x)ζj (y),
x, y ∈ S,
j=1
where {ζj }j∈N is an L2 (S, υ)-orthonormal basis of continuous functions on S,
{aj (K)}j∈N decreases to 0 and the series converges uniformly and absolutely on
compact subsets of S × S.
For the rest of this manuscript we will consider S to be a topological space
endowed with a strictly positive measure υ, that is, a complete Borel measure
on S for which two properties hold: every open nonempty subset of S has positive
measure and every x ∈ S belongs to an open subset of S having finite measure.
Besides, as in Section 2, (Ω, F, P) denotes a complete probability space.
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
24
Theorem 7.1. Let (Xs )s∈S be a complex-valued centred stochastic process with
continuous covariance function and such that
Z
E(|Xs |2 )dυ(s) < +∞.
(7.1)
S
Then, the kernel K associated with the covariance function of X is a Mercer’s
Kernel. Therefore, X admits a Karhunen-Loève expansion
X=
+∞
X
λj ζj
(7.2)
j=1
where {ζj }j∈N is an orthonormal basis of L2 (S, υ),
Z
Xs ζj (s) dυ(s),
λj =
S
with E[λj ] = 0, and there exists a sequence {aj }j∈N of non-negative real numbers
such that E[λj λk ] = δjk aj and Var[λj ] = aj .
The series expansion (7.2) converges in L2 (Ω × S; C), i.e.,
2
Z
J
X
lim E Xs −
λj ζj (s) dυ(s) = 0.
J→+∞
S
j=1
The series expansion (7.2) converges in L2 (Ω; C) for all s ∈ S, i.e., for all s ∈ S
2
J
X
lim E Xs −
λj ζj (s) = 0.
J→+∞
j=1
The convergence of the series expansion (7.2) is absolute and uniform on compact subsets of S in the mean-square sense.
Proof. Let K be the kernel associated to the covariance of the stochastic process
X, i.e.
K :S×S
−→
(t, s) −→
C
K(t, s) = E(Xt Xs ),
and let K be it’s associated integral operator. From hypothesis (7.1) it is direct
to see that the mapping κ, such that s ∈ S 7→ C 3 κ(s) := K(s, s), belongs to
L1 (S, υ). Since K is positive definite in the usual sense, K(s, t) = K(t, s) and
the matrix
K(s, s) K(s, t)
K(s, t) K(t, t)
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
25
is positive definite, hence its determinant K(s, s)K(t, t) − K(s, t)K(s, t) is nonnegative. Thus
Z
Z
|K(s, t)|2 d(υ ⊗ υ)(s, t) ≤
K(s, s)K(t, t)d(υ ⊗ υ)(s, t)
S×S
S×S
Z
=
2
< ∞,
κ(s)dυ(s)
S
that is, K ∈ L2 (S × S, υ ⊗ υ).
Now, consider f, g ∈ L2 (S, υ) and the classic tensor product of functions, i.e.,
(f ⊗ g)(s, t) = f (s)g(t),
s, t ∈ S,
with inner product given by
hf1 ⊗ f2 , g1 ⊗ g2 iL2 (S×S,υ×υ) = hf1 , g1 iL2 (S,υ) hf2 , g2 iL2 (S,υ) .
Apparently, f ⊗ g ∈ L2 (S × S, υ ⊗ υ), and thus, by Cauchy-Schwarz inequality
we obtain K · f ⊗ g ∈ L1 (S × S, υ ⊗ υ). This last condition allows to use Fubini’s
theorem, so for f ∈ L2 (S, υ)
Z Z
hK(f ), f i2 =
K(s, t)f (t) dυ(t) f (s) dυ(s)
ZS Z S
=
K(s, t)f (t)f (s) dυ(t)d υ(s)
S S
Z Z
=
E(Xs Xt )f (t)f (s) dυ(t) dυ(s)
S S
Z Z
= E
Xs f (s) Xt f (t) dυ(t) dυ(s)
S
S
=
E hX, f i2 hf, Xi2
=
E hX, f i2 hX, f i2
=
E| hX, f i2 |2 ≥ 0,
hence, K ∈ L2 PD(S, υ).
In conclusion, the kernel K is continuous and L2 (S, υ)-positive definite on S,
and the mapping κ belongs to L1 (S, υ). Then, by theorem 3.1 in [12] K is a
Mercer’s kernel.
The rest of the proof concerns the Karhunen-Loève expansion of the process X
and it follows well-known arguments that we reproduce for the convenience of the
reader. We have that K has a L2 (S × S, υ ⊗ υ)-convergent series representation
in the form
+∞
X
K(t, s) =
aj (K)ζj (t)ζj (s), t, s ∈ S,
j=1
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
26
{aj (K)}j∈N decreases to 0, {ζj }j∈N is a L2 (S, υ)-orthonormal basis. The convergence of the series is absolute and uniform on compact subsets of S × S.
From condition (7.1) there exists a set Ω0 ⊆ Ω with P(Ω0 ) = 1 such that for
all ω ∈ Ω0 , the mapping s ∈ S 7→ Xs (ω) is in L2 (S, υ). Define the random
coefficients
Z
λj =
Xt ζj (t) dυ(t) = hX, ζj iL2 (S,υ) .
S
E(λ2j )
2
Note that
= hK(ζj ), ζj i2 , hence Cauchy-Scharwz inequality guarantees
that λj ∈ L (Ω) for all j. Also, Fubini’s theorem allows to see that E[λj ] = 0,
E[λj λk ] = δjk aj and Var[λj ] = aj .
Now, for any fixed ω ∈ Ω0 it is clear that
Z
Xs (ω) −
fJ (ω) :=
S
J
X
2
λj (ω)ζj dυ(s) −−−−−→ 0.
J→+∞
j=1
By orthogonality of the (ζj ) we observe that
Z
0 ≤ fJ (ω)
Xs (ω) −
:=
S
Z
=
J
X
2
λj (ω)ζj dυ(s)
j=1
|Xs (ω)|2 dυ(s) −
S
Z
≤
J
X
|λj (ω)|2
j=1
|Xs (ω)|2 dυ(s) −
S
J−1
X
|λj (ω)|2 := fJ−1 (ω),
j=1
therefore,
Z
|fJ (ω)| ≤ |f0 (ω)| :=
|Xs (ω)|2 dυ(s),
J ∈ N.
S
By condition (7.1) is clear that E(|f0 |) < ∞, hence, the dominated convergence
theorem allows us to conclude that
Z
J
2
X
E(|fJ |) = E
Xs (ω) −
λj (ω)ζj dυ(s) −−−−−→ 0.
S
j=1
J→+∞
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
27
Now, fix s ∈ S. Again by Fubini’s theorem we observe that,
E Xt −
n
X
2
λj ζj (t)
j=1
n
X
= E|Xt |2 −
n
n
X
X
ζj (t)E Xt λj −
ζj (t)E Xt λj +
ζj (t)ζk (t)E λj λk
j=1
= E|Xt |2 −
n
X
j=1
ζj (t)E Xt
n
X
Z
n
X
Xs ζj (s) dυ(s) +
ζj (t)ζk (t)δkj ak
ζj (t)E Xt
S
j=1
= E|Xt |2 −
n
X
Z
−
S
Z
ζj (t)
E Xt Xs ζj (s) dυ(s) +
S
j=1
=
K(t, t) −
n
X
−
ζj (t)
K(s, t)ζj (s) dυ(s) +
S
= K(t, t) −
n
X
K(t, s)ζj (s) dυ(s)
S
Z
= K(t, t) −
K(t, s)ζj (s) dυ(s) +
S
n
X
= K(t, t) − 2
n
X
|ζj (t)|2 aj
j=1
ζj (t)Kζj (t) −
j=1
n
X
= K(t, t) −
|ζj (t)|2 aj
j=1
ζj (t)
ζj (t)
j=1
n
X
Z
j=1
−
K(t, s)ζj (s) dυ(s)
S
ζj (t)
n
X
|ζj (t)|2 aj
j=1
Z
j=1
n
X
Z
j=1
n
X
j,k=1
E Xt Xs ζj (s) dυ(s)
ζj (t)
j=1
n
X
Xs ζj (s) dυ(s)
S
j=1
−
j,k=1
Z
n
X
ζj (t)Kζj (t) +
j=1
|ζj (t)|2 aj +
j=1
n
X
|ζj (t)|2 aj
j=1
n
X
n
X
|ζj (t)|2 aj
j=1
|ζj (t)|2 aj
j=1
−−−−−→ 0.
n→+∞
Thus, the proof is concluded.
Remark 7.1. Karhunen-Loève expansion (or Karhunen-Loève theorem), usually
require extra hypothesis, like compactness of the associated space S or some kind
of invariance of the field. In that line the Stochastic Peter-Weyl theorem (theorem 5.5 introduced in [18]) may be understood as a Karhunen-Loève theorem
for 2-weakly isotropic fields over G a topological compact group with associated
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
28
Haar measure of unit mass. Theorem 7.1 only require condition (7.1) and the
continuity of the covariance function.
Acknowledgments
We are indebted to the Editor and to three Referees, whose thorough reviews
allowed for a considerably improved version of the manuscript.
References
[1] Adams, R. A. and Fournier, J. F. (2003), Sobolev Spaces. Second edition. Pure and
Applied Mathematics (Amsterdam). MR2424078
[2] Bateman, H. and Erdélyi, A. (1953), Higher Transcendental Functions. Vol. I - II.
McGraw-Hill, New York.
[3] Berg, C. and Porcu, E. (2016), From Schoenberg coefficients to Schoenberg functions. Constr. Approx., 1–25.
[4] Bergh, J. and Löfström J. (1976), Interpolation Spaces. An introduction.
Grundlehren der Mathematischen Wissenschaften, No. 223. Springer-Verlag, BerlinNew York. MR0482275
[5] Bonan, S. S. and Clark, D. S. (1990), Estimates of the Hermite and the Freud
polynomials. J. Approx. Theory, 63, 210–224. MR1079851
[6] Christakos, G. (2005), Random Field Models in Earth Sciences. Elsevier
[7] Christakos, G. and Hristopulos, T. (1998), Spatiotemporal Environmental Health
Modelling: a Tractatus Stochasticus. (With a foreword by William L. Roper). Kluwer
Academic Publishers, Boston, MA. MR1648348
[8] Christensen, J. P. R. (1970), On some measures analogous to Haar measure. Math.
Scand., 26, 103–106. MR0260979
[9] Dai, F. and Xu, Y. (2013), Approximation Theory and Harmonic Analysis on
Spheres and Balls. Springer Monographs in Mathematics. Springer, New York.
MR3060033
[10] Dimitrakopoulos, R. (1994) Geostatistics for the Next Century. Springer Netherlands.
[11] Ferreira, J. C. and Menegatto, V. A. (2009), Eigenvalues of integral operators
defined by smooth positive definite kernels. Integral Equations Operator Theory,
64, 61–81. MR2501172
[12] Ferreira, J. C. and Menegatto, V. A. (2013), Positive definiteness, reproducing
kernel Hilbert spaces and beyond. Ann. Funct. Anal. 4, 64–88. MR3004212
[13] Gneiting, T. (2002), Nonseparable, stationary covariance functions for SpaceTime data. J. Amer. Statist. Assoc., 97, 590–600. MR1941475
[14] Gneiting, T. (2013), Strictly and non-strictly positive definite functions on
spheres. Bernoulli, 19, 1327–1349. MR3102554
[15] Jones, R. H. (1963), Stochastic processes on a sphere. Ann. Math. Statist., 34,
213–218. MR3102554
[16] Kozachenko, Yu. V. and Kozachenko, L. F. (2001), Modelling Gaussian isotropic
random fields on a sphere. J. Math. Sci., 2, 3751–3757. MR1874654
J. Clarke et al./Regularity and simulations of GRFs on Spheres cross Time.
29
[17] Lang, A. and Schwab, C. (2015), Isotropic Gaussian random fields on the sphere:
regularity, fast simulation and stochastic partial differential equations. Ann. Appl.
Probab., 25, 3047–3094. MR3404631
[18] Marinucci, D. and Peccati, G. (2011), Random Fields on the Sphere. Representation, Limit Theorems and Cosmological Applications. London Mathematical Society
Lecture Note Series, 389. Cambridge University Press, Cambridge. MR2840154
[19] Morimoto, Mitsuo (1998), Analytic Functionals on the Sphere. Translations of
Mathematical Monographs, 178. American Mathematical Society, Providence, RI.
MR1641900,
[20] Porcu, E. Alegrı́a, A. and Furrer R. (2017), Modeling temporally evolving and
spatially globally dependent data. arXiv:1706.09233v1
[21] Porcu, E. Bevilacqua, M. and Genton, M. (2015), Spatio-Temporal covariance
and cross-covariance functions of the great circle distance on a sphere. J. Amer.
Statist. Assoc., 111, 888–898.
[22] Sasvári, Z. (1994), Positive Definite and Definitizable Functions. Mathematical
Topics, 2. Akademie Verlag, Berlin. MR1270018
[23] Schoenberg, I. J. (1942), Positive definite functions on spheres. Duke Math. J., 9,
96–108. MR0005922
[24] Stein, M. (2005), Space-time covariance functions. J. Amer. Statist. Assoc., 100,
310–321. MR2156840
[25] Triebel, H. (1983), Theory of Function Spaces. Monographs in Mathematics, 78.
Birkhäuser Verlag, Basel. MR0781540
[26] Triebel, H. (1999), Interpolation Theory, Function Spaces, Differential Operators.
Wiley.
[27] Zastavnyi, V. P. and Porcu, E. (2011), Characterization theorems for the Gneiting
class of space-time covariances. Bernoulli, 17, 456–465. MR2797999
| 10math.ST
|
arXiv:1701.03308v2 [cs.DS] 6 Sep 2017
Sampling and Reconstruction Using Bloom Filters
Neha Sengupta
IIT-Delhi
[email protected]
Amitabha Bagchi
IIT-Delhi
[email protected]
Srikanta Bedathur
IBM-IRL
[email protected]
Maya Ramanath
IIT-Delhi
[email protected]
Abstract
In this paper, we address the problem of sampling from a set and reconstructing a set stored
as a Bloom filter. To the best of our knowledge our work is the first to address this question.
We introduce a novel hierarchical data structure called BloomSampleTree that helps us design
efficient algorithms to extract an almost uniform sample from the set stored in a Bloom filter and
also allows us to reconstruct the set efficiently. In the case where the hash functions used in the
Bloom filter implementation are partially invertible, in the sense that it is easy to calculate the
set of elements that map to a particular hash value, we propose a second, more space-efficient
method called HashInvert for the reconstruction. We study the properties of these two methods
both analytically as well as experimentally. We provide bounds on run times for both methods
and sample quality for the BloomSampleTree based algorithm, and show through an extensive
experimental evaluation that our methods are efficient and effective.
1
Introduction
Bloom filters, introduced by Bloom in the 1970’s [1], are space-efficient structures for the setmembership problem. They have found numerous applications in a diverse array of settings because
of the tremendous advantages they offer in terms of space. Broder and Mitzenmacher surveyed a
host of these applications in 2003 [2], and, since then, the usage of Bloom filters has grown and
diversified. Typically, these applications rely on the set-membership query being answered correctly
with good probability, and are able to deal with the drawback that with some probability a false
positive will occur.
However, one fundamental question has not yet been addressed: How do we sample an element
from a set stored in a Bloom filter? A related question – How do we retrieve the set stored in the
Bloom filter? – has also not been addressed. We believe addressing these two problems will open
up the possibility of using Bloom filters in applications that need to store, retrieve and/or sample
from a large number of sets. For example, storing and subsequently sampling from a large number
of dynamic, online communities that form on social networks such as Twitter, Flickr, etc. ([3], [4],
[5]), that could help advertisers determine where to target their products. Or storing and retrieving
all call records associated with specific locations in crime-related investigations [6].
We note that other compact structures, such as sketches, have been used as compact storage
structures from which samples can later be obtained [7, 8, 9]. However, a limitation of this approach
is that the sketches that are proposed to be created are specifically for the problem of sampling
and tend to be output sensitive in their design (and do not support reconstruction). Our work, on
the other hand, shows how to draw samples as well as reconstruct sets from a widely-used generic
synopsis structure, the Bloom filter, that is also useful for several other applications.
Problem Statement. Formally speaking, if we are given a set S, drawn from a universe or name
space U , that is stored in a Bloom filter B (referred to as the query Bloom filter ), if we denote by
1
S(B) those elements of U \ S that are false positives of B (i.e., the query “Is x ∈ S?” answered by
B returns YES for all x ∈ S ∪ S(B)), then:
1. an algorithm that samples from B is one that returns an element chosen uniformly at random
from S ∪ S(B), and,
2. an algorithm that reconstructs the set stored in B returns the set S ∪ S(B).
Since Bloom filters hide information about the elements stored in them providing only (partially
correct) answers to membership queries, the natural way of trying to sample from a set stored in a
Bloom filter is to fire membership queries with different elements of the name space at the Bloom
filter. Such a method, referred to as a Dictionary Attack is not scalable since its running time is
linear in the size of the name space, which may be huge.
Solution Overview. In this paper, we outline a method that approaches this task much more efficiently. Conceptually, we design a data structure, the BloomSampleTree, that organizes the namespace as a binary search tree. That is, each node of the tree stores a subset of the namespace, but
at each level of the tree, the union of these subsets yields the entire namespace. While the root of
the tree, by itself, stores all the elements of the namespace, each leaf stores only a small subset of
this namespace. Once this binary search tree is constructed, the key idea is to now locate only those
leaves which potentially contain elements present in the given query Bloom filter B. This is done
by intersecting B starting from the root of the search tree and working our way towards the leaves.
Entire subtrees are pruned away because they yield empty intersections, thus eliminating large parts
of the namespace. Once we identify the relevant leaves, we can efficiently sample or reconstruct the
original set using the dictionary attack method explained above. Note that this search tree needs
to be constructed only once and will be repeatedly used for different query Bloom filters.
A drawback in this approach is that we are storing the entire namespace in the BloomSampleTree,
even though only a small part of it may be actually occupied. Sparse occupancy of a namespace
is a regular occurence, especially when we consider non-numeric keys such as strings, where the
namespace is typically of the order of 264 , but the actual occupancy is likely to be of the order of 230
(a little over 1 billion) or perhaps less. Therefore, it is space-inefficient to construct a tree for the
entire namespace, when a large number of the nodes are going to be empty. In order to address this
we present a dynamic version of the BloomSampleTree, we call it Pruned-BloomSampleTree which
takes into account the occupancy and can dynamically change its size and structure as the occupancy
changes.
The BloomSampleTree-based algorithms we provide for sampling and reconstruction have one
very important feature: they do not require the hash functions used by the Bloom filter to be
invertible. Our method only needs to be able to use those hash functions and will work if we are
given the implementation of the Bloom filter used to store the set. It is also important to note
that we do not distinguish between true elements of the set stored in the Bloom filter and the false
positives that are created in the process of insertion. We approach the Bloom filter as is without
any prior knowledge of what has been inserted in it, and without any method of distinguishing true
elements from false positives. In summary, our method is designed to work efficiently in a scenario
where: i) the namespace is potentially large, even dynamic, ii) the no. of interesting subsets is large
(in the millions or billions) and may continue to grow indefinitely, iii) we need to either sample from
or reconstruct a subset(s) from the set of interesting subsets, stored in the form of Bloom filters
(specifically, these are our query Bloom filters).
We present our methods as an aid to the engineer who has chosen to use Bloom filters for a
particular application and has optimised parameters to achieve a given level of accuracy (ie. ratio
of true elements to all the elements that return a true answer to a membership query) and who has
a way of dealing with false positives.
Contributions (i) We introduce a novel data-structure called BloomSampleTree that can be used
sample from a set stored in a Bloom filter as well as reconstruct that set. The BloomSampleTree
takes into account the occupancy of the namespace and can change size as the occupancy changes,
2
(ii) We provide theoretical bounds on the runtime and on the quality of samples generated by our
BloomSampleTree-based algorithm, show them to be near-uniform. (iii) We show through extensive evaluations that our BloomSampleTree-based algorithms are efficient and provide good quality
samples.
Organization In Section 2 we review the literature. We provide a brief background on Bloom
filters and outline the framework in which our methods operate in Section 3. Section 4 outlines two
baseline techniques for sampling from Bloom filters, along with a discussion on their limitations and
the need for our BloomSampleTree method. The BloomSampleTree for sampling and reconstruction
are described in detail in Sections 5 and 6. The results of our detailed experimental analysis are
presented in Sections 7 and 8.
2
Related Work
Bloom filters are one of the most widely used data structures for approximately answering set
membership queries. Their compact storage and efficient querying through simple bit operators
has made them valuable in many different settings. A thorough survey of Bloom filters and their
applications are available in [2, 10]. Despite their widespread use, we are not aware of any work that
systematically addresses the problems of generating provably uniform samples using Bloom filters
and reconstruct the original set at a given level of accuracy in an efficient way.
The problem of identifying at least one true positive from the Bloom filter has been considered in
an adverserial setting to study how resilient the Bloom filters are for dictionary-based attacks [11, 12].
Given a Bloom filter, an adversary can mount an attack to obtain some elements of the original set by
repeatedly posing queries on the Bloom filter – potentially obtaining a large number of false positives,
but also some true positive elements. In our work, we do not operate in an adversarial setting–we
assume complete knowledge of the domain of values represented by the Bloom filter and the hash
functions used. Given an accuracy level, our aim is to efficiently generate provably uniform random
samples from the original set as well as to reconstruct the set as per the accuracy requirements.
We systematically solve these problems and back up our solutions with detailed analysis of time
complexity and accuracy.
Sketches for Handling Large Datasets Bloom filters belong to a general class of approximation
datastructures called sketches or data synopses, which compactly represent massive volumes of data
while preserving some vital properties of the data needed for further analysis [13]. Some of the
sketches used frequently in databases community include histograms, wavelets, samples, frequency
and distinct-value based sketches, and so on.
However, most of these synopses datastructures are used under the assumption that the underlying database is always accessible (e.g., in the case of histograms and samples) or not required (e.g.,
streaming scenarios). Reconstructing the underlying set of data values at a given level of accuracy
in an efficient manner is not their objective to begin with. Only recently, there have been some
results which show how sketches can be used for generating samples, called Lp -samplers [8, 9] which
generalize earlier work on inverse sampling [7]. In these, the goal is to maintain a synopses structure
for a stream of updates (i.e., addition and deletion of counts) over a given domain of size M , such
that at any time it is possible to sample with high accuracy the elements in probability proportional
to their number of occurrances.
Unlike these techniques, our approaches are not focused on streaming setting, and are not designed for specific forms of sampling. The proposed BloomSampleTree approach can be used to
generate uniform samples from Bloom filters, a widely-used and generic synopsis structure.
Trees and Bloom filters In this paper we present the BloomSampleTree that comprises a complete binary tree with Bloom filters stored in every node for the purposes of sampling and reconstruction. Yoon et. al. [14] also propose a structure that comprises a complete d-ary tree with Bloom
3
filters at every node to address the multiset membership problem. Similar in flavor to Yoon et. al.’s
structure is Bloofi proposed by Crainiceanu and Lemire who also address the multiset membership
problem by representing each set as a Bloom filter stored at a leaf of a tree, and building the tree
by combining these Bloom filters hierarchically [15]. While the flavour of both these structures is
similar to our BloomSampleTree but their concern is the problem of multiset membership testing and
so the principle on which their trees are built is completely different from the principle on which we
build our tree and the contents of the Bloom filters stored at each node both bear no relationship to
what we store in each node. Another work that combines Bloom filters and trees is by Athanassoulis
and Ailamaki [16] where the authors modify B+ -trees by placing Bloom filters at their leaves to
create approximate tree indexes that seek to exploit data ordering to improve storage performance.
Their structure is completely different from ours in intent and design.
3
Preliminaries
In this section, we briefly provide the necessary background in Bloom filters, and subsequently
describe the framework which our methods operate in.
3.1
Bloom Filters
A Bloom filter is a probabilistic data structure used to space-efficiently store the elements of a
set. It comprises a bit array of m bits, along with k independent hash functions, h1 . . . hk . An empty
set is represented by a Bloom filter each of whose bits is 0. For each element x in a non-empty set,
the k array positions indicated by h1 (x) . . . hk (x) are set to 1.
A Bloom filter supports membership queries, i.e. a Bloom filter B(S) storing a set S can answer
queries of the form “is x ∈ S” for any x, with a false positive probability that depends on the number
of bits in B(S) and S. x is hashed using each of the k hash functions to obtain k array positions.
If the bit at each of these positions is set, then the result is positive. Since these bits could have
been set due to the insertion of other elements, the probability of a false positive is non-zero and
evaluates to ≈ (1 − e(−kn/m) )k . A Bloom filter is incapable of false negatives.
Other than the membership query, the operation of union and intersection on a pair of Bloom
filters is also supported and can be implemented using bitwise OR and AN D operations respectively.
If B(A), B(B), and B(A ∪ B) use the same m, the same set of hash functions, and are over the same
namespace of values then, B(A ∪ B) = B(A) ∪ B(B). Also, if B(A), B(B), and B(A ∩ B) use the
same m and the same set of hash functions, then B(A ∩ B) = B(A) ∩ B(B), with probability
1 k2 |A−A∩B||B−A∩B|
(1 − m
)
[17].
For two fixed, disjoint sets S1 , S2 ⊂ U , each represented by Bloom filters of m bits and hash
functions h, the false set overlap predicate F SO∩ (S1 , S2 , h) is true if B(S1 ) ∩ B(S2 ) 6= φ even though
S1 ∩ S2 = φ. A false set overlap of S1 and S2 by Bloom filter intersection of B(S1 ) and B(S2 ) is
reported with probability [18],
k2 |S1 ||S2 |
1
P[F SO∩ |h] = 1 − 1 −
m
3.2
(1)
Framework
Our methods operate on a database D = {Xi | i = 1, . . .} of Xi = {xj |xj ∈ M} which are subsets
of elements drawn from a namespace M of size M . Instead of operating on the D directly,we
assume we are only given with a compact approximation D̄ where each Xi is represented by a
Bloom filter B(Xi ), for a given length of the filter (in bits), m, and the set of hash functions used in
its construction, H. Such collections of subsets of elements are commonly seen in many application
settings including graph databases – to represent the adjacency list of each vertex, information
retrieval – to represent the list of documents where a keyword occurs, etc.
4
The first task we are interested in tackling in this setting is that of generating a random sample
given D̄. Specifically, given information about other parameters used in building this approximation
viz., m, H and M, we would like to obtain a provably uniform random sample from a given X ∈ D
- the original database. Since we are operating on an approximate representation, it is also expected
that a fixed amount of inaccuracy (measured as the probability of sampling an element which is not
in X) is tolerable and is specified as an input to the system to begin with. It should be noted that
this inaccuracy is naturally linked to the probability of false positives in Bloom filters, and thus for a
given level of inaccuracy (or accuracy) the Bloom filters used in D̂ have to be designed. The second
task, a natural extension of the above, is to reconstruct the original entry X in the database with
high accuracy.
4
Sampling and Reconstruction
We describe two approaches to sample an element from a set and reconstruct a set stored as a
Bloom filter. The first of these is a simple ”dictionary attack”-based method (DictionaryAttack).
The second uses the weakly invertible property of certain types of hash functions to do sampling
(HashInvert). While both methods can be used to sample from as well as reconstruct a Bloom filter,
the DictionaryAttack method suffers from high runtime inefficiencies, while the HashInvert method
provides no guarantees on the quality of the sample. We compare our BloomSampleTree algorithm
against these two baselines and highlight the advantages and disadvantages of each approach in
detail in Section 7.
DictionaryAttack: Sampling with Membership Queries The DictionaryAttack algorithm
relies on reservoir sampling to guarantee a uniform sample. This is equivalent to reconstructing the
input set and sampling an element from it. It proceeds as follows. A membership query is fired on
the input set for each element in the namespace. When a positive is reported for an element, that
element is retained as the sample with diminishing probability proportional to the size of the set
reconstructed so far [19]. In particular, if n0 is the number of positives reported until now, then the
(n0 + 1)th positive is retained as the sample with probability 1/(n0 + 1). Clearly, the complexity of
this algorithm is O(M ), where M is the size of the namespace.
Note that it is straightforward to use this method to reconstruct the original set.
HashInvert - Sampling with Invertible Hash Functions This method assumes that the hash
functions are weakly invertible. A hash function h is weakly invertible if given the value of h(x),
one can find a set of values S such that ∀y ∈ S, h(y) = h(x). An example of a weakly invertible
hash function is h(x) = (ax + b)%c, where a, b, and c are constants. Knowing the namespace
M , it is straightforward to find a set of elements that all hash to h(x). Given a Bloom filter B,
it exploits the weak invertibility of the hash functions to invert a randomly sampled SET bit s
into k candidate sets P1 (s), P2 (s) . . . Pk (s), each obtained using a different hash function. The k
candidate sets are subsequently pruned using the membership queries on the Bloom filter to obtain
Sk
P10 (s), P20 (s) . . . , Pk0 (s). A value sampled uniformly at random from i=1 Pi0 (s) is the final sample
returned.
Analysis When sampling from the obtained candidate sets is done using a method such as reservoir
sampling, the HashInvert method occupies no extra space.
Sampling a set bit takes O(m) time, where m is the size of the Bloom filter. Once a set bit is
chosen, inversion using a hash function takes O M
m time. The overall time taken for sampling is
O m + kM
.
m
Note that, in contrast to DictionaryAttack, which provides uniformly random samples, no bounds
are given regarding the quality of the samples in the case of HashInvert. However, the algorithm
can be used to reconstruct the original set by exhaustively running the HashInvert algorithm on all
set bits of the Bloom filter.
5
A simple trick gives us more benefits from the HashInvert algorithm. If the Bloom filter is dense,
then the number of UNSET bits (0-bits) are potentially less than the number of SET bits. Therefore,
instead of inverting the set bits, we can invert the unset bits. This results in a set of elements which
are not present in the original set. Therefore, the original set can be recovered from a set difference
operation.
5
Bloom Sample Tree
In this section we define the BloomSampleTree data structure that will help us sample from
and reconstruct a set stored in a Bloom filter. The BloomSampleTree basically organises the entire
namespace. Note that the BloomSampleTree is built once and is then used repeatedly to sample
from any given query Bloom filter B.
5.1
Definition
The BloomSampleTree is a complete binary tree, denoted as T , with log M/M⊥ levels, where M⊥
is a threshold whose choice we discuss later in this section. Every node in the BloomSampleTree
has a Bloom filter that stores a subset of the namespace. Every level of the tree contains the
entire namespace partitioned uniformly amongst the nodes of that level. Hierarchically speaking
the organisation is laminar in the sense that the union of the subsets of the namespace stored in
two sibling nodes gives us the set stored in their parent node. All the Bloom filters used in the
BloomSampleTree have the same parameters – viz., m, the number of bits and, H, the set of hash
functions, as the Bloom filters used for the sets we are sampling from (or trying to reconstruct).
The reason for this is that we will be frequently intersecting the Bloom filter B of the set of interest
with the Bloom filters Bi stored at various nodes in the BloomSampleTree. We now present a more
formal definition.
Definition 5.1. Given a namespace M of size M , the size of the Bloom filter m, a set of hash
functions used for construction H of the form h : M → {0, 1, . . . , m − 1}, and an integer parameter
M⊥ < M the BloomSampleTree, T (M, m, H, M⊥ ), is a collection of Bloom filters
M
Bi,j : 0 ≤ i ≤ log
, 0 ≤ j < 2i ,
M⊥
such that each of these Bloom filters uses a bit vector of size m and the hash functions H, and with
the property that the Bloom filter Bi,j stores the elements
M
M
` : j · i ≤ ` < (j + 1) · i .
2
2
Note that,
• The collection of Bloom filters forms a tree structure. Since the portion of the name space
stored in Bi,j is partitioned equally amongst the nodes Bi+1,2j and Bi+1,2j+1 , Bi,j is the parent of
these two nodes in the tree.
• The leaves of the tree all store sets of size M⊥ . The namespace is not further subdivided.
Figure 1 shows an example BloomSampleTree for a namespace of M = 16. Each node in the tree,
except for the root, consist of Bloom filters of size m = 10 storing the range of elements depicted at
the node. A set S = {4, 6}, stored as Bloom filter b is the query set that we need to sample from or
reconstruct. Note that the Bloom filters in T are constructed with the same m and H as the Bloom
filter b for set S.
6
Bloom Sample Tree bT (M
= 16, m = 10, k = 2)
(0..15)
1111101101
1111111110
(0..7)
(8..15)
(0..3)
(4..7)
(8..11)
(12..15)
1101101100
1011011001
0110111010
1000011100
Bloom filter b
0011100001
S = {4,6}
Figure 1: A BloomSampleTree T with 3 levels and the query Bloom filter b representing the set S
from which we want to sample
5.2
Pruned-BloomSampleTree
As mentioned in the introduction, even though the namespace itself may be large, it is likely that
only a small portion of it is occupied. Therefore, building a complete BloomSampleTree as explained
in the previous section potentially wastes a huge amount of space. For example the real-world data
set on which we experimented (see Section 8) is taken from Twitter and contains 7.2 million user ids
distributed in a namespace of size 2.2 billion, i.e., the fraction of the namespace occupied is of the
order of 10−1 . Therefore, in practice we build the tree only for those portions of the namespace that
are actually occupied; we call this condensed version the “Pruned-BloomSampleTree”. This tree can
dynamically change its structure, based on the change in the occupancy of the namespace. That is,
if more of the namespace is assigned, then the tree potentially contains more nodes to reflect that.
An overview of the algorithm to build this tree is as follows: Let M0 ⊆ M be the set of identifiers
that are currently in use (M is the namespace, M 0 = |M0 | ).
• Initialise queue with Node(0, log M ).
• Repeat until the queue is empty
– Dequeue Node(a, b).
/* b is the level, a is the offset within that level */
– Check if the range (a, a + 2b − 1) has a non-empty intersection with M 0 .
7
∗ If yes, then create Bb,a/2b and attach in the tree; insert elements from M0 in the range
(a, b) in Bb,a/(2b ) ; if b > log M⊥ then enqueue Node(a, b − 1), Node((a + 2b−1 ), b − 1).
/* Create the Bloom filter corresponding to the subrange and grow the next level at this
point */
∗ If no, then do nothing.
The above algorithm essentially goes down the tree building subtrees where required to accommodate elements of M 0 and ignoring subtrees corresponding to ranges that have no overlap with M 0 .
Although this algorithm constructs the search tree when the M 0 is known ahead of time, it is easy
to see how to evolve the Pruned-BloomSampleTree when M 0 grows (e.g. when new Twitter accounts
are made)–either we need to insert this new element into already existing nodes in the tree, or we
need to create a new node (and potentially its subtree).
The time taken to build the Pruned-BloomSampleTree offline is proportional to the size of the
final tree constructed multiplied by the time for the range query on M 0 . The time taken to update
the tree is proportional to the height of the tree.
5.3
Sampling with the BloomSampleTree
Given a query Bloom filter B to sample from, the algorithm proceeds from the root in the
following recursive manner and relies on the pruning of the search space for performance gains.
• At a given (non-leaf) node, compute the intersection of the Bloom filters stored in the left and
right children of this node with b. If for both child nodes, the intersection is empty, then b does not
contain any element belonging in the range associated with this node. Therefore the subtree rooted
at this node is pruned from the search.
• If intersection with only one child is non empty, then the search proceeds along that child
node. The other child node and the subtree rooted at it are pruned from the search.
• If intersection with both child nodes is non empty, then one of the child nodes is selected
with probability directly proportional to the estimated number of elements in their corresponding
intersections and the search proceeds along that child. Note that it is possible that the intersection
was a false positive and this is discovered further down this subtree. In that case, the search then
backtracks and proceeds along the other child node.
The estimated number of elements in the intersection of two Bloom filters B1 and B2 is given by the
following expression [20]:
∧ ×m−t1 ×t2
ln m − tm−t
− ln(m)
1 −t2 +t∧
Ŝ −1 (t1 , t2 , t∧ ) =
1
k × ln 1 − m
where t1 is the number of bits set in B1 , t2 is the number of bits set in B2 , m is the size of both
Bloom filters, k is number of hash functions used in both, and t∧ is the number of bits set in the
bitwise AND of B1 and B2 . We recall that Equation (1) gives us the probability that this intersection
is incorrectly estimated to be non-empty when the two sets stored in B1 and B2 are disjoint. We
discuss this issue further in Section 5.6.
• At a leaf node, every element in the range of the node is checked for membership in b. The
sample at this leaf node is a value sampled uniformly at random from the set of values that satisfy
the membership test of b. If none of the elements within this range satisfy the membership query it
indicates that the search has reached this leaf node due to a (string of) false set overlap(s). In this
case, the sample at this node is N U LL.
This algorithm is called BSTSample, and a formal description is in Algorithm 1
Figure 2 shows a typical scenario that is encountered when sampling with the BloomSampleTree.
The numbers to the side of the node indicate the order in which the nodes are traversed. As shown,
the algorithm ultimately generates a sample from a leaf node by following one ”true” path out of
several false positive paths that may branch out at multiple places. Note, for example, that node
7, is ultimately determined to have led to several false positive paths discovered subsequently in
8
False Positive Path
Empty Intersection
Potential Path
True Path
(1 ... 10M)
1
3
2
4
5
7
6
8
9
10
14
12
Subtree
pruned
from
search
11
15
25
13
Subtree
not
visited
at all
37
38
Figure 2: A typical scenario: Sampling with BloomSampleTree. A False Positive Path is chosen
because of errors in determining the empty intersection. The Empty Intersection immediately results
in the pruning of a subtree. Potential paths are left unexplored when there is choice of following
either subtree. The True Path is the path actually taken by the algorithm to generate the sample.
its subtree. In contrast, the whole subtree at node 4 is immediately pruned from the search space.
Once the search reaches a leaf, a brute force search is conducted and there is no scope of a false set
overlap due to Bloom filter intersection.
Sampling multiple items The algorithm presented for sampling outputs a single sample. To
sample multiple items we could run this algorithm multiple times. However, these multiple runs can
be done together in one pass down the BloomSampleTree as we now explain. Given an integer r that
is less than the size of the set stored, we send r independent search paths down the BloomSampleTree
according to the algorithm BSTSample. These paths are sent down the BloomSampleTree in a single
pass since all the paths arriving at any internal node or leaf can be processed at the node or leaf
before we move on. If at a node we find that the Bloom filters at both its children intersect with the
query Bloom filter, we take each of these r paths and, independently of the other paths choose one
of the children at random as in BSTSample and send the path down to that child. This continues till
each of the r paths reaches a leaf. Let us take a concrete example to illustrate this process: Assume
we are given a query Bloom filter B and r = 3. We intersect B with the Bloom filters B1,1 and
B1,2 stored in the left and right child of the root of BloomSampleTree and estimate the size of both
intersections, let us say they are k1 and k2 . Now throw three independent coins biased to come up
heads with probability k1 /(k1 + k2 ). Suppose two of these coins come up heads and one comes up
tails, recursively call two instances of the multiple sampling method with B on B1,1 with r = 2 and
on B1,2 with r = 1.
It is easy to see that given the tree structure of the BloomSampleTree, such an extension of the
algorithm BSTSample will, in general, perform better than r times the running time for the case
when we ask for a single sample as output. Since all the paths behave like a single sampling path of
BSTSample the guarantee on sample quality is maintained. Finally, if two or more paths happen to
reach the same leaf we can sample at that leaf with or without replacement depending on whether
the r samples are to be generated with or without replacement.
5.4
Summary of Analyses
Given the BloomSampleTree structure and the algorithm for sampling, we briefly summarise the
analyses we performed and the effect of the various parameters.
Quality of samples The first question we answer is whether our method generates a uniformly
random sample. The answer is that a uniformly random sample in indeed generated with high
probability. We prove this property in Section 5.5 and show this empirically as well in Section 7.2.
9
Accuracy Given that Bloom filters are approximate data structures, it is possible that the
samples we generate do not actually belong to the original set (recall that a sample is generated by
membership queries at a leaf). We quantify the accuracy of our samples as follows:
acc =
n
n + (M − n) ∗ F P
where n is the number of elements in the query set, M is the size of the namespace and F P is the
probability of false positives in our Bloom filter implementation. The accuracy defined here simply
computes the ratio of correct outcomes to all potential outcomes of the algorithm.
Clearly the size of the Bloom filter m has an effect on accuracy and we can determine m based on
the desired accuracy. We show the performance of our method for various values of accuracy in
Section 7.
Runtime analysis The runtime of the algorithm depends on the number of false paths it may
follow. We analytically show the expected number of nodes visited in Section 5.5, given a BloomSampleTree. However, we also address a practical issue here with regard to runtime – the cost of
performing intersections at a node as opposed to the cost of performing a number of membership
queries. Note that it is possible that based on the hash function used, the cost of membership queries
may be cheaper or more expensive than the cost of intersections. These two costs are directly related
M
to the no. of elements stored at the leaf, M⊥ and the height of the BloomSampleTree is log M
. We
⊥
tradeoff the costs as follows:
If mcost is the cost of one membership query to a Bloom filter of size m with k hash functions,
and icost is the cost of an intersection between a pair of bloom filters of size m, then, at a current
node N , storing N⊥ values, we would like to determine whether it is better to perform membership
queries over N or perform intersections until the leaf which is at most at level log N⊥ below N . If
performing membership queries is preferred over traversing further down in the tree, we can truncate
the tree such that N is the leaf of the tree. Hence, we determine M⊥ = max N⊥ , such that
N⊥
icost
≤
.
log N⊥
mcost
We empirically show the runtime costs throughout Section 7.
Memory requirements The memory required to store the BloomSampleTree (which is constructed only once and used repeatedly) depends on the size of the Bloom filter m and the number
M
of levels in the tree, log M
. An interesting observation here is that, in our framework, there is no
⊥
tradeoff between memory and accuracy or memory and runtime. The tradeoff is between accuracy
and runtime, as explained in the previous paragraphs. Therefore, while we set the best possible
M
m and log M
in order to optimize the runtime, the memory required may actually reduce for in⊥
creased accuracy. The cause for this is that while we will need to use a larger Bloom filter in the
BloomSampleTree for increased accuracy, we would potentially reduce the number of levels so as to
reduce the intersection cost (as described in the previous paragraph). The effect of this is that we
end up reducing the space used, while increasing accuracy, but also increasing the runtime. We
discuss empirical results about this in more detail in Section 7.2.
5.5
Sample quality and running time
The first question that arises is: what is the distribution of samples BSTSample produces? Our
aim is to produce a uniform distribution from the set stored in the Bloom filter. We present a
theoretical result that shows that the samples produced are near uniform. We first state the result
and then discuss its implications.
Proposition 5.2. Given a set S with |S| = n taken from a name space of size M , if we run
BSTSample
on S with a BloomSampleTree T (M, m, H, M⊥ ) such that |H| = k, define (m) =
√
2nk(log m+log log m+log n)
. Then the probability that the sampling algorithm finally samples from
m
an L ⊆ S of size ` that is stored in a Bloom filter in a leaf of the BloomSampleTree lies between
10
M
(1−(m))· n` and (1+(m))· n` with probability at least 1− log4 m , as long as f (m) = 2·(m)·log M
→0
⊥
as m → ∞.
1 n
. Let ẑ be a random
Proof. Probability of a bit being zero after insertion of n elements = 1 − m
variable indicating the number of zero bits. We have that,
nk
1
E(ẑ) = m 1 −
m
or that, E(ẑ) = mp, where p = 1 −
1 nk
.
m
From Theorem 1 of [21], we have that
P [|ẑ − mp| > m] < 2e
√
We set =
2nk(log m+log log m+log n)
.
m
−2 m2
2nk
The estimated size of the population of a bloom filter is n̂ =
c=
1
.
1
k log(1− m
)
2
nm log m = o(1).
log (ẑ/m)
, or n̂ =
1
k log(1− m
)
Then P[|ẑ − mp| > m] <
c log(ẑ/m), where
Also, from the above bound, we have that with probability at least 1 − 2/nm log m,
mp − m ≤ ẑ ≤ mp + m.
Therefore, with probability at least 1 − 2/nm log m,
c log(p − ) ≤ n̂ ≤ c log(p + ).
For small values of ,
Substituing c =
1
1
k log(1− m
)
c (1 − ) log p ≤ n̂ ≤ c(1 + ) log p.
1 nk
, and p = log 1 − m
,
n (1 − ) ≤ n̂ ≤ n (1 + ) ,
with probability at least 1 − 2/nm log m.
Returning to the setting of the BloomTree, let the number of elements in the intersection of the
query Bloom filter with the root node be K. Similarly, let the number of elements in the intersection
of the query Bloom filter with the left and right child of the root node be K1 and K −K1 respectively.
In the sampling process, the probability of proceeding along the left child is directly proportional
to the estimated number of elements in the intersection of the left child and the query Bloom filter.
Let n̂l and n̂r be the estimated number of elements in the left and right child of the root node
respectively.
n̂l
P[selecting the left child] =
n̂l + n̂r
With probability at least 1 − 4/nm log m,
n̂l
(1 − ) K1
≤
,
(1 + ) K1 + (1 + ) (K − K1 )
n̂l + n̂r
and
n̂l
(1 + ) K1
≤
,
n̂l + n̂r
(1 − ) K1 + (1 − ) (K − K1 )
i.e.,
(1 − ) K1
n̂l
(1 + ) K1
≤
≤
.
(1 + ) K
n̂l + n̂r
(1 − ) K
11
Now, for || < 1/2, (1 − )/(1 + ) ≥ 1 − 2 and (1 + )(1 − ) ≤ 1 + 4 and so
(1 − 2)
K1
n̂l
K1
≤
≤ (1 + 4)
K
n̂l + n̂r
K
Consider now a given leaf node L of the Bloom Tree that contains a subset S 0 of size `. The
probability that an unbiased process of sampling should reach this leaf is `/n. We now estimate
what the probability of reaching this leaf is in the BloomTreeSample method.
Consider the path down the BloomTree to the given leaf: L0 , L1 , . . . Llog M/M⊥ = L and the
sequence of subsets of S stored in these nodes be S = S0 , S1 , . . . , Slog M/M⊥ = S 0 . Then we have
that
|Slog M/M⊥ |
|S1 | |S2 |
`
·
···
= .
|S0 | |S1 |
|Slog M/M⊥ −1 |
n
Now, using the analysis done above, we can argue that for the choice of above, the probability of
moving from Si to Si+1 is at least (1 − 2)|Si+1 |/|Si | and at most (1 + 4)|Si+1 |/|Si | with probability
1 − 4/nm log m. Repeating this argument over log M/M⊥ levels we get that
(1 − 2)log M/M⊥ ·
`
≤ P[BloomTreeSample reaches L],
n
and
P[BloomTreeSample reaches L] ≤ (1 + 4)log M/M⊥ ·
`
,
n
with probability at least
1−
4 log M/M⊥
4
≥1−
.
nm log m
n log m
, where the last inequality applies because log M/M⊥ < m whenever f (m) → 0 as m → ∞.
Since the set S can be present in at most n such leaf nodes, hence for any leaf node containing an
element of S, the probability that BloomTreeSample reaches that leaf node is at least
1−
4
.
log m
Now, we recall that (1 + x) < ex and for x < 1/2, e−x/2 < (1 − x) and so In other words, with high
probability
`
e− log M/M⊥ · ≤ P[BloomTreeSample reaches L],
n
and
`
P[BloomTreeSample reaches L] ≤ e4 log M/M⊥ · ,
n
for any leaf L that contains an element of S. Now, whenever log M/M⊥ is o(1), i.e., it goes to 0 as
m grows, both the bounds go to 1 proving the result.
Discussion on sample quality We note that since f (m) grows faster than (m), the condition
that f (m) → 0 as m → ∞ implies that (m) → 0 as m → ∞. The import of Proposition 5.2 is that
the eventual leaf that BSTSample selects for sampling is chosen with probability that is very close to
being proportional to the number of elements of the set S that belong to the segment of the name
space stored in that leaf. In the absence of false positives, which happens in the limit as the size of
the Bloom filter m → ∞, this would lead to perfectly uniform sampling.
We now move on to analysing the running time of the algorithm. Clearly the running time
depends on a number of factors. We provide a theoretical analysis of the number of nodes that
BSTSample visits as it moves down the tree to reach a leaf from where it generates a sample. Clearly
the lower bound for this number is the height of the tree, i.e., log M/M⊥ . We are not able to match
12
this lower bound but we are able to control the number of extra nodes visited and give a result
which can guide us on how to choose our system’s parameters to ensure a good asymptotic running
time. We show the following result:
Proposition 5.3. The expected number of BloomSampleTree nodes visited by the algorithm
BSTSample when sampling from a set of size n using a BloomSampleTree T (M, m, H, M⊥ ) where
|H| = k is
M
M k2 n
.
O log
+
M⊥
m
of Proposition 5.3. The path BloomTreeSample takes down to the leaf node from which it generates
a sample has length equal to the height of the tree, i.e., log(M/M⊥ ). Along this way, there are many
branches that are caused by false set overlaps, i.e., the intersection Bloom filter has at least k bits
set but no element of the name space has all k of its bits set in this Bloom filter. These branches
need to be followed since it is not possible to distinguish them from the branch that takes us to
genuine true or false positives of the set stored in the Bloom filter that we are sampling from. In
order to estimate the number of nodes visited over and above the log(M/M⊥ ) in the true path, we
need to estimate how many such false set overlap nodes we visit. We proceed by showing that below
a certain depth in the BloomTree a false set overlap branch leads to a constant number of false set
overlap nodes being visited below it. Hence, below that depth the number of extra nodes being
visited is just a constant factor more than the necessary nodes being visited along the true path.
Above that level, however, we cannot make any such claim so we just assume we visit every node.
Formally, we proceed by making the following claim:
Claim 5.4. Given a node u of depth d in a BloomTree T in which M⊥ names are stored at every
leaf, and a query set S such that S ∩ M [u] = ∅ where M [u] denotes the subset of the namespace
M that is represented in the subtree of T rooted at u, if the probability that the intersection of two
Bloom filters of size m containing S and M [u], i.e., the quantity
k2 n|M [u]|
1
αS (d) = 1 − 1 −
,
m
is at most 1/2, and if L(d) is the number of nodes that algorithm BloomTreeSample visits in the
subtree of u conditioned on the event that it reaches u, then
E(L(d)) =
αS (d)
.
1 − 2αS (d)
√
Proof. The proof of the claim follows by noting that if αS (d) = 1 − x, then αS (d − 1) = 1 − x,
since the number of names stored in any node of a child is exactly half of those stored by its parent
in the BloomTree. Noting further that it is only if the sampling algorithm finds an overlap at u that
it goes into the subtrees of u’s children, we get
E(L(d)) = αS (d) [1 + 2L(d − 1)] ,
Taking expectations repeatedly we get
1
1
1
E[L(d)] = (1 − x) + 2(1 − x)(1 − x 2 ) + 4(1 − x)(1 − x 2 )(1 − x 4 ) . . . ,
√
summed till the leaf level, where x = 1 − αS (d). Since x < 1, 1 − x < 1 − x, which allows us to
say that
log(M/M⊥ )−d
X
1
E[L(d)] =
(2αS (d))i ,
2
i=1
from where the result follows whenever αS (d) < 1/2. Note that this argument is basically the
same as saying that the sampling algorithms visits to nodes below a level d for which αS (d) < 1/2
is dominated by a subcritical branching process, which goes extinct with probability 1 and yields
1/(1 − µ) nodes in expectation, where µ is the mean of the progeny distribution (c.f. [22])
13
Noting that since |M [u]| = M/2d whenever the depth of u is d, the condition that αS (d) < 1/2
is satisfied whenever
M k2 n
d > d∗ = log
,
m ln 2
∗
and that the BloomTree up to d∗ levels contains 2d
+1
− 1 nodes, we get the result.
Discussion on running time Looking at the result we note that the ratio of name space to
Bloom filter size is a critical element in the number of nodes visited, i.e., raising the number of bits
in the Bloom filter will benefit the running time (at least up to the point where the second term in
the running time analysis continues to dominate the first term). The number of hash functions used
and the size of the set being sampled from are also correlated with the running time, which follows
intuition.
5.6
Determining the empty intersection
The BloomSampleTree data structure and the algorithm for sampling are both straightforward
to implement. However, one practical problem we encounter is that at each node that the algorithm
visits, a set intersection needs to be performed that determines whether to prune that branch or
not. Unfortunately, there is no reliable way to determine that the size of a set intersection is empty,
since even a single set bit results in a non-zero size estimation. Therefore, we use thresholding to
overcome this problem. That is, if the estimated size of the set intersection is below a particular
threshold, we consider the intersection to be NULL. Note that this heuristic can potentially affect
the theoretical guarantee offered in Proposition 5.2, but in effect it will not since the probability of
making a wrong decision here, i.e. assuming a set if empty when in fact it is not, is very small if
we choose the correct threshold. A wrong decision here implies that certain elements of the set are
never presented as samples, but, as we see in Section 7.2, this does not happen in practice.
6
Reconstruction with BloomSampleTree
A recursive traversal of the tree results in a reconstruction of the set. Given a Bloom filter B,
if the intersection with B at a non-leaf node is empty, then the reconstructed set at this node is
the empty set, and the subtree rooted at this node can be pruned from the search. However, if the
intersection is not empty, then the search continues along the left and the right children and the
final reconstructed set at this node is the UNION of the reconstructed sets obtained from the two
child nodes.
If the intersection with B at a leaf node is not empty, we conduct a brute force search on the
range of this node as before. However, instead of sampling a value from the set of elements thus
obtained, we return the set itself as the reconstructed set at this node.
We note that the expected number of nodes of the BloomSampleTree visited by the reconstruction
algorithm can be analysed in a manner similar to BloomTreeSample. This expected number will
come to
M
M⊥ k 2
O n · log
+
.
M⊥
m
We note that extracting a single element of a set from the treelike structure of the BloomSampleTree
would take log M/M⊥ and, in the worst case, assuming the different elements of the set are widely
distributed in the name space, the worst case number of nodes visited for reconstruction will be
n log M/M⊥ at least. Our algorithm is, unlike in the case of sampling, able to meet this lower bound
exactly in an asymptotic sense. Also, since the second term above is directly proportional to k 2
and M⊥ and inversely proportional to the size of the Bloom filters used, m, we can choose these
parameters appropriately to minimize the time taken.
14
7
Experimental Evaluation with Static Namespace
In this section, we describe several experiments that we conducted to determine the effectiveness
of our techniques for both sampling as well as reconstruction when the namespace is static. In
Section 8, we describe our experiments where only a fraction of the namespace is actually used.
7.1
Setup
We experimented with both synthetic as well as real datasets. We made extensive use of synthetic
datasets to generate controlled micro-benchmarks. We varied the namespace size (M ), from which
the elements are drawn, between 105 -107 . We also varied the size of the sets (n) between 100 to
50, 000, and generated them either by uniformly sampling from the namespace or by forming random
local clusters (details below). As we pointed out earlier in Section 5.4, the desired accuracy levels can
be used to determine the size of the Bloom filter m to construct the BloomSampleTree. We varied the
accuracy requirements between 0.5 – 1.0 and accordingly designed the Bloom filters. For simplicity
of experiments we kept the number of hash functions to 3, although we experimented with different
classes of hash functions viz., simple, Murmur3 and MD5. Table 1 summarizes the parameter choices
used in our experiments. Unless mentioned explicitly, the default values of parameters mentioned
in this table were used in our experimental evaluations.
Generating clustered and uniform query sets We report results on two kinds of randomly
generated query sets. Uniform sets are constructed by generating elements uniformly at random,
without replacement, from a given range.
The idea for generating clustered query sets comes from the observations in Web graphs where
neighbour sets of vertices typically have their ids clustered around a few nodes [23]. To generate
clustered query sets, n elements were iteratively sampled from the namespace using a pdf that is
updated after each sample is drawn. Initially, the pdf begins as that of the uniform distribution.
After a sample s is drawn, we identify x = max{i|i < s, pdf (i) > 0} and y = min{i|i > s, pdf (i) > 0}
as the neighbors of s. We divide the pdf (s) equally into pdf (x) and pdf (y), and set pdf (s) = 0. To
generate more aggressively clustered sets, one can subtract p% from the probability of each element
and equally divide the accumulated probability into x and y. Here, p controls the degree of clustering.
For our experiments, we have used p = 10.
Algorithms Our baseline method is the brute-force dictionary attack (referred to as DA). Additionally, for evaluating the performance of set reconstruction we use HashInvert (HI). Both these
baselines are described in Section 4. These methods are compared against our BloomSampleTree
(BST ) approach.
Metrics and Methodology We report on the following:
• Number of intersections and set membership operations: This is our main metric
of interest. We compute the depth of BloomSampleTree and the size of the Bloom filters based on
accuracy and the relative costs of intersection and membership operations, as discussed in Section 5.4.
For uniform and clustered query sets, we report the average number of intersection and membership
operations on Bloom filters over 10, 000 samples.
• Average time taken: With the same setup as above, we report on the average time over
10, 000 samples.
• Memory: We analytically computed the overall size required by the BloomSampleTree based
on the size of the Bloom filters used and the tree depth.
• Quality of uniform samples: We report the χ2 -statistic for the samples generated. In
addition, we also show the empirically observed distribution of samples.
15
7.2
Sampling Experiments
Figures 3 and 4 show the number of intersections and membership operations over uniformly
random and clustered query sets respectively. The DA method always uses M membership operations
and no intersection operations. On the other hand, BloomSampleTrees try to offset a large number of
membership operations with few intersections of Bloom filters. Note that as the sampling accuracy
increases, the size of Bloom filters, m, increases as well resulting in more expensive intersection
operations.
Runtime Performance Both intersection and membership operations become more expensive
– intersections more so than membership operations – as the Bloom filter size increases. The
Bloom filter size, in turn, is determined by the namespace size as well as the accuracy requirements.
Thus, the overall efficiency of BloomSampleTree depends on careful balance of the number of these
operations. Figure 5 shows the average time taken by BloomSampleTree and DA methods, for a
namespace size of 10-million. As these plots show, BloomSampleTrees achieve improvements in
efficiency over DA for a single sampling round.
Another implementation choice which can significantly affect the performance numbers is that
of hash functions. Figure 7 shows the time taken to generate samples with different hash function
families.
The Dictionary Attack suffers most when the cost of computing the hash function increases
– for instance with MD5-based hash functions the performance goes down by almost an order of
magnitude. On the other hand, the BloomSampleTree sampling procedure defers membership queries
to lower levels of the tree, by which time most of the tree has already been pruned from the search.
When using fast hash functions like Murmur3 or Simple, BloomSampleTree automatically leverages
their efficiency to reduce the overall time taken.
Memory requirement Finally, we turn our attention to the amount of memory footprint needed
by each method. The memory requirements (in MBs) are shown in Tables 2 and 3 when the number
of elements in the query set is n = 103 . In the Tables 2 and 3, Depth is log M/M⊥ where M⊥ is
computed as discussed in Section 5.4, and memory is analytically computed using m∗ number of
nodes in the BloomSampleTree. The memory of the BloomSampleTree thus computed was further
affirmed by empirical measurement during program execution. It is evident from this table that
memory requirement might actually reduce with increasing accuracy. This is primarily because
when the depth of the BloomSampleTree decreases, the total memory occupied reduces. This is
in spite of increased Bloom filter sizes, since lower levels have much larger memory footprint than
higher ones. Since memory can reduce with increasing accuracy, the overall trade-off is between
accuracy and memory on one hand, and running time on the other.
While the BloomSampleTree allows for very fast sampling, it requires small additional storage than
the other methods described in this paper. Moreover, one does not need to store a BloomSampleTree
for each possible query set. There is only one BloomSampleTree for a given size of namespace, Bloom
filter size, and choice of hash functions.
Quality of Sampling We use the Pearson’s chi-squared test, which we briefly describe here, to
empirically validate the sample quality. We conduct T sampling rounds from a Bloom filter B storing
a set S = (S1 , S2 , . . . Sn ). Now, ∀i ∈ [1, n], let oi be the number of times element Si is sampled.
Similarly, let ei be the expected number of times element Si should be sampled. Our null hypothesis,
H0 , is that the sampling is uniform, or restated, that ∀i ∈ [1, n], ei = Tn . The goal of the chi-squared
test would be to see if the null hypothesis should be rejected given the observations oi . We define a
2
Pn
i)
random variable Q = i=1 (oi −e
. Then Q follows a χ2 distribution with n − 1 degrees of freedom.
ei
Given an observation (o1 , o2 , . . . on ), we compute the value of Q. Let this value be q. The p-value
is defined as P (Q ≥ q|H0 ). Clearly, smaller the p-value, higher is the value of q, indicating greater
deviation from the expectation. In other words, a smaller p-value indicates that the observation has
lesser support for H0 . If the p-value falls below a threshold s, known as the significance level, then
16
H0 is rejected, otherwise it is not. The significance level is typically set around 0.05. We set it to
0.08 and use T = 130 × n, the recommended sample size for this significance level [24]. For M = 106 ,
the p-values thus obtained are reported for sets of different sizes in Table 5. All of the entries in this
table are > 0.08, and therefore the null hypothesis is not rejected in any case. For higher values of
accuracy, it is clear that the distribution of the elements is close to the uniform distribution.
Accuracy While the value of m was determined based on accuracy, we verified the accuracy
obtained from the sampling process using the expression for measured accuracy in Table 1. For all
cases, measured accuracy was found to be close to the expected value. Table 6 shows measured
accuracy values for n = 1000.
Creation Time: In Table 4, we measure the time taken to create the BloomSampleTree for
different sizes of the namespace and desired accuracy. Note, for the case of M = 105 , with increased
accuracy from 0.5 to 0.6, the required value of m increases, resulting in a decreased depth of the
BloomSampleTree (Section 5.4), and consequently lower creation time. For M = 107 and desired
accuracy 0.9, creation takes < 2.5s. In the application scenarios the BloomSampleTree works with,
the number of sets that must be constructed in Bloom filters is assumed to be massive, and changing
Bloom filter parameters requires creating each of the sets in the database again which is prohibitively
expensive, overshadowing the time taken to create the BloomSampleTree.
17
Algorithm 1: Sampling with BloomTrees
1 Algorithm BSTSample(bST :BloomSampleTree, b:Query Bloom filter )
2
[s, e] ← bST.range;
3
/* At a leaf, exhaustively check the interval of the leaf for membership in
b
*/
4
if bST is leaf then
5
S←φ;
6
for i in (s, e) do
7
if i ∈ b then
8
S ← S ∪ {i} ;
9
end
10
end
11
x ← uniformly sampled from S ;
12
return x ;
13
else
14
lFlag = EstimatedSize(bST.lef t ∩ b);
15
rFlag = EstimatedSize(bST.right ∩ b);
16
/* If no child intersects, we have reached here due to a false positive
and should return N U LL
*/
17
if (not lFlag and not rFlag) then
18
return N U LL ;
19
else
20
/* Randomly select one to proceed search along, with probability
proportional to estimated number of elements
*/
21
z = uniform(0, 1);
22
if z < lF lag/(lF lag + rF lag) then
23
sample ← BSTSample(bST.lef t, b) ;
24
/* In case no sample is found from that child, search along the
other child - Backtracking
*/
25
if sample = N U LL then
26
sample ← BSTSample(bST.right, b);
27
return sample;
28
else
29
sample ← BSTSample(bST.right, b) ;
30
if sample = N U LL then
31
sample ← BSTSample(bST.lef t, b);
32
return sample;
33
end
34
end
Table 1: Parameters of our experiments
Parameter
Range (Default value)
Size of the namespace (M )
Size of the query set (n)
Sampling accuracy(s)
Hash families
105 – 107 (107 )
100 – 50, 000 (1000)
0.5 – 1.0 (0.9)
Simple ((ax + b)%m), Murmur3, MD5
18
Intersection
20
100-BST
1K-BST
10K-BST
50K-BST
DA
0
1010
1000
Membership
(x103)
990
15
10
5
0
0.5
0.6
0.7
0.8
Sampling Accuracy
0.9
1.0
(a) M = 105
10K-BST
50K-BST
DA
Membership
(x103)
Intersection
40
30
20
10
0
1010
100-BST
1K-BST
1000
990
10
0
0.5
0.6
0.7
0.8
Sampling Accuracy
0.9
1.0
400
300
200
100
0
1010
Membership
(x104)
Intersection
(b) M = 106
100-BST
1K-BST
10K-BST
50K-BST
DA
1000
990
10
0
0.5
0.6
0.7
0.8
Sampling Accuracy
0.9
1.0
(c) M = 107
Figure 3: No. of intersections and set membership queries for uniformly random query sets. XXXBST in the legend refers to the cardinality of the query sets.
19
Intersection
50
100-BST
1K-BST
10K-BST
50K-BST
DA
Membership
(x103)
0
1010
1000
990
0
100
0.6
100-BST
1K-BST
0.7
0.8
Sampling Accuracy
(a) M = 105
10K-BST
50K-BST
0.9
1.0
DA
50
0
1010
1000
990
20
10
0
Membership
(x103)
Intersection
0.5
0.5
0.6
0.7
0.8
Sampling Accuracy
0.9
1.0
500
400
300
200
100
0
1010
1000
990
20
10
0
100-BST
1K-BST
10K-BST
50K-BST
DA
Membership
(x104)
Intersection
(b) M = 106
0.5
0.6
0.7
0.8
Sampling Accuracy
0.9
1.0
(c) M = 107
Figure 4: No. of intersections and set membership queries for clustered query sets. XXX-BST in
the legend refers to the cardinality of the query sets.
20
10K-BST
50K-BST
100-BST
1K-BST
100-DA
1K-DA
10K-DA
50K-DA
Time (ms)
625.0
125.0
25.0
5.0
1.0
0.5
0.6
0.7
0.8
0.9
1.0
Sampling Accuracy
(a) Uniformly random query set
10K-BST
50K-BST
100-BST
1K-BST
100-DA
1K-DA
10K-DA
50K-DA
Time (ms)
625.0
125.0
25.0
5.0
1.0
0.5
0.6
0.7
0.8
0.9
1.0
Sampling Accuracy
(b) Clustered query set
Figure 5: Avg. time taken for sampling with M = 107
Accuracy
0.5
0.6
0.7
0.8
0.9
1
m
28465
32808
38259
46000
60870
137230
Depth
10
10
10
9
9
6
M⊥
976
976
976
1953
1953
15625
Memory= m∗ #nodes
3.467
3.997
2.326
2.706
3.7
1.03
Table 2: Various parameters settings in Bloom Sample Tree implementation for n = 103 and M = 106
(Memory in MBs)
Accuracy
0.5
0.6
0.7
0.8
0.9
1.0
m
63120
72475
84215
101090
132933
297485
Depth
13
13
13
13
12
10
M⊥
1220
1220
1220
1220
2441
9765
Memory= m∗ #nodes
61.62
70.75
82.22
98.69
64.87
36.27
Table 3: Various parameters settings in Bloom Sample Tree implementation for n = 103 and M = 107
(Memory in MBs)
21
100-BST
1K-BST
10K-BST
50K-BST
100-DA
1K-DA
10K-DA
50K-DA
625.0
Time (ms)
125.0
25.0
5.0
1.0
0.2
0.0
0.5
0.6
0.7
0.8
0.9
1.0
Sampling Accuracy
(a) Uniformly random query set
100-BST
1K-BST
10K-BST
50K-BST
100-DA
1K-DA
10K-DA
50K-DA
625.0
Time (ms)
125.0
25.0
5.0
1.0
0.2
0.0
0.5
0.6
0.7
0.8
0.9
1.0
Sampling Accuracy
(b) Clustered query set
Figure 6: Avg. time taken for sampling with M = 106
BST-MD5
DA-MD5
BST-Murmur
DA-Murmur
Time (mS)
625.0
125.0
25.0
5.0
1.0
0.5
0.6
0.7
0.8
0.9
1.0
Sampling Accuracy
Figure 7: Effect of different hash function families on performance
22
a
0.5
0.6
0.7
0.8
0.9
M=10ˆ5
M=10ˆ6
M=10ˆ7
m
#Levels
STime
UTime
m
#Levels
STime
Utime
m
#Levels
Stime
Utime
12317
14334
16863
20494
27340
4
3
3
3
3
0.12
0.04
0.28
0.12
0.12
18.4
14.4
14.8
15.6
15.9
28464
32807
38258
46090
60869
6
6
6
5
5
0.96
0.68
0.64
0.72
0.56
156.5
161.7
165.2
158.9
163.4
63119
72474
84214
101088
132932
8
8
8
8
8
78.72
223.52
311
426.52
590.64
1603.8
1625
1616.3
1646.3
1720.2
Table 4: System and User Time taken (in mS) to create BloomSampleTree for different values of M
and desired accuracy
Accuracy / n 100
1K 10K 50K
0.5
1
0.99 0.52 0.78
0.6
1
0.92 0.75 0.88
0.7
0.99 0.15 0.87 0.63
0.8
0.93 0.49 0.51 0.12
0.9
0.93 0.75 0.28 0.47
1.0
0.84 0.48 0.43 0.64
Table 5: p-values for M = 106
7.3
Reconstruction Experiments
The setup of the reconstruction experiments follow that of the sampling experiments only adding
HashInvert as a baseline. Figures 9 and 10 show the number of intersections and set membership
queries to reconstruct sets which are uniformly random and clustered, drawn from namespaces of
size M = 106 and M = 107 respectively.
For the number of intersections with sampling accuracy, we see a trend that is similar to the ones
in the sampling experiments, and for the same reasons. One may note that the HashInvert procedure
performs more membership queries than the BloomSampleTree, but fewer than the Dictionary Attack.
Despite this, the overall cost for HashInvert is the most as can be seen in Figures 11 and 12, which
show the time taken for reconstruction. The overall cost for HashInvert essentially depends on
the number of set or reset bits in the Bloom filter. If the Bloom filter is extremely dense, then
reconstructing with the help of only reset bits efficiently reconstructs the set, whereas if it is very
sparse, then one can reconstruct using the set bits. However, HashInvert is inefficient if neither of
these cases apply, as is evident from the line for ’HI-10K’, which sets about 50% of the bits in the
Bloom Filter. The cause for this is the fact that HashInvert iterates through an inverted set for
each set or reset bit in the Bloom filter. Since some of these values may already have been checked,
it does save some membership queries. However, given that the membership query is very fast for
simple hash functions, this does not directly translate into smaller running times.
Accuracy / M
0.5
0.6
0.7
0.8
0.9
1.0
105
0.522
0.692
0.710
0.823
0.921
0.970
106
0.497
0.621
0.691
0.793
0.907
0.997
107
0.535
0.591
0.696
0.810
0.906
0.948
Table 6: Measured Accuracies for Uniform query sets of size n = 103
23
Intersection
(x100)
Membership
(x105)
10
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
0
4
0
0.5
0.6
0.7
0.8
0.9
1.0
Membership
Intersection
(x103)
Precision
(a) Uniformly Random Query Set
500
400
300
200
100
0
120
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
90
60
30
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(b) Clustered Query Set
Figure 8: Avg. No. of operations in reconstructing for M = 105
8
Experiments with Real-world Data with Low-Occupancy
Namespace
So far we presented results for the settings when the namespace is a contiguous and fixed. Now
we turn our attention to more practical settings where the size of the namespace we need to handle
is only a small fraction of a much larger domain and potentially spread throughout it.
8.1
Setup
Dataset We made use of a 34-day Twitter crawl consisting of 144 million tweets. There
are a
total of 7.2 million user ids in this tweet set, but they are distributed in a namespace of 0, 2 × 109
(a little over 2.2 billion).
Varying the namespace fractions Note that even though there are only 7.2 million unique ids
in our dataset, they could be distributed across the entire namespace of 2.2 billion. Suppose, for
example, we built a BloomSampleTree with 256 leaves – that is, the range of 2.2 billion is effectively
divided into 256 equal-sized ranges (of which some could be empty depending on the distribution
of the unique ids). From this hypothetical BloomSampleTree, we construct namespaces of different
namespace fractions as follows:
24
Intersection
(x100)
Membership
(x105)
40
30
20
10
0
15
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
10
5
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
Membership Intersection
(x100)
(x105)
(a) Uniformly Random Query Set
40
30
20
10
0
15
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
10
5
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(b) Clustered Query Set
Figure 9: Avg. No. of operations in reconstructing for M = 106
• Uniform Namespace: Following our example above, suppose we want to construct a namespace of namespace fraction 0.2, we uniformly sample 52 of 256 leaves. This gives us a set of
ranges, the union of which only occupy 0.2 fraction of the total namespace.
• Clustered Namespace: Again, for a namespace fraction of 0.2, we need to sample 52 of 256
leaves, but in a clustered way. We use the same technique as explained in Section 7 (in that
case, we were generating clustered query sets).
We fixed the desired accuracy, as discussed in section 5.4, at 0.8. Therefore, our hypothetical
BloomSampleTree has a depth of 7, with a Bloom filter size m = 1.2 × 106 . Correspondingly, the
pruned-BloomSampleTree has the same depth and Bloom filter size, but the number of nodes (and
therefore the space occupancy) is much smaller.
Query Bloom filters We identified 24, 000 unique hash tags that occurred at least 1, 000 times
in our dataset. The sets of users tweeting a particular hashtag is used to construct a query Bloom
filter. We therefore constructed 24,000 query Bloom filters. However, when experimenting with
varying namespace fractions, we simply ignore ids which do not belong in the namespace currently
under consideration and construct query Bloom filters without them.
Metrics
We report the following metrics.
25
Membership Intersection
(x100)
(x106)
90
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
60
30
0
15
10
5
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
Membership Intersection
(x100)
(x106)
(a) Uniformly Random Query Set
90
100-BST
1K-BST
10K-BST
50K-BST
HI-100
HI-1K
HI-10K
HI-50K
DA
60
30
0
15
10
5
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(b) Clustered Query Set
Figure 10: Avg. No. of operations in reconstructing for M = 107
• Average Time taken. At each namespace fraction, we run 1,000 sampling rounds on randomly chosen query Bloom filters and report the average time taken to generate a sample.
• Memory. The Pruned-BloomSampleTree occupies much less space than the full BloomSampleTree. We report on space usage at each namespace fraction.
• Accuracy. While the value of m, the Bloom filter size, was based on a desired accuracy for the
BloomSampleTree, the actual accuracy in a Pruned-BloomSampleTree is expected to be better,
since only those elements which occupy the namespace are stored. We report this accuracy for
various namespace fractions.
8.2
Sampling Experiments
Average time taken Figure 13 shows the average time taken to generate samples from our query
Bloom filters. At namespace fractions less than 0.1, the time taken is an order of magnitude smaller
than at full namespace occupancy. It is also expected that the sampling time in case of the clustered
namespace is smaller, since more leaves share common ancestors and there are far less paths in the
BloomSampleTree for the sampling algorithm to follow. The Dictionary Attack requires 100 seconds
on average for one sample to be drawn. This is natural since the size of the namespace is extremely
large in this case. As a result, we have not included the result of DA in Figure 13 to ensure that the
finer variations in the sampling time taken for random and clustered namespaces are clearly visible.
26
HI-100
HI-10K
BST-100
BST-10K
DA-100
DA-10K
600
Time (ms)
500
400
300
200
100
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(a) Uniform Random Query Set
HI-100
HI-10K
BST-100
BST-10K
DA-100
DA-10K
600
Time (ms)
500
400
300
200
100
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(b) Clustered Query Set
Figure 11: Avg. time taken for reconstruction with M = 106 for uniformly random and clustered
query sets.
Memory Figure 14 shows the memory usage at varying namespace fractions. Note that, if we built
the full BloomSampleTree for a namespace of 2.2 billion, the memory required would be approximately
36M B. In contrast, at a lower namespace fraction of 0.5, the memory usage of the BloomSampleTree
is about 71% for the uniform case, and much lower at 21.7% for the clustered case. For the same
reason as for sampling time, we expect the memory requirement of the BloomSampleTree to be
smaller for a clustered namespace.
Accuracy Figure 15 shows the sampling accuracy at various namespace fractions. Recall that we
had optimized the BloomSampleTree for an accuracy of 0.8. But, with our Pruned-BloomSampleTree,
we uniformly see a higher accuracy. Accuracy depends on the size of the namespace, as mentioned
in section 5.4, and the size of the effective namespace at a lower namespace fraction is smaller. This
shows that the BloomSampleTree is capable of producing higher accuracy results when the overall
namespace is large but the actually occupied effective namespace is small.
9
Conclusions
In this paper we described an efficient method to do sampling and reconstruction of sets stored
in Bloom filters. In particular, we described the BloomSampleTree data structure and analyzed its
27
BST-100
BST-10K
HI-100
HI-10K
DA-100
DA-10K
Time (ms)
2500
2000
1500
1000
500
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(a) Uniform Random Query Set
BST-100
BST-10K
HI-100
HI-10K
DA-100
DA-10K
Time (ms)
2500
2000
1500
1000
500
0
0.5
0.6
0.7
0.8
0.9
1.0
Precision
(b) Clustered Query Set
Figure 12: Avg. time taken for reconstruction with M = 107 for uniformly random and clustered
query sets.
properties both theoretically and experimentally. We compared our technique to the brute force
approach (Dictionary Attack) as well as HashInvert (useful when using invertible hash functions to
reconstruct sets). An extensive evaluation of our algorithm in various settings demonstrated its wide
applicability and significant advantages.
References
[1] B. H. Bloom, “Space/time trade-offs in hash coding with allowable errors,” Commun. ACM,
vol. 13, no. 7, pp. 422–426, 1970.
[2] A. Z. Broder and M. Mitzenmacher, “Network applications of bloom filters: A survey,” Internet
Math., vol. 1, no. 4, pp. 485–509, 2003.
[3] D. M. Romero, B. Meeder, and J. Kleinberg, “Differences in the mechanics of information
diffusion across topics: idioms, political hashtags, and complex contagion on twitter,” in WWW,
2011.
[4] R. Ghosh and K. Lerman, “A framework for quantitative analysis of cascades on networks,” in
WSDM, 2011.
28
Uniform
Time(s)
8
Clustered
4
0
-4
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Namespace Fraction
Memory (MB)
Figure 13: Time taken to generate a uniform sample for varying namespace fractions
36
32
28
24
20
16
12
8
4
0
Uniform
0
Clustered
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Namespace Fraction
Figure 14: Memory usage at varying namespace fractions
Uniform
1.1
Clustered
Accuracy
1
0.9
0.8
0.7
0.6
0.5
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Namespace Fraction
Figure 15: Sampling accuracy at varying namespace fractions
29
[5] J. Cheng, L. A. Adamic, P. A. Dow, J. M. Kleinberg, and J. Leskovec, “Can cascades be
predicted?” in WWW, 2014.
[6] J. E. R. MacMillan, W. B. Glisson, and M. Bromby, “Investigating the increase in mobile phone
evidence in criminal activities,” in HICSS, 2013, pp. 4900–4909.
[7] G. Cormode, S. Muthukrishnan, and I. Rozenbaum, “Summarizing and mining inverse distributions on data streams via dynamic inverse sampling,” in VLDB, 2005.
[8] M. Monemizadeh and D. P. Woodruff, “1-pass relative-error lp -sampling with applications,” in
SODA, 2010.
[9] H. Jowhari, M. Saglam, and G. Tardos, “Tight bounds for lp samplers, finding duplicates in
streams, and related problems,” in PODS, 2011.
[10] S. Tarkoma and C. E. Rothenberg and E. Lagerspetz, “Theory and Practice of Bloom Filters
for Distributed Systems,” IEEE Comm. Surveys and Tutorials, vol. 14, no. 1, 2012.
[11] S. Bellovin and W. Cheswick, “Privacy-Enhanced Searches using Encrypted Bloom Filters,”
Columbia University, Tech. Rep. CUCS-034-07, 2007.
[12] M. Naor and E. Yogev, “Bloom filters in adversarial environments,” CoRR, vol. abs/1412.8356,
2014. [Online]. Available: http://arxiv.org/abs/1412.8356
[13] G. Cormode, M. N. Garofalakis, P. J. Haas, and C. Jermaine, “Synopses for massive data:
Samples, histograms, wavelets, sketches,” Found. Trends. Databases, vol. 4, no. 1-3, 2012.
[14] M. Yoon, J. Son, and S.-H. Shin, “Bloom tree: A search tree based on Bloom filters for multipleset membership testing,” in Proc. 2014 IEEE Conference on Computer Communications (INFOCOM ’14), 2014, pp. 1429–1437.
[15] A. Crainiceanu and D. Lemire, “Bloofi: Multidimensional bloom filters,” Inf. Syst., vol. 54, pp.
311–324, 2015.
[16] M. Athanassoulis and A. Ailamaki, “Bf-tree: approximate tree indexing,” Proc. VLDB, vol. 7,
no. 14, pp. 1881–1892, October 2014.
[17] D. Guo, J. Wu, H. Chen, Y. Yuan, and X. Luo, “The dynamic bloom filters,” IEEE Trans.
Knowl. Data Eng., vol. 22, no. 1, pp. 120–133, 2010.
[18] M. C. Jeffrey and J. G. Steffan, “Understanding bloom filter intersection for lazy address-set
disambiguation,” in SPAA. New York, NY, USA: ACM, 2011, pp. 345–354.
[19] J. Vitter, “Random sampling with a reservoir,” ACM Trans. Math. Softw., vol. 11, no. 1, pp.
37–57, 1985. [Online]. Available: http://doi.acm.org/10.1145/3147.3165
[20] O. Papapetrou, W. Siberski, and W. Nejdl, “Cardinality estimation and dynamic length adaptation for bloom filters,” Dist. and Parallel Databases, vol. 28, no. 2-3, pp. 119–156, 2010.
[Online]. Available: http://dx.doi.org/10.1007/s10619-010-7067-2
[21] M. Mitzenmacher, “Compressed bloom filters,” in PODC.
[22] K. B. Athreya and P. E. Ney, Branching Processes.
New York, NY, USA: ACM, 2001.
Springer, 1972.
[23] P. Boldi, “Algorithmic gems in the data miner’s cave,” in Proc. Fun with Algorithms. Springer,
2014, pp. 1–15. [Online]. Available: http://dx.doi.org/10.1007/978-3-319-07890-8_1
[24] D. H. Stamatis, Six sigma and beyond: design for six sigma.
30
CRC Press, 2002, vol. 6.
| 8cs.DS
|
Degree-constrained 2-partitions of graphs
J. Bang-Jensen∗
Stéphane Bessy†
arXiv:1801.06216v1 [cs.DS] 18 Jan 2018
January 22, 2018
Abstract
A (δ ≥ k1 , δ ≥ k2 )-partition of a graph G is a vertex-partition (V1 , V2 ) of G satisfying that
δ(G[Vi ]) ≥ ki for i = 1, 2. We determine, for all positive integers k1 , k2 , the complexity of deciding
whether a given graph has a (δ ≥ k1 , δ ≥ k2 )-partition.
We also address the problem of finding a function g(k1 , k2 ) such that the (δ ≥ k1 , δ ≥ k2 )-partition
problem is N P-complete for the class of graphs of minimum degree less than g(k1 , k2 ) and polynomial for all graphs with minimum degree at least g(k1 , k2 ). We prove that g(1, k) = k for k ≥ 3,
that g(2, 2) = 3 and that g(2, 3), if it exists, has value 4 or 5.
Keywords: N P-complete, polynomial, 2-partition, minimum degree.
1
Introduction
A 2-partition of a graph G is a partition of V (G) into two disjoint sets. Let P1 , P2 be two graph
properties, then a (P1 , P2 )-partition of a graph G is a 2-partition (V1 , V2 ) where V1 induces a graph
with property P1 and V2 a graph with property P2 . For example a (δ ≥ k1 , δ ≥ k2 )-partition of a
graph G is a 2-partition (V1 , V2 ) where δ(G[Vi ]) ≥ ki , for i = 1, 2
There are many papers dealing with vertex-partition problems on (di)graphs. Examples (from a
long list) are [1, 2, 4, 6, 7, 8, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 24, 26, 27, 28, 29, 30, 31, 33, 34].
Examples of 2-partition problems are recognizing bipartite graphs (those having has a 2-partition into
two independent sets) and split graphs (those having a 2-partition into a clique and an independent
set) [15]. It is well known and easy to show that there are linear algorithms for deciding whether a
graph is bipartite, respectively, a split graph. It is an easy exercise to show that every graph G has a
2-partition (V1 , V2 ) such that the degree of each vertex in G[Vi ], i ∈ [2] is at most half of its original
degree. Furthermore such a partition can be found efficiently by a greedy algorithm. In [16, 17]
and several other papers the opposite condition for a 2-partition was studied. Here we require the
that each vertex has at least half of its neighbours inside the set it belongs to in the partition. This
problem, known as the satisfactory partition problem, is N P-complete for general graphs [5].
A partition problem that has received particular attention is that of finding sufficient conditions
for a graph to possess a (δ ≥ k1 , δ ≥ k2 )-partition. Thomassen [31] proved the existence of a function
f (k1 , k2 ) so that every graph of minimum degree at least f (k1 , k2 ) has a (δ ≥ k1 , δ ≥ k2 )-partition.
He proved that f (k1 , k2 ) ≤ 12 · max{k1 , k2 }. This was later improved by Hajnal [19] and Häggkvist
(see [31]). Thomassen [31, 32] asked whether it would hold that f (k1 , k2 ) = k1 + k2 + 1 which would
be best possible because of the complete graph Kk1 +k2 +1 . Stiebitz [29] proved that indeed we have
f (k1 , k2 ) = k1 + k2 + 1. Since this result was published, several groups of researchers have tried to find
extra conditions on the graph that would allow for a smaller minimum degree requirement. Among
others the following results were obtained.
Theorem 1.1 [20] For all integers k1 , k2 ≥ 1 every triangle-free graph G with δ(G) ≥ k1 + k2 has a
(δ ≥ k1 , δ ≥ k2 )-partition.
∗ Department
of Mathematics and Computer Science, University of Southern Denmark, Odense DK-5230, Denmark
(email:[email protected]). This work was done while the first author was visiting LIRMM, Université de Montpellier,
France. Hospitality is gratefully acknowledged. The research of Bang-Jensen was supported by the Danish research
council under grant number 7014-00037B.
† LIRMM, Université de Montpellier, France (email: [email protected]).
1
Theorem 1.2 [26] For all integers k1 , k2 ≥ 2 every graph G with no 4-cycle and with δ(G) ≥ k1 +k2 −1
has a (δ ≥ k1 , δ ≥ k2 )-partition.
Theorem 1.3 [23]
• For all integers k1 , k2 ≥ 1, except for K3 , every graph G with no K4 − e and with δ(G) ≥ k1 + k2
has a (δ ≥ k1 , δ ≥ k2 )-partition.
• For all integers k1 , k2 ≥ 2 every triangle-free graph G in which no two 4-cycles share an edge
and with δ(G) ≥ k1 + k2 − 1 has a (δ ≥ k1 , δ ≥ k2 )-partition.
The original proof that f (k1 , k2 ) = k1 + k2 + 1 in [29] is not constructive and neither are those
of Theorems 1.2 and 1.3. In [7] Bazgan et al. gave a polynomial algorithm for constructing a (δ ≥
k1 , δ ≥ k2 )-partition of a graph with minimum degree at least k1 + k2 + 1 or at least k1 + k2 when the
input is triangle-free.
The main result of this paper is a full characterization of the complexity of the (δ ≥ k1 , δ ≥ k2 )partition problem.
Theorem 1.4 Let k1 , k2 ≥ 1 with k1 ≤ k2 be integers. When k1 + k2 ≤ 3 it is polynomial to decide
whether a graph has a 2-partition (V1 , V2 ) such that δ(G[Vi ]) ≥ ki for i = 1, 2. For all other values of
k1 , k2 it is NP-complete to decide the existence of such a partition.
A result in [13] implies that (δ ≥ 3, δ ≥ 3)-partition is N P-complete already for 4-regular graphs
and there are other results about the complexity of finding partitions with lower and/or upper bounds
on the degrees inside each partition, such as [5, 6, 16, 17, 34], but we did not find anything which
implies Theorem 1.4.
The result of Stiebitz [29] insures that if the minimum degree of the input graph is large enough,
at least k1 + k2 + 1, then the (δ ≥ k1 , δ ≥ k2 )-partition always exists. We conjecture that if this
minimum degree is large but less than k1 + k2 + 1 then the (δ ≥ k1 , δ ≥ k2 )-partition is not always
trivial but can be solved in polynomial time.
Conjecture 1.5 There exists a function g(k1 , k2 ) so that for all 1 ≤ k1 ≤ k2 with k1 + k2 ≥ 3 the
(δ ≥ k1 , δ ≥ k2 )-partition problem is N P-complete for the class of graphs of minimum degree less than
g(k1 , k2 ) and polynomial for all graphs with minimum degree at least g(k1 , k2 ).
In the next section we introduce notions and tools that will be used later. In Section 3 we give
the proof of Theorem 1.4 and in Section 4 we provide some partial results concerning Conjecture 1.5.
In particular we prove that g(1, k) = k for k ≥ 3, that g(2, 2) = 3 and that g(2, 3), if it exists, has
value 4 or 5. Finally in Section 5 we address some other partition problems mainly dealing with
(edge-)connectivity in each part of the partition.
Notice that regarding the results that we establish in this paper the first open case of Conjecture 1.5
is the following problem.
Problem 1.6 What is the complexity of the (δ ≥ 2, δ ≥ 3)-partition problem for graphs of minimum
degree 4?
2
Notation, definitions and preliminary results
Notation is standard and follows [3, 12]. In this paper graphs have no parallel edges and no loops.
We use the shorthand notation [k] for the set {1, 2, . . . , k}.
2.1
Special Graphs
We first define some graphs that will be used frequently in our proofs to ensure that certain vertices
have a sufficiently high degree.
For all k ≥ 3 we let Xk,2 be the graph we obtain from Kk+1 by subdividing one edge by a vertex
x. Let X3,1 be obtained from X3,2 by adding a vertex x0 adjacent to the degree 2 vertex x of X3,2 .
2
Let Y4,1 be the graph on 7 vertices which we obtain from a 5-wheel by adding a new edge e linking 2
non adjacent vertices of the outer cycle of the 5-wheel, a new vertex joined to the 3 vertices of this
outer cycle not incident to e and to another new vertex y. For k = 3 let Zk be the graph X3,1 that we
defined above and let z = x0 . For k ≥ 4 let Zk be the graph that we obtain from Kk−2,k−1 by adding
a cycle on the k − 1 vertices of degree k − 2 and then adding a new vertex z adjacent to all the k − 2
vertices of degree k − 1. And finally let Wk be the graph we obtain from Kk+1 by deleting one edge
u0 v 0 and the adding two new vertices u, v and the edges uu0 , vv 0 .
All these graphs are depicted in Figure 1.
y
x0
x
Y4,1
X3,1
X3,2
x
z
u
v
0
v0
u
k − 2 vertices
Xk,2
Zk
Kk+1
Wk
Kk+1
k − 1 vertices
Figure 1: The graphs Y4,1 , X3,2 , X3,1 , Zk , Xk,2 and Wk .
2.2
Connected instances of 3-SAT
For a given instance F of 3-SAT with clauses C1 , C2 , . . . , Cm and variables x1 , . . . , xn , where each
variable xi occurs at least once as the literal xi and at least once as x̄i , we define the bipartite graph
B(F) as the graph with vertex set {v1 , v¯1 , v2 , v̄2 , . . . , vn , v̄n } ∪ {c1 , c2 , . . . , cm }, where the first set
corresponds to the literals and the second one to the clauses, and edge set containing an edge between
the vertex cj and each of the 3 vertices corresponding to the literals of Cj for every j ∈ [m]. We say
that F is a connected instance if B(F) is connected.
Lemma 2.1 3-SAT is N P-complete for instances F where B(F) is connected
Proof: Suppose B(F) has connected components X1 , X2 , . . . , Xk , where k ≥ 2. Fix a literal
0
vertex `i ∈ Xi for each i ∈ [k], add a new variable y and k − 1 new clauses C10 , . . . , Ck−1
where
0
0
Cj = (`j−1 ∨ y ∨ `j ), j ∈ [k − 1]. Let F be the new formula obtained by adding the variable y and the
0
. It is easy to check that F 0 is equivalent to F and that B(F 0 ) is connected.
clauses C10 , . . . , Ck−1
By adding a few extra variables, if necessary, we can also obtain an equivalent connected instance
in which each literal occurs at least twice. We leave the easy details to the interested reader.
2.3
Ring graphs and 3-SAT
We first introduce an important class of graphs that will play a central role in our proofs. The
directed analogue of these graphs was used in [2, 4]. A ring graph is the graph that one obtains
by taking two or more copies of the complete bipartite graph on 4 vertices {a1 , a2 , b1 , b2 } and edges
{a1 b1 , a1 b2 , a2 b1 , a2 b2 } and joining these in a circular manner by adding a path Pi,1 from the vertex
3
bi,1 to ai+1,1 and a path Pi,2 from bi,2 to ai+1,2 where bi,1 is the ith copy of b1 etc and indices are
’modulo’ n (bn+1,j = b1,j for j ∈ [2] etc). Our proofs are all reductions from N P-complete variants of
the 3-SAT problems. We call the copies of {a1 b1 , a1 b2 , a2 b1 , a2 b2 } switch vertices.
We start by showing how we can associate a ring graph to a given 3-SAT formula. Let F =
C1 ∧ C2 ∧ . . . ∧ Cm be an instance of 3-SAT consisting on m clauses C1 , . . . , Cm over the same set
of n boolean variables x1 , . . . , xn . Each clause Ci is of the form Ci = (`i,1 ∨ `i,2 ∨ `i,3 ) where each
`i,j belongs to {x1 , x2 , . . . , xn , x̄1 , x̄2 , . . . , x̄n } and x̄i is the negation of variable xi . By adding extra
clauses to obtain an equivalent formula, if necessary, we can ensure that every literal occurs at least
twice in F. We shall use this fact in one of our proofs.
For each variable xi the ordering of the clauses above induces an ordering of the occurrences of
xi , resp x̄i , in the clauses. Let qi (resp. pi ) denote the number of times xi (resp. x̄i ) occurs in the
clauses Let R(F) = (V, E) be the ring graph defined as follows. Its vertex set is
V = {a1,1 , . . . an,1 , a1,2 . . . , an,2 } ∪ {b1,1 , . . . , bn,1 , b1,2 , . . . , bn,2 } ∪
n
[
0
0
{vi,1 , . . . , vi,qi , vi,1
, . . . , vi,p
}
i
i=1
Its edge set E consists of the following edges:
Sn
• i=1 {ai,1 bi,1 , ai,1 bi,2 , ai,2 bi,1 , ai,2 bi,2 }
• the edges of the paths Pi,1 , Pi,2 , i ∈ [n] where Pi,1 = bi,1 vi,1 . . . vi,qi ai+1,1 and
0
0
Pi,2 = bi,2 vi,1
. . . vi,p
a
.
i i+1,2
For 1 ≤ j ≤ m, we associate the clause Cj = (`j,1 ∨ `j,2 ∨ `j,3 ) with the set Oj consisting of three
vertices of R(F) representing the occurrences of the literals of Cj in F: if `j,1 = xi for some i ∈ [n]
and this is the r’th occurrence of xi in the clauses, then Oj contains the vertex vi,r . If `j,1 = x̄i for
0
some i ∈ [n] and this is the r’th occurrence of x̄i in the clauses, then Oj contains the vertex vi,r
.
The other two vertices of Oj are defined similarly. In our proofs below we will often add a vertex ci
adjacent to all the vertices of Oi for i ∈ [n]. An example is depicted in Figure 2.
a2,1
b2,1
v1,2
v2,1
v1,1
b1,1
0
v1,1
a2,2
b2,2
b1,2
c2
c1
a1,1
c4
a1,2
0
v2,1
c3
a3,2
a3,1
b3,2
b3,1
0
v3,1
v4,2
0
v4,1
0
v3,2
a4,2
b4,2
0
v3,3
v3,1
v4,1
b4,1
a4,1
Figure 2: The ring graph R(F) corresponding to the formula F = (x1 ∨ x4 ∨ x̄3 ) ∧ (x1 ∨ x̄2 ∨ x̄3 ) ∧
(x2 ∨ x̄3 ∨ x4 ) ∧ (x̄1 ∨ x3 ∨ x̄4 ). The grey boxes contain the switch vertices, the white vertices are the
variable vertices and we added the clauses vertices c1 , c2 , c3 and c4 (these are not part of the ring
graph R(F)).
4
The following observation which forms the base of many of our proofs is easy to prove (for a proof
of result a very similar to this see [4]).
Theorem 2.2 Let F be a 3-SAT formula and let R(F) be the corresponding ring graph. Then R =
R(F) contains a cycle C which intersects all the sets O1 , . . . , Om so that R − C is a cycle C 0 if and
only if F is a ‘Yes’-instance of 3-SAT.
3
Proof of Theorem 1.4
3.1
The case k1 + k2 ≤ 3
We start with a trivial observation.
Proposition 3.1 Every graph G with δ(G) ≥ 1 and at least 4 vertices has a (δ ≥ 1, δ ≥ 1)-partition
except if G is a star.
Proposition 3.2 There is a polynomial algorithm for testing whether a graph has a (δ ≥ 1, δ ≥ 2)partition
Proof: We try for every choice of adjacent vertices u, v whether there is a solution with u, v ∈ V1 .
Clearly G is a ’yes’-instance if and only if at least one of these O(n2 ) attempts will succeed. Hence
by starting with V1 = {u, v} and then moving vertices with at most one neighbour in V − V1 to V1 we
either end with a good partition or V1 = V in which case no partition exists for that choice {u, v}.
3.2
The case k1 = 1 and k2 ≥ 3
The following variant of satisfiability, which we call ≤ 3-SAT(3), is known to be NP-complete: Given
a boolean CNF formula F consisting of clauses C1 , C2 , . . . , Cm over variables x1 , x2 , . . . , xn such that
each clause has 2 or 3 literals, no variable occurs in more than 3 clauses and no literal appears more
than twice; decide whether F can be satisfied.
As we could not find a proper reference for a proof that ≤ 3-SAT(3) is N P-complete, we give one
here as it is presented on pages 281-283 in a set of course notes1 by Prof. Yuh-Dauh Lyuu, National
Taiwan University:
Assume that F is an instance of 3-SAT in which the variable x occurs a total of r ≥ 4 times in
the formula (as x or x̄) in clauses Ci1 , . . . , Cir . Introduce new variables x1 , . . . , xr and replace the
first occurrence of x (in Ci1 ) by x1 is x is not negated in Ci1 and otherwise replace it by x̄1 in Ci1 .
Similarly we replace the occurrence of x in Cij , j ≥ 2 by xj or x̄j . Finally we add the new clauses
(x̄1 ∨ x2 ) ∧ (x̄2 ∨ x3 ) ∧ . . . ∧ (x̄r ∨ x1 ). These clauses (which have size 2) will force all the variables
x1 , . . . , xr to take the same value under any satisfying truth assignment. Repeating this replacement
for all variables of the original formula F we obtain an equivalent instance F 0 of ≤ 3-SAT(3).
Below we will need another variant which we call ≤ 3-SAT(5) where clauses still have size 2 or 3
and each variable is allowed to occur at most 5 times and at most 3 times as the same literal. By
following the scheme above and for each original variable occurring at least 4 times adding r extra
clauses (x1 ∨ x̄2 ) ∧ (x2 ∨ x̄3 ) ∧ . . . ∧ (xr ∨ x̄1 ) we obtain an equivalent instance F 00 and because the
2r new clauses will form a cycle in the bipartite graph B(F 00 ) of F 00 it is easy to see that F 00 is
a connected instance of ≤ 3-SAT(5) if F is a connected instance of 3-SAT. Hence, by Lemma 2.1,
connected ≤ 3-SAT(5) is N P-complete.
Theorem 3.3 For all k ≥ 3 it is N P-complete to decide whether a graph has a (δ ≥ 1, δ ≥ k)partition.
Proof: We show how to reduce an instance of connected ≤ 3-SAT(5) to (δ ≥ 1, δ ≥ k)-partition
where k ∈ {3, 4} in polynomial time and then show how to extend the construction to higher values
of k. We start the construction for k = 3.
Below we will use several disjoint copies of the graphs X3,1 and X3,2 to achieve our construction.
Let F be a connected instance of ≤ 3-SAT(5) with clauses C1 , . . . , Cm and variables x1 , . . . xn . We
1 https://www.csie.ntu.edu.tw/
~lyuu/complexity/2008a/20080403.pdf
5
may assume that each of the 2n literals occur at least once in F (this follows from the fact that we
may assume this for any instance of normal 3-SAT and the reduction above to ≤ 3-SAT(5) preserves
this property). We will construct G = G(F) as follows:
• For each variable xi , i ∈ [n] we introduce three new vertices yi , vi , v¯i and two the edges yi vi , yi v̄i .
• For each i ∈ [n]: if the literal xi (x̄i ) occurs precisely once in F, then we identify vi (v̄i ) with
the vertex x in a private copy of X3,2 . If xi (x̄i ) occurs precisely twice, then we identify vi (v̄i )
with the vertex x0 in a private copy of X3,1 .
• Now we add new vertices c1 , . . . , cm , where ci corresponds to the clause Ci , i ∈ [m], and join
each cj by an edge to those (2 or 3) vertices from {v1 , . . . , vn , v̄1 , . . . , v̄n } which correspond to
its literals. If cj gets only two edges this way, we identify it with the vertex x0 in a private copy
of X3,1 .
• Add 2m new vertices z1 , z2 , . . . , z2m and the edges of the 2m-cycle z1 z2 , z2 z3 , . . . , z2m−1 z2m , z2m z1 .
• Finally we add, for each j ∈ [m] the edges cj z2j−1 , cj z2j .
We claim that G(F) has a (δ ≥ 1, δ ≥ 3)-partition if and only if F can be satisfied. Suppose first
that t is a satisfying truth assignment. Then it is easy to check that (V1 , V2 ) is a good 2-partition
if we take V1 to be the union of {y1 , . . . , yn } and the n vertices from {v1 , . . . , vn , v̄1 , . . . , v̄n } which
corresponds to the false literals. Note that if a vertex vi (v̄i ) is in V2 then it will have degree 3 via
its private copy of one of the graphs X1 , X2 or because the corresponding literal occurred 3 times in F.
Conversely assume that (V1 , V2 ) is a (δ ≥ 1, δ ≥ 3)-partition. Then we claim that we must have
all the vertices z1 , z2 , . . . , z2m in V2 : If one of these is in V1 , then they all are as they have degree
exactly 3. Clearly we also have {y1 , . . . , yn } ⊂ V1 . However, by construction, the literal and clause
vertices all have degree 3 and they induce a connected graph (here we use that the instance F has a
connected bipartite graph B(F)). Thus all of these vertices must be in V2 , but then each vertex yi
is isolated, contradiction. Hence all the vertices z1 , z2 , . . . , z2m are in V2 and this implies that all of
c1 , . . . , cm are also in V2 . The vertices y1 , . . . , yn are in V1 and hence, for each i ∈ [n], at least one of
the vertices vi , v̄i is also in V1 . Now define a truth assignment a follows. For each i ∈ [n]: If both vi
and v̄i are in V1 , or vi is in V2 we put xi true; otherwise we put xi false. Since each cj must have a
neighbour from {v1 , . . . , vn , v̄1 , . . . , v̄n } in V2 this is a satisfying truth assignment.
To obtain the construction for k = 4 we replace each copy of X3,1 above by a copy of Y4,1 , each
copy of X3,2 by a copy of X4,2 and identify each of the literal and clause vertices with the vertex y
in an extra private copy of Y4,1 . Finally we identify each vertex zt , t ∈ [2m] with the vertex y in a
private copy of Y4,1 . Now it is easy to see that we can complete the proof as we did for the case k = 3.
For all k ∈ {3 + 2a, 4 + 2a|a ≥ 1} we can increase the degree of all literal, clause and zj vertices by
2a by identifying these with the x vertices of a private copies of Xk,2 and repeat the proof above.
Corollary 3.4 The (δ ≥ 1, δ ≥ k)-partition problem is N P-complete for graphs of minimum degree
k − 1.
Proof: Recall that in our proof above the vertices corresponding to clauses must always belong
to V2 in any good partition (V1 , V2 ) hence if we connect each vertex yi , i ∈ [n] to c1 , . . . , ck−3 by edges
we obtain a graph of minimum degree k − 1 which has a good partition if and only if F is satisfiable
(The vertices yi , i ∈ [n] must belong to V1 as they have degree k − 1).
3.3
(δ ≥ k1 , δ ≥ k2 )-partition when 2 ≤ k1 ≤ k2
Theorem 3.5 For every choice of natural numbers 2 ≤ k1 ≤ k2 it is N P-complete to decide whether
a graph has a (δ ≥ k1 , δ ≥ k2 )-partition.
Proof: We show how to reduce 3-SAT to (δ ≥ k1 , δ ≥ k2 )-partition. Given an instance F of
3-SAT with clauses C1 , C2 , . . . , Cm and variables x1 , . . . , xn we proceed as follows. Start from a copy
of the ring graph R = R(F) and then add the following:
6
• If k2 ≥ 3 we identify each vertex of R with the vertex z in a private copy of Zk2 .
• For each i ∈ [n] add a new vertex ui and join it by edges to the vertices ai,1 , ai,2 . If k1 ≥ 3 we
identify ui with the vertex z in a private copy of Zk1 .
• For each i ∈ [n] add a new vertex u0i and join it by edges to the vertices bi,1 , bi,2 . If k1 ≥ 3 we
identify u0i with the vertex z in a private copy of Zk1 .
• For each j ∈ [m] we add a new vertex cj , identify cj with the vertex z in a private copy of Zk1
if k1 ≥ 3 and add three edges from cj to the three vertices of R which correspond to the literals
of Cm .
• Finally add a new vertex r and join this to all of the vertices in {u1 , . . . , un , u01 , . . . , u0n , c1 , . . . , cm }
via private copies of Wk1 by identifying the vertex u with r and v with the chosen vertex from
{u1 , . . . , un , u01 , . . . , u0n c1 , . . . , cm }.
We claim that the final graph G has a (δ ≥ k1 , δ ≥ k2 )-partition if and only if F is satisfiable.
First we make some observations about G:
• Every vertex v of R which is not a switch vertex has degree exactly k2 + 1 as it has degree 2 in
R(F), is adjacent to exactly one cj , j ∈ [m] and if k2 ≥ 3 then v has been identified with one
vertex z of a private copy of Zk2 .
• Switch vertices all have degree exactly k2 + 2.
• The vertices u1 , . . . , un , u01 , . . . , u0n have degree exactly k1 + 1.
• All vertices in copies of Za have degree exactly a when a ≥ 3.
• All vertices in copies of Wk1 have degree exactly k1
• All vertices cj , j ∈ [m] have degree k1 + 2.
• The vertex r has degree m + 2n which we may clearly assume is at least k1 .
For convenience in writing, below we define Z2 to be the empty graph so that we can talk about Za ’s
without having to condition this on a being at least 3. Suppose first that F is satisfiable. By Theorem
2.2 this means that R(F) has a cycle C which intersects the neighbourhood of each cj , j ∈ [m] and so
that R − C is another cycle C 0 . Now we let V1 consist of the vertices of C, their corresponding private
copies of Zk2 , all the vertices {u1 , . . . , un , u01 , . . . , u0n , c1 , . . . , cm } along with their private copies of Zk1
and finally the vertex r and the vertices of all copies of Wk1 that we used. Let V2 = V (G) − V1 , that
is V2 contains the vertices of C 0 and their private copies of Zk2 . It is easy to check that δ(G[V1 ]) ≥ k1
and that δ(G[V2 ]) ≥ k2 so (V1 , V2 ) is a good partition. Now assume that G has a good 2-partition
(V1 , V2 ). The way we connected r to the vertices in {u1 , . . . , un , u01 , . . . , u0n , c1 , . . . , cm } via copies of
Wk1 implies that these must belong to the same set Vi as r. If k1 < k2 this must be V1 and otherwise
we can rename the sets so that i = 1. Since each cj , j ∈ [m] has degree k1 + 2 at least one of the
vertices corresponding to a literal of Cj must belong to V1 . Suppose that some vertex corresponding
a literal ` is in V1 , then all the vertices of the path in R corresponding to that literal, including the
two end vertices which are switch vertices, must belong to V1 . This follows from the fact that all
these vertices have degree k2 + 1 and have a neighbour in {u1 , . . . , un , u01 , . . . , u0n , c1 , . . . , cm } ⊂ V1 .
Moreover if ai,j (resp. bi,j ) belongs to V2 then, as ai,j (resp. bi,j ) has degree k2 + 2 and ui (resp. u0i )
belongs to V1 , at least one of the vertices of {bi,1 , bi,2 } (resp. {ai,1 , ai,2 }) belongs to V2 . And as ui
(resp. u0i ) belongs to V1 , one of {ai,1 , ai,2 } (resp. {bi,1 , bi,2 }) must lie in V1 . So since V2 is not empty
this implies that the restriction of V2 to R is a cycle consisting of paths Q1 , . . . , Qn where Qi is either
the path Pi,1 or the path Pi,2 . Hence R[V (R) ∩ V1 ] is a cycle intersecting each of the neighbourhoods
of the vertices cj , j ∈ [m] and hence F is satisfiable by Theorem 2.2.
Combining the results of this section concludes the proof of Theorem 1.4.
7
4
Higher degrees
We study the borderline between polynomial and N P-complete instances of the partition problems.
That is, we try to see how close we can get to the bound k1 + k2 + 1 on the minimum degree and still
have an N P-complete instance. For k1 = 1 we can give the precise answer by combining Corollary
3.4 and the result below.
Proposition 4.1 There is a polynomial algorithm for checking whether a graph G of minimum degree
at least k has a (δ ≥ 1, δ ≥ k)-partition.
Proof: It suffices to see that we can test for a given edge uv of G whether there is a (δ ≥ 1, δ ≥ k)partition (V1 , V2 ) with u, v ∈ V1 . This is done by starting with V1 = {u, v} and then moving vertices
from V − V1 to V1 when these vertices do not have at least k neighbours in V − V1 . Note that this
process preserves the invariant that δ(G[V1 ]) ≥ 1. Hence if the process terminates before V1 = V we
have found the desired partition and otherwise we proceed to the next choice for an edge to start
from.
For the (δ ≥ 2, δ ≥ 2)-partition problem we can also give the precise borderline between polynomial
and N P-complete instances.
Proposition 4.2 There exists a polynomial algorithm for checking whether a given graph of minimum
degree at least 3 has a (δ ≥ 2, δ ≥ 2)-partition.
Proof: First test whether G has two disjoint cycles C1 , C2 . This can be done in polynomial time
[9, 25]. If no such pair exists G is a ’no’-instance, so assume that we found a pair of disjoint cycles
C1 , C2 . Now put the vertices of C1 in V1 and continue to move vertices of V − V1 − V (C2 ) to V1
if they have at least two neighbours in the current V1 . When this process stops the remaining set
V2 = V − V1 induces a graph of minimum degree at least 2, since the vertices we did not move have
at most one neighbour in V1 .
We now proceed to partitions where 2 ≤ k1 ≤ k2 and try to raise the minimum degree above k1
to see whether we can still prove N P-completeness.
Theorem 4.3 For every a ≥ 3 it is N P-complete to decide whether a graph of minimum degree a + 1
has a (δ ≥ a, δ ≥ a)-partition.
Proof: We give the proof for a = 3 and then explain how to extend it to larger a. Let F be an
instance of 3-SAT with n variable and m clauses C1 , . . . , Cm . Let R0 = R0 (F) be obtained from R(F)
by adding, for all i ∈ [n], an edge between all vertices at distance 2 in one of the paths Pi,1 , Pi,2 , i ∈ [n]
(that is, we replace each of these paths by their square). Now we construct the graph H = H(F)
starting from R0 as follows:
• For each j ∈ [m]: add two vertices cj,1 , cj,2 and join them to the vertices in R0 which correspond
to the literals of Cj .
• add the vertices of a 2m -cycle y1 y2 . . . y2m y1 .
• For each j ∈ [m] add the two edges y2j−1 cj,1 , y2j−1 cj,2 and the two edges y2j cj,2 , y2j , cj+1,1 ,
where cm+1,1 = c1,1 .
The resulting graph H has minimum degree 4 and we claim that H has a (δ ≥ 3, δ ≥ 3)-partition
if and only if F is satisfiable, which we know, by Theorem 2.2 and the previous proofs, where we
used the same approach, is if and only if the vertex set of R (which is the same as that of R0 ) can be
partitioned into two cycles C, C 0 so that C contains a neighbour of each of the vertices cj,1 , cj,2 , j ∈ [m].
Again the proof is easy when F is satisfiable: Let C, C 0 be as above and let V2 = V (C 0 ) and
V1 = V (H) − V2 . It is easy to check that δ(H[Vi ]) ≥ 3 for i = 1, 2, because C contains a neighbour
of each cj,1 , cj,2 , j ∈ [m] (and we assume that each literal appears at least twice in F to insure that
δ(H[V2 ]) ≥ 3). Suppose now that H has a (δ ≥ 3, δ ≥ 3)-partition (V1 , V2 ). Since adjacent vertices in
{y1 , y2 , . . . , y2m } have degree 4 and share a neighbour they must all belong to the same set Vi , i ∈ [2]
8
and this set must also contain all the vertices cj,1 , cj,2 , j ∈ [m]. Without loss of generality we have
i = 1. Thus V2 is a subset of V (R0 ). The vertices of R have degree at most 4 in R0 and the initial and
terminal vertex of each path Pi,1 or Pi,2 has degree 3. Using this is not difficult to see that if some
vertex of a path Pi,1 or Pi,2 is in V2 then all the vertices of that path and the two adjacent switch
vertices are in V2 . If there is some i ∈ [n] so that both of the vertices ai,1 , ai,2 or both of the vertices
bi,1 , bi,2 are in in V2 , then, using the observation we just made, all vertices of R0 would be in V2 which
is impossible. Similarly we can show that V1 cannot contain both of the vertices ai,1 , ai,2 or both of
the vertices bi,1 , bi,2 for some i ∈ [n]. Hence for each switch {ai,1 , ai,2 , bi,1 , bi,2 }, exactly one of the
vertices ai,1 , ai,2 and exactly one of the vertices bi,1 , bi,2 is in V2 . Now we see that the vertices in V1
and V2 both induces a cycle in R and as the vertices cj,1 , cj,2 have degree 2 outside R, the cycle in R
which is in V1 must contain a neighbour of each of cj,1 , cj,2 , j ∈ [m]. Hence, by Theorem 2.2, F is
satisfiable.
We obtain the result for higher values of a by induction where we just proved the base case a = 3
above. Assume we have already constructed Ha = Ha (F) with δ(Ha ) ≥ a + 1 such that Ha has a
(δ ≥ a, δ ≥ a)-partition if and only if F is satisfiable. Construct Ha+1 from two copies of Ha by joining
copies of the same vertex by an edge. It is easy to check that Ha+1 has a (δ ≥ a+1, δ ≥ a+1)-partition
if and only if Ha has a (δ ≥ a, δ ≥ a)-partition.
Theorem 4.4 Deciding whether a graph of minimum degree 3 has a (δ ≥ 2, δ ≥ 3)-partition is N Pcomplete.
Proof: Let X = X(F) be the graph we obtain by starting from the ring graph R(F) and then
adding the following:
• Add the vertices of a tree T whose internal vertices have all degree 3 and which has 2m + 4n
leaves denoted by
u1,1 , u1,2 . . . , un,1 , un,2 , u01,1 , u01,2 , . . . , u0n,1 , u0n,2 , c1,1 , c1,2 , c2,1 , c2,2 , . . . , cm,1 , cm,2 , where the pairs
ui,1 , ui,2 and u0i,1 , u0i,2 have the same parent in T for i ∈ [n] and so do each of the pairs cj,1 , cj,2 ,
j ∈ [m]. Here the vertices cj,1 , cj,2 correspond to the clause Cj
• Join each vertex cj,1 , cj,2 , j ∈ [m] to the 3 vertices in R which correspond to the literals which
correspond to Cj and add the edge cj,1 cj,2 .
0
0
0
0
• Add 4n new vertices wi,1 , wi,2 , wi,1
, wi,2
, i ∈ [n]. Add the edges wi,1 wi,2 , wi,1
wi,2
, i ∈ [n].
0
0
• For each i ∈ [n] add the edges wi,1 ai,1 , wi,2 ai,2 , wi,1
bi,1 , wi,2
bi,2 .
• For each i ∈ [n] join each of ui,1 and ui,2 by edges to the vertices wi,1 , wi,2 and add the edge
ui,1 ui,2 .
0
0
• For each i ∈ [n] join each of u0i,1 and u0i,2 by edges to the vertices wi,1
, wi,2
and add the edge
u0i,1 u0i,2 .
We first prove that in any (δ ≥ 2, δ ≥ 3)-partition (V1 , V2 ) of X all the vertices of T must be in
the same set Vi . Note that we can not have cj,1 and cj,2 in different sets of the partition because then
one of their 3 neighbours in V (R) must be in both sets. By a similar argument, for each i ∈ [n] the
vertices ui,1 and ui,2 must belong to the same set in the partition and the vertices u0i,1 and u0i,2 must
belong to the same set in the partition. It is easy to check that this implies our claim for T .
If F is satisfiable, then by Theorem 2.2, we can find vertex disjoint cycles C, C 0 in R such that C
contains a vertex corresponding to a literal of Cj for each j and V (R) = V (C) ∪ V (C 0 ). Now we let
V1 = V (C 0 ) and V2 = V (X) − V1 . It is easy to check that this is a (δ ≥ 2, δ ≥ 3)-partition because
C must contain exactly one of the vertices ai,1 , ai,2 and exactly one of the vertices bi,1 , bi,2 for each
i ∈ [n].
Suppose now that (V1 , V2 ) is a good partition of X. By the argument above we have V (T ) ⊂ Vi for
i = 1 or i = 2. This must be i = 2 since in the graph X − V (T ) all vertices except the switch vertices
have degree 2. Thus we have V (T ) ⊂ V2 and each of the vertices cj,1 , cj,2 , j ∈ [m] have a neighbour in
9
V (R) which is also in V2 . As in earlier proofs it is easy to check that if some vertex of one of the paths
Pi,1 , Pi,2 is in V2 then all vertices of that path are in V2 . As in the proof of the previous theorem, we
now conclude that for each switch {ai,1 , ai,2 , bi,1 , bi,2 }, exactly one of the vertices ai,1 , ai,2 and exactly
one of the vertices bi,1 , bi,2 is in V2 . Now we see that the vertices in V1 and V2 both induce a cycle in
R and as the vertices cj,1 , cj,2 have degree 2 outside R, the cycle in R which is in V2 must contain a
neighbour of each of cj,1 , cj,2 , j ∈ [m]. Hence, by Theorem 2.2, F is satisfiable.
Corollary 4.5 For every k ≥ 2 it is N P-complete to decide if a given graph with minimum degree at
least k + 1 has a (δ ≥ k, δ ≥ k + 1)-partition.
Proof: This follows by induction on k with Theorem 4.4 as the base case in the same way as we
proved the last part of Theorem 4.3.
Proposition 4.6 There is a polynomial algorithm for deciding whether a given graph of minimum
degree 5 has a (δ ≥ 2, δ ≥ 3)-partition.
Proof: Let G have δ(G) ≥ 5. By Theorem 1.1 and the algorithmic version of this result from [7]
we may assume that G has a 3-cycle C. Denote its vertex set by {a, b, c}. If δ(G[V − V (C)] ≥ 3)
we are done as we can take V1 = V (C), so assume there is a vertex d which adjacent to all vertices
of C. Then {a, b, c, d} induce a K4 . If δ(G[V − {a, b, c, d}) ≥ 2 we can take V2 = {a, b, c, d}, so
we can assume that there is a vertex e which is adjacent to all 4 vertices in {a, b, c, d} and now
{a, b, c, d, e} induce a K5 . Now if G[V − {a, b, c, d, e}] contains a cycle C 0 we can conclude by starting
with V2 = {a, b, c, d, e} and adding vertices of V − V (C 0 ) − {a, b, c, d, e} to V2 as long as there is one
with at least 3 neighbours in V2 . When the process stops (V − V2 , V2 ) is a good partition. Hence we
can assume that G[V − {a, b, c, d, e}] is acyclic. If one connected component of G[V − {a, b, c, d, e}] is
non trivial with a spanning tree T , then two leaves u, v of T will share a neighbour in {a, b, c, d, e}.
Without loss of generality this is e and now the K4 induced by a, b, c, d and the cycle formed by e, u, v
and the path between u and v in T are disjoint and we can find a good partition as we did above.
Hence if we have not found the partition yet we must have that G[V − {a, b, c, d, e}] is an independent
set I, all of whose vertices are joined to all vertices in {a, b, c, d, e}. If |I| ≥ 2 it is easy to find a good
partition consisting of a 3-cycle on a, b and one vertex from I as V1 and the remaining vertices as V2 .
Finally if |I| = 1 there is no solution.
5
Further 2-partition problems
In [31] Thomassen proved that every graph G of connectivity at least k1 + k2 − 1 and minimum degree
at least 4k1 + 4k2 + 1 has a 2-partition (V1 , V2 ) so that G[Vi ] is ki -connected for i = 1, 2.
It is natural to ask about the complexity of deciding whether a graph has a 2-partition (V1 , V2 )
with prescribed lower bounds on the (edge-)connectivity of G[Vi ], i ∈ [2].
We start with a simple observation.
Proposition 5.1 There exits a polynomial algorithm for deciding whether a given graph has a 2partition (V1 , V2 ) such that G[V1 ] is connected and G[V2 ] is 2-edge-connected.
Proof: Suppose first that G is not 2-edge-connected. If G has more than two connected components it is a ’no’-instance. If it has two components, it is a ’yes’-instance if and only if one of these
is 2-edge-connected. Hence we can assume that G is connected but not 2-edge-connected. Now it is
easy to see that there is a good partition if and only if the block-cutvertex tree of G has a nontrivial
block which is a leaf in the block-cutvertex tree. Thus assume below that G is 2-edge-connected. Now
consider an ear-decomposition (sometimes called a handle-decomposition) of G where we start from
an arbitrary cycle C. Let P be the last non-trivial ear that we add and let u, v be the end vertices of
P . Then V1 = V (P ) − {u, v} and V2 = V − V1 is a good partition.
Perhaps a bit surprisingly, if we require just a bit more for the connected part, the problem becomes
N P-complete.
Both N P-completeness proofs below use reductions from a given 3-SAT formula F so we only
describe the necessary modifications of R(F).
10
Theorem 5.2 It is N P-complete to decide whether an undirected graph G = (V, E) has a vertex
partition (V1 , V2 ) so that G[V1 ] is 2-edge-connected and G[V2 ] is connected and non-acyclic.
Proof: We add vertices and edges to R = R(F) as follows:
• For each clause Cj , j ∈ [m] we add a vertex cj and join it by three edges to the three literal
vertices of R corresponding to Cj (as we did in several proofs above).
• Add new vertices c01 , c02 , . . . , c0m and edges cj c0j , j ∈ [m].
• Add new vertices α1 , . . . , αn , α01 , . . . , α0n and the edges αi ai,1 , αi ai,2 , αi α0i , i ∈ [n]
• Add new vertices β1 , . . . , βn , β 01 , . . . , β 0n and the edges βi bi,1 , βi bi,2 , β i βi0 , i ∈ [n].
We claim that the resulting graph G has a vertex partition (V1 , V2 ) such that G[V1 ] is 2-edgeconnected and G[V2 ] is connected and non-acyclic if and only if F is satisfiable. Note that, by
construction, for every good partition every vertex of G which is not in R must belong to V2 . In
particular if a path Pi,j contains a vertex of V2 then all the vertices of Pi,j are in V2 . Since we want
G[V2 ] to be connected, the edges between αi , βi , i ∈ [n] and R imply that for every i ∈ [n] at most
one of the vertices ai,1 , ai,2 and a most one of the vertices bi,1 , bi,2 can belong to V1 . This implies that
exactly one of ai,1 , ai,2 and exactly one of the vertices bi,1 , bi,2 belong to V1 as otherwise V1 would
be empty. Now it is easy to check that the desired partition exists if and only if R contains a cycle
C 0 which uses precisely one of the paths Pi,1 , Pi,2 for i ∈ [n] and avoids at least one literal vertex for
every clause of F. Thus, by Theorem 2.2, F is satisfiable if and only if G has a good partition.
Since we can decide whether a graph has two vertex disjoint cycles in polynomial time [9, 25] the
following result, whose easy proof we leave to the interested reader, implies that it is polynomial to
decide whether a graph has a 2-partition into two connected and non-acyclic graphs.
Proposition 5.3 A graph G has a 2-partition (V1 , V2 ) such that G[Vi ] is connected and has a cycle
for i = 1, 2 if and only if G has a pair of disjoint cycles and either G is connected or it has exactly
two connected components, each of which contain a cycle.
Theorem 5.4 It is N P-complete to decide whether a graph G = (V, E) has a vertex partition (V1 , V2 )
so that each of G[V1 ] and G[V2 ] are 2-edge-connected.
Proof: Let F be a 3-SAT formula and let G = G(F) be the graph we constructed in the proof above.
Let G1 be the graph obtained by adding the following vertices and edges to G:
• add new vertices c00j , j ∈ [m], qj , j ∈ {0} ∪ [m] and γ
• add the edges c00j c0j , j ∈ [m]
• add the edges of the path q0 c001 q1 c002 q2 . . . c00m qm
• complete this path into a cycle W by adding the edges γq0 , γqm
• add an edge from γ to all vertices in {α01 , . . . , α0n , β 01 , . . . , β 0n }.
We claim that G0 has a vertex-partition into two 2-edge-connected graphs if and only if F is
satisfiable. First observe that in any good partition (V1 , V2 ) we must have all vertices of W inside V1
or V2 . This follows from the fact that each qi has degree 2 so it needs both its neighbours in the same
set. Without loss of generality, W is a cycle in V2 . After deleting the vertices of W we have exactly
the graph G in the proof of Theorem 5.2 above and it is easy to see that all vertices not in V (R) must
belong to V2 in any good partition. This implies that G0 has the desired vertex-partition if and only
if the graph G has a partition (V1 , V20 ) so that G[V1 ] is 2-edge-connected and G[V20 ] is connected and
non-acyclic. This problem is N P-complete by Theorem 5.2 so the proof is complete.
By inspecting the proof above it is not difficult to see that the following holds.
Theorem 5.5 It N P-complete to decide whether a graph has a 2-partition (V1 , V2 ) such that each of
the graphs G[Vi ], i = 1, 2 are 2-connected.
It may be worth while to try and extend the results of this section to higher (edge)-connectivities.
11
References
[1] N. Alon. Splitting digraphs. Combin. Prob. Comput., 15:933–937, 2006.
[2] J. Bang-Jensen, N. Cohen, and F. Havet. Finding good 2-partitions of digraphs II. Enumerable
properties. Theor. Comput. Sci., 640:1–19, 2016.
[3] J. Bang-Jensen and G. Gutin. Digraphs: Theory, Algorithms and Applications. Springer-Verlag,
London, 2nd edition, 2009.
[4] J. Bang-Jensen and F. Havet. Finding good 2-partitions of digraphs I. Hereditary properties.
Theor. Comput. Sci., 636:85–94, 2016.
[5] C. Bazgan, Z. Tuza, and D.Vanderpooten. The satisfactory partition problem. Discrete Appl.
Math., 154(8):1236–1245, 2006.
[6] C. Bazgan, Z. Tuza, and D. Vanderpooten. Degree-constrained decompositions of graphs:
Bounded treewidth and planarity. Theor. Comput. Sci., 355(3):389–395, 2006.
[7] C. Bazgan, Z. Tuza, and D. Vanderpooten. Efficient algorithms for decomposing graphs under
degree constraints. Discrete Applied Mathematics, 155(8):979 – 988, 2007.
[8] J. Bensmail. On the complexity of partitioning a graph into a few connected subgraphs. J.
Combin. Optim., 30:174–187, 2015.
[9] B. Bollobás. Extremal Graph Theory. Academic Press, London, 1978.
[10] M. Bonamy, K.K. Dabrowski, C. Feghali, M. Johnson, and D. Paulusma. Recognizing graphs
close to bipartite graphs. In 42nd International Symposium on Mathematical Foundations of
Computer Science, MFCS 2017, August 21-25, 2017 - Aalborg, Denmark, pages 70:1–70:14, 2017.
[11] M. Bonamy, K.K. Dabrowski, C. Feghali, M. Johnson, and D. Paulusma. Independent feedback
vertex sets for graphs of bounded diameter. Information Proc. Letters., 131:26–32, 2018.
[12] J.A. Bondy and U.S.R. Murty. Graph Theory, volume 244 of Graduate Texts in Mathematics.
Springer-Verlag, Berlin, 2008.
[13] V. Chvátal. Recognizing decomposable graphs. J. Graph Theory, 8(1):51–53, 1984.
[14] M.E. Dyer and A.M. Frieze. On the complexity of partitioning graphs into connected subgraphs.
Discrete. Appl. Math., 10:139–153, 1985.
[15] S. Földes and P. Hammer. Split graphs. Congress. Numer., 19:311–315, 1977.
[16] M. Gerber and D. Kobler. Algorithmic approach to the satisfactory graph partitioning problem.
European J. of Operation Research, 125:283–291, 2000.
[17] M.U. Gerber and D. Kobler. Classes of graphs that can be partitioned to satisfy all their vertices.
Australasian J. of Combin., 29:201–214, 2004.
[18] A. Grigoriev and R. Sitters. Connected feedback vertex set in planar graphs. In WG 2009:
Graph Theoretical concepts in Computer Science, volume 5911 of Lect. Notes Comp. Sci., pages
143–153. Springer Verlag, Berlin, 2009.
[19] P. Hajnal. Partition of graphs with condition on the connectivity and minimum degree. Combinatorica, 3:95–99, 1983.
[20] A. Kaneko. On decomposition of triangle-free graphs under degree constraints. J. Graph Theory,
27(1):7–9, 1998.
[21] D. Kühn and D. Osthus. Partitions of graphs with high minimum degree or connectivity. J.
Combin. Theory Ser. B, 88:29–43, 2003.
12
[22] H.-O. Le, V.B. Le, and H. Müller. Splitting a graph into disjoint induced paths or cycles. Discrete.
Appl. Math., 131:199–212, 2003.
[23] M. Liu and B. Xu. On partitions of graphs under degree constraints. Disc. Appl. Math., 226:87–93,
2017.
[24] M.H. Liu and B.G. Xu. Bipartition of graph under degree constraints. Science China Mathematics, 58:869–874, 2015.
[25] L. Lovasź. On graphs not containing independent circuits (in Hungarian). MatLapok, 16:289299,
1965.
[26] J. Ma and T. Yang. Decomposing C4 -free graphs under degree constraints. arXiv:1706.07292,
2017.
[27] N. Misra, G. Philip, V. Raman, S. Saurabh, and S. Sikdar. FPT algorithms for connected feedback
vertex set. J. Combin. Optim., 24:131–146, 2012.
[28] M. Stiebitz. Decomposition of graphs and digraphs. In KAM Series in Discrete MathematicsCombinatorics-Operations Research-Optimization 95-309. Charles University Prague, 1995.
[29] M. Stiebitz. Decomposing graphs under degree constraints. J. Graph Theory, 23:321–324, 1996.
[30] H. Suzuki, N. Takahashi, and T. Nishizeki. A linear algorithm for bipartion of biconnected graphs.
Inform. Process. Lett., 33:227–231, 1990.
[31] C. Thomassen. Graph decomposition with constraints on the connectivity and minimum degree.
J. Graph Theory, 7:165–167, 1983.
[32] C. Thomassen. Paths, circuits and subdivisions. In Selected topics in graph theory Vol. 3, pages
97–131. Academic Press, 1988.
[33] P. van’t Hof, D. Paulusma, and G.J. Woeginger. Partitioning graphs into connected parts. Theor.
Comput. Sci., 410:4834–4843, 2009.
[34] M. Xiao and H. Nagamochi. Complexity and kernels for bipartition into degree-bounded induced
graphs. Theor. Comput. Sci., 659:72–82, 2017.
13
| 8cs.DS
|
How Crossover Speeds Up Building-Block
Assembly in Genetic Algorithms
arXiv:1403.6600v2 [cs.NE] 26 Nov 2014
Dirk Sudholt
Department of Computer Science, University of Sheffield, United Kingdom
Abstract
We re-investigate a fundamental question: how effective is crossover in Genetic Algorithms in combining building blocks of good solutions? Although this has been
discussed controversially for decades, we are still lacking a rigorous and intuitive answer. We provide such answers for royal road functions and O NE M AX, where every bit is a building block. For the latter we show that using crossover makes every
(µ+λ) Genetic Algorithm at least twice as fast as the fastest evolutionary algorithm using only standard bit mutation, up to small-order terms and for moderate µ and λ.
Crossover is beneficial because it effectively turns fitness-neutral mutations into improvements by combining the right building blocks at a later stage. Compared to
mutation-based evolutionary algorithms, this makes multi-bit mutations more useful. Introducing
crossover changes the optimal mutation rate on O NE M AX from 1/n
√
to (1 + 5)/2 · 1/n ≈ 1.618/n. This holds both for uniform crossover and k-point
crossover. Experiments and statistical tests confirm that our findings apply to a broad
class of building-block functions.
Keywords
Genetic algorithms, crossover, recombination, mutation rate, runtime analysis, theory.
1
Introduction
Ever since the early days of genetic algorithms (GAs), researchers have wondered when
and why crossover is an effective search operator. It has been folklore that crossover
is useful if it can combine building blocks, i. e., schema of high fitness, to give better
solutions [30]. But, as Watson and Jansen [50] put it, there has been a considerable difficulty
in demonstrating this rigorously and intuitively.
Many attempts at understanding crossover have been made in the past. Mitchell,
Forrest, and Holland [30] presented so-called royal road functions as an example where,
supposedly, genetic algorithms outperform other search algorithms due to the use of
crossover. Royal roads divide a bit string into disjoint blocks. Each block makes a positive contribution to the fitness in case all bits therein are set to 1. Blocks thus represent
schemata, and all-ones configurations are building blocks of optimal solutions. However, the same authors later concluded that simple randomized hill climbers performed
better than GAs [17, 31].
The role of crossover has been studied from multiple angles, including algebra [42],
Markov chain models [49], infinite population models and dynamical systems (see [7,
Chapter 6] for an overview) and statistical mechanics (see, e. g. [15, Chapter 11]).
Also in biology the role of crossover is far from settled. In population genetics,
exploring the advantages of recombination, or sexual reproduction, is a famous open
question [5] and has been called “the queen of problems in evolutionary biology” by
Graham Bell [6] and others. Evolutionary processes were found to be harder to analyze than those using only asexual reproduction as they represent quadratic dynamical
systems [2, 40].
Recent work in population genetics has focussed on studying the “speed of adaptation”, which describes the efficiency of evolution, in a similar vein to research in
evolutionary computation [52, 53]. Furthermore, a new theory of mixability has been
proposed recently from the perspective of theoretical computer science [28, 29], arguing that recombination favours individuals that are good “mixers”, that is, individuals
that create good offspring when being recombined with others.
Several researchers recently and independently reported empirical observations
that using crossover improves the performance
of evolutionary algorithms (EAs) on
Pn
the simple function O NE M AX (x) = i=1 xi [24, 41], but were unable to explain why.
The fact that even settings as simple as O NE M AX are not well understood demonstrates
the need for a solid theory and serves as motivation for this work.
Runtime analysis has become a major area of research that can give rigorous evidence and proven theorems [35, 3, 18]. However, studies so far have eluded the most
fundamental setting of building-block functions. Crossover was proven to be superior
to mutation only on constructed artificial examples like Jumpk [20, 23] and “Real Royal
Road” functions [21, 43], the H-IFF problem [8], coloring problems inspired by the Ising
model from physics [16, 44]1 , computing unique input-output sequences for finite state
machines [26], selected problems from multi-objective optimization [39], and the allpairs shortest path problem [11, 48, 34]. H-IFF [8] and the Ising model on trees [44]
consist of hierarchical building blocks. But none of the above papers addresses singlelevel building blocks in a setting as simple as royal roads.
Watson and Jansen [50] presented a constructed building-block function and
proved exponential performance gaps between EAs using only mutation and a GA.
However, the definition of the internal structure of building blocks is complicated and
artificial, and they used a tailored multi-deme GA to get the necessary diversity. With
regard to the question on how GAs combine building blocks, their approach does not
give the intuitive explanation one is hoping for.
This paper presents such an intuitive explanation, supported by rigorous analyses. We consider royal roads and other functions
Pn composed of building blocks, such
as monotone polynomials. O NE M AX(x) = i=1 xi is a special case where every bit
is a building block. We give rigorous proofs for O NE M AX and show how the main
proof arguments transfer to broader classes of building-block functions. Experiments
support the latter.
Our main results are as follows.
1. We show in Section 3 that on O NE M AX every (µ+λ) GA with uniform crossover and
standard bit mutation is at least twice as fast as every evolutionary algorithm (EA)
that only uses standard bit mutations (up to small-order terms). More precisely, the
dominating term in the expected number of function evaluations decreases from
e·n ln n to e/2·n ln n. This holds provided that the parent population and offspring
population sizes µ and λ are moderate, so that the inertia of a large population does
not slow down exploitation. The reason for this speedup is that the GA can store
a neutral mutation (a mutation not altering the parent’s fitness) in the population,
along with the respective parent. It can then use crossover to combine the good
building blocks between these two individuals, improving the current best fitness.
1 For
bipartite graphs, the problem is equivalent to the classical Graph Coloring problem with 2 colors.
2
In other words, crossover can capitalize on mutations that have both beneficial and
disruptive effects on building blocks.
2. The use of uniform crossover leads to a shift in the optimal mutation rate on O NE M AX. We demonstrate this in Section 4 for a simple “greedy” (2+1) GA that always
selects parents among the current best individuals. While for mutation-based EAs
1/n is the optimal
mutation rate [55], the greedy (2+1) GA has an optimal mutation
√
rate of (1 + 5)/2 · 1/n ≈ 1.618/n (ignoring small-order terms). This is because
introducing crossover makes neutral mutations more useful and larger mutation
rates increase the chance of a neutral mutation. Optimality is proved by means of
a matching lower bound on the expected optimization time of the greedy (2+1) GA
that applies to all mask-based crossover operators (where each bit value is taken
from either parent). Using the optimal mutation rate, the expected number of function evaluations is 1.19n ln n ± O(n log log n).
3. These results are not limited to uniform crossover or the absence of linkage. Section 5 shows that the same results hold for GAs using k-point crossover, for arbitrary k, under slightly stronger conditions on µ and λ, if the crossover probability
pc is set to an appropriately small value.
4. The reasoning for O NE M AX carries over to other functions with a clear building
block structure. Experiments in Section 6 reveal similar performance differences as
on O NE M AX for royal road functions and random polynomials with unweighted,
positive coefficients. This is largely confirmed by statistical tests. There is evidence that findings also transfer to weighted building-block functions like linear
functions, provided that the population can store solutions with different fitness
values and different building blocks until crossover is able to combine them. This
is not the case for the greedy (2+1) GA, but a simple (5+1) GA is significantly faster
on random linear functions than the optimal mutation-based EA for this class of
functions, the (1+1) EA [55].
The first result, the analysis for uniform crossover, is remarkably simple and intuitive.
It gives direct insight into the working principles of GAs. Its simplicity also makes it
very well suited for teaching purposes.
This work extends a preliminary conference paper [46] with parts of the results,
where results were restricted to one particular GA, the greedy (2+1) GA. This extended
version presents a general analytical framework that applies to all (µ+λ) GAs, subject
to mild conditions, and includes the greedy (2+1) GA as a special case. To this end, we
provide tools for analyzing parent and offspring populations in (µ+λ) GAs, which we
believe are of independent interest.
Moreover, results for k-point crossover have been improved. The leading constant
2c
larger
in the upper bound for k-point crossover in [46] was by an additive term of 3+3c
than that for uniform crossover, for mutation rates of c/n. This left open the question
whether k-point crossover is as effective as uniform crossover for assembling building blocks in O NE M AX. Here we provide a new and refined analysis, which gives an
affirmative answer, under mild conditions on the crossover probability.
1.1
Related Work
Kötzing, Sudholt, and Theile [23] considered the search behaviour of an idealized GA
on O NE M AX, to highlight the potential benefits of crossover under ideal circumstances.
3
If a GA was able to recombine two individuals with equal fitness that√result from independent evolutionary lineages, the fitness gain can be of order Ω( n). The
√ idealized GA would therefore be able to optimize O NE M AX in expected time O( n) [23].
However, this idealization cannot reasonably be achieved in realistic EAs with common search operators, hence the result should be regarded an academic study on the
potential benefit of crossover.
A related strand of research deals with the analysis of the Simple GA on O NE M AX.
The Simple GA is one of the best known and best researched GAs in the field. It uses
a generational model where parents are selected using fitness-proportional selection
and the generated offspring form the next population. Neumann, Oliveto and Witt [33]
showed that the Simple GA without crossover with high probability cannot optimize
O NE M AX in less than exponential time. The reason is that the population typically contains individuals of similar fitness, and then fitness-proportional selection is similar to
uniform selection. Oliveto and Witt [38] extended this result to uniform crossover: the
Simple GA with uniform crossover and population size µ ≤ n1/8−ε , ε > 0, still needs
exponential time on O NE M AX. It even needs exponential time to reach a solution of
fitness larger than (1 + c) · n/2 for an arbitrary constant c > 0. In [37] the same authors
relaxed their condition on the population size to µ ≤ n1/4−ε . Their work does not exclude that crossover is advantageous, particularly since under the right circumstances
crossover may lead to a large increase in fitness (cf. [23]). But if there is an advantage,
it is not noticeable as the Simple GA with crossover still fails badly on O NE M AX.
One year after [46] was published, Doerr, Doerr, and Ebel [9] presented a groundbreaking result: they designed an EA that was
√ proven to optimise O NE M AX (and any
simple transformation thereof) in time O(n log n). This is a spectacular result as all
black-box search algorithms using only unbiased unary operators—operators modifying one individual only, and not exhibiting any inherent search bias—need time
Ω(n log n) as shown by Lehre and Witt [27]. So their EA shows that crossover can lower
the expected running time by more than a constant factor. They call their algorithm a
(1+(λ, λ)) EA: starting with one parent, it first creates λ offspring by mutation, with a
random and potentially high mutation rate. Then it selects the best mutant, and crosses
it λ times with the original parent, using parameterized uniform crossover (the probability of taking a bit from the first parent
√ is not always 1/2, but a parameter of the algorithm). This leads to a number of O(n log n) expected function evaluations, which can
be further decreased to O(n) with a scheme adapting λ according to the current fitness.
The (1+(λ, λ)) EA from [9] is very cleverly designed to work efficiently on O NE M AX and similar functions. It uses a non-standard EA design because of its two phases
of environmental selection. Other differences are that mutation is performed before
crossover, and mutation is not fully independent for all offspring: the number of flipping bits is a random variable determined as for standard bit mutations, but the same
number of flipping bits is then used in all offspring. The focus of this work is different
as our goal is to understand how standard EAs operate, and how crossover can be used
to speed up building-block assembly in commonly used (µ+λ) EAs.
2
Preliminaries
We measure the performance of the algorithm with respect to the number of function
evaluations performed until an optimum is found, and refer to this as optimization time.
For steady-state algorithms this equals the number of generations (apart from the initialization), and for EAs with offspring populations such as (µ+λ) EAs or (µ+λ) GAs
the optimization time is by a factor of λ larger than the number of generations. Note
4
that the number of generations needed to optimize a fitness function can often be easily decreased by using offspring populations or parallel evolutionary algorithms [25].
But this significantly increases the computational effort within one generation, so the
number of function evaluations is a more fair and widely used measure.
Looking at function evaluations is often motivated by the fact that this operation
dominates the execution time of the algorithm. Then the number of function evaluations is a reliable measure for wall clock time. However, the wall clock time might
increase when introducing crossover as an additional search operator. Also when increasing the mutation rate, more pseudo-random numbers might be required. Jansen
and Zarges [22] point out a case where this effect leads to a discrepancy between the
number of function evaluations and wall clock time. This concern must be taken seriously when aiming at reducing wall clock time. However, each implementation must
be checked individually in this respect [22]. Therefore, we keep this concern in mind,
but still use the number of function evaluations in the following.
3
Uniform Crossover Makes (µ+λ) EAs Twice as Fast
We show that, under mild conditions, every (µ+λ) GA is at least twice as fast as its
counterpart without crossover. For the latter, that is, evolutionary algorithms using
only standard bit mutation, the author recently proved the following lower bound on
the running time of a very broad class of mutation-based EAs [47]. It covers all possible
selection mechanisms, parent or offspring populations, and even parallel evolutionary
algorithms. We slightly rephrase this result.
Theorem 1 (Sudholt [47]). Let n ≥ 2. Every EA that uses only standard bit mutation with
mutation rate p to create new solutions has expected optimization time at least
min{ln n, ln(1/(p2 n))} − ln ln n − 3
p(1 − p)n
on O NE M AX and every other function with a unique optimum, if 2−n/3 ≤ p ≤
p = c/n, c > 0 constant, this is at least
√ 1
.
n log n
If
c · ec · n ln n · (1 − o(1)).
In fact, for O NE M AX the author proved that among all evolutionary algorithms
that start with one random solution and only use standard bit mutations the expected
number of function evaluations is minimized by the simple (1+1) EA [47, Theorem 13].
Also the mutation rate p = 1/n is the best possible choice for O NE M AX, leading to a
lower bound of
en ln n − en ln ln n − 3en.
For the special case of p = 1/n, Doerr, Fouz, and Witt [10] recently improved the above
bound towards en ln n − O(n).
We show that for a range of (µ+λ) EAs, as defined in the following, introducing
uniform crossover can cut the dominant term of the running time in half, for the standard mutation rate p = 1/n.
The only requirement on the parent selection mechanism is that selection does not
favor inferior solutions over fitter ones. Formally, for maximizing a fitness function f ,
∀x, y : f (x) ≥ f (y) ⇒ Prob(select x) ≥ Prob(select y).
5
(1)
This in particular implies that equally fit solutions are selected with the same probability. Condition (1) is satisfied for all common selection mechanisms: uniform selection, fitness-proportional selection, tournament selection, cut selection, and rank-based
mechanisms.
The class of (µ+λ) EAs covered in this work is defined in Algorithm 1. All
(µ+λ) EAs therein create λ offspring through crossover and mutation, or just mutation, and then pick the best out of the µ previous search points and the λ new offspring.
Algorithm 1: Scheme of a (µ+λ) GA with mutation rate p and uniform crossover with
crossover probability pc for maximizing f : {0, 1}n → R.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Initialize population P of size µ ∈ N u. a. r.
while true do
Let P 0 = ∅.
for i = 1, . . . , λ do
With probability pc do
Select x1 , x2 with an operator respecting (1).
Let y := uniform crossover(x1 , x2 ).
otherwise do
Select y with an operator respecting (1).
end
Flip each bit in y independently with probability p.
Add y to P 0 .
end
Let P contain the µ best individuals from P ∪ P 0 ; break ties towards
including individuals with the fewest duplicates in P ∪ P 0 .
end
In the case of ties, we pick solutions that have the fewest duplicates among the considered search points. This strategy has already been used by Jansen and Wegener [21]
in their groundbreaking work on Real Royal Roads; it ensures a sufficient degree of
diversity whenever the population contains different search points of the same fitness.
Before stating the main result of this section, we provide two lemmas showing
how to analyse population dynamics. Both lemmas are of independent interest and
may prove useful in other studies of population-based EAs.
The following lemma estimates the expected time until individuals with fitness at
least i take over the whole population. It generalizes Lemma 3 in [45], which in turn
goes back to Witt’s analysis of the (µ+1) EA [54]. Note that the lemma applies to arbitrary fitness functions, arbitrary values for µ and λ, and arbitrary crossover operators;
it merely relies on fundamental and universal properties of cut selection and standard
bit mutations.
Lemma 2. Consider any (µ+λ) GA implementing Algorithm 1, with any crossover operator,
on any n-bit fitness function. Assume the current population contains at least one individual
of fitness i. The expected number of function evaluations needed for the (µ+λ) GA before all
individuals in its current population have fitness at least i is at most
O((µ + λ) log µ)
.
(1 − pc )(1 − p)n
6
This holds for any tie-breaking rule used in the environmental selection.
Proof. Call an individual fit if it has fitness at least i. We now estimate the expected
number of generations until the population is taken over by fit individuals, which we
call the expected takeover time. As fit individuals are always preferred to non-fit individuals in the environmental selection, the expected takeover time equals the expected
number of generations until µ fit individuals have been created, starting with one fit
individual.
For each offspring being created, there is a chance that the (µ+λ) GA will simply
create a clone of a fit individual. This happens if, during the creation of an offspring,
the (µ+λ) GA decides not to perform crossover, it selects a fit individual as parent to be
mutated, and mutation does not flip any bit. The probability for this event is at least
(1 − pc ) · (1 − p)n ·
number of fit individuals in population
µ
since each fit individual is selected as parent with probability at least 1/µ.
Now we divide the run of the (µ+λ) GA into phases in order to get a lower bound
on the number of fit individuals at certain time steps. The j-th phase, 0 ≤ j ≤
dlog5 µe − 1, starts with the first offspring creation in the first generation where the
number of fit individuals is at least 5j . It ends in the first generation where this number
is increased to min{5j+1 , µ}. Let Tj describe the random number of generations spent
in the j-th phase. Starting with a new generation with µ ≥ 5j fit individuals in the parent population, we now consider a phase of 8µ/((1 − pc )(1 − p)n ) offspring creations,
disregarding generation bounds.
Let Ni denote the random number of new fit offspring created in the phase, then
E(Ni ) ≥
i
8µ
n 5
= 8 · 5i
·
(1
−
p
)(1
−
p)
·
c
(1 − pc )(1 − p)n
µ
and by classical Chernoff bounds (see, e. g. [32, Chapter 4])
i
Prob(Ni < 4 · 5i ) ≤ e−E(Ni )/8 ≤ e−5 ≤ e−1 .
If Ni < 4 · 5i the phase is called unsuccessful and we consider another phase of
8µ/((1 − pc )(1 − p)n ) offspring creations. The expected waiting time for a successful
phase is at most 1/(1 − e−1 ) and the expected number of offspring creations until
Ni ≥ 4 · 5i is at most 8µ/((1 − pc )(1 − p)n (1 − e−1 )).
Since phases start at generation bounds, we may need to account for up to λ − 1
further offspring creations in between phases. This implies
E(Ti ) ≤
8µ
+λ
(1 − pc )(1 − p)n (1 − e−1 )
and the expected takeover time is at most
dlog5 µe−1
X
i=0
E(Ti ) ≤ dlog5 µe ·
=
8µ
+
λ
(1 − pc )(1 − p)n (1 − e−1 )
O((µ + λ) log µ)
.
(1 − pc )(1 − p)n
7
We also provide the following simple but handy lemma, which relates success
probabilities for created offspring to the expected number of function evaluations
needed to complete a generation where such an event has first happened.
Lemma 3. Consider any (µ+λ) GA implementing Algorithm 1, and assume that in each offspring creation there is a probability at least q that some specific event occurs. Then the expected
number of function evaluations to complete a generation where this event first occurs is at most
1
λ−1+ .
q
Proof. The expected number of trials for an event with probability q to occur is 1/q. To
complete the generation, at most λ − 1 further function evaluations are required.
Now we are able to prove the main result of this section.
Theorem 4. The expected optimization time of every (µ+λ) GA implementing Algorithm 1
with 0 < pc < 1 constant, mutation probability 0 < p < 1 and µ ≥ 2 on O NE M AX is at most
O((µ + λ)n log µ)
ln(n2 p + n) + 1 + p
+
.
n−1
p(1 − p)
· (1 + np)
(1 − p)n
(2)
If p = c/n, c > 0 constant, and µ, λ = o((log n)/(log log n)), this bound simplifies to
n ln n
· (1 + o(1)).
c · e−c · (1 + c)
(3)
Both statements hold for arbitrary initial populations.
The main difference between the upper bound for (µ+λ) GAs and the lower bound
for all mutation-based EAs is an additional factor of 1 + pn in the denominator of the
upper bound. This is a factor of 2 for p = 1/n and an even larger gain for larger
mutation rates.
For the default value of p = 1/n, this shows that introducing crossover makes EAs
at least twice as fast as the fastest EA using only standard bit mutation. It also implies
that introducing crossover makes EAs at least twice as fast as their counterparts without
crossover (i. e. where pc = 0).
Proof of Theorem 4. Bound (3) can be derived from (2) using (1−1/x)x−1 ≥ 1/e for x > 1
to estimate
c n−1
c (n/c−1)·c
c c−1
1
c2
1−
= 1−
· 1−
≥ c · 1−
= e−c − O(1/n)
n
n
n
e
n
as well as ln(cn + n) + 1 + c/n = (ln n) + O(1). Note that (µ + λ)n log µ = o(n log n) by
conditions on µ, λ, hence this and all other small-order terms are absorbed in the term
o(1).
In order to prove the general bound (2), we consider canonical fitness levels, i. e.,
the i-th fitness level contains all search points with fitness i. We estimate the time spent
on each level i, i. e., when the best fitness in the current population is i. For each fitness level we consider three cases. The first case applies when the population contains
individuals on fitness levels less than i. The second case is when the population only
contains copies of a single individual on level i. The third case occurs when the population contains more than one individual on level i; then the population contains different
“building blocks” that can be recombined effectively by crossover.
8
All these cases capture the typical behaviour of a (µ+λ) GA, albeit some of these
cases, and even whole fitness levels, may be skipped. We obtain an upper bound on its
expected optimization time by summing up expected times the (µ+λ) GA may spend
in all cases and on all fitness levels.
Case i.1: The population contains an individual on level i and at least one individual on a lower fitness level.
A sufficient condition for leaving this case is that all individuals in the population
obtain fitness at least i. Since the (µ+λ) GA never accepts worsenings, the case is left
for good.
The time for all individuals reaching fitness at least i has already been estimated
in Lemma 2. Applying this lemma to all fitness levels i, the overall time spent in all
cases i.1 is at most
O((µ + λ)n log µ)
O((µ + λ)n log µ)
=
.
n
(1 − pc )(1 − p)
(1 − p)n
Case i.2: The population contains µ copies of the same individual x on level i.
In this case, each offspring created by the (µ+λ) GA will be a standard mutation
of x. This is obvious for offspring where the (µ+λ) GA decides not to use crossover. If
crossover is used, the (µ+λ) GA will pick x1 , x2 = x, create y = x by crossover, and
hence perform a mutation on x.
The (µ+λ) GA leaves this case for good if either a better search point is created or if
it creates another search point with i ones. In the latter case we will create a population
with two different individuals on level i. Note that due to the choice of the tie-breaking
rule in the environmental selection, the (µ+λ) GA will always maintain at least two
individuals on level i, unless an improvement with larger fitness is found.
The probability of creating a better search point in one mutation is at least (n − i) ·
p(1 − p)n−1 as there are n − i suitable 1-bit flips. The probability of creating a different
search point on level i is at least i(n − i) · p2 (1 − p)n−2 as it is sufficient to flip one of i
1-bits, to flip one of n − i 0-bits, and not to flip any other bit. The probability of either
event happening in one offspring creation is thus at least
(n − i) · p(1 − p)n−1 + i(n − i) · p2 (1 − p)n−2
≥ p(1 − p)n−1 · (n − i)(1 + ip).
By Lemma 3, the expected number of function evaluations in Case i.2 is at most
λ+
1
.
p(1 − p)n−1 · (n − i)(1 + ip)
The expected number of functions evaluations made in all cases i.2 is hence at most
λn +
n−1
X
i=0
= λn +
1
p(1 − p)n−1 · (n − i)(1 + ip)
n−1
X
1
1
·
.
n−1
p(1 − p)
(n − i)(1 + ip)
i=0
9
(4)
The last sum can be estimated as follows. Separating the summand for i = n − 1,
n−2
X
1
1
+
(n − i)(1 + ip) 1 + (n − 1)p
i=0
n−1
≤
Z
i=0
1
1+p
di +
.
(n − i)(1 + ip)
1 + np
We use equation 3.3.20 in [1] to simplify the integral and get
n−1
1+p
1 + ip
1
· ln
+
1 + np
n−i
1
+ np
0
ln(np + 1 − p) + ln(n)
1+p
=
+
1 + np
1 + np
ln(n2 p + n) + 1 + p
≤
.
1 + np
Plugging this into (4) yields that the expected time in all cases i.2 is at most
λn +
ln(n2 p + n) + 1 + p
.
p(1 − p)n−1 · (1 + np)
Case i.3: The population only contains individuals on level i, not all of which are
identical.
In this case we can rely on crossover recombining two different individuals on
level i. As they both have different “building blocks”, i. e., different bits are set to 1,
there is a good chance that crossover will generate an offspring with a higher number
of 1-bits.
The probability of performing a crossover with two different parents in one offspring creation is at least
µ−1
pc ·
µ2
as in the worst case the population contains µ − 1 copies of one particular individual.
Assuming two different parents are selected for crossover, let these have Hamming distance 2d and let X denote the number of 1-bits among these positions in
the offspring. Note that X is binomially distributed with parameters 2d and 1/2
and its expectation is d. We estimate the probability of getting a surplus of 1-bits as
this leads to an improvement in fitness. This estimate holds for any d ∈ N. Since
Prob(X < d) = Prob(X > d), we have
1
1
1
−2d 2d
Prob(X > d) = (1 − Prob(X = d)) =
1−2
≥ .
2
2
d
4
Mutation keeps all 1-bits with probability at least (1 − p)n . Together, the probability of
increasing the current best fitness in one offspring creation is at least
pc ·
µ − 1 (1 − p)n
·
.
µ2
4
By Lemma 3, the expected number of function evaluations in Case i.3 is at most
λ+
4µ2
.
pc · (µ − 1) · (1 − p)n
10
The total expected time spent in all cases i.3 is hence at most
λn +
O(µn)
4µ2 n
= λn +
n
pc · (µ − 1) · (1 − p)
(1 − p)n
as pc = Ω(1).
Summing up all expected times yields a total time bound of
ln(n2 p + n) + 1 + p
O(µn) + O((µ + λ)n log µ)
+ 2λn +
p(1 − p)n−1 · (1 + np)
(1 − p)n
2
ln(n p + n) + 1 + p
O((µ + λ)n log µ)
=
.
+
p(1 − p)n−1 · (1 + np)
(1 − p)n
Remark 1 (On conditions for µ and λ). The second statement of Theorem 4 requires
µ, λ = o((log n)/(log log n)) in order to establish the upper bound in (3). This condition
seems necessary as for larger values of µ and λ the inertia of a large population slows
down exploitation, at least in the absence of crossover. Note not all EAs covered by
Theorem 4 (after removing crossover) optimize O NE M AX in time O(n log n).
Witt [54] showed that a (µ+1) EA with uniform parent selection has an expected
optimization time of Ω(µn + n log n) on O NE M AX. For µ = ω(log n), this lower bound
is ω(n log n). Jansen, De Jong, and Wegener [19] showed that a (1+λ) EA needs time
ω(n log n) on O NE M AX if λ = ω((log n)(log log n)/(log log log n)). Badkobeh, Lehre,
and Sudholt [4] showed that every black-box algorithm creating λ offspring, using
only standard bit mutation or other unary unbiased operators, needs time ω(n log n)
on O NE M AX for λ = ω((log n)(log log n)). This indicates that the threshold in the condition µ, λ = o((log n)/(log log n)) is tight up to polynomials of log log n.
Remark 2 (On conditions for pc ). Theorem 4 assumes 0 < pc < 1 constant, which reflects
the most common choices in applications of EAs. The theorem can be extended towards
smaller or larger values as follows. If pc = o(1) the upper bound on the time spent in
Cases i.3 increases as it contains a factor of 1/pc . The other cases remain unaffected,
and if ((µ + λ) log µ)/pc = o(log n) we still get the upper bound from (3).
For high crossover probabilities, that is, pc = 1 − o(1) or pc = 1, only Cases i.1
need to be revisited. The time in those cases was derived from Lemma 2, which can
be adapted as follows: the probability for increasing the number of fit individuals is at
least
(number of fit individuals in population)2
pc · (1 − p)n ·
2µ2
as it suffices to select two fit individuals and generate an average or above-average
number of 1-bits in the offspring, which happens with probability at least 1/2. The
time bound from Lemma 2 then becomes
O(µ2 + λ log µ)
(1 − p)n
and the time bound in Theorem 4 becomes
ln(n2 p + n) + 1 + p
O(n(µ2 + λ log µ))
.
+
n−1
p(1 − p)
· (1 + np)
(1 − p)n
√
For p = c/n, c > 0 constant, and µ, λ = o( log n), this also establishes the upper bound
from (3).
11
It is remarkable that the waiting time for successful crossovers in Cases i.3 is only
of order O((µ + λ)n). For small values of µ and λ, e. g. µ, λ = O(1), the time spent in
all Cases i.3 is O(n), which is negligible compared to the overall time bound of order
Θ(n log n). This shows how effective crossover is in recombining building blocks.
Also note that the proof of Theorem 4 is relatively simple as it uses only elementary
arguments and, along with Lemmas 2 and 3, it is fully self-contained. The analysis
therefore lends itself for teaching purposes on the behavior of evolutionary algorithms
and the benefits of crossover.
Our analysis has revealed that fitness-neutral mutations, that is, mutations creating
a different search point of the same fitness, can help to escape from the case of a population with identical individuals. Even though these mutations do not immediately
yield an improvement in terms of fitness, they increase the diversity in the population.
Crossover is very efficient in exploiting this gained diversity by combining two different search points at a later stage. This means that crossover can capitalize on mutations
that have both beneficial and disruptive effects on building blocks.
An interesting consequence is that this affects the optimal mutation rate on O NE M AX. For EAs using only standard bit mutations Witt [55] recently proved that 1/n
is the optimal mutation rate for the (1+1) EA on all linear functions. Recall that the
(1+1) EA is the optimal mutation-based EA (in the sense of Theorem 1) on O NE M AX [47].
For mutation-based EAs on O NE M AX neutral mutations are neither helpful nor
detrimental. With crossover neutral mutations now become helpful. Increasing the
mutation rate increases the likelihood of neutral mutations. In fact, we can easily derive
better upper bounds from Theorem 4 for slightly larger mutation rates, thanks to the
additional term 1 + np in the denominator of the upper bound.
The dominant term in (3),
n ln n
−c
c · e · (1 + c)
√
is minimized for c being the golden ratio c = ( 5 + 1)/2 ≈ 1.618. This leads to the
following.
Corollary√5. The asymptotically best running time bound from Theorem 4 is obtained for
p = (1 + 5)/(2n). For this choice the dominant term in (3) becomes
√
e( 5+1)/2
√
· n ln n ≈ 1.19n ln n.
5+2
4
The Optimal Mutation Rate
Corollary 5 gives the mutation rate that yields the best upper bound on the running
time that can be obtained with the proof of Theorem 4. However, it does not establish
that this mutation rate is indeed optimal for any GA. After all, we cannot exclude that
another mutation rate leads to a smaller expected optimization time.
In the following, we show for a simple (2+1) GA (Algorithm 2) that the upper
bound from
√ Theorem 4 is indeed tight up to small-order terms, which establishes
p = (1+ 5)/(2n) as the optimal mutation rate for that (2+1) GA. Proving lower bounds
on expected optimization times is often a notoriously hard task, hence we restrict ourselves to a simple “bare-bones” GA that captures the characteristics of GAs covered by
Theorem 4 and is easy to analyze. The latter is achieved by fixing as many parameters
as possible.
12
As the upper bound from Theorem 4 grows with µ and λ, we pick the smallest
possible values: µ = 2 and λ = 1. The parent selection is made as simple as possible:
we select parents uniformly at random from the current best individuals in the population. In other words, if we define the parent population as the set of individuals that
have a positive probability to be chosen as parents, the parent population only contains
individuals of the current best fitness. We call this parent selection “greedy” because it
is a greedy strategy to choose the current best search points as parents.
In the context of the proof of Theorem 4 greedy parent selection implies that
Cases i.1 are never reached as the parent population never spans more than one fitness
level. So the time spent in these cases is 0. This also allows us to eliminate one further
parameter by setting pc = 1, as lower values for pc were only beneficial in Cases i.1.
Setting pc = 1 minimizes our estimate for the time spent in Cases i.3. So Theorem 4
extends towards this GA (see also Remark 2).
We call the resulting GA “greedy (2+1) GA” because its main characteristics is the
greedy parent selection. The greedy (2+1) GA is defined in Algorithm 22 .
Algorithm 2: Greedy (2+1) GA with mutation rate p for maximizing f : {0, 1}n → R.
1
2
3
4
5
6
7
Initialize population P of size 2 u. a. r.
while true do
Select x1 , x2 u. a. r. from {x ∈ P | ∀y ∈ P : f (x) ≥ f (y)}.
Let y := crossover(x1 , x2 ).
Flip each bit in y independently with probability p.
Let P contain the 2 best individuals from P ∪ {y}; break ties towards
including individuals with the fewest duplicates in P ∪ {y}.
end
The following result applies to the greedy (2+1) GA using any kind of mask-based
crossover. A mask-based crossover is a recombination operator where each bit value is
taken from either parent; that is, it is not possible to introduce a bit value which is not
represented in any parent. All common crossovers are mask-based crossovers: uniform
crossover, including parameterized uniform crossover, as well as k-point crossovers for
any k. The following result even includes biased operators like a bit-wise OR, which
induces a tendency to increase the number of 1-bits.
√
Theorem 6. Consider the greedy (2+1) GA with mutation rate 0 < p ≤ 1/( n log n) using
an arbitrary mask-based crossover operator. Its expected optimization time on O NE M AX is at
least
min{ln n, ln(1/(p2 n))} − O(log log n)
.
k
n
(1 + maxk { (pn)
k!k! }) · p(1 − p)
Before giving the proof, note that for p = c/n with 0 < c ≤ 4 constant,
k
i+1
i
i
(pn)
(pn)
(pn)
pn
maxk { (pn)
k!k! } = pn as for 0 < pn ≤ 4 and i ∈ N (i+1)!(i+1)! = (i+1)2 · i!i! ≤ i!i! ,
hence a maximum is attained for k = 1. Then the lower bound from Theorem 6 is
n ln n
− O(n log log n).
c · e−c · (1 + c)
2 Note that in [46] the greedy (2+1) GA was defined slightly differently as there duplicate genotypes are
always rejected. Algorithm 2 is equivalent to the greedy (2+1) GA from [46] for the following reasons. If the
current population contains two different individuals of equal fitness and a duplicate of one of the parents is
created, both algorithms reject a duplicate genotype. If the population contains two individuals of different
fitness, both behave like the population only contained the fitter individual.
13
This matches the upper bound (3) up to small order terms, showing for the greedy
(2+1) GA that the new term 1 + c in the denominator of the bound from Theorem 4 was
not a coincidence. For p > 4/n, the lower bound is at least
(e + Ω(1)) · n ln n.
Together, this establishes the optimal mutation rate for the greedy (2+1) GA on O NE M AX.
Theorem√7. For the greedy (2+1) GA with uniform crossover on O NE M AX mutation rate
p = (1 + 5)/(2n) minimizes the expected number of function evaluations, up to small-order
terms.
For the proof of Theorem 6 we use the following lower-bound technique based on
fitness levels by the author [47].
Theorem 8 (Sudholt [47]). Consider a partition of the search space into non-empty sets
A1 , . . . , Am . For a search algorithm A we say that it is in Ai or on level i if the best individual created so far is in Ai . If there are χ, ui , γi,j for i < j where
1. the probability of traversing from level i to level j in one step is at most ui γi,j for all i < j,
Pm
2.
j=i+1 γi,j = 1 for all i, and
Pm
3. γi,j ≥ χ k=j γi,k for all i < j and some 0 ≤ χ ≤ 1,
then the expected hitting time of Am is at least
m−1
X
i=1
Prob(A starts in Ai ) · χ
m−1
X
j=i
1
.
uj
(5)
Proof of Theorem 6. We prove a lower bound for the following sped-up GA instead of
the original greedy (2+1) GA. Whenever it creates a new offspring with the same fitness,
but a different bit string as the current best individual, we assume the following. The
algorithm automatically performs a crossover between the two. Also, we assume that
this crossover leads to the best possible offspring in a sense that all bits where both
parents differ are set to 1 (i. e., the algorithm performs a bit-wise OR). That is, if both
search points have i 1-bits and Hamming distance 2k, then the resulting offspring has
i + k 1-bits.
Due to our assumptions, at the end of each generation there is always a single best
individual. For this reason we can model the algorithm by a Markov chain representing
the current best fitness.
The analysis follows a lower bound for EAs on O NE M AX [47, Theorem 9]. As
in [47] we consider the following
fitness-level partition that focuses only on the very
last fitness values. Let ` = n − min{n/log n, 1/(p2 n log n)} . Let Ai = {x | |x|1 = i}
for i > ` and A` contain all remaining search points. We know from [47] that the GA is
initialized in A` with probability at least 1 − 1/ log n if n is large enough.
The probability pi,i+k that the sped-up GA makes a transition from fitness i to
fitness i + k equals
pi,i+k = Prob(k more 0-bits than 1-bits flip) +
Prob(k 0-bits and k 1-bits flip)
14
According to [47, Lemma 2], for the considered fitness levels i > ` the former probability is bounded by
pk (1 − p)n−k ·
3 i(n − i)p2
(n − i)k
.
· 1+ ·
k!
5 (1 − p)2
The latter probability is bounded by
Prob(k 0-bits flip) · Prob(k 1-bits flip)
(n − i)k k
ik k
· p (1 − p)n−i−k ·
· p (1 − p)i−k
k!
k!
(n − i)k k
(pn)k
≤
· p (1 − p)n ·
.
k!
(1 − p)2k · k!
≤
Together, pi,i+k is at most
(pn)k
3 i(n − i)p2
.
+
(p(n − i))k (1 − p)n 1 + ·
5 (1 − p)2+k
(1 − p)2k · k!k!
We need to find variables ui and γi,i+k along with some 0 ≤ χ ≤ 1 such that all
conditions of Theorem 8 are fulfilled. Define
1
(pn)k
3 i(n − i)p2
u0i := p(1 − p)n (n − i) 1 + ·
+
·
max
k
5 (1 − p)3
(1 − p)2
k!k!
and
0
γi,i+k
:=
Observe that, for every k ∈ N,
p(n − i)
(1 − p)2
k−1
.
3 i(n − i)p2
1
(pn)k
0
u0i γi,i+k
≥ pk (1 − p)n (n − i)k 1 + ·
+
·
5 (1 − p)1+2k
(1 − p)2k k!k!
(pn)k
3 i(n − i)p2
+
≥ pk (1 − p)n (n − i)k 1 + ·
5 (1 − p)2+k
(1 − p)2k · k!k!
≥ pi,i+k .
In order to fulfill the second condition in Theorem 8, we consider the following norPn
γ0
0
0
malized variables: ui := u0i · j=i+1 γi,j
and γi,j := Pn i,j γ 0 . As ui γi,j = u0i γi,j
≥ pi,j ,
j=i+1 i,j
this proves the first condition of Theorem 8.
Following the proof of Theorem
Pm 9 in [47], it is easy to show that for
χ := 1 − (1−p)12 log n we get γi,j ≥ χ k=j γi,k for all i, j with j > i (the calculations
on [47, pp. 427–428] carry over by replacing “(1 − p)” with “(1 − p)2 ”). This establishes
the third and last
Pcondition.
Pm
m
0
0
As γi,j ≥ χ k=j γi,k is equivalent to γi,j
≥ χ k=j γi,k
, we get
n
X
j=i+1
0
γi,j
≤
0
γi,i+1
1
≤ ,
χ
χ
15
which implies, using i(n − i)p2 ≤ n(n − `)p2 ≤ log1 n [47, (12)] as well as 1 + x ≤ 1/(1 − x)
for x < 1,
1
(pn)k
3 i(n − i)p2
1
+
·
max
ui ≤ p(1 − p)n · (n − i) · · 1 + ·
k
χ
5 (1 − p)3
(1 − p)2
k!k!
k
1
3
(pn)
≤ p(1 − p)n−3 · (n − i) · · 1 +
+ max
k
χ
5 log n
k!k!
!
(pn)k
1
1
n−3
+ max
≤ p(1 − p)
· (n − i) · ·
3
k
χ
k!k!
1 − 5 log
n
k
(pn)
1 1 + maxk k!k!
n−3
.
≤ p(1 − p)
· (n − i) · ·
3
χ
1 − 5 log
n
Invoking Theorem 8 and recalling that the first fitness level is reached with probability at least 1 − 1/ log n, we get a lower bound of
n−1
X 1
χ
ui
i=`
3
n−1
1 − 5 log
(1 − p)3 X 1
1
n k ·
χ2 ·
≥ 1−
log n
p(1 − p)n
n−i
1 + maxk (pn)
i=`
k!k!
n−1
X 1
1
1
1
≥ 1−O
·
·
k
log n
p(1 − p)n
n−i
1 + maxk (pn)
i=`
k!k!
1−
1
log n
3
1
where in the last step we used that all factors χ, 1 − 5 log
,
and
1
−
p
are
1
−
O
n
log n ,
d
P
n−1 1
cd
and 1 − logc n
≥ 1 − log
i=` n−i ≤
n for any positive constants c, d. Bounding
ln(min{n, 1/(p2 n)}) − ln(log n) as in [47] and absorbing all small-order terms in the
−O(log log n) term from the statement gives the claimed bound.
We also ran experiments to see whether the outcome matches our inspection of the
dominating terms in the running time bounds for realistic problem dimensions. We
chose n = 1000 bits and recorded the average optimization time over 1000 runs. The
mutation rate p was set to c/n with c ∈ {0.1, 0.2, . . . , 4}. The result is shown in Figure 1.
One can see that for every mutation rate the greedy (2+1) GA has a lower average optimization time. As predicted, the performance difference becomes larger as the
mutation rate increases. The optimal mutation rates for both algorithms match with
minimal average optimization times. Note that also the variance/standard deviation
was much lower for the GA for higher mutation rates. Preliminary runs for n = 100
and n = 10000 bits gave very similar results. More experiments and statistical tests are
given in Section 6.1.
5
k-Point Crossover
The k-point crossover operator picks k cutting points from {1, . . . , n − 1} uniformly at
random without replacement. These cutting points divide both parents into segments
that are then assembled from alternating parents. That is, for parents x, y and cutting
16
number of evaluations
·104
(1+1) EA
(2+1) GA+uniform
8
6
4
2
0
0
1
√
5+1
2
2
3
mutation rate
4
·1/n
Figure 1: Average optimization times for the (1+1) EA and the greedy (2+1) GA with
uniform crossover on OneMax with n = 1000 bits. The mutation rate p is set to c/n
with c ∈ {0.1, 0.2, . . . , 4}. The thin lines show mean ± standard deviation.
points 1 ≤ `1 < `2 < · · · < `k ≤ n − 1 the offspring will be:
x1 . . . x`1 y`1 +1 . . . y`2 x`2 +1 . . . x`3 y`3 +1 . . . y`4
...
the suffix being y`k +1 . . . yn if k is odd and x`k +1 . . . xn if k is even.
For uniform crossover we have seen that populations containing different search
points of equal fitness are beneficial as uniform crossover can easily combine the good
“building blocks”. This holds regardless of the Hamming distance between these different individuals, and the position of bits where individuals differ.
The (µ+λ) GA with k-point crossover is harder to analyse as there the probability
of crossover creating an improvement depends on the Hamming distance of parents
and the position of differing bits.
Consider parents that differ in two bits, where these bit positions are quite close.
Then 1-point crossover has a high probability of taking both bits from the same parent.
In order to recombine those building blocks, the cutting point has to be chosen in between the two bit positions. A similar effect occurs for 2-point crossover also if the two
bit positions are on opposite ends of the bit string.
The following lemma gives a lower bound on the probability that k-point crossover
combines the right building blocks on O NE M AX, if two parents are equally fit and differ
in two bits. The lemma and its proof may be of independent interest.
Lemma 9. Consider two search points x, y with xi = 1, xi+d = 0, yi = 0, yi+d = 1 for 1 ≤
i < i + d ≤ n and xs = ys for s ∈
/ {i, i + d}. The probability of k-point crossover of x and y, for
any 1 ≤ k ≤ N − 1, where N := n − 1 ≥ 4 is the number of possible cutting points, creating
an offspring with a larger number of 1-bits is at least
d(N − d)
N (N − 1)
and exactly d/N for k = 1.
17
Proof. We identify cutting points with bits such that cutting point a results in two
strings x1 . . . xa and xa+1 . . . xn . We say that a cutting point a separates i and i + d
if a ∈ {i, . . . , i + d − 1}. Note that the prefix is always taken from x. The claim now
follows from showing that the number of separating cutting points is odd with the
claimed probability.
Let XN,d,k be the random variables that describes the number of cutting points
separating i and i+d. This variable follows a hypergeometric distribution Hyp(N, d, k),
illustrated by the following urn model with red and white balls. The urn contains N
balls, d of which are red. We draw k balls uniformly at random, without replacement.
Then XN,d,k describes the number of red balls drawn. We define the probability of
XN,d,k being odd, for 1 ≤ d ≤ N − 1 and 1 ≤ k ≤ N − 1 as
d N −d
k
k
X
X
x k−x
P (N, d, k) :=
Prob(XN,d,k = x) =
.
N
x=1, x odd
x=1, x odd
Note that for k = 1
P (N, d, 1) =
and for k = 2
d
1
P (N, d, 2) =
d
1
N −d
0
N
1
N −d
1
N
2
=
=
k
d
N
2d(N − d)
.
N (N − 1)
For all 1 ≤ d ≤ N − 1 and all 1 ≤ k ≤ N − 1 the following recurrence holds. Imagine drawing the first cutting point separately. With probability d/N , the cutting point
is a separating cutting point, and then we need an even number of further separating cutting points among the remaining k − 1 cutting points, drawn from a random
variable XN −1,d−1,k−1 . With the remaining probability (N − d)/N , the number of remaining cutting points must be even, and this number is drawn from a random variable
XN −1,d,k−1 . Hence
d
N −d
· (1 − P (N − 1, d − 1, k − 1)) +
· P (N − 1, d, k − 1).
N
N
Assume for an induction that for all 2 ≤ k 0 < k
P (N, d, k) =
d(N − d)
d(N − d)
≤ P (N, d, k 0 ) ≤ 1 −
.
N (N − 1)
N (N − 1)
This is true for k 0 = 2 as, using 3d(N − d) ≤ 3 · (N/2)2 ≤ N (N − 1) for N ≥ 4,
P (N, d, 2) =
2d(N − d)
3d(N − d) − d(N − d)
d(N − d)
=
≤1−
.
N (N − 1)
N (N − 1)
N (N − 1)
For k > 2, combining (6) and (7) yields
d
N −d
· (1 − P (N − 1, d − 1, k − 1)) +
· P (N − 1, d, k − 1)
N
N
d (d − 1)(N − d)
N −d
d(N − d − 1)
≥
·
+
·
N (N − 1)(N − 2)
N
(N − 1)(N − 2)
d(N − d)(d − 1 + N − d − 1)
=
N (N − 1)(N − 2)
d(N − d)
=
.
N (N − 1)
P (N, d, k) =
18
(6)
(7)
The upper bound follows similarly:
d
(d − 1)(N − d)
N −d
d(N − d − 1)
P (N, d, k) ≤
· 1−
+
· 1−
N
(N − 1)(N − 2)
N
(N − 1)(N − 2)
d(N − d)(d − 1 + N − d − 1)
=1−
N (N − 1)(N − 2)
d(N − d)
=1−
.
N (N − 1)
By induction, the claim follows.
In the setting of Lemma 9, the probability of k-point crossover creating an improvement depends on the distance between the two differing bits. Fortunately, for
search points that result from a mutation of one another, this distance has a favourable
distribution. This is made precise in the following lemma.
Lemma 10. Let x0 result from x by a mutation flipping one 1-bit and one 0-bit, where the positions i, j of these bits are chosen uniformly among all 1-bits and 0-bits, respectively. Then for
d := |i−j| the random variable min{d, n−d} stochastically dominates the uniform distribution
on {1, . . . , n/4}.
Proof. We first show the following. For any fixed index i and any integer 1 ≤ z < n/2
there are exactly two positions j such that min{d, n − d} = z. If i ∈ {1, . . . , n} and z ∈ N
are fixed, the only values for j that result in either |i − j| = z or n − |i − j| = z are
i + z, i − z, i + z − n, and i − z + n. Note that at most two of these values are in {1, . . . , n}.
Hence, there are at most 2 feasible values for j for every d ∈ N.
Let ` denote the number of 1-bits in x. If ` ≥ n/2, we assume that first the 0-bit
is chosen uniformly at random, and then consider the uniform random choice of a
corresponding 1-bit. Without loss of generality assume xi = 0 and xj = 1.
When i has been fixed and j is chosen uniformly at random, a worst case distribution for min{|i − j|, n − |i − j|} is attained when the 1-bits are distributed such that for
each 1 ≤ d ≤ b`/2c both feasible bit positions are 1. The worst case is hence a uniform
distribution on {1, . . . , b`/2c}, which stochastically dominates the uniform distribution
on {1, . . . , n/4}.
The case ` < n/2 is symmetrical: exchanging the roles of xj and xi as well as the
roles of zeros and ones yields a uniform distribution on the set {1, . . . , b(n − `)/2c}
as worst case, which again stochastically dominates the uniform distribution on
{1, . . . , n/4}.
Taken together, Lemma 9 and Lemma 10 indicate that k-point crossover has a good
chance of finding improvements through recombining the right “building blocks”.
However, this is based on the population containing potential parents of equal fitness
that only differ in two bits.
The following analysis shows that the population is likely to contain such a
favourable pair of parents. However, such a pair might get lost again if other individuals of the same fitness are being created, after all duplicates have been removed
from the population. For parents that differ in more than 2 bits, Lemma 9 does not
apply, hence we do not have an estimate of how likely such a crossover will find an
improvement.
In order to avoid this problem, we consider a more detailed tie-breaking rule. As
before, individuals with fewer duplicates are being preferred. In case there are still ties
19
number of evaluations
·104
6
(2+1) GA +1-point, dup-rnd
(2+1) GA +2-point, dup-rnd
(2+1) GA +1-point, dup-old
(2+1) GA +2-point, dup-old
4
2
0
1
2
mutation rate
3
4
·1/n
Figure 2: Average optimization times on O NE M AX with n = 1000 bits over 1000 runs
for the greedy (2+1) GA with 1- and 2-point crossover using different tie-breaking rules
if individuals are tied with regard to fitness and the number of duplicates. “dup-rnd”
breaks these ties randomly, whereas “dup-old” (Algorithm 3) prefers older individuals.
The mutation rate p is set to c/n with c ∈ {0.1, 0.2, . . . , 4}.
after considering the number of duplicates, the (µ+λ) GA will retain older individuals.
This refined tie-breaking rule is shown in Algorithm 3. As will be shown in the remainder, it implies that once a favourable pair of parents with Hamming distance 2 has been
created, this pair will never get lost.
Algorithm 3: Refined tie-breaking rule “dup-old”.
14
Let P contain the µ best individuals from P ∪ P 0 ; break ties towards
including individuals with the fewest duplicates in P ∪ P 0 . If there are still
ties, break them towards including older individuals.
This tie-breaking rule, called “dup-old” differs from the one used for the experiments in Figure 1 and those in Section 6. There, we broke ties uniformly at random
in case individuals are tied with respect to both fitness and the number of duplicates.
We call the latter rule “dup-rnd”. Experiments for the greedy (2+1) GA comparing
tie-breaking rules dup-old and dup-rnd over 1000 runs indicate that performance differences are very small, see Figure 2.3
Note, however, that on functions with plateaus, like royal road functions, retaining the older individuals prevents the (µ+λ) GA from performing random walks on
the plateau, once the population has spread such that there are no duplicates of any
individual. In this case we expect that performance will deteriorate when breaking ties
towards older individuals.
With the refined tie-breaking rule, the performance of (µ+λ) GAs is as follows.
3 Even though differences are small, one-sided Mann-Whitney U tests reveal some statistically significant
differences: for 1-point crossover dup-rnd is significantly faster than dup-old on a significance level of 0.001
for mutation rates at least 2.4/n (with two exceptions, 2.8/n and 3.6/n, with p-values still below 0.003).
Contrarily, dup-old was significantly faster for 2-point crossover for mutation rates in the range of 0.8/n to
3/n.
20
Theorem 11. The expected optimization time of every (µ+λ) GA implementing Algorithm 1
with tie-breaking rule dup-old from Algorithm 3, 2 ≤ µ = O(1), λ < µ, pc = o(1) and pc =
ω(1/ log n), p = c/n for some constant c > 0, and k-point crossover with any 1 ≤ k ≤ n − 2,
on O NE M AX is at most
n ln n
· (1 + o(1)).
−c
c · e · (1 + c)
This bound equals the upper bound (3) for (µ+λ) GAs with uniform crossover. It
improves upon the previous upper bound for the greedy (2+1) GA from [46, Theo2c
· n ln n larger. The reason
rem 8], whose dominant term was by an additive term of 3+3c
was that for the (2+1) GA favourable parents could get lost, which is now prevented by
the dup-old tie-breaking rule and conditions on pc .
The conditions pc = o(1) as well as µ, λ = O(1) are useful because they allow us to
estimate the probability that a single good individual takes over the whole population
with copies of itself.
In the remainder of this section we work towards proving Theorem 11 and assume
that n ≥ n0 for some n0 chosen such that all asymptotic statements that require a large
enough value of n hold true. For n < n0 there is nothing to prove, as the statement
holds trivially for bounded n.
We again estimate the time spent on each fitness level i, i. e., when the best fitness in the current population is i. To this end, we focus on the higher fitness levels
i ≥ n − n/ log n where the probability of creating an offspring on the same level can be
estimated nicely. The time for reaching these higher fitness levels only constitutes a
small-order term, compared to the claimed running time bound. The following lemma
proves this claim in a more general setting than needed for the proof of Theorem 11. In
particular, it holds for arbitrary tie-breaking rules and crossover operators.
Lemma 12. For every (µ+λ) GA implementing Algorithm 1 with µ, λ = O(1), pc = 1 − Ω(1),
and p = c/n for a constant c > 0, using any initialization and any crossover operator, the
expected time until a fitness level i ≥ n − n/ log n is reached for the first time is o(n log n).
A proof is given in the appendix.
In the remainder of the section we focus on higher fitness levels i ≥ n − n/ log n
and specify the different cases on each such fitness level. The cases i.1, i.2, and i.3 are
similar to the ones for uniform crossover, with additional conditions on the similarity of
individuals in Cases i.2 and i.3. We also have an additional error state that accounts for
undesirable and unexpected behavior. We pessimistically assume that the error state
cannot be left towards other cases on level i.
Case i.1: The population contains an individual on level i and at least one individual on a lower fitness level.
Case i.2: The population contains µ copies of an individual x on level i.
Case i.3: The population contains two search points x, y with current best fitness i,
where y resulted from a mutation of x and the Hamming distance of x and y is 2.
Case i.error: An error state reached from any Case i.· when the best fitness is i and
none of the prior cases applies.
The difference to the analysis of uniform crossover is that in Case i.2 we rely on
the population collapsing to copies of a single individual. This helps to estimate the
probability of creating a favourable parent-offspring pair in Case i.3 as the (µ+λ) GA
effectively only performs mutations of x while being in Case i.2.
Lemma 13. Consider any (µ+λ) GA as defined in Theorem 11, with parameters 2 ≤ µ = O(1),
λ < µ, pc = o(1) and pc = ω(1/ log n), p = c/n for some constant c > 0. The total expected
21
time spent in all Cases i.1, i.2, and i.3 across all i ≥ n − n/ log n is at most
n ln n
+ o(n log n).
c · e−c · (1 + c)
Proof. We have already analyzed the expected time in Cases i.1 and i.2, across all fitness
levels. As in the proof of Theorem 4, we use Lemma 2 and get that the expected time
spent in all Cases i.1 is at most
O((µ + λ)n log µ)
= O(n).
(1 − pc )(1 − p)n
In Case i.2 the algorithm behaves like the one using uniform crossover described in
Theorem 4 as both crossover operators are working on identical individuals. As before,
Case i.2 is left if either a better offspring is created, or if a different offspring with i ones
is created. In the latter case, either Case i.3 or the error state i.error is reached. By the
proof of Theorem 4 we know that the expected time spent in Cases i.2 across all levels i
is bounded by
λn +
ln(n2 p + n) + 1 + p
n ln n
=
+ O(n).
p(1 − p)n−1 · (1 + np)
c · e−c · (1 + c)
Now we estimate the total time spent in all cases i.3. As this time turns out to be
comparably small, we allow ourselves to ignore that fact that not all these cases are
actually reached.
Case i.3 implies that the population contains a parent-offspring pair x, y with
Hamming distance 2. Consider the mutation that has created this offspring and note
that this mutation flips each 1-bit and each 0-bit with the same probability. If a, b with
a < b denote the bit positions where x and y differ, then D := b − a is a random variable
with support {1, . . . , n − 1}. By the law of total expectation,
E(Ti.3 ) =
n−1
X
d=1
E(Ti.3 | D = d) · Prob(D = d).
(8)
We first bound the conditional expectation by considering probabilities for improvements. If D = d then by crossover is successful if crossover is performed (probability pc ), if the search point where bit a is 1 is selected as first parent (probability at least
1/µ), if the remaining search point in {x, y} is selected as second parent (probability
at least 1/µ), and if cutting points are chosen that lead to a fitness improvement. The
latter event has probability at least d(N − d)/(N (N − 1)) by Lemma 9, with N := n − 1.
Finally, we need to assume that the following mutation does not destroy any fitness
improvements (probability at least (1 − p)n ). The probability of a successful crossover
is then at least, using d(N − d) = min(d, N − d) · max(d, N − d) ≥ min(d, N − d) · N/2,
pc (1 − p)n d(N − d)
pc (1 − p)n min(d, (N − d))
pc (1 − p)n min(d, n − d) − 1
·
≥
·
≥
·
.
2
2
µ
N (N − 1)
µ
2(N − 1)
µ2
2n
Another means of escaping from Case i.3 is by not using crossover, but having mutation
create an improvement. The probability for this is at least
(1 − pc ) · (n − i)p(1 − p)n−1 ≥ γ ·
22
n−i
n
(9)
for a constant γ > 0. Applying Lemma 3, we have
E(Ti.3 | D = d) ≤ λ +
1
pc (1−p)n
µ2
·
min(d,n−d)−1
2n
+γ·
n−i
n
.
(10)
Note that this upper bound is non-increasing with min(d, n − d). We are therefore
pessimistic when replacing min(d, n − d) by a different random variable that is stochastically dominated by it. According to Lemma 10, min{d, n − d} dominates the uniform
distribution on {1, . . . , n/4} (assume w. l. o. g. that n is a multiple of 4). Combining this
with (8) and (10) yields
E(Ti.3 ) ≤ λ +
n/4
X
1
·
n/4
d=1
1
pc (1−p)n
µ2
2
≤ λ + O(µ /pc ) ·
n/4
X
d=1
·
d−1
2n
+γ·
n−i
n
1
.
d−1+n−i
The last sum is estimated as follows.
n/4−1
n/4−1
X
X
1
1
1
1
=
=
+
d−1+n−i
d+n−i
n−i
d+n−i
d=0
d=1
d=1
Z n/4
1
dd
≤1+
d=0 d + n − i
n/4
= 1 + ln 1 +
.
n−i
n/4
X
Along with λ = O(1), n/4 ≤ n, and O(µ2 /pc ) = o(log n), we get
n
.
E(Ti.3 ) ≤ o(log n) · 1 + ln 1 +
n−i
Pn−1
For the sum T·,3 = i=0 Ti,3 we then have the following.
E(T·.3 ) ≤ o(n log n) + o(log n) ·
n−1
X
i=0
n
X
ln 1 +
n
n−i
n
ln 1 +
i
i=1
Z n
n
≤ o(n log n) + o(log n) ·
ln 1 +
di
i
i=0
= o(n log n)
= o(n log n) + o(log n) ·
as the integral is 2 ln(2)n. This completes the proof.
The remainder of the proof is devoted to estimating the expected time spent in the
error state. To this end we need to consider events that take the (µ+λ) GA “off course”,
that is, deviating from situations described in Cases i.1, i.2, and i.3.
Since Case i.3 is based on offspring with Hamming distance 2 to their parents, one
potential failure is that an offspring with fitness i, but Hamming distance greater than 2
to its parent is being created. This probability is estimated in the following lemma.
23
Lemma 14. For i ∈ {1, . . . , n − 1} let p(i) denote the probability that standard bit mutation
with mutation rate 0 < p ≤ 1/2 of a search point with i 1-bits creates a different offspring with
i 1-bits. If i(n − i)p2 (1 − p)−2 ≤ 1/2 then
2
n−2
i(n − i)p (1 − p)
≤ p
(i)
2
≤ i(n − i)p (1 − p)
n−2
2i(n − i)p2
· 1+
(1 − p)2
.
The probability that, additionally, the offspring has Hamming distance larger than 2 to its parent
is at most
2i2 (n − i)2 p4 (1 − p)n−4 .
The proof is found in the appendix.
Another potential failure occurs if the population does not collapse to copies of a
single search point, that is, the transition from Case i.1 to Case i.2 is not made. We first
estimate the probability of mutation unexpectedly creating an individual with fitness i.
Lemma 15. The probability that a standard bit mutation with mutation probability 0 < p < 1
creates a search point with i ones out of a parent with less than i ones, is at most
2
p(n − i + 1) · e(pn)
/4+1
.
Note that for the special case p = 1/n Lemma 13 in [12] gives an upper bound of
(n − i + 1)/n as the highest probability for a jump to fitness level i is attained if the
parent is on level i − 1. However, for larger mutation probabilities this is no longer
true in general; there are cases where the probability of jumping to level i is maximized
for parents on lower fitness levels. Hence, a closer inspection of transition probabilities
between different fitness levels is required, see the proof in the Appendix.
Using Lemma 15, we can now estimate the probability of the (µ+λ) GA not collapsing to copies of a single search point as described in Case i.2.
Lemma 16. Consider any (µ+λ) GA as defined in Theorem 11, with parameters 2 ≤ µ = O(1),
λ < µ, pc = o(1) and pc = ω(1/ log n), p = c/n for some constant c > 0 and fix a fitness
level i < n. The probability that the (µ+λ) GA will reach a population containing different
individuals with fitness i before either reaching a population containing only copies of the same
individual on level i or reaching a higher fitness level, is at most
n−i
.
O(µ log µ) · pc +
n
Proof. We show that there is a good probability of repeatedly creating clones of individuals with fitness i (or finding an improvement) and avoiding the following bad event. A
bad event happens if an individual on fitness level i is created in one offspring creation
by means other than through cloning an existing individual on level i.
The probability of a bad event is bounded as follows. In case crossover is being
used, which happens with probability pc , we bound the probability of a bad event
by the trivial bound 1. Otherwise, such an individual needs to be created through
mutation from either a worst fitness level, or by mutating a parent on level i. The
probability for the former is bounded from above by Lemma 15. The probability for
the latter is at most p(n − i) as it is necessary to flip one out of n − i 0-bits. Using
n − i + 1 ≤ 2(n − i), the probability of a bad event on level i is hence bounded from
24
above by
2
pc + (1 − pc ) · p(n − i + 1) · e(pn) /4+1 + p(n − i)
2c(n − i) c2 /4+1 c(n − i)
n−i
≤ pc +
·e
+
= pc + κ ·
,
n
n
n
2
where κ := 2c · ec /4+1 + c is a constant. The (µ+λ) GA will only reach a population
containing different individuals with fitness i as stated if a bad event happens before
the population has collapsed to copies of a single search point or moved on to a higher
fitness level.
Consider the first generation where an individual of fitness i is reached for the first
time. Since it might be possible to create several such individuals in one generation, we
consider all offspring creations being executed sequentially and consider the possibility
of bad events for all offspring creations following the first offspring on level i. Let X be
the number of function evaluations following this generation, before all individuals in
the population have fitness at least i. By Lemma 2 we have
(µ + λ) log µ
= O(µ log µ).
E(X) = O
(1 − pc )(1 − p)n
Considering up to λ further offspring creations in the first generation leading to level i,
and completing the generation at the end of the X function evaluations, we have less
than X + 2λ trials for bad events. The probability that one of these is bad is bounded
by
!
X
X
n−i
n−i
t · Prob(X = t) + 2λ · pc + κ ·
= (E(X) + 2λ) · pc + κ ·
n
n
t=1
n−i
= O(µ log µ) · pc + κ ·
.
n
Absorbing κ in the O-term yields the claimed result.
Now we are prepared to estimate the expected time spent in all error states i.error
for i ≥ n − n/ log n.
Lemma 17. Consider any (µ+λ) GA as defined in Theorem 11, with parameters 2 ≤ µ = O(1),
λ < µ, pc = o(1) and pc = ω(1/ log n), p = c/n for some constant c > 0. The expected time
spent in all states i.error, for i ≥ n − n/ log n, is at most
O(n + pc · n ln n) = o(n log n).
Proof. The (µ+λ) GA only spends time in an error state if it is actually reached. So we
first calculate the probability that state i.error is reached from either Case i.1, i.2, or i.3.
Lemma 16 states that the probability of reaching a population with different individuals on level i before reaching Case i.2 or a better fitness level is
n−i
n−i
O(µ log µ) · pc +
= O pc +
.
n
n
We pessimistically ignore the possibility that Case i.3 might be reached if this happens;
thus the above is an upper bound for the probability of reaching i.error from Case i.1.
25
Recall that in Case i.2 all individuals are identical, so crossover has no effect and
the (µ+λ) GA only performs mutations. First consider the case λ = 1. Note that i ≥
n − n/ log n, along with p = c/n, implies that i(n − i)p2 (1 − p)−2 ≤ 1/2, hence Lemma 14
is in force. According to Lemma 14 the probability of leaving Case i.2 by creating a
different individual with fitness i is at least i(n − i)p2 (1 − p)n−2 . The probability of
doing this with an offspring of Hamming distance greater than 2 to its parent is at most
2i2 (n − i)2 p4 (1 − p)n−4 (second statement of Lemma 14). So the conditional probability
of reaching the error state when leaving Case i.2 towards another case on level i is at
most
2i2 (n − i)2 p4 (1 − p)n−4
= 2i(n − i)p2 (1 − p)−2 .
(11)
i(n − i)p2 (1 − p)n−2
In case λ > 1 note that Case i.3 is reached in case there is a single offspring with fitness i
and Hamming distance 2 to its parent. Such an offspring is guaranteed to survive
as we assume λ < µ and offspring with many duplicates are removed first. Thus,
in case several offspring with fitness i and differing from their parent are created, all
of them need to have Hamming distance larger than 2 in order to reach i.error from
Case i.2. This probability decreases with increasing λ, hence the probability bound (11)
also holds for λ > 1.
Finally, Case i.3 implies that there exists a parent-offspring pair x, y with Hamming
distance 2. In a new generation these two offspring – or at least one copy of each – will
always survive: individuals with multiple duplicates are removed first, and in case
among current parents and offspring more than µ individuals exist with no duplicates,
x and y will be preferred over newly created offspring. So the probability of reaching
the error state from Case i.3 is 0.
In case the error state is reached, according to (9) we have a probability of at least γ ·
n−i
of
finding a better individual in one offspring creation, for a constant γ > 0. Using
n
n
Lemma 3 as before, this translates to at most λ + γ1 · n−i
expected function evaluations.
So the expected time spent in Case i.error is at most
1
n
n−i
· ·
λ + 2i(n − i)p2 (1 − p)−2 + O pc +
n
γ n−i
2
n
=λ+
· inp2 (1 − p)−2 + O pc ·
+1
γ
n−i
n
= O pc ·
+ O(1)
n−i
as both λ = O(1) and inp2 (1 − p)−2 ≤ (pn)2 · (1 − p)−2 = O(1). The total expected time
across all error states is at most
!
n−1
X 1
= O(n + pc · n ln n).
O n + pc · n ·
n−i
i=0
Now Theorem 11 follows from all previous lemmas.
Proof of Theorem 11. The claimed upper bound now follows from adding the upper
bounds on the expected time on the smaller fitness levels (Lemma 12) to the expected
times spent in all considered cases (Lemma 13 and Lemma 17).
We believe that some of the technical conditions from Theorem 11 involving µ, λ,
and pc could be relaxed if it was possible to generalize Lemmas 9 and 10 towards more
than 2 differing bits between individuals of equal fitness.
26
8
·104
(1+1) EA
(2+1) GA+uniform
(2+1) GA+1-point
(2+1) GA+2-point
6
4
2
0
1
2
3
4
·1/n
mutation rate
(a) O NE M AX
·105
·104
(1+1) EA
GA+uniform
GA+1-point
GA+2-point
5
4
(1+1) EA
GA+uniform
GA+1-point
GA+2-point
8
6
3
4
2
2
1
0
1
2
mutation rate
3
4
·1/n
(b) Royal road
0
1
2
3
mutation rate
4
·1/n
(c) Random polynomials
Figure 3: Average optimization times over 1000 runs for (1+1) EA and the
greedy (2+1) GA with various crossover operators on functions with n = 1000 bits:
O NE M AX, a royal road function with block size 5, and random polynomials with 1000
unweighted monomials of degree 5. The mutation rate is c/n with c ∈ {0.1, 0.2, . . . , 4}.
Figure 3, discussed in the following Section 6, presents further experiments
and statistical tests, which includes a comparison of uniform crossover and k-point
crossover in the greedy (2+1) GA.
6
6.1
Extensions to Other Building-Block Functions
Royal Roads and Monotone Polynomials
So far, our theorems and proofs have been focused on O NE M AX only. This is because
we do have very strong results about the performance of EAs on O NE M AX at hand.
However, the insights gained stretch far beyond O NE M AX. Royal road functions generally consist of larger blocks of bits. All bits in a block need to be set to 1 in order to
contribute to the fitness; otherwise the contribution is 0. All blocks contribute the same
amount to the fitness, and the fitness is just the sum of all contributions.
27
O NE M AX
Royal road
Random
Polynomial
uniform
1-point
2-point
uniform
1-point
2-point
uniform
1-point
2-point
(1+1) EA
p < 10−3
p < 10−3
p < 10−3
p < 10−3
p < 10−3
p < 10−3
p < 10−3
p < 10−3
p < 10−3
for c ≥ 0.6
for c ≥ 0.6
for c ≥ 0.6
(1 ex.)
for c ≥ 0.3 (3 ex.)
for c ≥ 0.4 (1 ex.)
uniform
1-point
p < 10−3 for c ≥ 0.4
p < 10−3 for c ≥ 0.3
p > 10−3 (11 ex.)
p < 10−3 for c ≥ 0.8 (1 ex.)
p < 10−3 for c ≥ 1.4 (5 ex.)
p > 10−3 (6 ex.)
p > 10−3 (13 ex.)
p > 10−3 (6 ex.)
p > 10−3 (6 ex.)
Table 1: Summary of the results of two-sided Mann-Whitney U tests on the data from
Figure 3. For each function the table shows pairwise comparisons between the (1+1) EA
and the greedy GA with uniform, 1-point, and 2-point crossover, resp. Here p is the
p-value output by the statistics package R (version 2.8.1) and c is the constant in the
mutation rate c/n. Each cell describes a rule for p subject to a minimum value of c and
gives the number of exceptions from this rule where applicable.
The fundamental insight we have gained for neutral mutations also applies to
royal road functions. If there is a mutation that completes one block, but destroys
another block, this is a neutral mutation and the offspring will be stored in the population of a (µ+λ) GA. Then crossover can recombine all finished blocks in the same
way as for O NE M AX. The only difference is that the destroyed block may evolve further. More neutral mutations can occur that only alter bits in the destroyed block. Then
the population can be dominated by many similar solutions, and it becomes harder for
crossover to find a good pair for recombination. However, as crossover has generally
a very high probability of finding improvements, the last effect probably plays only a
minor role.
A theoretical analysis of general royal roads up to the same level of detail as
for O NE M AX is harder, but not impossible. So far results on royal roads and monotone polynomials have been mostly asymptotic [51, 14]. Only recently, Doerr and
Künnemann [13] presented a tighter runtime analysis of offspring populations for royal
road functions, which may lend itself to a generalization of our results on O NE M AX in
future work.
For now, we use experiments to see whether the performance is similar to that on
O NE M AX. We use royal roads with n = 1000 bits and block size 5, i. e., we have 200
pairwise disjoint blocks of 5 bits each. We also consider random monotone polynomials. Instead of using disjoint blocks, we use 1000 monomials of degree 5 (conjunctions
of 5 bits): each monomial is made up of 5 bit positions chosen uniformly at random,
without replacement. This leads to a function similar to royal roads, but “blocks” are
broken up and can share bits; bit positions are completely random. Figure 3 shows the
average optimization times in 1000 runs on all these functions, for the (1+1) EA and the
greedy (2+1) GA with uniform, 1-point, and 2-point crossover. We chose the last two
because k-point crossovers for odd k treat ends of bit strings differently from those for
even k: for odd k two bits close to opposite ends of a bitstring have a high probability
to be taken from different parents, whereas for even k there is a high chance that both
will be taken from the same parent (cf. Lemma 9 for k = 2 and the special case of k = 1).
For consistency and simplicity, we use pc = 1 and the tie-breaking rule dup-rnd in
all settings, that is, ties in fitness are broken towards minimum numbers of duplicates
and any remaining ties are broken uniformly at random. For O NE M AX this does not
perfectly match the conditions of Theorem 11 as they require a lower crossover proba28
bility, pc = o(1), and tie-breaking rule dup-old. But the experiments show that k-point
crossover is still effective when these conditions are not met.
On O NE M AX both k-point crossovers are better than the (1+1) EA, but slightly
worse than uniform crossover. This is in accordance with the observation from our
analyses that improvements with k-point crossover might be harder to find, in case the
differing bits are in close proximity.
For royal roads the curves are very similar. The difference between the (1+1) EA
and the greedy (2+1) GA is just a bit smaller. For random polynomials there are visible differences, albeit smaller. Mann-Whitney U tests confirm that wherever there is
a noticeable gap between the curves, there is a statistically significant difference on a
significance level of 0.001. The outcome of Mann-Whitney U tests is summarized in
Table 1.
For very small mutation rates c/n the tests were not significant. For mutation rates
no less than 0.6/n all differences between the (1+1) EA and all greedy (2+1) GAs were
statistically significant, apart from a few exceptions on random polynomials. For O NE M AX the difference between uniform crossover vs. k-point crossover was significant
for c ≥ 0.4. For royal roads the majority of such comparisons showed statistical significance, with a number of exceptions. However, for random polynomials the majority of
comparisons were not statistically significant. Most comparisons between 1-point and
2-point crossover did not show statistical significance.
These findings give strong evidence that the insights drawn from the analysis on
O NE M AX transfer to broader classes of functions where building blocks need to be
assembled.
6.2
Linear Functions
Another interesting question is in how far the theoretical analyses in this work extend
to cases where building blocks have different weights. The simplest such case is the
class of linear functions, defined as
f (x) =
n
X
wi xi
i=1
where wi > 0 are positive real-valued weights.
Doerr, Doerr, and Ebel [9] provided empirical evidence that their (1+(λ, λ)) EA is
faster than the (1+1) EA on linear functions with weights drawn uniformly at random
from [1, 2].
It is an open question whether this also holds for more common GAs, that is, those
implementing Algorithm 1. Experiments in [9] on the greedy (2+1) GA found that on
random linear functions “no advantage of the (2+1) GA over the (1+1) EA is visible”.
We provide an explanation for this observation and reveal why the (2+1) GA is not well
suited for weighted building blocks, whereas other GAs might be.
The reason why the (2+1) GA behaves like the (1+1) EA in the presence of weights
is that in case the current population of the (2+1) GA contains two members with different fitness, the (2+1) GA ignores the inferior one. So it behaves as if the population
only contained the fitter individual. Since the (2+1) GA will select the fitter individual
twice for crossover, followed by mutation, it essentially just mutates the fitter individual. This behavior of the (2+1) GA then equals that of a (1+1) EA working on the fitter
individual.
The (2+1) GA is more efficient than the (1+1) EA on O NE M AX (and other buildingblock functions where all building blocks are equally important) as it can easily gen29
8
·104
8
(1+1) EA
(5+1) GA+uniform
(5+1) GA+1-point
(5+1) GA+2-point
6
·104
(1+1) EA
(5+1) GA+uniform
(5+1) GA+1-point
(5+1) GA+2-point
6
4
4
2
2
0
1
2
mutation rate
3
4
·1/n
0
1
2
3
mutation rate
(a) O NE M AX
4
·1/n
(b) Random linear functions [1, 2]
Figure 4: Average optimization times over 1000 runs for (1+1) EA and a (5+1) GA with
uniform parent selection and various crossover operators on functions with n = 1000
bits: O NE M AX and random linear functions with weights drawn independently, uniformly at random from [1, 2], and anew for each run.
erate and store individuals with equal fitness in the population, and recombine their
different building blocks. However, in the presence of weights, chances of creating individuals of equal fitness might be very slim, and then the (2+1) GA behaves like the
(1+1) EA.
Theorem 18. As long as the population of the (2+1) GA does not contain two different individuals with the same fitness, the (2+1) GA is equivalent to the (1+1) EA.
On functions where all search points have different fitness values, the (2+1) GA is equivalent to the (1+1) EA. This includes linear functions with extreme weights like
BinVal(x) :=
n
X
2n−i xi
i=1
Pn
and, more generally, functions where w(i) > j=i+1 w(j) for all 1 ≤ i ≤ n, where w(i) denotes
the i-th largest weight. It also includes, almost surely, random linear functions with weights
being drawn from some real-valued interval [a, b] with a < b.
Proof. The first two statements have
Pn been established in the preceding discussion.
For functions where w(i) > j=i+1 w(j) for all 1 ≤ i ≤ n, all search points with a 1
on the bit of weight w(i) have a higher fitness than all search points where this bit is 0,
provided that all bits with larger weights are being fixed. It follows inductively that all
search points have different fitness values.
For random linear functions consider the function being constructed sequentially
by adding new bits with randomly drawn weights. Assume that after adding i bits,
all 2i bit patterns have different fitness values. This is trivially true for 0 bits. When
adding a new bit i + 1, a fitness value can only be duplicated with these i + 1 bits if the
i + 1-st weight is equal to any selection of weights from the first i bits. Since there are
at most 2i selections, which is finite, the i + 1-st weight will almost surely be different
from all of these. The statement then follows by induction.
30
In a sense, the (2+1) GA is not able to benefit from crossover in the settings from
Theorem 18 since its greedy parent selection suppresses diversity in the population.
So, in order for a GA to benefit from crossover, the population needs to be able
to maintain and select individuals with different building blocks and slightly different
fitness values for long enough, so that crossover has a good chance of combining those
building blocks. The (1+(λ, λ)) EA [9] achieves this using a cleverly designed twostage offspring creation process: mutation first creates diversity and the best among λ
mutants is retained and recombined with its parent λ times. However, this does not
explain why crossover is beneficial in common GA designs.
A promising common GA design does not need to be sophisticated – Figure 4
shows that already a simple (5+1) GA with uniform parent selection performs significantly better than the (1+1) EA (and hence the greedy (2+1) GA). The benefit of
crossover is smaller than that on O NE M AX, but the main qualitative observations are
the same: the average optimization time is smaller with crossover, and mutation rates
slightly larger than 1/n further improve performance.
One-sided Mann-Whitney U tests on a significance level of 0.001 showed that the
(5+1) GA with uniform crossover was significantly faster than the (1+1) EA on random
linear functions, for mutation rates no less than 0.6/n. Both k-point crossovers gave
mixed results: they were slower than the (1+1) EA for low mutation rates (0.4 ≤ c ≤ 1.2,
except for c = 0.6, for 1-point crossover and 0.9 ≤ c ≤ 1.1 for 2-point crossover), but
faster for high mutation rates (c = 2.3 and c ≥ 2.6 for 1-point crossover, c = 2.0 and
c ≥ 2.2 for 2-point crossover).
This shows that uniform crossover can speed up building-block assembly for
weighted building blocks – albeit not for all (µ+λ) GAs, and in particular not for the
greedy (2+1) GA. Proving this rigorously for random or arbitrary linear functions remains a challenging open problem, and so is identifying characteristics of (µ+λ) GAs
for which crossover is beneficial in these cases.
7
Conclusions and Future Work
We have demonstrated rigorously and intuitively that crossover can speed up building
block assembly on O NE M AX, with evidence that the same holds for a broad class of
functions. The basic insight is that mutations that create new building blocks while
destroying others can still be useful: mutants can be stored in the population and lead
to a successful recombination with their parents in a later generation. This effect makes
every (µ+λ) GA with cut selection and moderate population sizes twice as fast as every mutation-based EA on O NE M AX. In other words, adding crossover to any such
(µ+λ) EA halves the expected optimization time (up to small-order terms). This applies
to uniform crossover and to k-point crossover, for arbitrary values of k.
Furthermore, we have demonstrated how to analyze parent and offspring populations as in (µ+λ) EAs and (µ+λ) GAs. As long as both µ and λ are moderate, so that
exploitation is not slowed down, we obtained essentially the same results for arbitrary
(µ+λ) GAs as for the simple greedy (2+1) GA analyzed in [46]. This work provides
novel techniques for the analysis of (µ+λ)-type algorithms, including Lemmas 2 and 3,
which may prove useful in further studies of EAs.
Another intriguing conclusion following naturally from our analysis is that the
optimal
√ mutation rate for GAs such as the greedy (2+1) GA changes from 1/n to
(1 + 5)/2 · 1/n ≈ 1.618/n when using uniform crossover. This is simply because
neutral mutations and hence multi-bit mutations become more useful. Experiments
are in perfect accordance with the theoretical results for O NE M AX. For other functions
31
like royal roads and random polynomials they indicate that the performance differences also hold in a much more general sense. We have empirical evidence that this
might also extend to linear functions, and weighted building blocks in general, albeit
this does not apply to the greedy (2+1) GA. The discussion from Section 6.2 has shown
that the population must be able to store individuals with different building blocks for
long enough so that crossover can combine them, even though some individuals might
have inferior fitness values and be subject to replacement.
Our results give novel, intuitive and rigorous answers to a question that has been
discussed controversially for decades.
There are plenty of avenues for future work. We would like to extend the theoretical analysis of (µ+λ) GAs to royal road functions and monotone polynomials. Also
investigating weighted building blocks, like in linear functions, is an interesting and
challenging topic for future work.
Our (µ+λ) GAs benefit from crossover and an increased mutation rate because
cut selection removes offspring with inferior fitness. As such, cut selection counteracts disruptive effects of crossover and an increase of the mutation rate. The situation
is entirely different in generational GAs, where Ochoa, Harvey, and Buxton reported
that introducing crossover can decrease the optimal mutation rate [36]. Future work
could deal with complementing these different settings and investigating the balance
between selection pressure for replacement selection and the optimal mutation rate.
Acknowledgments
The author was partially supported by EPSRC grant EP/D052785/1 while being a
member of CERCIA, University of Birmingham, UK. The research leading to these results has received funding from the European Union Seventh Framework Programme
(FP7/2007-2013) under grant agreement no 618091 (SAGE). The author would like to
thank the reviewers for their detailed and constructive comments that helped to improve the manuscript.
References
[1] M. Abramowitz and I. A. Stegun. Handbook of Mathematical Functions with Formulas,
Graphs, and Mathematical Tables. Dover, New York, ninth Dover printing, tenth
GPO printing edition, 1964.
[2] S. Arora, Y. Rabani, and U. V. Vazirani. Simulating quadratic dynamical systems
is PSPACE-complete. In Proceedings of the 26th ACM Symposium on the Theory of
Computing (STOC), pages 459–467, 1994.
[3] A. Auger and B. Doerr, editors. Theory of Randomized Search Heuristics – Foundations
and Recent Developments. Number 1 in Series on Theoretical Computer Science.
World Scientific, 2011.
[4] G. Badkobeh, P. K. Lehre, and D. Sudholt. Unbiased black-box complexity of parallel search. In 13th International Conference on Parallel Problem Solving from Nature
(PPSN 2014), volume 8672 of LNCS, pages 892–901. Springer, 2014.
[5] N. H. Barton and B. Charlesworth.
281(5385):1986–1990, 1998.
Why sex and recombination?
Science,
[6] G. Bell. The masterpiece of nature the evolution and genetics of sexuality. Univ. of
California Press, 1982.
32
[7] K. A. De Jong. Evolutionary Computation – A Unified Approach. MIT press, 2006.
[8] M. Dietzfelbinger, B. Naudts, C. Van Hoyweghen, and I. Wegener. The analysis of
a recombinative hill-climber on H-IFF. IEEE Transactions on Evolutionary Computation, 7(5):417–423, 2003.
[9] B. Doerr, C. Doerr, and F. Ebel. Lessons from the black-box: fast crossover-based
genetic algorithms. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO ’13), pages 781–788. ACM, 2013.
[10] B. Doerr, M. Fouz, and C. Witt. Sharp bounds by probability-generating functions and variable drift. In Proceedings of the 13th Annual Genetic and Evolutionary
Computation Conference (GECCO ’11), pages 2083–2090. ACM Press, 2011.
[11] B. Doerr, E. Happ, and C. Klein. Crossover can provably be useful in evolutionary
computation. Theoretical Computer Science, 425(0):17–33, 2012.
[12] B. Doerr, D. Johannsen, and C. Winzen. Multiplicative drift analysis. Algorithmica,
64:673–697, 2012.
[13] B. Doerr and M. Künnemann. Royal road functions and the (1+λ) evolutionary algorithm: Almost no speed-up from larger offspring populations. In IEEE Congress
on Evolutionary Computation (CEC 2013), pages 424–431, 2013.
[14] B. Doerr, D. Sudholt, and C. Witt. When do evolutionary algorithms optimize
separable functions in parallel? In Foundations of Genetic Algorithms (FOGA 2013),
pages 51–64. ACM, 2013.
[15] A. E. Eiben and J. E. Smith. Introduction to Evolutionary Computing. Springer, 2007.
[16] S. Fischer and I. Wegener. The one-dimensional Ising model: Mutation versus
recombination. Theoretical Computer Science, 344(2–3):208–225, 2005.
[17] S. Forrest and M. Mitchell. Relative building block fitness and the building block
hypotheses. In Proc. of FOGA 2, pages 198–226. Morgan Kaufmann, 1993.
[18] T. Jansen. Analyzing Evolutionary Algorithms – The Computer Science Perspective.
Springer, 2013.
[19] T. Jansen, K. A. De Jong, and I. Wegener. On the choice of the offspring population
size in evolutionary algorithms. Evolutionary Computation, 13:413–440, 2005.
[20] T. Jansen and I. Wegener. On the analysis of evolutionary algorithms—a proof that
crossover really can help. Algorithmica, 34(1):47–66, 2002.
[21] T. Jansen and I. Wegener. Real royal road functions—where crossover provably is
essential. Discrete Applied Mathematics, 149:111–125, 2005.
[22] T. Jansen and C. Zarges. Analysis of evolutionary algorithms: from computational
complexity analysis to algorithm engineering. In Proceedings of the 11th Workshop
on Foundations of Genetic Algorithms (FOGA ’11), pages 1–14. ACM, 2011.
[23] T. Kötzing, D. Sudholt, and M. Theile. How crossover helps in pseudo-Boolean
optimization. In Proceedings of the 13th annual conference on Genetic and evolutionary
computation (GECCO ’11), pages 989–996. ACM Press, 2011.
33
[24] J. Lässig. Personal communication, 2009.
[25] J. Lässig and D. Sudholt. General upper bounds on the running time of parallel
evolutionary algorithms. Evolutionary Computation, 22(3):405–437, 2014.
[26] P. Lehre and X. Yao. Crossover can be constructive when computing unique input–
output sequences. Soft Computing, 15(9):1675–1687, 2011.
[27] P. K. Lehre and C. Witt. Black-box search by unbiased variation. Algorithmica,
64(4):623–642, 2012.
[28] A. Livnat, C. Papadimitriou, J. Dushoff, and M. W. Feldman. A mixability theory
for the role of sex in evolution. Proceedings of the National Academy of Sciences,
105(50):19803–19808, 2008.
[29] A. Livnat, C. Papadimitriou, N. Pippenger, and M. W. Feldman. Sex, mixability,
and modularity. Proceedings of the National Academy of Sciences, 107(4):1452–1457,
2010.
[30] M. Mitchell, S. Forrest, and J. H. Holland. The royal road function for genetic
algorithms: fitness landscapes and GA performance. In Proc. of the 1st European
Conference on Artificial Life, pages 245–254. MIT Press, 1992.
[31] M. Mitchell, J. H. Holland, and S. Forrest. When will a genetic algorithm outperform hill climbing? In Advances in Neural Information Processing Systems, pages
51–58. Morgan Kaufmann, 1994.
[32] M. Mitzenmacher and E. Upfal. Probability and Computing. Cambridge University
Press, 2005.
[33] F. Neumann, P. S. Oliveto, and C. Witt. Theoretical analysis of fitness-proportional
selection: landscapes and efficiency. In Genetic and Evolutionary Computation Conference (GECCO’09), pages 835–842. ACM Press, 2009.
[34] F. Neumann and M. Theile. How crossover speeds up evolutionary algorithms for
the multi-criteria all-pairs-shortest-path problem. In 11th International Conference
on Parallel Problem Solving from Nature (PPSN 2010), pages 667–676. Springer, 2010.
[35] F. Neumann and C. Witt. Bioinspired Computation in Combinatorial Optimization –
Algorithms and Their Computational Complexity. Springer, 2010.
[36] G. Ochoa, I. Harvey, and H. Buxton. Error thresholds and their relation to optimal
mutation rates. In D. Floreano, J.-D. Nicoud, and F. Mondada, editors, Advances
in Artificial Life, volume 1674 of Lecture Notes in Computer Science, pages 54–63.
Springer, 1999.
[37] P. S. Oliveto and C. Witt. Improved runtime analysis of the simple genetic algorithm. In Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO ’13), pages 1621–1628. ACM, 2013.
[38] P. S. Oliveto and C. Witt. On the runtime analysis of the simple genetic algorithm.
Theoretical Computer Science, 545:2–19, 2014.
[39] C. Qian, Y. Yu, and Z.-H. Zhou. An analysis on recombination in multi-objective
evolutionary optimization. Artificial Intelligence, 204:99–119, 2013.
34
[40] Y. Rabani, Y. Rabinovich, and A. Sinclair. A computational view of population
genetics. Random Structures and Algorithms, 12(4):313–334, 1998.
[41] J. E. Rowe. Genetic algorithms. In J. Kacprzyk and W. Pedrycz, editors, Handbook
of Computational Intelligence. Springer, 2015. To appear.
[42] J. E. Rowe, M. D. Vose, and A. H. Wright. Group properties of crossover and
mutation. Evolutionary Computation, 10(2):151–184, 2002.
[43] T. Storch and I. Wegener. Real royal road functions for constant population size.
Theoretical Computer Science, 320:123–134, 2004.
[44] D. Sudholt. Crossover is provably essential for the Ising model on trees. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO ’05), pages
1161–1167. ACM Press, 2005.
[45] D. Sudholt. The impact of parametrization in memetic evolutionary algorithms.
Theoretical Computer Science, 410(26):2511–2528, 2009.
[46] D. Sudholt. Crossover speeds up building-block assembly. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO 2012), pages 689–696. ACM
Press, 2012.
[47] D. Sudholt. A new method for lower bounds on the running time of evolutionary
algorithms. IEEE Transactions on Evolutionary Computation, 17(3):418–435, 2013.
[48] D. Sudholt and C. Thyssen. Running time analysis of ant colony optimization for
shortest path problems. Journal of Discrete Algorithms, 10:165–180, 2012.
[49] M. D. Vose. The Simple Genetic Algorithm: Foundations and Theory. MIT press, 1999.
[50] R. A. Watson and T. Jansen. A building-block royal road where crossover is provably essential. In Proceedings of the Genetic and Evolutionary Computation Conference
(GECCO ’07), pages 1452–1459. ACM, 2007.
[51] I. Wegener and C. Witt. On the optimization of monotone polynomials by simple
randomized search heuristics. Combinatorics, Probability and Computing, 14:225–
247, 2005.
[52] D. B. Weissman and N. H. Barton. Limits to the rate of adaptive substitution in
sexual populations. PLoS Genetics, 8(6):e1002740, 2012.
[53] D. B. Weissman, M. W. Feldman, and D. S. Fisher. The rate of fitness-valley crossing in sexual populations. Genetics, 186(4):1389–1410, 2010.
[54] C. Witt. Runtime analysis of the (µ+1) EA on simple pseudo-Boolean functions.
Evolutionary Computation, 14(1):65–86, 2006.
[55] C. Witt. Tight bounds on the optimization time of a randomized search heuristic
on linear functions. Combinatorics, Probability and Computing, 22:294–318, 2 2013.
35
A
Appendix
The appendix contains proofs of lemmas omitted from the main part.
Proof of Lemma 12. If the current population has a best individual of fitness j < i, by
Lemma 2 after an expected number of O((µ + λ) log µ) = O(1) function evaluations
all individuals will have fitness at least j. Then one offspring creation results in an
improvement if no crossover is being used, and mutation flips exactly one out of n − j
0-bits. The probability for this event is
(1 − pc ) · (n − j)p(1 − p)n−1 ≥ γ ·
n−j
n
for some constant γ > 0, due to our conditions for p and pc .
Using Lemma 3, the expected time until a fitness level i ≥ n − n/ log n is reached
for the first time is therefore at most
n−(n/ log n)−1
n
X
X
n
1
n
= O(n) + ·
O(1) + λ +
γ(n
−
j)
γ
j
j=0
j=(n/ log n)+1
Z n
n
1
≤ O(n) + ·
dj
γ j=n/ log n j
n
= O(n) + · (ln n − ln(n/ log n))
γ
n
= O(n) + · ln(log n)
γ
= o(n log n).
Proof of Lemma 14. In order to create a different search point on the same fitness level,
there must be some integer ` ∈ {1, . . . , min{i, n − i}} such that ` 1-bits flip to 0 and `
0-bits flip to 1. This is a necessary and sufficient condition, so
p(i) =
min{i,n−i}
X
`=1
i
n − i 2`
p (1 − p)n−2` .
`
`
(12)
The case ` = 1 yields
the claimed lower bound. For the upper bound we bound the
above term, using nk ≤ nk /(k!) to bound both binomial coefficients:
p
(i)
min{i,n−i} `
X
i (n − i)` 2`
· p (1 − p)−2`
`!`!
`=1
`
∞
X
i(n − i)p2
≤ (1 − p)n
(1 − p)2
≤ (1 − p)
n
`=1
= (1 − p)
Applying
yields
1
1−x
= 1+
x
1−x
n
1
i(n−i)p2
(1−p)2
2
− i(n−i)p
(1−p)2
.
≤ 1 + 2x for x ≤ 1/2 to x :=
i(n−i)p2
(1−p)2
i(n − i)p2
2i(n − i)p2
(1 − p) ·
· 1+
(1 − p)2
(1 − p)2
n
36
in the above formula
and hence the claimed upper bound.
The second statement follows from the upper bound and the fact that the offspring
has Hamming distance 2 in the case ` = 1, i. e., with probability i(n−i)p2 (1−p)n−2 .
Proof of Lemma 15. A search point with i ones is created from a parent with i − d < i
ones if there is a value ` such that d + ` 0-bits flip to 1 and ` 1-bits flip to 0. The sought
probability therefore is
!
X n − i + di − d
d+2`
n−d−2`
max
p
(1 − p)
d≥1
d+`
`
`=0
!
X (n − i + d)d+` (i − d)`
d+2`
≤ max
·
·p
d≥1
(d + `)!
`!
`=0
!
X (p(n − i + d))d (p2 · (i − d)(n − i + d))`
·
= max
d≥1
(d + `)!
`!
`=0
X ((pn)2 /4)`
(p(n − i + d))d
≤ max
·
d≥1
d!
`!
`=0
2
(p(n − i + d))d
· e(pn) /4 .
= max
d≥1
d!
Using 1/(d!) ≤ (e/d)d , we bound the max term as
max
d≥1
(p(n − i + d))d
d!
≤ max
d≥1
(ep(n − i + d))
d
d
d
≤ max (ep(n − i + 1)) .
d≥1
Now, if ep(n − i + 1) ≤ 1, the maximum is attained for d = 1, in which case we get a
2
probability bound of ep(n − i + 1) · e(pn) /4 as claimed. If ep(n − i + 1) > 1 we trivially
bound the sought probability by
1 < ep(n − i + 1) ≤ ep(n − i + 1) · e(pn)
37
2
/4
.
| 9cs.NE
|
Modeling Musical Context Using Word2vec
D. Herremans∗1 and C.-H. Chuan†2
1
Queen Mary University of London, London, UK
University of North Florida, Jacksonville, USA
arXiv:1706.09088v2 [cs.SD] 29 Jun 2017
2
We present a semantic vector space model for capturing complex polyphonic musical context. A word2vec model based on a skip-gram representation with negative
sampling was used to model slices of music from a dataset of Beethoven’s piano
sonatas. A visualization of the reduced vector space using t-distributed stochastic
neighbor embedding shows that the resulting embedded vector space captures tonal
relationships, even without any explicit information about the musical contents of
the slices. Secondly, an excerpt of the Moonlight Sonata from Beethoven was altered by replacing slices based on context similarity. The resulting music shows
that the selected slice based on similar word2vec context also has a relatively short
tonal distance from the original slice.
Keywords: music context, word2vec, music, neural networks, semantic vector
space
1 Introduction
In this paper, we explore the semantic similarity that can be derived by looking solely at
the context in which a musical slice appears. In past research, music has often been modeled through Recursive Neural Networks (RNNs) combined with Restricted Bolzmann Machines [Boulanger-Lewandowski et al., 2012], Long-Short Term RNN models [Eck and Schmidhuber, 2002, Sak et al., 2014], Markov models [Conklin and Witten, 1995] and other statistical
models, using a representation that incorporates musical information (i.e., pitch, pitch class,
duration, intervals, etc.). In this research, we focus on modeling the context, over the content.
Vector space models [Rumelhart et al., 1988] are typically used in natural language processing
(NLP) to represent (or embed) words in a continuous vector space [Turney and Pantel, 2010,
McGregor et al., 2015, Agres et al., 2016, Liddy et al., 1999]. Within this space, semantically
similar words are represented geographically close to each other [Turney and Pantel, 2010]. A
recent very efficient approach to creating these vector spaces for natural language processing
is word2vec [Mikolov et al., 2013c].
∗
†
[email protected]
[email protected]
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
Although music is not the same as language, it possesses many of the same types of characteristics. Besson and Schön [2001] discuss the similarity of music and language in terms of,
among others, structural aspects and the expectancy generated by both a word and a note.
We can therefore use a model from NLP: word2vec. More specifically a skip-gram model with
negative sampling is used to create and train a model that captures musical context.
There have been only few attempts at modeling musical context with semantic vector space
models. For example, Huang et al. [2016] use word2vec to model chord sequences in order
to recommend chords other than the ‘ordinary’ to novice composers. In this paper, we aim
to use word2vec for modeling musical context in a more generic way as opposed to a reduced
representation as chord sequences. We represent complex polyphonic music as a sequence of
equal-length slices without any additional processing for musical concepts such as beat, time
signature, chord tones and etc.
In the next sections we will first discuss the implemented word2vec model, followed by a
discussion of how music was represented. Finally, the resulting model is evaluated.
2 Word2vec
Word2vec refers to a group of models developed by Mikolov et al. [2013c]. They are used to
create and train semantic vector spaces, often consisting of several hundred dimensions, based
on a corpus of text [Mikolov et al., 2013a]. In this vector space, each word from the corpus is
represented as a vector. Words that share a context are geographically close to each other in
this space. The word2vec architecture can be based on two approaches: a continuous bag-ofwords, or a continuous skip-gram model (CBOW). The former uses the context to predict the
current word, whereas the latter uses the current word to predict surrounding words [Mikolov
et al., 2013b]. Both models have a low computational complexity, so they can easily handle a
corpus with a size ranging in the billions of words in a matter of hours. While CBOW models
are faster, it has been observed that skip-gram performs better on small datasets [Mikolov
et al., 2013a]. We therefore opted to work with the latter model.
Skip-gram with negative sampling The architecture of a skip-gram model is represented in
Figure 1. For each word wt in a corpus of size T at position t, the network tries to predict the
surrounding words in a window c (c = 2 in the figure). The training objective is thus defined
as:
T
1X
T
X
log p(wt+i |wt ),
(1)
t=1 −c≤i≤c,i6=0
whereby the term p(wt+i |wt ) is calculated by a softmax function. Calculating the gradient of
this term is, however, computationally very expensive. Alternatives to circumvent this problem
include hierarchical softmax [Morin and Bengio, 2005] and noise contrastive estimation [Gutmann and Hyvärinen, 2012]. The word2vec model used in this research implements a variant
of the latter, namely negative sampling.
Figure 1: A skip-gram model with n-dimensions for word wt at position t.
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
The idea behind negative sampling is that a well trained model should be able to distinguish
between data and noise [Goldberg and Levy, 2014]. The original training objective is thus
approximated by a new, more efficient, formulation that implements a binary logistic regression
to classify between data and noise samples. When the model is able to assign high probabilities
to real words and low probabilities to noise samples, the objective is optimized Mikolov et al.
[2013c].
Cosine similarity was used as a similarity metric between two musical-slice vectors in our
vector space. For two non-zero vectors A and B in n dimensional space, with an angle θ, it is
defined as [Tan et al., 2005]:
Pn
Ai × Bi
Similarity(A, B) = cos(θ) = Pn i=12 Pn
(2)
2
i=1 Ai ×
i=1 Bi
In this research, we port the above discussed model and techniques to the field of music.
We do this by replacing ‘words’ with ‘slices of polyphonic music’. The manner in which this is
done is discussed in the next section.
3 Musical slices as words
In order to study the extend to which word2vec can model musical context, polyphonic musical
pieces are represented with as little injected musical knowledge as possible. Each piece is simply
segmented into equal-length, non-overlapping slices. The duration of these slices is calculated
for each piece based on the distribution of time between note onsets. The smallest amount of
time between consecutive onsets that occurs in more than 5% of all cases is selected as the
slice-size. The slices capture all pitches that sound in a slice: those that have their onset in the
slice, and those that are played and held over the slice. The slicing process does not depend
on musical concepts such as beat or time signature; instead, it is completely data-driven. Our
vocabulary of words, will thus consist of a collection of musical slices.
In addition, we do not label pitches as chords. All sounding pitches, including chord tones,
non-chord tones, and ornaments, are all recorded in the slice. We do not reduce pitches into
pitch classes either, i.e., pitches C4 and C5 are considered different pitches. The only musical
knowledge we use is the global key, as we transpose all pieces to either C major or A minor
before segmentation. This enables the functional role of pitches in tonality to stay the same
across compositions, which in turn causes there to be more repeated slices over the dataset
and allows the model to be better trained on less data. In the next section, the performance
of the resulting model is discussed.
4 Results
In order to evaluate how well the proposed model captures musical context, a few experiments
were performed on a dataset consisting of Beethoven’s piano sonatas. The resulting dataset
consists of 70,305 words, with a total of 14,315 unique occurrences. As discussed above,
word2vec models are very efficient to train. Within minutes, the model was trained on the
CPU of a MacBook Pro.
We trained the model a number of times, with a different number of dimensions of the vector
space (see Figure 2a). The more dimensions there are, the more accurate the model becomes,
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
however, the time to train the model also becomes longer. In the rest of the experiments, we
decided to use 128 dimensions. In a second experiment, we varied the size of the skip window,
i.e., how many words to consider to the left and right of the current word in the skip-gram.
The results are displayed in Figure 2b, and show that a skip window of 1 is most ideal for our
dataset.
(a) Results for varying the number of dimen- (b) Results for varying the size of the skip
sions of the vector space.
window.
Figure 2: Evolution of the average loss during training. A step represents 2000 training windows.
4.1 Visualizing the semantic vector space
In order to better understand and evaluate the proposed model, we created visualizations
of selected musical slices in a dimensionally reduced space. We use t-Distributed Stochastic
Neighbor Embedding (t-SNE), a technique developed by Maaten and Hinton [2008] for visualizing high-dimensional data. t-SNE has previously been used in a music analysis context for
visualizing clusters of musical genres based on musical features [Hamel and Eck, 2010].
In this case, we identified the ‘chord’ to which each slice of the dataset belongs based on a
simple template-matching method. We expect that tonally close chords occur together in the
semantic vector space. Figure 3 confirms this hypothesis. When examining slices that contain
C and G chords (a perfect fifth apart), the space looks very dispersed, as they often co-occur
(see Figure 3c). The same occurs for the chord pair Eb and Bb in Figure 3d. On the other
hand, when looking at the tonally distant chord pair E and Eb (Figure 3a), we see that clusters
appear in the reduced vector space. The same happens for the tonally distant chords Eb , Bb
and B in Figure 3b.
4.2 Content versus context
In order to further examine if word2vec captures semantic meaning in music via the modeling
of context, we modify a piece by replacing some of its original slices with the most similar one
as captured by the cosine similarity in the vector space model. If word2vec is really able to
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
(a) E (green) and Eb (blue).
(b) Eb (black), Db (green) and B (gray).
(c) C (green) and G (blue).
(d) Eb (green) and Bb (blue).
Figure 3: Reduced vector space with t-SNE for different slices (labeled by the most close chord)
capture this, the modified piece should sound similar to the original. This allows us to evaluate
the effectiveness of using word2vec for modeling music.
Figure 4 shows the first 17 measures of Beethoven’s piano sonata Op. 27 No. 2 (Moonlight),
2nd movement in (a) and the measures with modified pitch slices in the dashed box in (b).
An audio version of this score is available online1 . The modified slices in (b) are produced by
replacing the original with the slice that has the highest cosine similarity based on the word2vec
embeddings. The tonal distance between the original and modified slices is presented below
each slice pair. This is calculated as the average of the number of steps between each pair of
pitches in the two slices in a tonnetz representation [Cohn, 1997], extended with pitch register.
It can be observed that even thought the cosine similarity is around 0.5, the tonal distance of
1
http://dorienherremans.com/word2vec
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
the selected slice remains relatively low in most of the cases. For example, the tonal distance
in the third dashed box between the modified slice (Db major triad with pitches Db 4 , F4 , and
Ab 4 ) and the original slice of a single pitch Ab 4 is 1.25. However, we notice that word2vec does
not necessarily model musical context for voice leading. For example, better voice leading can
be achieved if the pitch D4 in the last dashed box is replaced with pitch D5 .
Figure 4: (a) An excerpt of Beethoven’s piano sonata Op. 27 No. 2, 2nd movement with (b)
modified measures by replacing with slices that report the highest word2vec cosine
similarity.
In Figure 4b, a number of notes are marked in a different color (orange). These are the
held notes, i.e., their onsets are played previously and the notes remain being played over the
current slice. These notes create a unique situation in music generation using word2vec. For
example, the orange note (pitch Db 5 ) in the first dashed box is a held note, which indicates
that the pitch should have been played in the previous slice. However, word2vec does not
capture this relation; it only considers the similarity between the original and modified slices.
5 Conclusions
A skip-gram model with negative sampling was used to build a semantic vector space model
for complex polyphonic music. By representing the resulting vector space in a reduced twodimensional graph with t-SNE, we show that musical features such as a notion of tonal proximity are captured by the model. Music generated by replacing slices based on word2vec context
similarity also presents close tonal distance compared to the original.
In the future, an embedded model that combines both word2vec with, for instance, a longshort term memory recurrent neural network based on musical features, would offer a more
complete way to more completely model music. The TensorFlow code used in this research is
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
available online2 .
Acknowledgements
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 658914.
References
Kat R Agres, Stephen McGregor, Karolina Rataj, Matthew Purver, and Geraint A Wiggins.
Modeling metaphor perception with distributional semantics vector space models. In Workshop on Computational Creativity, Concept Invention, and General Intelligence. Proceedings
of 5 th International Workshop, C3GI at ESSLI, pages 1–14, 2016.
Mireille Besson and Daniele Schön. Comparison between language and music. Annals of the
New York Academy of Sciences, 930(1):232–258, 2001.
Nicolas Boulanger-Lewandowski, Yoshua Bengio, and Pascal Vincent. Modeling temporal dependencies in high-dimensional sequences: Application to polyphonic music generation and
transcription. arXiv preprint arXiv:1206.6392, 2012.
Richard Cohn. Neo-riemannian operations, parsimonious trichords, and their” tonnetz” representations. Journal of Music Theory, 41(1):1–66, 1997.
Darrell Conklin and Ian H Witten. Multiple viewpoint systems for music prediction. Journal
of New Music Research, 24(1):51–73, 1995.
Douglas Eck and Juergen Schmidhuber. Finding temporal structure in music: Blues improvisation with lstm recurrent networks. In Neural Networks for Signal Processing, 2002.
Proceedings of the 2002 12th IEEE Workshop on, pages 747–756. IEEE, 2002.
Yoav Goldberg and Omer Levy. word2vec explained: Deriving mikolov et al.’s negativesampling word-embedding method. arXiv preprint arXiv:1402.3722, 2014.
Michael U Gutmann and Aapo Hyvärinen. Noise-contrastive estimation of unnormalized statistical models, with applications to natural image statistics. Journal of Machine Learning
Research, 13(Feb):307–361, 2012.
Philippe Hamel and Douglas Eck. Learning features from music audio with deep belief networks. In ISMIR, volume 10, pages 339–344. Utrecht, The Netherlands, 2010.
Cheng-Zhi Anna Huang, David Duvenaud, and Krzysztof Z Gajos. Chordripple: Recommending chords to help novice composers go beyond the ordinary. In Proceedings of the 21st
International Conference on Intelligent User Interfaces, pages 241–250. ACM, 2016.
Elizabeth D Liddy, Woojin Paik, S Yu Edmund, and Ming Li. Multilingual document retrieval
system and method using semantic vector matching, December 21 1999. US Patent 6,006,221.
2
http://dorienherremans.com/word2vec
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine
Learning Research, 9(Nov):2579–2605, 2008.
Stephen McGregor, Kat Agres, Matthew Purver, and Geraint A Wiggins. From distributional
semantics to conceptual spaces: A novel computational method for concept creation. Journal
of Artificial General Intelligence, 6(1):55–86, 2015.
Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word
representations in vector space. arXiv preprint arXiv:1301.3781, 2013a.
Tomas Mikolov, Quoc V Le, and Ilya Sutskever. Exploiting similarities among languages for
machine translation. arXiv preprint arXiv:1309.4168, 2013b.
Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119, 2013c.
Frederic Morin and Yoshua Bengio. Hierarchical probabilistic neural network language model.
In Aistats, volume 5, pages 246–252. Citeseer, 2005.
David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by
back-propagating errors. Cognitive modeling, 5(3):1, 1988.
Hasim Sak, Andrew W Senior, and Françoise Beaufays. Long short-term memory recurrent
neural network architectures for large scale acoustic modeling. In Interspeech, pages 338–342,
2014.
Pang-Ning Tan, Michael Steinbach, and Vipin Kumar. Introduction to Data Mining, (First
Edition). Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 2005. ISBN
0321321367.
Peter D Turney and Patrick Pantel. From frequency to meaning: Vector space models of
semantics. Journal of artificial intelligence research, 37:141–188, 2010.
D. Herremans, and C.H. Chuan. 2017. Modeling Musical Context Using Word2vec. Proceedings of the First
International Workshop on Deep Learning and Music joint with IJCNN. Anchorage, US. May, 2017. 1(1). pp
11-18
| 9cs.NE
|
arXiv:1711.02002v1 [math.AC] 6 Nov 2017
REGULARITY AND H-POLYNOMIALS OF MONOMIAL IDEALS
TAKAYUKI HIBI AND KAZUNORI MATSUDA
A BSTRACT. Let S = K[x1 , . . . , xn ] denote the polynomial ring in n variables over a field K
with each deg xi = 1 and I ⊂ S a homogeneous ideal of S with dim S/I = d. The Hilbert series of S/I is of the form hS/I (λ )/(1 − λ )d , where hS/I (λ ) = h0 + h1 λ + h2λ 2 + · · · + hs λ s
with hs 6= 0 is the h-polynomial of S/I. It is known that, when S/I is Cohen–Macaulay,
one has reg(S/I) = deg hS/I (λ ), where reg(S/I) is the (Castelnuovo–Mumford) regularity
of S/I. In the present paper, given arbitrary integers r and s with r ≥ 1 and s ≥ 1, a monomial ideal I of S = K[x1 , . . . , xn ] with n ≫ 0 for which reg(S/I) = r and deg hS/I (λ ) = s
will be constructed. Furthermore, we give a class of edge ideals I ⊂ S of Cameron–Walker
graphs with reg(S/I) = deghS/I (λ ) for which S/I is not Cohen–Macaulay.
I NTRODUCTION
Let S = K[x1 , . . . , xn ] denote the polynomial ring in n variables over a field K with each
deg xi = 1 and I ⊂ S a homogeneous ideal of S with dim S/I = d. The Hilbert series
HS/I (λ ) of S/I is of the form HS/I (λ ) = (h0 + h1 λ + h2 λ 2 + · · · + hs λ s )/(1 − λ )d , where
each hi ∈ Z ([2, Proposition 4.4.1]). We say that hS/I (λ ) = h0 + h1 λ + h2 λ 2 + · · · + hs λ s
with hs 6= 0 is the h-polynomial of S/I. Let reg(S/I) denote the (Castelnuovo–Mumford )
regularity [2, p. 168] of S/I. A well-known fact (e.g., [3, Lemma 2.5]) is that, when S/I
is Cohen–Macaulay, one has
reg(S/I) = deg hS/I (λ ).
Its converse is false. The following example was found by Jürgen Herzog. Let I ⊂
S = K[x1 , x2 , x3 , x4 ] be the monomial ideal (x32 , x22 x3 , x2 x23 , x33 , x21 , x1 x2 , x1 x3 , x1 x4 ), which is
strongly stable ([7, p. 103]). Then dim S/I = 1, depth S/I = 0, reg(S/I) = 2 and hS/I (λ ) =
1 + 3λ + 2λ 2 . At this stage, it is reasonable to discover a natural class of monomial ideals
I ⊂ S = K[x1 , . . ., xn ] for which S/I is not Cohen–Macaulay with reg(S/I) = deg hS/I (λ ).
Furthermore, one cannot escape from the temptation to present the following
Conjecture 0.1. Given arbitrary integers r and s with r ≥ 1 and s ≥ 1, there exists
a strongly stable ideal I of S = K[x1 , . . . , xn ] with n ≫ 0 for which reg(S/I) = r and
deg hS/I (λ ) = s.
It follows from [2, Lemma 4.1.3] that if I has a pure resolution ([2, p. 153]), then
deg hS/I (λ ) − reg(S/I) = dim S/I − depth S/I.
2010 Mathematics Subject Classification. 05E40, 13H10.
Key words and phrases. Castelnuovo–Mumford regularity, h-polynomial, Cameron–Walker graph.
1
As a result, when 1 ≤ r ≤ s, a desired ideal can be found in the class of squarefree lexsegment ideals ([1], [7, p. 124]).
The purpose of the present paper is to give an affirmative answer to a weak version of
Conjecture 0.1, that is to say, a monomial ideal I of S = K[x1 , . . ., xn ] for which reg(S/I) =
r and deg hS/I (λ ) = s will be constructed.
Theorem 0.2. Given arbitrary integers r and s with r ≥ 1 and s ≥ 1, there exists a monomial ideal I of S = K[x1 , . . . , xn ] with n ≫ 0 for which reg(S/I) = r and deg hS/I (λ ) = s.
When r > s, a basic process in order to obtain a required ideal of Theorem 0.2 is to find
a monomial ideal I = IN ⊂ S with reg(S/I) = N + 1 and deg hS/I (λ ) = 1 for an arbitrary
integer N > 0. A proof of Theorem 0.2 will be achieved in Section 1.
On the other hand, in Section 2, we give a class of edge ideals I ⊂ S of Cameron–Walker
graphs ([8]) with reg(S/I) = deg hS/I (λ ) for which S/I is not Cohen–Macaulay.
1. P ROOF
OF
T HEOREM 0.2
Before giving a proof of Theorem 0.2, several lemmata will be prepared. Let, as before, S = K[x1 , . . ., xn ] denote the polynomial ring in n variables over a field K with each
deg xi = 1. Lemma 1.1 below follows immediately from the definition of regularity in
terms of graded Betti numbers ([7, p. 48]).
Lemma 1.1. Let I ⊂ S be a proper homogeneous ideal. Then reg(S/I) = reg(I) − 1.
Lemma 1.2 ([9, Lemma 3.2]). Let S1 = K[x1 , . . ., xm ] and S2 = K[y1 , . . . , yn ] be polynomial
rings over a field K. Let I1 be a nonzero homogeneous ideal of S1 and I2 that of S2 . Write S
for S1 ⊗K S2 = K[x1 , . . ., xm , y1 , . . ., yn ] and regard I1 + I2 and I1 I2 as homogeneous ideals
of S. Then
(1) reg(I1 I2 ) = reg(I1 ) + reg(I2 ) ;
(2) reg(I1 + I2 ) = reg(I1 ) + reg(I2 ) − 1 ;
(3) reg(S/(I1 + I2 )) = reg(S1 /I1 ) + reg(S2 /I2 ).
Lemma 1.3 ([6, Lemma 2.10]). Let I ⊂ S be a monomial ideal and xi a variable of S which
appears in a monomial belonging to the unique minimal system of monomial generators
of I. Then
reg(I) ≤ max{ reg(I : (x)) + 1, reg(I + (x)) }.
In the first step of a proof of Theorem 0.2, given integers 1 ≤ r ≤ s, we construct a
monomial ideal I ⊂ S with reg(S/I) = r and hS/I (λ ) = s. As was mensioned in Introduction, a desired ideal I can be found in the class of squarefree lexsegment ideals.
Let <lex denote the lexicographic order ([7, p. 24]) on S = K[x1 , . . ., xn ] induced from
x1 > x2 > · · · > xn . A monomial ideal I ⊂ S is called squarefree lexsegment if I is generated
by squarefree monomials and if, for all squarefree monomials u ∈ I and for all squarefree
monomials v ∈ S with deg u = deg v and u <lex v, one has v ∈ I.
2
Fix integers r and s with 1 ≤ r ≤ s and consider the squarefree lexsegment ideal
Ir,s = (u1 u2 · · · ur ur+1 , u1 u2 · · · ur ur+2 , . . . , u1 u2 · · · ur us+1 )
of the polynomial ring K[u1 , . . ., us+1 ] in (s + 1) variables over a field K.
Proposition 1.4. One has
(1) reg(K[u1, . . . , us+1 ]/Ir,s ) = r ;
1 + λ + · · · + λ r−1 + λ r (1 − λ )s−r
(2) HK[u1 ,...,us+1 ]/Ir,s (λ ) =
.
(1 − λ )s
Thus, in particular, deg hK[u1 ,...,us+1 ]/Ir,s (λ ) = s.
Proof. (1) Since Ir,s = (u1 u2 · · · ur )(ur+1, . . . , us+1 ), it follows from Lemma 1.2 (1) that
reg(Ir,s ) = r + 1. Hence Lemma 1.1 says that reg(K[u1 , . . ., us+1 ]/Ir,s) = r, as desired.
(2) Let u = u1 u2 · · · ur . Then Ir,s + (u) = (u) and Ir,s : (u) = (ur+1 , . . . , us+1 ). Thus the
short exact sequence
K[u1, . . . , us+1 ]
K[u1 , . . . , us+1 ]
×u K[u1 , . . . , us+1 ]
(−r) −−→
→
→ 0,
0→
Ir,s : (u)
Ir,s
Ir,s + (u)
yields
HK[u1 ,...,us+1 ]/Ir,s (λ ) = HK[u1 ,...,us+1 ]/(u) (λ ) + λ r · HK[u1,...,us+1 ]/(ur+1 ,...,us+1 ) (λ )
=
λr
1 + λ + · · · + λ r−1
+
(1 − λ )s
(1 − λ )r
=
1 + λ + · · · + λ r−1 + λ r (1 − λ )s−r
,
(1 − λ )s
as required.
Proposition 1.4 guarantees that, when 1 ≤ r ≤ s, Conjecture 0.1 is true.
Now, in the second step of a proof of Theorem 0.2, we turn to the discussion of finding
a desired monomial ideal for 1 ≤ s < r.
Let n ≥ 2 and
Sn = K[x, y1 , . . . , yn , z1 , . . . , zn+1 ]
the polynomial ring in 2(n + 1) variables over a field K. We then introduce the monomial
ideals In , Jn , Kn and Ln defined as follows:
3
n−2
In = (xy1 y2 · · · yn , xz1 z2 · · · zn+1 ) + ∑ (y1 , . . ., yi , yi+1 · · · yn )(zi )
i=1
+ (y1 , . . . , yn )(zn−1 , zn , zn+1 ),
n−1
Jn = (xz1 z2 · · · zn+1 ) + ∑ (zi , . . . , zn+1 )(yi ),
i=1
n−2
Kn = (xy1 y2 · · · yn−1 ) + ∑ (y1 , . . ., yi , yi+1 · · · yn−1 )(zi ),
i=1
n−1
Ln =
∑ (zi, . . ., zn)(yi).
i=1
Remark 1.5. One has
(1) In + (yn ) = (yn ) + Jn ;
(2) In : (yn ) = (zn−1 , zn , zn+1 ) + Kn ;
(3) Jn + (zn+1 ) = (zn+1 ) + Ln ;
(4) Jn : (zn+1 ) = (xz1 z2 · · · zn ) + (y1 , . . . , yn−1 ) ;
(5) Kn + (yn−1 ) = (yn−1 ) + ∑n−2
i=1 (y1 , . . ., yi )(zi ) if n ≥ 3 ;
(6) Kn : (yn−1 ) = (zn−2 ) + Kn−1 if n ≥ 3 ;
(7) Ln can be regarded as the edge ideal ([7, p. 156]) of the Ferrers graph associated
with the partition (n − 1, n − 1, n − 2, . . . , 2, 1) and ∑n−2
i=1 (y1 , . . ., yi )(zi ) appearing
in (5) can be regarded as that with the partition (n − 2, n − 1, . . ., 2, 1)), see [4].
By using [4, Theorem 2.1], one has
!
n−2
reg(Ln ) = reg
∑ (y1, . . . , yi)(zi)
= 2,
i=1
1 + (n − 1)λ − (n − 1)λ 2
,
(1 − λ )n
1 + (n − 2)λ
.
HK[y1 ,...,yn−2 ,z1 ,...,zn−2 ]/ ∑n−2 (y1 ,...,yi )(zi ) (λ ) =
i=1
(1 − λ )n−2
HK[y1 ,...,yn−1 ,z1 ,...,zn]/Ln (λ ) =
Lemma 1.6. One has
(1) reg(Jn ) = n + 2 ;
(2) HK[x,y1 ,...,yn−1 ,z1 ,...,zn+1 ]/Jn (λ ) =
1 + nλ − (n − 2)λ 2 + λ 3 + · · · + λ n+1
.
(1 − λ )n+1
4
Proof. (1) Since xz1 z2 · · · zn+1 belongs to the unique minimal system of monomial generators of Jn , it follows that n + 2 ≤ reg(Jn ). We claim reg(Jn ) ≤ n + 2. By using Lemma
1.2 (2) together with Remark 1.5 (3) and (7), one has
reg(Jn + (zn+1 )) = reg((zn+1 ) + Ln )) = reg(Ln ) = 2.
Furthermore, Lemma 1.2 (2) together with Remark 1.5 (4) says that reg(Jn : (zn+1 )) =
n + 1. Hence, reg(Jn ) ≤ n + 2 follows from Lemma 1.3.
(2) Let T = K[x, y1 , . . ., yn−1 , z1 , . . . , zn+1 ] and consider the short exact sequence
T
T
×zn+1 T
(−1) −−−−→ →
→ 0.
0→
Jn : (zn+1 )
Jn
Jn + (zn+1 )
Remark 1.5 (3) and (7) yield
HT /Jn +(zn+1 ) (λ ) = HT /(zn+1 )+Ln (λ )
= HK[y1,...,yn−1 ,z1 ,...,zn ]/Ln (λ ) ·
1
(1 − λ )
1 + (n − 1)λ − (n − 1)λ 2
.
=
(1 − λ )n+1
Furthermore, Remark 1.5 (4) yeilds
HT /Jn :(zn+1 ) (λ ) = HT /(xz1 z2 ···zn )+(y1 ,...,yn−1 ) (λ )
= HK[x,z1,...,zn ]/(xz1 z2 ···zn ) (λ ) ·
=
1
(1 − λ )
1+λ +···+λn
.
(1 − λ )n+1
Hence
HT /Jn (λ ) = HT /Jn +(zn+1 ) (λ ) + λ · HT /Jn :(zn+1 ) (λ )
=
1 + nλ − (n − 2)λ 2 + λ 3 + · · · + λ n+1
,
(1 − λ )n+1
as desired.
Lemma 1.7. One has
(1) reg(Kn ) = n;
(2) HK[x,y1 ,...,yn−1 ,z1 ,...,zn−2 ]/Kn (λ ) =
i
1 + ∑n−1
i=1 (n − i)λ
.
(1 − λ )n−1
5
Proof. Since xy1 y2 · · · yn−1 belongs to the unique minimal system of monomial generators
of Jn , one has n ≤ reg(Kn ). We claim reg(Kn ) ≤ n and (2) by using induction on n. Since
K2 = (xy1 ), each of the assertion is trivial for n = 2. Let n > 2. Lemma 1.2 (2) together
with Remark 1.5 (5) and (7) guarantees that
!
n−2
reg(Kn + (yn−1 )) = reg (yn−1 ) + ∑ (y1 , . . ., yi )(zi )
= 2.
i=1
Moreover, by virtue of Lemma 1.2 (2), Remark 1.5 (6) as well as the induction hypothesis,
it follows that
reg(Kn : (yn−1 )) = reg((zn−2 ) + Kn−1 ) = reg(Kn−1 ) = n − 1.
Hence, Lemma 1.3 says that reg(Kn ) ≤ n.
Now, consider the short exact sequence
0→
′
T′
T′
×yn−1 T
(−1) −−−−→
→
→ 0,
Kn : (yn−1 )
Kn
Kn + (yn−1 )
where T ′ = K[x, y1 , . . . , yn−1 , z1 , . . ., zn−2 ]. It follows from Remark 1.5 (5) and (7) that
HT ′ /Kn +(yn−1 ) (λ ) = HT ′ /(yn−1 )+∑n−2 (y1 ,...,yi )(zi ) (λ )
i=1
= HK[y1 ,...,yn−2 ,z1 ,...,zn−2 ]/ ∑n−2 (y1 ,...,yi )(zi ) (λ ) ·
i=1
1
1−λ
1 + (n − 2)λ
.
(1 − λ )n−1
Furthermore, Remark 1.5 (6) as well as the induction hypothesis guarantees that
=
HT ′ /Kn :(yn−1 ) (λ ) = KT ′ /(zn−2 )+Kn−1 (λ )
= HK[x,y1,...,yn−2 ,z1 ,...,zn−3 ]/Kn−1 (λ ) ·
i
1 + ∑n−2
i=1 (n − 1 − i)λ
.
=
(1 − λ )n−1
6
1
1−λ
Hence, one has
HT ′ /Kn (λ ) = HT ′ /Kn +(yn−1 ) (λ ) + λ · HT ′ /Kn :(yn−1 ) (λ )
i
1 + (n − 2)λ + λ + λ · ∑n−2
i=1 (n − 1 − i)λ
=
(1 − λ )n−1
i+1
1 + (n − 1)λ + ∑n−2
i=1 (n − 1 − i)λ
=
(1 − λ )n−1
=
i
1 + (n − 1)λ + ∑n−1
i=2 (n − i)λ
(1 − λ )n−1
=
i
1 + ∑n−1
i=1 (n − i)λ
,
(1 − λ )n−1
as desired.
The monomial ideal In plays an important role in our proof of Theorem 0.2.
Proposition 1.8. One has
(1) reg(Sn /In ) = n + 1 ;
1 + (n + 1)λ
.
(2) HSn /In (λ ) =
(1 − λ )n+1
Thus, in particular, deg hSn /In (λ ) = 1.
Proof. (1) By virtue of Lemma 1.1, it is sufficient to show that reg(In ) = n + 2. Since
xz1 z2 · · · zn+1 belongs to the unique minimal system of monomial generators of In , one has
n + 2 ≤ reg(In ). We claim reg(In ) ≤ n + 2. It follows from Lemma 1.2 (2), Remark 1.5
(1) together with Lemma 1.6 (1) that
reg(In + (yn )) = reg((yn ) + Jn ) = reg(Jn ) = n + 2.
By using Lemma 1.2 (2), Remark 1.5 (2) together with Lemma 1.7 (1), one has
reg(In : (yn )) = reg((zn−1 , zn , zn+1 ) + Kn ) = reg(Kn ) = n.
Hence Lemma 1.3 says that reg(In ) ≤ n + 2, as desired.
(2) Considering the short exact sequence
0→
Sn
Sn
×yn Sn
(−1) −−−→
→
→ 0.
In : (yn )
In
In + (yn )
7
Remark 1.5 (1) together with Lemma 1.6 (2) yields
HSn /In +(yn ) (λ ) = HSn /(yn )+Jn (λ )
= HK[x,y1 ,...,yn−1 ,z1 ,...,zn+1 ]/Jn (λ )
1 + nλ − (n − 2)λ 2 + λ 3 + · · · + λ n+1
.
=
(1 − λ )n+1
Furthermore, Remark 1.5 (2) together with Lemma 1.7 (2) yields
HSn /In :(yn ) (λ ) = HSn /(zn−1 ,zn ,zn+1 )+Kn (λ )
= HK[x,y1,...,yn−1 ,z1 ,...,zn−2 ]/Kn (λ ) ·
=
1
1−λ
i
1 + ∑n−1
i=1 (n − i)λ
.
(1 − λ )n
It then follows that
HSn /In (λ ) = HSn /In +(yn ) (λ ) + λ · HSn/In :(yn ) (λ )
i+1
1 + nλ − (n − 2)λ 2 + λ 3 + · · · + λ n+1 t + ∑n−1
i=1 (n − i)λ
=
+
(1 − λ )n+1
(1 − λ )n
=
i+1
1 + nλ − (n − 2)λ 2 + λ 3 + · · · + λ n+1 + λ (1 − λ ) + (1 − λ ) ∑n−1
i=1 (n − i)λ
(1 − λ )n+1
=
i+1 − n−1 (n − i)λ i+2
1 + (n + 1)λ − (n − 1)λ 2 + ∑ni=2 λ i+1 + ∑n−1
∑i=1
i=1 (n − i)λ
n+1
(1 − λ )
=
i+1 − n (n − i + 1)λ i+1
1 + (n + 1)λ − (n − 1)λ 2 + ∑ni=2 λ i+1 + ∑n−1
∑i=2
i=1 (n − i)λ
n+1
(1 − λ )
=
1 + (n + 1)λ − (n − 1)λ 2 + ∑ni=2 λ i+1 + (n − 1)λ 2 − ∑ni=2 λ i+1
(1 − λ )n+1
1 + (n + 1)λ
,
(1 − λ )n+1
as required.
=
We are now in the position to finish a proof of Theorem 0.2.
8
Proof. (Proof of Theorem 0.2.) Let r and s be positive integers with r, s ≥ 1. By virtue of
Proposition 1.4, only the case of r > s ≥ 1 will be discussed. Let S1 = K[x, y1 , z1 , z2 ] and
1+2λ
I1 = (xy1 , xz1 z2 , y1 z1 , y1 z2 ). It then follows that reg(S1 /I1 ) = 2 and HS1 /I1 (λ ) = (1−
.
λ )2
By virtue of this example and of Proposition 1.8, one has
reg(Sr−s /Ir−s) = r − s + 1, HSr−s /Ir−s (λ ) =
1 + (r − s + 1)λ
.
(1 − λ )r−s+1
Let S = Sr−s ⊗K K[u1, . . . , us ] and I = Ir−s + (u1 u2 · · · us ). Lemma 1.2 (3) together with
Proposition 1.4 yields
reg(S/I) = r − s + 1 + s − 1 = r
and
{1 + (r − s + 1)λ }(1 + λ + · · · + λ s−1 )
.
(1 − λ )r
Hence deg hS/I (λ ) = s and I is a desired monomial ideal.
HS/I (λ ) =
2. E XAMPLES
The purpose of this section is to give a class of edge ideals I ⊂ S of Cameron–Walker
graphs ([8]) with reg(S/I) = deg hS/I (λ ) for which S/I is not Cohen–Macaulay.
Let G be a finite simple graph on the vertex set [n] = {1, . . ., n} and E(G) its edge set.
(A finite graph is called simple if it possesses no loop and no multiple edge.) The edge
ideal I(G) of G is the monomial ideal of S = K[x1 , . . . .xn ] generated by those quadratic
monomials xi x j with {i, j} ∈ E(G):
I(G) = ( xi x j : {i, j} ∈ E(G) ) ⊂ S.
In general, it is quite difficult to compute the regularity of an edge ideal. However, one
can compute reg(I(G)) easily if G is a Cameron–Walker graph. The notion of CameronWalker graph was introduced by [5]. We refer the reader to [8, p. 258] for a classification
of Cameron–Walker graphs.
Example 2.1. Fix m ≥ 1 and write G1m for the star triangle joining m triangles at one
common vertex. Then S/I(G1m) is Cohen–Macaulay if and only if m = 1 ([8, Theorem
1.3]). Hence
• If m = 2k, then reg(S/I(G1m)) = 2k > 2k − 1 = deg hS/I(G1m ) (λ ).
• If m = 2k + 1, then reg(S/I(G1m)) = deg hS/I(G1m ) (λ ) = 2k + 1.
9
Example 2.2. Fix m ≥ 1 and write G2m for the graph drawn below on the vertex set [2m +
3]:
G2m =
1
2m + 3
2m + 2
❄▲▲2m + 1
r⑧r⑧r⑧ ❄❄▲❄▲▲▲
r
r
❄❄ ▲▲
rr ⑧⑧
❄❄ ▲▲▲
rrr ⑧⑧⑧ · · ·
❄❄ ▲
r
❄
r❄ ⑧⑧
❄❄ ⑧
⑧
⑧
⑧
2m
2m − 1
2
Then [8, Theorem 1.3] says that G2m is not Cohen–Macaulay if m ≥ 2. However,
reg(S/I(G2m)) = deg hS/I(G2m ) (λ ) = m + 1.
On the other hand, this graph is of interest from the viewpoint of h-vector (h0 , h1 , . . ., hs ),
which is the sequence of coefficients of h-polynomial. It is known [10, Theorem 4.4]
that the h-vector of Gorenstein ring is symmetric, but the converse is false. A routine
computation shows that
hS/I(G2m ) (λ ) = (1 + λ )m+1 + λ (1 − λ )m−1 .
Hence h-vector of S/I(G2m) is symmetric if m is odd, but not necessary unimodal. For
example, the h-vector of S/I(G23) is (1, 5, 4, 5, 1), which is not unimodal. In general, the
h-vector of S/I(G2m ) is not unimodal if m = 4k + 3.
Acknowledgment. The first author was partially supported by JSPS KAKENHI 26220701.
The second author was partially supported by JSPS KAKENHI 17K14165.
R EFERENCES
[1] A. Aramova, J. Herzog and T. Hibi, Squarefree lexsegment ideals, Math. Z. 228 (1998), 353–378.
[2] W. Bruns and J. Herzog, Cohen-Macaulay Rings, Revised ED., Cambridge Stud. Adv. Math., vol. 39,
Cambridge University Press, Cambridge, 1998.
[3] B. Benedetti and M. Varbaro, On the dual graph of Cohen-Macaulay algebras, Int. Math. Res. Not.
IMRN 17 (2015), 8085–8115.
[4] A. Corso and U. Nagel, Monomial and toric ideals associated to Ferrers graphs, Trans. Amer. Math.
Soc. 361 (2009), 1371–1395.
[5] K. Cameron and T. Walker, The graphs with maximum induced matching and maximum matching the
same size, Discrete Math. 299 (2005), 49–55.
[6] H. Dao, C. Huneke and J. Schweig, Bounds on the regularity and projective dimension of ideals
associated to graphs, J. Algebraic Combin. 38 (2013), 37–55.
[7] J. Herzog and T. Hibi, Monomial ideals, Graduate Texts in Mathematics 260, Springer, London, 2010.
10
[8] T. Hibi, A. Higashitani, K. Kimura and A. B. O’Keefe, Algebraic study on Cameron–Walker graphs,
J. of Algebra 422 (2015), 257–269.
[9] L. T. Hoa and N. D. Tam, On some invariants of a mixed product of ideals, Arch. Math. (Basel) 94
(2010), 327–337.
[10] R. Stanley, Hilbert functions of graded algebras, Adv. Math. 28 (1978), 57–83.
TAKAYUKI H IBI , D EPARTMENT OF P URE AND A PPLIED M ATHEMATICS , G RADUATE S CHOOL OF
I NFORMATION S CIENCE AND T ECHNOLOGY, O SAKA U NIVERSITY, S UITA , O SAKA 565-0871, JAPAN
E-mail address: [email protected]
K AZUNORI M ATSUDA , D EPARTMENT OF P URE AND A PPLIED M ATHEMATICS , G RADUATE S CHOOL
OF I NFORMATION S CIENCE AND T ECHNOLOGY, O SAKA U NIVERSITY, S UITA , O SAKA 565-0871, JAPAN
E-mail address: [email protected]
11
| 0math.AC
|
1
Correntropy Maximization via ADMM
– Application to Robust Hyperspectral Unmixing –
arXiv:1602.01729v1 [stat.ML] 4 Feb 2016
Fei Zhu, Abderrahim Halimi, Paul Honeine, Badong Chen, Nanning Zheng
Abstract
In hyperspectral images, some spectral bands suffer from low signal-to-noise ratio due to noisy acquisition and atmospheric
effects, thus requiring robust techniques for the unmixing problem. This paper presents a robust supervised spectral unmixing
approach for hyperspectral images. The robustness is achieved by writing the unmixing problem as the maximization of the
correntropy criterion subject to the most commonly used constraints. Two unmixing problems are derived: the first problem
considers the fully-constrained unmixing, with both the non-negativity and sum-to-one constraints, while the second one deals with
the non-negativity and the sparsity-promoting of the abundances. The corresponding optimization problems are solved efficiently
using an alternating direction method of multipliers (ADMM) approach. Experiments on synthetic and real hyperspectral images
validate the performance of the proposed algorithms for different scenarios, demonstrating that the correntropy-based unmixing
is robust to outlier bands.
Index Terms
Correntropy, maximum correntropy estimation, alternating direction method of multipliers, hyperspectral image, unmixing problem.
F. Zhu is with the Institut Charles Delaunay (CNRS), Université de Technologie de Troyes, France. ([email protected])
A. Halimi is with the School of Engineering and Physical Sciences, Heriot-Watt University, U.K. ([email protected])
P. Honeine is with the LITIS lab, Université de Rouen, France. ([email protected])
Badong Chen and Nanning Zheng are with the Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an, China.
(chenbd; [email protected])
2
I. I NTRODUCTION
S
Pectral unmixing is an essential issue in many disciplines, including signal and image processing, with a wide range of
applications, such as classification, segmentation, material identification and target detection. Typically, a hyperspectral
image corresponds to a scene taken at many continuous and narrow bands across a certain wavelength range; namely, each
pixel is a spectrum. Assuming that each spectrum is a mixture of several pure materials, the unmixing problem consists
in two tasks: (i) identifying these pure materials (the so-called endmembers); (ii) estimating their proportions (the so-called
abundances) at each pixel [1]. In practice, these two steps can be performed either sequentially or simultaneously [2]. Wellknown endmember extraction algorithms include the pure-pixel-based ones, e.g., the vertex component analysis (VCA) [3]
and the N-FINDR [4], as well as the minimum-volume-based ones, e.g., the minimum simplex analysis [5] and the minimum
volume constrained nonnegative matrix factorization [6]. While the endmember extraction is relatively easy from geometry,
the abundance estimation remains an open problem. Indeed, the abundances can be estimated using least-squares methods,
geometric approaches [2], or by tackling recently-raised issues such as nonlinearity [7], [8]. In this paper, we consider the
abundance estimation problem.
The linear mixture model (LMM) is the most investigated over the past decades [6], [9], [10]. Its underlying premise is
that each pixel/spectrum is a linear combination of the endmembers. To be physically interpretable, two constraints are often
enforced in the estimation problem: the abundance non-negativity constraint (ANC) and the abundance sum-to-one constraint
(ASC) for each pixel [11]. Considering both constraints, the fully-constrained least-squares method (FCLS) was presented
in [9]. A more recently proposed unmixing algorithm is the so-called SUnSAL, for Sparse Unmixing by variable Splitting
and Augmented Lagrangian [12]. It addresses the same optimization problem by taking advantage of the alternating direction
method of multipliers (ADMM) [13]. A constrained-version of SUnSAL was also proposed to solve the constrained sparse
regression problem, where the sum-to-one constraint (ASC) is relaxed and the ℓ1 -norm regularizer is added.
All these unmixing algorithms hugely suffer from noisy data and outliers within bands. Indeed, in real hyperspectral images
for remote sensing, a considerable proportion (about 20%) of the spectral bands are noisy with low SNR, due to the atmospheric
effect such as water absorption [14]. These bands need to be removed prior to applying any existing unmixing method; otherwise,
the unmixing quality drastically decreases. Such sensitivity to outliers is due to the investigated ℓ2 -norm as a cost function in
the FCLS and SUnSAL algorithms, as well as all unmixing algorithms that explore least-squares solutions. It is worth noting
that nonlinear unmixing algorithms also suffer from this drawback, including the kernel-based fully-constrained least-squares
(KFCLS) [15], nonlinear fluctuation methods [7] and post-nonlinear methods [16].
Information theoretic learning provides an elegant alternative to the conventional minimization of the ℓ2 -norm in least-squares
3
problems, by considering the maximization of the so-called correntropy [17], [18]. Due to its stability and robustness to noise
and outliers, the correntropy maximization is based on theoretical foundations and has been successfully applied to a wide class
of applications, including cancer clustering [19], face recognition [20], and recently hyperspectral unmixing [21], to name a
few. In these works, the resulting problem is optimized by the half-quadratic technique [22], either in a supervised manner [20]
or as an unsupervised nonnegative matrix factorization [19], [21].
In this paper, we consider the hyperspectral unmixing problem by defining an appropriate correntropy-based criterion, thus
taking advantage of its robustness to large outliers, as opposed to the conventional ℓ2 -norm criteria. By including constraints
commonly used for physical interpretation, we propose to solve the resulting constrained optimization problems with alternating
direction method of multipliers (ADMM) algorithms. Indeed, the ADMM approach splits a hard problem into a sequence of
small and handful ones [13]. Its relevance to solve nonconvex problems was studied in [13, Section 9]. We show that ADMM
provides a relevant framework for incorporating different constraints raised in the unmixing problem. We present the socalled CUSAL (for Correntropy-based Unmixing by variable Splitting and Augmented Lagrangian), and study in particularly
two algorithms: CUSAL-FC to solve the fully-constrained (ANC and ASC) correntropy-based unmixing problem, and the
CUSAL-SP to solve the sparsity-promoting correntropy-based unmixing problem.
The rest of the paper is organized as follows. We first provide a succinct survey on the classical unmixing problems in
Section II. In Section III, we propose the correntropy-based unmixing problems subject to the aforementioned constraints,
and study the robustness. The resulting optimization problems are solved by the ADMM algorithms described in Section IV.
Experiments on synthetic and real hyperspectral images are presented in Sections V and VI, respectively. Finally, Section VII
provides some conclusions and future works.
II. C LASSICAL U NMIXING P ROBLEMS
The linear mixture model (LMM) assumes that each spectrum can be expressed as a linear combination of a set of pure
material spectra, termed endmembers [1]. Consider a hyperspectral image and let Y ∈ RL×T denote the matrix of the T
pixels/spectra of L spectral bands. Let y ∗t be its t-th column and y l∗ its l-th row, representing the l-th band of all pixels. For
notation simplicity, we denote y t = y ∗t , for t = 1, . . . , T . The LMM can be written as
yt =
R
X
xrt mr + nt = M xt + nt ,
(1)
r=1
where M = [m1 · · · mR ] ∈ RL×R is the matrix composed by the R endmembers with mr = [m1r · · · mLr ]⊤ , xt =
[x1t · · · xRt ]⊤ is the abundance vector associated with the t-th pixel, and nt ∈ RL is the additive noise. In matrix form for
all pixels, we have Y = M X + N , where X = [x1 · · · xT ] ∈ RR×T and N is the noise matrix.
4
In the following, the endmembers are assumed known, either from ground-truth information or by using any endmember
extraction technique. The spectral unmixing problem consists in estimating the abundances for each pixel, often by solving
the least-squares optimization problem
min kyt − M xt k22 ,
xt
(2)
for each t = 1, . . . , T , where k · k2 denotes the conventional ℓ2 -norm. The solution to this conventional least-squares problem
is given by the pseudo-inverse of the (tall) endmember matrix, with xt = (M ⊤ M )−1 M ⊤ y t . The least-squares optimization
problems (2), for all t = 1, . . . , T , are often written in a single optimization problem using the following matrix formulation
min kY − M Xk2F ,
X
(3)
where k · k2F denotes the Frobenius norm. Its solution is
X LS = (M ⊤ M )−1 M ⊤ Y .
(4)
Finally, this optimization problem can be also tackled by considering all the image pixels at each spectral band, which yields
the following least-squares optimization problem
min
X
L
X
l=1
ky l∗ − (M X)l∗ k22 ,
where (·)l∗ denotes the l-th row of its argument. While all these problem formulations have a closed-form solution, they suffer
from two major drawbacks. The first one is that several constraints need to be imposed in order to have a physical meaning
of the results. The second drawback is its sensitivity to noise and outliers, due to the use of the ℓ2 -norm as a fitness measure.
These two drawbacks are detailed in the following.
To be physically interpretable, the abundances should be nonnegative (ANC) and satisfy the sum-to-one constraint (ASC).
Considering both constraints, the fully-constrained least-squares problem is formulated as, for each t = 1, . . . , T ,
min kyt − M xt k22 , subject to xt 0 and 1⊤ xt = 1,
xt
where 1 ∈ RR×1 denotes the column vector of ones and 0 is the non-negativity applied element-wise; In matrix form:
min kY − M Xk2F , subject to X 0
X
and 1⊤ xt = 1, for t = 1, . . . , T.
Since there is no closed-form solution when dealing with the non-negativity constraint, several iterative techniques have been
proposed, such as the active set scheme with the Lawson and Hanson’s algorithm [23], the multiplicative iterative strategies
[24], and the fully-constrained least-squares (FCLS) technique [9]. More recently, the alternating direction method of multipliers
(ADMM) was applied with success for hyperspectral unmixing problem, with the SUnSAL algorithm [12].
5
Recent work in hyperspectral unmixing have advocated the sparsity of the abundance vectors [12], [25], [26]. In this case,
each spectrum is fitted by a sparse linear mixture of endmembers, namely only the abundances with respect to a small number
of endmembers are nonzero. To this end, the sparsity-promoting regularization with the ℓ1 -norm is included in the cost function,
yielding the following constrained sparse regression problem [12], for each t = 1, . . . , T ,
min kyt − M xt k22 + λkxt k1 , subject to xt 0,
xt
where the parameter λ balances the fitness of the least-squares solution and the sparsity level. It is worth noting that the ASC
is relaxed when the ℓ1 -norm is included. This problem is often considered by using the following matrix formulation
min kY − M Xk2F + λ
X
T
X
t=1
kxt k1 , subject to X 0.
Sensitivity to outliers
All the aforementioned algorithms rely on solving a (constrained) least-squares optimization problem, thus inheriting the
drawbacks of using the ℓ2 -norm as the fitness measure. A major drawback is its sensitivity to outliers, where outliers are
some spectral bands that largely deviate from the rest of the bands. Indeed, considering all the image pixels, the least-squares
optimization problems take the form
min
X
L
X
l=1
ky l∗ − (M X)l∗ k22 ,
(5)
subject to any of the aforementioned constraints. From this formulation, it is easy to see how the squared ℓ2 -norm gives
more weight to large residuals, namely to outliers in which predicted values (M X)l∗ are far from actual observations y l∗ .
Moreover, it is common for hyperspectral images to present up to 20% of unusable spectral bands due to low signal-to-noise
ratio essentially from atmospheric effects, such as water absorption. In the following section, we overcome this difficulty
by considering the correntropy maximization principle from the information theoretic learning, which yields an optimization
problem that is robust to outliers.
III. C ORRENTROPY- BASED U NMIXING P ROBLEMS
In this section, we examine the correntropy and write the unmixing problems as correntropy maximization ones. Algorithms
for solving these problems are derived in Section IV.
A. Correntropy
The correntropy, studied in [17], [18], is a nonlinear local similarity measure. For two random variables, Y and its estimation
Yb using some model/algorithm, it is defined by
b
IE[κ(Y, Y)],
(6)
6
where IE[·] is the expectation operator, and κ(·, ·) is a shift-invariant kernel satisfying the Mercer theorem [27]. In practice, while
the joint distribution function of Y and Yb is unavailable, the sample estimator of correntropy is adopted instead. Employing a
b l∗ )}L
finite number of data {(y l∗ , y
l=1 , it is estimated by
L
1X
b l∗ ),
κ(y l∗ , y
L
(7)
l=1
up to a normalization factor. The Gaussian kernel is the most commonly-used kernel for correntropy [17], [20], [28]. This
leads to the following expression for the correntropy
L
1X
exp
L
l=1
−1
2σ2 ky l∗
where σ denotes the bandwidth of the Gaussian kernel.
b l∗ k22 ,
−y
(8)
The maximization of the correntropy, given by
L
max
b1∗ ,...,b
y
yL∗
1X
b l∗ ),
κ(y l∗ , y
L
l=1
is termed the maximum correntropy criterion [17]. It is noteworthy that well-known second-order statistics, such as the mean
square error (MSE) depends heavily on the Gaussian and linear assumptions [17]. However, in presence of non-Gaussian
noise and in particular large outliers, i.e., observations greatly deviated from the data bulk, the effectiveness of the MSE-based
algorithms will significantly deteriorate [29]. By contrast, the maximization of the correntropy criterion is appropriate for
non-Gaussian signal processing, and is robust in particular against large outliers, as shown next.
B. The underlying robustness of the correntropy criterion
In this section, we study the sensitivity to outliers of the correntropy maximization principle, by showing the robustness of
the underlying mechanism. To this end, we examine the behavior of the correntropy in terms of the residual error defined by
bl∗ k2 . Thus, the correntropy (8) becomes
ǫl = ky l∗ − y
L
1X
exp
L
l=1
−1 2
2σ2 ǫl
.
Compared with second-order statistics, e.g. MSE, the correntropy is more robust with respect to the outliers, as shown in
Fig. 1 illustrating the second-order and the correntropy objective functions in terms of the residual error. As the residual error
increases, the second-order function keeps increasing dramatically. On the contrary, the correntropy is only sensitive within a
region of small residual errors, this region being controlled by the kernel bandwidth. For large magnitudes of residual error,
the correntropy falls to zero. Consequently, the correntropy criterion is robust to large outliers.
7
1.5
second-order function
correntropy (σ = 0.5)
correntropy (σ = 2)
correntropy (σ = 5)
1
0.5
0
1
2
3
4
5
6
residual error
Fig. 1.
Illustration of second-order objective function (solid line) and the correntropy objective function (dashed lines) with different values of the kernel
bandwidth.
C. Correntropy-based unmixing problems
The correntropy-based unmixing problem consists in estimating the unknown abundance matrix X, by minimizing the
objective function C (the negative of correntropy), given by
C(X) = −
L
X
exp
l=1
−1
2σ2 ky l∗
− (M X)l∗ k22 ,
(9)
where the Gaussian kernel was considered, or equivalently
C(X) = −
L
X
exp
−1
2σ2
T
X
t=1
l=1
ylt −
R
X
r=1
xrt ml r
2
!
.
(10)
Considering both the ANC and ASC constraints, the fully-constrained correntropy unmixing problem becomes
min C(X), subject to X 0
X
(11)
⊤
and 1 xt = 1, for t = 1, . . . , T.
For the sake of promoting sparse representations, the objective function (9)-(10) can be augmented by the ℓ1 -norm penalty on
the abundance matrix X, leading to the following problem
min C(X) + λ
X
T
X
t=1
kxt k1 , subject to X 0.
(12)
8
IV. ADMM FOR S OLVING
THE
C ORRENTROPY- BASED U NMIXING P ROBLEMS
We first briefly review the alternating direction method of multipliers (ADMM), following the expressions in [13, Chap. 3].
Consider an optimization problem of the form
min f (x) + g(x),
x
where the functions f and g are closed, proper and convex. The ADMM solves the equivalent constrained problem
min f (x) + g(z) subject to Ax + Bz = c,
(13)
x,z
such as having the particular constraint x = z for instance. While this formulation may seem trivial, the optimization problem
can now be tackled using the augmented Lagrangian method where the objective function is separable in x and z. By alternating
on each variable separately, the ADMM repeats a direct update of the dual variable. In its scaled form, the ADMM algorithm
is summarized in Algorithm 1.
Algorithm 1 The ADMM algorithm [13]
Input: functions f and g, matrices A and B, vector c, parameter ρ
1:
Initialize k = 0, x0 , z 0 and u0
2:
repeat
3:
xk+1 = arg minx f (x) + ρ2 kAx + Bz k − c + uk k22 ;
4:
z k+1 = arg minz g(z) + ρ2 kAxk+1 + Bz − c + uk k22 ;
5:
uk+1 = uk + Axk+1 + Bz k+1 − c;
6:
k = k + 1;
7:
until stopping criterion
A. Correntropy-based unmixing with full-constraints
In the following, we apply the ADMM algorithm to solve the correntropy-based unmixing problem in the fully-constrained
case, presented in (11). The main steps are summarized in Algorithm 2. Rewrite the variables to be optimized in a vector
⊤ ⊤
x ∈ RRT ×1 , which is stacked by the columns of the matrix X, namely x = [x⊤
1 · · · xT ] . Rewrite also the following vectors
⊤ ⊤
⊤
⊤ ⊤
⊤
⊤
in RRT ×1 : z = [z ⊤
1 · · · z T ] and u = [u1 · · · uT ] , where, for t = 1, . . . , T , z t = [z1t · · · zRt ] and ut = [u1t · · · uRt ] .
9
By following the formulation of the ADMM in Algorithm 1, we set
f (x) = C(x) +
T
X
ι{1} (1⊤ xt )
(14)
t=1
g(z) = ιRRT
(z)
+
A = −I, B = I and c = 0,
where I is the identity matrix, 0 ∈ RRT ×1 is the zero vector and ιS (u) is the indicator function of the set S defined by
0
if u ∈ S;
ιS (u) =
∞
otherwise.
In this case, the subproblem of the x-update (in line 3 of Algorithm 1) addresses a nonconvex problem without any closed-
form solution. To overcome this difficulty, we apply an inexact ADMM variant in lines 3-5 of Algorithm 2, which solves the
subproblem iteratively using the gradient descent method, instead of solving it exactly and explicitly.
Before that, we eliminate the T equality constraints, i.e., the sum-to-one constraints, by replacing xRt with
xRt = 1 −
R−1
X
xrt ,
r=1
for t = 1, . . . , T . Let x ∈ R(R−1)T ×1 be the reduced vector of (R − 1) unknowns to be estimated, stacked by
⊤
xt = x1t · · · x(R−1)t ,
for t = 1, . . . , T . By this means, the objective function in (14) is transformed from (10) into the reduced-form
!
L
T
X
−1 X
f1 (x) = −
exp
ǫl (xt )2 ,
2σ 2 t=1
(15)
l=1
where ǫl (xt ) = ylt − mlR −
where
∂f1
∂xt
=
h
∂f1
∂x1t
···
PR−1
p=1
∂f1
∂x(R−1)t
(mlp − mlR )xpt , for l = 1, . . . , L. The gradient of (15) with respect to x is stacked as
#⊤
"
∂f1
∂f1 ⊤
∂f1 ⊤
∈ R(R−1)T ×1 ,
···
=
∂x
∂x1
∂xT
i⊤
, with the entries given by
L
T
X
∂f1 (x) 1 X
−1
= σ2 (mlR − mlr ) exp 2σ
ǫl (xs )2 ǫl (xt ),
2
∂xrt
s=1
l=1
for all r = 1, . . . , (R − 1) and t = 1, . . . , T . Similarly, the function ρ2 kx − z k − uk k22 is expressed with respect to x as
φ(x) =
with the entries in its gradient
T
R−1
2 R−1
X
ρ X X
2
1−
xpt − zRt,k − uRt,k +
(xpt − zpt,k − upt,k )
2 t=1
p=1
p=1
∂φ
∂x
given by
R−1
X
∂φ(x)
= ρ xrt +
xpt − 1 + zRt,k − zrt,k + uRt,k − urt,k ,
∂xrt
p=1
(16)
10
for all r = 1, . . . , R − 1 and t = 1, . . . , T .
The solution of the z-update in line 4 Algorithm 1 becomes the projection of xk+1 − uk onto the first orthant, as shown
in line 7 of Algorithm 2.
Algorithm 2 Correntropy-based unmixing with full-constraints (CUSAL-FC)
1: Initialize k = 0, ρ > 0, η > 0, σ > 0; x0 , z 0 and u0 ;
2:
3:
4:
repeat
repeat
1
+
xk+1 = xk+1 − η( ∂x∂fk+1
5:
until convergence
6:
reform xk+1 using xk+1 ;
7:
z k+1 = max(0, xk+1 − uk );
8:
uk+1 = uk − (xk+1 − z k+1 );
9:
k = k + 1;
10:
∂φ
∂xk+1 );
until stopping criterion
B. Sparsity-promoting unmixing algorithm
In order to apply the ADMM algorithm, we express the constrained optimization problem (12) as follows
f (x) = C(x)
(17)
g(x) = ιRRT
(x) + λkxk1
+
A = −I, B = I and c = 0.
By analogy with the previous case, the x-update in line 3 of Algorithm 1 is solved iteratively with the gradient descent method
and is given in Algorithm 3 lines 3-5. The gradient of (17) with respect to xis stacked by
!
L
T
∂f
1 X
−1 X
2
=− 2
ǫl (xt ) exp
(ǫl (xs ))
m⊤
l ,
∂xt
σ
2σ 2 s=1
∂f
∂xt ,
where
l=1
for t = 1, . . . , T , where ǫl (xt ) = ylt −
PR
r=1
xrt mlr .
The z-update in line 4 Algorithm 1 involves solving
1
z k+1 = arg min ιRRT
(z) + (λ/ρ)kzk1 + kz − xk+1 − uk k22 .
+
z
2
(18)
In [13], the ADMM has been applied to solve various ℓ1 -norm problems, including the well-known LASSO [30]. The only
difference between (18) and the z-update in LASSO is that in the latter, no non-negativity term ιR+RT (z) is enforced. In this
11
case, the z-update in LASSO is the element-wise soft thresholding operation
z k+1 = Sλ/ρ (xk+1 − uk ),
where the soft thresholding operator [13] is defined by
ζ −b
Sb (ζ) =
0
ζ + b
if ζ > b;
if kζk < b;
if ζ < −b.
Following [12], it is straightforward to project the result onto the nonnegative orthant in order to include the non-negativity
constraint, thus yielding
z k+1 = max(0, Sλ/ρ (xk+1 − uk )),
where the maximum function is element-wise. All these results lead to the correntropy-based unmixing algorithm with sparsitypromoting, as summarized in Algorithm 3.
Algorithm 3 Correntropy-based unmixing with sparsity-promoting (CUSAL-SP))
1: Initialize k = 0, ρ > 0, σ > 0, η > 0, λ > 0; x0 , z 0 and u0 ;
2:
3:
4:
repeat
repeat
xk+1 = xk+1 − η
∂f
∂xk+1
+ ρ(xk+1 − z k − uk ) ;
5:
until convergence
6:
z k+1 = max(0, Sλ/ρ (xk+1 − uk ));
7:
uk+1 = uk − (xk+1 − z k+1 );
8:
k = k + 1;
9:
until stopping criterion
C. On the initialisation and the bandwidth determination
We apply a three-fold stopping criterion for Algorithms 2 and 3, according to [13], [12]: (i) the primal and dual residuals
are small enough, namely kxk+1 − z k+1 k2 ≤ ǫ1 and ρkz k+1 − z k k2 ≤ ǫ2 , where ǫ1 = ǫ2 =
√
RT × 10−5 as in [12], (ii) the
primal residual starts to increase, i.e., kxk+1 − z k+1 k2 > kxk − z k k2 , or (iii) the maximum iteration number is attained.
The bandwidth σ in the Gaussian kernel should be well-tuned. Note that a small value for this parameter punishes harder the
outlier bands, thus increasing the robustness of the algorithm to outliers [20]. Note that, in this study, the ADMM is applied to
12
address a nonconvex objective function, thus no convergence is guaranteed theoretically, according to [13]. Considering these
issues, we propose to fix the bandwidth empirically as summarized in Algorithm 4 and described next.
Following [20], [21], we first initialize the bandwidth parameter as a function of the reconstruction error, given by
σ02 =
R
kY − M X LS k2F ,
8L
(19)
where X LS is the least-squares solution (4). In the case of a result too apart from that of least-squares solution, the parameter
is augmented by σ = 1.2σ, until that the condition
kY −M XkF
kY −M X LS kF
< 2 is satisfied. The algorithm divergence occurs if the
stopping criterion (ii) is satisfied, namely the primal residual increases over iterations. In this case, either the parameter is too
large due to an overestimated initialization, or it is too small. Accordingly, we either decrease it by σ = σ0 /p, or increase it
by σ = 1.2σ, until that the ADMM converges.
Algorithm 4 Tuning the bandwidth parameter σ
1: Initialize σ = σ0 using (19); p = 1;
2:
Do CUSAL with Algorithm 2 or Algorithm 3;
3:
if stopping criterion (i) or (iii) is satisfied then
4:
if condition
else
increase σ = 1.2σ, and go to line 2
7:
8:
9:
10:
< 2 is satisfied, then
σ ∗ = σ (optimal value)
5:
6:
kY −MXk2
kY −M X LS k2
end if
else
if σ > 1000σ0 (due to the overestimated σ0 ) then
11:
p = p + 1;
12:
decrease σ = σ0 /p, and go to line 2
13:
14:
15:
16:
else
increase σ = 1.2σ, and go to line 2
end if
end if
13
V. E XPERIMENTS WITH S YNTHETIC DATA
In this section, the performance of the proposed fully-constrained (CUSAL-FC) and sparsity-promoting (CUSAL-SP)
algorithms is evaluated on synthetic data. A comparative study is performed considering six state-of-the-art methods proposed
for linear and nonlinear unmixing models.
•
Fully-Constrained Least-Squares (FCLS) [9]: The FCLS is developed for the linear model. Enforcing both ANC and ASC
constraints, this technique yields the optimal abundance matrix in the least-squares sense.
•
Sparse Unmixing by variable Splitting and Augmented Lagrangian (SUnSAL) [12]: This method is based on the ADMM.
Several variants are developed by including different constraints, with the fully-constrained SUnSAL-FCLS and the
sparsity-promoting SUnSAL-sparse.
•
The Bayesian algorithm for Generalized Bilinear Model (BayGBM) [31], [32]: This method estimates the abundances
with the generalized bilinear model (GBM), which adds second-order interactions between endmembers to the linear
model, yielding the model
y t = M xt +
R−1
X
R
X
i=1 j=i+1
γij,t xit xjt (mi ⊙ mj ) + nt ,
where 0 ≤ γij,t ≤ 1 controls the interactions between endmembers mi and mj , and ⊙ is the element-wise product. The
BayGBM considers both ANC and ASC.
•
The Bayesian algorithm for Polynomial Post-Nonlinear Mixing Model (BayPPNMM) [33]: This algorithm estimates the
parameters by assuming that the pixel reflectances are nonlinear functions of endmembers using
y t = M xt + bt (M xt ) ⊙ (M xt ) + nt ,
(20)
where the nonlinear terms are characterized by bt ∈ R, and both ANC and ASC are required.
•
Kernel Fully-Constrained Least-Squares (KFCLS) [15]: This method generalizes FCLS, by replacing the inner product
with a kernel function. In the following, the Gaussian kernel is applied for simulation.
•
Robust nonnegative matrix factorization (rNMF) [34]: To capture the nonlinear effect (outliers), this NMF-based method
introduces a group-sparse regularization term into the linear model. Accounting for both constraints, the problem is
optimized by a block-coordinate descent strategy. For fair comparison in this paper, the endmembers are fixed with the
real ones. The regularization parameter is set with the degree of sparsity as suggested in [33].
We first compare the fully-constrained CUSAL-FC, presented in IV-A, with the state-of-the-art methods. A series of
experiments are performed, mainly considering the influences of four aspects: (i) mixture model, (ii) noise level, (iii) number
of corrupted bands and (iv) number of endmembers.
14
0.9
1
0.8
0.9
0.8
0.7
0.7
Reflectance
Reflectance
0.6
0.5
0.4
0.3
0.4
0.2
0.1
0.1
0
50
100
150
200
Bands
Fig. 2.
0.5
0.3
0.2
0
0.6
0
0
50
100
150
200
Bands
The R = 3 (left) and 6 (right) USGS signatures chosen for simulation.
Each image, of 50 × 50 pixels, is generated using either the linear mixing model (1) or the polynomial post-nonlinear mixing
model (PPNMM) (20), where nt is a Gaussian noise of SNR ∈ {15, 35}dB. The R ∈ {3, 6} endmembers, as shown in Fig. 2,
are drawn from the USGS digital spectral library [35]. These endmembers are defined over L = 244 continuous bands with the
wavelength ranging from 0.2µm to 3.0µm. The abundance vectors xt are uniformly generated using a Dirichlet distribution
as in [35], [36]. For PPNMM, the values of bt are generated uniformly in the set (−3, 3) according to [33]. To imitate the
noisy bands in the real hyperspectral images, several bands in the generated data are corrupted by replacing the corresponding
rows of Y with random values within [0, 1]. The number of corrupted bands varies in the set {0, 20, 40, 60}.
The unmixing performance is evaluated using the abundance root mean square error (RMSE) [31], [37], defined by
v
u
T
u 1 X
b t k2 ,
kxt − x
RMSE = t
RT t=1
b t is the estimated abundance vector. Fig. 3 and 4 illustrates the average of RMSE over 10 Monte-Carlo realizations,
where x
respectively on the LMM and PPNMM data. It is easy to see that, in presence of outlier bands, the proposed CUSAL-FC
algorithm outperforms all the comparing methods in terms of RMSE, for different mixture models, noise levels and numbers
of endmembers. It is also shown that the performance of the proposed algorithm improves when increasing the SNR.
The performance of the proposed the sparsity-promoting CUSAL-SP, presented in IV-B, is compared with the sparsitypromoting SUnSAL-sparse, as well as the FCLS, on a series of data with sparse abundance matrices. The influences of (i) the
number of corrupted bands and (ii) the sparsity level of the abundances, are studied. Each image, of 15 × 15 pixels, is generated
by the linear mixture model. The endmember matrix is composed by R = 62 USGS signatures, where the angle between
any two different endmembers is larger than 10◦ [25]. The K nonzero entries in each abundance vector xt are generated
by a Dirichlet distribution. The value of K (i.e., the indicator of sparsity level) ranges from 4 to 20, while the number of
15
0.18
0.14
0.12
0.1
0.09
RMSE
0.16
RMSE
0.11
CUSAL-FC
SUnSAL-FCLS
FCLS
BayGBM
BayPPNMM
KFCLS
rNMF
0.1
0.08
0.07
0.08
0.06
0.06
0.04
0
20
40
0.05
60
0
20
Number of corrupted bands
40
60
Number of corrupted bands
(a) SNR = 15, R = 3
(b) SNR = 15, R = 6
0.12
0.16
0.1
0.14
0.12
RMSE
RMSE
0.08
0.1
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0
Fig. 3.
0
20
40
0
60
0
20
40
Number of corrupted bands
Number of corrupted bands
(c) SNR = 35, R = 3
(d) SNR = 35, R = 6
60
LMM data: The root mean square error (RMSE) with respect to the number of corrupted bands, averaged over ten Monte-Carlo realizations, for
different number of endmembers and SNR.
corrupted bands varies in {0, 20, 40, 60}. We set the Gaussian noise by SNR = 30dB, a level that that is commonly present in
real hyperspectral images according to [25]. For both sparsity-promoting algorithms, the regularization parameter λ is adjusted
using the set {10−5 , 5 · 10−5 , 10−4 , 5 · 10−4 , 10−3 , 10−2 , 10−1 }.
The unmixing performance with the sparsity-promoting algorithms is evaluated using the signal-to-reconstruction error,
measured in decibels, according to [12], [25]. It is defined by
SRE = 10 log10
PT
PT
t=1
t=1
kxt k22
b t k22
kxt − x
!
.
The results, averaged over ten Monte-Carlo realizations, are illustrated in Fig. 5. Considering that the abundance matrix under
estimation is sparse at different levels, we conclude the following: Concerning the case without outlier bands, the CUSAL-SP
outperforms the SUnSAL-SAL for K > 8 and FCLS for K > 12. When the number of outlier bands is increases, the proposed
CUSAL-SP algorithm generally provides the best unmixing quality with the highest SRE value, especially for K > 6.
16
0.18
0.14
0.12
0.1
0.09
RMSE
0.16
RMSE
0.11
CUSAL-FC
SUnSAL-FCLS
FCLS
BayGBM
BayPPNMM
KFCLS
rNMF
0.1
0.08
0.07
0.08
0.06
0.06
0.04
0
20
40
0.05
60
0
Number of corrupted bands
20
40
60
Number of corrupted bands
(b) SNR = 15, R = 6
(a) SNR = 15, R = 3
0.2
0.12
0.18
0.1
0.16
0.08
0.12
RMSE
RMSE
0.14
0.1
0.06
0.08
0.04
0.06
0.04
0.02
0.02
0
Fig. 4.
0
20
40
0
60
0
20
40
Number of corrupted bands
Number of corrupted bands
(c) SNR = 35, R = 3
(d) SNR = 35, R = 6
60
PPNMM data: The root mean square error (RMSE) with respect to the number of corrupted bands, averaged over ten Monte-Carlo realizations,
for different number of endmembers and SNR.
VI. E XPERIMENTS
WITH
R EAL DATA
This section presents the performance of the proposed algorithms on a real hyperspectral image. We consider a 250 × 190
sub-image taken from the Cuprite mining image, acquired by the AVIRIS sensor when flying over Las Vegas, Nevada, USA.
The image has been widely investigated in the literature [7], [25]. The raw data contains L = 224 bands, covering a wavelength
range 0.4 − 2.5µm. Among, there are 37 relatively noisy ones with low SNR, namely the bands 1 − 3, 105 − 115, 150 − 170,
and 223 − 224. The geographic composition of this area is estimated to include up to 14 minerals [3]. Neglecting the similar
signatures, we consider 12 endmembers as often investigated in the literature [7], [38]. The VCA technique is first applied to
extract these endmembers on the clean image with L = 187 bands. Starting from L = 187 bands, the noisy bands, randomly
chosen from the bands 1 − 3, 105 − 115, 150 − 170, and 223 − 224, are gradually included to form a series of input data.
Therefore, the experiments are conducted with L = 187, 193, 199, 205, 211, 217, 223 and 224 bands.
17
3.5
CUSAL-SP
9
SUnSAL-sparse
3
FCLS
7
SRE(dB)
SRE(dB)
8
6
2.5
2
5
1.5
4
1
3
4
6
8
10
12
14
16
18
0.5
20
4
6
8
10
K
12
14
16
18
20
16
18
20
K
(a) 0 corrupted band
(b) 20 corrupted bands
2
1.6
1.5
1.3
1
SRE(dB)
SRE(dB)
1.9
1
0.7
0.5
0
0.4
−0.5
0.1
−0.2
4
6
8
10
12
14
16
18
20
−1
4
6
K
10
12
14
K
(c) 40 corrupted bands
Fig. 5.
8
(d) 60 corrupted bands
LMM data: The averaged signal-to-reconstruction error (SRE) with respect to the sparsity level K, averaged over ten Monte-Carlo realizations.
Comparison for various number of corrupted bands at SNR = 30.
Since ground-truth abundances are unknown, the performance is measured with the averaged spectral angle distance (SAD)
b t , as illustrated in Fig. 6, where the SAD is defined by
between the input spectra y t and the reconstructed ones y
T
1 X
SAD =
arccos
T t=1
!
bt
y⊤
t y
.
ky t kkb
yt k
The estimated abundance maps using 187, 205 and 224 bands are given in Fig. 7, Fig. 8, and Fig. 9, respectively. In absence of
noisy bands (i.e., L = 187 bands), all the considered methods lead to satisfactory abundance maps, with BayPPNMM providing
the smallest SAD. As the number of noisy bands increases, especially from L = 199 to L = 224, the unmixing performance
of the state-of-the-art methods deteriorates drastically, while the proposed CUSAL yields stable SAD. The obtained results
confirm the good behavior of the proposed CUSAL algorithms and their robustness in presence of corrupted spectral bands.
18
7
6.5
6
5.5
SAD
5
CUSAL-FC
SUnSAL−FCLS
FCLS
BayGBM
BayPPNMM
KFCLS
rNMF
4.5
4
3.5
3
2.5
2
187
Fig. 6.
193
199
205
211
Number of bands
217
224
Cuprite image: The averaged spectral angle distance (SAD) using different number of bands, computed without the noisy bands 1 − 3, 105 −
115, 150 − 170, and 223 − 224.
VII. C ONCLUSION
This paper presented a supervised unmixing algorithm based on the correntropy maximization principle. Two correntropybased unmixing problems were addressed, the first with the non-negativity and sum-to-one constraints, and the second with
the non-negativity constraint and a sparsity-promoting term. The alternating direction method of multipliers (ADMM) was
investigated in order to solve the correntropy-based unmixing problems. The effectiveness and robustness of the proposed
unmixing method were validated on synthetic and real hyperspectral images. Future works include the generalization of
the correntropy criterion to account for the multiple reflection phenomenon [31], [39], as well as incorporating nonlinear
models [40].
ACKNOWLEDGMENT
This work was supported by the French ANR, grant HYPANEMA: ANR-12BS03-0033.
19
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Fig. 7.
Cuprite image: Estimated abundance maps using 187 clean bands. Left to right: sphene, alunite, buddingtonite, kaolinite, chalcedony, highway.
Top to bottom: SUnSAL-FCLS, FCLS, BayGBM, BayPPNMM, KFCLS, rNMF, CUSAL-FC.
20
1
0.8
0.6
0.4
0.2
10
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Fig. 8.
Cuprite image: Estimated abundance maps using 205 bands, with 187 clean bands. Same legend as Fig. 7.
21
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Fig. 9.
Cuprite image: Estimated abundance maps using all the 224 bands, with 187 clean bands. Same legend as Fig. 7.
22
R EFERENCES
[1] N. Keshava and J. F. Mustard, “Spectral unmixing,” IEEE Signal Processing Magazine, vol. 19, no. 1, pp. 44–57, Jan 2002.
[2] P. Honeine and C. Richard, “Geometric unmixing of large hyperspectral images: a barycentric coordinate approach,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 50, no. 6, pp. 2185–2195, Jun. 2012.
[3] J. Nascimento and J. M. Bioucas-Dias, “Vertex component analysis: a fast algorithm to unmix hyperspectral data,” IEEE Transactions on Geoscience
and Remote Sensing, vol. 43, no. 4, pp. 898–910, Apr. 2005.
[4] M. Winter, “N-FINDR: an algorithm for fast autonomous spectral end-member determination in hyperspectral data: an algorithm for fast autonomous
spectral end-member determination in hyperspectral data,” Proc. of SPIE: Imaging Spectrometry V, vol. 3753, no. 10, 1999.
[5] J. Li, A. Agathos, D. Zaharie, J. Bioucas-Dias, A. Plaza, and X. Li, “Minimum volume simplex analysis: A fast algorithm for linear hyperspectral
unmixing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 9, pp. 5067–5082, Sept 2015.
[6] L. Miao and H. Qi, “Endmember extraction from highly mixed data using minimum volume constrained nonnegative matrix factorization,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 45, no. 3, pp. 765–777, March 2007.
[7] J. Chen, C. Richard, and P. Honeine, “Nonlinear unmixing of hyperspectral data based on a linear-mixture/nonlinear-fluctuation model,” IEEE Transactions
on Signal Processing, vol. 61, no. 2, pp. 480–492, Jan. 2013.
[8] ——, “Nonlinear estimation of material abundances of hyperspectral images with ℓ1 -norm spatial regularization,” IEEE Transactions on Geoscience and
Remote Sensing, vol. 52, no. 5, pp. 2654–2665, May 2014.
[9] D. Heinz and C. Chang, “Fully constrained least squares linear spectral mixture analysis method for material quantification in hyperspectral imagery,”
IEEE Transactions on Geoscience and Remote Sensing, vol. 39, no. 3, pp. 529–545, Mar. 2001.
[10] A. Huck, M. Guillaume, and J. Blanc-Talon, “Minimum dispersion constrained nonnegative matrix factorization to unmix hyperspectral data,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 48, no. 6, pp. 2590–2602, Jun. 2010.
[11] J. M. Bioucas-Dias, A. Plaza, N. Dobigeon, M. Parente, Q. Du, P. Gader, and J. Chanussot, “Hyperspectral unmixing overview: Geometrical, statistical,
and sparse regression-based approaches,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 5, no. 2, pp. 354–379,
2012.
[12] J. M. Bioucas-Dias and M. A. Figueiredo, “Alternating direction algorithms for constrained sparse regression: Application to hyperspectral unmixing,”
in IEEE Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2010, pp. 1–4.
[13] S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein, “Distributed optimization and statistical learning via the alternating direction method of
multipliers,” Foundations and Trends R in Machine Learning, vol. 3, no. 1, pp. 1–122, 2011.
[14] A. Zelinski and V. Goyal, “Denoising hyperspectral imagery and recovering junk bands using wavelets and sparse approximation,” in IEEE International
Conference on Geoscience and Remote Sensing Symposium (IGARSS), July 2006, pp. 387–390.
[15] J. Broadwater, R. Chellappa, A. Banerjee, and P. Burlina, “Kernel fully constrained least squares abundance estimates,” in IEEE International Geoscience
and Remote Sensing Symposium (IGARSS), 2007, pp. 4041–4044.
[16] J. Chen, C. Richard, and P. Honeine, “Estimating abundance fractions of materials in hyperspectral images by fitting a post-nonlinear mixing model,”
in Proc. IEEE Workshop on Hyperspectral Image and Signal Processing : Evolution in Remote Sensing, Jun. 2013.
[17] W. Liu, P. Pokharel, and J. C. Prı́ncipe, “Correntropy: properties and applications in non-gaussian signal processing,” IEEE Transactions on Signal
Processing, vol. 55, no. 11, pp. 5286–5298, 2007.
[18] J. C. Principe, Information theoretic learning: Renyi’s entropy and kernel perspectives.
Springer Science & Business Media, 2010.
[19] J. J. Wang, X. Wang, and X. Gao, “Non-negative matrix factorization by maximizing correntropy for cancer clustering,” BMC bioinformatics, vol. 14,
no. 1, p. 107, 2013.
23
[20] R. He, W. Zheng, and B. Hu, “Maximum correntropy criterion for robust face recognition,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 33, no. 8, pp. 1561–1576, 2011.
[21] Y. Wang, C. Pan, S. Xiang, and F. Zhu, “Robust hyperspectral unmixing with correntropy-based metric,” IEEE Transactions on Image Processing,
vol. 24, no. 11, pp. 4027–4040, Nov 2015.
[22] M. Nikolova and M. Ng, “Analysis of half-quadratic minimization methods for signal and image recovery,” SIAM Journal on Scientific computing,
vol. 27, no. 3, pp. 937–966, 2005.
[23] C. L. Lawson and R. J. Hanson, Solving Least Squares Problems (Classics in Applied Mathematics).
Society for Industrial Mathematics, 1987.
[24] H. Lantéri, M. Roche, O. Cuevas, and C. Aime, “A general method to devise maximum-likelihood signal restoration multiplicative algorithms with
non-negativity constraints,” Signal Processing, vol. 81, pp. 945–974, May 2001.
[25] M. D. Iordache, J. M. Bioucas-Dias, and A. Plaza, “Sparse unmixing of hyperspectral data,” IEEE Transactions on Geoscience and Remote Sensing,
vol. 49, no. 6, pp. 2014–2039, June 2011.
[26] ——, “Total variation spatial regularization for sparse hyperspectral unmixing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 50, no. 11,
pp. 4484–4502, Nov. 2012.
[27] V. Vapnik, The Nature of Statistical Learning Theory.
New York, NY, USA: Springer-Verlag, 1995.
[28] B. Chen and J. C. Prı́ncipe, “Maximum correntropy estimation is a smoothed map estimation,” IEEE Signal Processing Letters, vol. 19, no. 8, pp.
491–494, 2012.
[29] Z. Wu, S. Peng, B. Chen, and H. Zhao, “Robust hammerstein adaptive filtering under maximum correntropy criterion,” Entropy, vol. 17, no. 10, p. 7149,
2015.
[30] R. Tibshirani, “Regression shrinkage and selection via the lasso,” Journal of the Royal Statistical Society. Series B (Methodological), pp. 267–288, 1996.
[31] A. Halimi, Y. Altmann, N. Dobigeon, and J. Y. Tourneret, “Nonlinear unmixing of hyperspectral images using a generalized bilinear model,” IEEE
Transactions on Geoscience and Remote Sensing, vol. 49, no. 11, pp. 4153–4162, Nov. 2011.
[32] ——, “Unmixing hyperspectral images using the generalized bilinear model.” in IEEE International Conference on Geoscience and Remote Sensing
Symposium (IGARSS), 2011, pp. 1886–1889.
[33] Y. Altmann, A. Halimi, N. Dobigeon, and J. Y. Tourneret, “Supervised nonlinear spectral unmixing using a postnonlinear mixing model for hyperspectral
imagery,” IEEE Transactions on Image Processing, vol. 21, no. 6, pp. 3017–3025, 2012.
[34] C. Fevotte and N. Dobigeon, “Nonlinear hyperspectral unmixing with robust nonnegative matrix factorization,” IEEE Transactions on Image Processing,
vol. 24, no. 12, pp. 4810–4819, Dec 2015.
[35] J. M. Bioucas-Dias and J. M. P. Nascimento, “Hyperspectral subspace identification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 46,
no. 8, pp. 2435–2445, Aug 2008.
[36] A. Halimi, N. Dobigeon, and J.-Y. Tourneret, “Unsupervised unmixing of hyperspectral images accounting for endmember variability,” IEEE Transactions
on Image Processing, vol. 24, no. 12, pp. 4904–4917, december 2015.
[37] N. Yokoya, J. Chanussot, and A. Iwasaki, “Nonlinear unmixing of hyperspectral data using semi-nonnegative matrix factorization,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 52, no. 2, pp. 1430–1437, Feb. 2014.
[38] X. Lu, H. Wu, Y. Yuan, P. Yan, and X. Li, “Manifold regularized sparse nmf for hyperspectral unmixing,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 51, no. 5, pp. 2815–2826, 2013.
[39] W. Fan, B. Hu, J. Miller, and M. Li, “Comparative study between a new nonlinear model and common linear model for analysing laboratory simulatedforest hyperspectral data,” International Journal of Remote Sensing, vol. 30, no. 11, pp. 2951–2962, 2009.
[40] A. Halimi, P. Honeine, and J. Bioucas-Dias, “Hyperspectral unmixing in presence of endmember variability, nonlinearity or mismodelling effects,”
http://arxiv.org/abs/1511.05698, pp. 1–32, Nov. 2015.
| 9cs.NE
|
arXiv:1502.04893v3 [math.CO] 10 Sep 2015
DETECTING BINOMIALITY
CARSTEN CONRADI AND THOMAS KAHLE
Abstract. Binomial ideals are special polynomial ideals with many algorithmically
and theoretically nice properties. We discuss the problem of deciding if a given polynomial ideal is binomial. While the methods are general, our main motivation and
source of examples is the simplification of steady state equations of chemical reaction
networks. For homogeneous ideals we give an efficient, Gröbner-free algorithm for
binomiality detection, based on linear algebra only. On inhomogeneous input the algorithm can only give a sufficient condition for binomiality. As a remedy we construct a
heuristic toolbox that can lead to simplifications even if the given ideal is not binomial.
Contents
1. Introduction
2. Gröbner-free criteria for binomiality
3. The homogeneous case
4. Heuristics for the inhomogeneous case
References
1
4
7
10
15
1. Introduction
Non-linear algebra is a mainstay in modern applied mathematics and across the
sciences. Very often non-linearity comes in the form of polynomial equations which
are much more flexible than linear equations in modeling complex phenomena. The
price to be paid is that their mathematical theory—commutative algebra and algebraic
geometry—is much more involved than linear algebra. Fortunately, polynomial systems
in applications often have special structures. In this paper we focus on sparsity, that is,
polynomials having few terms.
The sparsest polynomials are monomials. Systems of monomial equations are a big
topic in algebraic combinatorics, but in the view of modeling they are not much help.
Date: August 2015.
2010 Mathematics Subject Classification. Primary: 13P15, 37N25; Secondary: 92C42, 13P25, 13P10.
Key words and phrases. polynomial systems in biology, binomial ideal, steady states, chemical reaction networks.
The authors are supported by the research focus dynamical systems of the state Saxony-Anhalt.
1
2
CARSTEN CONRADI AND THOMAS KAHLE
Their solution sets are unions of coordinate hyperplanes. The next and more interesting
class are binomial systems in which each polynomial is allowed to have two terms.
Binomials are flexible enough to model many interesting phenomena, but sparse enough
to allow a specialized theory [8]. The strongest classical results about binomial systems
require one to seek solutions in algebraically closed field such as the complex numbers.
However, for the objects in applications (think of concentrations or probabilities) this
assumption is prohibitive. One often works with non-negative real numbers and this
leads to the fields of real and semi-algebraic geometry. New theory in combinatorial
commutative algebra shows that for binomial equations field assumptions can be skirted
and that the dependence of binomial systems on their coefficients is quite weak [14]. For
binomial equations one can hope for results that do not depend on the explicit values
of the parameters and are thus robust in the presence of uncertainty.
The main theme of this paper is how to detect binomiality, that is, how to decide if a
given polynomial system is equivalent to a binomial system. The common way to decide
binomiality is to compute a Gröbner basis since an ideal can be generated by binomials
if and only if any reduced Gröbner basis is binomial [8, Corollary 1.2]. For polynomial
systems arising in applications, however, computing a reduced Gröbner basis is often too
demanding: as parameter values are unknown, computations have to be performed over
the field of rational functions in the parameters. Even though this is computationally
feasible, it is time consuming and usually yields an output that is hard to digest for
humans. This added complexity comes from the fact that Gröbner bases contain a
lot more information than what may be needed for a specific task such as deciding
binomiality of a polynomial system. Hence Gröbner-free methods are desirable.
Gröbner-free methods. Gröbner bases started as a generalization of Gauss elimination to polynomials. They have since come back to their roots in linear algebra by
the advent of F4 and F5 type algorithms which try to arrange computations so that
sparse linear algebra can exploited [7]. Our method draws on linear algebra in bases of
monomials too, and is inspired by these developments in computer algebra.
Deciding if a set of polynomials can be brought into binomial form using linear algebra is the question whether the coefficient matrix has a partitioning kernel basis (Definition 2.1 and Proposition 2.5). Deciding this property requires only row reductions and
hence is computationally cheap compared to Gröbner bases. It was shown in [18] that,
if the coefficient matrix of a suitably extended polynomial system admits a partitioning
kernel basis, then the polynomial system is generated by binomials. As a first insight
we show that the converse of this need not hold (Example 2.8).
In general computer algebra profits from homogeneity. This is true for Gröbner bases
where, for example, Hilbert function driven algorithms can be used to convert a basis
for a term-order that is cheap to compute into one for an expensive order, such as lex.
We also observe this phenomenon in our Gröbner-free approach: a satisfying answer
DETECTING BINOMIALITY
3
to the binomial detection problem can be found if the given system of polynomials is
homogeneous. In Section 3 we discuss this case which eventually leads to Algorithm 3.3.
In the inhomogeneous case things are more complicated. Gröbner basis computations
can be reduced to the homogeneous case by an easy trick. Detection of binomiality can
not (Example 4.1). We address this problem by collecting heuristic approaches that, in
the best case, establish binomiality without Gröbner bases (Recipe 4.5). Our approaches
can also be used if the system is not entirely binomial, but has some binomials. In
Example 4.4 we demonstrate this on a polynomial system from [5].
Binomial steady state ideals. While binomiality detection can be applied to any
polynomial system, our motivation comes from chemical reaction network theory where
ordinary differential equations with polynomial right-hand sides are used to model dynamic processes in systems biology. The mathematics of these systems is extremely
challenging, in particular since realistic models are huge and involve uncertain parameters. As a consequence of the latter, studying dynamical systems arising in biological
applications often amounts to studying parameterized families of polynomial ODEs.
The first order of business (and concern of a large part of the work in the area) is to
determine steady states which are thus the non-negative real zeros of families of parameterized polynomial equations. Moreover, the structure of the polynomial ODEs
entails the existence of affine linear subspaces that are invariant for solutions. Hence
questions concerning existence and uniqueness of steady states or existence of multiple
steady states are equivalent to questions regarding the intersection of the zero set of a
parameterized family of polynomials with a family of affine linear subspaces.
If the polynomial equations describing steady states are equivalent to binomial equations (that is, generate a binomial ideal), then their mathematical analysis becomes
much easier. This is the main theme of [18]. If a system is binomial, then, for instance, one can decide efficiently if positive steady states exist. If so, then a monomial
parametrization can be found using only linear algebra over the integers [8, Section 2],
and the steady states are toric: they are the positive real part of a toric variety. A
sufficient criterion for toric steady states appears in [13, Theorem 4.1]. Since zero sets
of general polynomial systems need not have parametrizations at all, we view the task
of detecting binomiality as an important step in analyzing polynomials—in systems
biology, or other areas like algebraic statistics, control theory, economics, etc.
A frequent and challenging problem in the analysis of dynamical systems in biology
is to decide multistationarity, that is, the existence of parameter values leading to more
than one isolated steady state. A variety of results for precluding multistationarity has
appeared in recent years. See, for instance, [12, 21, 3, 1] for methods employing the
Jacobian. Similarly, several sufficient criteria for multistationarity have emerged (for
example [4, 5]). In general this problem remains very hard. However, in the case of binomial steady state equations, the question of multistationarity can often be answered
4
CARSTEN CONRADI AND THOMAS KAHLE
effectively, for example positively by [18, Theorem 5.5], or negatively by [17, Theorem 1.4]. Both of these results require only the study of systems of linear inequalities.
Notation. In this paper we work with the polynomial ring k[x1 , . . . , xn ] in n variables.
The coefficient field k is usually R, or the field of real rational functions in a set of
parameters. Our methods are agnostic towards the field. A system of polynomial
equations f1 = f2 = · · · = fs = 0 in the variables x1 , . . . , xn is encoded in the ideal
hf1 , . . . , fs i ⊂ k[x1 , . . . , xn ]. A polynomial is homogeneous if all its terms have the
same total degree, and an ideal is homogeneous if it can be generated by homogeneous
polynomials. A binomial is a polynomial with at most two terms. In particular, a
monomial is a binomial. It is important to distinguish between binomial ideals and
binomial systems. A binomial system f1 = · · · = fs = 0 is a polynomial system such
that each fi is a binomial. In contrast, an ideal hf1 , . . . , fs i is a binomial ideal if there
exist binomials that generate the same ideal. Thus general non-binomials do not form
a binomial system, even if they generate a binomial ideal. For the sake of brevity we
will not give an introduction to commutative algebra here, but refer to standard text
books like [6, 2]. The very modest amount of matroid theory necessary in Section 2 can
be picked up from the first pages of [20].
Acknowledgment. We thank David Cox for discussions on the role of homogeneity in
computer algebra. We are grateful to Alicia Dickenstein for pointing out a crucial error
in an earlier version of this paper. TK is supported by CDS the Center for Dynamical
Systems at Otto-von-Guericke University.
2. Gröbner-free criteria for binomiality
The most basic criterion to decide if an ideal is binomial is to compute a Gröbner
basis. This works because the Buchberger algorithm is binomial-friendly: an S-pair
of binomials is a binomial. Since the reduced Gröbner basis is unique and must be
computable from the binomial generators, it consists of binomials if and only if the
ideal is binomial. However, Gröbner bases can be very hard to compute, so other
criteria using only linear algebra are also desirable. Linear algebra enters, when we
write a polynomial system as AΨ(x), the product of a coefficient matrix A with entries
in k, and a vector of monomials Ψ(x). Clearly, if we use row operations on the matrix
to bring it into a form where each row has at most two non-zero entries, then the ideal is
generated by binomials and monomials. This criterion is too naive to detect all binomial
ideals since it allows only k-linear combinations of the given polynomials. We show in
Section 3 that, at least for homogeneous ideals, it can be extended to a characterization.
Before we embark into the details, we formalize the condition on the matrix.
Definition 2.1. A matrix A has a partitioning kernel basis if its kernel admits a basis
of vectors with disjoint supports, that is, if there exists a basis b(1) , . . . , b(d) of ker(A)
such that supp(b(i) ) ∩ supp(b(j) ) = ∅ for all i 6= j.
DETECTING BINOMIALITY
5
The following proposition allows one to check for a partitioning kernel basis with
linear algebra. The underlying reason is the very restricted structure of the kernel,
expressed best in matroid language.
Proposition 2.2. The following are equivalent for any matrix A.
1) A has a partitioning kernel basis.
2) The column matroid of A is a direct sum of uniform matroids Ur−1,r of corank
one, and possibly several coloops U1,1 .
3) The reduced row echelon form of A has at most two non-zero entries in each
row.
Proof. 1 ⇒ 2: Let b1 , . . . , bk be the partitioning kernel basis. The supports of this basis
satisfy the circuit axioms and are thus equal to the circuits of the column matroid of A.
Indeed, non-containment and circuit elimination are satisfied trivially because there is
no overlap between any two circuits. For any non-zero element b̃ ∈ ker(A), we have
P
S
b̃ = i λi bi . By the partitioning kernel basis property supp(b̃) = i {supp(bi ) : λi 6= 0},
so either b̃ is proportional to one of the bi , or its support properly contains the support
of a circuit, so it cannot be a circuit. The columns of A which do not appear in any
circuit are coloops and the remaining columns form a direct sum of k uniform matroids
of corank one.
2 ⇒ 3: If the column matroid of A is a direct sum of matroids, then the unique reduced
row echelon form has block structure corresponding to the direct sum decomposition.
Therefore it suffices to consider a single block which has one-dimensional kernel of full
support (the coloops are (1 × 1)-identity blocks). Ignoring zero rows, the reduced row
r−1
echelon form of such a matrix is (Ir−1 |c) where r − 1 is the rank and c ∈ k6=
0 .
3 ⇒ 1: Rows of the reduced row echelon form with exactly one non-zero entry correspond to positions where every element of the kernel has a zero. Thus we can assume
that there are none and each row of A has exactly two non-zero entries. Let c be a
non-pivotal column with r − 1 non-zero entries. The restriction of A to c and the corresponding pivotal columns yields a block containing (Ir−1 |c) and some zero rows. The
unique kernel vector corresponding to the dependencies in this block is orthogonal to
the kernel of the remaining columns. This procedure can be applied to any non-pivotal
column. The thus constructed basis is a partitioning kernel basis.
Remark 2.3. Proposition 2.2 shows that the complexity of deciding if a matrix has
a partitioning kernel basis is essentially the same as that of Gauss-Jordan elimination.
One needs O(n3 ) field operations where n is the larger of the dimensions of the matrix.
Remark 2.4. A direct sum of (arbitrary) uniform matroids is called a partition matroid.
We now translate Proposition 2.2 to polynomial systems.
Proposition 2.5. If A has a partitioning kernel basis, and Ψ(x) is a vector of monomials of appropriate length, then the ideal hAΨ(x)i ⊂ k[x1 , . . . , xn ] is binomial. If
6
CARSTEN CONRADI AND THOMAS KAHLE
AΨ(x) is any system that can be transformed into a binomial system using only k-linear
combinations, then A has a partitioning kernel basis.
Proof. Up to coloops the first part is [18, Theorem 3.3] and the coloops only give
monomials. The second statement is clear since k-linear combinations of polynomials
are row operations on the coefficient matrix and those do not change the kernel.
Our general strategy is to suitably extend a given system AΨ(x) with redundant
polynomials such that Proposition 2.5 yields binomiality of an extended system A′ Ψ′ (x).
This happens in the following example.
Example 2.6. Let f1 = x − y, f2 = z − w, and f3 = x(f1 + f2 ) = x2 − xy + xz − xw.
Ordering the monomials ΨT = (x, y, z, w, x2, xy, xz, xw), the system linearizes as
1 −1 0 0 0 0 0 0
f = (f1 , f2 , f3 ) = 0 0 1 −1 0 0 0 0 · Ψ.
0 0 0 0 1 −1 1 −1
The coefficient matrix is in reduced row echelon form and does not have a partitioning
kernel basis by Proposition 2.2. Algorithm 3.3 takes this problem into account, working
degree by degree. The ideal is binomial since f3 is a binomial (in fact, zero) in the
quotient ring k[x, y, z]/hx − y, z − wi.
The first hint into how to extend A and Ψ(x) is the following theorem due to Pérez
Millán et al.
Theorem 2.7 ([18, Theorem 3.19]). Let f1 = f2 = · · · = fs = 0 be a polynomial system.
If there exist monomials xα1 , . . . , xαm such that, for some i1 , . . . , im ∈ [s], the system
(2.1)
f1 = f2 = · · · = fs = xα1 fi1 = · · · = xαm fim = 0
has a coefficient matrix with a partitioning kernel basis, then hf1 , . . . , fs i is binomial.
Theorem 2.7 is true since the additional generators in (2.1) do not change the ideal
that the system generates. This together with the explicit description of the binomial
generators in the case of a partitioning kernel basis [18, Theorem 3.3] yields the result.
If the condition in Theorem 2.7 were also necessary, then a test for binomiality could
be built on trying to systematically identify the monomials xα . However, the converse
of Theorem 2.7 is not true.
Example 2.8. Let I = hf1 , f2 i be the homogeneous binomial ideal generated by the
non-binomials f1 = x − y + x2 + y 2 + z 2 , f2 = x2 + y 2 + z 2 . For no choice of monomials
m11 , . . . , m1r , m21 , . . . , m2s ∈ k[x, y, z] does the coefficient matrix of the system
(2.2)
f1 = m11 f1 = · · · = m1r f1 = f2 = m21 f2 = · · · = m2s f2 = 0
have a partitioning kernel basis.
For the proof Example 2.8 we first need the following curious little fact.
DETECTING BINOMIALITY
7
Lemma 2.9. The ideal I = hx2 + y 2 + z 2 i ⊂ k[x, y, z] does not contain a non-zero
binomial.
Proof. We can assume that k is algebraically closed, since if I contains a non-zero
binomial, then so does its extension to the algebraic closure. Assume that for some
f ∈ k[x, y, z], the product b = f (x2 + y 2 + z 2 ) ∈ I is a binomial. We can assume that
b is not divisible by any variable. Indeed, if a variable divides b, then it divides f and
we find a lower degree binomial in I. Since I is homogeneous, we can also assume that
b is homogeneous. Potentially renaming the variables, we can assume b = xd − λy s z d−s
for some 0 ≤ s < d and λ ∈ k. Since k is algebraically closed, there is a solution ξ to
the equation x2 = −1. The generator x2 + y 2 + z 2 vanishes at (ξ, 1, 0) but b does not
vanish there. This contradiction shows that I cannot contain a binomial.
Proof of Example 2.8. Let d be the highest total degree among monomials in the system (2.2) and consider the restriction of all involved polynomials to degree d. Since the
highest degree part of both f1 and f2 equals (x2 + y 2 + z 2 ), only monomial multiples
of (x2 + y 2 + z 2 ) can contribute to the degree d part of the system (2.2). If the whole
coefficient matrix of (2.2) had a partitioning kernel basis, then also the submatrix with
only the columns for degree d monomials had one (Proposition 2.2). In this case row
reductions on the submatrix would yield a binomial in degree d in the ideal generated
by (x2 + y 2 + z 2 ). This is impossible by Lemma 2.9.
Example 2.8 may seem contrived, but this kind of “trivial obfuscation” of binomials
does happen in applications. Of course, for humans it is obvious that one should first
isolate the linear binomial x − y and then search for implied quadratic binomials which
reduce the trinomial. Our next aim is Algorithm 3.3 which implements this idea, at least
in the homogeneous case. The homogeneity assumption cannot be skirted, unfortunately.
It is true that an ideal is binomial if and only if its homogenization is binomial [8,
Corollary 1.4], but the homogenization is not accessible without a Gröbner basis. It
would be superb for our purposes if homogenizing the generators of a binomial ideal
would always yield a binomial ideal. Unfortunately this is not the case as Example 4.1
shows.
3. The homogeneous case
If a given ideal I is homogeneous, the graded vector space structure of the quotient
k[x1 , . . . , xn ]/I allows one to check binomiality degree by degree. For this we need some
basic facts about quotients modulo binomials (see [8, Section 1] for details). Any set of
binomials B in k[x1 , . . . , xn ] induces an equivalence relation on the set of monomials in
k[x1 , . . . , xn ] under which m1 ∼ m2 if and only if m1 −λm2 ∈ hBi for some non-zero λ ∈
k. As a k-vector space the quotient ring k[x1 , . . . , xn ]/hBi is spanned by the equivalence
classes of monomials and those are all linearly independent [8, Proposition 1.11]. If the
binomials are homogeneous, then the situation is particularly nice. For example, the
8
CARSTEN CONRADI AND THOMAS KAHLE
equivalence classes are finite and elements of the quotient have well-defined degrees.
The notions of monomial, binomial, and polynomial are extended to the quotient ring.
For example, a binomial is a polynomial that uses at most two equivalence classes of
monomials. The unified mathematical framework to treat quotients modulo binomials
are monoid algebras, but we refrain from introducing this notion here.
As a consequence of the discussion above, a polynomial system f1 = · · · = fs = 0
can be considered modulo binomials, and the coefficient matrix of the quotient system
is well-defined. It arises from the coefficient matrix of the original system by summing
columns for monomials in the same equivalence class.
Example 3.1. In k[x, y], let b = x2 −y 2. Among monomials of total degree three, x3 and
xy 2 , as well as x2 y and y 3 become equal in k[x, y]/hbi. Thus the degree three part in the
quotient is two-dimensional with one basis vector per equivalence class. Consequently,
the trinomial f = x3 + xy 2 + y 3 maps to a binomial with coefficient matrix [2, 1]. This
matrix arises from the matrix [1, 1, 1, 0] by summing the columns corresponding to x3
and xy 2 , as well as those for x2 y and y 3.
The reduction modulo lower degree binomials in Example 3.1 can be done in general.
Lemma 3.2. Let f1 , . . . , fs ∈ k[x1 , . . . , xn ] be homogeneous polynomials of degree d,
and B ⊂ k[x1 , . . . , xn ] a set of homogeneous binomials of degree at most d. Then in
the quotient ring k[x1 , . . . , xn ]/hBi the ideal hf1 , . . . , fs i/hBi is binomial if and only if
the coefficient matrix of the images of f1 , . . . , fs in k[x1 , . . . , xn ]/hBi has a partitioning
kernel basis.
Proof. The graded version of Nakayama’s lemma (see Corollary 4.8b together with Exercise 4.6 in [9]) implies that the ideal hf1 , . . . , fs i/hBi ⊂ k[x1 . . . , xn ]/hBi has a welldefined number of minimal generators in each degree. Therefore any minimal generating
set consists only of degree d polynomials and Proposition 2.5 applied to the finitedimensional vector space of degree d polynomials correctly decides binomiality.
Lemma 3.2 is the basis for the following binomial detection algorithm.
Algorithm 3.3.
Input: Homogeneous polynomials f1 , . . . , fs ∈ k[x1 , . . . , xn ].
Output: Yes and a binomial generating set of hf1 , . . . , fs i if one exists, No otherwise.
1) Let
• B := ∅,
• R := k[x1 , . . . , xn ],
• F := {f1 , . . . , fs }.
2) While F is not empty,
a) Let Fmin be the set of elements of minimal degree in F .
b) Redefine F := F \ Fmin .
c) Compute the reduced row echelon form A of the coefficient matrix of Fmin .
DETECTING BINOMIALITY
9
d) If A has a row with three or more non-zero entries, output No and stop.
e) Find a set B ′ of binomials in k[x1 , . . . , xn ] whose images in R generate
hFmin i and redefine B := B ∪ B ′ .
f ) Redefine R := k[x1 , . . . , xn ]/hBi.
g) Redefine F as its image in R.
3) Output Yes and B.
Proof of correctness and termination. Termination is obvious. In fact, the maximum
number of iterations in the while loop equals the number of distinct total degrees among
f1 , . . . , fs . Step 2.d relies on Proposition 2.2. In step 2.e, binomials that generate hFmini
in R can be read off the reduced row echelon form via Proposition 2.2. Then any
preimages in k[x1 , . . . , xn ] of those binomials suffice for B ′ . Lemma 3.2 shows that the
while loop either exhausts F if hf1 , . . . , fs i is binomial, or stops when this is not the
case.
Remark 3.4. In the homogeneous case there is a natural choice of finite-dimensional
vector spaces to work in: polynomials of a fixed degree. In each iteration of the while
loop in Algorithm 3.3, the rows of A span the vector space of polynomials of degree d in
the ideal (modulo the binomials in hBi). In the general inhomogeneous situation extra
work is needed to construct a suitable finite-dimensional vector space. In particular, one
needs to select from the infinite list of binomials in the ideal not too many, but enough
to reduce all given polynomials to binomials whenever this is possible. An interesting
problem for the future is to adapt one of the selection strategies from the F4 algorithm
for Gröbner bases [10] for this task.
Remark 3.5. Coefficient matrices of polynomial systems are typically very sparse. An
efficient implementation of Algorithm 3.3 has to take this into account.
Remark 3.6. Algorithm 3.3 could also be written completely in the polynomial ring
without any quotients. Then in each new degree, one would have to consider the
coefficient matrix of Fmin together with all binomials of degree d in the ideal hBi. This
list grows very quickly and so does the list of monomials appearing in these binomials.
Thus it is not only more elegant to work with the quotient, but also more efficient.
To implement Algorithm 3.3 completely without Gröbner bases some refinements are
necessary. Simply using R = k[x1 , . . . , xn ]/B in Macaulay2 will make it compute a
Gröbner basis of B to effectively work with the quotient. For our purposes, however,
this is not necessary.
Proposition 3.7. Algorithm 3.3 can be implemented without Gröbner bases.
Proof. The critical step is 2.g, when the algorithm reduces F modulo the binomials
already found. For the following step 2.c elements of Fmin need to be written in terms
of a basis of the finite-dimensional vector space Rdeg(Fmin ) of degree deg(Fmin ) monomials
modulo the binomials in B. The equivalence relation introduced in the beginning of this
10
CARSTEN CONRADI AND THOMAS KAHLE
section can also be thought of as a graph on monomials, and thus these reductions can
be carried out with graph enumeration algorithms like breadth first search. Restricting
to monomials of degree deg(Fmin), the connected components are a vector space basis
of Rdeg(Fmin ) and can thus be used to gather coefficients in step 2.g.
Remark 3.8. The feasibility of graph-theoretic computations in cases where Gröbner
bases cannot be computed has been demonstrated in [15]. Example 4.9 there contains
a binomial ideal whose Gröbner basis cannot be computed, but whose non-radicality
was proved using a graph-theoretic computation. This yielded a negative answer to the
question of radicality of conditional independence ideals in algebraic statistics.
Remark 3.9. Using Gröbner bases one represents each connected component of the
graph defined by hBi by its least monomial with respect to the term order. Our philosophy is that this is not necessary: one should work with the connected components per
se. Why bother with picking and finding a specific representative in each component
if any representative works? In an implementation one could choose a data structure
that for each monomial stores an index of the connected component it belongs to.
Remark 3.10. It is trivial to generate classes of examples where Gröbner bases methods fail, but Algorithm 3.3 is quick. For example, take any set of binomials whose
Gröbner basis cannot be computed and add any polynomial in the ideal. Algorithm 3.3
immediately goes to work on reducing the polynomial modulo the binomials, while any
implementation of Gröbner bases embarks into its hopeless task.
Remark 3.11. Remark 3.10 highlights the Gröbner-free spirit of our method. The
Gröbner basis of an ideal contains much more information than binomiality. One should
avoid expensive computation to decide this simple question.
4. Heuristics for the inhomogeneous case
The ideals one encounters in chemical reaction network theory are often not homogeneous, so that the results from Section 3 do not apply. The first idea one may have for
the inhomogeneous case is to work with some (partial) homogenization. Gröbner bases
are quite robust in relation to homogenization. For example, to compute a Gröbner
basis of a non-homogeneous ideal it suffices to homogenize the generators, compute a
Gröbner basis of this homogeneous ideal, and then dehomogenize. Although the intermediate homogeneous ideal is generally not equal to the homogenization of the original
ideal, the dehomogenized Gröbner basis is a Gröbner basis of the dehomogenized ideal
[2, Exercise 1.7.8].
Unfortunately the notion of binomiality does not lend itself to that kind of tricks.
Geometrically, homogenizing (all polynomials in) an ideal yields the projective closure
and dehomogenizing restricts to one affine piece. Homogenizing only the generators
creates extra components at infinity and these components need not be binomial. Even
DETECTING BINOMIALITY
11
if they are binomial, the intersection need not be binomial (see also [14, Problem 17.1]).
This is the case in the following example.
Example 4.1. The ideal hab − x, ab − y, x + y + 1i ⊂ k[a, b, x, y] is binomial as it equals
h2y + 1, 2x + 1, 2ab + 1i. Homogenizing the generators, however, yields the non-binomial
ideal hab − xz, ab − yz, x + y + zi.
We now present some alternatives that do not give complete answers but are quick to
check. They can be applied before resorting to an expensive Gröbner basis computation.
The quickest (but least likely to be successful) approach is to try linear algebraic
manipulations of the given polynomials. Equivalently one applies row operations to the
coefficient matrix, for instance, computing the reduced row echelon form. If it has a
partitioning kernel basis, then the ideal is binomial and all non-binomial generators are
k-linear combinations of the binomials. While it may seem very much to ask for this, it
does happen for the family of networks in [18, Section 4].
If just linear algebra is not successful, one can homogenize the generators and run
Algorithm 3.3. If the resulting homogeneous ideal comes out binomial, then the original ideal was binomial by the following simple fact, proven for instance in [2, Corollary A.4.16].
Proposition 4.2. Let I ⊂ k[x1 , . . . , xn ] be an ideal and I ′ ⊂ k[x0 , x1 , . . . , xn ] the homogeneous ideal generated by the homogenizations of the generators of I (using variable
x0 ). Then I is generated by the dehomogenization of any generating set of I ′ .
We now illustrate a phenomenon leading to failure of the above heuristics.
Example 4.3. Consider the network from [18, Example 3.15]. The steady states are
non-negative real zeros of the following polynomials.
f1
f2
f3
f4
f5
= −k12 x1 + k21 x2 − k1112 x1 x7 + (k1211 + k1213 )x9 ,
= k12 x1 − k21 x2 − k23 x2 + k32 x3 + k67 x6 ,
= k23 x2 − k32 x3 − k34 x3 − k89 x3 x7 + k910 x8 + k98 x8 ,
= k34 x3 − k56 x4 x5 + k65 x6 ,
= −k56 x4 x5 + k65 x6 + k910 x8 + k1213 x9 ,
f6 = k56 x4 x5 − (k65 + k67 )x6 ,
f7 = k67 x6 − k1112 x1 x7 − k89 x3 x7 + k98 x8 + k1211 x9 ,
f8 = k89 x3 x7 − (k910 + k98 )x8 ,
f9 = k1112 x1 x7 − (k1211 + k1213 )x9 .
The binomials f6 , f8 , and f9 can be used to eliminate one of every pair (x6 , x4 x5 ),
(x8 , x3 x7 ), and (x9 , x1 x7 ). We eliminate x4 x5 , x8 , and x9 . It can be checked that eliminating x1 x7 instead of x9 does not lead to binomials immediately (although it leads to
12
CARSTEN CONRADI AND THOMAS KAHLE
linear trinomials).
f1′ = −k12 x1 + k21 x2 ,
f2′ = k12 x1 − (k21 + k23 )x2 + k32 x3 + k67 x6 ,
f3′ = k23 x2 − (k32 + k34 )x3 ,
f4′ = k34 x3 − k67 x6 ,
k1112 k1213 x1 x7 k89 k910 x3 x7
f5′ = −k67 x6 +
+
,
k1211 + k1213
k910 + k98
(k1112 k1213 (k910 + k98 )x1 + (k1211 + k1213 )k89 k910 x3 )x7
.
f7′ = k67 x6 −
(k1211 + k1213 )(k910 + k98 )
Using the linear relations f1′ and f3′ the remaining system is recognized to consist of
only two independent binomials:
f2′′ = −k34 x3 + k67 x6
f4′′ = k34 x3 − k67 x6 ,
k1112 k1213 k21
k23 k89 k910
′′
f5 = −k67 x6 +
x2 x7 ,
+
k12 (k1211 + k1213 ) (k32 + k34 )(k910 + k98 )
k23 k89 k910
k1112 k1213 k21
′′
x2 x7 .
−
f7 = k67 x6 + −
k12 (k1211 + k1213 ) (k32 + k34 )(k910 + k98 )
This analysis shows that the steady state ideal under consideration equals the binomial
ideal hf1′ , f2′′ , f3′ , f5′′ , f6 , f8 , f9 i. The Gröbner basis computation in [18, Example 3.15]
also yields the result, but it is arguably less instructive. Note also that naive homogenization does not yield binomiality. The element f2 is linear. After homogenization,
Algorithm 3.3 would pick only this element as Fmin and stop since it is not a binomial.
The effect in Example 4.3 motivates our final method: term replacements using known
binomials. We expect this to be very useful in applications from system biology for the
following reasons.
• It often happens that non-binomial generators are linear combinations of binomials as in Example 4.3 where f1 = f1′ + f9 .
• Steady state ideals of networks with enzyme-substrate complexes always have
some binomial generators. These complexes are produced by only one reaction
and thus their rate of change is binomial.
• In MAPK networks, which describe certain types of cellular signaling, one often
finds binomials of the form kxa xb − k ′ xc .
• Frequently binomials in steady state ideals are linear. Equivalently some of
the concentrations are equal up to a scaling (which may depend on kinetic
parameters). This happens for all examples in [19].
DETECTING BINOMIALITY
13
We now illustrate term replacements in a larger example which comes from the network for ERK activation embedded in two negative feedback loops (see [5, Section 5]
for pointers to the relevant biology).
Example 4.4. Consider the following steady state ideal generated by 29 polynomials.
f1 = −k1 x1 x2 + k2 x3 + k6 x6 ,
f2 = −k1 x1 x2 + k2 x3 + k3 x3 ,
f3 = k1 x1 x2 − k2 x3 − k3 x3 ,
f4 = k11 x10 + k12 x10 + k38 x25 + k42 x27 + k3 x3 − k37 x18 x4 ,
− k4 x4 x5 + k5 x6 − k7 x4 x7 + k8 x8 + k9 x8 − k10 x4 x9
f5 = k14 x12 + k15 x12 + k17 x13 + k18 x13 + k35 x24 + k36 x24 + k41 x27 + k42 x27 − k13 x11 x5 − k34 x16 x5
− k40 x26 x5 − k4 x4 x5 + k5 x6 + k6 x6 − k16 x5 x9 ,
f6 = k4 x4 x5 − k5 x6 − k6 x6 ,
f7 = k18 x13 − k7 x4 x7 + k8 x8 ,
f9 = k11 x10 + k15 x12 + k17 x13 + k9 x8 − k10 x4 x9 − k16 x5 x9 ,
f8 = k7 x4 x7 − k8 x8 − k9 x8 ,
f10 = −k11 x10 − k12 x10 + k10 x4 x9 ,
f11 = k12 x10 + k14 x12 − k19 x11 x14 + k20 x15 + k21 x15 − k22 x11 x16 + k23 x17 + k24 x17 − k13 x11 x5 ,
f12 = −k14 x12 − k15 x12 + k13 x11 x5 ,
f13 = −k17 x13 − k18 x13 + k16 x5 x9 ,
f14 = −k19 x11 x14 + k20 x15 + k30 x21 + k36 x24 ,
f15 = k19 x11 x14 − k20 x15 − k21 x15 ,
f16 = k21 x15 − k22 x11 x16 + k23 x17 − k28 x16 x19 + k27 x20 + k29 x21 + k33 x23 + k35 x24 − k34 x16 x5 ,
f17 = k22 x11 x16 − k23 x17 − k24 x17 ,
f18 = k24 x17 − k25 x18 x19 + k26 x20 − k31 x18 x22 + k32 x23 + k38 x25 + k39 x25
− k43 x18 x28 + k44 x29 + k45 x29 − k37 x18 x4 ,
f19 = −k46 x19 − k28 x16 x19 − k25 x18 x19 + k26 x20 + k27 x20 + k29 x21 + k30 x21 + k45 x29 ,
f20 = k25 x18 x19 − k26 x20 − k27 x20 ,
f22 = −k31 x18 x22 + k32 x23 + k33 x23 ,
f21 = k28 x16 x19 − k29 x21 − k30 x21 ,
f23 = k31 x18 x22 − k32 x23 − k33 x23 ,
f24 = −k35 x24 − k36 x24 + k34 x16 x5 ,
f26 = k39 x25 + k41 x27 − k40 x26 x5 ,
f25 = −k38 x25 − k39 x25 + k37 x18 x4 ,
f27 = −k41 x27 − k42 x27 + k40 x26 x5
f28 = k46 x19 − k43 x18 x28 + k44 x29 ,
f29 = k43 x18 x28 − k44 x29 − k45 x29 .
After some obvious factorization, the following elements are binomials: f2 , f6 , f8 , f10 ,
f12 , f13 , f15 , f17 , f20 , f21 , f22 , f23 , f24 , f25 , f27 , f29 . The system has seven conservation
relations, which can be found by linear algebra. According to our strategy to use
binomials to simplify the system, we eliminate, if possible, non-binomials using the
conservation relations. This is not always possible, as some of the conservation relations
stem from duplicate equations like f2 = −f3 . We eliminate f3 , f4 , f5 , f8 , f9 , f18 , and f19 .
The remaining non-binomial part consists of f1 , f7 , f11 , f14 , f16 , f26 , and f28 . Dividing
by reaction constants, each of the binomials is of the form xi = Kxj xl for some rational
expression K involving only reaction constants. Using these in the non-binomials yields
k1 k3 x1 x2
k4 k6 x4 x5
k7 k9 x4 x7
k16 k18 x5 x9
+
,
f7′ = −
+
,
k2 + k3
k5 + k6
k8 + k9
k17 + k18
k10 k12 x4 x9
k13 k15 x11 x5
+
,
=−
k14 + k15
k11 + k12
f1′ = −
′
f11
14
CARSTEN CONRADI AND THOMAS KAHLE
k19 k21 x11 x14
k28 k30 x16 x19
k34 k36 x16 x5
+
+
,
k20 + k21
k29 + k30
k35 + k36
k22 k24 x11 x16
k28 k29 x16 x19
k25 k27 x18 x19
k19 k21 x11 x14
−
− k28 x16 x19 +
+
=
k20 + k21
k23 + k24
k29 + k30
k26 + k27
k31 k33 x18 x22
k34 k35 x16 x5
+
− k34 x16 x5 +
,
k32 + k33
k35 + k36
k37 k39 x18 x4
k40 k42 x26 x5
k43 k45 x18 x28
′
=
−
,
f28
= k46 x19 −
.
k38 + k39
k41 + k42
k44 + k45
′
f14
=−
′
f16
′
f26
′
′
′
′
′
In particular, we find five new binomials f1′ , f7′ , f11
, f26
, and f28
. Adding f14
to f16
yields the trinomial
′′
f16
=−
k22 k24 x11 x16 k25 k27 x18 x19 k31 k33 x18 x22
+
+
.
k23 + k24
k26 + k27
k32 + k33
Consequently, the original system is equivalent to a system consisting of 27 binomials
and two trinomials of a relatively simple shape. For comparison we computed the
Gröbner basis in Macaulay2 with rational functions in the reaction rates as coefficients.
Although the computation finished in just 18 minutes, the result is practically unusable.
The Gröbner basis consists of 169 elements each of it with huge rational functions as
coefficients. The structure that we observed above is completely lost.
The lesson learned from Example 4.4 is that term replacements using binomials are
useful in solving a polynomial system, even if the end result is not binomial. Especially
in the non-homogeneous case where the notion of minimal generators is absent, computations with the Binomials package [16] in Macaulay2 [11] can probably only assist,
but not automatically do useful reductions. For example, a natural general choice would
be to replace higher degree monomials by lower degree ones, but this would not directly
reveal binomiality in Example 4.3.
Finally, we summarize a possible strategy to deal with inhomogeneous ideals. Example 4.1, for instance, is solved already by item 1, but also by item 3.
Recipe 4.5.
1) Try linear algebra and Proposition 2.5.
2) Homogenize the given ideal and run Algorithm 3.3. If the algorithm returns binomials, then by Proposition 4.2 the original (dehomogenized) ideal is binomial.
The homogenization should be carried out after linear algebra reductions to possibly detect homogeneity already at an earlier stage (compare Example 2.8).
3) If Algorithm 3.3 returns a negative answer, dehomogenize and use known binomials for term replacements (as in Example 4.4). Potentially homogenize again
with an enlarged generating set.
4) Compute a reduced Gröbner basis.
DETECTING BINOMIALITY
15
References
[1] Banaji M., Craciun G., Graph-theoretic criteria for injectivity and unique equilibria in general
chemical reaction systems, Adv. in Appl. Math. 44 (2010) 168 – 184.
[2] Bachmann O., Greuel G.-M., Lossen C., Pfister G., Schönemann H., A Singular introduction to
commutative algebra, Springer, Berlin, 2007.
[3] Craciun G., Feinberg M., Multiple equilibria in complex chemical reaction networks: II. The
species-reaction graph, SIAM J. Appl. Math. 66 (2006) 1321 – 1338.
[4] Craciun G., Feinberg M., Multiple equilibria in complex chemical reaction networks: Semiopen
mass action systems, SIAM J. Appl. Math. 70 (2010) 1859 – 1877.
[5] Conradi C., Flockerzi D., Multistationarity in mass action networks with applications to ERK
activation, J. Math. Biol. 65 (2012) 107 – 156.
[6] Cox D. A., Little J. B., O’Shea D., Ideals, Varieties, and Algorithms, Springer, New York, 1996.
[7] Eder C., Faugère J.-C., A survey on signature-based Gröbner basis computations, preprint,
arXiv:1404.1774.
[8] Eisenbud D., Sturmfels B., Binomial Ideals, Duke Math. J. 84 (1996) 1 – 45.
[9] Eisenbud D., Commutative Algebra with a View Toward Algebraic Geometry, Springer, New York,
1995.
[10] Faugère J.-C., A new efficient algorithm for computing Gröbner bases (F4), J. Pure Appl. Algebra
139 (1999) 61 – 88.
[11] Grayson D. R., Stillman M. E., Macaulay2, a software system for research in algebraic geometry,
Available at http://www.math.uiuc.edu/Macaulay2/.
[12] Joshi B., Shiu A., Simplifying the Jacobian criterion for precluding multistationarity in chemical
reaction networks, SIAM J. Appl. Math. 72 (2012) 857 – 876.
[13] Johnston M. D., Translated chemical reaction networks, Bull. Math. Biol. 76 (2014) 1081 – 1116.
[14] Kahle T., Miller E., Decompositions of Commutative Monoid Congruences and Binomial Ideals,
Algebra Number Theory 8 (2014) 1297 – 1364.
[15] Kahle T., Rauh J., Sullivant S., Positive margins and primary decomposition, J. Commut. Algebra
6 (2014) 173 – 208.
[16] Kahle T., Decompositions of Binomial Ideals, J. Softw. Algebra Geom. 4 (2012) 1 – 5.
[17] Müller S., Feliu E., Regensburger G., Conradi C., Shiu A., Dickenstein A., Sign conditions for
injectivity of generalized polynomial maps with applications to chemical reaction networks and
real algebraic geometry, Found. Comput. Math., published onlineu under doi:10.1007/s10208-0149239-3 (2015).
[18] Pérez Millán M., Dickenstein A., Shiu A., Conradi C., Chemical Reaction Systems with Toric
Steady States, Bull. Math. Biol. 74 (2012) 1027 – 1065.
[19] Pérez Millán M., Turjanski A. G., MAPK’s networks and their capacity for multistationarity due
to toric steady states, Math. Biosci. 262 (2015) 125 – 137.
[20] Oxley J. G., Matroid Theory, Oxford University Press, Oxford, 2006.
[21] Wiuf C., Feliu E., Power-law kinetics and determinant criteria for the preclusion of multistationarity in networks of interacting species, SIAM J. Appl. Dyn. Syst. 12 (2013) 1685 – 1721.
Hochschule für Technik und Wirtschaft, Berlin, Germany
E-mail address: [email protected]
Otto-von-Guericke Universität, Magdeburg, Germany
Website: http://www.thomas-kahle.de
| 0math.AC
|
1
On Repetitive Scenario Design
Giuseppe C. Calafiore
arXiv:1602.03796v1 [cs.SY] 11 Feb 2016
Abstract
Repetitive Scenario Design (RSD) is a randomized approach to robust design based on iterating two phases: a
standard scenario design phase that uses N scenarios (design samples), followed by randomized feasibility phase that
uses No test samples on the scenario solution. We give a full and exact probabilistic characterization of the number
of iterations required by the RSD approach for returning a solution, as a function of N , No , and of the desired levels
of probabilistic robustness in the solution. This novel approach broadens the applicability of the scenario technology,
since the user is now presented with a clear tradeoff between the number N of design samples and the ensuing
expected number of repetitions required by the RSD algorithm. The plain (one-shot) scenario design becomes just
one of the possibilities, sitting at one extreme of the tradeoff curve, in which one insists in finding a solution in a
single repetition: this comes at the cost of possibly high N . Other possibilities along the tradeoff curve use lower N
values, but possibly require more than one repetition.
Keywords
Scenario design, probabilistic robustness, randomized algorithms, random convex programs.
I.
I NTRODUCTION
The purpose of the approach described in this paper is to obtain a probabilistically reliable solution for some
design problem affected by uncertainty. The concept of “probabilistic design” has been discussed extensively in the
control community in the last decade, and it is now well accepted as a standard tool for tacking difficult robust design
problems; we refer the reader to the survey paper [6] and to the book [18] for many pointers to the related literature.
The essential elements of a probabilistic design approach are the following ones:
1) A spec function, f (θ, q) : Rn × Q → R, which associates a real value to each pair (θ, q) of a design parameter
θ ∈ Rn and uncertainty instance q ∈ Q, where Q ⊆ Rnq . Function f represents the design constraints and
specifications of the problem and, in particular, we shall say that a design θ is a robust design, if f (θ, q) ≤ 0,
∀q ∈ Q. In this paper, we make the standing assumption that f is convex.
2) A probability measure Prob defined on Q, which describes the probability distribution of the uncertainty.
Equipped with these two essential elements, for given ∈ (0, 1), and given design vector θ, we are in position to
define the probability of violation for the spec function at θ:
.
V (θ) = Prob{q ∈ Q : f (θ, q) > 0}.
(1)
We say that θ is an -probabilistic robust design, if it holds that V (θ) ≤ . Further, a designer also typically seeks
to minimize some cost function of θ (which can be considered of the linear form c> θ, without loss of generality;
see, e.g., Section 8.3.4.4 in [2]), while guaranteeing that V (θ) ≤ . Finding such an -probabilistic robust design
amounts to solving a so-called chance-constrained optimization problem, which is computationally hard in general,
and perhaps harder than finding a classical deterministic robust design. Chance-constrained optimization problems
can be solved exactly only in very restrictive cases (e.g., when f is linear, and q has some specific distribution,
such as Normal; see, e.g., [16]). Deterministic convex approximations of chance-constrained problems are discussed
in [13] for some special classes of problems where f is affine in q and the entries of q are independent. Also, the
sampling average approximation (SAA) method replaces the probability constraint V (θ) ≤ with one involving the
empirical probability of violation based on N sampled values of q ; see, e.g., [14], [12]. The optimization problem
resulting from SAA, however, remains non-convex and intractable, in general.
Giuseppe C. Calafiore, Dipartimento di Automatica e Informatica, Politecnico di Torino, Italy. [email protected]
2
A. The standard scenario theory
While effective approximation schemes for chance-constrained optimization problems remain to date hard to tackle
numerically, an alternative and efficient randomized scheme emerged in the last decade for finding -probabilistic
robust designs. This technique, which is now a well-established technology (see, e.g., the recent surveys [9], [15]) in
the area of robust control, is called “scenario design,” and was introduced in [4]. In scenario design one considers N
.
i.i.d. random samples of the uncertainty {q (1) , . . . , q (N ) } = ω , and builds a scenario random convex program (RCP):
min
θ∈Θ
c> θ
(2)
s.t.: f (θ, q (i) ) ≤ 0, i = 1, . . . , N,
where Θ is some given convex and compact domain, and c is the given objective direction. An optimal solution θ∗
to this problem, if it exists, is a random variable which depends on the multiextraction ω , i.e., θ∗ = θ∗ (ω). As a
consequence, the violation probability relative to a scenario solution, V (θ∗ ), is itself, a priori, a random variable.
Scenario design lies somewhere in between worst-case robust design (where c> θ is minimized subject to f (θ, q) ≤ 0
for all q ∈ Q) and chance-constrained design (where c> θ is minimized subject to V (θ) ≤ ). Indeed, the optimal
objective value resulting from a scenario design is lower than the worst-case optimal objective and it is (with
high probability) higher than a suitable chance-constrained optimal objective (see, e.g., Section 6 in [3]). Moreover,
a fundamental feature of scenario design is that its optimal solution θ∗ (ω) is feasible with high probability for the
chance-constrained problem. This key result is recalled next for the sake of clarity. We shall work under the following
simplifying assumption, which is routinely made in the literature on scenario design; see [4], [7].
Assumption 1: With probability (w.p.) one with respect to the multi-extraction ω = {q (1) , . . . , q (N ) }, problem (2)
is feasible and it attains a unique optimal solution θ∗ (ω).
?
Also, we need the following standard definition (see, e.g., Definition 4 in [4])
Definition 1: Let J ∗ = c> θ∗ denote the optimal objective value of problem (2). Also, for j = 1, . . . , N , define
.
c> θ
Jj∗ = min
θ∈Θ
s.t.: f (θ, q (i) ) ≤ 0, i ∈ {1, . . . , N } \ j.
The i-th constraint in (2) is said to be a support constraint if Jj∗ < J ∗ .
?
A key fact is that, regardless of the problem structure and of N , the number of support constraints for problem
(2) cannot exceed n (the number of decision variables); see, e.g., Theorem 3 in [4]. If an instance of problem (2)
happens to have precisely n support constraints, then the problem instance is said to be fully supported (f.s.); see
Definition 3 in [7], and Definition 2.5 in [3]. If the instances of problem (2) are fully supported almost surely with
respect to the random extraction ω of the N constraints, then we say that problem (2) is fully supported w.p. one.
The following key result holds, see Theorem 1 in [7], and Corollary 3.4 in [3].
Theorem 1: Let Assumption 1 hold. Then, for given ∈ [0, 1] and N ≥ n, it holds that
.
FV () = ProbN {ω : V (θ∗ (ω)) ≤ }
(3)
N
X
N i
≥
(1 − )N −i
(4)
i
i=n
.
= 1 − β (N ).
Moreover, the bound (4) is tight, since it holds with equality for the class of problems of the form (2) that are fully
supported with probability one.
?
A remarkable feature of the result in (4) is that it holds irrespective of the probability distribution assumed on q ,
and that it depends on the problem structure only through the dimension parameter n.
B. Scenario problems and Bernoulli trials
For given ∈ [0, 1] and N ≥ n, let us consider the following Bernoulli variable associated to problem (2):
1, if V (θ∗ (ω)) ≤
z = z(ω) =
0, otherwise.
3
By the definition in eq. (3), the event z = 1 happens w.p. FV (). One interpretation of eq. (4) is thus that each time
we solve a scenario problem (2) we have an a priori probability ≥ 1 − β (N ) of realizing a “successful design,”
that is of finding a solution θ∗ which is an -probabilistic robust design, and a probability ≤ β (N ) of realizing a
“failure,” that is of finding a solution θ∗ which is not -probabilistic robust.
In the classical scenario theory it is usually prescribed to choose N so to make β (N ) very small (values as low as
−12
10
are common). This guarantees that the event {V (θ∗ (ω)) ≤ } will happen with “practical certainty.” In other
words, in such regime, the scenario problem will return an -probabilistic robust solution with practical certainty.
Moreover, a key feature of scenario theory is that such high level of confidence can be reached at a relatively
“cheap” computational price. Indeed, considering the condition β (N ) ≤ β for some given desired probability level
β ∈ (0, 1), and using some fairly standard techniques for bounding the Binomial tail (see, e.g., Corollary 5.1 in [3]
for the details), one can prove that the condition is satisfied for1
2
ln β −1 + n − 1 .
(5)
N≥
Since β −1 appears in the above bound under a logarithm, we indeed see that N grows gracefully with the required
certainty level β −1 . However, there are cases in which the number N of constraints prescribed by (5) for reaching the
desired confidence levels is just too high for practical numerical solution. Convex optimization solvers are certainly
efficient, but there are practical limits on the number of constraints they can deal with; these limits depend on the
actual type of convex problem (say, a linear program (LP), or a semidefinite program (SDP)) one deals with. A critical
situation is, for instance, when problem (2) is a semidefinite program (formally, f can be taken as the maximum
eigenvalue function of the matrices describing the linear inequality constraints): dealing with an SDP with many
thousands of LMI constraints can pose serious practical issues.
C. Contribution
In this paper we discuss how a variation of the scenario approach can be used for obtaining an -probabilistic
robust solution with high confidence, using “small” values of N . More precisely, we are interested in using scenario
optimization in a regime of N for which the right-hand side of eq. (4) is not close to one. We shall do so by solving
repeatedly instances of the scenario problem, and checking the result via a suitable “violation oracle.” This novel
approach, named repeated scenario design (RSD), is discussed in Section II, which contains all the relevant results.
Section III describes two numerical examples of robust control design where the proposed approach is applied. For
improving readability, technical proofs are reported in the Appendix.
D. Notation and preliminaries
We shall make intensive use of the beta and related probability distributions. Some definitions and standard facts
are recalled next. We denote by beta(α, β) the beta density function with parameters α > 0, β > 0:
.
beta(α, β; t) =
1
tα−1 (1 − t)β−1 ,
B(α, β)
t ∈ [0, 1],
.
−1 = α
where B(α, β) = Γ(α)Γ(β)
Γ(α+β) , and Γ is the Gamma function (for α, β integers, it holds that B(α, β)
Also, we denote by Fbeta(α, β) the cumulative distribution function of the beta(α, β) density:
Z t
.
Fbeta(α, β; t) =
beta(α, β; ϑ)dϑ, t ∈ [0, 1].
α+β−1
β−1
).
0
Fbeta(α, β; t) is the regularized incomplete beta function, and a standard result establishes that, for α, β integers, it
holds that
α+β−1
X α + β − 1
Fbeta(α, β; t) =
ti (1 − t)α+β−1−i .
i
i=α
1
Notice that the expression
the exact minimal value of N can be easily found numerically by searching for the
may be conservative;
P in (5)
N i
N −i
least integer N such that N
≥ 1 − β.
i=n i (1 − )
4
The number x of successes in d independent Bernoulli trials each having success probability p is a random variable
with Binomial distribution (which we denote by Bin(d, p)); its cumulative distribution is given by
bzc
X
d i
Prob{x ≤ z} = Prob{x ≤ bzc} =
t (1 − t)d−i
i
i=0
d
X
d i
t (1 − t)d−i
= 1−
i
i=bzc+1
= 1 − Fbeta(bzc + 1, d − bzc; t)
≤ 1 − Fbeta(z + 1, d − z; t)
(6)
= Fbeta(d − z, z + 1; 1 − t),
where bzc denotes the largest integer no larger than z . The number x of successes in d binary trials, where each trial
has success probability p, and p is itself a random variable with beta(α, β) distribution, is a random variable with
so-called beta-Binomial distribution: for i = 0, 1, . . . , d,
. d B(i + α, d − i + β)
f bb (d, α, β; i) =
.
(7)
i
B(α, β)
The cumulative distribution of a beta-Binomial random variable is given by (see, e.g., [11], [19])
bzc
X
.
f bb (d, α, β; i)
F bb (d, α, β; z) = Prob{x ≤ z} =
i=0
1 B(β + d − z − 1, α + z + 1)
=1−
Fd (a, b; z),
d + 1 B(α, β)B(d − z, z + 2)
where Fd (a, b; z) is the generalized hypergeometric function
3 F2 (1, α
+ z + 1, −n + z + 1; z + 2, −β − n + z + 2; 1).
II.
R EPETITIVE SCENARIO DESIGN
This section develops the main idea of this paper. By repetitive scenario design (RSD) we here mean an iterative
computational approach in which, at each iteration k , the scenario problem (2) is solved and then the ensuing solution
θk∗ is checked by a violation oracle (either deterministic, or randomized, as illustrated next). If the oracle returns
false, another iteration is performed; if instead the oracle returns true, the algorithm is terminated and the current
solution θk∗ is returned.
In the RSD the user selects a desired probabilistic feasibility level ∈ (0, 1), and a number N ≥ n of scenarios
to be used in (2). We have from Theorem 1 that, at any iteration k , it holds that
ProbN {ω (k) : V (θk∗ ) ≤ } = FV () ≥ 1 − β (N ),
(1)
(N )
(8)
where ω (k) denotes the multisample {qk , . . . , qk }. In very elementary terms, each iteration of the RSD method
can be thought of as a biased “coin toss,” where the probability of a success in a toss (that is, getting θk∗ such that
V (θk∗ ) ≤ ) is at least 1 − β (N ). In our setting, this probability need not be too close to one: the simple idea behind
the RSD method is to repeat the coin toss until we obtain a success, where success is detected by the violation oracle.
As one may easily argue intuitively, the probability of obtaining a success at some point in the algorithm is much
higher than the probability of obtaining a success in a single toss. A similar idea has been recently proposed in [8],
where the authors solve repeatedly a “reduced-size” scenario problem, followed by a randomized test of feasibility.
The approach and the results in [8], however, are distinctively different from the ones proposed here. In [8], the
scenario problems are solved using a number Nk of scenarios that grows with the iteration count k , up to the value
N plain that corresponds to the plain, one-shot, scenario design. The major shortcoming of the approach and analysis
in [8] is that the number of iterations is not bounded a-priori, either in a deterministic or in a probabilistic sense,
and no tradeoff curve is proposed for the choice of Nk in function of the expected running time of the algorithm.
As a result, there is no a-priori guarantee that the algorithm does not reach the final iteration, in which Nk equals
5
N plain , hence the worst-case complexity of the algorithm in [8] can be worse than the one of the plain scenario
design method, and an actual reduction of the number of design samples is not theoretically guaranteed.
We shall next analyze precisely the probabilistic features of our RSD algorithm in two cases. In the first case
we assume that an ideal exact feasibility oracle is available for checking the current solution θk∗ ; this case may be
unrealistic in general, but serves for providing an insightful preliminary analysis of the RSD approach. In the second
case, we analyze the RSD approach when a practically implementable randomized feasibility oracle is used.
A. Violation oracles
A deterministic -violation oracle (-DVO) is a “black box” which, when given in input a value of the design
variable θ, returns as output a flag value which is true if V (θ) ≤ , and false otherwise. Such an oracle may
not realizable computationally in practice, since computing the probability in (1) is numerically hard, in general. For
this reason, we next also introduce a randomized 0 -violation oracle (0 -RVO), which is defined by means of the
randomized scheme described next.
0 -RVO (Randomized 0 -violation oracle) Input data: integer No , level 0 ∈ [0, 1], and θ ∈ Rn . Output data: a logic
flag, true or false.
.
1) Generate No i.i.d. samples ωo = {q (1) , . . . , q (No ) }, according to Prob.
2) ForPi = 1, . . . , No , let vi = 1 if f (θ, q (i) ) > 0 and vi = 0 otherwise.
3) If i vi ≤ 0 No , return true, else return false.
The 0 -RVO simply evaluates the empirical probability of violation on No samples, and returns true if it is below
and false otherwise. A similar type of randomized feasibility oracle has been previously introduced in [5], and
used in a probabilistic design setting also in [6]; see also Section 11.1 in [18], and the “validation” step proposed
in [8]. However, the 0 -RVO we propose in this paper is different from the one used in the cited references: the
latter exits with a false flag as soon as one infeasible sample is found, whereas the 0 -RVO allows up to b0 No c
infeasible samples before exit. Also, the kind of a priori analysis we develop here for the repetitive scenario design
based on the 0 -RVO is entirely novel.
0 ,
B. Repetitive scenario design with ideal oracle
We consider the following RSD algorithm, where each repetition consists of a plain scenario optimization step,
followed by a feasibility check of the ensuing solution, performed by an exact feasibility oracle.
Algorithm 1 (RSD with -DVO): Input data: integer N ≥ n, level ∈ [0, 1]. Output data: solution θ∗ . Initialization:
set iteration counter to k = 1.
.
(1)
(N )
1) (Scenario step) Generate N i.i.d. samples ω (k) = {qk , . . . , qk } according to Prob, and solve scenario
problem (2). Let θk∗ be the resulting optimal solution.
2) (-DVO step) If V (θk∗ ) ≤ , then set flag to true, else set it to false.
3) (Exit condition) If flag is true, then exit and return current solution θ∗ ← θk∗ ; else set k ← k + 1 and goto 1.
?
The following theorem holds.
Theorem 2: Let Assumption 1 hold. Given ∈ [0, 1] and N ≥ n, define the running time K of Algorithm 1 as
the value of the iteration counter k when the algorithm exits. Then:
1) The solution θ∗ returned by Algorithm 1 is an -probabilistic robust design, i.e., V (θ∗ ) ≤ .
2) The expected running time of Algorithm 1 is ≤ (1 − β (N ))−1 , and equality holds if the scenario problem is
f.s. w.p. 1.
3) The running time of Algorithm 1 is ≤ k with probability ≥ 1 − β (N )k , and equality holds if the scenario
problem is f.s. w.p. 1.
?
See Section A in the Appendix for a proof of Theorem 2.
Remark 1 (Potential and limits of the RSD approach): The preliminary results in Theorem 2 show the potential of
the RSD approach. Suppose that N is chosen so that β (N ) is about, say, 0.4. This means that a plain (i.e., one-shot)
6
scenario approach has only at least a 0.6 chance of returning a “good” solution (i.e., an -probabilistic robust design: a
θ∗ such that V (θ∗ ) ≤ ). However, we see from point 3 of Theorem 2 that there is more than 1−10−9 probability that
Algorithm 1 returns an -probabilistic robust design within 23 iterations. Further, the eventual outcome of Algorithm 1
is -probabilistic robust with probability one, and the expected number of iterations of the RSD algorithm is just
(1 − 0.4)−1 = 1.67, in the worst case of a f.s. problem.
Theorem 2 also shows a fundamental limit of the RSD approach: we can decrease N (and hence increase β (N ))
with respect to a plain scenario design approach, but we cannot decrease N too much, for otherwise β (N ) → 1,
and the expected number of iterations of Algorithm 1 tends to ∞. There is thus a fundamental tradeoff between
the reduction of N (which reduces the effort needed for solving the scenario problem) and the increase of the
number of iterations of Algorithm 1. This tradeoff can be fully captured by plotting the expected running time bound
(1 − β (N ))−1 versus the number N of scenarios.
C. Repetitive scenario design with randomized oracle
This section contains the main contribution of this paper, where we consider a realistically implementable version
of the RSD approach, in which a randomized oracle is used instead of the ideal deterministic one.
Algorithm 2 (RSD with 0 -RVO): Input data: integers N , No , level 0 ∈ [0, 1]. Output data: solution θ∗ . Initialization: set iteration counter to k = 1.
.
(1)
(N )
1) (Scenario step) Generate N i.i.d. samples ω (k) = {qk , . . . , qk } according to Prob, and solve scenario
∗
problem (2). Let θk be the resulting optimal solution.
2) (0 -RVO step) Call the 0 -RVO with current θk∗ as input, and set flag to true or false according to the
output of the 0 -RVO.
3) (Exit condition) If flag is true, then exit and return current solution θ∗ ← θk∗ ; else set k ← k + 1 and goto 1.
?
A generic iteration, or stage, k , of Algorithm 2 is illustrated in Figure 1.
()
()
‘
Fig. 1.
Generic stage k of Algorithm 2.
We next analyze Algorithm 2 along two directions. First, contrary to Algorithm 1, the present algorithm may exit
with a solution which is not -probabilistic robust. This is due to the randomized nature of the oracle, which may
detect a “false positive,” by misclassifying as good a solution θk∗ for which V (θk∗ ) > . We show that the probability
of such a “bad exit” event can be made arbitrarily small. Second, we fully characterize the probabilistic running time
(iterations to exit) of the algorithm. We start with the following key preliminary lemma, which is the backbone of
the whole paper.
Lemma 1: Let Assumption 1 hold and, for any given iteration k , define the events
True = {0 -RVO returns true}
GoodTrue = {0 -RVO returns true ∩ V (θk∗ ) ≤ }
BadTrue = {0 -RVO returns true ∩ V (θk∗ ) > }
7
Let
f bb (No , n, N + 1 − n; i)
. No B(i + n, No − i + N − n + 1)
=
,
i
B(n, N + 1 − n)
H,0 (N, No )
0
b No c
X
.
=1−
f bb (No , n, N + 1 − n; i) ·
i=0
· Fbeta(n + i, N + No − n − i + 1; ),
H1,0 (N, No )
0
b No c
X
.
=1−
f bb (No , n, N + 1 − n; i),
i=0
β̄,0 (N, No )
.
= Fbeta(N + (1 − 0 )No − n + 1, n + 0 No ; 1 − ).
At any iteration k of Algorithm 2, it holds that
,ProbN +No {True} ≥ 1 − H1,0 (N, No )
N +No
Prob
(9)
{GoodTrue} ≥ 1 − H,0 (N, No )
(10)
≥ (1 − β̄,0 (N, No ))(1 − H1,0 (N, No )).
(11)
N +No
(12)
Prob
{BadTrue}
0
0
≤ Fbeta((1 − )No , No + 1; 1 − )β (N ).
Moreover, if problem (2) is f.s. w.p. one, then bounds (9) and (10) hold with equality, and
ProbN +No {BadTrue} = H,0 (N, No ) − H1,0 (N, No )
≤ β̄,0 (N, No )(1 − H1,0 (N, No )).
(13)
?
See Section B in the Appendix for a proof of Lemma 1.
We can now state the main result concerning Algorithm 2.
Theorem 3: Let Assumption 1 hold. Let , 0 ∈ [0, 1], 0 ≤ , and N ≥ n be given. Let all the notation be set as in
Lemma 1, and let Prob×× denote the product probability ProbN +No × ProbN +No × · · · . Define the event BadExit
in which Algorithm 2 exits returning a “bad” solution θ∗ :
.
BadExit = {Algorithm 2 returns θ∗ : V (θ∗ ) > }.
The following statements hold.
1)
Prob×× {BadExit}
Fbeta((1 − 0 )No , 0 No + 1; 1 − )
≤
β (N ).
1 − H1,0 (N, No )
If problem (2) is f.s. w.p. one, then it actually holds that
Prob×× {BadExit} ≤ β̄,0 (N, No )).
))−1 ,
(14)
(15)
and equality holds if the scenario
The expected running time of Algorithm 2 is ≤ (1 − H1,0 (N, No
problem is f.s. w.p. 1.
3) The running time of Algorithm 2 is ≤ k with probability ≥ 1 − H1,0 (N, No )k , and equality holds if the
scenario problem is f.s. w.p. 1.
?
See Section C in the Appendix for a proof of Theorem 3.
2)
8
1) Asymptotic bounds: A key quantity related to the expected running time of Algorithm 2 is H1,0 (N, No ), which
is the upper tail of a beta-Binomial distribution. This quantity is related to the hypergeometric function 3 F2 , and
to ratios of Gamma functions, which may be delicate to evaluate numerically for large values of the arguments. It
is therefore useful to have a more “manageable,” albeit approximate, expression for H1,0 (N, No ). The following
corollary gives an asymptotic expression for H1,0 (N, No ), see Section D in the Appendix for a proof.
Corollary 1: For No → ∞ it holds that
H1,0 (N, No ) → β0 (N ).
(16)
?
An interesting consequence of Corollary 1 is that, for large No , and 0 ≤ , we have H1,0 (N, No ) ' β0 (N ) ≥ β (N ),
from which we conclude that
1
1
1
.
'
≥
.
(17)
K̂ =
1 − H1,0 (N, No )
1 − β0 (N )
1 − β (N )
This last equation gives us an approximate, asymptotic, expression for the upper bound K̂ on the expected running
time of Algorithm 2, and also tells us that, for 0 ≤ , this bound cannot be better (smaller) than the corresponding
bound of the “ideal” Algorithm 1.
D. Practical dimensioning of the scenario and oracle blocks
In a typical probabilistic design problem we are given the dimension n of the decision variable and the level
∈ (0, 1) of probabilistic robustness we require from our design. If we intend to use a randomized approach, we also
set a confidence level 1 − β ∈ (0, 1), which is the a-priori level of probability with which our randomized approach
will be successfull in returning an -probabilistic robust design. In a plain (i.e., non repetitive) scenario design setting,
this requires dimensioning the number N of scenarios so to guarantee that β (N ) ≤ β ; this can be done, for instance,
by using the bound in (5), or via a simple numerical search over N . However, if the required N turns out to be too
large in practice (e.g., the ensuing scenario optimization problem becomes impractical to deal with numerically), we
can switch to a repetitive scenario design approach. In such a case, we suggest the following route for designing
the scenario and oracle blocks. Let us first select a level 0 ≤ to be used in the oracle. Qualitatively, decreasing 0
increases the expected running time and decreases the required No , and the converse happens for increasing 0 . We
here suggest to set 0 in the range [0.5, 0.9].
1) Dimensioning the scenario block: We dimension the scenario optimization block by choosing N so to achieve
a good tradeoff between the complexity of the scenario program (which grows with N ) and the expected number
of iterations required by the RSD approach (which decreases with N ). This choice can be made, for instance, by
plotting the approximate expression (which becomes exact as No → ∞) in (17) for the upper bound on the expected
running time of Algorithm 2, (1 − β0 (N ))−1 , versus N , and selecting a value of N for which this running time is
acceptable.
2) Dimensioning the oracle block: Once N has been selected according to the approach described above, we
consider point 1 and point 2 in Theorem 3 and we dimension the 0 -RVO block by searching numerically for an No
such that the right-hand side of (14) (or of (15), if the problem is f.s.) is ≤ β .
Remark 2: We observe that, in general, the bound in (14) should be used for the design of the 0 -RVO block.
However, the expression in (15) is easier to deal with than the one in (14). It is hence advisable to use the former
in a preliminary dimensioning phase; the so-obtained values can then be verified ex-post against the actual bound in
(14). Another advantage of (15) is that, using a bounding technique analogous to the one described in Section 5 of
[3], we can “invert” the condition β̄,0 (N, No ) ≤ β , finding (after some manipulation) that this condition is satisfied
if
.
No δ + N (δ/2 + 0 ) ≥ ln β −1 + n − 1, δ = − 0 > 0.
(18)
δ
With a choice of the pair (N, No ) such that (18) is satisfied, we guarantee a priori that our randomized Algorithm 2
may fail in returning an -probabilistic robust design w.p. at most β , as desired (rigorously, this only holds under
the assumption that the scenario problem is f.s. w.p. one). The nice feature highlighted by (18) is that now the
“workload” necessary to achieve the desired failure level β is subdivided between N (samples in the scenario
problem) and No (samples in the oracle): a lower complexity scenario problem can be employed, as long as it is
9
paired with a randomized oracle having a suitable No . Notice, however, that, in making the choice of the (N, No )
pair, the expected running time of Algorithm 2 should also taken into consideration, and that this places a lower limit
on how small N can be, see also the discussion in Section II-C1.
Remark 3: We further observe that, in typical cases, dealing with large No is a milder problem than dealing with
large N . This is due to the fact that merely checking satisfaction of inequality f (θk∗ , q (i) ) > 0, for i = 1, . . . , No ,
is generally easier than solving a related optimization problem with as many constraints. Also, we remark that the
0 -RVO algorithm is inherently parallel, so an M -fold speedup can potentially be gained if M processors are available
in parallel for the randomized feasibility test. Actually, the whole approach can be formulated in a fully parallel –
instead of sequential – form, where W workers solve in parallel W instances of scenario problems, and each worker
has its own M parallel sub-workers to be used in the randomized oracle. Such a parallel version of the RSD method
can be easily analyzed using the probabilistic tools developed in this paper.
III.
N UMERICAL EXAMPLES
We exemplify the steps of the RSD approach, from algorithm dimensioning to numerical results, using two examples
of robust control design. The first example deals with robust finite-horizon input design for an uncertain linear system,
while the second example deals with robust performance design for a positive linear system.
A. Robust finite-horizon input design
We consider a system of the form
x(t + 1) = A(q)x(t) + Bu(t),
t = 0, 1, . . . ; x(0) = 0,
where u(t) is a scalar input signal, and A(q) ∈ Rna ,na is an interval uncertain matrix of the form
A(q) = A0 +
na
X
qij ei e>
j ,
|qij | ≤ ρ, ρ > 0,
i,j=1
where ei is a vector of all zeros, except for a one in the i-th entry. Given a final time T ≥ 1 and a target state x̄, the
problem is to determine an input sequence {u(0), . . . , u(T − 1)}
P such that (i) the state x(T ) is robustly contained in a
small ball around the target state x̄, and (ii) the input energy k u(k)2 is not too large. We write x(T ) = x(T ; q) =
.
R(q)u, where R(q) is the T -reachability matrix of the system (for a given q ), and u = (u(0), . . . , u(T − 1)). Then,
we formally express our design goals in the form of minimization of a level γ such that
kx(T ; q) − x̄k22 + λ
T
−1
X
u(t)2 ≤ γ,
t=0
where λ ≥ 0 is a tradeoff parameter. Letting θ = (u, γ), the problem is formally stated in our framework by setting
.
f (θ, q) ≤ 0, where f (θ, q) = kR(q)u − x̄k22 + λkuk22 − γ.
Assuming that the uncertain parameter q is random and uniformly distributed in the hypercube Q = [−ρ, ρ]na ×na ,
our scenario design problem takes the following form:
min
γ
θ=(u,γ)
s.t.: f (θ, q (i) ) ≤ 0,
i = 1, . . . , N.
a) Dimensioning the RSD algorithm: We set T = 10, thus the size of the decision variable θ = (u, γ) of the
scenario problem is n = 11. We set the desired level of probabilistic robustness to 1 − = 0.995, i.e., = 0.005, and
require a level of failure of the randomized method below β = 10−12 , that is, we require the randomized method to
return a good solution with “practical certainty.” Using a plain (one-shot) scenario approach, imposing β (N ) ≤ β
would require N ≥ 10440 scenarios. Let us now see how we can reduce this N figure by resorting to a repetitive
scenario design approach.
Let us fix 0 = 0.7 = 0.0035, thus δ = − 0 = 0.0015. A plot of the (asymptotic) bound on expected number
of iterations, (1 − β0 (N ))−1 , as a function of N is shown in Figure 2. We see from this plot, for instance, that the
10
3
10
2
10
1
10
0
10 3
10
Fig. 2.
4
10
Example in Section III-A: Log-log plot of (1 − β0 (N ))−1 vs. N .
A0 =
−0.7214
0.5704
−1.3983
−0.6308
0.7539
−1.3571
−0.0578
0.1762
−0.1795
−0.0058
0.1423
−0.1769
0.2757 0.7255
0.2171
0.3901
0.3684 −0.0971 0.6822 −0.5604
0.1511 1.0531 −0.1601 0.9031
0.4422 0.8169
0.512
0.2105
0.2039 −0.3757 0.5088 −0.6081
0.1076 1.0032 −0.1781 0.9151
,
B=
0
1
0
1
0
1
choice N = 2000 corresponds to a value of about 10 for the upper bound on the expected number of iterations in
Algorithm 2. Let us choose this value of N for the scenario block.
For β = 10−12 , the simplified condition in (18) tells us that No ≥ 62403. Let us choose No = 63000 samples to
be used in the oracle. With the above choices we have H1,0 (N, No ) = 0.8963, thus the algorithm’s upper bound on
average running time is K̂ = (1 − H1,0 (N, No ))−1 = 9.64. Notice that this upper bound is tight for f.s. problems,
but it is conservative for problems that are not necessarily f.s. Thus, in general, we may expect a performance which
is in practice better than the one predicted by the theoretical worst-case bound.
b) Numerical test: We considered the nominal matrix A0 of dimension na = 6 and B matrix shown on top of
this page, with target state x̄ = [1, −1/2, 2, 1, −1, 2]> , ρ = 0.001, and λ = 0.005. We run Algorithm 2 for 100 times,
and on each test run we recorded the number of iterations and the solution returned upon exit. Figure 3(a) shows the
number of repetitions in the test runs: we see that the algorithm exited most of the times in a single repetition, with a
maximum of 4 repetitions, which is below the figure predicted by the upper bound K̂ = 9.64: practical performance
was thus better than predicted, which suggests that the problem at hand is not fully supported w.p. 1. Figure 3(b)
shows the level of empirical violation probability evaluated by the oracle upon exit. Finally, Figure 4(a) shows the
optimal γ level returned by the algorithm in the test runs, and Figure 4(b) shown the optimal input signal returned
by the algorithm, averaged over the 100 test runs.
c) Computational improvements: In this example, the RSD approach permitted a substantial reduction of the
number of design samples (from the 10440 samples required by the plain scenario method, to just 2000 samples), at
the price of a very moderate number of repetitions (the average number of repetitions in the 100 test runs was 1.27).
The numerical experiments were carried out on an Intel Xeon X5650 machine using CVX under Matlab; [10]. On
average over the 100 test experiments, the RSD method (with N = 2000, No = 63000) required 224 s to return a
solution. For comparison purposes, we also run a plain, one-shot, scenario optimization with the N = 10440 scenarios
that are required to attain the desired β = 10−12 level: the time required for obtaining such a solution was 2790
s. Using the RSD approach instead of a plain one-shot scenario design thus yielded a reduction in computing time
of about one order of magnitude. The reason for this improvement is due to the fact that the scenario optimization
problem in the RSD approach (which uses N = 2000 scenarios) took about 173 s to be solved on a typical run, and
the subsequent randomized oracle test (with No = 63000) is computationally cheap, taking only about 3.16 s.
11
(a)
4
3.5
3.5
3
3
Number of iterations
(b)
−3
x 10
2.5
2.5
2
2
1.5
1.5
1
1
0.5
0.5
0
0
10
20
30
40
50
Test run
60
70
80
90
0
100
0
10
20
30
40
50
Test run
60
70
80
90
100
Fig. 3. Example in Section III-A: (a) Repetitions of Algorithm 2 in the 100 test runs. (b) Levels of empirical violation probability evaluated
by the oracle upon exit, in the 100 test runs.
(a)
0.0565
0.056
0.3
0.0555
0.25
0.2
0.0545
u(t)
γ
0.055
0.054
0.15
0.1
0.0535
0.05
0.053
0.0525
(b)
0.35
0
10
20
30
40
50
Test run
60
70
80
90
0
100
0
1
2
3
4
t
5
6
7
8
9
Fig. 4. Example in Section III-A: (a) Optimal γ level returned by Algorithm 2 in the 100 test runs. (b) Average over the 100 test runs of the
optimal input u(t) returned by Algorithm 2.
w
x2
x1
x4
x3
Fig. 5.
Example in Section III-B: A network model.
B. An uncertain linear transportation network
As a second example, we consider a variation on a transportation network model introduced in Section 3 of [17];
see Figure 5.
The model is described by the state equations where the states xi , i = 1, . . . , 4, represent the contents of four
buffers, the parameters `ij ≥ 0 represent the rate of transfer from buffer j to buffer i, w(t) ≥ 0 is an input flow on
12
−1 − `31
`12
0
0
0
−`12 − `32
`23
0
x +
ẋ =
`31
`32
−`23 − `43
`34
0
0
`43
−4 − `34
1 1 1 1 x.
y =
min
0
1
w
0
0
γ
`12 ,`23 ,`32 ∈[0,1];γ ξ≥0
s.t.:
(i)
−3 − q1
0
2 + q1
0
min
`12
0
0
−`12 − `32
`23
0
(i)
(i)
`32
−`23 − 2 − q3
1 + q2
(i)
(i)
0
2 + q3
−5 − q2
Cξ < γ1.
s.t.:
(i)
−3 − q1
0
(i)
2 + q1
0
(20)
(21)
ξ + B1 < 0, i = 1, . . . , N
γ
µ≥0;ξ≥0;γ
(19)
(22)
0
0
0
1
0
0
0
0
0
−1 −1 1
(i)
(i) ξ +
0
1 −1
0 −2 − q3
1 + q2
(i)
(i)
0
0
0
0 2 + q3
−5 − q2
Cξ < γ1, µ12 ≤ ξ2 , µ32 ≤ ξ2 , µ23 ≤ ξ3 .
µ + B1 < 0, i = 1, . . . , N
the second buffer, and we take as output y the total content of the buffers; see eq. (19)-(20).
We consider the situation in which `31 = 2 + q1 , `34 = 1 + q2 , `43 = 2 + q3 , where ` = [`12 `23 `32 ]> ∈ [0, 1]3
is a vector of parameters to be designed, and q = [q1 q2 q3 ]> is an uncertainty term, which is assumed to be a
truncated Normal random vector with zero mean, covariance matrix Σ = 0.22 I , and kqk∞ ≤ 1. This system has the
form ẋ = A(`, q)x + Bw, y = Cx, where B ≥ 0, C ≥ 0 (element-wise), and the A(`, q) matrix is Metzler (i.e., the
off-diagonal entries of A are nonnegative). Theorem 4 in [17] states that, for given `, q , this system is stable and the
peak-to-peak gain from w to y is smaller that some given γ if and only if there exist ξ ≥ 0 such that
A(`, q) B
ξ
0
<
,
C
0
1
γ1
where 1 is a vector of ones. By taking N i.d.d. samples q (i) of q , a robust scenario design problem is one in which
one seeks to minimize the peak-to-peak gain γ subject to the above constraint on the scenarios; see eq. (21). This
problem is a “robustified” version of the one discussed in Section V of [17]. The problem as stated is not convex, due
to the product terms between entries in ξ and `. However, by introducing new variables µ12 = `12 ξ2 , µ32 = `32 ξ2 ,
µ23 = `23 ξ3 , we rewrite the problem as an LP in the variables ξ , µ = [µ12 µ32 µ23 ]> , and γ ; see eq. (22).
d) Dimensioning the RSD algorithm: The size of the decision variable θ = (ξ, µ, γ) of the scenario problem
is n = 8. As in the previous example, we set the desired level of probabilistic robustness to 1 − = 0.995, i.e.,
= 0.005, and require a level of failure of the randomized method below β = 10−12 . Using a plain (one-shot)
scenario approach, imposing β (N ) ≤ β would require N ≥ 9197 scenarios. We next reduce this N figure by
resorting to a repetitive scenario design approach.
Let us fix 0 = 0.7 = 0.0035, thus δ = − 0 = 0.0015. Plotting the asymptotic bound on expected number of
iterations, (1 − β0 (N ))−1 as a function of N (as we did in Figure 2 for the previous example), we see that the
13
choice N = 1340 corresponds to a value of about 10 for the upper bound on the expected number of iterations in
Algorithm 2. Let us choose this value of N for the scenario block.
For β = 10−12 , the simplified condition in (18) tells us that No ≥ 62273 samples can be used in the randomized
feasibility oracle. With the above choices we have H1,0 (N, No ) = 0.8931, thus the algorithm’s upper bound on
average running time is K̂ = (1−H1,0 (N, No ))−1 = 9.36 (notice again that, in general, we may expect a performance
which is in practice better than the one predicted by this theoretical worst-case bound, since the the actual problem
may not be fully supported).
e) Numerical test and computational performance: We first solved the problem via a plain scenario approach,
using N = 9197 scenarios. The computational time was of about 50 s, resulting in the following optimal solution:
0.2314
0.5000
0.5000
ξ=
1.7206 , µ = 0.5000 , γ = 3.4283.
0.0000
0.9763
Next, we run the RSD method (Algorithm 2, with N = 1340, No = 62273) for 100 times, and on each test run we
recorded the number of iterations and the solution returned upon exit. Figure 6(a) shows the number of repetitions
in the test runs: we see that the algorithm exited most of the times in a single repetition, with a maximum of 3
repetitions; average 1.24 repetitions. Figure 6(b) shows the level of empirical violation probability evaluated by the
oracle upon exit. Finally, Figure 7 shows the optimal γ level returned by the algorithm in the test runs.
The average (over the 100 test trials) running time of the RSD method was about 6.4 s. Since the plain scenario
approach required about 50 s, it was about 680% slower than the newly proposed RSD approach, in this test example.
Each repetition of the RSD method required about 4.6 s for solving the scenario problem (with N = 1340), and 0.6
s for the randomized oracle check (with No = 62273); once again, we observe that the oracle time was much lower
than the scenario optimization time.
(b)
(a)
3
3.5
-3
3
2.5
Number of iterations
# 10
2.5
2
2
1.5
1.5
1
1
0.5
0
0.5
0
10
20
30
40
50
Test run
60
70
80
90
100
0
0
10
20
30
40
50
Test run
60
70
80
90
100
Fig. 6. Example in Section III-B: (a) Repetitions of Algorithm 2 in the 100 test runs. (b) Levels of empirical violation probability evaluated
by the oracle upon exit, in the 100 test runs.
IV.
C ONCLUSIONS
Repetitive scenario design generalizes the scenario approach to robust design by setting up an iterative procedure
whereby scenario design trials are followed by a randomized check on the feasibility level of the solution. The
expected number of repetitions (or trials) in this procedure is dictated by the key quantity H1,0 (N, No ), which is
well approximated, for large No , by β0 (N ). For H1,0 (N, No ) → 0 we recover the extreme situation of the standard,
one-shot, scenario design, in which a valid solution is found in a single repetition, at the cost of possibly large N .
For smaller N values, we can trade off complexity in the solution of the scenario problem for additional iterations
in the RSD algorithm. The extent to which N can be reduced is however limited by the upper bound K̂ we impose
14
3.8
3.6
3.4
.
3.2
3
2.8
2.6
2.4
0
10
20
30
40
50
60
70
80
90
100
Test run
Fig. 7.
Example in Section III-B: Optimal γ level returned by Algorithm 2 in the 100 test runs.
on the expected running time, since eq. (17) tells us that H1,0 (N, No ) ' β0 (N ) ≤ 1 − K̂ −1 . Numerical examples
showed that the proposed RSD approach may lead to improvements in computational time of about one order of
magnitude, compared to a plain scenario approach.
A PPENDIX
A. Proof of Theorem 2
The first point of the theorem is obvious, sice the algorithm terminates if and only if true is returned by the deterministic
oracle, which happens if and only if the condition V (θk∗ ) ≤ is satisfied.
For point two, let zk = zk (ω (k) ), k = 1, . . ., be i.i.d. Bernoulli variables representing the outcome of the -DVO step at
each iteration, i.e., zk = 1 if V (θk∗ ) ≤ (oracle returns true), and zk = 0 otherwise (oracle returns false). From eq. (8) we
observe that the probability of zk = 1 is FV () ≥ 1 − β (N ). Since the algorithm terminates as soon as a true is returned by
the oracle, the running time of the algorithm is defined as the random variable
.
K = {iteration k at which true is returned for the first time}.
Clearly, K has a geometric distribution
Prob× {K = k} = (1 − FV ())k−1 FV (),
where Prob× denotes the product probability measure over ω (1) , ω (2) , . . . The mean of this geometric distribution is 1/FV (),
whence
1
1
≤
,
E{K} =
FV ()
1 − β (N )
which proves the second point (note that equality holds if the scenario problem is f.s. w.p. one). The cumulative of the above
geometric distribution is
Prob× {K ≤ k} = 1 − (1 − FV ())k .
This function is increasing in FV (), thus FV () ≥ 1 − β (N ) implies
Prob× {K ≤ k} ≥ 1 − β (N )k ,
which proves the third point.
B. Proof of Lemma 1
At any given iteration k of Algorithm 2, let us consider the sequence of binary random variables appearing inside the 0 -RVO:
1 if f (θk∗ , q (i) ) > 0
vi =
i = 1, . . . , No
0 otherwise,
By definition, we have that Prob{q : f (θk∗ , q) > 0} = V (θk∗ ), and V (θk∗ ) is a random variable with cumulative distribution
function given by FV . Therefore, for given V (θk∗ ) = p, the vi s form an i.i.d. Bernoulli sequence with success probability p.
However, p is itself a random variable having cumulative distribution FV . Therefore, the vi s form a so-called conditionally i.i.d.
Bernoulli sequence [1], having FV as the directing de Finetti measure. In simpler terms, the vi s are described by a compound
distribution: first a success probability p is extracted at random according to its directing distribution FV , and then the vi s are
15
. PNo
vi . Conditional on V (θk∗ ) = p,
generated according to an i.i.d. Bernoulli distribution with success probability p. Let S = i=1
the random variable S has Binomial distribution Bin(No , p) thus, from (6),
ProbNo {S ≤ z|V (θk∗ ) = p} =
bzc
X
No i
p (1 − p)No −i
i
i=0
= Fbeta(No − bzc, bzc + 1, 1 − p)
= 1 − Fbeta(bzc + 1, No − bzc, p)
≤ 1 − Fbeta(z + 1, No − z, p).
Considering eq. (4), we next let
.
FV (t) = Fbeta(n, N + 1 − n; t) + Ψ(t),
t ∈ [0, 1],
(23)
(24)
where Ψ(t) is some unknown function such that 0 ≤ Ψ(t) ≤ 1 − Fbeta(n, N + 1 − n; t), for all t ∈ [0, 1], and Ψ(0) = Ψ(1) = 0.
Observe that Ψ(t) is identically zero if the scenario problem is f.s. w.p. one. Consider the event
.
GoodTrue = {True ∩ V (θk∗ ) ≤ }
{S ≤ b0 No c ∩ V (θk∗ ) ≤ }.
=
.
Leting z = b0 No c, we have that
ProbN +No {GoodTrue} = ProbN +No {S ≤ z ∩ V (θk∗ ) ≤ }
Z
=
ProbNo {S ≤ z|V (θk∗ ) = t}dFV (t)
0
Z
[using (23)] =
Fbeta(No − z, z + 1; 1 − t)dFV (t)
0
[using (24)] = (1 − H,0 (N, No )) + R(),
(25)
where we defined
.
H,0 (N, No ) =
Z
1−
Fbeta(No − z, z + 1; 1 − t) ·
0
.
R() =
· beta(n, N + 1 − n; t)dt
Z
Fbeta(No − z, z + 1; 1 − t)dΨ(t).
(26)
(27)
0
We next analyze the above two terms. For the first term, we have
1 − H,0 (N, No )
Z
=
Fbeta(No − z, z + 1; 1 − t)beta(n, N + 1 − n; t)dt
0
z Z
X
No i
t (1 − t)No −i ·
[using (23)] =
i
0
i=0
· beta(n, N + 1 − n; t)dt
z Z
X
No
1
=
ti+n−1 (1 − t)No −i+N −n dt
i
B(n,
N
+
1
−
n)
0
i=0
z Z
X
No B(i + n, No − i + N − n + 1)
=
·
i
B(n, N + 1 − n)
i=0 0
· beta(i + n, No − i + N − n + 1; t)dt
z Z
X
[by def. in (7)] =
f bb (No , n, N + 1 − n; i) ·
i=0
0
· beta(i + n, No − i + N − n + 1; t)dt
=
z
X
f bb (No , n, N + 1 − n; i) ·
i=0
· Fbeta(n + i, N + No − n − i + 1; ).
(28)
16
Observe that, for all i = 0, . . . , z, it holds that
Fbeta(n + i, N + No − n − i + 1; )
NX
+No
N + No j
=
(1 − )N +No −j
j
j=n+i
NX
+No
N + No j
≥
(1 − )N +No −j
j
j=n+z
= Fbeta(n + z, N + No − n − z + 1; ).
Therefore, we obtain following bound
1 − H,0 (N, No )
≥ Fbeta(n + z, N + No − n − z + 1; )
z
X
·
f bb (No , n, N + 1 − n; i)
i=0
= (1 − Fbeta(N + No − n − z + 1, n + z; 1 − )) ·
· (1 − H1,0 (N, No )) .
For z = b0 No c, we have, in particular, that
1 − H,0 (N, No )
≥ (1 − Fbeta(N + No − n − b0 No c + 1, n + b0 No c; 1 − ))
· (1 − H1,0 (N, No ))
≥ (1 − Fbeta(N + (1 − 0 )No − n + 1, n + 0 No ; 1 − ))
· (1 − H1,0 (N, No ))
.
= 1 − β̄,0 (N, No ) · (1 − H1,0 (N, No )) .
We next consider the R() term in (27). We have that
Z
R() =
(1 − Fbeta(z + 1, No − z; t))dΨ(t)
0
Z
= Ψ() −
Fbeta(z + 1, No − z; t)dΨ(t)
(29)
(30)
0
[integrating by parts] = Ψ()Fbeta(No − z, z + 1; 1 − )
Z
+
Ψ(t)beta(z + 1, No − z; t)dt.
0
Since Ψ(t) ≥ 0 forall t ∈ [0, 1], the above expression shows that R() ≥ 0 for all ∈ [0, 1], with R() being identically zero
for problems that are f.s. Considering (25), this fact permits us to conclude that
ProbN +No {GoodTrue}
= ProbN +No {S ≤ z ∩ V (θk∗ ) ≤ }
=
[from (29)]
1 − H,0 (N, No ) + R()
≥ 1 − H,0 (N, No )
≥ 1 − β̄ (N, No ) ·
· (1 − H1,0 (N, No )) ,
which proves (10) and (11). Also, we obtain that
ProbN +No {True}
.
= ProbN +No {S ≤ z}
= ProbN +No {S ≤ z ∩ V (θk∗ ) ≤ 1}
=
1 − H1,0 (N, No ) + R(1)
≥ 1 − H1,0 (N, No ),
17
which proves (9). Further, using (25), we have that
ProbN +No {BadTrue}
.
= ProbN +No {S ≤ z ∩ V (θk∗ ) > }
= ProbN +No {True}
−ProbN +No {GoodTrue}
= H,0 (N, No ) − H1,0 (N, No )
+(R(1) − R()).
If the scenario problem is f.s., then R(1) = R() = 0, hence
ProbN +No {BadTrue}
=
H,0 (N, No ) − H1,0 (N, No )
+(R(1) − R())
[if problem is f. s.] =
[using (29)] ≤
H,0 (N, No ) − H1,0 (N, No )
β̄,0 (N, No )(1 − H1,0 (N, No )).
All the above proves (13). To upper bound the probability of BadTrue in the non-fully supported case, we reason instead as
follows:
ProbN +No {BadTrue}
Z 1
=
ProbNo {S ≤ z|V (θk∗ ) = t}dFV (t)
Z 1
=
Fbeta(No − z, z + 1; 1 − t)dFV (t)
[integrand is decreasing in t]
Z 1
≤
Fbeta(No − z, z + 1; 1 − )dFV (t)
Z 1
= Fbeta(No − z, z + 1; 1 − )
dFV (t)
= Fbeta(No − z, z + 1; 1 − ) · ProbN {V (θk∗ ) > }
[from (4)]
≤ Fbeta(No − z, z + 1; 1 − ) · Fbeta(N + 1 − n, n; 1 − )
[since z = b0 No c]
≤ Fbeta((1 − 0 )No , 0 No + 1; 1 − ) · β (N ),
which proves (13).
C. Proof of Theorem 3
Let us define the event BadExitk as the one where the algorithm reaches the k-th iteration, and then exits with a “bad”
solution, i.e., with a solution θk∗ for which V (θk∗ ) > . The probability of this event is the probability that the 0 -RVO returns
false precisely k − 1 times (for this guarantees that we reach the k-th iteration), and then the event BadTrue happens at the
k-th iteration. Therefore, letting q denote the probability of BadTrue, and p denote the probability of True (events defined as
in Lemma 1) we have that
Prob×× {BadExitk }
=
(1 − p)k−1 q.
The event BadExit in which the algorithm terminates with a bad solution is the union of the non-overlapping events BadExitk ,
k = 1, 2, . . ., therefore
Prob×× {BadExit}
=
=
∞
X
Prob×× {BadExitk }
k=1
∞
X
(1 − p)k−1 q = q
k=1
(1 − p)k
k=0
N +No
=
∞
X
Prob
{BadTrue}
q
=
.
p
ProbN +No {True}
18
We now use (12) to upper bound q, and then use (9) to conclude that
Prob×× {BadExit}
Fbeta((1 − 0 )No , 0 No + 1; 1 − )β (N )
≤
,
1 − H1,0 (N, No )
which proves (14). In the fully supported case, we can instead use (13) to upper bound q, and hence conclude that
Prob×× {BadExit} ≤ β̄,0 (N, No ),
which proves (15).
Let next K denote the running time of Algorithm 2, that is the value of the iteration count when the algorithm terminates.
Since the algorithm terminates as soon as a True event happens, and since the True events are statistically independent among
iterations, we have that {K = k} has geometric probability (1 − p)k−1 p, where p is the probability of True. Therefore, the
expected value of K is 1/p ≤ 1/(1 − H1,0 (N, No )), where the inequality follows from (9), and this proves point 2 in the
theorem. Via the same reasoning, {K > k} has probability (1 − p)k , and hence we conclude that
Prob×× {K ≤ k} = 1 − (1 − p)k ≥ 1 − H1,0 (N, No )k ,
which proves the third point in the theorem.
D. Proof of Corollary 1
.
From from eq. (28) we have that, for z = b0 No c,
1 − H1,0 (N, No ) =
Z 1
Fbeta(No − z, z + 1; 1 − t)beta(n, N + 1 − n; t)dt.
(31)
0
We recall that a beta(α, β) density has mean α/(α + β), peak (mode) at (α − 1)/(α + β − 2), and variance σ 2 = αβ/((α +
β)2 (α + β + 1)). Then, we observe that Fbeta(No − z, z + 1; 1 − t) = 1 − Fbeta(z + 1, No − z; t), where Fbeta(z + 1, No − z; t)
is the cumulative distribution of a beta(z + 1, No − z) density. The peak of this density is at z/(No − 1), which tends to 0 for
No → ∞; further, the variance of this distribution goes to zero as O(No−1 ), which permits us to argue that, for large No , the
function Fbeta(No − z, z + 1; 1 − t) has an inflection point near 0 and decreases rapidly from value ' 1 to value ' 0 as t
crosses 0 . That is, as No → ∞, the function Fbeta(No − z, z + 1; 1 − t) tends to a step function which is one for t < 0 and
zero for t > 0 . Therefore, we have for the integral in (31) that
Z 0
1 − H1,0 (N, No ) →
1 · beta(n, N + 1 − n; t)dt
0
= Fbeta(n, N + 1 − n; 0 ) = 1 − β0 (N ),
which proves (16).
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
D. Aldous. Exchangeability and related topics. In P.L. Hennequin, editor, École d’Été de Probabilités de Saint-Flour XIII, volume 1117
of Lecture Notes in Math., pages 1–198. Springer, 1983.
G. Calafiore and L. El Ghaoui. Optimization Models. Cambridge University Press, 2014.
G.C. Calafiore. Random convex programs. SIAM Journal on Optimization, 20(6):3427–3464, 2010.
G.C. Calafiore and M.C. Campi. The scenario approach to robust control design. IEEE Transactions on Automatic Control, 51(5):742–753,
2006.
G.C. Calafiore and F. Dabbene. A probabilistic analytic center cutting plane method for feasibility of uncertain LMIs. Automatica,
43:2022–2033, 2007.
G.C. Calafiore, F. Dabbene, and R. Tempo. Research on probabilistic methods for control system design. Automatica, 47:1279–1293,
2011.
M.C. Campi and S. Garatti. The exact feasibility of randomized solutions of uncertain convex programs. SIAM Journal on Optimization,
19(3):1211–1230, 2008.
M. Chamanbaz, F. Dabbene, R. Tempo, V. Venkataramanan, and Q.-G. Wang. Sequential randomized algorithms for convex optimization
in the presence of uncertainty. arXiv:1304.2222v2, 2014.
S. Garatti and M.C. Campi. Modulating robustness in control design: principles and algorithms. IEEE Control Systems Magazine,
33:36–51, 2013.
19
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
Michael Grant and Stephen Boyd. CVX: Matlab software for disciplined convex programming, version 2.1. http://cvxr.com/cvx, March
2014.
J.C. Lee and D.J. Sabavala. Bayesian estimation and prediction for the beta-binomial model. Journal of Business and Economic Statistics,
5(3):357–367, 1987.
J. Luedtke and S. Ahmed. A sample approximation approach for optimization with probabilistic constraints. SIAM Journal on
Optimization, 19(2):674–699, 2008.
A. Nemirovski and A. Shapiro. Convex approximations of chance constrained programs. SIAM Journal on Optimization, 17(4):969–996,
2009.
B.K. Pagnoncelli, S. Ahmed, and A. Shapiro. Sample average approximation method for chance constrained programming: Theory and
applications. J. Optim. Theory Appl., 142:399–416, 2009.
I. Petersen and R. Tempo. Robust control of uncertain systems: Classical results and recent developments. Automatica, 50:1315–1335,
2014.
A. Prékopa. Probabilistic programming. In A. Rusczyński and A. Shapiro, editors, Stochastic Programming, volume 10 of Handbooks
in Operations Research and Management Science. Elsevier, Amsterdam, 2003.
A. Rantzer. Distributed control of positive system. arXiv:1203.0047v3, 2014.
R. Tempo, G.C. Calafiore, and F. Dabbene. Randomized Algorithms for Analysis and Control of Uncertain Systems: With Applications.
Communications and Control Engineering. Springer-Verlag; 2nd Ed., 2013.
Eric W. Weisstein. Beta binomial distribution. MathWorld–A Wolfram Web Resource.
| 3cs.SY
|
Modeling Duct Flow for Molecular Communication
Wayan Wicke∗ , Tobias Schwering∗ , Arman Ahmadzadeh∗ , Vahid Jamali∗ , Adam Noel† , and Robert Schober∗
∗ Institute
for Digital Communications, Friedrich-Alexander-Universität Erlangen-Nürnberg
of Electrical Engineering and Computer Science, University of Ottawa
arXiv:1711.01479v1 [cs.ET] 4 Nov 2017
† School
[10] and is now known as Taylor dispersion [6]. Via an effective
diffusion coefficient, the particle distribution can be derived
in the regime of large release-observation distances where
the interaction of cross-sectional diffusion and non-uniform
flow will have averaged to a uniform particle distribution in
the cross-section and a Gaussian spread along the axis. For
molecular communication, some authors adapted this model
to keep their analysis analytically tractable but the conditions
under which such simplifications are justified have not been
considered in detail [11]–[14]. In particular, it is not clear
how the channel behaves and can be described outside of
this dispersion regime. Moreover, it is not clear when the
conditions for Taylor dispersion are satisfied and what their
impact on molecular communication is. Potentially, theoretical
I. I NTRODUCTION
insight into these phenomena could be exploited for molecular
Using molecules for conveying digital messages has recently communication system design.
The focus of this paper is twofold. First, we introduce the
been recognized as a key communication strategy for nanoscale
devices such as artificial cells cooperatively fighting a disease notion of dispersion in a systematic manner which is in contrast
[1]. As the entities involved in this molecular communication to previous works. Second, we analyze and highlight the major
are in the micro- and nanoscale, diffusion plays a significant effects of the advection-diffusion particle transport on molecular
communication systems for two different regimes, namely the
role in the transport of messages [1].
However, diffusion has a limited effective range that renders dispersion regime and the flow-dominated regime, the latter of
molecular communication inefficient over extended distances. which has not been considered in the literature but is prevailing
This limitation can be overcome by exploiting fluid flow in for example in blood vessels [6].
addition to diffusion. For example, in blood vessels it is
The two key messages of this paper are as follows:
the interplay of fluid flow and diffusion that governs the
1) There is a regime where one-dimensional diffusion with
supply of oxygen from the lungs to tissues. Consequently,
constant drift can accurately capture the channel charthe molecular communication literature has considered basic
acteristics by means of an effective diffusion coefficient
models of these fundamental mechanisms [1]. In particular, the
and the cross-sectional mean velocity. In this regime, the
basic channel characteristics of diffusion in three-dimensional
initial spatial release pattern at the transmitter does not
(3D) unbounded space with uniform flow in the context of
affect the resulting particle distribution.
molecular communication have been investigated for example
2) Non-uniform flow as encountered in ducts can cause
in [2]. Such a model might be applicable when the boundaries
significant intersymbol interference (ISI), especially in a
are far from the nanonetwork. Our previous work [3] considered
flow-dominated regime. Diffusion tends to decrease ISI
a 2D environment with uniform flow to study the impact of
by enabling slowly-moving particles to move away from
bounded drift-diffusion in more detail. On the other hand,
the boundary.
1D diffusion with drift has been studied in [4], [5]. It is not
This paper is structured as follows. In Section II, we
clear when such a simplified model is applicable in molecular
introduce the system model and present preliminaries. Section
communication since flow in blood vessels or in microfluidic
III analyzes the duct channel and the different flow regimes.
channels, i.e., in ducts, especially at the microscale, is far from
Numerical results are presented in Section IV. Finally, in
uniform [6], [7]. Hence, in general, a reduction of the 3D reality
Section V, we draw some conclusions.
to a 1D model is not justified. In particular, the marginal axial
and cross-sectional particle distributions are inherently coupled,
II. S YSTEM M ODEL AND P RELIMINARIES
which makes a mathematical analysis of the channel response
A. System Model
difficult. For blood vessels simulations are shown in [8].
The notion of dispersion as the interaction of diffusion and
We consider a straight impermeable cylindrical duct of
non-uniform laminar flow was principally investigated in [9], infinite axial extent and radius a which can be described
Abstract—Active transport is sought in molecular communication to extend coverage, improve reliability, and mitigate
interference. One such active mechanism inherent to many liquid
environments is fluid flow. Flow models are often over-simplified,
e.g., assuming one-dimensional diffusion with constant drift.
However, diffusion and flow are usually encountered in threedimensional bounded environments where the flow is highly
non-uniform such as in blood vessels or microfluidic channels.
For a qualitative understanding of the relevant physical effects
inherent to these channels a systematic framework is provided
based on the Péclet number and the ratio of transmitter-receiver
distance to duct radius. We review the relevant laws of physics
and highlight when simplified models of uniform flow and
advection-only transport are applicable. For several molecular
communication setups, we highlight the effect of different flow
scenarios on the channel impulse response.
y
z
a
cr
cϕ
y
(a)
cx
a
cr
point
release
uniform
release
x
x = v(r) · t
(b)
Fig. 1. System model geometry (a) in the cross-section, and (b) along the
axis. The red shading in (a) reflects the flow velocity which is maximum in
the center and vanishes at the boundary. The corresponding parabolic shape,
x = v(r) · t, on which released particles reside when not diffusing after a
uniform release, is sketched in (b) for three different time instances. Point
and uniform transmitter release are shown as black dot and as a blue line,
respectively.
in cylindrical coordinates by the points (x, r, ϕ), where
x ∈ (−∞, ∞) is the axial position, r ∈ [0, a] is the radial
distance, and ϕ ∈ (−π, π] is the azimuth angle. The duct
is filled with a fluid of viscosity η that is subject to steady
laminar flow in the x-direction where the flow velocity v(r)
is a function of r only and is given by a parabolic function;
see Fig. 1.
We consider a transmitter (TX) that releases NTX molecules
either 1) uniformly and randomly distributed over the cross
section at x = 0, or 2) from the point (0, r0 , ϕ0 ). Moreover,
we assume a transparent receiver (RX) specified by the points
(x, r, ϕ) satisfying |x − d| ≤ cx /2, a − cr ≤ r ≤ a, |ϕ| ≤ cϕ /2,
i.e., the receiver is mounted on the duct wall with axial TX-RX
distance given by d, radial extent cr , and spanning an angle of
cϕ (see Fig. 1). We model the particle release from the TX as
instantaneous and the released particles are transported by the
fluid flow and Brownian motion. For simplicity, we assume
that particles do not interact with each other nor influence the
flow field. Because of their small size, other forces such as
gravity acting on the particles are negligible [7].
B. Preliminaries
In molecular communication, information is conveyed by
mass transfer. Mass transfer in fluids is mediated by flow
and Brownian motion which is referred to as advection and
diffusion, respectively. Thereby, mass transfer can be described
by a time-varying spatial probability density function (PDF)
p(r; t) which can be interpreted as a normalized concentration
where dV · p(r; t) gives the average fraction of particles within
the differential volume dV at r at time t. The PDF p(r; t) can
be found as the solution to the following partial differential
equation (PDE), which we will also refer to as the advectiondiffusion equation [7, Eq. (5.22)]
∂t p = D∇2 p − ∇ · pv,
(1)
∂
where ∂t p = ∂t
p denotes the partial derivative of p with respect
to t and ∇ is the Nabla operator. Moreover, D is the diffusion
coefficient and v(r) is the velocity vector at point r.
To solve (1), we need to know the velocity field v(r). In
general, the velocity field can be obtained by solving the NavierStokes equation, which provides a fundamental description of
flow by relating the velocity field to the local pressure [7,
Ch. 2]. Applied to rigid and straight channels with no-slip
boundary conditions, i.e., where the velocity at the boundary
is zero, and subject to pressure-driven flow in the steady-state,
the velocity profile is referred to as Poiseuille flow [7, Ch. 3].
Thereby, assuming a Newtonian fluid, i.e., a fluid which can
be described by the viscosity η, we obtain [7, Eq. (3.32)]
r2
v(r) = 2veff 1 − 2 ,
(2)
a
where veff is the mean velocity in the channel and is a function
of the applied pressure gradient ∂x P . In particular, veff can be
obtained as [7, Eq. (3.34)]
veff =
|∂x P |a2
.
8η
We note that the maximum velocity vmax = 2veff occurs at
the center and can be found using r = 0 in (2).
By thermodynamic reasoning and assuming spherical particles, the diffusion coefficient satisfies the Einstein relation [6,
Eq. (5.2.24)]
kB Tf
D=
,
(3)
6πηRp
where η [kg m−1 s−1 ] and Tf [K] are the viscosity and
temperature of the fluid, respectively, and Rp [m] is the particle
radius. We note that (3) is strictly valid only if the solvent can
be modeled as continuous with respect to the solute particles,
i.e., the Einstein relation gives physical insight but is not
to be understood as absolute. The diffusion coefficient is a
fundamental characteristic of the diffusion process and governs
the microscopic particle motion [6].
An important parameter is the Péclet number, which gives
an estimate of the importance of diffusion over advection. This
dimensionless number is defined as [6, Eq. (4.6.8)]
veff a
Pe =
,
(4)
D
when considering the duct radius a as the length scale of
interest. Intuitively, Pe ∈ [0, ∞) increases and decreases when
veff and D increase as the importance of particle transport by
flow and diffusion becomes more relevant, respectively.
III. A NALYSIS OF THE D UCT C HANNEL
The advection-diffusion equation (1) for the environment in
Fig. 1 simplifies to the following PDE:
∂t p = D∇2 p − v(r)∂x p,
(5)
for t > 0 because the velocity field is independent of the
axial position. At the boundary r = a, ∂r p = 0 has to
hold and p(x, r, ϕ; 0) is initially given by δ(x)/(πa2 ) and
δ(x)δ(r − r0 )δ(ϕ − ϕ0 )/r for uniform and point release,
respectively. Eq. (5) is still difficult to solve in general because
of the nonlinear velocity (2) and the inherent coupling of p
in the x- and r-directions. Interestingly, (5) can be solved in
a quasi-steady-state where enough time has passed such that
cross-sectional diffusion and the non-uniform flow profile have
averaged out. We will refer to this behavior as the dispersion
regime. Trivially, this regime includes the special case where
veff → 0, i.e., when there is pure diffusion. Another special
case where (5) can be solved is when D → 0, i.e., when we
are in the flow-dominated regime. For each of these regimes,
we seek the time-dependent observation probability
Z
Pob (t) =
p(x, r, ϕ; t) dVRX ,
(6)
expense of a larger required distance for dispersion to take
place. In fact, we can rearrange (7) as
a2 · veff
,
(11)
4d
which approximates the minimum diffusion coefficient required
for dispersion to occur. For small and large D, we can neglect
the first and second parts in the sum on the right-hand side of
(10), respectively.
By substituting (4) in (10) and minimizing with respect to D,
we obtain the minimum possible effective diffusion coefficient
veff a
VRX
Deff,min = √ ,
(12)
2 3
where VRX is the RX volume. We will also refer to Pob (t) as
√
the impulse response. The impulse response is a fundamental for D = DT = Dmin = veff a/(4 3). For a given Deff >
characteristic of the molecular communication channel and Deff,min , there are two values D < Dmin and D > Dmin ,
heavily influences for example the symbol error rate [2].
which achieve the same effective diffusion coefficient. These
values are given by
A. Dispersion Regime
Deff
1q 2
2
D± =
±
Deff − Deff,min
.
(13)
Dispersion is the result of the interaction of cross-sectional
2
2
diffusion and the non-uniform advection due to the flow profile.
The observation probability obtained by integrating (9) over
This interaction can lead to a particle distribution that is uniform
the receiver volume is given by (6)
in each cross-section, i.e., the spatial PDF can be written
cϕ
cr
c2r
d − cx /2 − veff t
as p(x, r, ϕ; t) = p(x; t)/(πa2 ). In this regime, the particle
Pob,d (t) =
· 2 − 2 × Q
distribution does depend only on the initial x-position, i.e.,
2π
a
a
2Deff t
there is no difference for point and uniform release. This
d + cx /2 − veff t
−Q
,
(14)
behavior occurs if [6, Chapter 4.6]
2Deff t
d
Pe 4 ,
(7) where Q(·) is the Gaussian Q-function.
a
It is of interest to derive the time at which Pob,d (t) attains
e.g., when the TX-RX distance is large, the diffusion coefficient its maximum. As maximizing Pob,d (t) with respect to t is
is large, or the duct radius is small. Naturally, this includes cumbersome, we resort to maximizing (9) for x = d, which
the special case of pure diffusion where flow is not present. yields
s
!
2
If (7) is satisfied, (5) can be written as the following 1D
veff
Deff
−1 + 1 + 2 d2 .
(15)
tmax = 2
advection-diffusion equation [6, Eq. (4.6.30)]
veff
Deff
∂t p = Deff ∂x2 p − veff ∂x p,
(8)
D
In this approximation, the peak height follows as pmax =
Pob,d (tmax ). We note that because of diffusion tmax < d/veff ,
with effective diffusion coefficient Deff and mean velocity veff . where d/v is the time when particles moving with the mean
eff
For an instantaneous uniform release at x = 0, the solution to velocity will reach the RX.
(8) is given by
B. Flow-dominated Regime
1
1
(x − veff t)2
p(x, r, ϕ; t) =
×√
exp −
, (9)
In this subsection, we directly determine the observation
πa2
4Deff t
4πDeff t
probability Pob,f (t) = Pob (t) by neglecting diffusion and
where 1/(πa2 ) is the cross-sectional distribution in the r- and without resorting to (6).
1) Uniform Release: First, we assume a uniform release
ϕ-directions.
Following [10, Eq. (26)], the Taylor-Aris effective diffusion at x = 0. In this case, all particles will lie on the surface
of a paraboloid that extends along the axis over time and
coefficient Deff is obtained as [6, Eq. (4.6.35)]
exhibits rotational symmetry. Thereby, the marginal distribution
1
2
Deff = D 1 +
Pe ,
(10) in the cross-section, i.e., for the r- and ϕ-coordinates, remains
48
uniform because the flow is in the x-direction. The geometric
where D is given in (3). We will also refer to D as the diffusion shape is given by x = v(r) · t, cf. Fig. 1. Hence, the time
D
coefficient in pure diffusion and to DT = 48
Pe2 as the Taylor t when the points (x, r, ϕ) lie on the paraboloid is simply
effective diffusion coefficient, i.e., we can write Deff = D +DT . given by t = x/v(r). From this, we also obtain the following
We note that in general Deff > D and moreover Deff auxiliary relationship:
r
D when D is decreased to very small values, which by (4)
x
rt (x) = a · 1 −
,
(16)
increases Pe. However, by (7), decreasing D comes at the
2veff t
which is obtained by re-arranging (2).
i.e., with probability Pob (t) and 1 − Pob (t) each of the NTX
The fraction of particles within the receiver volume is particles can be independently and cannot be observed within
determined by those particles that lie within a segment of the RX volume at time t, respectively. The reason for this
angular extent cϕ and within a ring-region with inner and outer is that for both uniform and point release the initial particle
radii ri (t) and ro (t), respectively. Following the uniform release position can be understood as independently and uniformly
at x = 0, the observation probability Pob,f (t) can consequently random distributed within the available TX area (the whole
be written as Pob,f (t) = cϕ /(2π) · (ro2 − ri2 )/a2 . Variables cross section or one point).
ri (t) and ro (t) depend on how the paraboloid intersects with
IV. N UMERICAL E VALUATION
the RX volume. In fact, there are three scenarios, which are
By using particle-based simulation, we validate our derived
shown in Fig. 1b for the x-y-plane. In the first case, the
paraboloid has not yet reached the RX, cf. the green line analytical expressions and explore those regimes for which
in Fig. 1b. The first intersection occurs at t = t1 when the mathematical analysis is difficult. Thereby, unless explicitly
paraboloid reaches the points (x = d − cx /2, r = a − cr , ϕ) stated otherwise, we employ the following physical parameter
where ϕ ∈ [−cϕ /2, cϕ /2]. In the second case, for t1 < t < t2 , values. As diffusion coefficient we choose D = 10−10 m2 /s
the paraboloid intersects with the RX at ri (t) = rt (d − cx /2) which is a reasonable estimate for small proteins [6]. Two
and ro (t) = a − cr , cf. the orange line in Fig. 1b. Time t2 is values for the duct radius a = 10 µm and a = 200 µm
characterized by the paraboloid intersecting with the RX at the are considered, which is reasonable for small capillaries [6]
points (d + cx /2, a − cr , ϕ) where ϕ ∈ [−cϕ /2, cϕ /2]. The and microfluidic ducts [7], respectively. Thereby, two TXlast case occurs for t ≥ t2 , cf. the red line in Fig. 1b. In this RX distances are considered with values d = 200 µm and
case, we have ri (t) = rt (d − cx /2) and ro (t) = rt (d + cx /2). d = 800 µm. Moreover, we choose the receiver dimensions
Now, we only require t1 and t2 in order to obtain the impulse as cx = a/2, cr = a/2, cϕ = π/2, i.e., the receiver size
scales with the duct radius. The microscopic simulation time
response. From x = v(r) · t and (2), we obtain
step is set to ∆t = 10−3 s. The fluid flow mean velocity is
d − cx /2
t1 =
(17) assumed to be veff = 1 mm s−1 , which is reasonable for small
2veff (1 − (1 − cr /a)2 )
capillaries [6].
d + cx /2
We show in Fig. 2 (adapted from [6]) the considered
t2 =
.
(18)
2veff (1 − (1 − cr /a)2 )
parameter values in terms of the Péclet number and the ratio
In summary, the impulse response in the flow-dominated of the TX-RX distance to the duct radius. In particular, we
regime following a uniform release can be written as
have shaded the two regimes for which the obtained analytical
results from Section III are expected to be applicable. Eq. (7)
0,
t ≤ t1
separates these two regimes and is shown as black line. The
cϕ
d − cx /2
cr 2
derived analytical results are valid for parameter values well
·
−
−
1
−
,
t
<
t
<
t
1
2
Pob,f (t) = 2π
2veff t
a
within
the dispersion or the flow-dominated regime. However,
c
c
ϕ · x ,
the analytical results cannot be expected to be accurate close the
t ≥ t2 .
2π 2veff t
(19) boundary set by (7). For the two duct radii a = 10 µm, 200 µm
Pob,f (t) is maximized for t = t2 . Also, at time t = t2 /α the and the two TX-RX distances d = 200 µm, 800 µm, we show
fraction α ∈ (0, 1] of Pob,f (t2 ) can be observed. The tail of the resulting four combinations of d/a and Pe as black dots in
the impulse response decays only polynomial with time, which Fig. 2. We see that the two scenarios for a = 10 µm lie close to
may give rise to significant ISI in molecular communication the boundary of both regions. These parameter values have been
chosen such that simulations can reveal the deviations from
systems.
2) Point Release: For a point release with r0 ∈ [a − cr , a] either regime. On the other hand, the scenarios for a = 200 µm
and ϕ0 ∈ [−cϕ /2, cϕ /2], i.e., when the TX coordinates are lie well within the flow-dominated regime and we expect no
within the r- and ϕ-coordinates of the RX, we observe all deviations from the developed theory. We note that changing
particles with certainty if d − cx /2 ≤ v(r0 )t ≤ d + cx /2, i.e., the duct radius a affects both d/a and Pe whereas a change in
d influences only d/a. Considering the parameter values chosen
the impulse response is given by
in this paper, from Fig. 2, we can conclude that the dispersion
v(r0 ) t − d
•
Pob,f
(t) = rect
,
(20) regime is most applicable for small microscale ducts. On the
cx
other hand, we also see that there is a large set of parameters
where rect(x) = 1 if −1/2 ≤ x ≤ 1/2 and zero otherwise. for which the flow-dominated regime is more appropriate,
When the release point is not within the r- and ϕ-coordinates especially for medium to large ducts.
In Fig. 3, we numerically evaluate the effective diffusion
of the RX then the impulse response is zero for all times.
Eqs. (19) and (20) are intuitively valid when D → 0. More coefficient as given in (10) as a function of the molecular
generally, the solution is applicable when Pe d/a [6], e.g., diffusion coefficient for different duct radii a. Thereby, we
when the duct radius is large or the TX-RX distance is small. show Deff = D and Deff = DT as the limiting expressions for
We note that (19) and (20) still give the observation large and small D, respectively. Furthermore, we distinguish
probability ∈ [0, 1] even though the flow is deterministic, between small (shown as orange line) and large (shown as
r2 [µm2 ]
log10 (d/a)
a = 10 µm
dispersion
d = 800 µm
2
d = 200 µm
0.6
flow-dominated
log10 (Pe)
a = 200 µm
Fig. 2. Sketch of regions of different transport regimes. Adapted from [6].
All four simulation scenarios are shown as black dots. For a = 10 µm, we
have Pe = 100 and d/a = 20, 80. For a = 200 µm, we have Pe = 2000
and d/a = 1, 4.
100
50
0
0
500
1,000
1,500
x [µm]
Fig. 4.
Snapshot of particle positions for a = 10 µm and at t =
0.02, 0.2, 0.8 s after uniform release at x = 0 and t = 0 shown in different
colors and starting from left to right, respectively. In total, NTX = 103 are
released.
Deff [m2 /s]
d = 800 µm. For the largest time shown, t = √
0.8 s, the red lines
show the standard deviation positions veff t ± 2Deff t from the
mean when assuming the Gaussian distribution in (9) due to
2
·veff
D > a 4d
10−5
dispersion. For small times, e.g., t = 0.02 s, the particles follow
a=
a2 ·veff
the parabolic profile (16) closely. For slightly larger times, e.g.,
D
<
20
4d
0µ
10−6
a=
Deff = DT
t = 0.2 s, the particles start to spread because of diffusion.
m
50
minimum
For large times, e.g., t = 0.8 s, particles become uniformly
µ
−7
m
a=
10
distributed along the r2 dimension within the duct due to
10
µm
dispersion. We note that for veff = 1 mm s−1 , at t = 0.2 s and
10−8
t = 0.8 s, the mean particle position has arrived at d = 200 µm
and d = 800 µm, respectively. In summary, at small times after
D
10−9
=
ff
the release the flow-dominated regime and at large times after
e
D
10−10 −11
the release the dispersion regime accurately model the actual
10
10−10 10−9
10−8
10−7
10−6
behavior.
D [m2 /s]
In Figs. 5a and 5b, we show the impulse response for a =
10
µm and a = 200 µm, respectively. In each case, we consider
Fig. 3. Effective diffusion coefficient Deff in (10) as a function of the
both d = 200 µm and d = 800 µm as well as uniform and point
molecular diffusion coefficient D.
release. For the point release, the position (0, 0.75a, 0) was
blue line) D, which for d = 800 µm satisfy and do not satisfy chosen such that particles can arrive at the receiver when not
(11), respectively. Finally, Deff,min from (12) is shown as a diffusing. We simulate the impulse responses and investigate
red dot and the value D = 10−10 m2 /s is highlighted by a which of the developed analytical models provides the best fit
black vertical line. As a increases, Deff also increases. At the in each case.
same time, as can be seen from the transition from the orange
In Fig. 5a, for comparison, both Pob,d (t) (applicable for both
to the blue line, the (soft) threshold given in (11) delineating point and uniform release) in (14) and Pob,f (t) (applicable
the validity of the dispersion regime also increases with a. only for uniform release) in (19) are shown. Thereby, for
For a = 10 µm, for D = 10−10 m2 /s, we obtain Deff = Pob,d (t), the peaks are also highlighted with large dots. When
2.1 × 10−8 m2 /s, which is a value otherwise unattainable for d = 200 µm, the simulated impulse response following a point
the diffusion coefficient of small proteins [6]. For a = 50 µm release is significantly larger than that for a simulated uniform
and a = 200 µm, for D = 10−10 m2 /s and d = 800 µm by release. In this case, both simulated impulse responses neither
(11) Deff is not meaningful anymore, cf. Fig. 2. For D > Dmin , match Pob,d (t) nor Pob,f (t). However, especially considering
an increase in D leads to an increase in dispersion. On the other the long-time behavior, e.g., for t > 0.5 s, the simulated data
hand, for D < Dmin , an increase in D leads to less dispersion points tend to be better described by Pob,d (t) than by Pob,f (t).
because the flow cannot broaden the particle distribution along For d = 800 µm, the deviations of the simulated impulse
the axis as effectively.
responses for point and uniform release from Pob,d (t) are
To gain a basic understanding of the particle evolution much smaller and the dispersion regime provides a much
towards dispersion, in Fig. 4, we show three snapshots of better fit despite the fact that (7) is not strictly satisfied, cf.
the particle positions corresponding to three different time Fig. 2. This is consistent with the green particle cloud in
instances and distinguished by different colors following a Fig. 4 which appears uniform in r2 . Comparing Pob,f (t) and
uniform release at t = 0. In particular, we plot r2 over x Pob,d (t), we see that the peak of Pob,f (t) is larger and smaller
motivated by the fact that the marginal distribution in r2 of than that of Pob,d (t) when d is small and large, respectively.
a uniform distribution within a circular disk is uniform. As For larger times, e.g., for t > 0.75 s, Pob,f (t) for d = 200 µm
a side effect, for the flow-dominated regime from (16), r2 and Pob,f (t) for d = 800 µm coincide as expected from (19)
becomes a simple linear function which for each considered which is independent of d for t > t2 . In conclusion, parameter
t is shown as a blue line. Furthermore, a = 10 µm such that values close to the boundary in Fig. 2 can still be applicable
according to Fig. 2, dispersion occurs for x in the vicinity of for the dispersion model.
20
0µ
m
=
0.5
0
0
0.5
a = 10 µm to a = 200 µm. We note that for both uniform
and point release the peak values of the impulse responses in
Fig. 5a are at least by an order of magnitude smaller than in
Fig. 5b. Moreover, for uniform and point release the simulated
impulse responses in Fig. 5a decay faster and slower from their
peak values than those in Fig. 5b, respectively.
Pob,f (t)
Pob,d (t)
sim. unif.
sim. point
peak
1
V. C ONCLUSION
d = 800 µm
d
fraction of particles
·10−2
1
time [s]
1.5
2
(a)
Pob,f (t)
• (t)
0.1 · Pob,f
sim. unif.
0.1 · sim. point
peak
=
20
0µ
m
1
0.5
d
fraction of particles
·10−1
d = 800 µm
0
0
0.5
1
time [s]
1.5
2
(b)
Fig. 5. Impulse responses for (a) a = 10 µm, and (b) a = 200 µm. Simulation
results are shown for both uniform and point release. For (b), additionally the
simulated and analytical impulse responses due to a point release are scaled
by the constant factor 0.1 for a better visualization. Simulation results are for
NTX = 106 .
•
In Fig. 5b, Pob,f (t) in (19) and Pob,f
(t) in (20) are shown.
For the former, the peak times t2 are highlighted. For both
uniform and point release, simulation results are also shown.
By Fig. 2, the dispersion approximation is not applicable in
this scenario and for clarity is not shown. Considering the
point release, the simulated curve for d = 200 µm matches the
•
rectangular shape of Pob,f
(t) in (20) reasonably well. On the
other hand, for d = 800 µm, the simulated impulse response
significantly deviates from the rectangular shape because
diffusion has had enough time to result in a spread of the
pulse. As expected from Fig. 2, we observe in general a good
agreement between Pob,f (t) in (19) and the simulation results
in the case of a uniform release, i.e., the flow-dominated regime
provides a reasonable description of the channel. Nevertheless,
for d = 200 µm, at larger times, e.g., for t > 1 s, there is
a small deviation because of residual particles close to the
RX when most particles have already passed. Comparing the
impulse responses for point release and uniform release, we
find that the tail of the impulse response strongly depends on
the initial distribution. There can be considerable ISI, especially
when a fraction of the particles is released close to the duct wall.
However, as long as the RX can be reached by the particles, an
initial release close to the center of the duct might reduce ISI.
Comparing Figs. 5a and 5b, we see that the channels behave
completely different when the duct radius is changed from
Dispersion generalizes the concept of diffusion which is
crucial for molecular communication. Thereby, non-uniform
flow can be accounted for by an effective diffusion coefficient.
This effective diffusion coefficient can be multiple orders
of magnitude larger than the molecular diffusion coefficient.
However, this description relies on a large TX-RX distance.
On the other hand, there are many practical scenarios at the
microscale that fall within the flow-dominated regime and
where dispersion is insignificant. In this regime, the initial
release pattern drastically influences ISI. For a uniform release,
the dispersion regime might be preferable over the flowdominated regime, because of the reduced ISI. For a given
duct radius either regime can be applicable depending on the
TX-RX distance and the Péclet number.
R EFERENCES
[1] N. Farsad, H. B. Yilmaz, A. Eckford, C. B. Chae, and W. Guo, “A comprehensive survey of recent advancements in molecular communication,”
IEEE Commun. Surv. Tutorials, vol. 18, no. 3, pp. 1887–1919, 2016.
[2] A. Noel, K. C. Cheung, and R. Schober, “Diffusive molecular communication with disruptive flows,” in Proc. IEEE ICC 2014, Jun. 2014, pp.
3600–3606.
[3] W. Wicke, A. Ahmadzadeh, V. Jamali, H. Unterweger, C. Alexiou, and
R. Schober, “Molecular communication using magnetic nanoparticles,”
arXiv:1704.04206, Apr. 2017.
[4] K. V. Srinivas, A. W. Eckford, and R. S. Adve, “Molecular communication
in fluid media: The additive inverse Gaussian noise channel,” IEEE Trans.
Inf. Theory, vol. 58, no. 7, pp. 4678–4692, Jul. 2012.
[5] N. R. Kim, A. W. Eckford, and C. B. Chae, “Symbol interval optimization
for molecular communication with drift,” IEEE Trans. Nanobioscience,
vol. 13, no. 3, pp. 223–229, Sep. 2014.
[6] R. F. Probstein, Physicochemical Hydrodynamics: An Introduction. John
Wiley & Sons, Feb. 2005.
[7] H. Bruus, Theoretical Microfluidics, 1st ed. Oxford University Press,
Nov. 2007.
[8] L. Felicetti, M. Femminella, and G. Reali, “Simulation of molecular signaling in blood vessels: Software design and application to atherogenesis,”
Nano Commun. Networks, vol. 4, no. 3, pp. 98–119, Sep. 2013.
[9] G. I. Taylor, “Dispersion of soluble matter in solvent flowing slowly
through a tube,” Proc. R. Soc. Lond. A, vol. 219, no. 1137, pp. 186–203,
Aug. 1953.
[10] R. Aris, “On the dispersion of a solute in a fluid flowing through a tube,”
Proc. R. Soc. Lond. A, vol. 235, no. 1200, pp. 67–77, Apr. 1956.
[11] P. He, Y. Mao, Q. Liu, P. Li, and K. Yang, “Channel modelling of
molecular communications across blood vessels and nerves,” in Proc.
IEEE ICC 2016, May 2016, pp. 1–6.
[12] Y. Chahibi, M. Pierobon, and I. F. Akyildiz, “Pharmacokinetic modeling
and biodistribution estimation through the molecular communication
paradigm,” IEEE Trans. Biomed. Eng., vol. 62, no. 10, pp. 2410–2420,
Oct. 2015.
[13] A. O. Bicen and I. F. Akyildiz, “System-theoretic analysis and leastsquares design of microfluidic channels for flow-induced molecular
communication,” IEEE Trans. Signal Process., vol. 61, no. 20, pp. 5000–
5013, Oct. 2013.
[14] Y. Sun, K. Yang, and Q. Liu, “Channel capacity modelling of blood
capillary-based molecular communication with blood flow drift,” in Proc.
ACM NanoCom 2017, 2017, pp. 19:1–19:6.
| 7cs.IT
|
New Tools and Connections for Exponential-time
Approximation
Nikhil Bansal1 , Parinya Chalermsook2 , Bundit Laekhanukit3 ,
Danupon Nanongkai4 , and Jesper Nederlof5
arXiv:1708.03515v1 [cs.DS] 11 Aug 2017
1
2
3
4
5
Eindhoven University of Technology, The Netherlands. [email protected].
Aalto University, Finland. [email protected].
Weizmann Institute of Science, Israel. [email protected].
KTH, Royal Institute of Technology, Sweden. [email protected]
Eindhoven University of Technology, The Netherlands. [email protected].
Abstract
In this paper, we develop new tools and connections for exponential time approximation. In this
setting, we are given a problem instance and a parameter α > 1, and the goal is to design an
α-approximation algorithm with the fastest possible running time. We show the following results:
An r-approximation for maximum independent set in O∗ (exp(Õ(n/r log2 r + r log2 r))) time,
An r-approximation for chromatic number in O∗ (exp(Õ(n/r log r + r log2 r))) time,
A (2 − 1/r)-approximation for minimum vertex cover in O∗ (exp(n/rΩ(r) )) time, and
A (k − 1/r)-approximation for minimum k-hypergraph vertex cover in O∗ (exp(n/(kr)Ω(kr) ))
time.
(Throughout, Õ and O∗ omit polyloglog(r) and factors polynomial in the input size, respectively.)
The best known time bounds for all problems were O∗ (2n/r ) [Bourgeois et al. 2009, 2011 &
Cygan et al. 2008]. For maximum independent set and chromatic number, these bounds were
complemented by exp(n1−o(1) /r1+o(1) ) lower bounds (under the Exponential Time Hypothesis
(ETH)) [Chalermsook et al., 2013 & Laekhanukit, 2014 (Ph.D. Thesis)]. Our results show that
the naturally-looking O∗ (2n/r ) bounds are not tight for all these problems. The key to these
algorithmic results is a sparsification procedure that reduces a problem to its bounded-degree
variant, allowing the use of better approximation algorithms for bounded degree graphs. For
obtaining the first two results, we introduce a new randomized branching rule.
Finally, we show a connection between PCP parameters and exponential-time approximation
algorithms. This connection together with our independent set algorithm refute the possibility
to overly reduce the size of Chan’s PCP [Chan, 2016]. It also implies that a (significant) improvement over our result will refute the gap-ETH conjecture [Dinur 2016 & Manurangsi and
Raghavendra, 2016].
1.
2.
3.
4.
1998 ACM Subject Classification F.2.2 Nonnumerical Algorithms and Problems
Keywords and phrases Approximations Algorithms, PCP’s, Exponential Time Algorithms
Digital Object Identifier 10.4230/LIPIcs...
1
Introduction
The Independent Set, Vertex Cover, and Coloring problems are central problems in combinatorial
optimization and have been extensively studied. Most of the classical results concern either
approximation algorithms that run in polynomial time or exact algorithms that run in
(sub)exponential-time. While these algorithms are useful in most scenarios, they lack
flexibility: Sometimes, we wish for a better approximation ratio with worse running time (e.g.
© Nikhil Bansal, Parinya Chalermsook, Bundit Laekhanukit, Danupon Nanongkai, Jesper Nederlof;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
2
New Tools and Connections for Exponential-time Approximation
computationally powerful devices), or faster algorithms with less accuracy. In particular, the
trade-off between the running time and approximation ratios are needed in these settings.
Algorithmic results on the trade-offs between approximation ratio have been studied
already in the literature in several settings, most notably in the context of Polynomialtime Approximation Schemes (PTAS). For instance, in planar graphs, Baker’s celebrated
approximation scheme for several NP-hard problems [1] gives an (1 + ε)-approximation for
e.g. Independent Set in time O∗ (exp(O(1/ε))) time. In graphs of small treewidth, Czumaj
et al. [14] give an O∗ (exp(tw/r)) time algorithm that given a graph along with a tree
decomposition of it of width at most tw, find an r-approximation for Independent Set. For
general graphs, approximation results for several problems have been studied in several works
(see e.g. [5, 6, 7, 13, 12, 11]). A basic building block that lies behind many of these results is
to partition the input instance in smaller parts in which the optimal (sub)solution can be
computed quickly (or at least faster than fully exponential-time). For example, to obtain
an r-approximation for Independent Set one may arbitrarily partition the vertex set in r
blocks and restrict attention to independent sets that are subsets of these blocks to get a
O∗ (exp(n/r)) time r-approximation algorithm.
While at first sight one might think that such a naïve algorithm should be easily improvable via more advanced techniques, it was shown in [9, 5] that almost linear-size
PCPs [15, 30] imply that r-approximating Independent Set [9] and Coloring [27] requires at
least exp(n1−o(1) /r1+o(1) ) time assuming the popular Exponential Time Hypothesis (ETH).
In the setting of the more sophisticated Baker-style approximation schemes for planar graphs,
Marx [29] showed that no (1 + ε)-approximating algorithm for planar Independent Set can
run in time O∗ (exp((1/ε)1−δ )) assuming ETH, which implies that the algorithm of Czumaj
cannot be improved to run in time O∗ (exp(tw/r1+ε )).
These lower bounds, despite being interesting, are far from tight and by no means answer
the question whether the known approximation trade-offs can be improved significantly, and
in fact in many settings we are far from understanding the full power of exponential time
approximation. For example we cannot exclude algorithms that 2-approximate k-Independent
Set1 in time O∗ (f (k)) for some function f (see e.g. [26]), nor do we know algorithms that
run asymptotically faster than the fastest exact algorithm that runs in time n0.792k time [31].
In this paper we aim to advance this understanding and study the question of designing
as fast as possible approximation algorithms for Independent Set, Coloring and Vertex Cover
in general (hyper)graphs.
Our Results.
For Independent Set our result is the following. Here we use Õ to omit log log factors in r.
I Theorem 1. There is a randomized algorithm that given an n-vertex graph G and integer r outputs an independent set that, with constant probability, has size at least α(G)/r,
where α(G) denotes the maximum independent set size of G. The algorithm runs in time
O∗ (exp(Õ(n/(r log2 r) + r log2 r))).
To prove this result we introduce a new randomized branching rule that we will now
introduce and put in context towards previous results. This follows a sparsification technique
that reduces the maximum degree to a given number. This technique was already studied
1
That is, given a graph and integer k answer YES if it has an independent set of size at least 2k and NO
if it has no independent set of size at least k.
N. Bansal et al.
before in the setting of exponential time approximation algorithms Independent Set by Cygan
et al. (see [11, paragraph ‘Search Tree Techniques’]) and Bourgeois et al. (see [7, Section
2.1]), but the authors did not obtain running times sub-exponential in n/r. Specifically, the
sparsification technique is to branch (e.g. select a vertex and try to both include v in an
independent set or discard and recurse for both possibilities) on vertices of sufficiently high
degree. The key property is that if we decide to include a vertex and the independent set, we
may discard all neighbors of v. If we generate instances by keep branching on
vertices of degree
n
at least d until the maximum degree is smaller than d, then at most n/d
/ exp(n log(d)/d)
instances are created. In each such instance, the maximum independent set can be easily
d-approximated by a greedy argument. Cygan et al. [11] note that this gives worse than
O∗ (2n/r ) running times.
Our algorithm works along this line but incorporates two (simple) ideas. Our first
observation is that instead of solving each leaf instance by greedy d-approximation algorithm,
one can use a recent Õ( logd2 d ) approximation algorithm by Bansal et al. [2] for Independent Set
on bounded degree graphs. If we choose d ≈ r log2 r, this immediately gives an improvement,
n
an r-approximation in time essentially exp( r log
r ). To improve this further we use present an
additional (more innovative) idea introducing randomization. This idea relies on the fact
that in the sparsification step we have (unexploited) slack as we aim for an approximation.2
Specifically, whenever we branch, we only consider the ‘include’ branch with probability 1/r.
This will lower the expected number of produced leaf instances in the sparsification step to
n
2n/d ≈ exp( r log
2 r ) and preserves the approximation factor with good probability.
Via fairly standard methods (see e.g. [4]) we show this also gives a faster algorithm for
coloring in the following sense:
I Theorem 2. There is a randomized algorithm that, given an n-vertex graph G and an
integer r > 0, outputs with constant probability a proper coloring of G using at most r · χ(G)
colors. The algorithm runs in time O∗ (exp(Õ(n/(r log r) + r log2 r))).
As a final indication that sparsification is a very powerful tool to obtain fast exponential
time approximation algorithms, we show that a combination of a result of Halperin [20] and
the sparsification Lemma [22] gives the following result for the Vertex Cover problem in
hypergraphs with edges of size at most k (or Set Cover problem with frequency at most k).
I Theorem 3. For every k, there is an r0 := r(k) such that for every r ≥ r0 there is an
O∗ (exp( (kr)nΩ(kr) )) time (k − 1r )-approximation algorithm for the Vertex Cover problem in
hypergraphs with edges of size at most k.
n
Note that for k = 2 (e.g. vertex cover in graphs), this gives an O∗ (exp( rΩ(r)
)) running
time, which gives an exponential improvement (in the denominator of the exponent) upon
the (2 − 1/r) approximation by Bonnet et al. [7] that runs in time O∗ (2n/r ). It was recently
brought to our attention that Williams and Yu [32] independently have unpublished results
for (hypergraph) vertex cover and independent set using sparsification techniques similar to
ours.
Connections to PCP parameters The question of approximating the maximum independent
set problem in sub-exponential time has close connections to the trade-off between three
important parameters of PCPs: size, gap and free-bit. We discuss the implications of our
algorithmic results in terms of these PCP parameters.
2
This observation was already made by Bourgeois et al. [7], but we exploit it in a new way.
3
4
New Tools and Connections for Exponential-time Approximation
Roughly speaking, the gap parameter is the ratio between completeness and soundness,
while the freeness parameter is the number of distinct proofs that would cause the verifier to
accept; the free-bit is simply a logarithm of freeness. For convenience, we will continue our
discussions in terms of freeness, instead of freebit.
Freebit v.s. gap: The dependency between freeness and gap has played important
role in hardness of approximation. Most notably, the existence of PCPs with freeness
g o(1) where g is a gap parameter is “equivalent” to n1−o(1) hardness of approximating
maximum independent set [21, 3]; this result is a building block for proving other hardness
of approximation for many other combinatorial problems, e.g., coloring [19], disjoint paths,
induced matching, cycle packing, and pricing. So it is fair to say that this PCP parameter
trade-off captures the approximability of many natural combinatorial problems.
Better parameter trade-off implies stronger hardness results. The existence of a PCP
with arbitrarily large gap and freeness 1 (lowest possible) is in fact equivalent to (2 − )
inapproximability for Vertex Cover. The best known trade-off is due to Chan [10]: For
any g > 0, there is a polynomial-sized PCP with gap g and freeness O(log g), yielding the
best known NP-hardness of approximating maximum independent set in sparse graphs,
i.e. Ω(d/ log4 d) NP-hardness of approximating maximum independent set in degree-d
graphs. 3
Size, freebit, and gap: When a polynomial-time approximation algorithm is the main
concern, polynomial size PCPs are the only thing that matter. But when it comes to
exponential time approximability, another important parameter, size of the PCPs, has
come into play. The trade-off between size, freebit, and gap tightly captures the (sub)exponential time approximability of many combinatorial problems. For instance, for
√ any
O( log g)
1+o(1)
g > 0, Moshkovitz and Raz [30] constructs a PCP of size n
and freeness 2
;
1−o(1)
/r 1+o(1)
this implies that r-approximating Independent Set requires time 2n
[9].
Our exponential-time approximation result for Independent Set implies the following
tradeoff results.
I Corollary 4. Unless ETH breaks, a freebit PCP with gap g, freeness F and size S must
satisfy F · S = Ω(n log2 g).
In particular, this implies that (i) Chan’s PCP cannot be made smaller size than o(n log g),
unless ETH breaks, and (ii) in light of the equivalence between gap-amplifying freebit PCPs
with freeness 1 and (2 − ) approximation for Vertex Cover, our result shows that such a PCP
must have size at least Ω(n log2 g). We remark that no such trade-off results are known for
polynomial-sized PCPs. To our knowledge, this is the first result of its kind.
Further related results The best known results for Independent Set in the polynomial-time
log n)2
regime are an O( n(log
)-approximation [17], and the hardness of n/exp(O(log3/4+o(1) n))
log3 n
(which also √
holds for Coloring) [24]. For Vertex Cover, the best known hardness of approximation is ( 2 − o(1)) NP-hardness [23] and (2 − ) hardness assuming the unique games
conjecture [25]. All three problems (Independent Set, Coloring, and Vertex Cover) do not
admit exact algorithms that run in time 2o(n) , unless ETH fails. Besides the aforementioned
works [7, 11] sparsification techniques for exponential time approximation were studied by
Bonnet and Paschos in [6], but mainly hardness results were obtained.
3
Roughly speaking, the existence of a PCP with freeness F (g) (where g is a gap) implies Ω( F (d)dlog3 d )
hardness of approximating independent set in degree-d graphs.
N. Bansal et al.
2
Preliminaries
We first formally define the three problems that we consider in this paper. Independent Set:
Given a graph G = (V, E), we say that J ⊆ V is an independent set if there is no edge
with both endpoints in J. The goal of Independent Set is to output an independent set J of
maximum cardinality. Denote by α(G), the cardinality of the maximum independent set.
Vertex Cover: Given a graph G = (V, E), we say that J ⊆ V is a vertex cover of G if every
edge is incident to at least one vertex in J. The goal of Vertex Cover is to output a vertex
cover of minimum size. A generalization of vertex cover, called k-Hypergraph Vertex Cover
k-Vertex Cover, is defined as follows. Given a hypergraph G = (V, E) where each hyperedge
h ∈ E has cardinality at most k, the goal is to find a collection of vertices J ⊆ V such that
each hyperedge is incident to at least one vertex in J, while minimizing |J|. The degree
∆(H) of hypergraph H is the maximum frequency of an element. Coloring: Given a graph
G = (V, E), a proper k-coloring of G is a function f : V → [k] such that f (u) 6= f (v) for all
uv ∈ E. The goal of Coloring is to compute a minimum integer k > 0 such that G admits a
(proper) k-coloring; this number is referred to as the chromatic number, denote χ(G).
For a graph G = (V, E), NG (v) denotes the set of neighbors of v and dG (v) denotes
|NG (v)|. If X ⊆ V we let G[X] denote the graph (X, E ∩ (X × X)) i.e. the subgraph of
G induced by X We use exp(x) to denote 2x in order to avoid superscripts. We use the
O∗ (·)-notation to suppress factors polynomial in the input size. We use Õ and Ω̃ to suppress
factors polyloglog in r in respectively upper and lower bounds and write Θ̃ for all functions
that are in both Õ and Ω̃.
3
3.1
Faster Approximation via Randomized Branching and Sparsification
Maximum Independent Set
In this section we prove Theorem 1. Below is our key lemma.
I Lemma 5. Suppose there is an approximation algorithm dIS(G, r) that runs in time
T (n, r) and outputs an Independent Set of G of size α(G)/r if G has maximum degree
d(r),
d(r) ≥ 2r). Then there
(where
is an algorithm IS(G, r) running in expected time
O∗ exp
n
d(r)
log(4d(r)/r) T (n, r) that outputs an independent set of expected size α(G)/r.
Proof. Consider the following algorithm.
Algorithm IS(G = (V, E), r)
1: if ∃v ∈ V : dG (v) ≥ d(r) then
2:
Draw a random Boolean variable b such that Pr[b = true] = 1/r.
3:
if b = true then
4:
return the largest of IS(G[V \ v]) and IS(G[V \ N (v)]) ∪ {v}.
5:
else
6:
return IS(G[V \ v]).
7: else
8:
return dIS(G).
Figure 1 Approximation algorithm for Independent Set using an approximation algorithm dIS
that works in bounded degree graphs.
5
6
New Tools and Connections for Exponential-time Approximation
For convenience, let us fix r and d := d(r). We start by analyzing the expected running
time of this algorithm. Per recursive call the algorithm clearly uses O∗ (T (n, r)) time. It
remains to bound the number of recursive calls R(n) made by IS(G, r) when G has n vertices.
We will bound R(n) ≤ 2λn for λ = log(4d/r)/d by induction on n. Note that here λ is chosen
such that
exp(−λ · d) = r/(4d) ≤
r log(4d/r)
,
2d
(1)
where we use d/r ≥ 2 for the inequality. For the base case of the induction, note that if the
condition at Line 1 does not hold, the algorithm does not use any recursive calls and the
statement is trivial as λ is clearly positive. For the inductive step, we see that
R(n) ≤ R(n − 1) + Pr[b = true] · R(n − d)
= R(n − 1) + R(n − d)/r
= exp(λ(n − 1)) + exp(λ(n − d))/r
= exp(λn) (exp(−λ) + exp(−λd)/r)
Using exp(−x) ≤ 1 − x/2 for x ∈ [0, 1]
≤ exp(λn) (1 − λ/2 + exp(−λd)/r)
Using exp(−λ · d(r)) ≤ λr/2 from (1)
≤ exp(λn).
We continue by analyzing the output of the algorithm. It clearly returns a valid independent
set as all neighbors of v are discarded when v is included in Line 4 and an independent set
is returned at Line 8. It remains to show E[|IS(G, r)|] ≥ α(G)/r which we do by induction
on n. In the base case in which no recursive call is made, note that on Line 8 we indeed
obtain an r-approximation as G has maximum degree d(r). For the inductive case, let X be
a maximum independent set of G and let v be the vertex as picked on Line 1. We distinguish
two cases based on whether v ∈ X. If v ∈
/ X, then α(G) = α(G[V \ v]) and the inductive step
follows as E[|IS(G[V \ v], r)|] ≥ α(G)/r by the induction hypothesis. Otherwise, if v ∈ X,
then E[|IS(G, r)|] is at least
Pr[b = false] · E[|IS(G[V \ {v}], r)|] + Pr[b = true] · E[|IS(G[N \ N (v)], r)| + 1]
α(G) − 1
1
1 α(G) − 1
+r
+1
≥ 1− r
r
r
α(G) − 1 1
=
+ r = α(G)/r,
r
as required. Here the first inequality uses the induction hypothesis twice.
J
We will invoke the above lemma by using the algorithm dIS(G) by Bansal et al. [2]
implied by the following theorem:
I Theorem 6 ([2], Theorem 1.3). There is an Õ(d/ log2 d) approximation algorithm dIS(G)
for Independent Set on graphs of maximum degree d running in time O∗ (exp(O(d))).
Proof of Theorem 1. We may apply Lemma 5 with r/3 and, by virtue of Theorem 6,
with d(r/3) = Θ̃(r log2 r), and T (n, r) = O∗ (exp(Õ(r log2 r))).
We obtain an
2
2
∗
O (exp(Õ(n/r log r + r log r))) expected time algorithm that outputs an independent
set of expected size 2α(G)/r.
Since the size of the output is upper bounded by α(G) we obtain an independent set of
size at least α(G)/r with probability at least 1/(3r), and we may boost this probability to
3/4 by O(r) repetitions.
N. Bansal et al.
7
By Markov’s inequality these repetitions together run in O∗ (exp(Õ(n/r log2 r + r log2 r)))
time with probability 3/4. The theorem statement follows by a union bound as these O(r)
repetitions run in the claimed running time and simultaneously some repetition finds an
independent set of size at least α(G)/r, with probability at least 1/2.
J
A deterministic algorithm: Additionally, we also show a deterministic r-approximation
algorithm that runs in time exp(Õ(n/r log r)). The algorithm utilizes Feige’s algorithm [17]
as a blackbox, and is deferred to Appendix A.
3.2
Graph Coloring
Now we use the approximation algorithm for Independent Set as a subroutine for an approximation algorithm for Coloring to prove Theorem 2 as follows:
Proof of Theorem 2. The algorithm combines the approximation algorithm IS from Section 3.1 for Independent Set with an exact algorithm optcol for Coloring (see, e.g., [4]) as
follows:
Algorithm CHR(G = (V, E), r)
1: Let n = |V |, c = 0.
2: while |V | ≥ n/(r log r) do
3:
c ← c + 1.
4:
Cc ← IS(G[V ], r/ ln(r log r)).
5:
V ← V \ Cc .
6: Let (Cc+1 , . . . , C` ) ← optcol(G[V ]) be some optimum coloring of the remaining
graph G(V ).
7: return (C1 , . . . , C` ).
Figure 2 Approximation algorithm for the chromatic number.
We claim that CHR(G, r) returns with high probability a proper coloring of G using
` ≤ (r + 2) · χ(G) colors. To prove the theorem, we invoke CHR(G, r − 2) which has
the same asymptotic running time. First, note that in each iteration of the while loop
log r)
(Line 2 of Algorithm 2), |V | is decreased by a multiplicative factor of at most 1 − ln(r
r·χ(G)
because G[V ] must have an independent set of size at least n/χ(G) and therefore |Cc | ≥
ln(r log r)n/(r · χ(G)). Before the last iteration, we have |V | ≥ n/(r ln r). Thus, the number
` of iterations must satisfy
1/(r log r) ≤
`−1
ln(r log r)
ln(r log r)(` − 1)
1−
≤ exp −
.
r · χ(G)
r · χ(G)
This implies that (` − 1) ≤ r · χ(G). Consequently, the number of colors used in the first
phase of the algorithm (Line 1 to Line 5) is c ≤ rχ(G) + 1. The claimed upper bound on `
follows because the number of colors used for G[V ] in the second phase (Line 6) is clearly
upper bounded by χ(G).
To upper bound the running time, note that Line 4 runs in time
n ln(r log r)
n
2
2
exp Õ
+
r
log
r
=
exp
Õ
+
r
lg
r
,
r log r
r log2 (r/ ln(r log r))
8
New Tools and Connections for Exponential-time Approximation
and implementing optcol(G = (V, E)) by using the O∗ (2|V | ) time algorithm from [4], Line 6
also takes O∗ (2n/(r log r) ) time and the running time follows.
J
3.3
Vertex Cover and Hypergraph Vertex Cover
In this section, we show an application of the sparsification technique to Vertex Cover to
obtain Theorem 3. Here the sparsification step is not applied explicitly. Instead, we utilize
the sparsification Lemma of Impagliazzo et al. [22] as a blackbox. Subsequently, we solve
each low-degree instance by using an algorithm of Halperin [20]. The sparsification lemma
due to Impagliazzo et al. [22], shows that an instance of the k-Hypergraph Vertex Cover
problem can be reduced to a (sub-)exponential number of low-degree instances.4
I Lemma 7 (Sparsification Lemma, [22, 8]). There is an algorithm that, given a hypergraph
H = (V, E) with edges of size at most k ≥ 2, a real number ε > 0, produces set systems
H1 = (V, E1 ), . . . , H` = (V, E` ) with edges of size at most k in O∗ (`) time such that
1. every subset X ⊆ V is a vertex cover of H if and only if X is a vertex cover of Hi for
some i,
2. for every i = 1, . . . , `, the degree ∆(Hi ) is at most (k/ε)3k ,
3. ` is at most exp(εn).
The next tool is an approximation algorithm for the k-Hypergraph Vertex Cover problem
when the input graph has low degree due to Halperin [20].
ln ln ∆
I Theorem 8 ([20]). There is a polynomial time k − (1 − o(1)) k(k−1)
-approximation
ln ∆
algorithm for the vertex cover problem in hypergraphs with edges of size at most k in which
every element has degree at most ∆, for large enough ∆ := ∆(k).
Now we complete the proof of the theorem by applying Lemma 7 with parameter
ε = k/(kr)kr . The number
of low-degree instances Hi produced by Lemma 7 is at most
exp(εn) = exp O
Note that
k
(kr)kr
2
. Each graph Hi has degree at most ∆(Hi ) ≤ (k/ε)3k = (kr)3k r .
ln ln ∆(Hi )
ln(3k 2 r ln(kr))
1
≥
≥ 2 .
ln ∆(Hi )
3k 2 r ln(kr)
3k r
Plugging this value of ∆(Hi ), Halperin’s algorithm gives the approximation factor of
k−
k(k − 1) ln ln ∆
1
≤k− .
ln ∆
6r
Thus this gives an k − 1/(6r)-approximation running in time O∗ (exp(nk/(kr)kr )) which
translates to an k − 1/r-approximation running in time O∗ (exp(nk/(kr/6)kr/6 )).
4
PCP Parameters and Exponential-time approximation hardness
Exponential-time approximation has connections to the trade-off questions between three
parameters of PCPs: size, freebit, and gap. To formally quantify this connection, we define
new terms, formally illustrating the ideas that have been already around in the literature.
4
The original formulation is for the Set Cover problem and the most popular formulation is for CNF-SAT
problem, but they are all equivalent by direct transformation.
N. Bansal et al.
9
We define a class of languages FGPCP which stands for Freebit and Gap-amplifiable PCP. Let
g be a positive real, and S, F be non-decreasing functions. A language L is in FGPCPc (S, F )
if there is a constant g0 > 1 such that, for all constants g ≥ g0 , there is a verifier Vg that, on
input x ∈ {0, 1}n , has access to a proof π : |π| = O(S(n, g)) and satisfies the properties:
The verifier runs in 2o(n) time.
If x ∈ L, then there is a proof π such that Vgπ (x) accepts with probability ≥ c.
If x 6∈ L, then for any proof π, Vgπ (x) accepts with probability ≤ c/g.
For each x and each random string r, the verifier has ≤ F (g) accepting configurations.
The parameters g, S and log F are referred to as gap, size and freebit of the PCPs
respectively. For convenience, we call F (g) the freeness of the PCP. An intuitive way to view
this PCP is as a class of PCPs parameterized by gap g. An interesting question in the PCPs
and hardness of approximation literature has been to find the smallest functions S and F .
I Theorem 9. If SAT ∈ FGPCPδ (S, F ) for some function S(n, g) that is at least linearly
growing in n, then for any constant r, r-approximating Independent Set, in input graph G,
−1
cannot be done in time 2o(S (|V (G)|,r)/rF (r)) unless ETH fails. (we think of r as a fixed
number, and therefore S(n, r) should be seen as a function on a single variable n.)
We prove the theorem later in this section.
I Corollary 10. Assuming that SAT has no 2o(n) -time randomized algorithm and that SAT ∈
log2 g
FGPCPδ (S, F ), then it must be the case that S(n, g) · F (g) = Ω(n · poly(log
log g) ).
Proof. Otherwise, S −1 (|V (G)|, r) = o(|V (G)| ·
o(|V
imply that there is no 2
approximation algorithm.
poly(log log r)
(G)|· r log2 r )
F (r)poly(log log r)
),
log2 r
and the Theorem 9 would
, contradicting the existence of our Independent Set
J
Now let us phrase the known PCPs in our framework of FGPCP. Chan’s PCPs [10] can
be stated that SAT ∈ FGPCP1−o(1) (poly, O(log g)). Applying our results, this means that if
one wants to keep the same freebit parameters given by Chan’s PCPs, then the size must be
at least Ω(n log g). Another interesting consequence is a connection between Vertex Cover
and Freebit PCPs in the polynomial time setting [3].
I Theorem 11 ([3]). Vertex Cover is (2 − ) hard to approximate if and only if SAT ∈
FGPCP1/2− (poly, 1).
The intended PCPs in Theorem 11 have arbitrary small soundness while the freeness
remains 1. Our Corollary 10 implies that such a PCP must have size at least Ω(n log2 g).
4.1
Proof of Theorem 9
Step 1: Creating a hard CSP We will need the following lemma that creates a “hard” CSP
from FGPCP. This CSP will be used later to construct a hard instance of Independent Set.
I Lemma 12. If SAT ∈ FGPCPδ (S, F ), then, for any g > 1, there is a randomized reduction
from an n-variable SAT φ to a CSP φ0 having the following properties (w.h.p.):
The
The
The
If φ
number of variables of φ0 is ≤ S(n).
number of clauses of φ0 is ≤ 10S(n)g/δ.
freeness of φ0 is ≤ F (g).
is satisfiable, then val(φ0 ) ≥ δ/2. Otherwise, val(φ0 ) ≤ 6δ/g.
10
New Tools and Connections for Exponential-time Approximation
Proof. Let g be any number and Vg be the corresponding verifier. On input φ, we create
a CSP φ0 as follows. For each proof bit Πi , we have variable xi . The set of variables is
X = {x1 , . . . , xS(n) }. We perform M = 10dS(n)g/δe iterations. In iteration j, the verifier
picks a random string rj and create a predicate Pj (xb1 , . . . , xbq ), where b1 , . . . , bq are the
proof bits read by the verifier VgΠ on random string rj . This predicate is true on assignment
γ if and only if the verifier accepts the local assignment where Πbi = γ(xi ) for all i ∈ [q].
∗
First, assume that φ is satisfiable. Then there is a proof Π∗ such that the verifier V Π (φ)
accepts with probability δ. Let γ : X → {0, 1} be an assignment that agrees with the proof
Π∗ . So γ satisfies each predicate Pj with probability δ, and therefore, the expected number
of satisfied predicates is δM . By Chernoff’s bound, the probability that γ satisfies less than
δM
−δM/8
≤ 2−n .
2 predicates is at most 2
Next, assume that φ is not satisfiable. For each assignment γ : X → {0, 1}, the fraction
of random strings satisfied by the corresponding proof Πγ is at most δ/g. When we pick
a random string rj , the probability that V Πγ (φ, rj ) accepts is then at most δ/g. So, over
all the choices of M strings, the expected number of satisfied predicates is δM/g ≥ 10S(n).
By Chernoff’s bound, the probability that γ satisfies more than δM/g predicates is at most
2−10S(n) . By union bound over all possible proofs of length S(n) (there are 2S(n) such proofs),
the probability that there is such a γ is at most 2S(n) 2−10S(n) ≤ 2−S(n) .
J
Step 2: FGLSS reduction The FGLSS reduction is a standard reduction from CSP to
Independent Set introduced by Feige et al. [18]. The reduction simply lists all possible
configurations (partial assignment) for each clause as vertices and adding edges if there
is a conflict between two configuration. In more detail, for each predicate Pi and each
partial assignment γ such that Pi (γ) is true, we have a vertex v(i, γ). For each pair of
vertices v(i, γ)v(i0 , γ 0 ) such that there is a variable appearing in both Pi and Pi0 for which
γ(xj ) 6= γ 0 (xj ), we have an edge between v(i, γ) and v(i0 , γ 0 ).
I Lemma 13 (FGLSS Reduction [18]). There is an algorithm that, given an input CSP φ
with m clauses, n variables, and freeness F , produces a graph G = (V, E) such that (i)
|V (G)| ≤ mF and (ii) α(G) = val(φ)m, where val(φ) denotes the maximum number of
predicates of φ that can be satisfied by an assignment.
Combining everything Assume that SAT ∈ FGPCPδ (S, F ). Let g > 0 be a constant and Vg
be the verifier of SAT that gives the gap of g. By invoking Lemma 12, we have a CSP φ1 with
S(n, g) variables and 100S(n, g)g/δ clauses. Moreover, the freeness and gap of φ1 are F (g)
and g respectively. Applying the FGLSS reduction, we have a graph G with N = |V (G)| =
100S(n, g)F (g)g/δ = O(S(n, g)F (g)g). Now assume that we have an algorithm A that gives
o(S −1 (N,g))
a g approximation in time 2 gF (g) . Notice that S −1 (N, g) ≤ O(ngF (g)) and therefore
algorithm A distinguishes between Yes- and No-instance in time 2o(n) , a contradiction.
Hardness under Gap-ETH: Dinur [16] and Manurangsi and Raghavendra [28] made a
conjecture that SAT does not admit an approximation scheme that runs in 2o(n) time. We
c
observe a Gap-ETH hardness of r-approximating Independent Set in time 2n/r for some
constant c. The proof uses a standard amplification technique and is deferred to Appendix B.
5
Further Research
Our work leaves ample opportunity for exciting research. An obvious open question is to
derandomize our branching, e.g. whether Theorem 1 can be proved without randomized
N. Bansal et al.
algorithms. While the probabilistic approximation guarantee can be easily derandomized
by using a random partition of the vertex set in r parts or splitters, it seems harder to
strengthen the expected running time bound to a worst-case running time bound.
Can we improve the running times of the other algorithms mentioned in the introduction
that use the partition argument, possibly using the randomized branching strategy? Specifically, can we (1 + ε)-approximate Independent Set on planar graphs in time O∗ (2(1/ε)/ log(1/ε) ),
or r-approximate Independent Set in time O∗ (2tw/r log r )? As mentioned in the introduction,
a result of Marx [29] still leaves room for such lower order improvements. Another open
question in this category is how fast we can r-approximate k-Independent Set, where the goal
is to find an independent st of size of k. For example no O(nk/f (r) ) time algorithm is known,
where f (r) is a non-trivial function of r, that distinguishes graphs G with α(G) ≥ 2k from
graphs with α(G) ≤ k. The partition argument gives only a running time of (n/r)0.792k , and
no strong lower bounds are known for this problem. Finally, a big open question in the area
is to find or exclude a (2 − ε)-approximation for Vertex Cover in graphs in subexponential
time for some fixed constant ε > 0.
Acknowledgment NB is supported by a NWO Vidi grant 639.022.211 and ERC consolidator
grant 617951. BL is supported by ISF Grant No. 621/12 and I-CORE Grant No. 4/11. DN
is supported by the European Research Council (ERC) under the European Union’s Horizon
2020 research and innovation programme under grant agreement No 715672 and the Swedish
Research Council (Reg. No. 2015-04659). JN is supported by NWO Veni grant 639.021.438.
References
1
2
3
4
5
6
7
8
9
10
Brenda S. Baker. Approximation algorithms for np-complete problems on planar graphs.
J. ACM, 41(1):153–180, 1994.
Nikhil Bansal, Anupam Gupta, and Guru Guruganesh. On the Lovász Theta Function for
Independent Sets in Sparse Graphs. In Symposium on Theory of Computing, STOC, pages
193–200, 2015.
Mihir Bellare, Oded Goldreich, and Madhu Sudan. Free bits, pcps, and nonapproximabilitytowards tight results. SIAM J. Comput., 27(3):804–915, 1998.
Andreas Björklund, Thore Husfeldt, and Mikko Koivisto. Set partitioning via inclusionexclusion. SIAM J. Comput., 39(2):546–563, 2009.
Édouard Bonnet, Michael Lampis, and Vangelis Th. Paschos. Time-approximation tradeoffs for inapproximable problems. In Symposium on Theoretical Aspects of Computer Science, STACS, pages 22:1–22:14, 2016.
Édouard Bonnet and Vangelis Th. Paschos. Sparsification and subexponential approximation. Acta Informatica, pages 1–15, 2016.
Nicolas Bourgeois, Bruno Escoffier, and Vangelis Th. Paschos. Approximation of max independent set, min vertex cover and related problems by moderately exponential algorithms.
Discrete Applied Mathematics, 159(17):1954 – 1970, 2011.
Chris Calabro, Russell Impagliazzo, and Ramamohan Paturi. A duality between clause
width and clause density for SAT. In Conference on Computational Complexity (CCC),
pages 252–260, 2006.
Parinya Chalermsook, Bundit Laekhanukit, and Danupon Nanongkai. Independent set, induced matching, and pricing: Connections and tight (subexponential time) approximation
hardnesses. In Foundations of Computer Science, FOCS, pages 370–379, 2013.
Siu On Chan. Approximation resistance from pairwise-independent subgroups. J. ACM,
63(3):27:1–27:32, 2016.
11
12
New Tools and Connections for Exponential-time Approximation
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
Marek Cygan, Lukasz Kowalik, Marcin Pilipczuk, and Mateusz Wykurz. Exponential-time
approximation of hard problems. CoRR, abs/0810.4934, 2008.
Marek Cygan, Lukasz Kowalik, and Mateusz Wykurz. Exponential-time approximation of
weighted set cover. Inf. Process. Lett., 109(16):957–961, 2009.
Marek Cygan and Marcin Pilipczuk. Exact and approximate bandwidth. Theor. Comput.
Sci., 411(40-42):3701–3713, 2010.
Artur Czumaj, Magnús M. Halldórsson, Andrzej Lingas, and Johan Nilsson. Approximation algorithms for optimization problems in graphs with superlogarithmic treewidth. Inf.
Process. Lett., 94(2):49–53, 2005.
Irit Dinur. The PCP theorem by gap amplification. J. ACM, 54(3):12, 2007.
Irit Dinur. Mildly exponential reduction from gap 3sat to polynomial-gap label-cover. Electronic Colloquium on Computational Complexity (ECCC), 23:128, 2016.
Uriel Feige. Approximating maximum clique by removing subgraphs. SIAM J. Discrete
Math., 18(2):219–225, 2004.
Uriel Feige, Shafi Goldwasser, László Lovász, Shmuel Safra, and Mario Szegedy. Interactive
proofs and the hardness of approximating cliques. J. ACM, 43(2):268–292, 1996.
Uriel Feige and Joe Kilian. Zero knowledge and the chromatic number. J. Comput. Syst.
Sci., 57(2):187–199, 1998.
Eran Halperin. Improved approximation algorithms for the vertex cover problem in graphs
and hypergraphs. SIAM J. Comput., 31(5):1608–1623, 2002.
Johan Håstad. Clique is hard to approximate within n1-epsilon . In 37th Annual Symposium
on Foundations of Computer Science, FOCS, pages 627–636, 1996.
Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly
exponential complexity? J. Comput. Syst. Sci., 63(4):512–530, 2001.
Subhash Khot, Dor Minzer, and Muli Safra. On independent sets, 2-to-2 games and grassmann graphs. Electronic Colloquium on Computational Complexity (ECCC), 23:124, 2016.
Subhash Khot and Ashok Kumar Ponnuswami. Better inapproximability results for maxclique, chromatic number and min-3lin-deletion. In Automata, Languages and Programming, International Colloquium, (ICALP), pages 226–237, 2006.
Subhash Khot and Oded Regev. Vertex cover might be hard to approximate to within
2-epsilon. J. Comput. Syst. Sci., 74(3):335–349, 2008.
Subhash Khot and Igor Shinkar. On hardness of approximating the parameterized clique
problem. In Innovations in Theoretical Computer Science (ITCS), pages 37–45, New York,
NY, USA, 2016. ACM. doi:10.1145/2840728.2840733.
Bundit Laekhanukit. Inapproximability of Combinatorial Problems in Subexponential-Time.
PhD thesis, McGill University, 2014.
Pasin Manurangsi and Prasad Raghavendra. A birthday repetition theorem and complexity
of approximating dense csps. CoRR, abs/1607.02986, 2016.
Dániel Marx. On the optimality of planar and geometric approximation schemes. In
Foundations of Computer Science (FOCS), pages 338–348, 2007.
Dana Moshkovitz and Ran Raz. Two-query PCP with subconstant error. J. ACM,
57(5):29:1–29:29, 2010.
Jaroslav Nešetřil and Svatopluk Poljak. On the complexity of the subgraph problem. Commentationes Mathematicae Universitatis Carolinae, 026(2):415–419, 1985.
Ryan Williams and Huacheng Yu. Personal communication.
N. Bansal et al.
A
13
A Deterministic Algorithm for Independent Set
In this section, we give a deterministic r-approximation algorithm that runs in time
2O(n/r log r) . This algorithm is a simple consequence of Feige’s algorithm [17], that we
restate below in a slightly different form.
α(G)
|V (G)|
n
Ω(t · logk ( kt
)) in
I Theorem 14 ([17]). Let G be a graph with independence ratio
= 1/k. Then, for
any parameter t, one can find an independent set of size
time poly(n)k O(t) .
Now, our algorithm proceeds as follows.
If α(G) < n/ log2 r, we can enumerate
sets of size n/(r log2 r) (this is an
n
all independent
n
2
2
r-approximation) in time n/(r log2 r) ≤ (er log r) r log r ≤ 2O(n/(r log r)) .
Otherwise, the independence ratio is at least 1/k where k = log2 r. We choose t =
n/(r log r), so Feige’s algorithm finds an independent set of size at least
n
n
Ω t · logk ( ) = Ω
· logk (r log r) = Ω(n/(r log log r))
kt
r log r
The running time is
k O(t) = 2O(
n(log log r)
)
r log r
If we redefine r0 = r log log r, then the algorithm is an r0 -approximation algorithm that
0 2
0
0
runs in time 2O(n(log log r ) /r log r ) .
B
Gap-ETH hardness of Independent Set (sketch)
We now sketch the proof. We are given an n-variable 3-CNF-SAT formula φ with perfect
completeness and soundness 1 − for some > 0. We first perform standard amplification
and sparsification to get φ0 with gap parameter g, the number of clauses is ng, and freeness
is g O(1/) . Now, we perform FGLSS reduction to get a graph G such that |V (G)| = ng O(1/) .
O(1/)
)
Therefore, g-approximation in time 2o(|V (G)|/g
would lead to an algorithm that satisfies
more than (1 − ) fraction of clauses in 3-CNF-SAT formula in time 2o(n) . In other words, any
c
2n/r -time algorithm that r-approximates Independent Set can be turned into a (1 + O(1/c))approximation algorithm for approximating 3-CNF-SAT in sub-exponential time.
| 8cs.DS
|
arXiv:1610.06783v1 [math.GR] 21 Oct 2016
LES GROUPES DE MATHIEU SONT-ILS AUSSI
SPORADIQUES QU’IL Y PARAÎT ?
LABIB HADDAD
À Yves Sureau, guide avisé, de haute
montagne, au pays des hypergroupes.
Abstract. This is a plea for the reopening of the building site for the
classification of finite simple groups in order to include the finite simple
hypergroups.
Hypergroups were first introduced by Frédéric Marty, in 1934, at a
Congress in Stockholm, not to be confused with a later and quite different notion to which the same name was given, inopportunely.
I am well aware that, probably, quite a few mathematicians must have
already felt uncomfortable about the presence of the so-called sporadic
simple groups in the large tableau of the classification of finite simple
groups, and might have wrote about it, though I do not have any
reference to mention.
In what follows, I will try to explain, step by step, what a hypergroup
is, and, then, suggest a notion of simplicity for hypergroups, in a simple
and natural way, to match the notion in the case of groups, hoping it
will be fruitful.
Examples and constructions are included.
2
LH
Introduction
Ce qui suit est un plaidoyer pour la réouverture du chantier de classification des groupes simples finis afin d’y inclure les hypergroupes
simples finis.
La notion d’hypergroupe a été introduite par Frédéric Marty, en 1934,
lors d’un congrès à Stockholm, à ne pas confondre avec une notion plus
tardive et tout à fait différente à laquelle on a, malencontreusement,
donné le même nom.
Je suis bien conscient qu’un bon nombre de mathématiciens ont déjà
dû se sentir mal à l’aise au sujet de la présence de ce que l’on nomme les
groupes simples sporadiques dans le grand tableau de la classification,
et certains ont même pu le coucher par écrit, bien que je ne puisse
mentionner aucune référence.
Dans ce qui suit, on va commencer par rappeller, pas à pas, ce qu’est
un hypergroupe puis comment y introduire la simplicité de manière
simple (et utile) et qui coı̈ncide avec la notion classique, dans le cas des
groupes.
On élabore en donnant une manière très générale pour la construction
d’hypergroupes et l’on donne des conditions pour qu’ils soient simples.
On fournit de nombreux exemples et caractérisations
On pose enfin la question : ≪À peine clos, le chantier de la classification
des groupes simples finis devra-t-il, éventuellement, rouvrir ses portes
afin d’entreprendre la classification des hypergroupes simples finis ?≫
Quelques définitions
Un hypergroupe est un ensemble muni d’une opération binaire associative multivalente dont les propriétés généralisent celles des groupes.
Plus précisément, une opération binaire multivalente sur un ensemble donné H est une application (x, y) 7→ x.y qui, à chaque couple
d’éléments (x, y) ∈ H × H, fait correspondre une partie x.y ⊂ H.
Groupes de Mathieu sporadiques ?
3
Pour X ⊂ H et Y ⊂ H, on désigne alors par X.Y la réunion de
toutes les parties x.y où x parcourt X et y parcourt Y .
Le maniement des opérations multivalentes nécesssite un tout petit
effort d’attention de plus que l’usage des opérations classiques, univalentes, n’en demande, mais l’habitude finit par estomper la difficulté.
Lorsque le risque de confusion est minime, on écrit, simplement, x.Y
au lieu de {x}.Y et, de même, X.y au lieu de X.{y}. Enfin, si une
des parties x.y est un singleton, au lieu de x.y = {z}, on écrit plus
simplement x.y = z et on retrouve l’écriture classique des opérations
binaires univalentes.
Définition. Un hypergroupe est un ensemble non vide, H, muni
d’une opération binaire multivalente qui vérifie les deux identités suivantes, pour tous x, y, z dans H,
x.(y.z) = (x.y).z : associativité.
x.H = H = H.x : reproductivité.
Ces deux propriétés impliquent que (x.y).H = x.(y.H) = x.H = H,
de sorte que le produit x.y n’est jamais vide.
Bien entendu, tout groupe est un hypergroupe dont l’opération est
univalente. Réciproquement, tout hypergroupe dont l’opération est univalente est un groupe puisqu’alors, pour chaque couple (a, b) donné, les
équations a.x = b et y.a = b possèdent des solutions en x et en y.
Un exemple élémentaire. Soient G un groupe, H l’un quelconque
de ses sous-groupes et soit G/H = {xH : x ∈ G} l’ensemble des classes
à droite de G modulo H.
L’ensemble xHyH est la réunion des classes à droite xhyH où h
parcourt H. En posant (xH).(yH) = {xhyH : h ∈ H}, on définit
une opération binaire multivalente sur l’ensemble G/H qui en fait un
hypergroupe, comme on le vérifie sans détour.
Cette opération est univalente si et seulement si H est un sous-groupe
invariant de G : dans ce cas l’hypergroupe G/H est le groupe quotient
classique de G par H. Lorsque le sous-groupe H n’est pas invariant
dans G, on obtient alors un exemple élémentaire d’un hypergroupe qui
n’est pas un groupe.
Traditionnellement, on appelle D-hypergroupe tout hypergroupe
isomorphe à un hypergroupe de classes à droite de la forme G/H.
Morphismes. Un morphisme est alors, par définition, une application f : H → K d’un hypergroupe H dans un hypergroupe K qui
vérifie l’identité d’homomorphisme, f (x.y) = f (x).f (y).
4
LH
On appelle isomorphisme entre H et K toute application bijective
f : H → K qui est un morphisme ainsi que sa réciproque f −1 . Comme
dans le cas des groupes, un morphisme bijectif f : H → K est (déjà) un
isomorphisme car, dans ce cas, f −1 est aussi un morphisme. En effet,
pour a = f (x) et b = f (y), on aura x.y = f −1 f (x.y) = f −1 (f (x).f (y)),
autrement dit, f −1 (a).f −1 (b) = f −1 (a.b).
Bien entendu, lorsque H et K sont des groupes, on retrouve ainsi les
notions classiques d’homomorphismes et d’isomorphismes de groupes.
Nombres premiers et groupes simples. Comme pour la primalité,
il y a deux manières équivalentes de définir la simplicité.
Un entier n > 0 est premier lorsqu’il a exactement deux diviseurs :
cela revient à dire qu’il est seulement multiple de deux entiers distincts.
Un groupe est simple lorsqu’il a exactement deux sous-groupes invariants : cela revient à dire qu’il possède seulement deux images homomorphes distinctes, à isomorphisme près.
Afin d’étendre aux hypergroupes la notion de simplicité, il faut trouver un substitut convenable à la notion d’image homomorphe d’un
groupe. Voici ce que nous proposons.
Reflets et simplicité. On dira que l’hypergroupe K est un reflet de
l’hypergroupe H lorsqu’il existe une application surjective f : H → K
qui vérifie les identités suivante :
f −1 f (x.y) = f −1 (f (x).f (y)) = x.f −1 f (y) = f −1 f (x).y.
On dira alors que l’application f est un réflecteur.
Tout réflecteur f est nécessairement un morphisme. En effet, f étant
surjective, on a f (x.y) = f f −1 f (x.y) = f f −1 (f (x).f (y)) = f (x).f (y).
On observera aussi, sans détour, qu’une application f : H → K est
un isomorphisme si et seulement si c’est un réflecteur injectif.
Bien entendu, si H est un groupe, l’hypergroupe K est un reflet du
groupe H si et seulement si c’est une image homomorphe de H.
Chaque hypergroupe H possède toujours, naturellement, les deux
reflets suivants : lui-même et l’hypergroupe trivial à un seul élément ;
Groupes de Mathieu sporadiques ?
5
ces deux reflets sont isomorphes si et seulement si H lui-même est
trivial.
On dira qu’un hypergroupe est simple lorsqu’il possède (à isomorphisme près) exactement deux reflets. Cette manière de procéder permet d’éviter les arcanes de la théorie des sous-hypergroupes. On peut
espérer que cette nouvelle notion soit un bon choix.
Il est clair qu’un groupe donné est simple si et seulement si c’est
un ≪hypergroupe simple≫. La classe des hypergroupes simples finis
contient ainsi celle des groupes simples finis. Elle la contient, et la
prolonge : voir, en effet, la conséquence (5) ci-dessous.
Invariance. On aura besoin de la définition suivante. Etant donnés
un groupe G ainsi que deux sous-groupes H et K, on dira que le sousgroupe K est invariant modulo H lorsque l’on a
(∗)
KxK = HxK = KxH, pour tout x ∈ G.
Cela généralise la notion classique de sous-groupe invariant car, bien
entendu, un sous-groupe est invariant (au sens classique) si et seulement
s’il est invariant modulo le sous-groupe trivial.
On observera, en passant aux inverses, que chacune des deux conditions suivantes est équivalente à l’autre :
KxK = HxK, pour tout x ∈ G,
KxK = KxH, pour tout x ∈ G.
Chacune d’elles est ainsi équivalente à la condition (∗). En outre,
lorsque K est invariant modulo H, on a nécessairement H ⊂ K puisque,
pour x ∈ K, la condition HxK = KxK entraı̂ne HK = K. La condition (∗) peut donc s’énoncer également comme ceci :
(∗)
H ⊂ K et Kx ⊂ HxK, pour tout x ∈ G.
Théorème. Soient G un groupe, H et K deux quelconques de ses
sous-groupes. Si K est invariant modulo H, le D-hypergroupe G/K est
un reflet du D-hypergroupe G/H. Réciproquement, tout reflet de G/H
est isomorphe à l’un de ces D-hypergroupes G/K où K est invariant
modulo H.
Afin de ne pas couper le fil de l’exposé, nous renvoyons la démonstration
de ce théorème à l’Appendice ci-dessous.
6
LH
Conséquences
1. Tout reflet d’un D-hypergroupe est donc un D-hypergroupe.
2. Le D-hypergroupe G/H est simple si et seulement si H est distinct
de G et que les seuls sous-groupes K de G invariants modulo H sont
G et H lui-même.
3. En particulier, le D-hypergroupe G/H est simple dès lors que H est
un sous-groupe maximal dans G.
4. Ainsi, lorsque H est un sous-groupe maximal du groupe G et qu’il
n’est pas invariant dans G, l’hypergroupe G/H est simple et n’est
pas un groupe.
5. Il existe donc des hypergroupes simples finis qui ne sont
pas des groupes.
Classes à gauche
On définit, semblablement, l’hypergroupe H\G = {Hx : x ∈ G} des
classes à gauche modulo H. Les hypergroupes isomorphes à ces hypergroupes de classes à gauche possèdent les mêmes propriétés, corrélatives,
que les D-hypergroupes. Par exemple, H\G est un groupe si et seulement si H est invariant dans G et, s’il en est ainsi, les deux groupes
G/H et H\G coı̈ncident, bien entendu. De même, étant donnés deux
sous-groupes H ⊂ K ⊂ G, l’hypergroupe K\G est un reflet de H\G si
K est invariant modulo H.
Le théorème et ses conséquences se transposent, corrélativement. En
particulier H\G est simple si et seulement si G/H l’est.
Hypergroupe opposé
Soit H un hypergroupe avec son opération (x, y) 7→ x.y. Comme
dans le cas des opérations univalentes, on définit l’opération opposée
(x, y) 7→ x ◦ y = y.x. Muni de cette nouvelle opération, H est de
nouveau un hypergroupe que l’on désigne par H ◦ et que l’on appelle
l’hypergoupe opposé.
Que dire de l’hypergroupe (G/H)◦ opposé d’un hypergroupe G/H ?
Groupes de Mathieu sporadiques ?
7
Théorème. Soient G un groupe et H un de ses sous-groupes.
(1) Les deux hypergroupes (G/H)◦ et H\G sont isomorphes.
(2) D’un autre côté , pour que les deux hypergroupes G/H et H\G
soient isomorphes, il faut et il suffit que H soit un sous-groupe invariant
dans G.
Autrement dit, lorsque le sous-groupe H n’est pas invariant, les deux
hypergroupes G/H et H\G ne sont pas isomorphes. Lorsque le sousgroupe H est invariant, G/H et H\G sont un seul et même groupe
quotient.
Démonstration. (1) Plus précisément, l’application Hx 7→ (Hx)−1
qui, à chaque classe à droite Hx ∈ H\G, fait correspondre la classe à
gauche (Hx)−1 = x−1 H ∈ G/H est un isomorphisme de l’hypergroupe
H\G sur l’hypergroupe opposé (G/H)◦, comme une simple vérification
le montre.
(2) Si H est invariant, alors G/H et H\G sont, tous deux, le même
groupe quotient classique de G par H.
Réciproquement, soit f : G/H → H\G un isomorphisme entre ces
deux hypergroupes. Dans chacun d’eux, H est la seule classe qui vérifie
l’identité H.H = H. On a donc f (H) = H. De plus, quelque soit x ∈ G,
on a (xH).H = xH dans l’hypergroupe G/H. Pour chaque y ∈ G, il
existe x ∈ G tel que f (xH) = Hy. On aura alors
Hy = f (xH) = f ((xH).H) = f (xH).f (H) = (Hy).H.
Ainsi, dans H\G, le produit (Hy).H = Hy est un singleton : cela
revient à dire que HyH = Hy, de sorte que Hy ⊂ yH pour tout
y ∈ G, donc H est invariant.
Ils vont par paires
Soit K un sous-groupe d’un groupe donné G. Si l’hypergroupe G/K
est simple, on sait que l’hypergroupe K\G est également simple. Lorsque,
de plus, ce ne sont pas des groupes, on sait qu’ils ne sont pas isomorphes. Ainsi, comme les racines imaginaires conjuguées d’un
polynôme réel, ces hyergroupes simples vont par paires !
En particulier, en prenant tous les couples (G, K) où K est un sousgroupe maximal, non invariant, d’indice fini, dans un groupe G, on
8
LH
obtient une famille de paires d’hypergroupes simples finis (G/K, K\G),
deux à deux opposés et non isomorphes : cette famille prolonge ainsi
la famille des groupes cycliques simples finis.
Les plus simples des exemples
On prend pour G le groupe de toutes les permutations d’un ensemble
E non vide, fini ou infini, ayant α pour cardinal. On distingue un point
particulier p ∈ E et on prend H = {s ∈ G : s(p) = p}, le sous-groupe
des permutations qui laissent le point p fixe. C’est un sous-groupe maximal de G. Pour α > 3, ce sous-groupe H n’est pas invariant, de sorte
que K = G/H est alors un D-hypergroupe simple qui n’est pas un
groupe et dont le cardinal est celui de E. En désignant par e la classe
H ∈ K, on vérifie aisément que sa table de multiplication se résume
alors ainsi :
x.e = x, pour tout x ∈ K,
x.y = K \ {x}, pour y 6= e.
On observera encore ceci : pour tout x 6= e, on a
x2 = x.x = K \ {x} et x3 = x.x.x = K.
Ainsi, contrairement au cas des groupes, l’ordre de chacun de ces
hypergroupes monogènes est toujours égal à trois et ne coı̈ncide avec
son cardinal que pour α = 3.
En particulier, en prenant pour α, successivement, chacun des entiers
n > 3, on obtient une suite de D-hypergroupes simples finis ayant n
éléments et qui ne sont pas des groupes.
Un pas de plus
Le procédé de construction des hypergroupes de classes, à droite ou
à gauche, de la forme G/H ou H\G, peut être grandement étendu en
une large généralisation que voici.
On se donne un ensemble T , une partie C ⊂ T × T ainsi qu’une
opération binaire univalente, Op : C → T, partiellement définie, pour
les seuls couples (x, y) ∈ C que l’on appellera les couples composables.
Le plus souvent, on écrira simplement xy au lieu de Op(x, y) et on dira
que (T, C) est une trame.
Groupes de Mathieu sporadiques ?
9
Soit R une relation d’équivalence sur l’ensemble T et, pour chaque
élément x ∈ T , soit x̄ sa classe modulo R. On définit alors une opération
multivalente (naturelle) sur l’ensemble des classes T /R, de la manière
suivante :
z̄ ∈ x̄.ȳ ⇐⇒ (∃(u, v) ∈ C)(uv ∈ z̄).
On dira que (T, C, R) est une présentation de la structure quotient T /R et, lorsque T /R est un hypergroupe, on dira que la relation
R est adéquate.
On peut donner, facilement, des conditions nécessaires et suffisantes
pour que la relation d’équivalence R soit adéquate, sous la forme d’un
ensemble d’énoncés du premier ordre. Il n’est pas besoin de rentrer dans
le détail, ici. Voir, pour cela, l’Appendice ci-dessous. Pour le moment,
il suffit de savoir que ces conditions explicites existent.
Une heureuse circonstance
Bien entendu, chaque hypergroupe de classes, de la forme G/H ou
H\G, possède une présentation naturelle dont la trame est (T, C) où
T = G est le groupe lui-même, C = G × G, (i.e., tous les couples sont
composables, l’opération étant celle du groupe). La relation d’équivalence
R est définie par
xRy ⇐⇒ x−1 y ∈ H,
respectivement,
xRy ⇐⇒ yx−1 ∈ H.
Plus généralement donc, chaque hypergroupe isomorphe à l’un quelconque de ces hypergroupes de classes possède également une présentation. Cette particularité n’est, cependant, pas due à une quelconque singularité de ces hypergroupes. Elle est le lot de tous les hypergroupes, quels qu’ils soient. Voici comment.
Tout hypergroupe possède une présentation
En effet, soit H un hypergroupe quelconque et T = H × H × H × H.
À chaque triplet t = (x, y, z) ∈ H × H × H pour lequel z ∈ x.y,
on associe le couple composable ((x, t), (y, t)) ∈ T × T et on pose
Op((x, t), (y, t)) = (z, t) : cela définit une opération Op : C → T où
C est l’ensemble de tous les couples composables. On obtient ainsi la
trame (T, C). On prend pour R la relation d’équivalence sur T dont les
10
LH
classes sont les parties {x} × H × H × H. On vérifie, sans détour, que
T /R est un hypergroupe isomorphe à l’hypergroupe donné H. cqfd
Une nouvelle famille d’exemples
Voici une première illustration du procédé de présentation qui fournit
des hypergroupes qui ne sont pas des D-hypergroupes.
On prend un ensemble K réunion d’une famille (Ai )i∈I d’ensembles
non vides, deux à deux disjoints. On suppose que 0 ∈ I et on distingue
un point particulier e ∈ A0 . On introduit l’ensemble T formé de toutes
les applications injectives f : X → K où e ∈ X ⊂ K et vérifiant la
condition suivante : il existe (au moins) une permutation s : I → I
telle que l’on ait, pour tout i ∈ I,
f (X ∩ Ai ) ⊂ As(i) .
On dira que le couple (f, g) formé de deux éléments f : X → K
et g : Y → K de T est composable lorsque l’on a g(e) ∈ X et on
prend pour composé f g l’application y 7→ f (g(y)) définie sur la partie
Y ∩ g −1(X) : cette application, on le voit sans détour, appartient bien
à T .
On obtient, ainsi, une trame (T, C). L’application f 7→ f (e) de T
dans K est surjective et définit une relation d’équivalence R sur T
dont les classes d’équivalences sont les parties x̄ = {f ∈ T : f (e) = x}.
L’application x 7→ x̄ permet d’identifier l’ensemble T /R des classes
modulo R à l’ensemble K lui-même. L’opération multivalente correspondant à la présentation (T, C, R) de T /R = K est donnée par
z ∈ x.y ⇐⇒ (∃(f, g) ∈ C)(f (e) = x , g(e) = y , f (y) = z).
On en établit facilement la table de multiplication sous la forme
suivante : quels que soient ai ∈ Ai , pour chaque indice i,
x.e = x, pour tout x ∈ K,
ai .y = Ai \ {ai }, pour tout y ∈ A0 \ {e},
ai .aj = K \ Ai , pour tout indice j 6= 0.
Dans le cas particulier où I = {0}, K = A0 , on reconnaı̂t l’exemple
du D-hypergroupe simple déjà rencontré plus haut.
Groupes de Mathieu sporadiques ?
11
Plus généralement, si toutes les parties Ai ont même cardinal, K est
encore un D-hypergroupe.
En effet, on le voit facilement en introduisant une nouvelle trame G,
une sous-trame de la trame de départ T . Du fait que les parties Ai
sont équipotentes, chaque élément f de T se prolonge en un élément
g de T qui est une application bijective g : K → K, autrement dit,
une permutation de l’ensemble K. Ce prolongement n’est cependant
pas nécessairement unique. L’ensemble G ⊂ T formé de toutes ces
permutations de K est un groupe. On désigne par S la restriction
de l’équivalence R au sous-ensemble G et par H le sous-groupe de G
formé des permutations qui laissent fixe le point e. On vérifie, sans
grand détour, que la structure T /R est isomorphe à la structure G/S
elle-même isomorphe au D-hypergroupe G/H.
cqfd
On trouvera davantage de précisions au sujet de ces structures dans
l’Appendice, ci-dessous.
Pour le moment, on notera, cependant, encore ceci : pour
I = {0, 1} , |A0 | = n > 3 , |A1 | > 3 , n 6= p,
on obtient une structure à n + p éléments qui est un hypergroupe, mais
pas un D-hypergroupe : voir le A5 de l’Appendice.
Scholie
On va explorer les liens qui peuvent exister entre deux présentations
ayant une même trame (T, C).
On se donne deux relations d’équivalence R et S sur l’ensemble T . On
désigne, respectivement, par x̄ et x̂ la classe de x modulo R et la classe
de x modulo S. On a ainsi deux présentations, (T, C, R) et (T, C, S), de
la structure quotient T /R et de la structure quotient T /S, respectivement, qui sont munies des opérations multivalentes correspondantes :
z̄ ∈ x̄.ȳ ⇐⇒ (∃(u, v) ∈ (x̄ × ȳ) ∩ C)(uv ∈ z̄),
ẑ ∈ x̂.ŷ ⇐⇒ (∃(u, v) ∈ (x̂ × ŷ) ∩ C)(uv ∈ ẑ).
Bien entendu, lorsque R ⊂ S, on a x̄ ⊂ x̂ et, dans ce cas, x̄ 7→ x̂
est une application surjective f : T /R → T /S, f (x̄) = x̂. De plus, il
12
LH
suffit de se reporter aux définitions pour constater que l’on aura alors,
toujours, l’inclusion
f (x̄.ȳ) ⊂ x̂.ŷ.
Pour avoir l’égalité
f (x̄.ȳ) = x̂.ŷ,
il faut et il suffit que l’on ait
(∗∗)((u, v) ∈ (x̂ × ŷ) ∩ C) =⇒
(∀(p, q) ∈ x̂ × ŷ)(∃(r, s) ∈ (p̄ × q̄) ∩ C)(rs
b =u
cv).
Pour examiner plus attentivement cette situation, commençons par
lui donner un nom commode : convenons de dire que l’équivalence S
est invariante modulo R lorsque l’on a R ⊂ S et que la condition
précédente, (∗∗), est satisfaite pour tous x, y dans T .
Supposons donc que S soit invariante modulo R. Il va en
découler trois conséquences importantes.
1. Pour tous x, y, z dans T , on aura
f −1 f (x̄.ȳ) = f −1 (f (x̄).f (ȳ)) = x̄.f −1 f (ȳ) = f −1 f (x̄).ȳ.
2. Si R est adéquate, il en sera de même de S. Autrement dit, si T /R
est un hypergroupe, T /S sera un reflet de T /R car f est alors un
réflecteur d’après 1.
3. En particulier, si T /R est un groupe, T /S sera une image homomorphe de T /S.
Démonstration.
1. Du fait de l’égalité
f (x̄.ȳ) = x̂.ŷ = f (x̄).f (ȳ),
on a aussi
f −1 f (x̄.ȳ) = f −1 (f (x̄).f (ȳ)).
Par définition, on a
f (x̄) = x̂ , f (ȳ) = ŷ.
Il nous faut établir l’identité
f −1 (f (x̄).f (ȳ)) = x̄.f −1 f (ȳ)
qui s’écrit donc
f −1 (x̂.ŷ) = x̄.f −1 (ŷ).
On procède par implications et équivalences successives. On a
z̄ ∈ f −1 (x̂.ŷ) ⇐⇒ ẑ ∈ x̂.ŷ ⇐⇒ ∃ (u, v) ∈ (x̂ × ŷ) ∩ C tel que uv ∈ ẑ,
Groupes de Mathieu sporadiques ?
13
z̄ ∈ x̄.f −1 (ŷ) ⇐⇒ ∃ t ∈ ŷ et ∃ (u, v) ∈ (x̄ × t̄) ∩ C tel que uv ∈ z̄.
Or,
t ∈ ŷ et (u, v) ∈ (x̂ × t̂) ∩ C =⇒ (u, v) ∈ (x̂ × ŷ) ∩ C,
donc
z̄ ∈ x̄.f −1 (ŷ) =⇒ z̄ ∈ f −1 (x̂.ŷ).
Pour la réciproque, on utilise la condition (∗∗) sous la forme suivante :
t ∈ ŷ et (u, v) ∈ (x̂× ŷ) ∩C =⇒ ∃ (r, s) ∈ (x̄× t̄) ∩C tel que rs
b =u
cv,
donc
z̄ ∈ f −1 (x̂.ŷ) =⇒ z̄ ∈ x̄.f −1 (ŷ).
D’où l’égalité annoncée
f −1 (f (x̄).f (ȳ)) = x̄.f −1 f (ȳ).
On étalit de même, l’égalité
f −1 (f (x̄).f (ȳ)) = f −1 f (x̄).ȳ.
Cela démontre le point 1.
2. Si T /R est un hypergroupe, il en va de même de T /S du fait que f
est surjective et vérifie l’identité
f (x̄.ȳ) = x̂.ŷ = f (x̄).f (ȳ).
En effet, l’opération multivalente de T /S est alors associative car
celle de T /R l’est :
(x̂.ŷ).ẑ = f (x̄.ȳ).f (z̄) = f (x̄.ȳ.z̄) = f (x̄).f (ȳ.z̄) = x̂.(ŷ.ẑ).
De même, elle possède également la propriété de reproductibilité
comme pour T /R.
Cela démontre le point 2 dont découle immédiatement le point 3.
On aura remarqué, sans doute, la similitude avec la démonstration
faite au A1 de l’Appendice pour le cas particulier des D-hypergroupes.
On en vient, à présent, à la généralisation complète suivante du
théorème qui caractérise les reflets des D-hypergroupes.
Théorème. Soit (T, C, R) une présentation d’un hypergroupe H quelconque. Tout reflet L de H possède une présentation (T, C, S) où S est
une équivalence invariante modulo R.
14
LH
Démonstration. Soient p : T → h = T /R la projection de T sur
l’ensemble quotient, h : H → L un réflecteur et S l’équivalence sur T
définie par l’application composée g = hp, autrement dit,
xSy ⇐⇒ h(p(x)) = h(p(y)).
L’hypergroupe L est ainsi identifié à l’hypergroupe T /S présenté par
(T, C, S). On adopte les notations précédentes, en posant
x̄ = p(x) , x̂ = h(p(x)).
On va montrer que S vérifie la condition (∗∗). On se donne
(u, v) ∈ (x̂ × ŷ) ∩ C et (p, q) ∈ x̂ × ŷ.
On a ainsi
p̂ = x̂, q̂ = ŷ et u
cv ∈ x̂.ŷ = h(p̄.q̄).
Puisque h est un réflecteur, il existe z tel que z̄ ∈ p̄.q̄ et ẑ = u
cv. Cela
veut dire qu’il existe (r, s) ∈ (p̄ × q̄) ∩ C tel que z̄ = rs et, par suite,
rs
b = ẑ = u
cv.
Corollaire. Soit (T, C, R) une présentation d’un hypergroupe H non
trivial. Cet hypergroupe H est simple si et seulement si les deux seules
équivalences invariantes modulo R sont l’équivalence totale T × T et R
elle-même.
Une remarque. Insistons :
tout hypergroupe possède une présentation à l’aide d’une
trame convenable.
La démonstration donnée plus haut établit, en fait, le résultat encore
plus général que voici. Soit H un ensemble muni d’une opération multivalente quelconque : autrement dit, certains des produits x.y peuvent
être vides et l’opération n’a nul besoin d’être associative, ni reproductive. La construction donnée plus haut fournit une trame (T, C) et une
relation d’équivalence R sur T telles que (T, C, R) soit une présentation
de H, autrement dit, H est isomorphe à la structure quotient T /R.
On peut concevoir cette présentation comme un déploiement de
l’opération multivalente sur H en une opération univalente sur T , mais
partiellement définie. Ce que l’on gagne en simplicité, en passant du
multivalent à l’univalent, on le perd, un tout petit peu, en passant
d’une opération partout définie à une opération uniquement définie
pour certains couples : le mal n’est pas si grand car cela sous-tend
la notion d’une opération univalente générale pour laquelle certains
produits x.y au lieu d’être des singletons, seraient simplement vides.
Groupes de Mathieu sporadiques ?
15
Dans ce déploiement, chaque élément x de H est, pour ainsi dire, éclaté
en une partie de T .
On peut aussi concevoir cette présentation comme un revêtement
de la structure H par la structure (T, C), un peu comme pour les
variétés topologiques.
Cette façon de faire est loin d’être rare. Elle est illustrée par de très
nombreux exemples. Citons-en uniquement deux : le revêtement d’un
groupe de Lie connexe par un groupe de Lie simplement connexe, et
la représentation d’un ensemble analytique de la droite réelle par la
projection d’un fermé du plan. Les chemins qui ne se distinguent pas
très bien dans le groupe, se voient bien plus clairement une fois relevés
dans le revêtement. De même, le schéma (complexe) de Souslin qui
définit l’ensemble analytique est déployé, à l’aide du fermé du plan,
comme une nappe froissée et entortillée que l’on étend sur une corde à
linge.
Bien entendu, le critère de simplicité qu’on vient de donner pour les
hypergoupes s’applique à toutes les structures multivalentes ou univalentes. Il s’applique, en particulier, aux groupes : un groupe est simple
si et seulement si l’une de ses présentations (T, C, R), vérifie la condition suivante : les deux seules équivalences invariantes modulo R sont
l’équivalence totale T × T et R elle-même. De plus, s’il en est ainsi,
alors toutes les présentations du groupe vérifie cette même condition.
La question se pose alors : est-il plus simple d’établir qu’un groupe
est simple, directement, en s’y tenant au plus près, ou bien peut-on
simplifier la preuve, parfois, en prenant de la hauteur au-dessus du
groupe, à l’aide d’une trame convenable ?
Constructions ≪à la Utumi ≫
On se donne un hypergroupe H dont on désigne l’opération par le
signe + (même si elle n’est pas ncessairement commutative). On se
donne une partition P de l’ensemble H et, pour chaque x ∈ H, on
désigne par x̄ la partie de la partition P à laquelle x appartient (sa
≪classe≫ ). On suppose que H possède un élément 0 ayant les propriétés
16
LH
suivantes :
0̄ = {0} , x + 0 = x et x ∈ 0 + x ⊂ x̄ pour tout x ∈ H.
On notera que l’on a alors 0 + x̄ = x̄.
On définit sur H une nouvelle opération multivalente par la règle :
x.y = x + ȳ.
Cette opération est toujours reproductive puisque
x.H = x + H = H et H.x = H + x̄ = H.
Elle est associative si et seulement si, pour tous x et y, on a
x̄ + ȳ = x + ȳ.
Cela se voit simplement, sans grand détour. On dira alors que cet hypergroupe (H, .) est construit ≪à la Utumi≫ à partir de l’hypergroupe
(H, +) et de la partition P.
L’exemple de Utumi.
Voici le premier exemple de cette construction, donné en 1949 par
Utumi [7]. On prend pour H le groupe cyclique d’ordre huit :
H = Z/8Z = {0, 1, 2, 3, 4, 5, 6, 7}
et la partition formée des trois classes :
I = {0}, A = {1, 4, 7} et B = {2, 3, 5, 6}.
Les conditions précédentes, édictées pour la construction, étant satisfaite, (H, .) est un hypergroupe. Utumi a montré que ce n’était pas un
D-hypergroupe bien qu’il possède la propriété suivante : pour chaque
élément fixé x ∈ H, la collection {x.y : y ∈ H} est une partition de H
et les cardinaux |x.y| sont tous égaux, pour y ∈ H.
En 1940, Eaton [1] a donné le nom de cogroupes (à droite) à cette
classe particulière d’hypergroupes et l’hypergroupe H de Utumi en fait
partie. Aussi lui a-t-on donné le nom de cogroupe de Utumi. Pour plus
de détail, on se reportera à [3].
On va établir, ici, le résultat nouveau que voici.
Théorème. Le cogroupe de Utumi est un hypergroupe simple.
En effet, soit f : H → K un réflecteur et pour chaque partie X ⊂ H,
posons X̂ = f −1 f (X) et, en particulier, x̂ = f −1 f (x) pour chaque
x ∈ H. Les propriétés du réflecteur se traduisent, ici, de la manière
suivante :
x.y
c = x̂.ŷ = x.ŷ = x̂.y.
Groupes de Mathieu sporadiques ?
17
De ce que x.0 = x, on tire :
x̂ = x̂.0̂ = x.0̂ = x̂0
et, en particulier,
0̂ = 0̂.0̂ = 0.0̂ = 0̂.0.
Deux cas seulement peuvent se présenter.
Premier cas : 0̂ = 0. Alors, pour chaque x, on a x̂ = x.0 = x et f est
un isomorphisme.
Second cas : 0̂ contient un élément a 6= 0. Alors 0̂ contient aussi
le produit 0.a = ā donc, également, le produit ā.a = ā + ā ainsi que
toutes les sommes ā + ā + · · · + ā. Une simple vérification montre que
A + A + A = H et B + B = H. Donc, quel que soit a 6= 0, on voit que
l’on a 0̂ = H et l’hypergoupe K est donc trivial, réduit à un singleton.
cqfd
Plus généralement, et en brodant sur ce même canevas, on obtient
donc ceci.
Un hypergroupe (H, .) construit ≪à la Utumi≫, à partir d’un groupe
(H, +) et d’une partition P, est simple dès que, pour chacune des
parties A ∈ P, différente du singleton {0}, l’une au moins des sommes
A + A + · · · + A est égale à H.
Appendice
Lorsque l’on traite d’hypergroupes, on adopte souvent une convention tacite : chaque fois que le produit x.y est un singleton, on écrit,
indifféremment, x.y = {z} ou x.y = z, pourvu que le risque de confusion soit minime.
Reflets d’un D-hypergroupe
A1. Lemme. Soient H et K deux sous-groupes d’un groupe G. On
suppose que K est invariant modulo H. Le D-hypergroupe G/K est
alors un reflet du D-hypergoupe G/H.
18
LH
Démonstration. Pour x ∈ G, posons x̄ = xH. Par hypothèse, H est
un sous-groupe de K. En associant à chaque classe x̄ = xH ∈ G/H
la classe x̂ = xK ∈ G/K, on obtient donc une application surjective
canonique f : G/H → G/K. On va vérifier que f est un réflecteur.
On a f (x̄) = x̂ = xK, de sorte que
f (x̄.ȳ) = {ẑ : z ∈ xHyH},
f (x̄).f (ȳ) = x̂.ŷ = {ẑ : z ∈ xKyK},
et ces deux ensembles sont égaux.
En effet, soit z ∈ xKyK alors z ∈ xHyK, puisque KyK = HyK ;
autrement dit, l’intersection xK ∩ (xHy) n’est pas vide ; soit t un
élément de cette intersection ; alors t ∈ xHyH et zK = tK. On a
ainsi
f (x̄.ȳ) = f (x̄).f (ȳ).
De même, on a
f −1 f (x̄) = f −1 (x̂) = {z̄ : z ∈ xK},
de sorte que
f −1 f (x̄.ȳ) = f −1 (f (x̄).f (ȳ)) = {z̄ : z ∈ xKyK},
f −1 f (x̄).ȳ = {z̄ : z ∈ xKyH},
x̄.f −1 f (ȳ) = {z̄ : z ∈ xHyK}.
Ces quatre ensembles sont égaux et l’application f est donc bien un
réflecteur.
A2. Deux remarques utiles. Pour établir le lemme suivant, on aura
besoin des deux résultats que voici.
R1. Tout réflecteur f vérifie l’identité
f −1 f (x).f −1 f (y) = f −1 (f (x).f (y)).
En effet, soit z ∈ f −1 f (x).f −1 f (y). Il existe alors deux éléments
u ∈ f −1 f (x) et v ∈ f −1 f (y) tels que z ∈ u.v.
On a donc f (z) ∈ f (u.v) = f (u).f (v) = f (x).f (y).
Autrement dit, on a f −1 f (x).f −1 f (y) ⊂ f −1 (f (x).f (y)).
On conclut en remarquant que
f −1 (f (x).f (y)) = x.f −1 f (y) ⊂ f −1 f (x).f −1 f (y). cqfd
Groupes de Mathieu sporadiques ?
19
R2. On se donne des hypergoupes, H, K, L, ainsi que des applications,
f : H → K , g : K → L, et leur composée h = gf . Si f et g
sont des morphismes, il est clair que h est également un morphisme.
Réciproquement, si f et h sont des morphismes, et pourvu que f soit
surjective, l’application g est aussi un morphisme.
En effet, pour x, y dans K, il existe a, b dans H tels que x = f (a) et
y = f (b). Ainsi f (a.b) = f (a).f (b) = x.y donc
g(x.y) = h(a.b) = h(a).h(b) = gf (a).gf (b) = g(x).g(y).
Autrement dit, g est bien un morphisme.
cqfd
A3. Lemme. Soient G un groupe, H l’un quelconque de ses sousgroupes et L un reflet du D-hyprgroupe G/H. Il existe alors un sousgroupe K de G, invariant modulo H, tel que L soit isomorphe au
D-hypergroupe G/K.
Démonstration. Soit h : G/H → L un réflecteur. Désignons par e
l’élément unité du goupe G puis, pour chaque x ∈ G, posons x̄ = xH.
Soient 1 = h(ē) et K = {x : h(x̄) = 1}. On procède par étapes. On
commence par montrer que K est un sous-groupe de G puis on montre
qu’il est invariant modulo H et on établit, enfin, que L est isomorphe
à l’hypergroupe G/K.
1. Pour x ∈ H, on a x̄ = ē, de sorte que h(x̄) = h(ē) = 1. On a donc
H ⊂ K.
On commence par utiliser le fait que h est un morphisme.
2. Pour x ∈ G et y ∈ G, on a xy ∈ xHyH donc xy ∈ (xH).(yH) = x̄.ȳ,
de sorte que
h(xy) ∈ h(x̄.ȳ) = h(x̄).h(ȳ).
3. Pour x ∈ G, on a aussi x̄.ē = x̄ d’où l’on tire h(x̄).1 = h(x̄). De
sorte que, pour y ∈ K, on a
h(xy) ∈ h(x̄).1 = h(x̄) qui est un singleton, donc h(xy) = h(x̄).
4. En particulier, pour y ∈ K, il vient
1 = h(y −1y) = h(y −1) donc y −1 ∈ K.
5. Plus généralement, si x ∈ K et y ∈ K, alors
h(x−1 y) = h(x−1 ) = 1 donc x−1 y ∈ K,
de sorte que
K −1 K ⊂ K.
Cela prouve bien que K est un sous-groupe de G.
20
LH
6. En procédant par équivalences successives, on a
z̄ ∈ h−1 h(ē) ⇐⇒ z̄ ∈ h−1 (1) ⇐⇒ h(z̄) = 1 ⇐⇒ z ∈ K.
Autrement dit
z̄ ∈ h−1 h(ē) ⇐⇒ z ∈ K.
On utilise, à présent, les identités spécifiques aux réflecteurs.
7. On a h−1 h(x̄) = h−1 h(x̄.ē) = x̄.h−1 h(ē). D’où
z̄ ∈ h−1 h(x̄) ⇐⇒ (∃ū ∈ h−1 h(ē))(z̄ ∈ x̄.ū) ⇐⇒
(∃u ∈ K)(z ∈ xHuH) ⇐⇒ z ∈ xK.
Ainsi, h(z̄) = h(x̄) ⇐⇒ z ∈ xK. Il en résulte que l’on a
h(x̄) = h(ȳ) ⇐⇒ xK = yK.
Il existe ainsi une application g : G/K → L qui, à chacune des classes
xK ∈ G/K, associe g(xK) = h(x̄) ∈ L et g est une application bijective.
8. On a l’identité ē.h−1 h(x̄) = h−1 h(ē).h−1 h(x̄), d’après R1.
Or, z̄ ∈ ē.h−1 h(x̄) ⇐⇒ z ∈ HxK.
De même, z̄ ∈ h−1 h(ē).h−1 h(x̄) ⇐⇒ z ∈ KxK.
Donc, pour tout x ∈ G, on a
HxK = KxK.
Cela achève de prouver que K est invariant modulo H.
9. Soit f : G/H → G/K l’application canonique qui fait correspondre,
à la classe xH ∈ G/H, la classe xK ∈ G/K. C’est un réflecteur, d’après
A1. On a g(f (x̄)) = g(xK) = h(x̄). Autrement dit, h est la composée
gf . Or, en particulier, f est un morphisme surjectif et h un morphisme.
Donc g est un morphisme, d’après R2. Puisque g est bijectif (voir le 7
ci-dessus) c’est donc un isomorphisme, comme annoncé.
A4. Conditions d’adéquation. Soient (T, C) une trame et R une
relation d’équivalence sur T . Pour chaque x ∈ T , on désignera par x̄
sa classe modulo R, de sorte que toutes les relations suivantes sont
synonymes :
xRy , yRx , y ∈ x̄ , x ∈ ȳ , x̄ = ȳ.
Afin de simplifier les écritures, pour r ∈ T et s ∈ T , on conviendra de
ceci : lorsque l’on écrit rs, on suppose (implicitement) que le couple
(r, s) est composable, autrement dit, que l’on a (r, s) ∈ C.
Voici deux énoncés relatifs à la structure quotient T /R, (ils sont tous
deux du premier ordre) :
(1) (∀x ∀y)(∃ r ∈ x̄ ∃ s ∈ T )(rs ∈ ȳ) et (∃ r ∈ T ∃ s ∈ ȳ)(rs ∈ x̄)}.
Groupes de Mathieu sporadiques ?
(2)
21
(∀x)(∀y)(∀z)(∀u)(∃r ∈ x̄ , s ∈ ȳ , t ∈ z̄)((rs)t ∈ ū ⇐⇒
(∃r ∈ x̄ , s ∈ ȳ , t ∈ z̄)(r(st) ∈ ū).
On voit, sans détour, que la condition (1) est nécessaire et suffisante pour assurer la reproductibilité. De même, la condition (2) est
nécessaire et suffisante pour assurer l’ associativité.
Autement dit, pour que l’équivalence R soit adéquate, il faut et il
suffit qu’elle satisfasse les deux énoncé s (1) et (2).
Sur la nouvelle famille d’exemples
On se donne un ensemble K, réunion d’une famille (Ai )i∈I d’ensembles non vides, deux à deux disjonts. On suppose que 0 ∈ I et on
distingue un point particulier e ∈ A0 .
Sur cet ensemble K, on a défini une opération multivalente dont voici
la table : quels que soient ai ∈ Ai , pour chaque indice i,
x.e = x, pour tout x ∈ K,
ai .y = Ai r {ai }, pour tout y ∈ A0 r {e},
ai .aj = K r Ai , pour tout indice j 6= 0.
Il est clair que la structure ainsi définie sur K dépend, uniquement, à
isomorphisme près, de la famille des cardinaux, finis ou infinis, n = |A0 |
et, pour i 6= 0, pi = |Ai | : on dira qu’elle est de type S(n, (pi )). Cela
généralise, quelque peu, les structures de type S(n, p), correspondant
au cas où l’on a |I| 6 2, introduite dans l’article [2].
A5. Lorsque l’on a n > 3, pi > 3, pour tout indice i, et pk 6= n,
pour un indice k, la structure est un hypergroupe qui n’est pas un
D-hypergroupe.
En effet, pour voir que la structure est un hypergroupe, il suffit de
s’assurer que les conditions d’adéquation sont satisfaites.
Pour la condition de reproductibilité , la vérification est sans détour.
Quant à la condition d’associativité, on reprend la trame (T, C) qui
sert à construire la structure. Soient alors ai , aj , ak , donnés dans K ainsi
que r, s, t, dans T tels que r(e) = ai , s(e) ∈ aj , t(e) ∈ ak . Si (r, s) ∈ C
22
LH
et (rs, t) ∈ C, on voit, simplement, que l’on a (s, t) ∈ C et (r, st) ∈ C,
de sorte que ((rs)t)(e) = (r(st))(e).
Réciproquement, soit (s, t) ∈ C et (r, st) ∈ C. Ainsi, s(t(e)) = s(ak )
est défini, de même que r(s(ak )) = al . On montre qu’il existe alors un
élément q ∈ T défini au point aj et tel que
q(e) = r(e) = ai , (q, s) ∈ C , q(s(ak )) = r(s(ak )) = al , (qs, t) ∈ C,
ce qui établira la réciproque. L’application r est déjà définie au point e
et au point s(ak ). Si aj est l’un des points e ou s(ak ), on prend q = r.
Sinon, on construit q : {e, s(ak ), aj } → K en posant
q(e) = r(e) = ai , q(s(ak )) = r(s(ak )) = al , q(aj ) = x,
avec un élément x soumis aux trois conditions suivantes :
(1) x ∈
/ {ai , al }.
(2) x ∈ Ai si e ∈ Aj .
(3) x ∈ Al si s(ak ) ∈ Aj .
Cela est possible car chacune des parties Am de la famille possède au
moins trois points. Enfin, ces conditions impliquent que l’on a bien
q ∈ T.
On pourrait également vérifier à l’aide de la table de multiplication,
avec un peu de patience, que l’opération est associative et reproductive.
Enfin, pour y ∈ A0 r {e} et a ∈ Ak , les deux parties e.y = A0 r {e}
et a.y = Ak r {a} ne sont pas équipotentes alors qu’elles le seraient
dans un D-hypergroupe.
cqfd
On peut ainsi ranger les différentes structures S(n, (pi )) dans quatre
classes distinctes.
Première classe : pi = n pour tout indice i. Cela comprend, bien
entendu, tous les cas où I = {0}. Toutes ces structures sont des Dhypergroupes, comme on l’a dit plus haut.
Deuxième classe : n > 3, pi > 3 pour tout indice i et pk 6= n pour
un indice k. Ces structures sont, toutes, des hypergroupes qui ne sont
pas des hypergroupes de classes.
Troisième classe : n > 2 et pk = 1 pour un indice k. L’un des produits
est vide : en particulier, A0 = {e, y, . . .}, Ak = {a} et a.y = ∅.
Groupes de Mathieu sporadiques ?
23
Quatrième classe : tous les autres cas. Ces structures ne sont pas des
hypergroupes car l’associativité est alors en défaut, comme on peut le
vérifier.
Vérifications. Il n’y a que les trois cas suivants :
1. n = 1 et pk > 2 pour un indice k.
2. n = 2 et pk > 3 pour un indice k.
3. n > 3 et pk = 2 pour un indice k.
1. On a A0 = {e}, Ak = {a, b, . . .} donc a.(a.a) 6= (a.a).a car
a.(a.a) = a.(K \ Ak ) ⊂ {a} ∪ (K \ Ak ) 6∋ b,
[
(a.a).a = (K \ Ak ).a = (K \ Ai ) ∋ b.
i6=k
2. On a A0 = {e, y}, Ak = {a, b, c, . . .}, donc a.(y.y) 6= (a.y).y car
a.(y.y) = a.e = a,
(a.y).y = (Ak \ {a}).y ⊃ (b.y) ∪ (c.y) = Ak .
3. On a A0 = {e, y, z}, Ak = {a, b} donc a.(y.y) 6= (a.y).y car
a.(y.y) = a.{e, z} = Ak ,
(a.y).y = b.y = a.
Pour la petite histoire
Les premières versions de ce texte remontent au siècle dernier ( !)
Très souvent remanié, il a été exposé, en partie, au Groupe de travail de
théorie des nombres du Département de mathématiques de l’Université
Blaise Pascal, de Clermont-Ferrand, sur invitation, le lundi 18 juin
2012, sous le titre :
≪Est-il encore possible d’avancer à la manière d’Emile Mathieu ?≫
avec le résumé suivant :
≪Courte promenade mathématique à travers champs, suivie d’un bref
plaidoyer en faveur de la révision de la liste des sporadiques.≫
24
LH
En guise de conclusion (provisoire)
On sait faire agir les hypergroupes sur des ensembles, comme on le
fait pour les groupes. Plus précisément, on sait les faire agir sur l’ensemble des racines d’un polynôme. Et c’est encore une autre histoire.
On peut dès lors se poser la question suivante. Est-il raisonnable, ou
non, de prédire l’existence d’hypergroupes de Mathieu simples ?
À peine clos, le chantier de la classification des groupes simples
finis devra-t-il, éventuellement, rouvrir ses portes afin d’entreprendre
la classification des hypergroupes simples finis ?≫
≪
Bibliographie
[1] Eaton, J. E., Theory of cogroups, Duke Math. J., 6 (1940) 101-107.
[2] Haddad, Labib et Sureau, Yves, Les cogroupes et les D-hypergroupes,
J. Algebra, 118 (1988) 468-476.
[3] Haddad, Labib et Sureau, Yves, Les cogroupes et la construction de
Utumi, Pacific J. Math., 145 (1990) 17-58.
[4] Krasner, M., Sur la théorie de la ramification des idéaux de corps
nongaloisiens de nombres algébriques, Acad. Belgique, Cl. Sci. Mém.
Coll. 4◦ , 11, Thèse, (1937), 110 pages.
[5] Marty, F., Sur une généralisation de la notion de groupe, 8th. Scand.
Math. Congr., Stockholm (1934) 45-49.
[6] Marty, F., Sur les groupes et hypergroupes attachés à une fraction
rationnelle, Ann. Sci. Éc. Norm. Supér. Ser. III. 53 (1936) 83-123.
[7] Utumi, Yuzo, On hypergroups of group right cosets, Osaka Math.
J., 1 (1949) 73-80.
120 rue de Charonne, 75011 Paris, France
E-mail address: [email protected]
| 4math.GR
|
Optimal Task Scheduling in
Communication-Constrained Mobile Edge
Computing Systems for Wireless Virtual Reality
Xiao Yang∗, Zhiyong Chen∗, Kuikui Li∗ , Yaping Sun∗ and Hongming Zheng†
Medianet Innovation Center, Shanghai Jiao Tong University, Shanghai, P. R. China
†Intel Corporation, P.R. China
[email protected]
Email: {yangxiao1652, zhiyongchen, kuikuili, yapingsun
}@sjtu.edu.cn,
arXiv:1708.00606v1 [cs.IT] 2 Aug 2017
∗ Cooperative
Abstract—Mobile edge computing (MEC) is expected to
be an effective solution to deliver 360-degree virtual reality
(VR) videos over wireless networks. In contrast to previous
computation-constrained MEC framework, which reduces the
computation-resource consumption at the mobile VR device
by increasing the communication-resource consumption, we develop a communications-constrained MEC framework to reduce communication-resource consumption by increasing the
computation-resource consumption and exploiting the caching
resources at the mobile VR device in this paper. Specifically,
according to the task modularization, the MEC server can only
deliver the components which have not been stored in the VR
device, and then the VR device uses the received components
and the corresponding cached components to construct the task,
resulting in low communication-resource consumption but high
delay. The MEC server can also compute the task by itself to
reduce the delay, however, it consumes more communicationresource due to the delivery of entire task. Therefore, we then
propose a task scheduling strategy to decide which computation
model should the MEC server operates, in order to minimize
the communication-resource consumption under the delay constraint. Finally, we discuss the tradeoffs between communications,
computing, and caching in the proposed system.
I. I NTRODUCTION
Virtual reality (VR) over wireless networks is gaining an unprecedented attention due to the ability of bringing immersive
experience to users. VR application is computational-intensive,
communications-intensive and delay-sensitive, leading to the
fact that most of VR are wired with cables. Current wireless
systems (e.g., LTE) cannot cope with the ultra-low latency and
ultra-high throughput requirements of wireless VR application
(e.g., 360-degree VR video) [1].
To address this problem, an effective solution is mobile edge
computing (MEC) [2], [3], which enables cloud computing
capabilities at the edge of wireless network, e.g., base stations
(BS). By deploying computation resource at the network edge,
MEC performs the computation tasks closer to the VR device,
improving the quality of computation experience.
Computation-Constrained MEC, where the VR user can
offload the computation tasks from the VR device to the MEC
server due to the limited computation capability at the VR
device, has attracted significant attention recently [6]–[8]. In
[6], a dynamic computation scheduling algorithm based on
Lyapunov theory was proposed to minimize the execution
!"
#
Fig. 1. Task can be modularized and then recovered at the user [4], [5].
delay and task failure. Similarly, [7] proposed an optimization
offloading framework to minimize the execution latency and
the user’s energy consumption. In order to minimize the
mobile energy consumption under the constraint of execution
delay, [8] optimized the offloading policy based on the channel
gains and the consumption of local computing energy. In
general, the computation-constrained MEC exploits the communication resource to reduce the requirement of the user’s
computation resource. Therefore, it is quite suitable for the
computational-intensive and delay-sensitive application with
low bandwidth consumption, e.g., online picture processing.
The video data rate on online 360-degree VR videos is
already many times greater than that of a high definition
TV, even though 5G networks maybe not able to satisfy this
requirement [1]. Meanwhile, the end-to-end latency through
the system is no more than 20 ms for online 360-degree
VR videos. We can also use the MEC solution to improve
network responsiveness and reduce latency, however, we cannot try to reduce the computation resource by increasing the
communications resource overhead. In contrast, we should try
to reduce the consumption of communications resource by
taking advantage of the computation and caching resources
at the mobile VR devices, where we term this solution as
Communications-Constrained MEC in this paper.
In this paper, we present a communications-constrained
MEC framework and develop an optimal task scheduling
policy to minimize the average transmission data per task
in this system, where the average transmission data per task
is adopted as the consumption of communications resource
metric. Our major contributions are summarized as follows:
•
We propose a communications-constrained MEC frame-
denoted by p = [p1 , . . . , pN ], where
N
P
pi = 1. We assume
i=1
Fig. 2. The proposed communication-constrained MEC system with a caching
enabled mobile VR device
work to reduce the consumption of communications resource byexploiting
the caching resource and increasing
the consumption of computing resource at the mobile VR
device. To address
this issue, we consider that a 360degree VR video (task) can be modularized today [4],
as
Media Transport (MMT) standard,
e.g., MPEG
[5],
shown in Fig. 1. Besides, the
contents
composing
each
whole task have popularity. As a result, for one intended
task, the MEC server only delivers the corresponding
components which have not been stored in the VR device,
and then the VR device uses the received components
and the corresponding cached components to construct
the task by exploiting the local computation resource.
• We develop an optimal task scheduling policy to minimize
%
%
the average transmission data per task. Of course,
the
MEC server can also select MEC computation model,
which the MEC server combines all corresponding con !
"#
$ !
tents as the target task and then deliver the task to the VR
device. The MEC computation mode is a reliable way to
the
latency due to the fast computation capability
reduce
at the MEC server, but this model delivers more data per
task to the user. Therefore, we formulate the transmission
data consumption minimization problem under the delay
constraint and propose a task scheduling strategy based
on Lyapunov theory.
• We discuss the tradeoffs between communications, computing, and caching. We present how to joint allocate
communications, computing and caching resources in the
proposed communications-constrained MEC system to
achieve a target end-to-end latency, e.g., 20 ms.
II. S YSTEM M ODEL
As shown in Fig.2, we consider an MEC system, where a
cache-enabled mobile VR device can access BS with an MEC
server to obtain task. The MEC server has an abundance of
computing and caching resource, while the mobile VR device
has limited computing ability and cache capacity.
A. Task Model
We consider each task consists of a number of contents,
e.g., MMT assets. All the contents composing each task come
from a set of N possible contents, which is denoted by F =
{F1 , F2 ......FN }. Note that one content may be used more than
once in a task. The popularity distribution of the contents is
all the contents are of equal size τ and the MEC server has
all N contents. The cache capacity of mobile VR device is M
with M < N , which can store at most M contents. We adopt
the most popular caching strategy and the stored content set
can be denoted as M = {F1 , F2 ......FM }.
The system is time-slotted with the time slot length ∆. Let
Ht be the task scheduled at time slot t, which consists of Kt
contents. We denote H(t) = [h1 (t), . . . , hKt (t)] as the content
index vector of the task Ht , where hkt (t) ∈ {1, . . . , N }
indicates that the kt -th content in Ht is Fhkt (t) . Thus the size
of Ht is D(t) = τ Kt . Let Gtn (1 ≤ n ≤ N ) denote whether
Fn ∈ F is requested in Ht and not cached in mobile VR
device, which can be given by
(
1,
for M + 1 ≤ n ≤ N, and hkt (t) = n
t
Gn =
(1)
0,
otherwise.
B. Computation Model
Each task can either be executed at the MEC server, or at the
mobile VR device. Let W denote the required CPU cycles for
computing one bit. The CPU frequency of the MEC server and
the mobile VR device is fc and fl , respectively. The wireless
transmission throughput is R.
1) MEC Computation Mode: When the mobile VR device
has a task request, the MEC server computes the contents as
the task, and then deliver the task. The size of transmission
data is Dct (t) = D(t). The size of computation data is
Dcc (t) = D(t). Therefore, Nc (t) = ⌈Dcc (t)W/(fc ∆) +
Dct (t)/(R∆)⌉ time slots are required to complete the task Ht .
Similar to [9], we use Sc (t) ∈ {0, 1, ..., Nc(t) − 1} to model
the MEC state, where Sc (t) = 0 means that the MEC server
is idle and Sc (t) = n (n 6= 0) indicates a task is processing
at the MEC server and Nc (t) − n time slots are required to
complete the computation.
2) Local Computation Mode: For this model, the MEC
server delivers the contents which do not been stored in the
mobile VR device, and the mobile VR device computes the
received contents with the cached contents as the task. Because
the MEC server does not transmit the stored contents of the
N
P
Gtn .
task, the size of transmission data is Dlt (t) = τ
n=M+1
The size of computation data also is Dlc (t) = D(t). As
a result, Nl (t) = ⌈Dlc (t)W/(fl ∆) + Dlt (t)/(R∆)⌉ time
slots are required to complete Ht . Similarly, let Sl (t) ∈
{0, 1, ..., Nl (t) − 1} denote the mobile VR device state. If
the system allocates a task to the mobile VR device at time
slot t, Sl (t) updates to Sl (t + 1) = Nl (t) − 1.
C. Task Queueing Model
The task arrival process is modeled as a Bernoulli process
with probability λ. When a task arrivals, the task first enters
into a task queue with infinite capacity. The number of the task
waiting in the queue is the queue state Q(t) = {0, 1, 2, 3......},
where Q(t) updates according to the following equation
Q(t+1)=(Q(t)−(u1l (t)+u2l (t)+u1c (t)+u2c (t)))+ +A(t), (2)
where A(t) denotes whether a task arriving in the time slot t.
Thus we have Pr{A(t) = 1} = λ and Pr{A(t) = 0} = 1 − λ.
Here, {u1l (t), u2l (t), u1c (t), u2c (t)} denotes the task scheduling
decision at time slot t.
Notice that at most two tasks can be scheduled at a
time slot. The first task should be scheduled before the
second task. If the first task is scheduled to do the computation at the mobile VR device (MEC server), we have
u1l (t) = 1 (u1c (t) = 1) and the second task can not be
scheduled to operate in the local computation mode (MEC
computation mode) because the CPU has been occupied by
the first task, yielding u2l = 0 (u2c (t) = 0). Otherwise,
we have u1l (t) = 0 (u1c (t) = 0) for the first task and the
local computation mode (MEC computation mode) could be
scheduled for the second task, i.e., u2l = 1 (u2c (t) = 1). As
a result, there are five possible states for the task scheduling
decision in each time slot, i.e., {u1l (t), u2l (t), u1c (t), u2c (t)} =
{(0, 0, 0, 0), (1, 0, 0, 0), (0, 0, 1, 0), (1, 0, 0, 1), (0, 1, 1, 0)}.
III. TASK S CHEDULING S TRATEGY AND P ROBLEM
F ORMULATION
The MEC server is a reliable way to reduce the computation
latency due to fc ≥ fl , but consumes more communication
resource due to Dc (t) ≥ Dl (t). Hence, the MEC server needs
to make the task scheduling decision at each time slot to
minimize the average transmission data per task under the
average delay constraint.
A. Task Scheduling Strategy
When the MEC server (mobile VR device) is idle, the
task can be scheduled to the MEC computation mode (local
computation mode). The queue state Q(t) = 0 denotes the task
queue is empty and there is no task will be scheduled in time
slot t+1, while Q(t) = ∞ indicates that there are infinite tasks
in the task queue, yielding the unstable system. According to
Q(t), Sl (t), and Sc (t), we can describe the system state.
Case 1: Sl (t) = Sc (t) = 0. Both the mobile VR device and
the MEC server are idle. The system can process at most two
tasks. If there are two tasks in Q(t) at least, i.e., Q(t) ≥ 2,
one task can be processed in the mobile VR device (MEC
server) and the other task remains wait in the task queue or
to be precessed in the MEC server (mobile VR device). The
task scheduling policy can be expressed as the follow:
(0, 0, 0, 0)
(0,
0, 1, 0)
1
u (t) = (1, 0, 0, 0) for Q(t) ≥ 2.
(3)
(1,
0,
0,
1)
(0, 1, 1, 0)
If there is only one task in Q(t), the task can be processed
in the mobile VR device, the MEC server or remains wait in
the task queue. We thus have
(0, 0, 0, 0)
u2 (t) = (0, 0, 1, 0) for Q(t) = 1.
(4)
(1, 0, 0, 0)
Case 2: Sl (t) 6= 0, Sc (t) = 0. In this case, the MEC server
is idle and the mobile VR device is busy so that the system
can process one task at most for the MEC server. The task
scheduling policy is:
(
(0, 0, 0, 0)
3
u (t) =
for Q(t) ≥ 1.
(5)
(0, 0, 1, 0)
Case 3: Sl (t) = 0, Sc (t) 6= 0. In this case, the mobile VR
device operates in idle mode and the MEC server is occupied.
Only one task can be scheduled for the mobile VR device.
The task scheduling policy can be expressed as following:
(
(0, 0, 0, 0)
4
u (t) =
for Q(t) ≥ 1.
(6)
(1, 0, 0, 0)
Case 4: Sc (t) 6= 0, Sl (t) 6= 0 or Q(t) = 0. If both the MEC
server and mobile device are busy, i.e., (Sc (t) 6= 0, Sl (t) 6= 0),
or there is no task in the task queue Q(t) = 0, no task is
scheduled. We then have
u5 (t) = (0, 0, 0, 0).
(7)
At the time slot t, Sl (t) and Sc (t) can be expressed as:
1
2
max(Sl (t)−1, 0) ul (t) = 0 or ul (t)=0,
1
1
(8)
Sl (t+1) = Nl (t) − 1
ul (t) = 1,
2
2
ul (t) = 1.
Nl (t) − 1
2
1
max(Sc (t)−1, 0) uc (t)=0 or uc (t)=0,
Sc (t+1)= Nc1 (t) − 1
u1c (t) = 1,
2
u2c (t) = 1.
Nc (t) − 1
(9)
where Nli (t) and Nci (t) denote Nl (t) and Nc (t) of the i-th
task for i = 1, 2, respectively. Sl (t + 1) = Nli (t) − 1 means
the mobile VR device is occupied by a task in the time slot
t + 1 and will be busy in the follow Nli (t) − 1 time slots.
Similarly, we have Sc (t + 1) = Nci (t) − 1.
B. Problem Formulation
When T → ∞ and the length of task queue is not infinite,
the total number of the task is close to λT . Therefore, the
average transmission data per task can be expressed as:
T −1 2
o
1 nXX i
i
i
ul (t)Dlt
(t) + uic (t)Dct
(t) .
lim
T →∞ λT
t=0 i=1
(10)
i
i
where Dlt
(t) and Dct
(t) denote Dlt (t) and Dct (t) of the i-th
task for i = 1, 2, respectively.
From the system model, we know that each task requires
transmission time, waiting time and processing time. The
computation processing time of the MEC server or the mobile
VR device is the dominant influence on the execution delay
[10]. Based on the Little Law [11], the execution delay,
including the waiting time and processing time, is proportional
to the average queue length of the task buffer. The execution
delay is written as:
T −1
i
1 hX
lim
E
Q(t) .
T →∞ T
t=0
(11)
Denote π(t) , {u1l (t), u2l (t), u1c (t), u2c (t)}. Thus, the
communication-resource consumption minimization problem
is formulated as:
T
2
o
1 nXX i
i
i
ul (t)Dlt
(t) + uic (t)Dct
(t)
P1 : min
lim
T →∞ λT
π(t)
t=1 i=1
s.t.
π(t) ∈ uk (t),
k ∈ {1, 2, 3, 4, 5},
T
i
1 hX
lim
E
Q(t) < ∞,
T →∞ T
t=1
(12)
(13)
where (13) indicates the delay constraint to ensure the task requires can be completed with a finite delay. Unfortunately, P1
is a stochastic optimization problem. The system state changes
after a offloading decision is made, and P1 is impossible to
be solved by convex optimization methods.
IV. O PTIMAL TASK S CHEDULING A LGORITHM BASED O N
LYAPUNOV T HEORY
A. Problem Transform
To simplify P1, we consider Lyapunov optimization theory.
We first define the Lyapunov function:
1 2
Q (t).
(14)
2
Consider the initial state Q(0) = 0, and then we have
L(Q(0)) = 0. If the queue is unstable, L(Q(t)) is more
volatile than Q(t). Thus the expectation of L(Q(t)) is:
L(Q(t)) =
t−1
o
nX
[L(Q(i + 1)) − L(Q(i))]
E[L(Q(t))] = E
=
E{L(Q(i + 1)) − L(Q(i))|Q(i)}.
(15)
i=0
The system is stable when E[L(Q(t))] < ∞. Therefore the
Lyapunov drift function can be given by:
n
o
∆L(Q(t)) = E L(Q(t + 1)) − L(Q(t))|Q(t) .
(16)
We can see from (15) and (16) that to maintain the stability
of the queue, we should minimize (16) in each time slot.
Therefore the expectation of the L(Q(t)) would not tend to
infinite. As a result, we have the following Lemma 1.
Lemma 1. Let us define the scheduling rate U (t) = u1l (t) +
u2l (t) + u1c (t) + u2c (t). In order to ensure E[L(Q(t))] < ∞,
we have:
∆L(Q(t)) ≤ Cmax − Q(t)E[U (t)|Q(t)],
state Q(t) can also approach a lower length. However, the
minimization of ∆L(Q(t) can not cause the minimization
of (10). Thus, we define the Lyapunov drift-plus-penalty
function:
∆L(Q(t)) + V E[D(t)|Q(t)] ≤Cmax − Q(t)E[U (t)|Q(t)]
+ V E[D(t)|Q(t)],
(17)
where we use Cmax = 21 {5 + 2Q(t)A(t)}.
Proof. Please refer to Appendix A.
According to Lyapunov theory [11], when we make the task
scheduling decision π(t) to minimize ∆L(Q(t), the queue
(18)
where V is a non-negative control parameter, which denotes
that the system is sensitive to the communication cost. When
V = 0, the system is only sensitive to the delay. With the
increase of V , the Lyapunov drift-plus-penalty becomes more
sensitive to the communication cost. Notice that the optimal
task scheduling decision π ∗ (t) for minimizing the right side
of (18) also minimize D(t) under the queue length stability
constraint. Therefore, we can solve P2 in each time slot t:
P2 : min
s.t.
i=0
t−1
X
Algorithm 1 Optimal Task Scheduling Algorithm Based On
Lyapunov Theory
1: Obtain the queue state Q(t), mobile device state Sl (t),
MEC server state Sc (t) at the beginning of each time
slot t.
2: Find the system case discussed in Section III.
3: Obtain the system case k.
4: Determine π(t) by solving:
5:
min
−Q(t)U (t) + V D(t)
6:
s.t.
(12), and (13)
7: Set t = t + 1 and update Q(t), Sl (t) , Sc (t) according
to (2), (8), (9) respectively.
− Q(t)U (t) + V D(t)
(19)
(12) and (13).
For each time slot t, we can obtain D(t) based on π(t) and
Q(t). Because there are only 5 possible choices for π(t), we
can solve P2 in each time slot t by an enumeration method.
Thus, we propose an optimal task scheduling algorithm based
on Lyapunov theory, as shown in Algorithm 1.
Lemma 2. P1 is not equivalent to P2, but if the control
parameter V is sufficiently large, the solution of P 1 is very
close to P2. Let D̄Alg and D̄Opt be the average transmission
data obtained by solving P2 and the optimal value of P 1,
respectively. We then have:
5
+ D̄Opt .
2V
Proof. Please refer to Appendix B.
D̄Opt ≤ D̄Alg ≤
(20)
B. Tradeoffs between Communications, Computing and
Caching
In this subsection, we reveal the tradeoff between the
average transmission data per task and the computing ability
of the mobile VR device. The tradeoff between the the average
transmission data per task and the cache capacity is also
discussed.
Proof. Due to page limitations, we skip the proof here.
E[Dlt (t)] = τ
N
X
X
220
200
180
160
140
0
5
10
15
20
25
30
35
40
45
50
Cache capacity
(22)
1 − (1 − pn )k Pr(Kt = k). (23)
n=M+1 k
Proof. Due to page limitations, we skip the proof here.
From proposition 2, we can see that D̄lt decreases with the
increase of the caching capacity M in the mobile VR device.
And we can know that D̄Opt decreases with the decrease of
Dlt based on Proposition 1. Hence the increase of M can
decrease DOpt . Similar to proposition 1, with large V , D̄Alg
approaches D̄opt , and D̄Alg decrease with the increase of M .
V. N UMERICAL R ESULTS
In this section, we evaluate the performance of the proposed
optimal scheduling policy by simulations. The number of
contents is N = 1000, the time slot is ∆ = 0.2 s, the length
of each content is τ = 5 M bits, and the average arrival
rate is λ = 0.4. The mobile VR device has fl = 1 GHz
CPU frequency, the cache capacity M = 50 and transmission throughput R = 500M bps, when the MEC server has
fc = 4 × 2.5 GHz CPU frequency unless otherwise specified.
Kt is distributed uniformly in [40, 60]. We assume the content
popularity distribution is identical among all elements of a
task, which follows the Zipf distribution with the parameter
α. We set α = 0.8 in simulations. We consider the MEC
computation policy as baselines, which executes all the tasks
in the MEC server.
Fig.3 shows that the average transmission data per task
achieved by the proposed optimal scheduling policy decreases with the cache capacity. That means the average
communication-resource consumption can be traded off by
the cache capacity to keep the queue length stable, which
verifies the tradeoff presented by Proposition 2. Moreover, the
260
Average transmission data per task (Mbits)
l
Proposition 2. Define K = E[Kt ], and then we have
E[Dct (t)] = D̄ct = τ K,
MEC computation policy
Optimal schedling, W=10
Optimal schedling, W=30
Optimal schedling, W=50
240
Fig. 3. The communication-resource consumption vs. cache capacity.
According to Lemma 2, with large V , D̄Alg approaches
5
+ D̄Opt based on (20). Because N̄l =
D̄opt , and D̄Alg ≤ 2V
E[Nl (t)] = E[⌈Dlc (t)W/(fl ∆) + Dlt (t)/R⌉], for large V , we
can observe from (21) that D̄Alg or D̄opt decreases with the
increase of fl when N̄1 + N̄1c ≥ λ, N̄1 < λ.
l
260
Average transmission data per task (Mbits)
Proposition 1. Let E[Dct (t)] = D̄ct and E[Dlt (t)] = D̄lt
denote the average transmission data of MEC computation
model and local computation model, respectively. And let
E[Dlc (t)] = D̄lc denote the average computation data of local
computation model. We thus have
1
1
1
1
D̄ct − λN̄l (D̄ct − D̄lt ) N̄l + N̄c ≥ λ, N̄l < λ,
1
D̄Opt = D̄lt
≥ λ,
N̄l
1
1
< λ.
+
inf easible
N̄l
N̄c
(21)
where we use E[Nc (t)] = N̄c and E[Nl (t)] = N̄l to denote the
exception of the time slots required to complete a task in MEC
computation model and local computation model, respectively.
240
MEC computation policy
Optimal scheduling, V=0.5
Optimal scheduling, V=0.01
220
Optimal DOpt
200
180
160
140
50
100
150
200
250
300
350
400
450
MD CPU frequency (MHz)
Fig. 4. The communication-resource consumption vs. the computing ability
of mobile VR device.
scheduling policy always outperforms the MEC computation
policy even when there are no contents cached in the mobile
VR device. This is because the optimal scheduling always
executes a part of tasks by local computation policy, and the
redundant transmission of contents needed in those tasks can
be avoided. However, the MEC computation policy delivers
all the contents requested by each task.
Fig.4 presents the average transmission data per task versus
the mobile VR device computing ability fl . The increase
of the mobile VR device computing ability fl decrease the
average communication-resource consumption. The reason for
it is that the increase of fl decreases the computing delay
of local execution, and more tasks will be executed by local
computation policy. With large V , the system is more sensitive
to the communication-resource consumption, and more tasks
are scheduled to mobile VR device. And DAlg is close to
DOpt when V is sufficiently large, which verify the lemma
2. Further, when fl is sufficiently large, the local computation
mode becomes the optimal schedule strategy.
Tradeoff between Communications, Computing and
Caching (3C): In Fig. 5, the average end-to-end latency is
20 ms of the proposed system with different 3C resources
allocation, e.g., {R, fl , M }={125 M bps, 100 M Hz, 5} or
{R, fl , M }={67 M bps, 500 M Hz, 45}. As one can see, the
communication throughput R decreases with increasing com-
B. Proof of Lemma 2
We assume P 1 is feasible, and there exists at least one π ∗ (t)
for satisfying the constraints of P 1. D̄Alg and D̄Opt satisfies
the following condition:
Transmission throughput (Mbps)
180
160
140
E[D(t)|Q(t)] = D̄Alg ≤ D̄Opt + γ,
120
where γ is a positive value. According to Little Theorem [11],
if the average arriving rate is larger than the average service
rate, the queue length tends to infinite with the increase of time
slot t. Therefore, if P 2 can be solved by proposed algorithm,
the following condition should be satisfied
100
80
60
0
0
20
200
40
400
60
600
E[U (t)|Q(t)] = λ + θ,
MD CPU frequency (MHz)
Caching capacity
Fig. 5. The incorporation of communications, computing and caching to
achieve a target delay 20 ms. Here, we set λ = 0.2.
puting capability fl and caching capacity M . As fl increases,
more task be scheduled to the mobile VR device, yielding
lower the communication cost. M is similar to fl . Therefore,
the system needs a smaller R. We observe that the computing
ability has more impact on the communication-resource consumption than that of the caching capacity. This is because
the cache capacity only has impact on the performance of
local computation mode. On the other hand, when the system
has small computing ability and caching capability, the large
transmission throughput is required. This result reconfirm the
importance of simultaneous exploitation of all 3Cs in wireless
networks [2].
VI. C ONCLUSION
In this paper, we investigated the communicationconstrained mobile edge computing systems for wireless virtual reality. A transmission data consumption minimization
problem with the execution delay constraints was formulated,
and we proposed a task scheduling strategy based on Lapunov
theory. The tradeoffs between communications, computing,
and caching in the proposed system was also dicussed. Finally
Simulation results shown that the proposed scheduling strategy
achieve a significant reduction in the average transmission data
consumption.
A PPENDIX
A. Proof of Lemma 1
According to (2), we first have
Q2 (t+1) ≤ Q2 (t)+U 2 (t)+A2 (t)−2Q(t)(U (t)−A(t)). (24)
Substituting (24) into (16), then (16) can be rewritten as
1
E[U 2 (t) + A2 (t)|Q(t)]
2
− Q(t)E[(U (t) − A(t))|Q(t)].
(26)
∆L(Q(t)) ≤
(25)
The task arriving rate A(t) is independent of Q(t). Finally,
according to U (t) ≤ 2 and A(t) ≤ 1, we have Lemma 1.
(27)
where θ is a positive value. Substituting (26) and (27) into
(18), and with γ → 0, we obtain:
5
∆L(Q(t)) + V {DAlg (t)|Q(t)} ≤ + V D̄opt .
(28)
2
Then taking iterated expectation and using the telescoping
sums over t ∈ {1......T }, we get
E[L(Q(T ))] − E[L(Q(1))] + V
T
X
E[DAlg (t)|Q(t)]
t=1
5
≤ T ( + V D̄opt ).
(29)
2
We divide (29) with V T and let T → ∞, then we have:
D̄Alg ≤
5
+ D̄Opt .
2V
(30)
R EFERENCES
[1] E. Bastug, M. Bennis, M. Medard, and M. Debbah, “Toward interconnected virtual reality: Opportunities, challenges, and enablers,” IEEE
Communications Magazine, vol. 55, no. 6, pp. 110–117, 2017.
[2] H. Liu, Z. Chen, and L. Qian, “The three primary colors of mobile
systems,” IEEE Communications Magazine, vol. 54, no. 9, pp. 15–21,
September 2016.
[3] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A
survey on mobile edge networks: Convergence of computing, caching
and communications,” IEEE Access, vol. 5, pp. 6757–6779, 2017.
[4] L. Zhang, Y. Xu, W. Huang, L. Yang, J. Sun, and W. Zhang, “A
mmt-based content classification scheme for vod service,” in 2015
IEEE International Symposium on Broadband Multimedia Systems and
Broadcasting, June 2015, pp. 1–5.
[5] J. Huang, Z. Chen, D. Ceylan, and H. Jin, “6-dof vr videos with a
single 360-camera,” in 2017 IEEE Virtual Reality (VR), March 2017,
pp. 37–44.
[6] Y. Mao, J. Zhang, Z. Chen, and K. B. Letaief, “Dynamic computation
offloading for mobile-edge computing with energy harvesting devices,”
IEEE Journal on Selected Areas in Communications, vol. 34, no. 12,
pp. 3590–3605, Dec. 2016.
[7] T. Q. Dinh, J. Tang, Q. D. La, and T. Q. S. Quek, “Offloading in mobile
edge computing: Task allocation and computational frequency scaling,”
IEEE Transactions on Communications, vol. PP, no. 99, pp. 1–1, 2017.
[8] C. You, K. Huang, H. Chae, and B. H. Kim, “Energy-efficient resource
allocation for mobile-edge computation offloading,” IEEE Transactions
on Wireless Communications, vol. 16, no. 3, pp. 1397–1411, March
2017.
[9] J. Liu, Y. Mao, J. Zhang, and K. B. Letaief, “Delay-optimal computation
task scheduling for mobile-edge computing systems,” in IEEE ISIT, July
2016, pp. 1451–1455.
[10] Y. Jian, Y. Qinghai, K. Kyungsup, and R. Ramesh, “Power-Delay Tradeoff in Wireless Powered Communication Networks,” IEEE Transactions
on Vehicular Technology, July. 2016.
[11] S. Ross, Introduce to probability models. Academic Press, 2014.
| 7cs.IT
|
Span-program-based quantum algorithms for
graph bipartiteness and connectivity ⋆
Agnis Āriņš
arXiv:1510.07825v1 [quant-ph] 27 Oct 2015
University of Latvia, Raiņa bulvāris 19, Riga, LV-1586, Latvia
Abstract. Span program is a linear-algebraic model of computation
which can be used to design quantum algorithms. For any Boolean function there exists a span program that leads to a quantum algorithm
with optimal quantum query complexity. In general, finding such span
programs is not an easy task.
In this work, given a query access to the adjacency matrix of a simple graph G with n vertices, we provide two new span-program-based
quantum algorithms:
√
– an algorithm for testing if the graph is bipartite that uses O(n n)
quantum queries;
√
– an algorithm for testing if the graph is connected that uses O(n n)
quantum queries.
1
Introduction
The concept of a span program as a linear-algebraic model of computation is
not new. It was introduced by Karchmer and Wigderson in 1993 [9] and has
many applications in classical complexity theory. Span programs can be used
to evaluate decision problems. In 2008 Reichardt and Spalek [12] introduced a
new complexity measure for span programs – witness size, which, as Reichardt
showed later in [10,11], has strong connection with the quantum query complexity. There is a quantum algorithm for evaluating span programs [10] and these
two complexity measures are essentially equivalent. The difficulty is to come up
with a span program with a good witness size complexity.
In [12] authors dealt with bounded-size span programs evaluating Boolean
functions each on O(1) bits and posed an open question – do there exist interesting quantum algorithms based directly on asymptotically large span program?
Belovs used span programs to construct learning graphs [4,3]. He also used span
program approach for the matrix rank problem [2]. In [1] Ambainis et al. came
up with a simple yet powerful span program for the graph collision problem.
In this paper, we extend the family of algorithms based on span programs.
√
We present two new span-program-based quantum algorithms
– an O(n n)
√
algorithm for the graph bipartiteness problem and an O(n n) algorithm for the
graph connectivity problem. Both algorithms in the quantum query sense are
⋆
This work has been supported by the ERC Advanced Grant Methods for Quantum
Computing.
optimal because the witness sizes match the quantum query complexity lower
bounds [6,7] for these problems. Thus we demonstrate that span programs can
be useful also for the problems with an asymptotically large input and possibly
our algorithms could be building blocks for bigger span programs in the future.
The graph connectivity
problem has been studied before [7] and there al√
ready exists a O(n n) quantum query algorithm which requires O(n) qubits of
quantum memory. The advantage of our algorithm is that it uses only O(log n)
qubits of quantum memory because the span program P2 uses a vector space
with O(n2 ) dimensions. Similarly for the graph bipartiteness problem. It can be
solved with the breadth-first search method [8] which uses O(n) qubits of quantum memory, but our approach with a span program requires O(log n) qubits of
quantum memory.
2
Preliminaries
In this paper, we present algorithms which work on simple graphs, given in
adjacency model. If the given graph has n vertices then the input size for an
algorithm is n × n and we assume that the input variable xi,j corresponds to the
value of entry in i-th row and j-th column of the adjacency matrix.
2.1
Span programs
Definition 1 ([1]). A span program P is a tuple P = (H, |ti , V ), where H
is a finite-dimensional Hilbert space, |ti ∈ H is called the target vector, and
V = {Vi,b |i ∈ [n], b ∈ {0,
S 1}}, where each Vi,b ⊆ H is a finite set of vectors.
Denote by V (x) = {Vi,b |i ∈ [n], xi = b}. The span program is said to compute function f : D → {0, 1}, where the domain D ⊆ {0, 1}n, if for all x ∈ D,
f (x) = 1 ⇐⇒ |ti ∈ span(V (x)).
Basically, what this definition says is that for each input variable xi we have
two sets of vectors (as the span program authors, we define these vectors in
advance) – Vi,0 and Vi,1 . If xi = b then we say that vectors from the set Vi,b
are available and vectors from the set Vi,1−b are not available. If some vector
is included in both sets Vi,0 and Vi,1 then we say that it is a free vector – it is
always available.
The function returns 1 iff the target vector can be expressed as a linear
combination of the available vectors, otherwise it returns 0.
Definition 2 ([1]).
−1
(1) A positive
P witness for x ∈ f (1) is a vector w = (wv ), v ∈ V (x), such that
|ti = v∈V (x) wv v. The positive witness size is
wsize1 (P ) :=
max
min
x∈f −1 (1) w:witness of x
kwk2 .
(2) A negative witness for x ∈ f −1 (0) is a vector w ∈ H, such that ht|wi = 1
and for all v ∈ V (x): hv|wi = 0. The negative witness size is
X
2
hv|wi .
wsize0 (P ) := max
min
−1
x∈f
(0) w:witness of x
v∈V
(3) The witness size of a program P is
p
wsize(P ) := wsize0 (P ) · wsize1 (P ).
(4) The witness size of a function f denoted by wsize(f ) is the minimum witness
size of a span program that computes f .
Theorem 1 ([10,11]). Q(f ) and wsize(f ) coincide up to a constant factor.
That is, there exists a constant c > 1 which does not depend on n or f such that
1
c wsize(f ) ≤ Q(f ) ≤ c · wsize(f ).
3
Span program for testing graph bipartiteness
A bipartite graph is a graph whose vertices can be divided into two disjoint sets
such that there is no edge that connects vertices of the same set. An undirected
graph is bipartite iff it has no odd cycles.
Algorithm 1. There exists a span program P which √
for a graph with n vertices
detects if the graph is bipartite with wsize(P ) = O(n n).
Proof. We will make a span program which detects if a graph has an odd cycle.
Let n = |G| be a number of vertices in the given graph G. Then the span
program is as follows:
Span program P1 for testing graph bipartiteness
– H is a (2n2 + 1) dimensional vector space with basis vectors {|0i} ∪
{|vk,b i |v, k ∈ [1..n], b ∈ {0, 1}}.
– The target vector is |0i.
– For every k ∈ [1..n] make available the free vector |0i + |kk,0 i + |kk,1 i.
– For every k ∈ [1..n], for every edge u − v (where input xu,v = 1), make
available the vectors |uk,0 i + |vk,1 i and |uk,1 i + |vk,0 i.
The states in the span program P1 are mostly in the form |vk,b i where v is
vertex index, k represents vertex from which we started our search for an odd
length cycle and b represents the parity of the current path length. The first
subindex k in state |vk,b i can also be considered as the subspace index for the
subspace Vk = span({|vk,b i |v ∈ [1..n], b ∈ {0, 1}}). Vectors corresponding to
edges are in the form |uk,b i + |vk,1−b i consisting from sum of two states which
both belong to same subspace Vk .
In the span program P1 the target vector |0i can only be expressed as a linear
combination of the available vectors if at least one of the vectors in the form
|kk,0 i + |kk,1 i can be expressed. Without loss of generality, if there is an odd
length cycle v1 − v2 − · · · − v(2j+1) − v1 then the target vector can be expressed
by taking the vectors corresponding to the edges of this cycle, alternatingly with
plus and minus sign
|0i = (|0i + |11,0 i + |11,1 i) − (|11,0 i + |21,1 i) + · · · − (|(2j + 1)1,0 i + |11,1 i)
therefore the span program P1 will always return 1 when the given graph is not
bipartite.
From the other side, if there is no odd length cycle then none of the vectors
in the form of |kk,0 i + |kk,1 i can be expressed using the available vectors from
P1 . To cancel out the state |kk,0 i we should be using a vector |kk,0 i + |vk,1 i
corresponding to some edge k − v where v is some vertex adjacent to k because
no other vector contains the state |kk,0 i. By doing so we move from the state
|kk,0 i to the state |vk,1 i (possibly with some coefficient other than 1) which has
the parity bit flipped. Similarly, to cancel out the state |vk,1 i we should be using
a vector corresponding to some edge going out from vertex v. To stop this process
we need to reach the state |kk,1 i. It can be done only if there is an odd cycle
because the path must be closed and the parity bit restricts it to odd length.
When there is no odd cycle, span program P1 will always return 0.
We can conclude that P1 indeed computes the expected function. It remains
to calculate the witness size of P1 .
For the case when there is an odd cycle we need to calculate the positive
witness size. If there is an odd cycle v1 − v2 − · · · − vd − v1 with length d then
the target vector can be expressed in this way
|0i = 1 · (|0i + |11,0 i + |11,1 i) + (−1) · (|11,0 i + |21,1 i) + · · · + (−1) · (|d1,0 i + |11,1 i)
and the positive witness w here consists only from d + 1 entries ±1 therefore
kwk2 = d + 1.
If v1 − v2 − · · · − vd − v1 is a cycle then also v2 − v3 − · · · − vd − v1 − v2 is a
cycle and therefore the target vector can also be expressed in this way
|0i = (|0i + |22,0 i + |22,1 i) − (|22,0 i + |32,1 i) + · · · − (|12,0 i + |22,1 i)
the same follows for all d vertices in this cycle and the target vector therefore
can be expressed in atleast d different ways. We can combine these d ways each
taken with coefficient 1/d and then we get that the positive witness size
wsize1 (P1 ) ≤ d ∗ (1/d)2 ∗ (d + 1) < 2
(1)
To estimate the negative witness size we must find a negative witness w′ . We
derive w′ by defining how it acts on basis vectors. From definition hw′ |0i = 1.
For every k we must have hw′ |(|0i + |kk,0 i + |kk,1 i)i = 0 therefore lets pick w′
in such a way that hw′ |kk,0 i = 0 and hw′ |kk,1 i = −1. Now repeat the following
steps until no changes happen:
– for every available vector |uk,0 i + |vk,1 i if hw′ |uk,0 i is defined then define
hw′ |vk,1 i = − hw′ |uk,0 i.
– for every available vector |uk,1 i + |vk,0 i if hw′ |uk,1 i is defined then define
hw′ |vk,0 i = − hw′ |uk,1 i.
For all not yet defined hw′ |vk,b i define hw′ |vk,b i = 0.
For any given vector v in span program P1 the value hw′ |vi2 ≤ 1. The total
number of vectors does not exceed n + n3 therefore the negative witness size is
wsize0 (P1 ) ≤ 1 · (n + n3 )
(2)
Combining positive and negative witness sizes we obtain the upper bound
for witness size which also corresponds to quantum query complexity
p
√
(3)
wsize(P1 ) = wsize0 (P1 ) · wsize1 (P1 ) = O n n
4
Span program for testing graph connectivity
A graph is said to be connected if every pair of vertices in the graph is connected.
If in an undirected graph one vertex is connected to all other vertices then by
transitivity the graph is connected.
Algorithm 2. There exists a span program P which for
√ a graph with n vertices
detects if the graph is connected with wsize(P ) = O(n n).
Proof. Let n = |G| be a number of vertices in the given graph G. Then the span
program is as follows:
Span program P2 for testing graph connectivity
– H is a n2 − 1 dimensional vector space with basis vectors {|vk i |v ∈
[0..n], k ∈ [2..n]}.
– The target vector is |ti = |02 i + |03 i + · · · + |0n i.
– For every k ∈ [2..n] make available the free vector |0k i + |1k i − |kk i.
– For every k ∈ [2..n], for every edge u − v (where u, v ∈ [1..n] and input
xu,v = 1), make available the vector |uk i − |vk i.
If all vertices are reachable from vertex with index 1 then the given graph is
connected. Here we use Belov’s [5] span program for s-t connectivity as subroutine. This subroutine checks if in a given graph there is a path from the vertex
s to the vertex t. The span program for it has the target vector |si − |ti and for
each edge i − j (input xi,j = 1) we can use the vector |ii − |ji.
In span program P2 , by using this subroutine n−1 times, we check if all other
vertices are connected to vertex with index 1. We create a separate subspace Vk =
span({|vk i |v ∈ [0..n]}) for each such subroutine call to avoid any interference
between them, which is a common technique [10] how to compose span programs.
The span program returns 1 when all vertices are connected, but otherwise it
returns 0.
For the case when the given graph is connected we need to calculate the
positive witness size. In each s-t subroutine the shortest path length from the
vertex s to the vertex t can not be larger than n − 1. Threfore each vector from
the set {|0k i |k ∈ [2..n]} requires no more than n vectors to express it. There are
n − 1 such subroutines. The positive witness size is
wsize1 (P2 ) ≤ n · (n − 1) ≤ n2
(4)
To estimate the negative witness size we must find a negative witness w′ . We
derive w′ by defining how it acts on the basis vectors. From definition hw′ |ti = 1.
We need to talk about negative witness only when some vertex v is not connected
to vertex with index 1. Then the vertex v belongs to different connected component than vertex with index 1. Lets name this connected component Cv and let
the count of vertices in this connected component be dv . Lets pick w′ in such a
way that for each vertex vk ∈ Cv set hw′ |0k i = 1/dv and for each vertex vz ∈
/ Cv
set hw′ |0z i = 0.
For k ∈ [2..n] we must have hw′ |(|0k i + |1k i − |kk i)i = 0 therefore set hw′ |1k i =
− hw′ |0k i and hw′ |kk i = 0. Now repeat the following step until no changes
happen: for every available vector |uk i − |vk i if hw′ |uk i is defined then define
hw′ |vk i = hw′ |uk i. For all other not yet defined basis vectors |vk i set hw′ |vk i = 0.
With such negative witness w′ choice the overall negative witness size will
only get increased by vectors which correspond to nonexistent graph edges which
connects Cv with other connected components in graph - i.e. border edges. An
edge u−v is a border edge if u ∈ Cv and v ∈
/ Cv . To a border edge u−v correspond
the vectors |uk i − |vk i where k ∈ [2..n] but only dv from these vectors will have
hw′ |uk i =
6 hw′ |vk i and each such vector increases the negative witness size by
value (1/dv )2 . For Cv there are at most dv · (n − 1) border edges therefore the
overall negative witness size is
wsize0 (P2 ) ≤ d2v · (n − 1) · (1/dv )2 ≤ n
(5)
Combining the positive and negative witness sizes we obtain the upper bound
for the witness size which also corresponds to the quantum query complexity
p
√
(6)
wsize(P2 ) = wsize0 (P2 ) · wsize1 (P2 ) = O n n
Acknowledgements
I am grateful to Andris Ambainis for the suggestion to solve the graph problems
with span programs, and for many useful comments during the development of
the paper.
References
1. Ambainis, A., Balodis, K., Iraids, J., Ozols, R., Smotrovs, J.: Parameterized
quantum query complexity of graph collision. CoRR abs/1305.1021 (2013),
http://arxiv.org/abs/1305.1021
2. Belovs, A.: Span-program-based quantum algorithm for the rank problem. CoRR
abs/1103.0842 (2011), http://arxiv.org/abs/1103.0842
3. Belovs,
A.:
Learning-graph-based
quantum
algorithm
for
kdistinctness.
In:
IEEE
53rd
Annual
Symposium
on
Foundations of Computer Science (FOCS). pp. 207–216. IEEE (2012),
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=6375298
4. Belovs, A.: Span programs for functions with constant-sized 1-certificates. In: Proceedings of the 44th symposium on Theory of Computing. pp. 77–84. ACM (2012),
http://dl.acm.org/citation.cfm?id=2213985
5. Belovs, A., Reichardt, B.: Span programs and quantum algorithms for stconnectivity and claw detection. In: Epstein, L., Ferragina, P. (eds.) Algorithms –
ESA 2012, Lecture Notes in Computer Science, vol. 7501, pp. 193–204. Springer
Berlin Heidelberg (2012), http://dx.doi.org/10.1007/978-3-642-33090-2_18
6. Berzina, A., Dubrovsky, A., Freivalds, R., Lace, L., Scegulnaja, O.: Quantum
query complexity for some graph problems. In: Van Emde Boas, P., Pokorný, J.,
Bieliková, M., Štuller, J. (eds.) SOFSEM 2004: Theory and Practice of Computer
Science, Lecture Notes in Computer Science, vol. 2932, pp. 140–150. Springer Berlin
Heidelberg (2004), http://dx.doi.org/10.1007/978-3-540-24618-3_11
7. Dürr, C., Heiligman, M., Høyer, P., Mhalla, M.: Quantum query complexity of some graph problems. In: Dı́az, J., Karhumäki, J., Lepistö, A., Sannella, D. (eds.) Automata, Languages and Programming, Lecture Notes in
Computer Science, vol. 3142, pp. 481–493. Springer Berlin Heidelberg (2004),
http://dx.doi.org/10.1007/978-3-540-27836-8_42
8. Furrow, B.: A panoply of quantum algorithms. Quantum Info. Comput. 8(8), 834–
859 (Sep 2008), http://dl.acm.org/citation.cfm?id=2017011.2017022
9. Karchmer, M., Wigderson, A.: On span programs. In: Proceedings of the Eighth
Annual Structure in Complexity Theory Conference. pp. 102–111. IEEE (1993),
http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=336536
10. Reichardt, B.W.: Span programs and quantum query complexity: The general adversary bound is nearly tight for every boolean function. In: 50th Annual IEEE
Symposium on Foundations of Computer Science (FOCS). pp. 544–551. IEEE
(2009), http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=5438598
11. Reichardt, B.W.: Reflections for quantum query algorithms. In: Proceedings of
the Twenty-Second Annual ACM-SIAM Symposium on Discrete Algorithms. pp.
560–569. SIAM (2011), http://dl.acm.org/citation.cfm?id=2133080
12. Reichardt, B.W., Spalek, R.: Span-program-based quantum algorithm for evaluating formulas. In: Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing. pp. 103–112. STOC ’08, ACM, New York, NY, USA (2008),
http://doi.acm.org/10.1145/1374376.1374394
| 8cs.DS
|
1
Decentralized Hypothesis Testing in
Energy Harvesting Wireless Sensor Networks
arXiv:1605.03316v2 [cs.IT] 10 Jun 2017
Alla Tarighati, James Gross, Senior Member, IEEE, and Joakim Jaldén, Senior Member, IEEE
Abstract—We consider the problem of decentralized hypothesis
testing in a network of energy harvesting sensors, where sensors
make noisy observations of a phenomenon and send quantized
information about the phenomenon towards a fusion center. The
fusion center makes a decision about the present hypothesis
using the aggregate received data during a time interval. We
explicitly consider a scenario under which the messages are sent
through parallel access channels towards the fusion center. To
avoid limited lifetime issues, we assume each sensor is capable
of harvesting all the energy it needs for the communication from
the environment. Each sensor has an energy buffer (battery) to
save its harvested energy for use in other time intervals. Our
key contribution is to formulate the problem of decentralized
detection in a sensor network with energy harvesting devices. Our
analysis is based on a queuing-theoretic model for the battery
and we propose a sensor decision design method by considering
long term energy management at the sensors. We show how
the performance of the system changes for different battery
capacities. We then numerically show how our findings can be
used in the design of sensor networks with energy harvesting
sensors.
I. I NTRODUCTION
Distributed detection problem formulations have traditionally addressed detection in sensor networks by considering
network performance measures like error probability and receiver operating characteristic [1]. In these setups, spatially
separated sensors make observations of the same phenomenon
and send a summary of their observations towards a fusion
center (FC) through rate-constrained channels. Each sensor can
be viewed as a quantizer which quantizes its observation, and
according to the network arrangement, sends its output either
to another sensor or to the FC. In many applications, sensors
send their outputs to the FC through a multiple access channel
[2]–[4] or through parallel access channels, commonly known
as the parallel topology [5], [6]. A survey of early works on
decentralized hypothesis testing in wireless sensor networks
can be found in [7]–[9].
A large number of sensors with small batteries and limited
life-time are often used in wireless sensor networks. A major
limitation of these sensors is their finite lifetime. In other
words, the sensors work as long as their battery last and
this implies that also the network has a limited lifetime.
Many solutions to increase the lifetime of battery-powered
sensor nodes have been proposed, see [10]–[12] and references
therein. While in all of these methods the aim is to find an
energy usage strategy to maximize the lifetime of a network,
the lifetime remains bounded and finite. An alternative way of
dealing with this problem is to use energy harvesting devices
at the sensor nodes. An energy harvesting device is capable of
acquiring energy from nature or from man-made sources [13],
[14].
Energy harvesting technologies provide a promising future
for wireless sensor networks, such as self sustainability and an
effectively perpetual network lifetime which is not limited by
the sensor battery lifetime [14]–[16]. While acquiring energy
from the environment makes it possible to deploy wireless
sensor networks in situations which are impossible using
conventional battery-powered sensors, it poses new challenges
related to the management of the harvested energy. These
new challenges are due to the fact that the amount of energy
available at a sensor is random, since the source of energy
might not be available at all times we may want to use the
sensor nodes.
We address the problem of detection in networks of sensors
arranged in parallel, where each sensor is an energy harvesting
device. At each time t = 1, 2, . . . the sensors send a message
towards the FC about the state of the current hypothesis Ht and
the FC makes a decision about the hypothesis at that time. The
sensors communicate with the FC using energy asymmetric
on-off keying (OOK), as a low communication rate scheme
for distributed detection applications [17], where a positive
message can be sent at the cost of one unit of energy, and a
negative message is conveyed through a non-transmission at no
cost in energy [18], [19]. It was previously shown in [20] that
OOK is the most energy efficient modulation scheme under
Rayleigh fading non-coherent transmission, though we are
not primarily concerned with explicitly modeling the fading
channels between the sensors and the FC herein. We assume
that each sensor is equipped with an internal battery and is
allowed to use a long term energy conversion policy. We
assume the observations at the sensors are, conditioned on
the true hypothesis, independent and our goal is to design
the sensors’ transmission rules in such a way that the error
probability at each time instance t, at the FC, is minimized.
To this end, we use the Bhattacharyya distance between the
conditional distributions at the FC input, which has been
frequently used in the past as a performance measure in the
design of distributed detection systems [5], [21], [22].
Using the Bhattacharyya distance, essentially reduces the
joint design of decision rules at the sensors to the design
of decision rules at single sensors. Doing so may risk not
capturing all the aspects of the joint design of the decision
rules at peripheral sensors. However, due to the analytical
tractability that follows from this choice of design rule, the
Bhattacharyya distance has been frequently considered as a
performance metric before in the literature [5].
The novelty of our work is in the formulation of a de-
2
un,t
0
Ht ∈ {0, 1}
1 − 0
yn,t
0
0
1
x1,t
S1
xN,t
xn,t
···
u1,t
···
Sn
un,t
y1,t
yn,t
SN
1
1 − 1
1
Fig. 2. Binary asymmetric channels between sensors and the FC.
uN,t
yN,t
Fusion Center
bt
H
harvesting sensor, and formulate the problem. In Section III
we study the performance of an energy harvesting sensor with
different battery capacities. In Section IV we will illustrate
our results in the design of sensor networks by presenting
numerical simulations, and finally in Section V we conclude
the paper.
Fig. 1. Decentralized hypothesis testing scheme in a parallel network.
centralized detection problem with system costs coupled to
the random behavior of the energy available at the sensors.
Concretely, we will find the depletion probability at the sensor
batteries, and evaluate the performance of the network for different battery capacities (buffer sizes). We will show how the
problem formulation changes (compared to the unconstrained
case) when we consider the energy features in the problem of
designing the sensors in the network.
Distributed inference using energy harvesting agents has
gained a lot of interest during the last years (see [23]–
[32] and references therein). In the context of distributed
estimation, Nayyar et al. studied the structure of optimal
communication scheduling [23], Zhao et al. [24], Huang et al.
[25], and Nourian et al. [26] studied optimal power allocation
schemes, while Li et al. [27] studied the performance of
an energy harvesting relay-aided cooperative network under
fading channels. Hong [28] considered the case where each
peripheral node sends a re-scaled version of its observation
towards a FC, if the required energy for that transmission is
available at the sensor; otherwise the sensor remains silent.
This non-transmission also conveys information regarding the
magnitude of the observed signal. Hua Liu et al. [29] considered adaptive quantization is distributed estimation using a
game-theoretic approach. Energy harvesting in relay systems
was also considered in [30], [31], where the nodes (or relays)
harvest the energy they need to transmit their amplified received messages towards a receiver. Medepally and Mehta [31]
considered a relay selection scheme where if multiple active
relays are available, one of them is selected to transmit. In
the context of distributed detection, however, finding optimal
decision rules at the remote energy harvesting sensors is
largely open. In [32] we studied the structure of decision
rules at agents with infinite battery size and in error-free
communication channels, while in this paper we shall extend
the results to a more general case of erroneous communication
channels.
The outline of this paper is as follows. In Section II we
describe the structure of a parallel network and an energy
II. P RELIMINARIES
In this section we first present the system model. We then
define an energy harvesting sensor and formulate the problem
of designing the energy harvesting sensors in the network.
Limitations of the proposed system model, and their influence
on the obtained results, are further discussed in the concluding
remarks in Section V.
A. System Model
We consider a decentralized hypothesis testing problem
where N sensors S1 , . . . , SN are arranged in parallel, according to Fig. 1. During each time interval t (defined as [t, t + 1))
each sensor Sn , n ∈ {1, . . . , N }, makes an observation
xn,t from the same phenomenon and sends a message un,t
towards the FC. Note that xn,t is just assumed to be an
element of an abstract space, so it could be a single scalar
measurement or a vector of measurements made during time
interval t at sensor Sn . We consider the case where different
observations at the sensor Sn , conditioned on the hypothesis,
are independent and identically distributed and Xn is a random
variable corresponding to observations at sensor Sn .
At each time interval the phenomenon Ht is modeled as
a random variable drawn from a binary set {0, 1} with apriori probabilities π0 and π1 , respectively, and gives rise
to conditionally independent observations Xn ∈ Xn with
conditional distribution fXn |Ht (xn,t |ht ) for ht ∈ {0, 1} at the
sensors. In this paper, we assume that the present hypothesis
Ht changes over time in an i.i.d. fashion while it is fixed during
each time interval. We also assume that the sensors are allowed
to use a long term energy usage policy managed together with
an internal battery. In other words, we consider the long term
behavior of a sensor when its battery state operates in steady
state.
The communication channels between the sensors and the
FC are one-way links from the sensors, and there is no
communication between the sensor nodes. The sensors communicate with the FC using an energy asymmetric on-off
keying, where a positive message (labeled by “1”) is conveyed
by transmission of a message and a negative message (labeled
by “0”) is conveyed by a non-transmission. We model the
3
channel between sensor Sn and the FC by the most general
binary discrete memoryless channel that is the so called binary
asymmetric channel (BAC) [33], where the 0-to-1 error occurs
with probability 0 < 0.5 and the 1-to-0 error occurs with
probability 1 < 0.5, as shown in Fig. 2, where yn,t ∈ {0, 1}
is the received message at the FC corresponding to the sensor
output un,t . When 0 = 1 the channel will be a binary
symmetric channel (BSC) and when 0 = 0, the channel will
reduce to the Z-channel. It is important to stress that although
we consider the sensor-to-FC channel to be unreliable, we
also consider it to be fixed and not subject to alteration by
the sensors, i.e., the optimization of the sensor decision rules
does not encompass the design of the over the air transmission
protocol.
The BAC is also a relevant high-level model for the case
where each sensor-to-FC channel is a fading channel and the
FC (as a non-coherent receiver) uses an energy detector to
detect its input yn,t , i.e., where
1 En,t > Γ
yn,t =
0 Otherwise ,
where En,t is the energy received at the FC resulting from
sensor un,t and Γ is some positive threshold [18]. For OOK
transmission over a fading channel with an energy threshold
detector at the FC, it may be reasonable to assume that 1
0 , reflecting the belief that deep fades causing the received
energy to not meet the specified threshold are more likely than
noise induced fluctuations that push the decision metric over
the threshold boundary.
The overall problem considered is structurally similar to
the classical binary hypothesis testing problem over one bit
channels. In other words, each sensor Sn , given a realization
xn,t of Xn , computes a message un,t ∈ {0, 1} using its
decision function γn : X → {0, 1}, i.e.,
γn (xn,t ) = un,t ,
and sends this message towards the FC. The FC, based on the
aggregate received BAC channel outputs y t , (y1,t , . . . , yN,t ),
makes a decision b
ht ∈ {0, 1} (with corresponding random
b t ) about the present hypothesis at each time interval,
variable H
using its decision function γ0 : {0, 1}N → {0, 1}, i.e.,
γ0 (y t ) = b
ht .
Note that the FC does not directly have access to the sensor
outputs ut , (u1,t , . . . , uN,t ), while it has access to the
corresponding BAC channel outputs y t .
The overall objective in this paper is to design the decision
functions γn , for n = 0, . . . , N , of the FC and the sensors
in such a way that some performance measure is optimized.
This problem was considered extensively in the classical
distributed detection literature [1], [5], [6] when energy is
always available at the sensors to send their messages and the
sensor-to-FC channels are error-free channels. However, the
problem of designing sensors’ decision functions when each
sensor is an energy harvesting device is largely open. We shall
herein consider the problem of designing decision functions of
the sensors when each sensor is an energy harvesting device.
In what follows, we first define our performance metric and
then, in Section II-B, we define an energy harvesting device.
Let Un be a random variable corresponding to output
messages of sensor Sn . Then, due to the independence of
observations, the conditional PMF associated with the message
vector U , (U1 , . . . , UN ) can be obtained according to
N Z
Y
PU |Ht (u|ht ) =
fXn |Ht (x|ht ) dx , (1)
n=1
−1
x∈γn
(un )
where u , (u1 , . . . , uN ), and where γn−1 (un ) is the set of
observations x ∈ X that satisfy γn (x) = un .
Let BTot,t be the total Bhattacharyya distance (BD) at the
FC at time instance t, and for given sensor decision functions
γ1 , γ2 , . . . , γN :
X q
PY |Ht (y t |0)PY |Ht (y t |1) ,
BTot,t = − log
y ∈{0,1}N
t
where Y , (Y1 , . . . , YN ) and Yn is a random variable
corresponding to BAC channel output
yn,t . The Bayesian error
b
probability PE,t , Pr Ht 6= Ht is minimized when the FC
applies the maximum a-posteriori (MAP) rule [34] and the
corresponding error probability can be upper bounded by the
total BD according to [35]
√
PE,t ≤ π0 π1 e−BTot,t .
In the context of distributed hypothesis testing, it has been
acknowledged that the Bayesian error probability criteria in
most cases does not lead to a tractable design procedure for
decision functions, even for small sized networks. This led
authors to consider the Bhattacharyya distance as a performance measure of the network [5], [21], [22]. In addition,
when the observations at the sensors, conditioned on the true
hypothesis, are independent, the total Bhattacharyya distance
at the FC simplifies greatly and decouples [22], i.e.,
BTot,t =
N
X
− log
n=1
=
N
X
X q
PYn |Ht (yn,t |0)PYn |Ht (yn,t |1)
yn,t =0,1
Bn,t .
n=1
In other words, at each time instance t, the total Bhattacharyya
distance at the FC of a parallel network is the summation of all
the Bhattacharyya distances delivered from different sensors.
Therefore, a network which maximizes the Bhattacharyya
distance at the FC is a network with individually optimized
sensors. Then, it suffices to maximize the BD received at the
FC from each sensor, separately. From now on we will focus
on a single energy harvesting sensor S and drop the subscript
n and denote its BD at time t by Bt . Throughout this paper,
we interchangeably use the terms “Bhattacharyya distance”
and “BD”.
B. Energy Harvesting Sensors
Consider an energy harvesting sensor S as in Fig. 3.
Let us assume that during time interval t energy et arrives
4
for a given Θ, where
xt
et
K
S
bt
wt
bt
wt
t
ut
et
l(xt ) , ln
bt+1
t+1
Fig. 3. Model of an energy harvesting node (left) and an example of a time
interval t (right).
stochastically at the node as a stationary and ergodic random
process Et , see [14]. Let bt be the battery state and let Bt
be its corresponding random process, which is in general
a correlated random process over time even when energy
harvesting process Et is i.i.d. Note that the actions of the
sensor affect the future of the battery state, and the sensor
knows the battery state. Assume that the battery size (capacity)
is K. Then the amount of available energy in the battery at
transmission time t + 1 is (cf. [13])
bt+1 = min {bt − wt + et , K} ,
(2)
where wt is the amount of energy used to send the message
ut at time t, with the corresponding random variable Wt . We
assume that the arrival energy et during time interval t can not
be used at the same time interval (see Fig. 3). The same as
in [36]–[39], we assume energy arrives in packets and at each
time interval the sensor is capable of harvesting at most one
packet of energy. We also assume et ∈ {0, 1} is drawn from
a Bernoulli distribution, with Pr(et = 1) = pe . We further
assume that only sending a message costs a packet of energy,
and the energy of making the observation and processing
is negligible. These assumptions, while rather simplistic, are
repeatedly used in the literature (see [36]–[39] and references
therein) for an energy harvesting system. We adopt this model
for the sake of analytical tractability and insight. Thus, wt = 1
if ut = 1 was sent during time interval t, otherwise wt = 0,
i.e.,
(
1 ut = 1 ,
wt = wt (ut ) =
0 ut = 0 .
As noted before, the received BD at the FC from sensor S at
time t is
"
#
X q
PY |Ht (yt |0)PY |Ht (yt |1) ,
(3)
Bt = − log
yt =0,1
is the log-likelihood ratio for a given observation xt . Let
lt , l(xt ) and let L , l(X). It was shown in [40] that, for
the unconstrained case1 , and error-free sensor-to-FC channels,
likelihood-ratio quantizers are optimal decision rules at the
local sensors. Chen and Willett [41] then generalized this result
for the case of non-ideal sensor-to-FC channels. They have
shown that likelihood-ratio thresholds are optimal for nonideal channels as well (see [41], [42] for the conditions under
which this result holds). Furthermore, for scalar observations
xt , when the likelihood-ratio of the observation at a sensor S is
monotone and increasing in xt , the likelihood-ratio threshold
can be directly translated to the sensor observation. This is an
immediate consequence of the Karlin-Rubin theorem [43].
Throughout this paper, we denote the optimal threshold of
an energy unconstrained sensor by Θ?u . In other words, Θ?u is
the threshold in (5) which maximizes (3), i.e.,
hp
[0 + δq0 ][0 + δq1 ] +
Θ?u = arg max − log
Θ
i
p
[1 − 0 − δq0 ][1 − 0 − δq1 ]
(6)
where
qh , Pr (L ≥ Θ|Ht = h) ,
δ , 1 − 0 − 1 ,
for h = 0, 1. Note that q0 and q1 are the false alarm and the
detection probability of an unconstrained sensor, respectively.
We say that the observation model at the sensor is separable,
or perfect, if there exists a threshold Θ for which
q0 = 0 ,
lim q0 = 0 ,
PY |Ht (yt = i|h) =PU |Ht (ut = i|h)(1 − i )+
PU |Ht (ut = 1 − i|h)1−i .
(4)
We say a sensor decision function γ is a likelihood-ratio
quantizer (or likelihood-ratio threshold) if
1 l(xt ) ≥ Θ,
ut = γ(xt ) ,
(5)
0 Otherwise ,
q1 = 1 .
(7)
In this situation, for an unconstrained sensor and error-free
channel, the BD will be infinity and the detection problem is
trivial. However, in real-world applications such observation
models do not exist and the observation models are nonseparable, i.e., there is no threshold Θ for which the conditions
in (7) are both satisfied. However, such a condition may hold
asymptotically in the high SNR limit. Thus, let us define E as
the signal-to-noise ratio (SNR) at the sensor. We say that the
observation model at the sensor is asymptotically separable
if, when SNR goes to infinity, for a sensor with non-separable
observations there exists a threshold Θ for which
E→∞
where
fX|Ht (xt |1)
fX|Ht (xt |0)
lim q1 = 1 ,
E→∞
(8)
and therefore the corresponding BD for the unconstrained sensor and error-free channel goes to infinity, limE→∞ Bt = ∞.
Many observation models that are considered in the literature
are asymptotically separable and in the following we will
introduce one of these observation models. Consider the case
where each observation is from a Rayleigh distribution with
1 By unconstrained we mean the situation where energy is always available
at the sensor.
5
scale parameter σ0 or from a Rician distribution with scale
parameter σ1 and noncentrality parameter s. The conditional
distributions at the sensor are therefore
x
x2
fX|Ht (x|0) = 2 exp − 2 , and
σ0
2σ
20 2
(9)
x
x +s
xs
fX|Ht (x|1) = 2 exp −
I0
,
σ1
2σ12
σ12
where I0 (z) is the modified Bessel function of the first kind
with order zero. This is a relevant observation model for low
complexity sensors in a wireless sensor network used to detect
the presence of a known signal in Gaussian noise based on the
received power. For simplicity, we assume σ0 = σ1 = 1 and
by definition E , s2 . Now, letting SNR go to infinity, there is a
threshold Θ under which the conditions in (8) are satisfied, and
therefore the observation model is asymptotically separable.
We shall herein assume that the energy constrained sensor S
is also a single-threshold likelihood-ratio quantizer that applies
the following threshold test:
1 lt ≥ Θ, bt > 0,
ut = γ(xt , bt ) ,
(10)
0 Otherwise.
In other words, a sensor S at each time t compares the
likelihood-ratio lt of its observation xt with a threshold Θ.
If lt ≥ Θ and the battery is not empty bt > 0, it sends
a message towards the FC, otherwise it remains silent.2 We
would like to note that, the FC in this setup is a static device,
in the sense that, it makes decision at each time t only based
on its input messages at the same time, and it does not use
previous input messages from the sensors. In other words, the
FC does not take into account any correlation in the received
messages from a sensor that could (possibly) be introduced
by the memory of the battery at that sensor. In what follows
we will find a threshold Θ? which maximizes the delivered
Bhattacharyya distance from a sensor to the FC. We will show
that for the problem at hand an optimum decision rule not
only depends on the observation model, it also depends on the
battery charge state, arrival energy features and the capacity
of the battery K. We further show that the resulting thresholds
in general differ from those of the unconstrained case.
In the following section, for different battery capacities, we
study the depletion probability of an energy harvesting sensor.
Furthermore, we will show how one can design an energy
harvesting sensor for different battery capacities. Note that by
designing energy harvesting sensors we mean the selection of
the decision threshold Θ in (10).
III. D ESIGN OF E NERGY H ARVESTING S ENSORS
To formulate the problem, let us first find conditional mass
probabilities resulting from a sensor decision ut ∈ {0, 1}.
PU |Ht (ut = 1|h) = Pr (L ≥ Θ ∩ Bt > 0|Ht = h)
= Pr (L ≥ Θ|Ht = h) Pr (Bt > 0) (11)
= qh [1 − Pr (Bt = 0)]
From (3), (4), and (11) we observe that the BD of an
energy harvesting sensor at time t depends on the observation
2 Note
that the condition bt > 0 is a consequence of wt = 1 assumption.
distribution at the sensor through qh and the battery depletion probability Pr (Bt = 0). Under the assumptions that the
energy harvesting probability is i.i.d. Bernuoulli over time
and space, and has a probability pe , and the observations at
the sensor are i.i.d. in time, we say that, the battery has a
Markovian behavior in the sense that, conditioned on Et and
Wt , its state at time t + 1 (i.e., Bt+1 ) only depends on its
state at time t (i.e., Bt ) and not the sequence of previous
states, {Bt0 }t−1
t0 =0 . Under the Markovian assumption, a steady
state probability for the battery charge state can be derived,
which allows us to consider the long term performance of an
energy harvesting sensor equipped with a battery. Let the state
probability vector of the battery charge at time t be
|
pt , Pr (Bt = 0) , . . . , Pr (Bt = K) ,
where the superscript | indicates transposition, and let us
define the transition probability matrix as
P = pi,j K×K ,
where K = K + 1 and pi,j , Pr(Bt = j|Bt−1 = i) , for
i, j = 0, . . . , K. When the battery is in steady state [44], the
state probability vector will satisfy
p∞ = P| p∞ ,
(12)
and the steady state probability of each battery state, i.e.,
pj , lim Pr(Bt = j),
t→∞
can be found using (12) and the fact that the summation of
the probabilities must equal unity.
In steady state, after dropping the subscript t, the conditional
probabilities in (11) are given by
PU |H (u = 1|h) = qh (1 − p0 ) ,
PU |H (u = 0|h) = 1 − qh (1 − p0 ) .
(13)
We can plug (13) into (4) and find the resulting Bhattacharyya
distance in (3) as
hp
[0 + δq0 (1 − p0 )][0 + δq1 (1 − p0 )] +
B = − log
i
p
[1 − 0 − δq0 (1 − p0 )][1 − 0 − δq1 (1 − p0 )] .
(14)
We observe that the resulting BD for the energy constrained
case (in steady state) depends on the depletion probability p0
of the battery, which itself is a function of energy features
and the battery capacity. Therefore, in the following we study
the performance of an energy harvesting sensor for different
battery capacities, K. To this end, for an arbitrary battery
capacity K, in Lemma 1, we will find the depletion probability
p0 . Then we can find a threshold Θ? which maximizes (14).
To this end, consider a K-slot-battery energy harvesting
sensor S. Assume that at transmission time t, its battery charge
is in state bt = k, where 0 < k < K. Since during each time
interval the sensor is capable of harvesting and consuming at
most one packet of energy, its state at transmission time t + 1
will be either bt+1 = k − 1, k or k + 1. Two exceptions to
this rule are when the battery charge is in state bt = 0 or in
state bt = K. In the former case the state of the battery at
6
0.5
0.14
rS
0.45
rS
rS
rS
rS
rS
bC
bC
0.25
bC
bC
bC
bC
bC
bC
bC
0.3
rS
bC
bC
rS
0.2
bC
0.15
bC
0.1
bC
rS
rS
bC
0
Using
Using
Using
Using
rS
rS
bC
rS
0.05
bC
bC
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
K
K
K
K
=2
=2
=1
=1
3
4
5
rS
bC
bC
bC
bC
6
7
8
rS
0.08
bCbC
bC
bC
rS
bC
0.06
bC
rS
bC
0.04
bC
Using
Using
Using
Using
rS
rS
0.02
bCrS
bC
bC
bC
rS
rS
2
rS
rS
0.1
rS
bCbCrS
1
rS
rS
rS
Bhattacharyya distance
0.35
rS
rSrS
0.12
rS
rS
0.4
Bhattacharyya distance
rS
rSrS
rS
9
10
0
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
K
K
K
K
=2
=2
=1
=1
bCbCrS
1
2
3
4
5
6
7
8
9
10
Non-centrality parameter s
Non-centrality parameter s
Fig. 4. Bhattacharyya distance of single- and double-slot-battery (K = 1, 2) energy harvesting sensors for different threshold tests and for noiseless (left)
and noisy (right) communication channels, for π1 = 0.2, pe = 0.15, as a function of non-centrality parameter.
transmission time t + 1 will be either zero or one, since the
battery charge can not be negative. In the latter case, the state
of the battery at transmission time t + 1 will be either K − 1
or K, since there is no space to save more energy packets.
Lemma 1: The depletion probability of a K-slot-battery
energy harvesting sensor is given by3
"
#−1
K
1 X k
p0 = 1 +
Ω
(15a)
1−q
k=1
pe − q
=
,
pe ΩK − q
where
Ω,
(15b)
pe (1 − q)
,
q(1 − pe )
and q , π0 q0 + π1 q1 .
Proof. See Appendix A.
Remark: The depletion probability of a K-slot-battery energy harvesting sensor (when pe < 1) is always above zero for
any K < ∞. However, when K = ∞ it admits the following
expression:
0
pe ≥ q ,
p0 =
(16)
1 − pqe Otherwise .
This can be seen by noting that, the summation
k
∞
∞
X
X
pe (1 − q)
k
Ω =
q(1 − pe )
k=1
k=1
converges if pe < q. Otherwise the sum diverges, which
makes the depletion probability equal to zero. This depletion
probability expression for an infinite capacity battery sensor
3 We would like to note that the difference between this model and an
M/M/1/K queue is that in this model each battery charge state k after each
transition can either change by one, or remain unchanged.
is in line with the expression found before by modeling the
battery state as a birth-death process in [32].
Note that the condition pe ≥ q, which results in zero
depletion probability, follows intuition in the sense that, when
the probability of energy arrival pe is higher than the probability of energy consumption (or the probability q that the
sensor observation is above the threshold and it decides to
send a message), the battery will accumulate energy and, with
probability one, in long term not be empty. In this situation
the problem will be the same as the unconstrained setup.
Now by plugging the expression for the depletion probability in (15) into that of the BD for energy constrained sensor
in (14), we can find a closed-form expression for the BD of
a K-slot-battery sensor as follows.
"s
pe (ΩK − 1)
pe (ΩK − 1)
B = −log
0 + δq0
+
δq
+
0
1
pe ΩK − q
p e ΩK − q
s
#
pe (ΩK − 1)
pe (ΩK − 1)
1 − 0 − δq0
1 − 0 − δq1
.
p e ΩK − q
pe ΩK − q
(17)
Note that since the BD is a function of observation models,
we can not simplify it more. In what follows we will be
considering different battery capacities K, and numerically
find the threshold Θ? which maximizes the corresponding
Bhattacharyya distance.
When a sensor is capable of saving only one packet of
energy, incoming energy et is saved in the battery if the battery
is empty. Otherwise, the sensor discards the incoming energy
packet. For a single-slot-battery sensor an optimal threshold
Θ? is found by maximizing the BD in (17) for K = 1. To do
this, we use a grid search to find the threshold Θ that maximizes the BD. When a sensor is capable of saving two packets
of energy (K = 2), incoming energy et is saved in the battery
if the battery is empty or has only one packet of energy in
the buffer. Otherwise, the sensor discards the incoming energy
packet. The same as for a single-slot-battery, by maximizing
7
(17) we can find an optimal threshold Θ? for a double-slotbattery sensor K = 2. Fig. 4 illustrates the resulting BD of
single- and double-slot-battery energy harvesting sensors using
the adapted (Θ? ) and the unconstrained (Θ?u ) thresholds, for
π1 = 0.2 and pe = 0.15. The figure on the left, considers
error-free communication channels (0 = 1 = 0) while the
figure on the right considers BAC channels with parameters
0 = 0.1 and 1 = 0.2.
We observe that the unconstrained threshold Θ?u is always
sub-optimal. In the single- and double-slot-battery cases the
Bhattacharyya distance is bounded when the SNR goes to
infinity. In Theorem 1 we will find a closed-form expression
for this upper bound for an arbitrary battery size K, and find
conditions under which the Bhattacharyya distance is upper
bounded.
Theorem 1: Consider a K-slot-battery energy harvesting
sensor S. Assume that the probability of harvesting energy
at each time interval is pe , and the a-priori probability of
hypothesis Ht = 1 is π1 and the sensor-to-FC channels are
BAC channels as in Fig. 2. The BD of this sensor at the input
of the FC can not exceed
hp
i
p
B , − log
0 (1 − 1 − p0 δ) + (1 − 0 ) (1 + p0 δ) ,
(18)
where
"
k
K
1 X pe (1 − π1 )
p0 = 1 +
1 − π1
π1 (1 − pe )
#−1
.
k=1
Proof. See Appendix B.
Remark: The upper bound in (18) is achievable under
any separable observation model by selecting Θ? so that
q0 = 0 and q1 = 1. It is also asymptotically achievable
(when E → ∞) for a sensor with an asymptotically separable
observation model. These asymptotes are also shown in Fig. 4
by dotted lines. Note that a, possibly unexpected, insight due
to this result is that when the sensors make (asymptotically)
separable observations, i.e., when the sensors are sure of the
true hypothesis, it is optimal for the sensor to act greedily and
always transmit whenever Ht = 1 and their internal battery
allows for a transmission. The FC will in this situation still
only receive a transmission from a subset of the sensors, as the
battery depletion events are independent across sensors due to
the independence of the energy arrival process.
Remark: The upper bound in (18) is an increasing function
of the battery capacity K. It is in line with the intuition that the
performance of an energy harvesting sensor S is improved by
increasing its battery capacity. The upper bound in (18) is also
an increasing function of the probability of harvesting energy
pe and is a decreasing function of the a-priori probability π1 .
These also follow the intuition in the sense that by increasing the probability of having energy available at the battery,
the performance of the sensor is improved. While pe affects
the amount of available energy directly, π1 affects the battery
content in a more complicated way: According to (10) an
optimally designed sensor aims to send a message “1” and
consume a packet of energy when its observation is above
Θ? . By increasing the a-priori probability of the hypothesis
Ht = 1, it will be more likely that the sensor aims to send
a “1” and consume energy. This itself increases the depletion
probability p0 and so decreases the performance of the sensor.
We would also like to note that, unlike previous works on
the design of sensor decision functions in a distributed detection network using the Bhattacharyya distance as performance
metric, our problem formulation comprises the affect of apriori probabilities πj , for j = 0, 1 through q in the depletion
probability p0 .
Remark: For any finite K, the BD never grows unboundedly
with E → ∞, while for the unconstrained case we have seen
that it can grow unboundedly for separable and asymptotically separable observation models for error-free channel by
increasing the SNR.
Using (17) and (18), one can analyze the BD performance
of a K-slot-battery sensor (when K < ∞) and its asymptote.
In Fig. 5 the optimum BD for an energy harvesting sensor
as a function of the sensor battery capacity is shown when
the observation model at the sensor is according to (9) with
s = 5. We observe from this figure that the maximum BD
and the upper bound (as discussed before) are both increasing
(non-decreasing) functions of sensor battery capacity K.
Remark: For an infinite battery capacity sensor K = ∞,
under some conditions the BD distance grows unboundedly
for an asymptotically separable model, as SNR increases.
Concretely, only if pe ≥ π1 and communication channels
are noiseless, 0 = 1 = 0, the Bhattacharyya distance
of an optimal threshold test (10) increases unboundedly for
separable or asymptotically separable observation models, as
SNR increases.
In Fig. 6 the Bhattacharyya distance for an infinite-slotbattery sensor is shown, when using different thresholds, and
for different setups. We observe from this figure that as the
non-centrality parameter (or SNR) increases, for the case
where pe ≥ π1 and 0 = 1 = 0 the Bhattacharyya distance
increases unboundedly, otherwise it is upper bounded with the
asymptote (shown by the dotted line) found in (18).
In the following section, we compare the error probability
performance of networks of energy harvesting sensors which
we have designed using conventional (unconstrained) formulation in (6) (for Θ?u ) and using our proposed formulation in (17)
(for Θ? ). Before considering the error probability performance
of the designed networks, consider again the choice of Θ?u and
Θ? . While using Θ?u , the sensor makes an optimal decision (in
the sense of the Bhattacharyya distance) for the case where
required energy for transmission of positive message is always
available at the sensor. However, when energy is not always
available at the sensor, the sensor should act more conservatively in the sense that: The sensor should remain silent
and preserve energy for future time slots, unless it receives
observation about the presence of hypothesis 1 with high
reliability. In our formulation for the Bhattacharyya distance,
the threshold Θ? determines if a received observation has high
reliability about the presence of hypothesis 1. According to our
simulation results, we always obtain Θ? ≥ Θ?u which further
confirms our discussion above.
8
0.7
1
rS
rS
rS
rS
rS
uT
rS
0.8
rS
rS
rS
Bhattacharyya distance
Bhattacharyya distance
rS
rS
rS
rS
0.9
rS
rS
0.6
0.5
rS
rS
Upperbound, ǫ0 = 0.1, ǫ1 = 0.2
Using Θ⋆ , ǫ0 = 0.1, ǫ1 = 0.2
Upperbound, ǫ0 = 0, ǫ1 = 0
Using Θ⋆ , ǫ0 = 0, ǫ1 = 0
rS
rS
0.4
rS
bC
bC
rS
0.3
rS
0.7
bC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bC
bCbC
bC
bC
rS
0.5
rS
uT
uT
uT
0.4
rS
rS
rS
rS
0.3
bC
uT
0.1
2
uT
3
4
5
6
7
8
9
10
Battery capacity K
Fig. 5. Bhattacharyya distance of an energy harvesting sensor as a function
of battery capacity K, for noisy and noiseless communication channels and
for π1 = 0.2, pe = 0.15, and corresponding upper bounds according to
Theorem 1.
IV. E RROR P ROBABILITY P ERFORMANCE OF N ETWORKS
In this section we illustrate the benefit of our results by
numerical examples. Consider a sensor network with N = 4
energy harvesting sensors. Suppose the sensors make observations from the same phenomenon Ht during each time
interval t, and send an OOK message to the FC. Let us
assume again that sending a positive message consumes a
packet of energy and a negative message is conveyed through
a non-transmission with no energy cost. Let the observation
model at each sensor be as in (9), and conditioned on the
true hypothesis, the observations be independent, and that the
sensors use the threshold test in (10). Note that, though in
this section we assume the same observation model at the
sensors, our results and conclusions drawn through this work
are generalized to non-identical observation models at the
sensors (and respectively non-identical sensor) case.
Using our results in the previous section, we design sensor
decision rules Θ for different battery sizes, and compare their
error probability performance with those of the unconstrained
case. The expected error probability, when the FC uses the
MAP criterion, at time t is found using [6]
n
o
X
PE,t = 1 −
max πj PY |H y t |j ,
y
rS
bC
bC
bC
j=0,1
t
which can be numerically computed, without the need for
Monte-Carlo simulations.
Fig. 7 shows the error probability performance of designed
sensor networks, with single-slot-battery sensors and doubleslot-battery sensors, when π1 = 0.2 and pe = 0.15. In both
cases (K = 1, 2) the adapted threshold Θ? leads to better
performance than the unconstrained threshold Θ?u . This is in
line with our results based on the Bhattacharyya distance:
Adapted threshold leads to a higher BD. As s → ∞, we also
observe that the error probability does not converge to zero
(lower bounded). It was shown for the Bhattacharyya distance
0
rSuT
bC
1
2
bCrS
Using
Using
Using
Using
Using
Using
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
(0.2, 0.2, 0, 0)
(0.2, 0.2, 0, 0)
(0.2, 0.15, 0, 0)
(0.2, 0.15, 0, 0)
(0.2, 0.15, 0.1, 0.2)
(0.2, 0.15, 0.1, 0.2)
bC
bCbC
bC
bC
bC
bC
5
6
7
8
9
10
bC
bC
bCrS
bCbC
1
0.2
rS
bC
rS
rS
0.6
uT
bC
0.1
rSrS
rS
rS
bC
0.2
rS
rS
rS
rS
uT
3
4
Non-centrality parameter s
Fig. 6. Bhattacharyya distance of an infinite-slot-battery (K = ∞) energy
harvesting sensor for different threshold tests and for different (π1 , pe , 0 , 1 ),
as a function of noncentrality parameter.
that Bt is upper bounded for single and double-slot-battery
sensors. We also observe from the figure that, while both using
the optimal threshold or the sub-optimal threshold, increasing
the battery capacity (here from one to two) can improve the
error probability performance of a network of sensors.
In Fig. 8 the error probability performance of the network of
N = 4 infinite-slot-battery sensors is shown for different sets
of (π1 , pe , 0 , 1 ). When π1 > pe and/or sensor-to-FC channels
are erroneous, the error probability of the network converges
to a fixed value as s → ∞, while when π1 ≤ pe and the
channels are error-free (0 = 1 = 0), the error probabilities
for both the adapted threshold and the energy unconstrained
threshold rapidly go to zero. These observations are also in
line with our results in terms of the Bhattacharyya distance:
When K = ∞, if pe ≥ π1 and channels are error-free, the BD
increases unboundedly, otherwise it converges to a non-zero
asymptote. Note that we have the same observations for other
choices of (π1 , pe , 0 , 1 ).
V. C ONCLUDING R EMARKS
In this paper we studied the problem of decentralized hypothesis testing in a network of energy harvesting sensors. The
sensors in the network make observations of a phenomenon
and harvest all the energy they need from the environment.
We consider the case where the sensors have different battery
capacities to save the harvested energy. Considering the Bhattacharyya distance as a performance metric, we formulated the
problem of designing sensors in the network by considering
the constraints imposed by energy harvesting and proposed a
method to design the sensors decision rules. We further studied
the performance of sensors for different battery capacities and
presented conditions under which the Bhattacharyya distance
is upper bounded, and therefore the error probability is lower
bounded and does not converge to zero.
In this paper, we considered the case where the observations and the energy charging processes at the sensors
9
−0.65
Using
Using
Using
Using
rS
rS
bCrS
bC
bC
bC
−0.75
=1
=1
=2
=2
uTrSbC
bC
bC
rS
bC
uTrS
bC
bC
rS
bC
uT
−1
bC
bCbC
uT
−1.5
bC
rS
rS
rS
uT
bCrS
rS
rS
−0.85
rS
bC
rS
rS
−0.9
rSrS
rS
rS
rS
rS
−2
rS
uT
rS
uT
−2.5
bC
bC
bC
rS
−0.8
log10 PE
−0.5
K
K
K
K
log10 PE
−0.7
rS
bCrS
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
uT
Using
Using
Using
Using
Using
rS
Using
bC
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
Θ⋆u ,
Θ⋆ ,
uT
bC
bC
(0.2, 0.15, 0.1, 0.2)
(0.2, 0.15, 0.1, 0.2)
(0.2, 0.15, 0, 0)
(0.2, 0.15, 0, 0)
(0.2, 0.2, 0, 0)
(0.2, 0.2, 0, 0)
rS
bC
−0.95
−3
bC
rS
rS
uT
rS
rS
rS
rS
rS
6
7
8
9
10
bC
bC
−1
−1.05
bC
bC
1
2
3
4
5
6
bC
bC
bC
bC
7
8
9
10
−3.5
−4
uT
1
2
3
4
Fig. 7. Error probability performance of networks with N = 4 energy
harvesting sensors with different battery capacities, when π1 = 0.2 and
pe = 0.15, and noisy communication channels with 0 = 0.1 and 1 = 0.2.
are independent. A possible extension to this work can be
to consider the case where the observations at the sensors,
or/and the energy charging processes are correlated. For the
sake of analytical tractability and insight, we only studied
the case where each sensor decision rule is a simple singlethreshold test. It is known that having multiple-thresholds at
the sensors can result in better BD performance. Concretely,
battery state dependent thresholds Θ(bt ) can improve the
performance of a single-threshold energy harvesting sensor,
and an extension to our study can be to study the performance
of such threshold tests. Although not presented herein, we have
numerically found optimal threshold tests Θ?1 , Θ?2 , where Θ?i
is the optimal threshold when bt = i for i = 1, 2, using a
grid search in the case of a double-slot-battery (K = 2) and
the observation model in (9). Our results confirm that having
multiple thresholds can improve the BD performance of the
sensor, but the gains are small compared to introducing the
single threshold in the first place.
We would also like to note that other members of Ali-Silvey
distance (like J-divergence) can also be used as a performance
metric for the design of decision rules at the sensors, with a
lot of similar derivations. In this paper, we chose to use the
BD as the performance metric since according to [5] the BD is
one of the most analytically tractable members of Ali-Silvey
distances and was reported as the most efficient metric among
those studied in [21].
Our work represents a first attempt to introduce energy harvesting considerations in the context of distributed detection
using established (but arguably simple) design metrics such as
the Bhattacharyya distance. Our approach furthermore utilizes
simplified models such as the energy arrival model taken from
[36]–[39]. While the simplifying assumptions make the results
more tractable and interpretable, they also naturally come with
some strong limitations. Apart from the extension to a batterydependent decision threshold, one could also consider the case
where the sensors optimize the OOK transmission energy in
5
Non-centrality parameter s
Non-centrality parameter s
Fig. 8. Error probability performance of networks with infinite-slot-battery
(K = ∞) energy harvesting sensors, and for different (π1 , pe , 0 , 1 ).
order to influence the error probability of the transmission,
i.e., the cross-over probability 0 in Fig. 2, based on some
given physical channel model. This would however arguably
be more relevant in conjunction with a more refined energy
arrival and storage model, which we cannot presently handle
directly given that we (for simplicity) assume that energy
arrives in quanta matched to the energy needed for a positive
(on) transmission. Furthermore, although the decoupling of
the sensor design caused by considering the Bhattacharyya
distance is often used to simplify the sensor design problem
in the distributed detection literature, this approach should be
scrutinized again when adding the energy harvesting aspect.
This is due to the fact that jointly designed sensor decision
rules can potentially lead to tangible benefits through more
advanced joint energy conservation rules across sensors than
what is implicitly provided by the differences in the random
battery state across the sensors.
ACKNOWLEDGMENT
The authors would like to thank the Associate Editor
and anonymous reviewers for their valuable comments and
suggestions that led to improvement of this paper. The authors
would especially like to thank the reviewers who prompted us
to consider non-perfect communication channels.
A PPENDIX A
P ROOF OF L EMMA 1
Using (12), state probabilities for p0 , . . . , pK−1 are found
as
p0 = p0,0 p0 + p1,0 p1 ,
pk = pk−1,k pk−1 + pk,k pk + pk+1,k pk+1 ,
1 ≤ k < K.
(19)
Using (2) and (10), for the transition probabilities we obtain
p0,0 = 1 − pe ,
p1,0 = q(1 − pe ) ,
p0,1 = pe ,
10
and for 1 ≤ k < K
by replacing α, β, ξ, and ζ in (24) by q0 , q1 , 0 , and (1−p0 )δ,
respectively.
pk,k+1 = (1 − q)pe ,
pk+1,k = q(1 − pe ) ,
(20)
pk,k = qpe + (1 − q)(1 − pe ) .
By re-arranging the equations in (19) and replacing the transition probabilities pi,j with those in (20) we get the following
equations
pe
p1 =
p0
(21a)
q(1 − pe )
p2 (1 − q)
p0
(21b)
p2 = 2 e
q (1 − pe )2
pk = −Ω pk−2 + (1 + Ω) pk−1 , 3 ≤ k ≤ K.
(21c)
The system of equations described by (21) is a homogeneous
difference equation [45]. Its characteristic polynomial is
In what follows, we will find another upper bound for B by
maximizing the right-hand-side of (25) which is equivalent to
minimizing Γ2 .
p
Γ2 , 0 [0 + q1 (1 − p0 )δ]+
p
(26)
[1 − 0 ][1 − 0 − q1 (1 − p0 )δ]
Minimizing Γ2 is equivalent to maximizing Σ , q1 (1 − p0 ),
since Γ2 is a decreasing function of Σ. Then by replacing the
depletion probability p0 from (15) we obtain
"
#−1
K
k
k−1
X
p
(1
−
q
π
−
q
π
)
0
0
1
1
e
.
Σ = q1 1 − 1 +
(1 − pe )k (q0 π0 + q1 π1 )k
k=1
(27)
2
ρ − (1 + Ω)ρ + Ω = 0 ,
whose roots are ρ = {1, Ω}. These lead to the general solution
for the difference equation as
k
pk = λ1 + λ2 Ω .
Ωk
p0 ,
1−q
1≤k≤K,
max
q0 ,q1
0≤q0 ≤q1 ≤1
(22)
Now, applying auxiliary conditions (21a) and (21b) to (22),
1
we obtain λ1 = 0 and λ2 = 1−q
p0 . Thus the solution to the
difference equation is found to be given by
pk =
Now, considering the optimization problem
(28)
the objective function Σ is a decreasing function of q0 (note
that q0 only appears in the argument of the summation) and
its maxima is attained when q0 = 0, and therefore the problem
in (28) reduces to the following problem,
(23)
max Λ ,
(29)
0≤q1 ≤1
which describes each state probability pk in terms of the depletion probability p0 . Now, using the fact that the summation
of the probabilities must equal unity, for a K-slot-battery we
obtain the desired depletion probability expression (15).
where,
Λ,Σ
q0 =0
"
K
X
pke (1 − q1 π1 )k−1
= q1 1 − 1 +
(1 − pe )k (q1 π1 )k
#−1
(30)
.
k=1
A PPENDIX B
P ROOF OF T HEOREM 1
To prove the theorem, first consider the following function
Γ1 where 0 ≤ α ≤ β ≤ 1, 0 ≤ ζ ≤ 1, and 0 ≤ ξ ≤ 0.5.
p
p
Γ1 , [ξ + ζα][ξ + ζβ] + [1 − ξ − ζα][1 − ξ − ζβ]
Taking the derivative of Γ1 with respect to α, we observe that
when 0 ≤ α ≤ β, Γ1 is an increasing function of α and its
minima is attained when α = 0. In other words
p
p
min Γ1 = ξ[ξ + ζβ] + [1 − ξ][1 − ξ − ζβ] ,
In what follows we will show that the objective function
Λ is an increasing function of q1 and therefore its maxima is
attained when q1 = 1, which proves the theorem. Our aim is
now to show
dΛ
≥ 0,
dq1
when 0 ≤ q1 ≤ 1. By evaluating the summation in (30) and
introducing
λ,
α
or equivalently
p
p
[ξ + ζα][ξ + ζβ] + [1 − ξ − ζα][1 − ξ − ζβ] ≥
p
p
ξ[ξ + ζβ] + [1 − ξ][1 − ξ − ζβ]
Σ,
pe (1 − q1 π1 )
,
(1 − pe )q1 π1
we obtain
(24)
for 0 ≤ α ≤ β ≤ 1, 0 ≤ ζ ≤ 1, and 0 ≤ ξ ≤ 0.5.
Using (24), we can conclude that for any given q1 , 0 , and
p0 , the BD in (14) is upperbounded by
hp
B ≤ − log
0 [0 + q1 (1 − p0 )δ]+
i
(25)
p
[1 − 0 ][1 − 0 − q1 (1 − p0 )δ] ,
Λ=
pe
π1
1 − λK
.
1 − pe λK − (1 − pe )λK+1
According to the chain rule we have
dΛ
dΛ dλ
=
,
dq1
dλ dq1
and using first principles we can straightforwardly show
dλ
≤ 0,
dq1
11
for 0 ≤ q1 ≤ 1. To complete the proof, we need to show
dΛ
≤ 0,
(31)
dλ
for λ ≥ 0. To this end, note that
K−1
(K + 1)λ − K − λK+1 (1 − p)
dΛ
pe λ
=
.
2
dλ
π1
(1 − p)λK+1 + pλK − 1
Thus, proving (31) is equivalent to proving that
g(λ) , (K + 1)λ − K − λK+1 ≤ 0 .
To prove this, we first show g(λ) (for any K ≥ 0) has
its optima at λ = 1 (where g (1) = 0), by setting its first
dg
equal to zero, i.e.,
derivative dλ
dg
= (K + 1) − (K + 1)λK = (K + 1)(1 − λK ) = 0 .
dλ
Then, we show this optima is a maxima (i.e., maxλ g(λ) =
g(1) = 0) by finding its second derivative, i.e.,
d2 g
= −K(K + 1)λK−1 ≤ 0 .
dλ2
This completes the proof of
of Theorem 1.
dΛ
dλ
≤ 0 and therefore the proof
R EFERENCES
[1] P. Varshney, Distributed Detection and Data Fusion. Springer-Verlag
New York, Inc., 1996.
[2] W. Li and H. Dai, “Distributed detection in wireless sensor networks
using a multiple access channel,” IEEE Trans. Signal Process., vol. 55,
no. 3, pp. 822–833, 2007.
[3] D. Ciuonzo, G. Romano, and P. S. Rossi, “Channel-aware decision
fusion in distributed mimo wireless sensor networks: Decode-and-fuse
vs. decode-then-fuse,” IEEE Trans. Wireless commun., vol. 11, no. 8,
pp. 2976–2985, 2012.
[4] D. Ciuonzo, P. S. Rossi, and S. Dey, “Massive MIMO channel-aware
decision fusion,” IEEE Trans. Signal Process., vol. 63, no. 3, pp. 604–
619, Feb 2015.
[5] M. Longo, T. Lookabaugh, and R. Gray, “Quantization for decentralized
hypothesis testing under communication constraints,” IEEE Trans. Inf.
Theory, vol. 36, no. 2, pp. 241–255, Mar. 1990.
[6] A. Tarighati and J. Jaldén, “Bayesian design of decentralized hypothesis
testing under communication constraints,” in Proc. IEEE Int. Conf.
Acoustics, Speech and Signal Process. (ICASSP), May 2014, pp. 7624–
7628.
[7] V. Veeravalli and P. Varshney, “Distributed inference in wireless sensor
networks,” Phil. Trans. A, Math. Phys. Eng. Sci., vol. 370, no. 1958, pp.
100–117, 2012.
[8] J.-F. Chamberland and V. Veeravalli, “Wireless sensors in distributed
detection applications,” IEEE Signal Process. Mag., vol. 24, no. 3, pp.
16–25, 2007.
[9] B. Chen, L. Tong, and P. Varshney, “Channel aware distributed detection
in wireless sensor networks,” IEEE Signal Process. Mag., 2006.
[10] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A survey
on sensor networks,” IEEE Commun. Mag., vol. 40, no. 8, pp. 102–114,
2002.
[11] S. J. Baek, G. De Veciana, and X. Su, “Minimizing energy consumption
in large-scale sensor networks through distributed data compression and
hierarchical aggregation,” IEEE J. sel. Areas Commun., vol. 22, no. 6,
pp. 1130–1140, 2004.
[12] P. Nuggehalli, V. Srinivasan, and R. R. Rao, “Energy efficient transmission scheduling for delay constrained wireless networks,” IEEE Trans.
Wireless Commun., vol. 5, no. 3, pp. 531–539, 2006.
[13] V. Sharma, U. Mukherji, V. Joseph, and S. Gupta, “Optimal energy
management policies for energy harvesting sensor nodes,” IEEE Trans.
Wireless Commun., vol. 9, no. 4, pp. 1326–1336, April 2010.
[14] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and
K. Huang, “Energy harvesting wireless communications: A review of
recent advances,” IEEE J. Sel. Areas Commun., 2015.
[15] D. Gunduz, K. Stamatiou, N. Michelusi, and M. Zorzi, “Designing
intelligent energy harvesting communication systems,” IEEE Commun.
Magazine, vol. 52, no. 1, pp. 210–216, 2014.
[16] S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes:
Survey and implications,” IEEE Communications Surveys & Tutorials,,
vol. 13, no. 3, pp. 443–461, 2011.
[17] C. Rago, P. Willett, and Y. Bar-Shalom, “Censoring sensors: a lowcommunication-rate scheme for distributed detection,” IEEE Tran.
Aerosp. Electron. Syst., vol. 32, no. 2, pp. 554–568, April 1996.
[18] C. R. Berger, M. Guerriero, S. Zhou, and P. Willett, “PAC vs. MAC
for decentralized detection using noncoherent modulation,” IEEE Trans.
Signal Process., vol. 57, no. 9, pp. 3562–3575, 2009.
[19] D. Ciuonzo, G. Romano, and P. S. Rossi, “Optimality of received energy
in decision fusion over rayleigh fading diversity mac with non-identical
sensors,” IEEE Trans. Signal Process., vol. 61, no. 1, pp. 22–27, 2013.
[20] F. Li, J. S. Evans, and S. Dey, “Decision fusion over noncoherent fading
multiaccess channels,” IEEE Trans. Signal Process., vol. 59, no. 9, pp.
4367–4380, 2011.
[21] H. Poor and J. B. Thomas, “Applications of ali-silvey distance measures
in the design generalized quantizers for binary decision systems,” IEEE
Trans. Commun., vol. 25, no. 9, pp. 893–900, 1977.
[22] A. Tarighati and J. Jaldén, “Optimality of rate balancing in wireless
sensor networks,” IEEE Trans. Signal Process., vol. 64, no. 14, pp.
3735–3749, July 2016.
[23] A. Nayyar, T. Başar, D. Teneketzis, and V. V. Veeravalli, “Optimal
strategies for communication and remote estimation with an energy
harvesting sensor,” IEEE Trans. Autom. Control, vol. 58, no. 9, pp.
2246–2260, Sept 2013.
[24] Y. Zhao, B. Chen, and R. Zhang, “Optimal power allocation for an energy harvesting estimation system,” in Proc. IEEE Int. Conf. Acoustics,
Speech and Signal Process., May 2013, pp. 4549–4553.
[25] C. Huang, Y. Zhou, T. Jiang, P. Zhang, and S. Cui, “Power allocation for
joint estimation with energy harvesting constraints,” in Proc. IEEE Int.
Conf. Acoustics, Speech and Signal Process., May 2013, pp. 4804–4808.
[26] M. Nourian, S. Dey, and A. Ahlén, “Distortion minimization in multisensor estimation with energy harvesting,” IEEE J. Sel. Areas Commun.,
vol. 33, no. 3, pp. 524–539, March 2015.
[27] T. Li, P. Fan, and K. B. Letaief, “Outage probability of energy harvesting
relay-aided cooperative networks over rayleigh fading channel,” IEEE
Trans. Vehicular Tech., vol. 65, no. 2, pp. 972–978, Feb 2016.
[28] Y.-W. P. Hong, “Distributed estimation with analog forwarding in
energy-harvesting wireless sensor networks,” in Proc. IEEE Int. Conf.
Commun. Syst. (ICCS), 2014, pp. 142–146.
[29] H. Liu, G. Liu, Y. Liu, L. Mo, and H. Chen, “Adaptive quantization for
distributed estimation in energy-harvesting wireless sensor networks: a
game-theoretic approach,” Int. J. Distributed Sensor Networks, vol. 2014,
2014.
[30] B. K. Chalise, Y. D. Zhang, and M. G. Amin, “Energy harvesting in
an ostbc based amplify-and-forward mimo relay system,” in Proc. IEEE
Int. Conf. Acoustics, Speech and Signal Process. (ICASSP), 2012, pp.
3201–3204.
[31] B. Medepally and N. B. Mehta, “Voluntary energy harvesting relays
and selection in cooperative wireless networks,” IEEE Trans. Wireless
Commun., vol. 9, no. 11, pp. 3543–3553, 2010.
[32] A. Tarighati, J. Gross, and J. Jaldén, “Decentralized detection in energy
harvesting wireless sensor networks,” in Proc. European Signal Process.
Conf. (EUSIPCO), Aug. 2016.
[33] J. M. Berger, “A note on error detection codes for asymmetric channels,”
Inf. and Control, vol. 4, no. 1, pp. 68–73, 1961.
[34] J. N. Tsitsiklis, “Decentralized detection by a large number of sensors,”
Math. Contr., Signals, Syst., vol. 1, no. 2, pp. 167–182, 1988.
[35] T. Kailath, “The divergence and bhattacharyya distance measures in
signal selection,” IEEE Trans. Commun., vol. 15, no. 1, pp. 52–60, Feb.
1967.
[36] R. Valentini, M. Levorato, and F. Santucci, “Aging aware random
channel access for battery-powered wireless networks,” IEEE Wireless
Commun. Letters, vol. 5, no. 2, pp. 176–179, 2016.
[37] N. Michelusi and M. Zorzi, “Optimal adaptive random multiaccess in
energy harvesting wireless sensor networks,” IEEE Trans. Commun.,
vol. 63, no. 4, pp. 1355–1372, 2015.
[38] N. Michelusi, K. Stamatiou, and M. Zorzi, “Transmission policies for
energy harvesting sensors with time-correlated energy supply,” IEEE
Trans. Commun., vol. 61, no. 7, pp. 2988–3001, 2013.
[39] K. Tutuncuoglu, O. Ozel, A. Yener, and S. Ulukus, “Improved capacity
bounds for the binary energy harvesting channel,” in Proc. IEEE Int.
Symp. Inf. Theory (ISIT), 2014, pp. 976–980.
12
[40] J. N. Tsitsiklis, “Extremal properties of likelihood-ratio quantizers,”
IEEE Trans. Commun., vol. 41, no. 4, pp. 550–558, 1993.
[41] B. Chen and P. K. Willett, “On the optimality of the likelihood-ratio test
for local sensor decision rules in the presence of nonideal channels,”
IEEE Trans. Inf. Theory, vol. 51, no. 2, pp. 693–699, Feb 2005.
[42] B. Liu and B. Chen, “Channel-optimized quantizers for decentralized
detection in sensor networks,” IEEE Trans. Inf. Theory, vol. 52, no. 7,
pp. 3349–3358, July 2006.
[43] E. L. Lehmann and J. P. Romano, Testing statistical hypotheses.
Springer Science & Business Media, 2006.
[44] D. Gross, J. F. Shortie, J. M. Thompson, and C. M. Harris, Fundamentals
of Queueing Theory. John Wiley & Sons, Inc., 2008.
[45] P. S. Diniz, E. A. Da Silva, and S. L. Netto, Digital signal processing:
system analysis and design. Cambridge University Press, 2010.
| 7cs.IT
|
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID
GROUPS
arXiv:1609.05611v1 [math.AT] 19 Sep 2016
ERIC RAMOS
Abstract. For a tree G, we study the changing behaviors in the homology groups Hi (Bn G)
as n varies, where Bn G := π1 (UConf n (G)). We prove that the ranks of these homologies
can be described by a single polynomial for all n, and construct this polynomialLexplicitly in
terms of invariants of the tree G. To accomplish this we prove that the group n Hi (Bn G)
can be endowed with the structure of a finitely generated graded module over an integral
polynomial ring, and further prove that it naturally decomposes as a direct sum of graded
shifts of squarefree monomial ideals. Following this, we spend time considering how our
methods might be generalized to braid groups of arbitrary graphs, and make various conjectures in this direction.
1. Introduction
1.1. Introductory remarks and statements of the main theorems. In recent years
there has been a push towards understanding the mechanisms connecting various well-known
asymptotic stability results in topology and algebra. For instance, let N be a connected
oriented manifold of dimension ≥ 2, which is the interior of a manifold with boundary, and
write Conf n (N ) for the n-strand configuration space
Conf n (N ) := {(x1 , . . . , xn ) ∈ N n | xi 6= xj }.
There is a natural action on Conf n (N ) by the symmetric group Sn , and we may therefore
define the n-strand unordered configuration space
UConf n (N ) := Conf n (N )/Sn .
A classical theorem of McDuff [McD, Theorem 1.2] implies that for each i, and any n ≫ i,
the group Hi (UConf n (N )) is independent of n. In contrast to the work of McDuff, it can be
seen that the analogous statement is not true for the ordered configuration spaces Conf n (N ).
What is true, however, is perhaps the next best thing. It follows from work of Church, Ellenberg, and Farb [CEF, Theorem 6.4.3] that for any i, there is a polynomial P ∈ Q[x] such
that the Betti number dimQ (Hi (Conf n (U); Q)) agrees with P (n) for n ≥ 0. Results of this
type fall under the heading of what one might call asymptotic algebra.
The modern philosophy of asymptotic algebra can be roughly stated as follows: a family
of algebraic objects which display asymptotic stability phenomena can often times be encoded into a single object, which is finitely generated in an appropriate abelian category.
In the case of McDuff, for each i and n the group Hi (UConf n (N )) can be realized as the
n-graded piece of some finitely generated graded module over Z[x]. The result of Church,
Ellenberg, and Farb involves showing that the Sn -representations Hi (Conf n (U); Q) are each
constituents of some finitely generated representation of the category FI, of finite sets and
injections (see [CEF] for more on this notion). This philosophy is also heavily featured in
The author was supported by NSF-RTG grant DMS-1502553.
1
2
ERIC RAMOS
Sam and Snowden’s recent resolution of Stembridge’s conjecture [SS]. The goal of this paper
is to apply similar techniques to the homologies of the unordered configuration spaces of
trees. Note that this problem was considered by Lütgehetmann in his Master’s thesis [L].
The results of that work are disjoint from the current work.
In this paper, a graph will always refer to a connected, compact CW-complex of dimension
1. A tree is a graph which is contractible as a topological space. An essential vertex of
G is a vertex of degree, or valency, at least 3, while an essential edge of G is a connected
component of the space obtained by removing all essential vertices from G. Note that both
the essential edges and vertices of a graph are unaffected by subdivision of edges, and can
be thought of as the topologically essential pieces of the graph.
Our main result relates to the asymptotic behavior of the homologies of the braid group of
a tree. To state this result, we first need to define a kind of connectivity invariant for trees.
Let G be a tree. Then we set
∆iG :=
max
{dimQ (H0 (G − {vj1 , . . . , vji }; Q))}.
{{vj1 ,...,vji }|vjk essential}
In words, ∆iG is the maximum number of connected components that G can be broken into
by removing exactly i essential vertices. Therefore ∆1G is the maximum degree of a vertex
G
in G, while, if G has NG essential vertices, ∆N
G is the number of essential edges of G. By
0
i
convention, ∆G = 1, while ∆G = 0 for i > NG .
Theorem A. Let G be a tree, and write Bn G to denote the braid group Bn G := π1 (UConf n (G)).
Then for each i there is a polynomial Pi ∈ Q[x] of degree ∆iG − 1, such that for all n ≥ 0,
Pi (n) = dimQ (Hi (Bn G; Q)).
Remark 1.1. It follows from a Theorem of Ghrist [Gh, Theorem 3.3] that given any graph G
not homeomorphic to S 1 , Hi (Bn G) = 0 for all i strictly greater than the number of essential
vertices of G. This is realized in the case where G is a tree in Theorem A by the fact that
∆iG − 1 = −1 in these cases.
The polynomial Pi of Theorem A is explicitly computed throughout the course of this work
(see Theorem 3.17). This computation implies something somewhat surprising about these
homology groups.
Corollary B. Let G be a tree, and let i ≥ 0. Then the homology groups Hi (Bn G) depend
only on i, n, and the degree sequence of G.
It is interesting to note that the rank of Hi (Bn G) agrees with a polynomial for all n ≥ 0,
as opposed to only agreeing for n sufficiently large. In the case of configuration spaces of
manifolds, a result of this kind does have precedent. We have already discussed the result of Church, Ellenberg, and Farb which states that if N is an oriented manifold, which
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
3
is the interior of a manifold with non-empty boundary, then for any i the dimension of
Hi (Conf n (N ); Q) agrees with a polynomial for all n [CEF, Theorem 6.4.3]. It is perhaps
an interesting observation that trees can be thought of as graphs with non-trivial boundary.
It is unclear whether the connection to the work of Church, Ellenberg, and Farb goes any
deeper than this, however.
1.2. An Outline of the proof. To prove Theorem A, we will rely on classical techniques
in commutative algebra, as well as more modern techniques in combinatorial topology. The
first key ingredient is the discrete Morse theory of Forman [Fo]. Given a regular CW complex
X (see Definition 2.1), write Ki for the set of i-cells of X. A discrete Morse function is a
map f from the cells of X to R satisfying the following two hypotheses for all cells σ ∈ Ki :
1. |{τ ∈ Ki+1 | σ ⊆ τ and f (σ) ≥ f (τ )}| ≤ 1;
2. |{τ ∈ Ki−1 | τ ⊆ σ and f (σ) ≤ f (τ )}| ≤ 1.
We call σ ∈ Ki a critical i-cell with respect to f , if the sets of conditions 1 and 2 are
both empty. The main consequence of discrete Morse theory is that the critical cells of X
determine its homotopy type. Formally, let a ∈ R and write X(a) for the subcomplex of
X comprised of the closures of cells σ with f (σ) ≤ a. If there is no critical cell σ with
a < f (σ) < b, then X(a) is a deformation retract of X(b). Otherwise, if there is a unique
critical cell σ with a < f (σ) < b, then X(b) is obtained from X(a) by attaching the cell σ.
Moreover, there is a complex
(1.1)
∂e
. . . → Mi → . . . → M1 → M0 → 0
where Mi is a free Z-module on the critical i-cells of X, whose homology is the homology
of the space X. We call the differential ∂e the Morse differential (see Definition 2.5).
Using work of Abrams [A], Farley and Sabalka were able to impose a discrete Morse structure on the spaces UConf n (G), where G is any graph [FS]. We will spend a good amount
of time recounting the construction of Farley and Sabalka in Section 2.2. Once we have accomplished this, our strategy will be to develop a strong understanding of the critical cells.
More specifically, we will work towards understanding the changing behaviors of the critical
cells as n varies.
Let G be a graph with E essential edges, and write SG := Z[x1 , . . . , xE ] for the integral
polynomial ring in E-variables. We will prove the following in Section 3.2.
(∗) For each i, there exists a finitely generated graded SG -module Mi,• for which Mi,n is
a free Z-module with basis vectors indexed by the critical i-cells of UConf n (G).
Specializing to the case where G is a tree, work of Farley [Fa] implies that the Morse differential is always trivial. Using the complex (1.1) we obtain the following.
Theorem C. Let G be a tree with E essential edges, and let SG denote the integral polynomial ring in E variables. Then for each i and n, the action of SG on the critical cells
of UConf n (G) described in Section
L 3.2 imposes the structure of a finitely generated graded
SG -module on the abelian group n≥0 Hi (UConf n (G)). In particular, there exists a finitely
4
ERIC RAMOS
generated graded SG -module Hi such that,
(Hi )n ∼
= Hi (UConf n (G)).
A result of Ghrist and Abrams (see Theorem 2.13) implies the spaces UConf n (G) are aspherical. It follows immediately from this that Hi (UConf n (G)) ∼
= Hi (Bn G). With this in mind,
the first part of Theorem A simply follows from the existence of the Hilbert polynomial. Of
course, the above theorem does not tell us anything about the degree of the Hilbert polynomial, nor does it bound its obstruction. To accomplish this, we must first prove a structure
theorem about the modules Hi .
To state this theorem, we first recall the definition of a squarefree monomial ideal. We
say an ideal I ⊆ Q[x1 , . . . , xd ] is a squarefree monomial ideal, if it contains a generating set
of monomials, none of which are divisible by a square in Q[x1 , . . . , xd ]. These ideals are the
subject of Stanley-Reisner theory, and have many desirable properties. For instance, much
is known about their Hilbert polynomial (see [MS] for a reference on the subject).
Theorem D. Let G be a tree, and let Hi denote the SG -module of Theorem C. Then Hi ⊗Z Q
is isomorphic to a direct sum of graded twists of squarefree monomial ideals, each having
dimension at most ∆iG .
We will find that this theorem implies the conclusions of Theorem A. In fact, we will be able
to compute the polynomial Pi associated to Hi explicitly in terms of invariants of the tree
G (see Theorem 3.17).
1.3. An overview of the paper. In the next section, we will spend time developing necessary background. This includes short summaries of discrete Morse theory (Section 2.1), and
configuration spaces of graphs (Sections 2.2 and 2.3). Following this, we use the machinery
developed in these preliminary sections to prove the statement (∗) (Section 3.2). Finally,
we specialize the the case of trees, and use enumerative combinatorial methods to prove
Theorem A via an explicit computation of the Hilbert polynomial (Sections 3.3 and 3.4).
To finish the paper, we briefly consider the case of a general graph G. Note that while most
of the explicit results in this paper are limited to the case where G is a tree, the statement (∗) will hold for any graph G. A result like Theorem C will therefore hold for general
graphs so long as we know that the Morse differential commutes with the action of SG on
Mi,• . It is the belief of the author that the action of SG , or perhaps a slight alteration
there of, will indeed commute with this differential. Unfortunately, it is known that this
differential can become tremendously complicated as G increases in complexity (see [KP]).
In any case, the methods in this work therefore provide at least a strategy for proving stability results for more general graphs. In the final sections, we discuss some implications of this.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
5
Acknowledgments
The author would like to send thanks to Jordan Ellenberg for his many insights, as well as
his help in editing this paper. The author would also like to send special thanks to Steven
Sam for his vital advice in approaching the primary problem of this work. A great amount of
gratitude must also be sent to Daniel Farley, whose expertise in the field was invaluable to the
author during his learning of the material. Finally, the author would like the acknowledge the
generous support of the National Science Foundation through NSF-RTG grant DMS-502553.
2. Preliminaries
2.1. Discrete Morse theory. We now take the time to briefly summarize the key points in
Forman’s discrete Morse theory [Fo]. We will largely be following the exposition of Forman
[Fo][Fo2], Farley and Sabalka [FS], and Ko and Park [KP].
In the introduction, we spent some time discussing the notion of discrete Morse function.
One thing that should stand out about this definition is that the literal values of the function
are immaterial. Namely, the classification of critical cells is unchanged by composition with
any strictly monotone function R → R. In many cases it is often easier to construct the
relationships between the cells, rather than the discrete Morse function itself. This hints
towards the construction of what are known as discrete vector fields. We will use this approach during the exposition of this, and all future sections.
To begin, we first must place certain light restrictions on the spaces we will be working with.
Definition 2.1. Let X be a CW complex. A cell of X will always refer to an open cell
in X. Given a cell σ of dimension i, we will often write σ (i) to indicate that σ has dimension i. We will write K to denote the set of cells of X, and Ki to denote the set of i-cells of X.
A cell τ (i) ⊆ σ i+1 is said to be a regular face of a cell σ (i+1) if, given a characteristic map
Φσ : D i+1 → X for σ (i+1) , Φ−1
is a homeomorphism.
σ (τ ) is a closed ball, and the map Φσ |Φ−1
σ (τ )
We say that the complex X is regular if given a pair of cells τ (i) ⊆ σ (i+1) , τ (i) is a regular
face of σ (i+1) . Equivalently, X is regular if and only if the attaching map of each of its cells
is a homeomorphism.
We will assume throughout most of our exposition that X is a regular CW complex. In
his original paper on discrete Morse theory [Fo], Forman proves some of his results without
the requirement that X be regular. The spaces UConf n (G), for any graph G, are actually
cubical complexes, which are certainly regular CW complexes. Therefore, the condition that
X be regular is not restrictive for what we need.
Definition 2.2. Let X be a regular CW complex. A discrete vector field V on X is
a collection of partially defined functions Vi : Ki → Ki+1 satisfying the following three
conditions for each i:
1. Vi is injective;
2. the image of Vi is disjoint from the domain of Vi+1 ;
3. for any σ (i) in the domain of Vi , σ (i) is a face of Vi (σ (i) ).
6
ERIC RAMOS
Given a regular CW complex X equipped with a discrete vector field V , a cellular path
between two cells α(i) and β (i) is a finite sequence of i-cells
(i)
(i)
(i)
(i)
α(i) = α0 , α1 , . . . , αl−1 , αl = β (i)
(i)
(i)
such that αj+1 is a face of Vi (αj ). We say that the path is closed if α(i) = β (i) , and we say
(i)
(i)
it is trivial if αj = αk for all j, k.
A discrete vector field V is said to be a discrete gradient vector field if it admits no
non-trivial closed cellular paths.
If V is a discrete gradient vector field on a regular CW complex X, then we call a cell σ of
X redundant if σ is in the domain of Vi for some i, collapsible if it is in the image of Vi
for some i, and critical otherwise.
Proposition 2.3 ([FS] Proposition 2.2, [Fo] Theorem 3.4). Let X be a regular CW complex
equipped with a discrete gradient vector field V . Consider the filtration
∅ = X0′′ ⊆ X0′ ⊆ X1′′ ⊆ X1′ ⊆ . . . ⊆ Xn′′ ⊆ Xn′ ⊆ . . .
where Xi′ is the i-skeleton of X with the redundant cells removed, and Xi′′ is the i-skeleton
of X with both the redundant and critical cells removed. Then:
1. For any i, Xi′ is obtained from Xi′′ by attaching mi i-cells to Xn′′ along their boundaries,
where mi is the number of critical i-cells of the discrete gradient vector field V .
′′
2. For any i, Xi+1
deformation retracts onto Xi′ .
The above proposition leads to one notable corollary, which we record now.
Corollary 2.4 ([FS] Proposition 2.3, [Fo] Corollary 3.5). Let X be a regular CW complex
equipped with a discrete gradient vector field V . Then X is homotopy equivalent to a CW
complex with precisely mi i-cells for each i, where mi is the number of critical i-cells of V .
Just as is the case with traditional Morse theory, the decomposition of the space X given
by Corollary 2.4 can be used to compute the homology groups of X. For simplicity, we will
state the construction for cubical complexes, although the general case is similar.
Definition 2.5. Let X be a cubical complex, and write Ci (X) for the free abelian group of
i-cells of X. If c ∼
= [0, 1], then we define
= I1 × . . . × Ii is an i-cell of X, with Ij ∼
cjτ = I1 × . . . × Ij−1 × {0} × Ij+1 × . . . × Ii ,
cjι = I1 × . . . × Ij−1 × {1} × Ij+1 × . . . × Ii .
This allows us to define a boundary morphism
∂ : Ci (X) → Ci−1 (X)
given by
∂(c) :=
X
j
cjτ − cjι ,
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
7
turning C• (X) into a chain complex. It is a well known fact that the homology of this chain
complex is the usual homology of the space X.
Further assume that X is equipped with a discrete gradient vector field V . Then we have a
map R : Ci (X) → Ci (X) defined by
if c is collapsible
0
R(c) = c
if c is critical
±∂(V (c)) + c otherwise,
i
where the sign of ∂(Vi (c)) in the above definition is chosen so that c has a negative coefficient.
The property that V has no non-trivial closed paths implies that Rm (c) = Rm+1 (c) for all
m ≫ 0 and all i-cells c [Fo]. We set R∞ (c) to be this stable value.
For each i, let Mi denote the free abelian group with basis indexed by the critical i-cells of
V . Then the Morse complex associated to V is defined to be
∂e
M• : . . . → Mn → . . . → M1 → M0 → 0,
where boundary map ∂e is given by
e := R∞ (∂(c))
∂(c)
The map ∂e is known as the Morse differential.
Theorem 2.6 (Theorem 8.2 [Fo], Theorem 7.3 [Fo2]). For all i there are isomorphisms,
Hi (X) ∼
= Hi (M• )
2.2. The configuration spaces of graphs. In this section we review necessary facts about
configuration spaces of graphs. In the next section, we will explain how the techniques of
discrete Morse theory apply to these spaces.
Definition 2.7. A graph is any compact, connected CW complex of dimension one. A tree
is a topologically contractible graph. Given a graph G with vertex v, we will write µ(v) to
denote the degree of the vertex v. Note that any loop on v contributes 2 to its degree count.
We say that v is essential if µ(v) ≥ 3. If µ(v) = 1, we say that v is a boundary vertex,
and the unique edge connected to v is called a boundary edge.
The configuration space on n points of G is the topological space,
Conf n (G) := {(x1 , . . . , xn ) ∈ Gn | xi 6= xj if i 6= j}.
We note that there is a natural fixed-point-free action on Conf n (G) by the symmetric group
Sn . The unordered configuration space on n points of G is the quotient space,
UConf n (G) := Conf n (G)/Sn .
8
ERIC RAMOS
For the majority of this paper, we will work with the spaces UConf n (G). Note that many
of the structural theorems discussed in this section will apply to both spaces.
In order to apply discrete Morse theory to questions about these configuration spaces, we
will first need to place a CW complex structure on them. To accomplish this, we use a
theorem of Abrams [A].
Definition 2.8. The discretized configuration space on n points over G, Dn (G), is
the CW subcomplex of Gn spanned by cells of the form
σ1 × . . . × σn
where σi ⊆ G is a cell (i.e. an edge or vertex) of G, and σi ∩ σj = ∅ whenever i 6= j. We
write UDn (G) to denote the quotient of Dn (G) by the action of the symmetric group.
Theorem 2.9 ([A] Theorem 2.1, [KKP] Theorem 2.4, [PS]). Let G be a graph, and assume
that G satisfies the following two properties:
1. each path connecting distinct vertices of degree 6= 2 has length at least n − 1;
2. each homotopically essential path connecting a vertex to itself has length at least n + 1.
Then the inclusions Dn (G) ֒→ Conf n (G), and UDn (G) ֒→ UConf n (G) are homotopy equivalences.
Remark 2.10. Note that in the first cited source, Abrams states the theorem assuming that
each path connecting distinct vertices of degree 6= 2 has length at least n + 1. It is noted
after the proof that the version of the theorem stated above is true, and a brief argument is
given for how it is proven. In the second source, Kim, Ko, and Park give a formal argument
for this improvement. In the third source, Prue and Scrimshaw provide a discrete Morse
theory argument, which is independent of the first two sources. For our purposes, the exact
number of vertices needed is unimportant, as we can always just subdivide the edges of G
more if needed.
Note that subdividing edges of a graph G does not impact the configuration spaces Conf n (G)
and UConf n (G). For much of what follows, we will often just assume, without explicit mention, G is subdivided enough so that the homotopy equivalence of Theorem 2.9 holds.
Theorem 2.9 implies that Conf n (G) and UConf n (G) are homotopy equivalent to cubical
complexes of dimension n. In fact, we will be able to do better than this.
Theorem 2.11 ([Gh] Theorem 3.3). Let G be a graph which is not homeomorphic to S 1 .
Then Conf n (G) and UConf n (G) are homotopy equivalent to CW complexes of dimension
NG , where NG is the number of essential vertices of G.
Remark 2.12. If G is homeomorphic to S 1 , then UConf n (G) is easily seen to be homotopy
equivalent to a circle for all n. Throughout this work, and the literature in general, it is a
recurring theme that certain theorems only apply to graphs which are neither S 1 nor the
interval. It is interesting to observe that these two graphs are precisely those which are
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
9
homeomorphic to compact manifolds.
Note that Ghrist originally proved Theorem 2.11 using more classical topological means. We
will later see it naturally falls out of the discrete Morse structure that Farley and Sabalka
placed on UDn (G). This was first noted by Farley and Sabalka in [FS].
One remarkable thing to note about Ghrist’s theorem is that the dimension of the configuration spaces of graphs is independent of n. This behavior is in stark contrast to the behavior
of configuration spaces of smooth manifolds of dimension ≥ 2.
In his paper [Gh], Ghrist also proves that configuration spaces of graphs are, in fact, aspherical. This result was later reproven by Abrams [A, Theorem 3.10], where he shows that both
Dn (G) and UDn (G) are universally covered by CAT(0) complexes.
Theorem 2.13 ([Gh] Theorem 3.1, [A] Corollary 3.11). Let G be a graph. Then Conf n (G),
and hence UConf n (G), are aspherical. That is to say, πk (Conf n (G)) = 0 for k > 1.
Note that this theorem is analogous to that which says the configuration spaces of the plane
are aspherical. In that case, the fundamental groups of the ordered and unordered spaces
are the Artin pure braid groups, and the Artin braid groups, respectively. We borrow this
terminology for our context as well.
Definition 2.14. Let G be a graph. The braid group on n strands of G is defined to be
Bn G := π1 (UConf n (G)).
We similarly define the pure braid group on n strands as
Pn G := π1 (Conf n (G))
The study of the braid groups of graphs is still a very active area of research. See [KKP][KP][FS2]
for more on these groups.
As an immediate corollary to Theorem 2.13, we obtain the following.
Corollary 2.15. Let G be a graph. then there are isomorphisms
Hi (Conf n (G)) ∼
= Hi (Pn G), Hi (UConf n (G)) ∼
= Hi (Bn G).
The goal for this paper is to establish a methodology for understanding stability phenomena
in the groups Hi (Bn G), in the spirit of modern trends of asymptotic algebra. Note that we
will spend very little time considering pure braid groups. In fact, there doesn’t seem to be
much literature about the homology of these groups, as they are vastly more complicated
than the braid groups [KP][BF].
10
ERIC RAMOS
The following theorem of Farley exactly computes the homology groups when G is a tree.
Theorem 2.16 ([Fa]). Let G be a tree. Then the group Hi (Bn G) is free for all n ≥ 0.
Note that the theorem of [Fa] is slightly more general than this. We will discuss the more
general version in later sections.
Expanding upon the work of Farley, the following theorem of Ko and Park suggests that the
tree case is indicative of a more general phenomenon.
Theorem 2.17 ([KP] Theorem 3.6). Let G be a graph. Then G is planar if and only if
H1 (Bn G) is torsion free for any, and therefore all, n ≥ 2. Moreover, in the case where G is
not planar, all torsion is 2-torsion.
Kim, Ko, and Park proved that G is planar if and only if H1 (B2 G, Z) is torsion free in [KKP,
Theorem 5.5]. It is also conjectured in that work that Theorem 2.17 is true. It should be
noted that [KP, Theorem 3.6] is far stronger than what we have written above. In fact,
their result explicitly computes the groups H1 (Bn G, Z) in terms of combinatorial invariants
of the graph G. It follows from their computation that the amount 2-torsion of the group is
unvarying in n. Moreover, if G is biconnected - that is, G requires the removal of at least 2
vertices to disconnect it - then H1 (Bn G, Z) ∼
= H1 (B2 G, Z) for all n ≥ 2 [KP, Lemma 3.12].
In the final section of this paper we will conjecture an extension of this fact to the higher
homologies (see Conjecture 4.3).
To finish this section, we record a result of Gal on the Euler characteristic of these spaces.
Note that Gal proves a more general theorem for computing the Euler characteristic of the
configuration space of any simplicial complex.
Theorem 2.18 ([Ga] Theorem 2). Let G be a graph, and set
X χ(Conf n (G))
e(t) :=
tn .
n!
n≥0
Then,
e(t) =
where E is the number of edges of G.
Q
v (1
− (1 − µ(v))t)
(1 − t)E
Remark 2.19. Conf n (G) is an n!-fold cover of UConf n (G). It follows from this that the
above formula can be easily used to compute the Euler characteristic of UConf n (G) as well.
Definition 2.20. Let G be a graph which has at least one essential vertex, and let E denote
the set of essential vertices of G. Then an essential edge of G is a connected component
of G − E.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
11
Corollary 2.21. Let G be a graph with at least one essential vertex, and let E denote the
number of essential edges of G. Then there is a polynomial P of degree E − 1, for all n ≥ 0,
such that the function
n 7→ χ(UConf n (G))
is equal P (n).
Proof
We first note that smoothing the degree 2 vertices of G does not impact the spaces UConf n (G).
We may therefore assume without loss of generality that G does not have any such vertices.
In this case, the Euler characteristic χ(UConf n (G)) is the coefficient of tn in the power series
expansion of
Q
v essential (1 − (1 − µ(v)t))
(1 − t)E
A straight forward enumerative combinatorics argument implies that the n-th coefficient of
the power series expansion of this expression is a polynomial of degree exactly E − 1 for n
sufficiently large. It remains to show that this agreement begins at n = 0.
Assume that G has NG essential vertices. It is a standard fact from enumerative combinatorics that the coefficients of the power series expansion of a rational function of the form
f (x)
will agree with a polynomial for all n so long as deg(f ) < l. This is the case in our
(1−t)l
specific instance, so long as NG > 1.
Theorem 2.18 was largely the inspiration for this work. It suggests that the asymptotically
the Betti numbers of Hi (Bn G) are polynomial. In this work we will prove this suggestion in
the case where G is a tree. We will also provide a setup in the case where G is a general
graph, which hopefully will be able to illustrate this behavior in that case as well.
2.3. The discrete Morse theory of UConf n (G). In this section we will outline the discrete Morse structure on UConf n (G) developed in the work of Farley and Sabalka [FS]. We
begin by fixing n, and assuming the graph G is sufficiently subdivided for Theorem 2.9 to
hold for UConf n (G).
Fix a spanning tree T for G, as well as an embedding of T into the plane. We will label the
vertices of T by applying a depth-first search. Concretely, begin by choosing a boundary
vertex v0 to be the root of T , and label it with the number 0. Continue down the boundary
edge adjacent to v0 , labeling any vertices encountered along the way with increasing labels.
If at any point an essential vertex is encountered, then travel down the leftmost - relative to
the current direction of travel - edge whose vertices have not been labeled. If a boundary
vertex is encountered, then one returns to the most recently passed essential vertex. An
example of a correctly labeled tree is given in Figure 1.
For the remainder of this section, G and T will be as in the previous paragraphs. As before
we will write K to denote the set of cells of UConf n (G), while Ki will denote the set of i-cells
of UConf n (G).
Definition 2.22. Given an edge e of G, we write ι(e) to denote its largest, with respect to
the given labeling, vertex, and τ (e) to denote its smallest vertex. If v is a vertex of G, which
is not v0 , then we write e(v) to denote the unique edge of T for which ι(e(v)) = v. We note
12
ERIC RAMOS
0 1
6
5
4
3
7 8
2
9
10
Figure 1. A tree which is properly labeled. Note that this tree is also sufficiently subdivided to apply Theorem 2.9 with n = 3.
that e(v) will be the first edge on the unique path within T from v to v0 .
Let c = {vj1 , . . . , vjn−i , er1 , . . . , eri } be an i-cell of UDn (G). If
6 ∅,
τ (e(vj )) ∩ {vj1 , . . . , vjn−i , ι(er1 ), τ (er1 ), . . . , ι(eri ), τ (eri )} =
then we say that vj is blocked in c. Note that by convention v0 is always blocked.
We define a partial function Vi : Ki → Ki+1 inductively in the following way. Given an i-cell
c = {vj1 , . . . , vjn−i , er1 , . . . , eri }, assume without loss of generality that vj1 is an unblocked
vertex of c which is smallest with respect to the labeling of T . Assuming that c is not in the
image of Vi−1 , we set
Vi ({vj1 , . . . , vjn−i , er1 , . . . , eri }) := {vj2 , . . . , vjn−i , er1 , . . . , eri , e(vj1 ))}
If c does not have any unblocked vertices, or if c is in the image of Vi−1 , then Vi (c) is undefined.
Theorem 2.23 ([FS] Sections 3.1 and 3.2). The collection of partial functions Vi : Ki → Ki+1
form a discrete gradient vector field on UDn (G).
While the spaces UConf n (G) can be high dimensional, we can still visualize their cells.
Indeed, it is often useful to think of the cells of UConf n (G) as being subsets of edges and
vertices of the graph G itself. In Figure 2 we see examples of a critical, a collapsible, and
a redundant 1-cell of V, where G is taken to be the tree in Figure 1 and n = 3. In these
examples, the cell {v0 , v1 , e9,10 } is collapsible, as
{v0 , v1 , e9,10 } = V0 ({v0 , v1 , v10 }),
and v10 is the smallest unblocked vertex. The cell {v1 , v5 , e4,7 } is redundant, as v1 is not
blocked, and it is not in the image of V0 . Indeed, the only cell that could possibly map to it
would be {v1 , v5 , v7 }. However, in this case v1 is the smallest unblocked vertex. Finally, the
cell {v0 , v5 , e4,7 } is critical, as its every vertex is blocked, and it is not collapsible.
From these examples, one can imagine there existing a way to classify all cells of the three
types. To write down such a classification, we first need a bit of nomenclature.
Definition 2.24. Let c = {vj1 , . . . , vjn−i , er1 , . . . , eri } be an i-cell of UDn (G). We say that
an edge e ∈ c is order respecting in c if e is in T , and for all v ∈ c such that τ (e(v)) = τ (e),
the label of v in T is larger than the label of ι(e).
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
0 1
6
5
4
3
9
10
7 8
2
0 1
6
5
4
3
9
10
7 8
2
0 1
6
5
4
3
13
7 8
2
9
10
Figure 2. The critical 1-cell {v0 , v5 , e4,7 }, the collapsible 1-cell {v0 , v1 , e9,10 },
and the redundant 1-cell {v1 , v5 , e4,7 }, respectively.
Remark 2.25. Note that edges e ∈ G−T - which we call deleted edges - are, by definition,
never order respecting.
Theorem 2.26 (The Classification Theorem, [FS] Theorem 3.6). Let c be a cell of UDn (G),
and let T be a choice of spanning tree of G equipped with a planar embedding. Then:
1. c is critical if and only if it contains no order respecting edges and all of its vertices
are blocked;
2. c is redundant if and only if
(a) it contains no order respecting edges and at least one of its vertices is unblocked
or,
(b) it contains an order respecting edge (and thus a minimal order respecting edge e)
and there is some unblocked vertex v such that the label of v in T is less than that
of ι(e);
3. c is collapsible if and only if it contains an order respecting edge (and thus a minimal
order respecting edge e) and, for any vertex v such that the label of v in T is less than
that of ι(e), v is blocked.
It is not hard to check that the previous examples satisfy the conditions of the Classification
Theorem.
One useful fact about the discrete gradient vector field V is that its critical cells are, in some
sense, the most restrictive of the three possible types. Indeed, we may expand upon the
classification of critical cells in the following way.
Lemma 2.27. Let c = {vj1 , . . . , vjn−i , er1 , . . . , eri } be a critical i-cell of V .
1. For each k in {1, . . . , i}, either erk is a deleted edge, or τ (erk ) is an essential vertex of
T.
2. Let c′ be a critical i-cell with edges er1 , . . . , eri , such that the number of its vertices in
each component of T − {er1 , . . . , eri } agrees with the number of vertices of c in these
same components. Then, c = c′ .
Proof
For the first statement, let e be an edge of c which is not deleted. If τ (e) = v0 , then it is
clear that e is order respecting. We note that τ (e) is not a boundary vertex in any other
case by how the labeling on T was defined. It therefore remains to show that τ (e) does not
have degree 2. Indeed, if this were the case, then for any vertex v of G, τ (e(v)) 6= τ (e). In
14
ERIC RAMOS
particular, e must be order preserving. This proves the first claim.
For the second claim, one uses the fact that the vertices of c must all be blocked.
Note that the observations made in the above lemma were originally pointed out by Farley and Sabalka when they defined the vector field V [FS, Section 3.3]. We collected these
observations in Lemma 2.27 so that they could be easily referred back to during future sections.
We will find that this lemma is critical in defining the polynomial ring structure on
L
n Hi (Bn G). In the next section, we will find that the deleted edges of G can be chosen
so that they remain unchanged as n increases and we repeatedly subdivide the graph. The
above lemma also suggests that those edges of critical cells contained in the tree T are unchanging in n. While this is not literally the case, as all edges of T are being subdivided as
n increases, Lemma 2.27 tells us that the important information encoded by such an edge is
the essential vertex of its tail, as well as the direction it is leaving this essential vertex. For
this reason we will often be a bit loose with our language and claim that two critical cells
with different choices of n have the “same” edges.
Definition 2.28. Let c be a critical cell of the discrete gradient vector field, and let e ∈ T be
an edge in c. The classification theorem implies that there must exist some vertex v ∈ c such
that the label of v in T is smaller than that of ι(e), while τ (e(v)) = τ (e). We refer to such
a vertex as a witness for e in c. If v is the only witness for e in c, then we say v is necessary.
3. The proof Theorem A
3.1. The setup. We use this section to fix the notation which will be used throughout the
proofs of the main theorems. Note that all of the constructions presented in the previous
sections were already well established in the literature prior to this work. To the knowledge
of the author, the main construction of the following sections - namely the action of SG on
the critical cells of UConf n (G) - does not appear elsewhere in the literature.
As was the case in all previous sections, let G denote a graph, which is neither a circle nor a
line segment. We will reserve E to denote the number of essential edges of G, while NG will
denote the number of essential vertices of G.
Next, we must construct our spanning tree T , as in Section 2.3. To do this, we first subdivide
every edge of G, which connects two essential vertices, once. Note that this includes any
loops of G. Once this is done, we choose a spanning tree T of G which satisfies the following:
(3.1)
Every edge of G not in T is adjacent to an essential vertex of G.
Note that it is not entirely obvious that such a T exists for an arbitrary graph G. This fact
was proven by Farley and Sabalka during the proof of [FS, Theorem 4.4].
Remark 3.1. It is noted by Farley and Sabalka that if one chooses T to satisfy (3.1), then
Lemma 2.27 and Corollary 2.4 imply Theorem 2.11.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
15
Having chosen our spanning tree T to satisfy (3.1), we observe that for n ≥ 2, we can sufficiently subdivide G for Theorem 2.9 by subdividing T . This follows from the fact that we
began by subdividing all edges of G which connected essential vertices. We will often not
differentiate the spanning trees chosen for each n for this very reason.
3.2. The modules Mi,• . Recall from Section 2.1 the Morse complex
. . . → Mi → . . . → M1 → M0 → 0
where Mi is the free Z-module with basis indexed by critical cells. In our case, we write Mi,n
to denote the free Z-module with basis indexed by the critical cells of the Farley-Sabalka
discrete gradient vector field of UDn (G).
Let e1 , . . . , eE denote the essential edges of the graph G. Then we set
SG := Z[xe1 , . . . , xeE ]
Our goal for the remainder of this section will be to argue, for each i, that Mi,• has the
structure of a finitely generated graded SG -module.
Definition 3.2. Fix i and n, and let e denote an essential edge of G. Given a critical i-cell
c = {vj1 , . . . , vjn−i , er1 , . . . , eri } of UDn (G), we define xe · c to be the unique critical i-cell of
UDn (G) obtained from c by adding a vertex to the connected component of T −{er1 , . . . , eri }
containing e. More precisely, xe · c is obtained from c by adding the smallest vertex on the
connected component containing e. This is well defined by Lemma 2.27, as well as the choice
of the tree T , as T will contain a unique representative of each essential edge of G by construction. This turns the collection {Mi,n }n into a graded SG -module, which we denote Mi,• .
One can visualize the action of SG described in the above definition in the following way.
The labeling on T induces a natural flow on T . Namely, all edges flow towards the root of
T . Let c = {vj1 , . . . , vjn−i , er1 , . . . , eri } be a critical cell. We imagine the vertices vj as being
particles drifting in the direction of the flow, while the edges ej , along with their endpoints,
are stationary blockades. For any essential edge e, the action of xe on c involves placing a
new particle somewhere on e, and allowing it to flow until it too is blocked.
We provide an illustration of the action of SG in Figure 3.
Lemma 3.3. For each i, the module Mi,• is finitely generated over SG .
Proof
We claim that every critical cell in Mi,n , with n > 2i, can be obtained from a critical cell
in Mi,n−1 . Indeed, if c is a critical cell in Mi,n , then at least one vertex is not a necessary
witness of some edge in e. Remove the vertex, among those which are not necessary witness
vertices, which occupies the maximal position with respect to the labeling on T . Removing
this vertex leaves us with a critical cell c′ in Mi,n−1, as this removal cannot create unblocked
vertices nor order respecting edges. It follows that c is the image of c′ under the action of
the appropriate essential edge.
16
ERIC RAMOS
0 1
6
5
4
3
9
10
7 8
2
0 1 2
9
8
7
6
5
4
10 11 12
3
13
14
15
Figure 3. An illustration of multiplication by xe , where e is the essential edge
containing the boundary vertex v6 of the tree G of Figure 1.
Remark 3.4. It follows from the above lemma that the number of critical i-cells grows, as
a function of n, like a polynomial of degree ≤ E − 1. In fact, if i < NG , we claim that this
polynomial must have degree strictly less than E − 1. First, we note that the action of SG
on a critical cell c does not affect the edges of c. Because of this it follows that Mi,• can
be expressed as a direct sum of graded SG -modules, where each summand corresponds to
a choice of edges in our critical i-cell. For any fixed critical i-cell c, one observes that two
variables xe and xe′ will act identically whenever e and e′ are on the same connected component of G with the edges of c removed. It follows that the summand of Mi,• containing
this cell has Hilbert polynomial of degree strictly less than E whenever i < NG . The Hilbert
polynomial of the whole of Mi,• is a sum of such polynomials, and therefore also has degree
strictly less than E. This will be used in the final section of the paper.
Lemma 3.3, and Theorem 2.6 now imply some asymptotic data about the homology groups
Hi (Bn G).
Theorem 3.5. Let G be a graph with E essential edges. Then for all i ≥ 0, there exists a
polynomial Pi ∈ Q[t], such that for all n ≫ i
dimQ (Hi (Bn G; Q)) ≤ Pi (n).
3.3. The case of trees. In this section, we begin to explore the specific case where G is a
tree. The work of Farley [Fa] will allow us to conclude quite a bit more than we are able to
in the general case. To begin with, we state a refined version of Theorem 2.16.
Theorem 3.6 ([Fa]). Let G be a tree. Then for each n the Morse complex Mi,n has trivial
differential. In particular,
Hi ∼
= Mi,• ,
as SG -modules, where Hi is the module of Theorem C.
To state our main stability theorem, we begin with the following definition.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
17
Definition 3.7. Let G be a tree. Then we define the quantity
∆iG :=
max
{dimQ (H0 (G − {vj1 , . . . , vji }; Q))}.
{{vj1 ,...,vji }|vjk essential}
G
For example, ∆1G is the maximum degree of G, while ∆N
is E, the number of essential
G
0
i
edges. By convention, ∆G = 1 while ∆G = 0 for i > NG .
Theorem 3.8. Let G be a tree, and fix i ≥ 0. Then the Betti number bi (n) := dimQ (Hi (Bn (G); Q))
is equal to a polynomial of degree ∆iG − 1 for n ≫ i.
Proof
The Theorem 3.6 and the work of the previous sections imply that bi (n) is eventually a
polynomial; it remains only to show that this polynomial has the claimed degree.
We begin by partitioning the basis vectors of Mi,n according to the collection of edges which
appear. The action of SG does not impact the edges, and so each of these partitions will
correspond to a summand of Mi,n . Moreover, considering any one of these summands, we
observe that the variables induce at most ∆iG distinct operators on Mi,n . It follows that the
degree of the Hilbert polynomial of Mi,• is at most ∆iG − 1. To finish the proof, we will show
that Mi,• has a summand whose Hilbert polynomial has degree ∆iG − 1.
Fix the i essential vertices of G whose absence realizes the maximum in Definition 3.7. For
each essential vertex in this fixed list, let e be an edge adjacent to it in the second leftmost
direction, relative to the direction of the root. Note that any critical cell containing these
edges has a unique essential edge which can house a witness vertex. There is therefore a
critical i-cell in Mi,2i constructed by choosing each of these edges, along with the unique
witness vertex for each. Denote this cell by c. The submodule generated by c is a summand
of Mi,• . We claim that this submodule our desired summand.
Any i-cell of Mi,n which shares the same edges of c - i.e. those basis vectors in the n-th
graded piece of SG · c - is entirely determined by distributing n − 2i vertices to one of ∆iG
components. Thus,
n − 2i + ∆iG − 1
.
rankZ ((SG · c)n ) =
∆iG − 1
as desired.
The only piece of Theorem A which remains to be proven is in showing that bi (n) agrees
with a polynomial for all n ≥ 0. To accomplish this, we must first prove Theorem D. As a
first simplification, we note that our answer is unaffected by changing basis to Q, and may
therefore assume we are working over a rational polynomial ring. This grants us access to
many classical approaches to solving the problem. We will also make heavy use of the direct
sum decomposition of Mi,• described in the proof of Theorem 3.8.
Definition 3.9. Let G be a tree. The notation
v := (vj1 , . . . , vji )
18
ERIC RAMOS
will always refer to an i-tuple of essential vertices of G, appearing in the order induced
by the labeling on G. We write µ(v) to denote the number of connected components of
G − {vj1 , . . . , vji }. Given v, let l := (l1 , . . . , li ) denote an i-tuple of positive integers such
that
1 ≤ lk ≤ µ(vjk ) − 2 for all k.
We will also set
X
|l| :=
lk .
k
Given a pair (v, l), we associate to it the summand of Mi,• ⊗Z Q generated in degree 2i by
critical cells c satisfying the following two properties:
1. If e ∈ c is an edge, then τ (e) = vjk for some k, and;
2. if e ∈ c has τ (e) = vjk , then there are precisely lk essential edges which can house
witness vertices of e.
Note that these two pieces of data uniquely determine the edges which are allowed to appear
in cells which form the bases for the graded pieces of the summand. Denote this summand
by N (v,l) .
We observe that two variables xe1 , xe2 of SG act identically on N (v,l) if and only if e1 and e2
are contained in the same connected component of G − {vj1 , . . . , vji }. Let S (v,l) denote the
quotient of SG ⊗Z Q by xe2 − xe2 . In particular, S (v,l) is isomorphic to a rational polynomial
ring in µ(v) variables, and N (v,l) is a finitely generated graded module over S (v,l) .
Remark 3.10. If i = 0, then v and l are empty tuples with µ(v) = 1. If i > NG , then by
convention N (v,l) = 0.
next goal will be to show that the Hilbert function
n 7→ dimQ (N (v,l) )
agrees with a polynomial for all n. Note that we already know that this Hilbert function is a
polynomial for n sufficiently large, and must only show that this agreement is the case for all
n. This will imply the same about the Hilbert function for Mi,•. We accomplish this through
the computations in the next section, although we spend time now to set some ground work
in this direction. Following this, we will explicitly compute this Hilbert polynomial. We
begin with the following.
Proposition 3.11. The module N (v,l) (i) is isomorphic to a squarefree monomial ideal of
S (v,l) .
Proof
We recall that a monomial ideal is called squarefree, if it is generated by monomials, none
of which are divisible by a square.
(v,l)
By definition, N (v,l) (i)i = N2i is spanned by critical cells containing the edges e1 , . . . , ei
determined by the pair (v, l) (see Definition 3.9), along with a single witness vertex for each.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
19
To define a map φ : N (v,l) (i) → S (v,l) , it suffices to specify where each of these cells is mapped.
We set for any such cell c
Y
(3.2)
φ(c) =
xe ,
e
where the product is over essential edges containing a witness vertex of c. Note that it is
impossible for a single essential edge to house a witness vertex for multiple edges, and so
the above monomial is indeed squarefree. Extending this map through the action of S (v,l)
defines our desired isomorphism.
Squarefree monomial ideals are some of the most well understood objects in commutative
algebra. They are also the subject of the so-called Stanley-Reisner theory (see [MS] for a
comprehensive reference on the subject).
3.4. Computing the Hilbert polynomial. We note that for i = 1, the polynomial describing the Betti number b1 (n) has been explicitly computed in terms of invariants of the
tree G. Indeed, this follows from the structure theorem of Ko, and Park [KP, Theorem 3.6],
and also from the work of Farley and Sabalka [FS]. The Hilbert polynomial has also been
computed for all i in the case where G is a tree with maximal degree 3 by Farley [Fa]. Aside
from these cases, no explicit description of the polynomial is known. That being said, the
work in this paper implies that computing this polynomial is no more difficult than computing the Hilbert polynomials of some squarefree monomial ideals. In other words, for any
tree G, the work of this paper reduces the task of computing Hi (Bn G) for all n to a finite
computation. We will proceed with this computation in this section.
Remark 3.12. Proposition 3.11 reveals that the monomial ideals which appear as graded
shifts of summands of Hi are quite simple. Because of this, we will find that the computation
of the Hilbert polynomial of Hi falls more in the realm of enumerative combinatorics rather
than Stanley-Reisner theory. It is likely that studying the cases of more general graphs will
require more robust commutative algebra, and we therefore proceed with the case that G is
a tree using that language. Our hope is that once the case of general graphs is completed,
this will aid in creating a uniform means of approaching asymptotic behaviors in the configuration spaces of graphs.
In this section we will use the decomposition of Proposition 3.11 to compute the Hilbert
polynomial of Mi,•. Proposition 3.11 implies that N (v,l) (i) is isomorphic to a squarefree
monomial ideal of S (v,l) . To ease computations, we reorder the variables of S (v,l) to satisfy
the following. Begin by labeling the essential edges adjacent to vj1 which house witness
vertices by x1 through xl1 , chosen in the order induced by our ordering of the vertices of G.
Next, label the essential edges housing a witness vertex adjacent to vj2 with the variables
xl1 +1 to xl1 +l2 , using the same rule. Continue in this fashion until |l| essential edges have
been labeled. We then use the remaining variables to label the final µ(v) − |l| essential edges,
chosen in any order. Our first observation is the following.
20
ERIC RAMOS
Lemma 3.13. There is an isomorphism of S (v,l) -modules,
N (v,l) (i) ∼
= (x1 , . . . , xl1 )⊗Q (xl1 +1 , . . . , xl1 +l2 )⊗Q . . .⊗Q (xPi−1 lj +1 , . . . , x|l| )⊗Q Q[x|l|+1 , . . . , xµ(v) ],
j=1
where (x1 , . . . , xl1 ) is considered as a module over Q[x1 , . . . , xl1 ], (xl1 +1 , . . . , xl1 +l2 ) a module
over Q[xl1 +1 , . . . , xl1 +l2 ], and so on.
Proof
Follows immediately from the isomorphism (3.2).
Definition 3.14. The Hilbert series associated to the module N (v,l) is the formal power
series
X
H(v,l) (t) :=
(dimQ Nn(v,l) )tn .
n≥0
Our next goal will be to express H(v,l) (t) as a rational function in t. From here standard
enumerative combinatorics will allow us to compute the Hilbert polynomial.
Proposition 3.15. The Hilbert series H(v,l) (t) can be expressed as the rational function,
Q
ti j (1 − (1 − t)lj )
.
H(v,l) (t) =
(1 − t)µ(v)
Proof
We will prove this using the fact that the Hilbert series is multiplicative over tensor products.
It is easy to see that the Hilbert series for the augmentation ideal of a rational polynomial
ring in d variables is equal to the rational function
1
1 − (1 − t)d
−
1
=
.
(1 − t)d
(1 − t)d
Lemma 3.13 therefore implies
i
H(v,l) (t) = t
Y 1 − (1 − t)lj
j
(1 − t)lj
!
1
(1 − t)µ(v)−|l|
=
ti
Q
− (1 − t)lj )
,
(1 − t)µ(v)
j (1
where the factor ti arises from the graded shift of N (v,l) .
This will allow us to prove our result on the obstruction to the Hilbert polynomial of Mi,• .
Corollary 3.16. Let G be a tree, and fix i ≥ 0. Then the dimension of Mi,n ⊗Z Q agrees
with a polynomial for all n ≥ 0.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
21
Proof
It suffices to prove the claim for the summands N (v,l) . Proposition 3.15 tells us that the
Hilbert series of N (v,l) takes the form
Q
ti j (1 − (1 − t)lj )
.
(1 − t)µ(v)
The
2.21 implies that it suffices to show the total degree of the numerator
Q proof of Corollary
i
lj
t j (1 − (1 − t) ) is strictly smaller than µ(v). Namely, we must argue
i + |l| < µ(v)
This follows from the fact that each essential vertex of v, say vjk , is adjacent to at least one
essential edge which can never house a witness vertex. Namely, the essential edge which is
maximal among all essential edges adjacent to vjk with respect to the ordering of the vertices
of G. Moreover, the essential edge containing the root can never house a witness vertex for
any essential vertex.
Put another way,
i + |l| ≤ i +
X
(µ(vjk ) − 2) = i + µ(v) − i − 1 = µ(v) − 1.
k
This concludes the proof.
We have now proven enough to explicitly compute the Hilbert polynomial of Mi,• , and consequently Hi (Bn G). The following theorem follows from the results of this section, as well
as the combinatorics of generating functions.
Theorem 3.17. Let G be a tree, fix i, r ≥ 0, and for any pair (v, l) write
X
li
(v,l)
r+i l1
.
···
ai,r :=
(−1)
di
d1
d1 +...+di =r
dj ≥1
Also let Piv (n) be the polynomial
Piv (n)
µ(v) − 1 + n − i
.
=
µ(v) − 1
Then for all n ≥ 0, the dimension of Hi (Bn , Q) is equal to the polynomial
(3.3)
|l|
XX
(v,l)
ai,r Piv (n − r).
(v,l) r=i
Remark 3.18. One observes from the above formulation that the polynomial Piv (n), as well
(v,l)
as the constants ai,r , only depend on the degree sequence of the graph G. It follows that the
homologies of the braid group of a tree only depends on its degree sequence, as we recorded
in Corollary B
Specializing Theorem 3.17 to the case wherein i = 1 yields a much simpler form.
22
ERIC RAMOS
Corollary 3.19. Let G be a tree. Then for all n ≥ 0, the dimension of H1 (Bn G, Q) agrees
with the polynomial
X µ(v) − 2 + n µ(v) − 2 + n − l
X µ(v)−2
.
−
µ(v)
−
1
−
l
µ(v)
−
1
v essential l=1
Proof
In this case we note that for an essential vertex v, and a chosen integer 0 ≤ l ≤ µ(v) − 2, we
have
v,l
r+1 l
.
a1,r = (−1)
r
Therefore the polynomial (3.3) can be written
l
X X
X µ(v)−2
µ(v) − 2 + n − r
r+1 l
.
(−1)
µ(v)
−
1
r
v essential l=1 r=1
Using the principle of inclusion-exclusion, it can be seen that
l
X
µ(v) − 2 + n − l
µ(v) − 2 + n
µ(v) − 2 + n − r
r+1 l
,
−
=
(−1)
µ(v) − 1 − l
µ(v) − 1
µ(v) − 1
r
r=1
as desired.
A result of this kind appears in the work of Farley and Sabalka, where they compute the
number of generators of the group Bn G [FS]. A formula for the case i = 1 is also found in
the results of Ko and Park [KP].
Another simple case is that in which every essential vertex of G has degree 3. In this case
Theorem 3.17 can be used to recover the following result of Farley [Fa].
Corollary 3.20. Let G be a tree whose every essential vertex has degree 3. Then for all
n ≥ 0, the dimension of Hi (Bn G; Q) agrees with the polynomial
n
NG
2i
i
Proof
The major thing to note in this case is that for any v, the only associated vector l is (1, . . . , 1).
Therefore,
(
1 if r = i
(v,l)
ai,r =
0 otherwise.
The formula (3.3) now becomes
n
NG
2i + 1 − 1 + n − 2i
µ(v) − 1 + n − 2i
NG
NG
NG
v
.
=
=
Pi (n−i) =
2i
i
2i
i
i
µ(v) − 1
i
This completes the proof.
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
23
4. Concluding remarks
In this section we consider what this work implies for the homology of graph braid groups
when G is a general graph rather than a tree.
Let G be a graph with NG essential vertices, and E essential edges. Also assume that G is
homeomorphic to neither S 1 nor a line segment. Then the work of Section 3.2 implies that
Mi,• carries the structure of a finitely generated SG -module. If G is not a tree, however, it
is no longer the case that the Morse differential is trivial. We therefore pose the following
question.
Question 1. Does the action of SG on Mi,• commute with the Morse differential? In other
words, is the Morse differential a morphism of graded SG -modules?
We note that altering our choice of spanning tree T has a non-trivial influence on the Morse
differential. It is therefore unclear whether the Morse differential will commute with the
action of SG for all choices of T . We therefore refine Question 1 as follows.
Question 2. Does there exist a choice of spanning tree T of G such that the action of SG
on Mi,• commutes with the Morse differential?
If the answer to these questions is affirmative, then we can immediately conclude many
things about the asymptotic behavior of these homology groups. Indeed, in this case the
Hilbert basis theorem would imply that there is a finitely generated SG -module Hi such that
(Hi )n ∼
= Hi (Bn G).
Corollary 4.1. Assume that Questions 1 and 2 have an affirmative answer. For each i, n
write
M M
(Z/pj Z)mp,i,j (n) .
Hi (Bn G) ∼
= Zbi (n) ⊕
p prime j≥1
Then:
1. there exists a polynomial Pi (t) ∈ Q[t] of degree ≤ E − 1 such that Pi (n) = bi (n) for all
n ≫ 0;
2. there exists a positive integer ei , independent of n, such that the exponent of Hi (Bn G)
is at most ei ;
3. for each prime p and each positive integer j, there is a polynomial Pi,pj (t) ∈ Q[t]
such that Pi,pj (n) = mp,i,j (n) for n ≫ 0. In particular, there is a finite set of primes
pi,1 , . . . , pi,l , and polynomials mi,pj (n) such that the order of the torsion subgroup of
mpi,1 (n)
Hi (Bn G) is precisely p1
mpi,l (n)
· · · pl
.
Proof
Follows from standard facts in the study of graded modules over integral polynomial rings.
The work of Ko and Park [KP, Theorem 3.6] that H1 (Bn G) has a torsion subgroup annihilated by 2, and that the multiplicity of this torsion is eventually constant in n. Their work
also implies that the rank of H1 (Bn G) will be constant whenever G is sufficiently connected.
In contrast to this, affirmative answers to Questions 1 and 2 imply that the top homology
of Bn G always grows as fast as possible.
24
ERIC RAMOS
Theorem 4.2. Assume that Questions 1 and 2 have an affirmative answer. Then the Hilbert
function
n 7→ dimQ (HNG (Bn G; Q))
is equal to a polynomial of degree E − 1 for n ≫ 0.
Proof
Theorem 2.21 implies that the Euler characteristic of UConf n (G) is a polynomial of degree
E − 1. On the other hand, Remark 3.4 implies the lower homology groups have Hilbert
functions which grow strictly slower than this. It follows that the rank of HNG (Bn G) must
eventually grow like a polynomial of degree E − 1.
Note that Theorem 2.6 implies that HNG (Bn G) is torsion-free. We have therefore seen that
H1 (Bn G) can only have 2 torsion, while the top homology HNG (Bn G) does not have any
torsion. It is an interesting question to ask what other types of torsion can appear in the
intermediate homologies.
While an affirmative answer to Questions 1 and 2 would be a step in the right direction, the
work of Ko and Park [KP] would suggest that the action of SG on Mi,• described in Section
3.2 is perhaps non-optimal. We have already discussed that the result [KP, Theorem 3.6]
implies that the braid groups of biconnected graphs have eventually constant first homologies.
However, the action of SG defined in Section 3.2 only detects the connectivity a chosen
spanning tree of G. In particular, is it possible that one can improve on this action by
accounting for connectivity granted by the deleted edges?
Question 3. Let G be a graph which is neither a circle, nor a line segment. Does there
exist an action of SG on Mi,• which interacts in a non-trivial way with the deleted edges of
G? More specifically, can one define this action so that if c is a critical i-cell, then there is
a Z-linear relation between xe1 · c and xe2 · c whenever the two essential edges e1 and e2 are
in the same connected component of G with the edges of c removed?
The existence of such an action would actually suggest something quite strong about what
kind of invariants of G are encoded by the homology of its braid groups. If G is a graph,
recall that we define ∆iG to be the maximum number of connected components G can be
broken into by removing exactly i-vertices.
Conjecture 4.3. Let G be a graph which is neither a line segment, nor a circle. Then for
all i, and all n ≫ i, the Betti number
bi (n) := dimQ (Hi (Bn G; Q))
agrees with a polynomial of degree ≤ ∆iG − 1.
Note that Theorem 4.2 is a specific case of this conjecture. Also note, if G is biconnected,
then the above implies that b1 (n) is eventually constant. This fact is proven in the work of
Ko and Park [KP, Theorem 3.6]. The above conjecture therefore proposes a generalization
of one aspect of that work.
Finally, we think about the ramifications of this work on the spaces UConf n (G). The action
of SG on Mi,• is defined in a purely formal way. However, in the study of more classical
STABILITY PHENOMENA IN THE HOMOLOGY OF TREE BRAID GROUPS
25
configuration spaces, stability phenomena often arise from maps on the underlying spaces
UConf n (G) → UConf n+1 (G) (see, for instance, [EW-G] and [McD]). Is that also the case
here?
Question 4. If G is a tree, is the action of SG on Mi,• induced by a map of topological
spaces UConf n (G) → UConf n+1 (G)? If G is a general graph, and Questions 1 and 2 have
affirmative answers, is the action of SG on the homologies of Bn G induced in a similar way?
Of course, Question 4 is highly relevant for Questions 1 and 2. If one understands the action
of SG topologically, then it might make the existence of such an action on the homologies of
general graph braid groups more apparent.
References
[A]
A. Abrams, Configuration spaces and braid groups of graphs, Ph.D thesis,
home.wlu.edu/~abramsa/publications/thesis.ps.
[BF] K. Barnett and M. Farber, Topology of configuration space of two particles on a graph, I,
Algebr. Geom. Topol. 9(1) (2009), 593624. arXiv:0903.2180.
[CEF] T. Church, J. S. Ellenberg and B. Farb, FI-modules and stability for representations of
symmetric groups, Duke Math. J. 164, no. 9 (2015), 1833-1910.
[EW-G] J. S. Ellenberg and J. D. Wiltshire-Gordon, Algebraic structures on cohomology of configuration spaces of manifolds with flows, arXiv:1508.02430.
[Fa] D. Farley, Homology of tree braid groups, Topological and asymptotic aspects of group
theory, 101-112, Contemp. Math., 394, Amer. Math. Soc., Providence, RI, 2006.
http://www.users.miamioh.edu/farleyds/grghom.pdf.
[Fo] R. Forman, Morse theory for cell complexes, Adv. in Math. 134 (1998), pp. 90-145.
[Fo2] R. Forman, A user’s guide to discrete Morse theory, Séminaire Lotharingien de Combinatoire
48 (2002), 35 p. http://www.emis.de/journals/SLC/wpapers/s48forman.pdf.
[FS] D.
Farley
and
L.
Sabalka,
Discrete Morse theory and graph braid
groups,
Algebr.
Geom.
Topol.
5
(2005),
1075-1109
(electronic).
http://www.users.miamioh.edu/farleyds/FS1.pdf.
[FS2] D. Farley and L. Sabalka, Presentations of graph braid groups, Forum Math. 24 (2012), no.
4, 827-859. http://www.users.miamioh.edu/farleyds/FS3final.pdf.
[Gh] R. Ghrist, Configuration spaces and braid groups on graphs in robotics, Knots, braids,
and mapping class groups - papers dedicated to Joan S. Birman (New York, 1998),
AMS/IP Stud. Adv. Math., 24, Amer. Math. Soc., Providence, RI (2001), 2940.
https://www.math.upenn.edu/~ghrist/preprints/birman.pdf.
[Ga] S. R. Gal, Euler characteristic of the configuration space of a complex, Colloq. Math. 89
(2001), 61-67. http://arxiv.org/pdf/math/0202143v1.pdf.
[KKP] J. H. Kim, K. H. Ko, and H. W. Park, Graph braid groups and right-angled Artin groups,
Trans. Amer. Math. Soc. 364 (2012), 309-360. arXiv:0805.0082.
[KP] K. H. Ko, and H. W. Park, Characteristics of graph braid groups, Discrete Comput Geom
(2012) 48: 915. arXiv:1101.2648.
[L]
D.
Lütgehetmann,
Configuration
spaaces
of
graphs,
Masters
Thesis,
http://luetge.userpage.fu-berlin.de/pdfs/masters-thesis-luetgehetmann.pdf.
[McD] D. McDuff, Configuration spaces of positive and negative particles, Topology, Volume 14,
Issue 1, March 1975, Pages 91-107.
[MS] E. Miller and B. Sturmfels, Combinatorial Commutative Algebra, Graduate Texts in Mathematics # 227, Springer-Verlag New York, 2005.
26
ERIC RAMOS
[PS] P. Prue and T. Scrimshaw, Abramss stable equivalence for graph braid groups,
arXiv:0909.5511.
[SS] S. Sam, and A. Snowden, Proof of Stembridge’s conjecture on stability of Kronecker coefficients, J. Algebraic Combin. 43 (2016), no. 1, 1-10, arXiv:1501.00333.
Department of Mathematics, University of Wisconsin - Madison.
E-mail address: [email protected]
| 0math.AC
|
End-to-end Analysis and Design of a Drone Flight
Controller
Zhuoqun Cheng1 , Richard West1 , and Craig Einstein1
1
Boston University, USA
[email protected]
[email protected]
[email protected]
arXiv:1802.05802v1 [cs.SY] 15 Feb 2018
Abstract
Timing guarantees are crucial to cyber-physical applications that must bound the end-to-end
delay between sensing, processing and actuation. For example, in a flight controller for a multirotor drone, the data from a gyro or inertial sensor must be gathered and processed to determine
the attitude of the aircraft. Sensor data fusion is followed by control decisions that adjust the
flight of a drone by altering motor speeds. If the processing pipeline between sensor input and
actuation is not bounded, the drone will lose control and possibly fail to maintain flight.
Motivated by the implementation of a multithreaded drone flight controller on the Quest
RTOS, we develop a composable pipe model based on the system’s task, scheduling and communication abstractions. This pipe model is used to analyze two semantics of end-to-end time:
reaction time and freshness time. We also argue that end-to-end timing properties should be
factored in at the early stage of application design. Thus, we provide a mathematical framework to derive feasible task periods that satisfy both a given set of end-to-end timing constraints
and the schedulability requirement. We demonstrate the applicability of our design approach
by using it to port the Cleanflight flight controller firmware to Quest on the Intel Aero board.
Experiments show that Cleanflight ported to Quest is able to achieve end-to-end latencies within
the predicted time bounds derived by analysis.
1998 ACM Subject Classification C.3 Real-Time and Embedded Systems
Keywords and phrases real-time systems, end-to-end timing analysis, flight controller
1
Introduction
Over the past few years, commercial and hobbyist multirotor drones have been rapidly
growing in popularity. The fast development of drone technology enables an ever widening
set of applications, including aerial photography [35], package delivery [3], and search and
rescue [39, 27]. One of the most commonly used control boards for drones in use today is
the STM32 family of systems-on-chip (SoCs), which are based on the ARM Cortex M-series
processors, and include integrated inertial sensors such as a gyroscope, accelerometer and
magnetometer.
While many of the existing flight control boards are perfectly adequate for drones operated via human-assisted radio control, they fall short of the processing capabilities needed for
fully autonomous operation. For this reason, we are developing a new approach to building
autonomous drones using emerging multicore platforms such as the Intel Aero board, Qualcomm Snapdragon flight development board [2], and Nvidia’s Jetson [1]. All these boards
offer multiple processing cores and integrated graphics processing capabilities, making them
capable of mission tasks that would be impossible on simpler hardware.
Our first step to building autonomous drones has so far involved a reimplementation
of the popular racing drone flight control firmware called Cleanflight [8] on the Intel Aero
board. We have also ported our in-house Quest real-time OS (RTOS) [31] to the Aero
board, to efficiently and predictably manage the multiple cores and I/O complexity. Our reimplementation of Cleanflight refactors the original single-threaded code running directly as
firmware on STM32 SoCs into a multi-threaded application running on Quest. The decoupling of software components into separate threads improves the modularity of Cleanflight,
and provides the capability for parallel task execution on platforms with multiple cores.
Flight management tasks are then able to leverage the availability of increased compute resources, potentially improving the controllability of a drone. Similarly, the cleaner interfaces
between software components eases our future plans to extend Cleanflight with advanced
features such as camera data processing, object detection and avoidance, and simultaneous
localization and mapping (SLAM) necessary for autonomous flight management.
The original Cleanflight code has a series of tasks that are executed as loops with predefined frequencies. These frequencies are based on a combination of the capabilities of the
hardware and the experiences of drone developers. However, a multithreaded Cleanflight is
subject to extra overheads and uncertainty, due to scheduling and inter-task communication.
It is critical to ensure the timing correctness of Cleanflight. For example, if a gyroscope reading fails to correctly influence a change in motor (and, hence, rotor) speed within a specific
time bound, the drone might not be able to stabilize.
Cleanflight is typical of many applications that process sensor inputs and require timebounded changes to actuators. With the recent development of multi-sensor data fusion
algorithms [23, 22, 17] and ever-increasing availability of open source hardware, multisensor/actuator cyber-physical applications are leading to a revolution in areas such as
3D printing, drones, robotics, driverless cars, and intelligent home automation systems. For
these applications, it is essential to guarantee two types of end-to-end timing requirements:
1) the maximum time it takes for an input sensor reading to flow through the whole system
to eventually affect an actuator output, and 2) the maximum time within which an input
sensor reading remains influential on output actuator commands.
While the real-time community has developed valuable approaches to scheduling and
response time analysis of tasks [10], end-to-end timing analysis has received only limited
attention. Most of the prior work has originated from real-time network communication
research [34, 30, 28, 40, 33, 42], and is based on event-triggered communication with FIFObased buffers. In a drone flight control program, however, single register-based buffers and
periodic sampling are more common. This paper, therefore, presents the end-to-end timing
analysis of a drone flight controller based on a combination of the periodic-sampling task
model, the register-based inter-task communication model and the Quest RTOS’s scheduling
model. We also show how the derived worst case end-to-end communication time can be, in
turn, used to guide the design of applications.
Contributions of this paper include: 1) the proposal of a composable pipe model to
capture the timing characteristics of end-to-end communication in the Quest RTOS; 2) a
demonstration of how to derive task periods from given end-to-end timing constraints in
the application design stage; 3) the re-implemention and evaluation of the Cleanflight flight
controller on the Intel Aero board.
The rest of the paper is organized as follows: Section 2 provides background on Cleanflight and the corresponding task, scheduling and communication models adopted throughout
the paper. Section 3 describes the end-to-end timing analysis of our proposed composable
pipe model. Section 4 shows how the end-to-end time is leveraged in the application design
stage, while Section 5 details the re-implementation and evaluation of Cleanflight on the
Aero board. Related work is discussed in Section 6, followed by conclusions and future work
in Section 7.
2
Execution Model
In this section, we first provide an overview of the application that motivated this work.
Secondly, we describe the application design model from three perspectives: i.e., the task
model, the scheduling model and the communication model.
2.1
Application and System Overview
This paper is motivated by our objective to implement an autonomous flight management
system for multirotor drones. An autonomous drone is one that is able to reason about
and adapt to changes in its surroundings, while accomplishing mission objectives without
remote assistance from a human being (once its objectives are established). As part of this
effort, we have undertaken a port of the Cleanflight firmware from a traditional single-core
STM32-based system-on-chip (SoC) to the Intel Aero compute board. The Aero board has
a quadcore x86 Atom x7-Z8750 processor, 4GB RAM, an integrated GPU, inertial measurement sensors and 3D camera connectivity. This makes it capable of flight management tasks
(e.g., package delivery, aerial photography, search and rescue) that would be impossible with
a less powerful single-core ARM Cortex M3 or M4 found in most STM32 SoCs.
Cleanflight is targeted at racing drones, which are operated by humans using radio
control. The core software components of Cleanflight consist of sensor and actuator drivers,
a PID controller, the Mahony Attitude and Heading Reference (AHRS 1 ) algorithm, various
communication stacks, and a logging system. Runtime entities of those components are
called tasks. There are 31 tasks in total, of which more than half are optional. The essential
ones are listed in Table 1. Tasks are scheduled from highest to lowest dynamic priority,
calculated as a function of a task’s static priority and the time since it was last executed.
Task Name
Period (µs)
System
100000
Battery Alert
200000
Battery Voltage
20000
Battery Current
20000
Gyro
1000
PID
1000
Accelerometer
1000
IMU Attitude
10000
RC Receiver
20000
Serial
10000
Magnetometer
100000
Table 1 List of Cleanflight Tasks
Static Priority
Medium-High
Medium
Medium
Medium
RealTime
RealTime
Medium
Medium
High
Low
Low
Description
Check system utilization
Alarm battery runout
Update battery voltage reading
Update battery current reading
Update gyro readings
Perform PID-based motor control
Update accelerometer readings
Calculate attitude
Process RC commands
Serial communication
Update magnetometer readings
Our port of Cleanflight to the Aero board runs on our in-house Quest RTOS. We have
developed Quest drivers for SPI, I2C, GPIO, UART, and inertial sensors on this and other
similar boards. Quest is an SMP system, providing both user and kernel level threads, as
1
The attitude is the orientation of the drone relative to a reference frame such as earth.
well as threaded interrupt handlers that are scheduled by a time-budgeted virtual CPU
(VCPU) scheduler detailed in Section 2.3.
2.2
Task Model
We model the flight controller program as a set of real-time periodic tasks {τ 1 , τ 2 , · · · , τ n }.
Each task τ j is characterized by its worst case execution time ej and period T j . ej and
T j are determined during the design stage and are fixed at runtime. ej is usually profiled
off-line under the worst case execution condition. Deciding the value of T j is a challenging
process, which is the major topic of this paper. It mainly depends on the end-to-end latency
constraints and the schedulability test. All the periodic tasks are implemented using Quest’s
user level threads. In this paper, we use term thread and task interchangeably.
Apart from user level threads, there are kernel threads dedicated to I/O interrupts, which
originate primarily from the SPI and I2C bus in Cleanflight. Quest executes interrupts in a
deferrable thread context, having a corresponding time budget. This way, the handling of
an interrupt does not steal CPU cycles from a currently running, potentially time-critical
task.
2.3
Scheduling Model
Threads in Quest are scheduled by a two-level scheduling hierarchy, with threads mapped
to virtual CPUs (VCPUs) that are mapped to physical CPUs. Each VCPU is specified
a processor capacity reserve [24] consisting of a budget capacity, C, and period, T . The
value of C and T are determined by the e and T of the task mapped to the VCPU. A
VCPU is required to receive at least C units of execution time every T time units when it is
runnable, as long as a schedulability test [20] is passed when creating new VCPUs. This way,
Quest’s scheduling subsystem guarantees temporal isolation between threads in the runtime
environment.
Conventional periodic tasks are assigned to Main VCPUs, which are implemented as
Sporadic Servers [38] and scheduled using Rate-Monotonic Scheduling (RMS) [21]. The
VCPU with the smallest period has the highest priority. Instead of using the Sporadic Server
model for both main tasks and bottom half threads, special I/O VCPUs are created for
threaded interrupt handlers. Each I/O VCPU operates as a Priority Inheritance Bandwidth
preserving Server (PIBS) [9]. A PIBS uses a single replenishment to avoid fragmentation of
replenishment list budgets caused by short-lived interrupt service routines (ISRs). By using
PIBS for interrupt threads, the scheduling overheads from context switching and timer
reprogramming are reduced [25].
2.4
Communication Model
Control flow within the flight controller is influenced by the path of data, which originates from sensory inputs and ends with actuation. Inputs include inertial sensors, optional
cameras and GPS devices, while actuators include motors that affect rotor speeds and the
attitude of the drone. Data flow involves a pipeline of communicating tasks, leading to a
communication model characterized by: (1) the interarrival times of tasks in the pipeline,
(2) inter-task buffering, and (3) the tasks’ access pattern to communication buffers.
Periodic vs. Aperiodic Tasks. Aperiodic tasks have irregular interarrival times,
influenced by the arrival of data. Periodic tasks have fixed interarrival times and operate
on whatever data is available at the time of their execution. A periodic task implements
asynchronous communication by not blocking to await the arrival of new data [11].
Register-based vs. FIFO-based Communication. A FIFO-based shared buffer
is used in scenarios where data history is an important factor. However, in a flight controller, data freshness outweighs the preservation of the full history of all sampled data.
For example, the motor commands should always be computed from the latest sensor data
and any stale data should be discarded. Moreover, FIFO-based communication results in
loosely synchronous communication: the producer is suspended when the FIFO buffer is
full and the consumer is suspended when the buffer is empty. Register-based communication achieves fully asynchronous communication between two communicating parties using
Simpson’s four-slot algorithm [36].
Implicit vs. Explicit Communication. Explicit communication allows access to
shared data at any time point during a task’s execution. This might lead to data inconsistency in the presence of task preemption. A task that reads the same shared data at
the beginning and the end of its execution might see two different values, if it is preempted
between the two reads by another task that changes the value of the shared data. Conversely,
the implicit communication model [13] essentially follows a read-before-execute paradigm to
avoid data inconsistency. It mandates a task to make a local copy of the shared data at the
beginning of its execution and to work on that copy throughout its execution.
This paper assumes a periodic task model, as this simplifies timing analysis. Applications
such as Cleanflight implement periodic tasks to sample data and perform control operations.
Our system also adopts register-based, implicit communication for data freshness and consistency.
3
End-to-end Communication Timing Analysis
In this section, we first distinguish two different timing semantics for end-to-end communication, which will be used as the basis for separate timing analyses. Secondly, we develop
a composable pipe model for communication, which is derived from separate latencies that
influence end-to-end delay. Lastly, we use the pipe model to derive the worst case end-to-end
communication time under various situations.
3.1
Semantics of End-to-end Time
To understanding the meaning of end-to-end time consider the following two constraints for
a flight controller:
Constraint 1: a change in motor speed must be within 2 ms of the gyro sensor reading
that caused the change.
Constraint 2: an update to a gyro sensor value must be within 2 ms of the corresponding
update in motor speed.
The values before and after a change differ, whereas they may stay the same before and after
an update. These semantics lead to two different constraints. To appreciate the difference,
imagine the two cases in Table 2. In Case 1, the task that reads the gyro runs every 10
ms and the one that controls the motors runs every 1 ms. Case 1 is guaranteed to meet
Constraint 1 because the motor task runs more than once within 2 ms, no matter whether
the gyro reading changes. However, it fails Constraint 2 frequently as the gyro task is not
likely to run even once in an interval of 2 ms. Conversely, Case 2 is guaranteed to meet
Constraint 2 but fails Constraint 1 frequently.
This example demonstrates the difference between the two semantics of end-to-end time,
which leads to the following formal definitions:
Case 1
Case 2
Gyro Period
10 ms
1 ms
Motor Period
1 ms
10 ms
Table 2 Example Periods
Figure 1 Task Chain
Reaction time is the time it takes for input data to flow through the system, and is
affected by the period of each consumer in a pipeline. A reaction timing constraint
bounds the time interval between a sensor input and the first corresponding actuator
output.
Freshness time is the time within which an instance of the input data has influence on
the system, and is affected by the period of each producer in a pipeline. A freshness timing constraint bounds the time interval between a sensor input and the last corresponding
actuator output.
Constraint 1, above, is a constraint on reaction time, while Constraint 2 is on freshness time.
We perform analysis of the two semantics of time in Section 3.5.2 and 3.5.3, respectively.
3.2
Latency Contributors
The end-to-end communication delay is influenced by several factors, which we will identify
as part of our analysis. To begin, we first consider the end-to-end communication pipeline
illustrated as a task chain in Figure 1. Task τ1 reads input data Din over channel Cin ,
processes it and produces data D1 . Task τ2 reads D1 and produces D2 , and τ3 eventually
writes output Dout to channel Cout after reading and processing D2 .
Each task handles data in three stages, i.e., read, process and write. The end-to-end
time should sum the latency of each stage in the task chain. Due to the asynchrony of
communication, however, we also need to consider one less obvious latency, which is the
waiting time it takes for an intermediate output to be read in as input, by the succeeding
task in the chain. In summary, the latency contributors are classified as follows:
Processing latency represents the time it takes for a task to translate a raw input
to a processed output. The actual processing latency depends not only on the absolute
processing time of a task without interruption, but also on the service constraints (i.e.,
CPU budget and period of the VCPU) associated with the task.
Communication latency represents the time to transfer data over a channel. The transfer data size, bandwidth and propagation delay of the communication channel, and the
software overheads of the communication protocol all contribute to the overall latency.
Since our communication model is asynchronous and register-based as described in Section 2, queuing latency is not a concern of this work.
Scheduling latency represents the time interval between the arrival of data on a channel from a sending task and when the receiving task begins reading that data. The
scheduling latency depends on the order of execution of tasks in the system, and therefore has significant influence on the end-to-end communication delay.
3.3
The Composable Pipe Model
In Section 3.2, we identified the factors that influence end-to-end communication delay.
Among them, the absolute processing time and the transfer data size are determined by the
nature of the task in question. To capture the rest of the timing characteristics, we develop
a composable pipe model, leveraging the scheduling approach described in Section 2.3. A
task and pipe have a one-to-one relationship, as illustrated in Figure 2.
Figure 2 Illustration of a Pipe
3.3.1
Terminology
A pipe is composed of three elements:
Pipe Terminal. A pipe has one terminal, which is an encapsulation of the data processing power reserved for this pipe. A pipe terminal is represented by a VCPU and its
timing characteristics are captured by the VCPU’s budget and period, thereby guaranteeing at least C units of execution time every T time units. Pipe terminals are associated
with conventional tasks bound to Main VCPUs and kernel control paths (including interrupt handlers and device drivers) bound to I/O VCPUs, as described in Section 2.3.
Pipe End. A pipe has two ends, one for input and one for output. A pipe end is an
interface to a communication channel, which is either an I/O bus or shared memory.
Theoretically, the physical timing characteristics of a pipe end consist of transmission
delay and propagation delay. As this work focuses on embedded systems where communicating parties are typically located within close proximity, we neglect propagation
delay. The transmission delay is modeled by the bandwidth parameter, W , of the communication channel. We also use a parameter δ to denote the software overheads of
a communication protocol. Though we are aware that δ depends on the data transfer
size, the time difference is negligible, compared to the time of actual data transfer and
processing. Therefore, for the sake of simplicity, δ is a constant in our model.
Note that in our definition of a single pipe there is only one terminal, not two at either
end of the pipe. This differs from the idea of a POSIX pipe, which comprises at task at
both sending and receiving ends. In our case, a pipe is represented by the single terminal
that takes input and produces output.
An example of two communicating pipes is shown in Figure 3. This is representative of a
communication path between a gyro task and attitude calculation in Cleanflight on the Aero
board. The gyro task is mapped to Pipe 1, whose input end is over the SPI bus connected
to the gyro sensor and output end is over a region of memory shared with Pipe 2. Pipe 1’s
terminal is an I/O VCPU because the gyro task is responsible for handling I/O interrupts
generated from the SPI bus. On the contrary, the terminal of Pipe 2 is a Main VCPU as the
AHRS task is CPU-intensive. The gyro task reads raw gyro readings from Pipe 1’s input
end, processes them, and writes filtered gyro readings to Pipe 1’s output end. Similarly, the
AHRS task reads the filtered gyro readings from Pipe 2’s input end and produces attitude
data for its output end over shared memory.
Figure 3 Illustration of Two Communicating Pipes
3.3.2
Notation
The timing characteristics of a pipe are denoted by the 3-tuple, π = ((Wi , δi ), (C, T ), (Wo , δo )),
where:
(Wi , δi ) and (Wo , δo ) denote the bandwidth and software overheads of the input and
output ends, respectively.
(C, T ) denotes the budget and period of the pipe terminal.
A task τ is also represented as a 3-tuple, τ = (di , p, do ), where:
di denotes the size of the raw data that is read in by τ in order to perform its job, and
do denotes the size of the processed data that is produced by τ .
p denotes the uninterrupted processing time it takes for τ to turn the raw data into the
processed data.
In addition, τ 7→ π denotes the mapping between task τ and pipe π. A task τ = (di , p, do )
is said to be mapped to a pipe π = ((Wi , δi ), (C, T ), (Wo , δo )) when
data of size di is read from the input end with parameters (Wi , δi ), and data of size do
is written to the output end with parameters (Wo , δo );
the pipe terminal with parameters (C, T ) is used for scheduling and accounting of the
read and write operations, as well as the processing that takes time p.
For the composition of a chain of pipes, the operator | connects a pipe’s output end to its
succeeding pipe’s input end. For example, Figure 3 is represented as τgyro 7→ π1 |τAHRS 7→ π2 .
The scheduling latency between two pipes is denoted by Sτ 7→π|τ ′ 7→π′ . Lastly, given a task
set T = {τ1 , τ2 , · · · , τn } identity mapped to a pipe set Π = {π1 , π2 , · · · , πn }, where pipes are
connected to each other in ascending order of subscript, Eτ1 7→π1 |τ2 7→π2 |···|τn 7→πn denotes the
end-to-end reaction time of the pipe chain, and Fτ1 7→π1 |τ2 7→π2 |···|τn 7→πn denotes the end-to-end
freshness time.
3.4
Reachability
Before mathematically analyzing end-to-end time, we introduce the concept of reachability,
inspired by the data-path reachability conditions proposed by Feiertag et al [11]. The necessity of introducing reachability is due to a subtle difference between our register-based
asynchronous communication model and the traditional FIFO-based synchronous communication. In the latter, data is guaranteed to be transferred without loss or repetition. This
way, end-to-end time is derived from the time interval between the arrival of a data input
and the departure of its corresponding data output. Unfortunately, this might result in
an infinitely large end-to-end time in the case of register-based asynchronous communication where not every input leads to an output. Instead, unprocessed input data might be
discarded (overwritten) when newer input data is available, as explained in Section 2.4.
An infinitely large end-to-end time, while mathematically correct, lacks practical use.
Therefore, the following timing analysis ignores all input data that fails to “reach” the exit
of the pipe chain it enters. Instead, only those data inputs that result in data outputs from
the pipe chain are considered. We define this latter class of inputs as being reachable.
3.5
Timing Analysis
As alluded to above, the execution of a task is divided into three stages, involving (1) reading,
(2) processing, and (3) writing data. To simplify the timing analysis, we assume that tasks
are able to finish the read and write stages within one period of the pipe terminal, to which
the task is mapped. This is not unrealistic for applications such as a flight controller, because:
1) data to be transferred is usually small, and 2) all three stages are typically able to finish
within one period. However, to maintain generality, we do not impose any restriction on the
length of the processing stage.
3.5.1
Worst Case End-to-end Time of a Single Pipe
First, we consider the case where there is a single pipe. Two key observations for this case
are: 1) the absence of scheduling latency due to the lack of a succeeding pipe, and 2) the
equivalence of the two end-to-end time semantics (reaction and freshness time) due to the
lack of a preceding pipe. We therefore use Lτ 7→π to unify the notation of Eτ 7→π and Fτ 7→π .
Given task τ = (di , p, do ) mapped to pipe π = ((Wi , δi ), (C, T ), (Wo , δo )), the worst case
end-to-end time is essentially the execution time of the three stages of τ on π. Due to the
timing property of π’s pipe terminal, τ is guaranteed C units of execution time within any
window of T time units. Hence, the worst-case latency Lwc
τ 7→π is bounded by the following:
∆in + p + ∆out
· T + (∆in + p + ∆out ) mod C
(1)
Lwc
=
τ 7→π
C
where ∆in =
3.5.2
di
Wi
+ δi and ∆out =
do
Wo
+ δo .
Worst Case End-to-end Reaction Time of a Pipe Chain
In this section, we extend the timing analysis of a single pipe to a pipe chain. For the
sake of simplicity, we start with a chain of length two. We show in Section 3.6 that the
mathematical framework is applicable to arbitrarily long pipe chains. To distinguish the
tasks mapped to the two pipes, we name the preceding task producer and the succeeding
consumer. The producer is denoted by τp = (dpi , pp , d) and its pipe is denoted by πp =
((Wip , δip ), (C p , T p ), (Wop , δop )). Similarly, the consumer task and pipe are denoted by τc =
(d, pc , dco ) and πc = ((Wic , δic ), (C c , T c ), (Woc , δoc )). Following the definition of the end-toend reaction time, Eτp 7→πp |τc 7→πc , in Section 3.1, we investigate the time interval between a
specific instance of input data, denoted by Di , being read by τp , and its first corresponding
output, denoted by Do , being written by τc .
It is of vital importance to recognize that end-to-end time of a pipe chain is not simply
the sum of the end-to-end time of each single pipe in the chain. We also need to account for
the scheduling latency resulting from each appended pipe. As described in Section 3.2,
the scheduling latency depends on the order of execution of tasks. We, therefore, perform
the timing analysis under two complementary cases: Case 1 - τc has shorter period and
thus higher priority than τp ; Case 2 - τp has shorter period and thus higher priority than
τc , according to rate-monotonic ordering.
3.5.2.1
Calculating the End-to-end Reaction Time
wc
wc
Case 1. The key to making use of Lwc
τp 7→πp and Lτc 7→πc in the timing analysis of Eτp 7→πp |τc 7→πc ,
wc
is to find the worst case scheduling latency, Sτp 7→πp |τc 7→πc . As illustrated in Figure 4, the
worst case scheduling latency occurs when τc preempts τp (Step 1 ) immediately before τp
produces the intermediate output Dint corresponding to Di . After preemption, τc uses up
πc ’s budget and gives the CPU back to τp . Upon being resumed, τp immediately produces
Dint (Step 2 ). For τc to become runnable again to read Dint in Step 3 , it has to wait
for its budget replenishment. The waiting time is exactly the worst case scheduling latency:
= T c − Cc − (
Sτwc
p 7→πp |τc 7→πc
d
+ δop )
Wop
(2)
Figure 4 End-to-end Reaction Time in Case 1
After replenishment, τc reads in Dint , processes it and eventually writes out Do . As
Eτp 7→πp |τc 7→πc is defined to be the time interval between the arrival of Di and the departure
of Do , the worst case of Eτp 7→πp |τc 7→πc is as follows:
wc
wc
= Lwc
Eτwc
τp 7→πp + Sτp 7→πp |τc 7→πc + Lτc 7→πc
p 7→πp |τc 7→πc
d
+ δop )
(3)
Wop
Note that if τc runs out of budget before writing Do , τp may overwrite Dint in the pipe with
new data (Step 4 ). However, the implicit communication property guarantees that τc only
works on its local copy of the shared data, which is Dint until τc initiates another read.
c
c
wc
= Lwc
τp 7→πp + Lτc 7→πc + T − C − (
Figure 5 End-to-end Reaction Time in Case 2
Case 2. The situation is more complicated when τp has higher priority than τc . The
worst case scenario in Case 1 does not hold in Case 2 primarily because τp might overwrite
Dint before τc has a budget replenishment. This is impossible in Case 1 because τp has a
larger period than τc , which is guaranteed to have its budget replenished before τp is able to
initiate another write. In other words, in Figure 4, Step 3 is guaranteed to happen before
Step 4 .
The data-overwrite problem in Case 2 is the reason for introducing reachability in Section 3.4. To find the worst case end-to-end reaction time in this case, we have to find the
scenario that not only leads to the worst case scheduling latency, but also originates from
a reachable input. Figure 5 illustrates a scenario that meets these requirements. In the
figure, τp preempts τc immediately after τc finishes reading τp ’s intermediate output (Step
3 ), Dint , corresponding to Di . It follows that the longest possible waiting time, between
Dint becoming available (Step 2 ) and τc reading the data (Step 3 ), is the period of τp
minus both its budget and the execution time of the read stage of τc . This waiting time is
exactly the worst case scheduling latency:
= T p − Cp − (
Sτwc
p 7→πp |τc 7→πc
d
+ δic )
Wic
(4)
Between reading Dint and writing Do , τc might experience more than one preemption
from τp , which repeatedly overwrites the shared data. This will not, however, affect τc ’s
processing on Dint either spatially or temporally, thanks to the VCPU model and the implicit
communication semantic. Therefore, similar to Case 1, the worst case end-to-end reaction
time is again the sum of Equation 1 of each pipe and Equation 4:
wc
wc
= Lwc
Eτwc
τp 7→πp + Sτp 7→πp |τc 7→πc + Lτc 7→πc
p 7→πp |τc 7→πc
wc
p
p
= Lwc
τp 7→πp + Lτc 7→πc + T − C − (
d
+ δic )
Wic
(5)
Since the output end of τp and the input end of τc share the same communication channel,
it is reasonable to assume that Wop = Wic and δop = δic . With that, we proceed to unify the
worst case end-to-end reaction time using one conditional equation:
(
d
wc
if T c < T p
+ δ) + Lwc
T c − Cc − ( W
τp 7→πp + Lτc 7→πc ,
wc
Eτp 7→πp |τc 7→πc =
d
wc
+ δ) + Lwc
T p − Cp − ( W
τp 7→πp + Lτc 7→πc , otherwise
(6)
where W = Wop = Wic and δ = δop = δic
3.5.2.2
Special Cases
Real-time systems are often profiled offline to obtain worst case execution times of their
tasks. In our case, this would enable CPU resources for pipe terminals to be provisioned so
that each task completes one iteration of all three stages (read, process, write) in one budget
allocation and, hence, period. This implies that ∆in + p + ∆out + ǫ = C in Equation 1, where
ǫ is an arbitrarily small positive number to account for surplus budget after completing all
task stages. With that, it is possible to simplify the worst case end-to-end reaction time
derived in Section 3.5.2.1. First, Equation 1 is simplified as follows:
C −ǫ
⌋ · T + [(C − ǫ) mod C]
C
= 0 · T + (C − ǫ) ≈ C
Lwc
τ 7→π = ⌊
(7)
Using Equation 7, Equation 5 reduces to:
wc
= T p − C p − ∆cin + Lwc
Eτwc
τp 7→πp + Lτc 7→πc
p 7→πp |τc 7→πc
= T p − C p − ∆cin + C p + C c
= T p + C c − ∆cin
The same simplification applied to Equation 3 of Case 1 reduces Equation 6 to:
(
T c + C p − ∆, if T c < T p
wc
Eτp 7→πp |τc 7→πc =
T p + C c − ∆, otherwise
(8)
(9)
d
+ δ.
where ∆ = W
If we further assume that πp and πc communicate data of small size over shared memory,
it is possible to discard communication overheads, such that ∆ = 0. With that, Equation 9
simplifies to:
(
T c + C p , if T c < T p
wc
Eτp 7→πp |τc 7→πc =
(10)
T p + C c , otherwise
Finally, notice that by appending τc 7→ πc to τp 7→ πp , the worst case end-to-end reaction
time is increased by the following:
− Cp
= Eτwc
− Eτwc
↑ E wc = Eτwc
p 7→πp
p 7→πp |τc 7→πc
p 7→πp |τc 7→πc
(
T c,
if T c < T p
=
p
p
c
T − C + C , otherwise
3.5.3
(11)
Worst Case End-to-end Freshness Time of a Pipe Chain
Techniques similar to those in Section 3.5.2 will be used to analyze end-to-end freshness
time. To avoid repetition, we abbreviate the end-to-end freshness timing analysis by only
focusing on the special cases described in Section 3.5.2.2.
Recall that freshness time is defined to be the interval between the arrival of an input
and the departure of its last corresponding output. Therefore, we investigate the interval
between a specific instance of input data, Di , being read by τp and its last corresponding
output, Do , being written by τc .
Figure 6 End-to-end Freshness Time in Case 1
Case 1. As illustrated in Figure 6, Di is read by the first instance of τp at time 0 and the
intermediate output, Dint , is written to the shared data (Step 1 ). After that, τc produces
three outputs corresponding to Dint (Steps 2 , 3 and 4 ), or to Di indirectly. The last
output, Do , is the one preceding τp ’s write of new data, Dnew (Step 5 ). Thus, the worst
, occurs when: 1) the two consecutive writes
case end-to-end freshness time, Fτwc
p 7→πp |τc 7→πc
(Steps 1 and 5 ) from τp have the longest possible time interval between them, and 2) the
write of Do happens as late as possible. The latest time to write Do is immediately before
the second write of τp , which is preempted by higher priority τc .
From Figure 6 that the worst case end-to-end freshness time is:
= 2 · T p − ∆pout
Fτwc
p 7→πp |τc 7→πc
(12)
Again, when communicating over shared memory, Equation 12 can be further simplified to:
= 2 · Tp
Fτwc
p 7→πp |τc 7→πc
(13)
Case 2. When τp has a smaller period than τc , it is impossible for τc to read the same
intermediate output of τp twice. In Figure 6, Step 5 is guaranteed to happen before 3 .
Thus, the worst case freshness time is essentially the worst case reaction time, shown in
Equation 5.
In summary, the worst case end-to-end freshness latency of two communicating pipes is
represented in the following conditional equation:
(
2 · T p,
if T c < T p
Fτp 7→πp |τc 7→πc =
(14)
T p + C c , otherwise
3.6
Composability
The timing analysis for two pipes in Section 3.5.2 extends to pipe chains of arbitrary length.
Every time an extra task τnew (mapped to πnew ) is appended to the tail end of a chain
(τtail 7→ πtail ), the worst case end-to-end reaction time increases by the worst case end-toend time of the newly appended pipe, plus the scheduling latency between the new pipe and
the tail pipe. The actual value of the increase, depending on the relative priority of the new
pipe and the tail pipe, is shown in Equation 11. Similarly, the added end-to-end freshness
time can be derived from Equation 14.
Composability is a crucial property of our pipe model, since it significantly eases the endto-end time calculation for any given pipeline. This provides the basis for a design framework
that derives task periods from given end-to-end timing constraints. This is detailed in the
following section.
4
End-to-end Design
There are significant challenges to porting a flight control firmware such as Cleanflight to
run as a multithreaded application on a real-time operating system. One of the major
issues is how to determine the period of each thread so that the application is able to
meet its end-to-end timing constraints. A naive approach would be to start by choosing
a tentative set of periods and use the timing analysis method in Section 3 to validate the
timing correctness. Upon failure, the periods are heuristically adjusted and the validation
step is repeated until end-to-end timing guarantees are met. This approach, however, is
potentially time-consuming and labor-intensive when the number of tasks or end-to-end
constraints increase.
Inspired by Gerber et al [12], we derive task periods from end-to-end timing constraints,
by combining the timing analysis of the pipe model with linear optimization techniques. In
this section, we generalize our method for use with a broader spectrum of cyber-physical
control applications.
4.1
Problem Definition
To precisely define the problem, first consider a set of tasks T = {τ1 , τ2 , · · · , τn } and
a set of pipes Π = {π1 , π2 , · · · , πn }, where τj = (dji , pj , djo ) and πj = ((Wij , δij ), (C j ,
T j ), (Woj , δoj )). We additionally require the following information:
the mapping between T and Π. For ease of notation, we assume tasks map to the pipe
with the same subscript, hence ∀j ∈ {1, 2, · · · , n}, τj 7→ πj ;
the topology of Π (an example is shown in Figure 7);
∀j ∈ {1, 2, · · · , n}, the value of dji , pj and djo ;
∀j ∈ {1, 2, · · · , n}, the value of Wij , δij , Woj and δoj ;
the end-to-end timing constraints, namely the value of Eτi 7→πi |τj 7→πj |···|τk 7→πk and/or
Fτp 7→πp |τq 7→πq |···|τr 7→πr where i, j, k, p, q, r ∈ {1, 2, · · · , n}.
The aim is to find a feasible set of {(C j , T j )} pairs for j ∈ {1, 2, · · · , n} that:
1. meets all the specified end-to-end timing constraints,
2. passes the task schedulability test, and
3. ideally but not necessarily minimizes CPU utilization. A task should not be run faster
than it needs to be, so that resources are made available for additional system objectives.
4.2
Solving the Constraints
Our solution is carried out in a three-step process. To make it easier to understand, we use a
concrete example with actual numbers to elaborate the process. Consider the pipe topology
Eτ1 7→π1 |τ4 7→π4 ≤ 10,
Eτ2 7→π2 |τ4 7→π4 ≤ 15,
Eτ2 7→π2 |τ5 7→π5 |τ6 7→π6 ≤ 25,
Eτ3 7→π3 |τ6 7→π6 ≤ 15;
Reaction
Fτ1 7→π1 |τ4 7→π4 ≤ 20,
Fτ2 7→π2 |τ4 7→π4 ≤ 30,
Fτ2 7→π2 |τ5 7→π5 |τ6 7→π6 ≤ 50,
Fτ3 7→π3 |τ6 7→π6 ≤ 20;
Freshness
Cj
j=1 T j
P6
√
≤ n( 6 2 − 1)
∀j ∈ {1, 2, · · · , 6},
dji = djo = 3,
Wij = Woj = 20,
δij = δoj = 0.1,
pj = 0.5;
Figure 7 Application Task Graph
Schedulability
Execution Times
Table 3 Application Timing Characteristics
graph shown in Figure 7, in which there are six tasks mapped to six pipes. Tasks 1, 2 and
3 read inputs from sensors, Tasks 4 and 6 write their outputs to actuators, and Task 5 is
an intermediary responsible for complicated processing such as PID control or sensor data
fusion. The timing characteristics of the tasks and pipes are shown in Table 3. Note that
the execution times are assumed to be identical for all tasks. In practice this would not
necessarily be the case but it does not affect the generality of the approach.
In Step 1, we use the given di , do , p, Wi , δi , Wo and δo to compute the budget of each
pipe terminal. The budget is set to a value that ensures the three stages (i.e., read, process
and write) finish in one period. To compute C 1 , for example, we aggregate the times for
d1
d1
τ1 to read, process and write data. Thus C 1 = Wi1 + δi1 + p1 + Wo1 + δo1 . All budgets are
o
i
computed in a similar way. When the input to a pipe terminal comes from multiple sources
the value di is aggregated from all input channels. For example, τ4 receives a maximum of
d4i = d1 + d2 amount of data every transfer from both τ1 and τ2 . Data from a pipe terminal
is not necessarily duplicated for all pipe terminals that are consumers. For example, τ2
generates a maximum of d2o = d2 amount of data every transfer, by placing a single copy of
the output in a shared memory region accessible to both τ4 and τ5 . If the communication
channels did not involve shared memory, then data would be duplicated, so that d2o = 2d2 .
In Step 2, we derive a list of inequations involving period variables from the given end-toend timing and scheduling constraints in Table 3. For simplicity, the scheduling constraint
is shown as a rate-monotonic utilization bound on the six pipe tasks. However, for sensor
inputs and some actuator outputs, our system would map those tasks to I/O VCPUs that
have a different utilization bound, as described in our earlier work [9]. The derivation is based
on Equations 9 and 14, and the composability property of the pipe model. According to the
conditional equations, however, every two pipes with undetermined priority can lead to two
possible inequations. This exponentially increases the search space for feasible periods. In
order to prune the search space, our strategy is to always start with the case where T p > T c .
This is based on the observation that tasks tend to over-sample inputs for the sake of better
overall responsiveness. Thus, the reaction constraint Eτ2 7→π2 |τ5 7→π5 |τ6 7→π6 ≤ 25, for example,
is translated to inequation T 5 + C 2 − ∆+ T 6 ≤ 25. This is derived by combining Equations 9
and 11. It is then possible to translate all timing constraints to inequations with only periods
as variables. In addition, periods are implicitly constrained by T j > C j , ∀j ∈ {1, 2, · · · , n}.
Given all the inequations, Step 3 attempts to find the maximum value for each period
so that the total CPU utilization is minimized. We are then left with a linear programming problem. Unfortunately, there is no polynomial time solution to the integer linear
programming problem, as it is known to be NP-hard. Though linear programming solutions
are still available under certain mathematical conditions [4], this is beyond the scope of
this paper. Instead, in practice, the problem can be simplified because 1) there are usually a small number of fan-in and fan-out pipe ends for each task, meaning that a period
variable is usually involved in a small number of inequations, and 2) a sensor task period
is usually pre-determined by a hardware sampling rate limit. For example, if we assume
T 3 is known to be 5, a feasible set of periods for the example in Table 3 is easily found:
{T 1 = 10, T 2 = 15, T 3 = 10, T 4 = 10, T 5 = 15, T 6 = 5}. If we ignore the integer requirement, it is possible to find a feasible solution in polynomial time using rational numbers that
are rounded to integers. Though rounding may lead to constraint violations, it is possible
to increase the time resolution to ensure system overheads exceed those of rounding errors.
In the worst case, the designer is always able to perform an exhaustive search of all possible
constraint solutions.
5
Evaluation
This section describes both simulations and experiments on the Intel Aero board with an
Atom x7-Z8750 1.6 GHz 4-core processor and 4GB RAM.
5.1
Simulation Experiments
We developed simulations for both Linux and Quest, to predict the worst-case end-to-end
time using the equations in Section 3. Each simulation consists of three tasks, τ1 , τ2 and
τ3 , mapped to pipes, π1 , π2 and π3 , respectively. All three tasks search for prime numbers
within a certain range and then communicate with one another to exchange their results. τ1
communicates with τ2 , which further communicates with τ3 . The communication channel
is shared memory with caches disabled and the data size is set to 6.7 KB to achieve a nonnegligible 1 millisecond communication overhead. Each task is assigned a different search
range and the profiled execution time is shown in Table 4 in milliseconds. The budget of
each pipe is set to be slightly larger than the execution time of its corresponding task, to
compensate for system overheads. The settings of each pipe terminal (PT) are also shown
in Table 4, again in milliseconds. Apart from the three main tasks, the system is loaded
with low priority background tasks that consume all the remaining CPU resources and serve
as potential interferences.
Case 1
Case 2
PT 1
(10,50)
(5,100)
Table 4 Simulation Settings
PT 2
(10,150)
(10,50)
PT 3
(5,100)
(10,150)
τ1
9.5
4.5
τ2
9.5
9.5
τ3
4.5
9.5
For each case, we measure the end-to-end reaction time and freshness time separately
and compare them to corresponding theoretical bounds. Figure 8 shows the results after
100,000 outputs are produced by τ3 . C1(E) and C1(F ) are the end-to-end reaction and
freshness times, respectively, for Case 1 in Table 4, while C2(E) and C2(F ) are for Case 2.
As can be seen, the observed values are always within the prediction bounds. As reference,
we also perform the two cases on Yocto Linux shipped with the Aero board. The kernel is
version 4.4.76 and patched with the PREEMPT_RT patch. While running the simulation,
the system also uncompresses Linux source code in the background. This places the same
load on the system as the background tasks in Quest. Figure 9 shows all (not only worst case)
end-to-end reaction and freshness times within the first 100 outputs. Compared to Linux,
there is less variance shown by the end-to-end times on Quest. Additionally, the freshness
and reaction times are generally lower with Quest than Linux. Figure 11 summarizes the
worst case reaction time (WCR), maximum variance of WCR (MaxRV), worst case freshness
time (WCF) and maximum variance of WCF (MaxFV) for both Quest and Linux.
400
350
300
Quest Reaction
Linux Reaction
Quest Freshness
Linux Freshness
Observed
Predicted
250
300
Time (ms)
Time (ms)
250
200
200
150
150
100
100
50
50
0
C1(E)
C2(E)
C1(F)
C2(F)
Figure 8 Observed vs. Predicted
5.2
0
20
40
60
80
100
Output #
Figure 9 Quest vs. Linux
The Cleanflight Experiment
Our next experiments apply the end-to-end design approach to determining the periods
of each task in the re-implementation of Cleanflight. We decouple software components
in the original Cleanflight firmware and re-build the flight controller as a multithreaded
application running on the Intel Aero board. The hardware and software architecture is
shown in Figure 10.
Hardware. We currently only use Core 0 to run Cleanflight on Quest. The remaining
three cores are reserved for our Quest-V separation kernel [41] to run a general purpose OS
such as Linux. Apart from the main processor, the Aero board also has an FPGA-based
I/O coprocessor. It provides FPGA-emulated I/O interfaces including analog-to-digital conversion (ADC), UART serial, and pulse-width modulation (PWM). Our system currently
uses the I/O hub to send PWM signals to electronic speed controllers (ESCs) that alter
motor and, hence, rotor speeds of the drone. We modified the FPGA logic to improve the
timing resolution of PWM signals, as well as control their duty cycle and periods. We make
additional use of an onboard Bosch BMI160 Inertial Measurement Unit (IMU). Both the
I/O hub and the IMU are connected to the main processor via SPI bus.
Software. To minimize the engineering efforts, we currently disable the auxiliary features of Cleanflight such as telemetry, blackbox data logging, and UART-based flight control configuration. The essential components are shown as circular tasks in Figure 10. The
Figure 10 Cleanflight Data Flow
AHRS sensor fusion task takes the input readings of the accelerometer and gyroscope (in
the BMI160 IMU) and calculates the current attitude of the drone. Then, the PID task
compares the calculated and target attitudes, and feeds the difference to the PID control
logic. In the original Cleanflight code, the target attitude is determined by radio-control
signals from a human flying the drone. In an autonomous setting, the target attitude would
be calculated according to on-board computations based on mission objectives and flight
conditions. The output is nonetheless mixed with the desired throttle and read in by the
PWM task, which translates it to motor commands. Motor commands are sent over the
SPI bus and ultimately delivered as PWM signals to each ESC associated with a separate
motor-rotor pair on the multirotor drone. We decouple all these tasks into separate threads.
For safety reasons, the sensor tasks and the PWM task are given individual address spaces.
To simplify the experiment setup, we instrument Cleanflight to use a synthetic radio input
value (20% throttle and 0 degree pitch, roll, and yaw angle) instead of reading from the real
radio driver. The tasks are profiled and execution times are shown in Table 5.
Exec Times (µs)
Gyro
174
AHRS
10
PID
2
PWM
970
Accl
167
Radio
12
Table 5 Task Execution Times
As can be seen in Figure 10, there are currently three data paths, originating from the
gyro, the accelerometer and the radio receiver, respectively. Unfortunately, there is little
information available on what end-to-end timing constraints should be imposed on each
path to guarantee a working drone. Most timing parameters in the original Cleanflight are
determined by trial and error. On the other hand, instead of determining the optimum
timing constraints, the focus of this paper is on guaranteeing given constraints. Therefore,
we first port Cleanflight to Yocto Linux on the Aero board, as a reference implementation.
The Linux version remains single-threaded and used to estimate the desired end-to-end time.
For example, for the gyro path, the worst case reaction and freshness times are measured
to be 9769 and 22972 µs, respectively. We round them to 10 and 23 ms, and use them
as end-to-end timing constraints for the Quest flight controller implementation. Using the
same approach, we determine end-to-end reaction and freshness times for the accelerometer
path, which are set to 10 and 23 ms, respectively. Finally, for the radio path, we set the
end-to-end reaction and freshness times to be 20 and 44 ms, respectively.
Using the execution times in Table 5 and timing constraints above, we apply the end-toend design approach to derive the periods. The results for each task are shown in Table 6.
Budget/Period (µs)
Gyro
200/1000
AHRS
100/5000
PID
100/2000
PWM
1000/5000
Accl
200/1000
Radio
100/10000
Table 6 Task Periods
Evaluation. To measure the actual end-to-end time, we focus on the longest pipe chain
highlighted in Figure 10. We instrument the Cleanflight code to append every gyro reading
with an incrementing ID, and also record a timestamp before the gyro input is read. The
timestamp is then stored in an array indexed by the ID. Every task is further instrumented
to maintain the ID when translating input data to output. This way, the ID is preserved
along the pipe chain, from the input gyro reading to the output motor command. After the
PWM task sends out motor commands, it looks up the timestamp using its ID and compares
it to the current time. By doing this, we are able to log both the reaction and freshness
end-to-end time for every input gyro reading. We then compare the observed end-to-end
time with the given timing constraints, as well as the predicted worst case value. Results
are shown in Figure 12. As can be seen, the observed values are always within the predicted
bounds, and always meet the timing constraints.
300
24000
Quest
Linux
250
21000
18000
Time (us)
200
Time (ms)
Observed
Predicted
Constraint
150
100
15000
12000
9000
6000
50
3000
0
WCR
MaxRV
WCF
MaxFV
Figure 11 Quest vs. Linux Worst-Case Times
6
Reaction
Freshness
Figure 12 Cleanflight Times
Related Work
Feiertag et al. [11] distinguish four semantics of end-to-end time and provides a generic
framework to determine all the valid data paths for each semantic. The authors do not
perform timing analysis as no scheduling model is assumed. Hamann et al. also discuss
end-to-end reaction and age time [13]. Their work focuses on integrating three different
communication models, including the implicit communication model, into existing timing
analysis tools such as SymTA/S [14]. While our composable pipe model is also based
on implicit communication, we perform timing analysis using the Quest RTOS’s task and
scheduling model. A large portion of end-to-end reaction time analysis is based on the
synchronous data-flow graph (SDFG) [19] where inter-task communication is driven by the
arrival of input data. In recent work [37], Singh et al. enhance the standard SDFG to allow
the specification of an end-to-end latency constraint.
Gerber et al. [12] propose a synthesis approach that determines tasks’ periods, offsets
and deadlines from end-to-end timing constraints. Their work relies on task precedence
constraints as there is no scheduling model used for the analysis. Our work uses a scheduling
model based on Quest to perform end-to-end timing analysis. We then derive task periods
and budgets to ensure specific reaction, freshness and schedulability constraints.
There are also efforts to develop programming languages, such as Prelude [29] and
Giotto [15], which are able to derive tasks’ periods based on user-specified timing constraints. Kirsch et al. [18] use Giotto to reimplement a helicopter control system. Others
have developed data-triggered rather than time-triggered drone flight control using reactive
programming languages [5].
In general operating systems, Scout [26] exposes paths that are similar to pipe chains
in our model, to offer Quality of Service guarantees to applications. Paths in Scout are
non-preemptive schedulable entities ordered according to an EDF policy.
Lastly, programming environments such as ROS [32], OROCOS [6] and LCM [16], already
widely adopted in the design of robotics and autonomous cars, are changing the development
of cyber-physical applications. Publisher-subscriber paradigms used by ROS, for example,
have influenced our thinking in the design cyber-physical systems with modularity and
robustness amongst software components. We aim to augment Cleanflight functionality
with ROS-style services as part of our ongoing efforts to build autonomous drones.
7
Conclusions and Future Work
In this paper, we identify two semantics of end-to-end time, namely reaction and freshness
time. We analyze them in the context of the Quest RTOS, for a port of the well-known
Cleanflight flight control firmware implemented as a collection of user-level tasks. This
paper describes a composable pipe model that is built on task, scheduling and communication abstractions. Using the pipe model, we derive the worst case end-to-end time for
data flow through a chain of tasks under various conditions. We argue that end-to-end
timing properties should be factored in at the early stage of application design. Thus, we
provide a mathematical framework to derive feasible task periods and budgets using the
Quest scheduling framework to satisfy a given set of end-to-end timing and schedulability
constraints. We demonstrate the applicability of our design approach by using it to port the
Cleanflight flight controller to Quest on the Intel Aero board.
Future work will integrate the end-to-end design approach into our Quest-Arduino (Qduino [7]) development environment. This will provide the basis for the design and implementation of an autonomous multicore flight management system on our Quest-V separation
kernel [41]. Quest-V enables non-time-critical tasks to run on a legacy Linux system in
parallel with real-time tasks running on our Quest RTOS.
References
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Nvidia Jetson Board, March 2017.
http://www.nvidia.com/object/embeddedsystems.html.
Qualcomm Snapdragon Flight Kit, March 2017. https://www.intrinsyc.com/verticaldevelopment-platforms/qualcomm-snapdragon-flight/.
Amazon Prime Air. http://www.amazon.com/b?ie=UTF8&node=8037720011.
S.P. Bradley, A.C. Hax, and T.L. Magnanti. Applied Mathematical Programming. AddisonWesley Publishing Company, 1977.
Endri Bregu, Nicola Casamassima, Daniel Cantoni, Luca Mottola, and Kamin Whitehouse.
Reactive Control of Autonomous Drones. In Proceedings of the 14th Annual International
Conference on Mobile Systems, Applications, and Services, MobiSys ’16, pages 207–219,
New York, NY, USA, 2016. ACM.
H. Bruyninckx. Open Robot Control Software: the OROCOS Project. In Proceedings 2001
ICRA. IEEE International Conference on Robotics and Automation (Cat. No.01CH37164),
volume 3, pages 2523–2528 vol.3, 2001.
Zhuoqun Cheng, Ye Li, and Richard West. Qduino: A Multithreaded Arduino System for
Embedded Computing. In Proceedings of the 2015 IEEE Real-Time Systems Symposium
(RTSS), RTSS ’15, pages 261–272, Washington, DC, USA, 2015. IEEE Computer Society.
Cleanflight: http://cleanflight.com/.
Matthew Danish, Ye Li, and Richard West. Virtual-CPU Scheduling in the Quest Operating
System. In Proceedings of the 17th Real-Time and Embedded Technology and Applications
Symposium, 2011.
Robert I. Davis, Sebastian Altmeyer, Leandro Indrusiak, Claire Maiza, Vincent Nelis, and
Jan Reineke. An Extensible Framework for Multicore Response Time Analysis. Real-Time
Systems, 2017.
Nico Feiertag, Kai Richter, Johan Nordlander, and Jan Jonsson. A Compositional Framework for End-to-End Path Delay Calculation of Automotive Systems under Different Path
Semantics. In Proceedings of the IEEE Real-Time System Symposium - Workshop on
Compositional Theory and Technology for Real-Time Embedded Systems, Barcelona, Spain,
November 30, 2008.
Richard Gerber, Seongsoo Hong, and Manas Saksena. Guaranteeing Real-Time Requirements With Resource-Based Calibration of Periodic Processes. IEEE Trans. Softw. Eng.,
July 1995.
Arne Hamann, Dakshina Dasari, Simon Kramer, Michael Pressler, and Falk Wurst. Communication Centric Design in Complex Automotive Embedded Systems. In 29th Euromicro
Conference on Real-Time Systems (ECRTS 2017), Leibniz International Proceedings in Informatics (LIPIcs), Dagstuhl, Germany, 2017.
Rafik Henia, Arne Hamann, Marek Jersak, Razvan Racu, Kai Richter, and Rolf Ernst. System Level Performance Analysis - the SymTA/S Approach. In IEEE Proceedings Computers
and Digital Techniques, 2005.
Thomas A. Henzinger, Benjamin Horowitz, and Christoph Meyer Kirsch. Giotto: A TimeTriggered Language for Embedded Programming, pages 166–184. Springer Berlin Heidelberg,
Berlin, Heidelberg, 2001.
Albert S. Huang, Edwin Olson, and David Moore. LCM: Lightweight Communications
and Marshalling. In Proceedings of the IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS), Taipei, Oct 2010.
Jonathan Kelly and Gaurav S Sukhatme. Visual-Inertial Sensor Fusion: Localization, Mapping and Sensor-to-Sensor Self-calibration. The International Journal of Robotics Research,
30(1):56–79, 2011.
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
Christoph M. Kirsch, Marco A. A. Sanvido, Thomas A. Henzinger, and Wolfgang Pree. A
Giotto-Based Helicopter Control System, pages 46–60. Springer Berlin Heidelberg, Berlin,
Heidelberg, 2002.
E. A. Lee and D. G. Messerschmitt. Synchronous data flow. Proceedings of the IEEE,
75(9):1235–1245, Sept 1987.
John Lehoczky, Lui Sha, and Ye Ding. The Rate Monotonic Scheduling Algorithm: Exact Characterization and Average Case Behavior. In Proceedings of the IEEE Real-Time
Systems Symposium (RTSS), 1989.
C. L. Liu and James W. Layland. Scheduling Algorithms for Multiprogramming in a Hard
Real-Time Environment. Journal of the ACM, 20(1):46–61, 1973.
Sebastian O.H. Madgwick. An Efficient Orientation Filter for Inertial and Inertial/Magnetic
Sensor Arrays. Technical report, University of Bristol, 2010.
Robert Mahony, Tarek Hamel, and Jean-Michel Pflimlin. Nonlinear Complementary Filters
on the Special Orthogonal Group. IEEE Transactions on Automatic Control, 53, June 2008.
Clifford W. Mercer, Stefan Savage, and Hideyuki Tokuda. Processor Capacity Reserves for
Multimedia Operating Systems. Technical report, Pittsburgh, PA, USA, 1993.
E. Missimer, K. Missimer, and R. West. Mixed-Criticality Scheduling with I/O. In 28th
Euromicro Conference on Real-Time Systems (ECRTS), pages 120–130, July 2016.
David Mosberger and Larry L. Peterson. Making Paths Explicit in the Scout Operating
System. In Proceedings of the Second USENIX Symposium on Operating Systems Design
and Implementation, OSDI ’96, pages 153–167, New York, NY, USA, 1996. ACM.
BBC News. Disaster Drones: How Robot Teams can Help in a Crisis. goo.gl/6efliV.
R. S. Oliver and G. Fohler. Probabilistic Estimation of End-to-end Path Latency in Wireless
Sensor Networks. In 2009 IEEE 6th International Conference on Mobile Adhoc and Sensor
Systems, pages 423–431, Oct 2009.
Claire Pagetti, Julien Forget, Frédéric Boniol, Mikel Cordovilla, and David Lesens. Multitask Implementation of Multi-periodic Synchronous Programs. Discrete Event Dynamic
Systems, Sep 2011.
Bo Peng, Nathan Fisher, and Thidapat Chantem. MILP-based Deadline Assignment for
End-to-End Flows in Distributed Real-Time Systems. In Proceedings of the 24th International Conference on Real-Time Networks and Systems, RTNS ’16, pages 13–22, New York,
NY, USA, 2016. ACM.
Quest RTOS: http://questos.org/.
Morgan Quigley, Ken Conley, Brian P. Gerkey, Josh Faust, Tully Foote, Jeremy Leibs, Rob
Wheeler, and Andrew Y. Ng. ROS: An Open-source Robot Operating System. In ICRA
Workshop on Open Source Software, 2009.
S. Quinton, T. T. Bone, J. Hennig, M. Neukirchner, M. Negrean, and R. Ernst. Typical
Worst Case Response-time Analysis and its use in Automotive Network Design. In 2014
51st ACM/EDAC/IEEE Design Automation Conference (DAC), pages 1–6, June 2014.
J. L. Scharbarg, F. Ridouard, and C. Fraboul. A Probabilistic Analysis of End-To-End
Delays on an AFDX Avionic Network. IEEE Transactions on Industrial Informatics,
5(1):38–49, Feb 2009.
Da-Jiang Innovations Science and Technology Co. DJI. http://dji.com/.
H.R. Simpson. Four-slot Fully Asynchronous Communication Mechanism. IEEE Computers
and Digital Techniques, 137:17–30, January 1990.
Abhishek Singh, Pontus Ekberg, and Sanjoy Baruah. Applying Real-Time Scheduling Theory to the Synchronous Data Flow Model of Computation. In 29th Euromicro Conference
on Real-Time Systems (ECRTS 2017), Leibniz International Proceedings in Informatics
(LIPIcs), Dagstuhl, Germany, 2017.
38
39
40
41
42
B. Sprunt. Scheduling Sporadic and Aperiodic Events in a Hard Real-Time System. Technical Report CMU/SEI-89-TR-011, Software Engineering Institute, Carnegie Mellon, 1989.
K. P. Valavanis. Advances in Unmanned Aerial Vehicles. Springer Science and Business
Media, 2008.
Y. Wang, M. C. Vuran, and S. Goddard. Cross-Layer Analysis of the End-to-End Delay Distribution in Wireless Sensor Networks. In 2009 30th IEEE Real-Time Systems Symposium,
pages 138–147, Dec 2009.
Richard West, Ye Li, Eric Missimer, and Matthew Danish. A Virtualized Separation Kernel
for Mixed-Criticality Systems. ACM Trans. Comput. Syst., 34(3):8:1–8:41, June 2016.
P. M. Yomsi, D. Bertrand, N. Navet, and R. I. Davis. Controller Area Network (CAN):
Response Time Analysis with Offsets. In 2012 9th IEEE International Workshop on Factory
Communication Systems, pages 43–52, May 2012.
| 3cs.SY
|
arXiv:1211.0638v3 [stat.ME] 1 Aug 2016
Why the Decision-Theoretic Perspective
Misrepresents Frequentist Inference: ‘Nuts and
Bolts’ vs. Learning from Data
Aris Spanos
Department of Economics,
Virginia Tech, USA
August 2016
Abstract
The primary objective of this paper is to revisit and make a case for
the merits of R.A. Fisher’s objections to the decision-theoretic framing of
frequentist inference. It is argued that this framing is congruent with the
Bayesian but incongruent with the frequentist inference. It provides the
Bayesian approach with a theory of optimal inference, but it misrepresents
the theory of optimal frequentist inference by framing inferences solely in
terms of the universal quantifier ‘for all values of θ in Θ’, denoted by
∀θ∈Θ. This framing is at odds with the primary objective of modelbased frequentist inference, which is to learn from data about the true
value θ∗ ; the one that gave rise to the particular data. The frequentist
approach relies on factual (estimation, prediction), as well as hypothetical
(testing) reasoning, both of which revolve around the existential quantifier
∃θ∗ ∈Θ. The paper calls into question the appropriateness of admissibility
and reassesses Stein’s paradox as it relates to the capacity of frequentist
estimators to pinpoint θ∗ . The paper also compares and contrasts lossbased errors with traditional frequentist errors, such as coverage, type
I and II; the former are attached to θ, but the latter to the inference
procedure itself.
Key words: decision theoretic framing; Bayesian vs. Frequentist
inference; Stein’s paradox; James-Stein estimator; loss functions; admissibility; error probabilities; risk functions
1
1
Introduction
Wald’s (1950) decision-theoretic framework is widely viewed as providing a
broad enough perspective to accommodate and compare the frequentist and
Bayesian approaches to inference, despite their well-known differences. It is
perceived as offering a neutral framing of inference that brings into focus their
common features and tones down their differences; see Berger (1985), Robert
(2001), O’Hagan (1994).
Historically, Wald (1939) proposed the original variant of the decision-theoretic
framework with a view to unify Neyman’s (1937) rendering of frequentist interval estimation and testing:
“The problem in this formulation is very general. It contains the problems of
testing hypotheses and of statistical estimation treated in the literature.” (p. 340)
Among the frequentist pioneers, Jerzy Neyman accepted enthusiastically this
broader perspective, primarily because the concepts of decision rules and action
spaces seemed to provide a better framing for his behavioristic interpretation
of Neyman-Pearson (N-P) testing based on the accept/reject rules; see Neyman (1952; 1971). Neyman’s attitude towards Wald’s (1950) framing was also
adopted wholeheartedly by some of his most influential students/colleagues at
Berkeley, including Lehmann (1959) and LeCam (1986). In a foreword of a
collection of Neyman’s early papers, his students/editors described the Wald’s
framing as (Neyman, 1967, p. vii):
“A natural but far reaching extension of their [N-P formulation] scope can be
found in Abraham Wald’s theory of statistical decision functions.”
At the other end of the argument, R. A. Fisher (1955) rejected Wald’s framing on the grounds that it seriously distorts his rendering of frequentist statistics:
“The attempt to reinterpret the common tests of significance used in scientific research as though they constituted some kind of acceptance procedure and led to
‘decisions’ in Wald’s sense, originated in several misapprehensions and has led, apparently, to several more.” (p. 69)
With a few exceptions, such as Cox (1958), Tukey (1960) and Birnbaum
(1977), Fisher’s (1955) viewpoint has been inadequately discussed and evaluated by the subsequent statistics literature. The primary aim of this paper is to
revisit Fisher’s minority view by taking a closer look at the decision-theoretic
framework with a view to re-evaluate the claim that it provides a neutral framework for comparing the frequentist and Bayesian approaches. It is argued that
Fisher’s view that the decision theoretic framing is germane to ”acceptance sampling”, but misrepresents frequentist inference, is not without merit. The key
argument of the discussion that follows is that the decision-theoretic notions of
loss function and admissibility are congruent with the Bayesian approach, but
incongruent with both the primary objective and the underlying reasoning of
the frequentist approach.
Section 2 introduces the basic elements of the decision theoretic set-up with
a view to bring out its links to the Bayesian and frequentist approaches, calling
into question the conventional wisdom concerning its neutrality. Section 3 takes
a closer look at the Bayesian approach and argues that had the decision-theoretic
2
apparatus not exist, Bayesians would have been forced to invent it in order to
establish a theory of optimal Bayesian inference. Section 4 discusses critically
the notions of loss functions and admissibility, focusing primarily on their role
in giving rise to Stein’s paradox and their incompatibility with the frequentist
approach. It is argued that the frequentist dimension of the notions of a loss
function and admissibility is more apparent than real. Section 5 makes a case
that the decision-theoretic framework misrepresents both the primary objective
and the underlying reasoning of the frequentist approach. Section 6 revisits
the notion of a loss function and its dependence on ‘information other than the
data’. It is argued that loss-based errors are both different and incompatible
with the traditional frequentist errors because they are attached to the unknown
parameters instead of the inference procedures themselves.
2
2.1
The decision theoretic set-up
Basic elements of the decision-theoretic framing
The current decision-theoretic set-up has three basic elements.
1. A prespecified (parametric) statistical model Mθ (x), generically specified
by:
Mθ (x)={f (x; θ), θ∈Θ}, x∈RnX , for θ∈Θ⊂Rm , m≪n,
(1)
where f (x; θ) denotes the (joint) distribution of the sample X:=(X1 , ..., Xn ), RnX
denotes the sample space and Θ the parameter space. This model represents the
stochastic mechanism assumed to have given rise to data x0 :=(x1 , ..., xn ).
2. A decision space D containing all mappings d(.): RnX → A, where A
denotes the set of all actions available to the statistician.
3. A loss function L(., .): [D × Θ] → R, representing the numerical loss if
the statistician takes action a∈A when the state of nature is θ∈Θ; see Ferguson
(1967), Berger (1985), Wasserman (2004).
The basic idea is that, when the decision-maker selects action a, he/she does
not know the ‘true’ state of nature, represented by θ ∗ . However, contingent on
each action a∈A, the decision-maker ‘knows’ the losses (gains, utilities) resulting
from different choices (d, θ) ∈[D × Θ]. The decision-maker observes data x0 ,
which provides some information about θ∗ , and then maps each x∈RnX to a
certain action a∈A guided solely by L(d, θ).
2.2
The original Wald framing
It is important to bring out the fact that the original Wald (1939) framing
was much narrower than the above (1-3) rendering, due to its aim to formalize
the Neyman-Pearson (N-P) approach; see Ferguson (1976). What were the key
differences?
(i) The decision (action) space D was defined exclusively in terms of subsets
of the parameter space Θ. For estimation D:={θ : θ∈Θ} is the set of all singleton points of Θ and for testing D:= (Θ0 , Θ1 ), the null and alternative regions,
respectively.
3
(ii) The original loss (weight) was a zero-positive function, with zero loss at
(
θ=θ∗ :
0 if b
θ(X) = θ∗
L0−c (θ, b
θ(X))=
(2)
cθ >0 if b
θ(X)=θ6=θ ∗ , θ∈Θ,
where θ∗ is the true value of θ in Θ. For the discussion that follows it is important
to note that (2) is non-operational in practice because θ ∗ is unknown.
The more general framing, introduced by Wald (1947; 1950) and broadened
by Le Cam (1955), extended the scope of the original set-up by generalizing the
notions of loss functions and decision spaces. In what follows it is argued that
these extensions created serious incompatibilities with both the objective and
the underlying reasoning of frequentist inference.
In addition, it is both of historical and methodological interest to note that
Wald (1939) introduced the notion of a prior distribution π(θ), ∀θ∈Θ, into
the original decision-theoretic machinery reluctantly, and justified it on being a
useful tool for proving certain theorems:
“The situation regarding the introduction of an a priori probability distribution
of θ is entirely different. First, the objection can be made against it, as Neyman
has pointed out, that θ is merely an unknown constant and not a variate, hence it
makes no sense to speak of the probability distribution of θ. Second, even if we
may assume that θ is a variate, we have in general no possibility of determining the
distribution of θ and any assumptions regarding this distribution are of hypothetical
character. The reason why we introduce here a hypothetical probability distribution
of θ is simply that it proves to be useful in deducing certain theorems and in the
calculation of the best system of regions of acceptance.” (p. 302)
He was also the first to highlight the extreme relativism of the decisiontheoretic notion of ‘optimality’ with respect a particular loss function:
“The ”best” system of regions of acceptance ... will depend only on the weight
function of the errors.” (p. 302) [emphasis added]
2.3
A shared neutral framework?
The frequentist, Bayesian and the decision-theoretic approaches share the notion
of a statistical model by viewing data x0 :=(x1 , ..., xn ) as a realization of a sample
X:=(X1 , ..., Xn ) from (1).
The key differences between the three approaches are:
(a) the frequentist approach relies exclusively on Mθ (x),
(b) the Bayesian approach adds a prior distribution, π(θ), ∀θ∈Θ (for all
θ∈Θ),
(c) the decision-theoretic framing revolves around a loss (gain or utility)
function:
L(d(x), θ), ∀θ∈Θ, ∀x∈RnX .
The loss function is often assumed to be an even, differentiable and convex
function of (d(x)−θ), and can take numerous functional forms; see Wasserman
(2004), Robert (2001), Bansal (2007) inter alia.
The claim that the decision-theoretic perspective provides a neutral ground
is often justified on account of the loss function being a function of the sample
and parameter spaces via the two universal quantifiers:
4
(i) ‘∀x∈RnX ’, associated with the distribution of the sample:
f (x; θ), ∀x∈RnX ,
frequentist:
(ii) ‘∀θ∈Θ’ associated with the posterior distribution:
Bayesian:
π(θ)·f (x0 |θ)
,
π(θ)·f (x0 |θ)dθ
π(θ|x0 )= R
θ∈Θ
∀θ∈Θ.
The idea is that allowing for all values of x in RnX goes beyond the Bayesian
perspective, which relies exclusively on a single point x0 . What is not obvious
is whether that is sufficient to do justice to the frequentist approach. A closer
scrutiny suggests that frequentist inference is misrepresented by the way both
quantifiers are used in the decision-theoretic framing of inference.
First, the quantifier ∀x∈RnX plays a minor role since its only relevance is in
defining the expectation for transforming a loss function, say L(θ, b
θ(x)), into a
risk function:
h
i R
θ(x))f (x; θ)dx, ∀θ∈Θ.
(3)
θ(X)) = x∈Rn L(θ, b
R(θ, b
θ)=EX L(θ, b
X
This is the only place where the underlying statistical model Mθ (x) enters the
decision-theoretic framing. Hence, from this perspective the only relevant part
of the behavior of b
θ(X) is how it affects the risk for different values of θ in
Θ. This undermines the pivotal role of the quantifier ∀x∈RnX in determining
the theory of optimal frequentist inference. For that the distribution of the
sample, f (x; θ), ∀x∈RnX , takes center stage since the sampling distribution of
any statistic Yn =g(X) (estimator, test, predictor) is derived via:
Z Z
Z
f (x; θ)dx.
F (y; θ):=P(Yn ≤y; θ)=
···
(4)
| {z }
{x: g(x)≤t; x∈Rn
X}
The relevant error probabilities that calibrate the optimality of frequentist procedures are defined in terms of (4).
Second, the decision-theoretic notion of optimality revolves around the universal quantifier ‘∀θ∈Θ’, rendering it congruent with the Bayesian but incongruent with the frequentist approach. To be more specific, since different risk
functions often intersect over Θ, an optimal rule is usually selected after the risk
function is reduced to a scalar. Two such choices of relevant risk are:
b
b
Maximum risk: Rmax (θ)=sup
R(θ, θ),
θ∈Θ
Bayes risk:
b
RB (θ)=
(5)
R
b
θ∈Θ R(θ, θ)π(θ)dθ.
Hence, an obvious way to choose among different rules is to find the one that
minimizes the relevant risk with respect to all possible estimates e
θ(x). In the
case of (5), this gives rise to two corresponding decision rules:
Minimax rule:
Bayes rule:
inf Rmax (b
θ)= inf [supR(θ, b
θ)],
e
θ(x)
e
θ(x) θ∈Θ
inf RB (b
θ)= inf
R
θ∈Θ
e
θ(x)
e
θ(x)
5
R(θ, b
θ)π(θ)dθ.
In this sense, a decision or a Bayes rule b
θ will be considered optimal when it
minimizes the relevant risk, no matter what the true state of nature θ ∗ happens
to be. This constitutes a key caveat that is often ignored in discussions of these
approaches. When viewed as a game against Nature, the decision maker selects
action a from A, irrespective of what value θ∗ Nature has chosen. That is, θ∗
plays no role in selecting the optimal rules since the latter have nothing to do
with the true value θ∗ of θ. To avoid any misreading of this line of reasoning,
it is important to emphasize that ‘the true value θ∗ ’ is shorthand for saying
that ‘data x0 constitute a typical realization of the sample X with distribution
f (x; θ∗ )’; see Spanos and Mayo (2015). This should be contrasted with the
notion of optimality in frequentist inference that gives θ∗ center stage, in the
sense that it evaluates the capacity of the inference procedure to inform the
modeler about θ ∗ ; no other value is relevant. According to Reid (2015):
“A statistical model is a family of probability distributions [Mθ (x)], the central
problem of statistical inference being to identify which member of the family [θ∗ ]
generated the data of interest.” (p. 418)
3
The Bayesian approach
To shed further light on the affinity between the decision-theoretic framework
and the Bayesian approach, let us take a closer look at the latter.
3.1
Bayesian inference and its primary objective
A key argument in favor of the Bayesian approach is often its simplicity in the
sense that all forms of inference revolve around a single function, the posterior distribution: π(θ|x0 )∝π(θ)·f (x0 |θ), ∀θ∈Θ. This, however, is only half the
story. The other half is how the posterior distribution is utilized to yield ‘optimal’ inferences. The issue of optimality, however, is intrinsically related to what
the primary objective of Bayesian inference is.
An outsider looking at Bayesian approach would surmise that its primary objective is to yield ‘the probabilistic ranking’ (ordering) of all values of θ in Θ. The
modeling begins with an a priori probabilistic ranking based on π(θ), ∀θ∈Θ,
which is revised after observing x0 to derive π(θ|x0 ), ∀θ∈Θ; hence the key
role of the quantifier ∀θ∈Θ. Indeed, O’Hagan’s (1994) argues that the revised
probabilistic ranking is the inference:
“Having obtained the posterior density π(θ|x0 ), the final step of the Bayesian
method is to derive from it suitable inference statements. The most usual inference
question is this: After seeing the data x0 , what do we now know about the parameter
θ. The only answer to this question is to present the entire posterior distribution.”
(p. 6)
In light of that, the question that naturally arises is: what does the revised
probabilistic ranking, based on π(θ|x0 ), ∀θ∈Θ, convey about the underlying
data generating mechanism in (1), assumed to have given rise to data x0 ? That
is, what does this ranking have to do with learning about the ‘true’ value θ ∗ ?
e 0 ) (mode) or some other
It is not obvious why the highest ranked value θ(x
feature of the posterior distribution pinpoints θ∗ better than any other value.
As mentioned above, θ∗ plays no role in selecting the optimal rule, since the
6
latter revolves exclusively around the relevant risk. In contrast, learning from
data x0 about the unique value θ∗ makes perfectly good sense in frequentist
inference since θ is viewed as an unknown constant that gave rise to x0 . This
issue highlights the key built-in tension between the frequentist and Bayesian
approaches.
If the primary objective of Bayesian inference is not the revised probabilistic ranking of all θ∈Θ, what is it? The answer is that the decision-theoretic
perspective came in to refocus and append to Bayesian inference. O’Hagan,
echoing earlier views by Lindley (1965) and Tiao and Box (1975) in contrasting
frequentist (classical) inferences with Bayesian inferences, argues:
“Classical inference theory is very concerned with constructing good inference
rules. The primary concern of Bayesian inference, ..., is entirely different. The
objective is to extract information concerning θ from the posterior distribution,
and to present it helpfully via effective summaries. There are two criteria in this
process. The first is to identify interesting features of the posterior distribution.
... The second criterion is good communication. Summaries should be chosen to
convey clearly and succinctly all the features of interest. ... In Bayesian terms,
therefore, a good inference is one which contributes effectively to appropriating the
information about θ which is conveyed by the posterior distribution.” (p. 14)
Clearly, O’Hagan’s attempt to define what is a ‘good’ Bayesian inference
begs the question: what does constitute ‘effective appropriation of information
about θ’ mean, apart from the probabilistic ranking? Putting that aside, in
his attempt to defend his stance that the entire posterior distribution is the
inference, O’Hagan argues that criteria for ‘optimal’ Bayesian inferences are
only parasitical on the Bayesian approach and enter the picture via the decision
theoretic perspective:
“... a study of decision theory has two potential benefits. First, it provides
a link to classical inference. It thereby shows to what extent classical estimators,
confidence intervals and hypotheses tests can be given a Bayesian interpretation or
motivation. Second, it helps identify suitable summaries to give Bayesian answers
to stylized inference questions which classical theory addresses.” (p. 14)
Both of the above mentioned potential benefits to the Bayesian approach,
are misleading for two reasons. First, the link between the decision-theoretic
and the classical (frequentist) inference is fraught with misleading definitions
and unclarities with respect to the reasoning and objectives of the latter. For
instance, the quantifier ‘∀θ∈Θ’ used to define ‘optimal’ estimators with respect
to particular loss functions is at odds with frequentist inference. Second, the
claim concerning Bayesian answers to frequentist questions of interest is misplaced because the former provides no real answers to the frequentist primary
question of interest which pertains to learning about θ ∗ . A Bayes rule offers very
little, if anything, relevant in learning what the value θ ∗ that gave rise to x0 is.
Let us unpack this answer in more detail.
Substituting the risk function in (3) into the Bayes risk in (5), one can show
7
that:
RB (b
θ) =
=
R
=
R
Rθ∈Θ
R
x∈Rn
X
R
x∈Rn
X
L(θ, b
θ(x))f (x; θ)dx π(θ)dθ=
R
L(θ, b
θ(x))f (x|θ)π(θ)dθdx=
nθ∈Θ
o
R
b
L(θ,
θ(x))π(θ|x)dθ
m(x)dx,
θ∈Θ
x∈Rn
X
(6)
where m(x)= θ∈Θ f (x; θ)dθ; see Bansal (2007). The second and third equalities
presume that one can reverse the order of integration (a technical issue), and
treat f (x; θ) as the joint distribution of X and θ so that:
f (x; θ)=f (x|θ)π(θ)=π(θ|x)m(x).
The latter raises a number of questionable hand-waving moves because it muddies the distinction between x, a generic value of RnX , and the particular value
x0 ; see Spanos (2015). In light of (6), a Bayesian estimate is ‘optimal’ relative
to a particular loss function L(b
θ(X), θ), when it minimizes RB (b
θ). This makes
it clear that what constitutes an ‘optimal’ Bayesian estimate is primarily determined by L(b
θ(X), θ) (Schervish, 1995):
(i) when L2 (b
θ, θ)=(b
θ − θ)2 the Bayes estimate b
θ is the mean of π(θ|x0 ),
(ii) when L1 (e
θ, θ)=|e
θ − θ| the Bayes estimate e
θ is the median of π(θ|x0 ),
0 for θ−θ < ε
(iii) when L0−1 (θ, θ)=δ(θ, θ)=
, for ε>0, the Bayes esti1 for θ−θ ≥ ε
mate θ is the mode of π(θ|x0 ).
In practice, the most widely used loss function is the square:
L2 (b
θ(X) − θ)2 , ∀θ∈Θ,
θ(X); θ)=(b
whose risk function is the decision-theoretic Mean Square Error (MSE 1 ):
R(θ, b
θ)=E(b
θ(X)−θ)2 =MSE1 (b
θ(X); θ), ∀θ∈Θ.
(7)
M SE(b
θn (X); θ ∗ )=E(b
θn (X)−θ∗ )2 .
(8)
R(θ, b
θ) ≤ R(θ, e
θ), ∀θ∈Θ,
(9)
Surprising, however, this definition of the MSE, denoted by MSE1 , is different
from the frequentist MSE, which is defined by:
∗
The key difference is that (8) is defined at the point θ=θ , as opposed to ∀θ∈Θ.
Unfortunately, statistics textbooks adopt one of the two definitions of the MSE
– either at θ=θ∗ or ∀θ∈Θ – and ignore (or seem unaware) of the other. At first
sight, his difference might appear pedantic, but it turns out that it has very
serious implications for the relevant theory of optimality for the frequentist vs.
Bayesian inference procedures. Indeed, reliance on ∀θ∈Θ undermines completely
the relevance of admissibility in frequentist inference.
Admissibility. An estimator e
θ(X) is inadmissible if there exists another
estimator b
θ(X) such that:
and the strict inequality (<) holds for at least one value of θ. Otherwise, e
θ(X)
b
is said to be admissible with respect to the loss function L(θ, θ).
8
3.2
The duality between loss functions and priors
The built-in affinity between the decision-theoretic set-up and Bayesian inference is cemented by a duality result between loss functions and prior distributions; see Robert (2001). This duality stems from the fact that minimizing the
R
Bayes risk:
θ, θ)π(θ)dθ,
RB (b
θ)= θ∈Θ R(b
is equivalent to minimizing the integral:
R
L(b
θ(X), θ)π(θ|x)dθ.
θ∈Θ
This result brings out two important features of Bayesian inference. First,
it confirms the minor role played by the quantifier x∈RnX in both Bayesian
and decision-theoretic optimality theory of inference. Second, it indicates that
L(θ, b
θ) and π(θ) are perfect substitutes with respect to any weight function
w(θ)>0, ∀θ∈Θ, in the derivation of Bayes rules. Modifying the loss function or
the prior yields the same result:
“... the problem of estimating θ with a modified (weighted) loss function is
identical to the problem with a simple loss but with modified hyperparameters of
the prior distribution while the form of the prior distribution does not change.”
(Srivastava et al., 2014, p. 522)
This implies that in practice a Bayesian could derive a particular Bayes
estimate by attaching the weight to the loss function or to the prior distribution
depending on which derivation is easier; see Bansal (2007), Srivastava et al.
(2014).
3.3
Bayes rule and admissibility
As argued in the sequel, it should come as no surprise to learn that Bayes rules
dominate all other rules when admissibility is given center stage.
A Bayes rule b
θ B (X) with respect to a prior distribution π(θ) is:
(i) Admissible, under certain regularity conditions, including when b
θ B (X) is
unique up to equivalence relative to the same risk function.
(ii) Minimax when R(θ, b
θB )=c<∞.
(iii) An admissible estimate b
θ(X) is either Bayes b
θB (X) or the limit of a
sequence of Bayes rules; see Wasserman (2004), Srivastava et al. (2014).
These results have been used as evidence for the superiority of the Bayesian
perspective, and led to the intimation that an effective way to generate optimal
frequentist procedures is to find the Bayes solution using a reasonable prior and
then examine their frequentist properties to see whether it is satisfactory from
the latter viewpoint; see Rubin (1984), Gelman et al (2004).
Even if one were to agree that Bayes rules and admissible estimators largely
coincide, the importance of such a result hinges on the appropriateness of admissibility for frequentist estimators.
4
Loss functions and admissibility revisited
The claim to be discussed in this section is that the notions of a ‘loss function’ and ‘admissibility’ are incompatible with the optimal theory of frequentist
estimation largely framed by Fisher; see Savage (1976).
9
4.1
Admissibility as a minimal property
The following example is used to call into question the notion of a loss function
and the associated property of admissibility for optimal frequentist estimators.
Example. In the context of the simple Normal model:
Xk ∽ NIID(θ, 1), k=1, 2, ..., n, for n > 2,
(10)
let us use the decision-theoretic notion of MSE1 in (7) to compare two estimators
of θ:
Pn
(i) the Maximum Likelihood Estimator (MLE): X n = n1 k=1 Xk ,
(ii) the ‘crystalball’ estimator: θcb =7405926, ∀x∈RnX .
When compared on admissibility grounds, both estimators are admissible, and
thus equally acceptable. Common sense, however, suggests that if a particular
criterion of optimality cannot distinguish between X n [a strongly consistent,
unbiased, fully efficient and sufficient estimator] and θcb , an arbitrarily chosen
real number that ignores the data altogether, is not much of a minimal property.
A moment’s reflection suggests that the inappropriateness of admissibility
stems from its reliance on the quantifier ‘∀θ∈Θ’. The admissibility of θcb arises
from the fact that for certain values of θ close to θcb , say θ∈(θcb ± √λn ), for
0<λ<1, θcb is ‘better’ than X n on MSE1 grounds:
M SE1 (X n ; θ)= n1 > M SE1 (θ cb ; θ) ≤
λ2
n
for θ∈(θcb ± √λn ).
(11)
Given that the primary objective of a frequentist estimator is to pin-point θ ∗ ,
the result in (11) seems totally irrelevant as a gauge of its capacity to achieve
that!
This example indicates that admissibility is totally ineffective as a minimal
property because it does not filter out θcb , the worst possible estimator! Instead,
it excludes potentially good estimators like the sample median; see Cox and
Hinkley (1974). This highlights the ‘extreme relativism’ of admissibility to
the particular loss function, L2 (b
θ(X); θ), in this case. For the absolute loss
function L1 (b
θ(X); θ)=|b
θ(X) − θ|, however, the sample median would have been
the optimal estimator. What determines which loss function is appropriate in
particular cases?
Despite his wholehearted embrace of the decision-theoretic framing, Lehmann
(1984) warned statisticians about the perils of arbitrary loss functions:
“It is argued that the choice of a loss function, while less crucial than that of the
model, exerts an important influence on the nature of the solution of a statistical
decision problem, and that an arbitrary choice such as squared error may be baldly
misleading as to the relative desirability of the competing procedures.” (p. 425)
A strong case can be made that the key minimal property (necessary but not
sufficient) for frequentist estimation is consistency, an extension of the Law of
Large Numbers to estimators. For instance, consistency would have eliminated
both e
θ and θ cb from consideration; they are both inconsistent. This makes
intuitive sense because if an estimator b
θ(X) cannot pinpoint θ ∗ with an infinite
data information, it should be considered impertinent. In contrast, there is
10
nothing in the notion of admissibility that advances learning from data about
θ∗ .
Further to relative (to particular loss functions) efficiency being a dubious
property for frequentist estimators, the pertinent measure of finite sample precision for frequentist estimators is full efficiency, which is defined relative to the
assumed statistical model (1), and not some arbitrary loss function based on
information other than the data.
4.2
Stein’s paradox and admissibility
The quintessential example that has bolstered the appeal of the Bayesian claims
concerning admissibility is the James-Stein estimator (Efron and Morris, 1973),
that gave rise to an extensive literature on shrinkage estimators; see Saleh
(2006).
Let X:=(X1 , X2 , ..., Xm ) be independent sample from a Normal distribution:
Xk ∽ NI(θ k , σ 2 ), k=1, 2, ..., m,
(12)
2
where σ is known. Using the notation θ:=(θ1 , θ2 , ..., θm ) and Im :=diag(1, 1, ..., 1),
this can be denoted by:
X ∽ N(θ, σ 2 Im ).
b
Find an optimal estimator θ(X) of θ with respect to the square ‘overall’ loss
function:
P
2
2
b
b
b
(13)
)= m
L2 (θ, θ(X))=(k
θ(X)−θk
k=1 (θ k (X)−θ k ) .
Stein (1956) astounded the statistical world by showing that for m=2 the LeastbLS (X)=X is admissible, but for m > 2 θ
bLS (X) is
Squares (LS) estimator θ
inadmissible. Indeed, James and Stein (1961) were able to come up with a
nonlinear estimator:
2
bJS (X)= 1 − (m−2)σ
θ
X,
(14)
2
kXk
bLS (X)=X
that became known as the James-Stein estimator, which dominates θ
in MSE1 terms by demonstrating that:
bJS (X); θ) < MSE1 (θ
bLS (X); θ), ∀θ∈Rm .
(15)
MSE1 (θ
b
It turns out that θ JM (X) is also inadmissible for m > 2 and dominated by the
modified James-Stein estimator that is admissible:
+
2
b+ (X)= 1 − (m−2)σ
θ
X,
2
JS
kXk
+
where (z) = max(0, z); see Wasserman (2004).
The traditional interpretation of this result is that for the Normal, Independent model in (12), the James–Stein estimator (14) of θ:=(θ1 , θ2 , ..., θm ), for
m > 2, reduces the overall MSE1 in (13). This result seems to imply that one
will ‘do better’ (in overall MSE1 terms) by using a combined nonlinear (shrinkage) estimator, instead of estimating these means separately. What is surprising
about this result is that there is no statistical reason (due to independence) to
connect the inferences pertaining to the different individual means, and yet the
obvious estimator (LS) is inadmissible.
As argued next, this result calls into question the appropriateness of the
notion of admissibility with respect to a particular loss function, and not the
judiciousness of frequentist estimation.
11
5
Frequentist inference and learning from data
The objectives and underlying reasoning of frequentist inference are inadequately discussed in the statistics literature. As a result some of its key differences with Bayesian inference remain beclouded.
5.1
Frequentist approach: primary objective and reasoning
All forms of parametric frequentist inference begin with a prespecified statistical model Mθ (x)={f (x; θ), θ∈Θ}, x∈RnX . This model is chosen from the
set of all possible models that could have given rise to data x0 :=(x1 , ..., xn ),
by selecting the probabilistic structure for the underlying stochastic process
{Xt , t∈N:=(1, 2, ..., n, ...)} in such a way so as to render the observed data x0
a ‘typical’ realization thereof. In light of the fact that each value of θ∈Θ represents a different element of the family of models represented by Mθ (x), the
primary objective of frequentist inference is to learn from data about the ‘true’
model:
M∗ (x)={f (x; θ∗ )}, x∈RnX ,
(16)
∗
where θ denotes the true value of θ in Θ. The ‘typicality’ is testable vis-a-vis
the data x0 using misspecification testing; see Spanos(2006).
The frequentist approach relies on two modes of reasoning for inference
purposes:
Factual (estimation, prediction):
Hypothetical (hypothesis testing):
f (x; θ∗ ), ∀x∈RnX ,
f (x; θ0 ), f (x; θ1 ), ∀x∈RnX ,
where θ ∗ denotes the true value of θ in Θ, and θi , i=0, 1 denote hypothesized
values of θ associated with the hypotheses, H0 : θ0 ∈Θ0 , H1 : θ1 ∈Θ1 , where Θ0
and Θ1 constitute a partition of Θ.
A frequentist estimator b
θ aims to pin-point θ ∗ , and its optimality is evaluated
by how effectively it achieves that. Similarly, a test statistic usually compares a
good estimator b
θ of θ with a prespecified value θ0 , but behind b
θ is the value θ∗
assumed to have generated data x0 . Hence, the hypothetical reasoning is used
in testing to learn about θ∗ , and has nothing to do with all possible values of θ
in Θ.
This contradicts misleading claims by Bayesian textbooks (Robert, 2001, p.
61):
“The frequentist paradigm relies on this criterion [risk function] to compare
estimators and, if possible, to select the best estimator, the reasoning being that
estimators are evaluated on their long-run performance for all possible values of the
parameter θ.”
Contrary to this claim, the only relevant value of θ in evaluating the ‘optimality’
of b
θ is θ ∗ . Such misleading claims stem from an apparent confusion between the
existential and universal quantifiers in framing certain inferential assertions.
The existence of θ∗ can be formally defined using the existential quantifier:
∃θ ∗ ∈Θ :
there exists a θ ∗ ∈Θ such that.
12
This introduces a potential conflict between the existential and the universal
quantifier ‘∀θ∈Θ’ because neither the decision theoretic nor the Bayesian approach explicitly invoke θ ∗ . Decision-theoretic and Bayesian rules are considered
optimal when they minimize the expected loss ∀θ∈Θ, no matter what θ∗ happens to be. How different the two quantifiers are, can be demonstrated using
elementary logic. The logical connective for negation (¬) can be used to define
the following equivalence relationships between the two quantifiers:
(i) ∃θ ∗ ∈Θ ⇐⇒ ¬∀θ ∈Θ,
/
(ii) ∀θ∈Θ ⇐⇒ ¬∃θ ∗ ∈Θ
/
Given that (i)-(ii) involve double negations, the two quantifiers could not be
further apart.
At first sight the quantifier ∀θ∈Θ seems rather innocuous and natural in
the context of statistical inference. It seems intuitively obvious that one should
care about the behavior of an estimator b
θ for all possible values of θ. This is
misleading intuition, however, since the behavior of b
θ, for all θ∈Θ, although
relevant, is not what determines how effective a frequentist estimator is at pinpointing θ∗ ; what matters is its behavior around θ∗ . Assessing its effectiveness
calls for evaluating (deductively) the sampling distribution of b
θ under θ=θ ∗ , or
hypothetical values θ0 and θ1 , and not for all possible values of θ in Θ. Let’s
unpack the details of this claim.
5.2
Frequentist estimation
The underlying reasoning for frequentist estimation is factual , in the sense the
optimality of an estimator is appraised in terms of its generic capacity of b
θn (X)
to zero-in on (pinpoint) the true value θ∗ , whatever the sample realization
X=x0 . Optimal properties like consistency, unbiasedness, full efficiency, sufficiency, etc., calibrate this generic capacity using its sampling distribution of
b
θn (X) evaluated under θ=θ∗ i.e., in terms of f (b
θn (x); θ ∗ ), for x∈RnX . For instance, strong consistency asserts that as n → ∞, b
θn (X) will zero-in on θ∗
almost surely:
P( lim b
θn (X)=θ∗ )=1.
n→∞
Similarly, unbiasedness asserts that the sampling distribution of b
θn (X) has a
mean equal to θ∗ :
E(b
θ n (X))=θ∗ .
In this sense both of these optimal properties are defined at the point θ=θ∗ . This
is achieved by using factual reasoning, i.e. evaluating the sampling distribution
of b
θ n (X) under the true state of nature (θ=θ∗ ), without having to know θ∗ . This
is in contrast to using loss functions, such as (2), which are defined in terms of
θ∗ but are rendered non-operational without knowing θ ∗ .
Example. In the case of the simple Normal model in (10) the point estimator, X n is consistent, unbiased, fully efficient, sufficient, with a sampling
distribution:
(17)
X n ∽ N(θ, n1 ).
What is not usually explicitly stated is that the evaluation of that distribution
is factual, i.e. θ=θ∗ , and should formally denoted by:
θ=θ ∗
X n ∽ N(θ∗ , n1 ).
13
When X n is standardized, it yields the pivotal function:
θ=θ∗
√
d(X;θ):= n X n − θ∗ ∽ N(0, 1),
(18)
whose distribution only holds for the true θ∗ , and no other value. This provides
the basis for constructing a (1−α) Confidence Interval (CI):
P X m − c α2 ( √1n ) ≤ θ ≤ X n + c α2 ( √1n ); θ=θ∗ =1−α,
(19)
which asserts that the random interval [X n −c α2 ( √sn ), X n +c α2 ( √sn )], will cover
(overlay) the true mean θ∗ , whatever that happens to be, with probability (1−α),
or equivalently, the error of coverage is α. Hence, frequentist estimation the
coverage error probability depends only on the sampling distribution of X n and
is attached to random interval for all values θ6=θ ∗ without requiring one to know
θ∗ .
The evaluation at θ=θ ∗ calls into question the decision-theoretic definition
of unbiasedness:
E(b
θ n (X))=θ, ∀θ∈Θ,
for frequentist estimation since this assertion makes no sense for all values θ
in Θ, but does make sense when defined at θ=θ∗ . Similarly, the appropriate
frequentist definition of the MSE for an estimator, initially proposed by Fisher
(1920), is defined at the point θ=θ∗ :
(20)
M SE(b
θn (X); θ ∗ )=E(b
θn (X)−θ∗ )2 , for θ ∗ in Θ.
Indeed, the well-known decomposition:
M SE(b
θ(X); θ ∗ )=V ar(b
θ(X))+[E(b
θ n (X))−θ ∗ ]2 , for θ∗ in Θ,
(21)
∗
is meaningful only when defined at the point θ=θ (true mean) since by definition:
V ar(b
θ(X))=E[b
θ n (X)−θm ]2 , θ m =E(b
θ n (X))
Bias(b
θn (X); θ ∗ )=E(b
θn (X))−θ ∗ ,
(22)
and thus, the variance and the bias involve only two values of θ in Θ, θ m and
θ∗ , and when θm =θ∗ the estimator is unbiased. This implies that the apparent
affinity between the MSE1 defined in (7) and the variance of an estimator is
more apparent than real because the latter makes frequentist sense only when
θm =E(b
θn (X)) is a single point.
5.3
James-Stein estimator from a frequentist perspective
For a proper frequentist evaluation of the above James-Stein result, it is important to bring out the conflict between the overall MSE (13) and the factual
reasoning underlying frequentist estimation. From the latter perspective, the
James-Stein estimator raises several issues of concern.
bLS (X) and the James-Stein θ
bJS (X) estimaFirst, both the Least-Squares θ
tors are inconsistent estimators of θ since the underlying model suffers from the
incidental parameter problem: there is essentially one observation (Xk ) for each
unknown parameter (θ k ), and as m → ∞ the number of unknown parameters
increases at the same rate. To bring out the futility of comparing these two
estimators more clearly, consider the following simpler example.
14
Example. Let X:=(X1 , X2 , ..., Xn ) be a sample from the simple Normal
model in (10). Comparing the two estimators b
θ 1 =Xn , b
θ2 = 21 (X1 +Xn ) and
b
b
inferring that θ2 is relatively more efficient than θ1 relative to a square loss
function, i.e.
MSE(b
θ 2 (X); θ)=1 < MSE(b
θ1 (X); θ)= 12 , ∀θ∈R,
is totally uninteresting because both estimators are inconsistent!
Second, to be able to discuss the role of admissibility in the Stein (1956)
result, we need to consider a consistent James-Stein estimator, by extending
the original data to a panel (longitudinal) data where the sample is:
Xt :=(X1t , X2t , ..., Xmt ), t=1, 2, ..., n. In this case the consistent Least-Squares
and James-Stein estimators are:
n
bLS (X)= X 1 , X 2 , ..., X m , where X k = 1 P Xkt , k=1, 2, ..., m,
θ
n
t=1
+
2
+
(m−2)σ
b
θ JS (X)= 1 − kXk2
X, where X:= X 1 , X 2 , ..., X m .
This enables us to evaluate the notion of ‘relatively better’ more objectively.
Admissibility relative to the overall loss function in (13) introduces a tradeoff between the accuracy of the estimators for individual parameters θ:=(θ 1 , θ2 , ..., θ m )
and the ‘overall’ expected loss. The question is: ‘In what sense the overall MSE
among a group of mean estimates provides a better measure of ‘error’ in learning about the true values θ∗ :=(θ ∗1 , θ∗2 , ..., θ∗m )?’ The short answer is: it doesn’t.
Indeed, the overall MSE will be irrelevant when the primary objective of estimation is to learn from data about θ∗ . This is because the particular loss function
penalizes the estimator’s capacity to pin-point θ ∗ by trading an increase in bias
for a decrease in the overall MSE in (13), when the latter is misleadingly evaluated over all θ in Θ:=Rm . That is, the James-Stein estimator flouts the primary
objective of pin-pointing θ∗ in favor of reducing the overall MSE ∀θ∈Θ.
In summary, the above discussion suggests that there is nothing paradoxical
about Stein’s (1956) original result. What is problematic is not the least-squares
estimator, but the choice of ‘better’ in terms of admissibility relative to an overall
MSE in evaluating the accuracy of the estimators of θ.
5.4
Frequentist hypothesis testing
Another frequentist inference procedure one can employ to learn from data
about θ∗ is hypothesis testing, where the question posed is whether θ∗ is close
enough to some prespecified value θ 0 . In contrast to estimation, the reasoning
underlying frequentist testing is hypothetical in nature.
5.4.1 Legitimate frequentist error probabilities
For testing the hypotheses:
H0 : θ ≤ θ 0 vs. H1 : θ > θ 0 , where θ 0 is a prespecified value,
one utilizes the same sampling distribution X n ∽N(θ, n1 ), but transforms the
√
pivot d(X;θ):= n X n −θ∗ into the test statistic by replacing θ∗ with the pre
√
specified value θ0 , yielding d(X):= n X n −θ0 . However, instead of evaluating
15
it under the factual θ=θ∗ , it is now evaluated under various hypothetical scenarios associated with H0 and H1 to yield two types of (hypothetical) sampling
distributions:
θ=θ
√
(I)
d(X):= n X n − θ 0 ∽ 0 N(0, 1),
θ=θ
√
√
(II) d(X):= n X n −θ 0 ∽ 1 N(δ 1 , 1), δ 1 = n (θ1 −θ0 ) for θ1 > θ0 .
In both cases (I)-(II) the underlying reasoning is hypothetical in the sense that
the factual in (18) is replaced by hypothesized values of θ, and the test statistic
d(X) provides a standardized distance between the hypothesized values (θ0 or
θ1 ) and θ∗ the true θ, assumed to underlie the generation of the data x0 , yielding d(x0 ). Using the sampling distribution in (I) one can define the following
legitimate error probabilities:
significance level: P(d(X) > cα ; H0 ) = α,
(23)
p-value: P(d(X) > d(x0 ); H0 )=p(x0 ).
Using the sampling distribution in (II) one can define:
type II error prob.: P(d(X)≤cα ; θ=θ1 )=β(θ 1 ), for θ 1 >θ 0 ,
(24)
power: P(d(X)>cα ; θ=θ1 )=̺(θ 1 ), for θ1 >θ0 .
It can be shown that the test Tα , defined by the test statistic d(X) and the
rejection region C1 (α)={x :d(x) > cα }, constitutes a Uniformly Most Powerful
(UMP) test for significance level α; see Lehmann (1959). The type I [II] error
probability is associated with test Tα erroneously rejecting [accepting] H0 . The
type I and II error probabilities evaluate the generic capacity [whatever the
sample realization x∈Rn ] of a test to reach correct inferences. Contrary to
Bayesian claims, these error probabilities have nothing to do with the temporal
or the physical dimension of the long-run metaphor associated with repeated
samples. The relevant feature of the long-run metaphor is the repeatability
(in principle) of the DGM represented by Mθ (x); this feature can be easily
operationalized using computer simulation; see Spanos (2013).
The key difference between the significance level α and the p-value is that
the former is a pre-data and the latter a post-data error probability. Indeed, the
p-value can be viewed as the smallest significance level α at which H0 would
have been rejected with data x0 . The legitimacy of post-data error probabilities underlying the hypothetical reasoning can be used to go beyond the N-P
accept/reject rules and provide an evidential interpretation pertaining to the
discrepancy γ from the null warranted by data x0 ; see Mayo and Spanos (2006).
Despite the fact that frequentist testing uses hypothetical reasoning, its main
∗
∗
n
objective is also to learn from data about the
M (x)={f (x; θ )}, x∈RX .
√ true model
This is because a test statistic like d(X):= n X n −θ0 constitutes nothing more
than a scaled distance between θ∗ [the value behind the generation of xn ], and
a hypothesized value θ0 , with θ∗ being replaced by its ‘best’ estimator X n .
6
Revisiting loss and risk functions
The above discussion raises serious questions about the role of loss functions
and admissibility in evaluating learning from data x0 about θ∗ . In particular:
16
(i) What does the extraneous information concerning costs associated with
different parameter values have to do with learning about θ∗ ?
(ii) In what sense is an inconsistent but relatively (to a particular loss function) efficient an ‘optimal’ estimator for learning about θ∗ ?
(iii) Why is the overall MSE more important than learning from data about
the true values of θ∗ ?
6.1
Where do loss functions come from?
A closer scrutiny of the decision-theoretic set up reveals that the loss function
needs to invoke ‘information from sources other than the data’, which is usually
not readily available. Indeed, such information is available in very restrictive
situations, such as acceptance sampling in quality control. In light of that,
a proper understanding of the intended scope of statistical inference calls for
distinguishing the special cases where the loss function is part and parcel of the
available substantive information from those that no such information is either
relevant or available.
Tiao and Box (1975), p. 624, reiterated Fisher’s (1935) distinction:
“Now it is undoubtedly true that on the one hand that situations exist where
the loss function is at least approximately known (for example certain problems in
business) and sampling inspection are of this sort. ... On the other hand, a vast
number of inferential problems occur, particularly in the analysis of scientific data,
where there is no way of knowing in advance to what use the results of research will
subsequently be put.”
Cox (1978), p. 45, went further and questioned this framing even in cases
where the inference might involve a decision:
“The reasons that the detailed techniques [decision-theoretic] seem of fairly limited
applicability, even when a fairly clear cut decision element is involved, may be (i)
that, except in such fields as control theory and acceptance sampling, a major
contribution of statistical technique is in presenting the evidence in incisive form for
discussion, rather than in providing mechanical presentation for the final decision.
This is especially the case when a single major decision is involved. (ii) The central
difficulty may be in formulating the elements required for the quantitative analysis,
rather than in combining these elements via a decision rule.”
Another important aspect of using loss functions in inference is that in practice they seem to be an add-on to the inference itself since they bring to the
problem the information other than the data. In particular, the same statistical
inference problem can give rise to very different decisions/actions depending
on one’s loss function. To illustrate that consider an example from Chatterjee
(2002):
“... consider the case of a new drug whose effects are studied by a research
scientist attached to the laboratory of a pharmaceutical company. The conclusion
of the study may have different bearings on the action to be taken by (a) the
scientist whose line of further investigation would depend on it, (b) the company
whose business decisions would determined by it, and (c) the Government whose
policies as to health care, drug control, etc. would take shape on that basis.” (p.
72)
17
In practice, each one of these different agents is likely to have a very different
loss function, but their inferences have a common denominator: the scientific
evidence which relating to the true θ that stems solely from the observed data?
Finally, the extreme relativism of loss function optimality renders decisiontheoretic and Bayes rules highly vulnerable to abuse. In practice, one can justify
any estimator as optimal, however lame in terms of other criteria, by selecting
the ”appropriate” loss function.
Example. Consider a manufacturer of high precision bolts and nuts who has
information that the buyer only checks the first and last box for quality control
when accepting an order. This suggests that to minimize losses, stemming from
the return of its products as defective, an appropriate loss function might be:
2
L(X; θ)= ([(X1 +Xn )/2] − θ) , θ∈(0, 1).
(25)
The ‘optimal’ estimator relative to (25) is e
θ=(X1 +Xn )/2, but e
θ is a terrible
∗
estimator for pinpointing θ because it is inconsistent!
6.2
Loss functions vs. inherent distance functions
The notion of a loss function stemming from ‘information other than the data’
raises another source of potential conflict. This stems from the fact that within
each statistical model Mθ (x) there exists an inherent statistical distance function, often relating to the log-likelihood and the score function, and hence stemming from information contained in the data; see Casella and Berger (2002).
It is well-known that when the distribution underlying Mθ (x) is Normal,
the inherent distance function for comparing estimators of the mean (θ) is the
square:
N D(b
θn (X); θ ∗ )=(b
θ n (X) − θ∗ )2 .
On the other hand, when the distribution is Laplace the relevant statistical
distance function is the Absolute Distance (see Shao, 2003):
AD(b
θ n (X); θ ∗ ) = |b
θ n (X) − θ∗ |.
Similarly, when the distribution underlying Mθ (x) is Uniform, the inherent
distance function is:
SU P (b
θn (X); θ ∗ ) = sup |b
θn (x) − θ∗ |.
x∈Rn
X
A key feature of all these distance functions is that they are defined at the point
θ=θ∗ and not for all θ in Θ, as the traditional loss functions.
The question that naturally arises is when it might make sense to ignore
these inherent distance functions and compare estimators using an externally
given loss function. The key difference between the two is that any assumptions
that comprise the likelihood function are testable vis-a-vis the data, but those
underlying the loss function are not. Moreover, the likelihood function gives
rise to a ‘global’ notion of optimality, known as full efficiency defined at θ=θ∗
in terms of Fisher’s information:
2
∂ ln L(θ)
∗
∗
.
(θ
),
I
(θ
):=E
−
CR(θ ∗ )=I−1
n
⊤
n
∂θ∂θ
What is an optimal estimator depends only on the information contained in the
statistical model Mθ (x). This contrasts with admissibility which is a property
defined in terms of ‘local’ optimality relative to a loss function based on outside
information and evaluated ∀θ∈Rm .
18
6.3
Decisions vs. inferences
The above discussion brings out the crucial distinction between a ‘decision’ and
an ‘inference’ stemming from data x0 .
Even before Wald (1939) introduced the decision-theoretic perspective, Fisher
(1935) perceptively argued:
“In the field of pure research no assessment of the cost of wrong conclusions,
or of delay in arriving at more correct conclusions can conceivably be more than a
pretence, and in any case such an assessment would be inadmissible and irrelevant
in judging the state of the scientific evidence.” (pp. 25-26)
Tukey (1960) echoed Fisher’s view by contrasting decisions vs. inferences:
“Like any other human endeavor, science involves many decisions, but it progresses by the building up of a fairly well established body of knowledge. This body
grows by the reaching of conclusions – by acts whose essential characteristics differ
widely from the making of decisions. Conclusions are established with careful regard to evidence, but without regard to consequences of specific actions in specific
circumstances.” (p. 425)
Tukey also recognized how decision theory distorts frequentist testing by
replacing error probabilities with losses and costs:
“Wald’s decision theory ... has given up fixed probability of errors of the first
kind, and has focused on gains, losses or regrets.” (p. 433)
Hacking (1965) brought out the key difference between an ‘inference pertaining to evidence’ for or against a hypothesis, and a ‘decision to do something’ as
a result of an inference:
“... to conclude that an hypothesis is best supported is, apparently, to decide
that the hypothesis in question is best supported. Hence it is a decision like any
other. But this inference is fallacious. Deciding that something is the case differs
from deciding to do something. ... Hence deciding to do something falls squarely in
the province of decision theory, but deciding that something is the case does not.”
(p. 31)
This issue was elaborated upon by Birnbaum (1977), p. 19:
“Two contrasting interpretations of the decision concept are formulated: behavioral, applicable to ‘decisions’ in a concrete literal sense as in acceptance sampling;
evidential,
and
applicable to ‘decisions’ such as ’reject H0 ’ in a research context, where
the pattern and strength of statistical evidence concerning statistical hypotheses is
of central interest.”
6.4
Acceptance sampling vs. learning from data
Let us bring out the key features of a situation where the above decisiontheoretic set up makes perfectly good sense. This is the situation Fisher (1955)
called acceptance sampling, such as an industrial production process where the
objective is quality control, i.e. to make a decision pertaining to shipping substandard products (e.g. nuts and bolts) to a buyer using the expected loss/gain
as the ultimate criterion.
In an acceptance sampling context, the MSE(b
θ(X); θ), or some other risk
function, are relevant because they evaluate genuine losses associated with a
decision related to the choice of an estimate b
θ(x0 ), say the cost of the observed
19
percentage of defective products, but that has nothing to do with type I and II
error probabilities.
Acceptance sampling differs from a scientific enquiry in two crucial respects:
[a] The primary aim is to use statistical rules to guide actions astutely, e.g.
use b
θ(x0 ) in order to minimize the expected loss associated with “a decision”,
and
[b] The sagacity of all actions is determined by the respective ‘losses’ stemming from “relevant information other than the data” (Cox and Hinkley, 1974,
p. 251).
The key difference between acceptance sampling and a scientific inquiry is
that the primary objective of the latter is not to minimize expected loss (costs,
utility) associated with different values of θ∈Θ, but to use data x0 to learn
about the ‘true’ model (16). The two situations are drastically different mainly
because the key notion of a ‘true θ’ calls into question the above acceptance
sampling set up. Indeed, the loss function being defined ‘∀θ∈Θ’, will penalize
θ∗ , since there is no reason to believe that the lowest ranked θ would coincide
with θ∗ , unless by accident.
Consider the case where acceptance sampling resembles hypothesis testing in
so far as final products are randomly selected for inspection during the production process. In such a situation the main objective can be viewed as operationalizing the probabilities of false acceptance/rejection with a view to minimize the
expected losses. The conventional wisdom has been that this situation is similar
enough to Neyman-Pearson (N-P) testing to render the latter as the appropriate
framing for the decision to ship this particular batch or not. However, a closer
look at some of the examples used to illustrate such a situation (Silvey, 1975),
reveals that the decisions are driven exclusively by the risk function and not
by any quest to learn from data about the true θ∗ . For instance, N-P way of
addressing the trade-off between the two types of error probabilities, fixing α to
a small value and seek a test that minimizes the type II error probability, seems
utterly irrelevant in such a context. One can easily think of a loss function
where the ‘optimal’ trade-off calls for a much larger type I than type II error
probability.
In light of the above discussion, what is different in acceptance sampling is
that:
[c] The trade-off between the two types of error probabilities is determined
by the risk function itself, and not by any attempt to learn from data about θ ∗ .
Indeed, this learning is deliberately undermined by certain loss function such as
the overall MSE (13) that favor biased estimators of the James-Stein type.
Given the crucial differences in [a]-[c], one can make a strong case that the
objectives and the underlying reasoning of acceptance sampling are drastically
different from those pertaining to learning from data in a scientific context.
6.5
Is expected loss a legitimate frequentist error?
The key question is whether expected loss is a legitimate frequentist error like
bias, MSE and the type I-II error. ‘What do these legitimate frequentist errors
have in common?’
20
First, they stem directly from the statistical model Mθ (x) since the underlying sampling distributions of estimators, test statistics and predictors are
derived exclusively from the distribution of the sample f (x; θ) via (4). In this
sense, the relevant error probabilities are directly related to statistical information pertaining to the data as summarized by the statistical model Mθ (x)
itself.
Second, they are attached to a particular frequentist inference procedure as
they relate to a relevant inferential claim. These error probabilities calibrate
the effectiveness of inference procedures in learning from data about the true
statistical model M∗ (x)={f (x; θ∗ )}, x∈RnX .
In light of these features, the question is: ‘in what sense a risk function could
potentially represent relevant frequentist errors?’ That argument that the risk
function represents legitimate frequentist errors because it is derived by taking
expectations with respect to f (x; θ), x∈RnX (Robert (2001), is misguided for
two reasons.
(a) The relevant errors in estimation, including the bias E(b
θ n (X))−θ∗ and
∗
∗
2
MSE E(b
θn (X)−θ ) , are evaluated with respect to f (x; θ ), x∈RnX , by invoking
factual reasoning (θ ∗ is assumed to be the state of Nature). Wald’s (1939)
original loss function in (2) represents an interesting case because it is defined
in terms of θ∗ , which renders it non-operational when evaluated for all θ in Θ,
since θ∗ is unknown in practice. In contrast, the errors associated with the bias
and MSE are rendered operational by the factual reasoning fashioned to forgo
knowing θ∗ .
(b) The expected losses stemming from the risk function R(θ, b
θ) are attached
to particular values of θ in Θ. Such an assignment is in direct conflict with all the
above legitimate error probabilities that are attached to the inference procedure
itself, and never to the particular values of θ in Θ. The expected loss assigned
to each value of θ in Θ has nothing to do with learning from data about θ ∗ .
Indeed, the risk function will penalize a procedure for pin-pointing θ∗ since the
latter is unknown in practice. This is in direct conflict with the main objective
of frequentist estimation but in sync with ‘acceptance sampling’, where the
objective of the inference has everything to do with expected losses.
7
Summary and conclusions
The paper makes a case for Fisher’s (1935; 1955) assertions concerning the appropriateness of the decision-theoretic framing for ‘acceptance sampling’ and its
inappropriateness for frequentist inference. A closer look at this framing reveals
that it is congruent with the Bayesian approach because provides it with a theory of optimal inference. Decision-theoretic and Bayesian rules are considered
optimal when they minimize the expected loss for all possible values of θ [∀θ∈Θ],
irrespective of what the true value θ ∗ happens to be. In contrast, the theory
of optimal frequentist inference revolves around the true value θ∗ , since it depends entirely on the capacity of the procedure to pinpoint θ∗ . The frequentist
approach relies on factual (estimation, prediction), as well as hypothetical (testing) reasoning, both of which revolve around the existential quantifier ∃θ ∗ ∈Θ.
21
The inappropriateness of the quantifier ∀θ∈Θ calls into question the relevance
of admissibility as a minimal property for frequentist estimators. A strong case
can be made that the relevant minimal property for frequentist estimators is
consistency. In addition, full efficiency provides the relevant measure of an estimator’s finite sample efficiency (accuracy) in pinpointing θ∗ . Both of these
properties stem from the underlying statistical model Mθ (x), in contrast to
admissibility which relies on loss functions based on information other than the
data.
It is argued that Stein’s (1956) result stems from the fact that admissibility
introduces a trade-off between the accuracy of the estimator in pinpointing θ∗
and the ‘overall’ expected loss. That is, the James-Stein estimator achieves
a higher overall MSE by blunting the capacity of a frequentist estimator to
pinpoint θ ∗ . Why would a frequentist care about the overall MSE defined for all
θ in Θ? After all, expected losses are not legitimate errors similar to bias and
MSE (when properly defined), as well as coverage, type I and II errors. The
latter are attached to the frequentist procedures themselves to calibrate their
capacity to achieve learning from data about θ ∗ . In contrast, expected losses
are assigned to different values of θ in Θ, using information other than the data.
References
[1] Bansal, A.K. (2007), Bayesian Parametric Inference, Alpha Science, Oxford.
[2] Berger, J.O. (1985), Statistical Decision Theory and Bayesian Analysis,
2nd edition, Springer, NY.
[3] Birnbaum, A. (1977), “The Neyman-Pearson Theory as Decision Theory,
and as Inference Theory; with a Criticism of the Lindley-Savage argument
for Bayesian Theory”, Synthese, 36: 19-49.
[4] Casella, G. and R. L. Berger (2002), Statistical Inference, 2nd ed., Duxbury,
CA.
[5] Chatterjee, S.K. (2002), Statistical Thought: A Perspective and History,
Oxford University Press, Oxford.
[6] Cox, D. R. (1958), “Some Problems Connected with Statistical Inference,”
The Annals of Mathematical Statistics, 29: 357-372.
[7] Cox, D.R. (1978), “Foundations of Statistical Inference: the Case for Eclecticism”, Australian Journal of Statistics, 20: 43-59.
[8] Cox, D.R. and D.V. Hinkley (1974), Theoretical Statistics, Chapman &
Hall, London.
[9] Efron, B. and C.N. Morris (1973), “Stein’s estimation rule and its
competitors–an empirical Bayes approach”, Journal of the American Statistical Association, 68: 117–130.
[10] Ferguson, T.S. (1967), Mathematical Statistics: A Decision Theoretic Approach, Academic Press, London.
22
[11] Ferguson, T.S. (1976), “Development of the Decision Model”, ch. 16 in On
the History of Statistics and Probability, edited by D. B. Owen, Marcel
Dekker, NY.
[12] Fisher, R.A. (1920), “A Mathematical Examination of the Methods of Determining the Accuracy of an Observation by the Mean Error, and by the
Mean Square Error”, Monthly Notices of the Royal Astronomical Society,
80: 758-770.
[13] Fisher, R.A. (1935), The Design of Experiments, Oliver and Boyd, Edinburgh.
[14] Fisher, R.A. (1955), “Statistical methods and scientific induction”, Journal
of the Royal Statistical Society, B, 17: 69-78.
[15] Gelman, A., J.B. Carlin and D.B. Rubin (2004), Bayesian Data Analysis,
2nd edition, Chapman & Hall, London.
[16] Hacking, I. (1965), Logic of Statistical Inference, Cambridge University
Press, Cambridge.
[17] James, W. and C. Stein (1961), “Estimation with quadratic loss”, Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and
Probability, 1: 361–379.
[18] LeCam, L. (1955), “An extension of Wald’s theory of statistical decision
functions”, Annals of Mathematical Statistics, 26: 69-81.
[19] LeCam, L. (1986), Asymptotic Methods in Statistical Decision Theory,
Springer, NY.
[20] Lehmann, E.L. (1959), Testing Statistical Hypotheses, Wiley, NY.
[21] Lehmann, E.L. (1984), “Specification Problems in the Neyman-PearsonWald Theory”, pp. 425-436 in Statistics: An Appraisal, edited by H.A.
David and H.T. David, The Iowa State University Press, Ames, IA.
[22] Lindley, D.V. (1965), Introduction to Probability and Statistics from a
Bayesian Viewpoint, Part 2: Inference, Cambridge University Press, Cambridge.
[23] Mayo, D. G. and A. Spanos, (2006), “Severe Testing as a Basic Concept in
a Neyman-Pearson Philosophy of Induction,” The British Journal for the
Philosophy of Science, 57: 323-357.
[24] Neyman, J. (1937), “Outline of a theory of statistical estimation based on
the classical theory of probability”, Philosophical Transactions of the Royal
Society of London, Series A, 236: 333-380.
[25] Neyman, J. (1952), Lectures and Conferences on Mathematical Statistics
and Probability, 2nd ed. U.S. Department of Agriculture, Washington.
[26] Neyman, J. (1967), A Selection of Early Statistical Papers by J. Neyman,
University of California Press, CA.
[27] Neyman, J. (1971), “Foundations of Behavioristic Statistics”, pp.1-13, in
Godambe, V. and Sprott, D. (1971), eds., Foundations of Statistical Inference, Holt, Rinehart and Winston of Canada, Toronto.
[28] O’Hagan, A. (1994), Bayesian Inference, Edward Arnold, London.
[29] Reid, N. (2015), “Statistical Sufficiency”, pp. 418–422 in International Encyclopedia of the Social & Behavioral Sciences, edited by J.D. Wright, 2nd
edition, Vol 23. Elsevier, Oxford.
23
[30] Robert, C.P. (2001), The Bayesian Choice: From Decision-Theoretic Foundations to Computational Implementation, 2nd ed., Springer, NY.
[31] Rubin, D.B. (1984), “Bayesianly justifiable and relevant frequency calculation for the applied statistician”, Annals of Statistics, 12: 1151-1172.
[32] Saleh, A.K. Md. E. (2006), Theory of Preliminary Test and Stein-Type
Estimation with Applications, Wiley-Interscience, NY.
[33] Savage, L.J. (1976), “On Rereading R.A. Fisher”, The Annals of Statistics,
4(3): 441-500.
[34] Schervish, M.J. (1995), Theory of Statistics, Springer-Verlag, NY.
[35] Shao, J. (2003), Mathematical Statistics, 2nd ed., Springer, NY.
[36] Silvey, S.D. (1975), Statistical Inference, Chapman & Hall, London.
[37] Spanos, A. (2006), “Where Do Statistical Models Come From? Revisiting
the Problem of Specification”, pp. 98-119 in Optimality: The Second Erich
L. Lehmann Symposium, edited by J. Rojo, Lecture Notes-Monograph Series, vol. 49, Institute of Mathematical Statistics.
[38] Spanos, A. (2013), “A Frequentist Interpretation of Probability for Modelbased Inductive Inference”, Synthese, 190: 1555-1585.
[39] Spanos, A. (2015), “Revisiting Bayes’ Rule and Evidence: Notional Events
vs. Real Data”, Virginia Tech, working paper.
[40] Spanos, A. and D.G. Mayo (2015), “Error statistical modeling and inference: Where methodology meets ontology”, Synthese, 192: 3533-3555.
[41] Srivastava, M.K., A.H. Khan and N. Srivastava (2014), Statistical Inference: Theory of Estimation, PHI Learning, Delhi, India.
[42] Stein, C. (1956), “Inadmissibility of the usual estimator for the mean of
a multivariate distribution”, Proceedings of the Third Berkeley Symposium
on Mathematical Statistics and Probability, 1: 197–206.
[43] Tiao, G.C. and G.E.P. Box (1975), “Some comments on ”Bayes” estimators”, pp. 619-626 in Studies in Bayesian Econometrics and Statistics, In
Honor of Leonard J. Savage, edited by S. E Fienberg and A. Zellner, NorthHolland, Amsterdam.
[44] Tukey, J.W. (1960), “Conclusions vs Decisions”, Technometrics, 2: 423433.
[45] Wald, A. (1939), “Contributions to the Theory of Statistical Estimation
and Testing Hypotheses”, Annals of Mathematical Statistics, 10: 299-326.
[46] Wald, A. (1947), “An essentially complete family class of admissible decision functions”, Annals of Mathematical Statistics, 18: 549-555.
[47] Wald, A. (1950), Statistical Decision Functions, Wiley, NY.
[48] Wasserman, L. (2004), All of Statistics, Springer, NY.
24
| 10math.ST
|
Facial Expression Recognition Based on Complexity Perception Classification
Algorithm
Tianyuan Chang , Guihua Wen* , Yang Hu , JiaJiong Ma
School of Computer Science and Engineering, South China University of Technology,Guangzhou, China
[email protected], [email protected]
Abstract
Facial expression recognition (FER) has always
been a challenging issue in computer vision. The
different expressions of emotion and uncontrolled
environmental factors lead to inconsistencies in the
complexity of FER and variability of between expression categories, which is often overlooked in
most facial expression recognition systems. In order
to solve this problem effectively, we presented a
simple and efficient CNN model to extract facial
features, and proposed a complexity perception
classification (CPC) algorithm for FER. The CPC
algorithm divided the dataset into an easy classification sample subspace and a complex classification
sample subspace by evaluating the complexity of
facial features that are suitable for classification.
The experimental results of our proposed algorithm
on Fer2013 and CK+ datasets demonstrated the
algorithm’s effectiveness and superiority over other
state-of-the-art approaches.
1
Introduction
Facial expression recognition (FER) has a wide range of
research prospects in human computer interaction and affective computing, including polygraph detection, intelligent
security, entertainment, Internet education, intelligent medical treatment, etc. As we all know facial expression is a
major way of expressing human emotions. Hence, the main
task in determining emotion is how to automatically, reliably,
and efficiently recognize the information conveyed by facial
expressions. In FER research, Ekman and Friensen first
proposed the Facial Action Coding System (FACS) [Friensen
and Ekman, 1983]. The six basic categories of expressions
(surprise, sadness, disgust, anger, happiness, and fear) are
defined in FACS, and are commonly used as the basic expression labels.
FER methods can be roughly divided into three types [Sun
et al., 2017; Yan et al., 2017]: geometry-based methods,
appearance-based methods, and hybrid methods. Geometry-based methods describe features mainly through geometric relationships such as key points on the face, face
positions and shapes. Appearance-based methods extract the
entire facial image as a whole. Lastly, hybrid methods com-
bine the first two methods to extract local and global facial
features of face images, respectively. Compared with appearance-based methods, the utility of geometry-based
methods requires further improvement because it is difficult
to accurately label the key points and effective positions of
the face in practical applications.
Most FER works focus on feature extraction and classifier
construction, which can be divided into static and dynamic
classification methods [Corneanu et al., 2016]. Static classification methods are applicable for a single static image, and
include SVM classifiers, Bayesian network classifiers,
Random forests, and Softmax classifiers. Dynamic classification is applied to facial video sequences, taking into account features independently extracted from each frame over
time as a basis; the main dynamic classification models are
HMM [Le et al., 2011] and VSL-CRF [Walecki et al., 2015].
In recent years, different methods of traditional machine
learning have been used in FER to extract appearance features of images, including Gabor filters, local binary patterns
(LBP) [Savran et al., 2015], local Gabor binary patterns
(LGBP) [Zhang et al., 2005], histograms of oriented gradients (HOG) [Dahmane et al., 2011], and scale-invariant
feature transform (SIFT) [Berretti et al., 2011]. These traditional methods tend to be more effective in specific small
sample sets, but are difficult to adjust to identify new face
images, which brings challenges to FER in uncontrolled
environments. This is because the extracted features often
belong to low-level features, and it is hard to extract and find
meaningful information that distinguishes the different categories from the data. However, a new recognition framework utilizing a convolutional neural network (CNN)
[Ranzato et al., 2011; Liu et al., 2014; Kim et al., 2015] based
on a deep learning network has achieved remarkable results
in FER, and has been used for feature extraction and recognition. Multiple convolution and pooling layers in a CNN
may extract higher and multi-level features of the entire face
or local area, and has good classification performance of
facial expression image features.
At present, there are still some deficiencies in FER. Many
related work focus on the improvement of classification
models and feature extraction methods, easily ignoring the
relevance of several basic expression categories and the
inconsistencies in the complexity of extracting features.
As described by [Lopes et al., 2017], it is difficult to defini-
tively partition each expression feature space. Some expressions, like happiness and surprise, belong to highly recognizable categories, which are easily distinguished by facial
features. Meanwhile, other expressions, like fear and sadness,
are very similar in some situations, making it difficult to
distinguish them effectively. In addition, facial images are
easily influenced by ethnicity, age, gender, hair, and other
uncontrolled factors, resulting in different facial feature distributions and facial feature complexity for classification.
Supposing that there are inconsistencies between the training
samples and test samples, no matter how suitable the feature
extraction of training samples in expression recognition, the
prediction result of the test samples will not be accurate. An
analogy of this is that we cannot ask an excellent student who
has only grasped primary mathematics knowledge to be
tested on knowledge of higher mathematics in a university. It
is obvious that the complexity of the unequal knowledge
needs to be distinguished for learning.
In order to effectively solve the above problems, in this
paper we proposed a complexity perception classification
(CPC) algorithm for FER referenced simple and effective
principle. Our algorithm firstly divided the dataset into two
parts: the difficult classification sub dataset, and the easy
classification sub dataset. This was done by evaluating the
complexity of facial features for expression recognition of
each sample. The next step was to separately learn the two
different feature distributions, that is, we trained two specific
classifiers for these two sub datasets to obtain the complexity
of different sample features. Instead of using a uniform
classifier to predict the facial expression, our algorithm divided the test sample into corresponding classifiers to complete FER through the facial feature complexity discriminator.
Our main contributions can be summarized as follows.
1) According to the scale of the facial expression dataset, we
proposed a simple and efficient CNN model based on ResNets for facial feature extraction, which effectively alleviated
the problem of gradient disappearance, and enhanced information flow in the deep network.
2) A new heuristic classification algorithm named complexity perception classification (CPC) was proposed; it not only
improved the recognition accuracy of higher recognizable
facial expression categories, but also alleviated the problem
of some misclassified expression categories.
3) We achieved state-of-the-art performance on static facial
expression recognition benchmarks, including the Fer2013
dataset and CK+ dataset.
work blocks, and fully connected layers. In order to extract
high level facial features, we selected the output of the second 1024 dimensional fully connected layer as the feature
representation.
2
Learning from the simple validity in Occam's razor principle,
we proposed a complexity perception classification (CPC)
algorithm to solve the complexity inconsistency problem of
FER.
In this section, we firstly illustrated how to evaluate the
complexity of training datasets for recognition.Then we
employed the discriminant learining for different disjoint sub
datasets and sample complexity discriminator to clearly
determine the complexity of test sample for recognition.
Figure 3 shows an overview and pipeline of the CPC system.
Feature Extraction
The classical improvement of CNN frameworks using ResNet [He et al., 2016] and DenseNet [Huang et al., 2016] not
only alleviated the problem that deep networks are prone to
gradient disappearance in backpropagation, but also remarkably improved the performance of image classification.
To this end and motivated by the above network architectures, we proposed a CNN framework for facial feature
extraction as shown in Figure 1. Our CNN framework contained multiple convolutional layers, modified residual net
Figure 1: Pipeline of the CNN framework
In RestNets, the output of the feature mapping of the residual block consists of a non-linearly transformed composite
function H(x) and an identity function x, which are combined
as in Equation 1:
F (x) H(x) x
(1)
This combination may hinder the flow of information
through deep networks. In order to improve the flow of information between layers, we improved the combination
mode of the residual block. Motivated by DenseNet, we no
longer summated the two inputs, but concatenated the two
feature mappings. The output function of the feature mapping
is shown in Equation 2.
F (x) [H(x), x]
(2)
Figure 2 shows the structure of the traditional residual
block and our modified residual block.
Figure 2: Left: Traditional residual block. Right: Modified residual
block.
3
Complexity Perception Classification
Algorithm for FER
Figure 3: Pipeline of the complexity perception classification (CPC) system
3.1 Evaluating the Complexity of Features for
Recognition
3.2 The Discriminant Learning
We adopted the feature extraction using the CNN as the input
of the FER system for each feature value fi (i 1,...,n) , where
fi is the i-th feature extracted from the original image. The
feature extraction of training samples X [x1 , x2 ,..., x n ] were
randomly divided into K folds, where xi Rd , i 1, 2,..., n .
x i is the i-th sample feature vectors and d is the feature dimension of each sample. In order to improve the generalization ability of the base classifier and to distinguish the easily
classifiable samples from the difficultly classifiable samples,
we first chose a fold of samples as the training set and the
remaining K-1 folds of samples as the test set. This resulted
in K base classifiers on different training datasets. We repeated the above process m times to obtain
N(N KM) trained base classifiers. Then, each training
sample feature vector was predicted by n base classifiers,
counting the correct number of prediction expression categories and computing the ease degree of classification
R(x i ) for the i-th sample:
R(x i )
N (x i )
N
(3)
where N denotes the number of the base classifiers, and
N(xi ) is the correct classification number in N base classifiers of the i-th training sample.
We evaluated the complexity of sample classification
features by the ease degree of classification R(x i ) . In addition, we set a parameter named ease threshold as the
boundary to distinguish the easily classifiable samples from
the difficultly classifiable samples.
S
xi E
S D
if R(x i )
if R(x i )
(4)
As shown in Equation 4, when the ease degree of classification for the i-th training sample R(xi ) , we divided it
into the easy classification sample subspace (SE ,LE ) . By
contrast, we divided it into the difficult classification sample
subspace (SD ,LD ) when the ease degree of classification for
the i-th training sample R(xi ) .
We divided the training datasets into two disjoint sub datasets
by evaluating the feature complexity for recognition. In order
to achieve discriminant learning in different sub datasets, we
trained an easy sample classifier C E in the easy classification
sample subspace (SE ,LE ) , and trained a difficult sample
classifier C D in the difficult classification sample subspace (SD ,LD ) . In this way, we could learn different feature
distributions in two different sample subspaces to provide
more accurate recognition performance for the test sample.
The easy sample classifier C E and the difficult sample classifier C D were as follows:
CE (SE ,LE ) ,
(5)
CD (SD ,LD ) ,
(6)
where is the specified classification method, such as Softmax, linear SVM, Random forest, etc. S E denotes the easy
classification dataset, and LE denotes the expression labels
of S E . Similarly, S D denotes the difficult classification dataset,
and LD denotes the expression labels of S D .
3.3 Sample Complexity Discriminator
In order to be able to clearly determine whether the test
sample belongs to the easy classification sample subspace or
difficult classification sample subspace, we dynamically
generated different sample complexity discriminators for
different test samples. The sample complexity discriminator
model is described as follows:
C s (SE S D ,{ , }) ,
(7)
where {+} is the label of the easy classification dataset, {-} is
the label of the difficult classification dataset. The sample
complexity discriminator model C s denotes a dynamic classification method of KNN-Softmax. We adopted the KNN
algorithm to find k neighbors for the i-th test sample in the
training set, and dynamically trained the sample complexity
discriminator through the k nearest neighbor training set. The
KNN-Softmax dynamic classification method could not only
reflect the advantages of Softmax, but also improve the dis-
crimination accuracy of the feature complexity for each test
sample.
If the i-th test sample predicted {+} in the sample complexity discriminator, we used the easy sample classifier to
recognize the facial expression. By contrast, if the label was
{-}, we used the difficult sample classifier to recognize the
facial expression.
Algorithm 1 elaborates the implementation method and the
details of the CPC algorithm.
Algorithm 1 Complexity perception classification (CPC)
Input: Feature vector x i from each sample, where i 1,..., n , classifier , Complexity discriminator , and ease threshold .
Output: The predictive expression label for the testing sample.
Training:
1: Partition data into training, validation, and test sets.
2: for i = 1 to m do
3: partition training data into K folds.
4: for j = 1 to k do
5:
apply to train j-th training data, which gets a base classifier.
6: end for
7: end for
8: Apply N base classifiers to compute the ease degree of classification R(x i ) for each sample.
9: for each training sample x i do
10: if R(x i )
11: Divide the sample x i into S E .
12: else
13: Divide the sample x i into S D .
14: end for
15: C E (SE ,LE )
16: C D (SD ,L D )
17: C I (SE S D ,{ , }) ]
Testing:
18: for each training sample do
19: y C I (x)
20: if y { } , C E (x)
21: if y { } , C D (x)
22: end for
4
Experiment
Our proposed algorithm was mainly applied to single-image
static facial expression recognition. We evaluated the performance of the proposed complexity perception classification (CPC) algorithm by applying it to the Fer2013 [Goodfellow et al., 2013] and CK+ [Lucey et al., 2010] datasets.
4.1 Datasets
The Fer2013 dataset is a facial expression recognition challenge dataset that ICML 2013 launched on Kaggle. The dataset contains 28709 training images, 3589 validation images,
and 3589 test images. The image size is 48 48 , and facial
expression is classified into seven different types (0=Angry,
1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise,
6=Neutral).
The CK+ dataset is an extension of the CK dataset. It
contains 327 labeled facial videos, with each classified into
(a) Fer2013_Baseline
(b) Fer2013_CPC
(c) CK+_Baseline
(d) CK+_CPC
Figure 4: Confusion matrices for facial expressions on two datasets.
The baseline is the result of feature extraction with the CNN predicted by Softmax. The CPC is the result of feature extraction with
the CNN predicted by Softmax and adding the CPC algorithm.
one of the seven categories mentioned above, i.e., the same
categories as those used in the Fer2013 dataset. We extracted
four frames from each sequence in the CK+ dataset, which
contains a total of 1308 facial expressions.
4.2 Experimental Settings
Before the feature extraction phase in the CNN, we preprocessed the dataset with ZCA whitening and global relative
normalization, which can effectively remove the redundant
information of the input image and reduce the correlation
between adjacent pixels in the image. In the pre-training
phase of feature extraction with the CNN designed in Figure
1, we used the Fer2013 training dataset to train the networks.
The initial network learning rate was set to 0.05, while the
decay of the learning rate was per epoch. The mini batch size
was 128, the momentum was set to 0.5, and the dropout was
set to 0.2. The activation function used in the convolutional
layers was the rectified linear unit (ReLU) activation function. The stochastic gradient descent (SGD) was used as the
optimization algorithm. In addition, the last fully connected
layer used Softmax as a multi-class activation function.
4.3 Baseline Classification vs. CPC
We firstly observed the impact of the CPC algorithm on the
recognition accuracy of different expression categories.
Figure 4 illustrates the resulting confusion matrices of testing
accuracies (%) on the two datasets; we employed five-fold
cross validation to ensure the stability of results. The effectiveness of the proposed algorithm was verified by compar-
ing the effect of the complexity perception classification
algorithm on recognition results after the feature extraction
of the CNN.
On Fer2013 (Figure 4a), we can clearly see that the
recognition rates of happiness were significantly higher than
those of the other expressions belonging to the easily distinguishable category; meanwhile, the recognition rates of
fear was the most difficult to distinguish. After adding the
complexity perception classification algorithm to Fer2013
(Figure 4b), we found the following. (1) The recognition rate
of happiness increased by 0.87% in the case of a high baseline recognition rate, the recognition rate of fear similarly
increased by 1.81%, and the error rate related to mistaking
fear for sadness decreased by 0.61%. These exciting recognition results demonstrated that the proposed algorithm not
only improved the recognition accuracy of easily distinguishable categories, but also alleviated the easy misclassification of difficultly distinguishable categories. (2) The
recognition rates of anger, sadness, and neutral also increased.
The proposed algorithm improved the recognition rates for
most categories. (3) None of the recognition rates of the
expression classes decreased due to the addition of the proposed algorithm. Notice that our proposed algorithm did not
raise the recognition accuracy of certain classes at the cost of
sacrificing the accuracy of other classes, which is meaningful
for practical application in facial expression recognition.
On CK+ (Figure 4c, 4d), the recognition accuracy of the
natural class increased from 95.71% to 97.14%, while the
other categories experienced no change in recognition accuracy upon adding the proposed algorithm. The results of the
confusion matrix on the CK+ dataset also validate our above
conclusion.
4.4 Performance of the CPC Algorithm
Table 1 show the recognition accuracy of different classifiers
on the two datasets. We extracted facial features with the
CNN, and compared the recognition effect of the CPC algorithm with the baseline recognition. In the experiment, we
respectively trained two normal classifiers C n as the baseline
recognition by employing the feature vectors of the training
set on two datasets. The experimental results showed that the
CPC algorithm had significant performance in facial expression recognition.
Compared to the baseline recognition rate on the Fer2013
test set, the average recognition rate of Softmax, linear SVM,
and random forest classifiers were enhanced by 1.58%,
1.06%, and 0.59%, respectively, by adding the CPC algorithm. The average baseline recognition rate of the CK+ test
set of the three different classifiers reached 97.76%. When
we introduced the CPC algorithm, this average recognition
increased to 98.27%.
4.5 Comparison of State-of-the-art Methods
We compared the proposed CNN feature extraction based on
the CPC algorithm with existing related methods for FER.
Table 2 and Table 3 respectively show the recognition
accuracies of the different FER methods and our proposed
Table 1: Comparison of recognition accuracies (%) of the baseline
with CPC with different classifiers on the Fer2013 and CK+ datasets.
Method
Fer2013
CNN+Softmax
CNN+LinearSVM
CNN+RandomForest
Baseline
69.77
70.02
69.97
CK+
CPC
71.35
71.08
70.56
Baseline
98.38
97.97
96.92
CPC
98.78
98.36
97.65
Table 2: Recognition rates (%) on the Fer2013 dataset for the
state-of-the-art methods
Method
Recognition rate (%)
Unsupervised [Goodfellow et
69.26
al., 2013]
Maxim Milakov [Goodfellow et 68.82
al., 2013]
[Mollahosseini et al., 2016]
66.40
[Tang et al., 2013]
71.16
[Liu et al., 2016]
65.03
DNNRL[Guo et al., 2016]
70.60
FC3072[Kim et al., 2016]
70.58
Proposed Approach
71.35
Table 3: Recognition rates (%) on the CK+ dataset for the
state-of-the-art methods
Method
Recognition rate (%)
[Mollahosseini et al., 2016]
93.20
[Liu et al., 2014]
94.19
[Wu et al., 2015]
98.54
[Happy et al., 2015]
94.09
[Sun et al., 2016]
94.87
[Yan et al., 2017]
96.60
[Barman et al., 2017]
98.30
[Lopes et al., 2017]
95.75
[Sariyanidi et al., 2017]
96.02
Proposed Approach
98.78
algorithm on the CK+ and Fer2013 datasets. As can be seen
in Table 2, our proposed algorithm achieved competitive
results in the Fer2013 dataset. Table 3 shows that our proposed algorithm outperformed all other compared FER
methods in the CK+ dataset.
4.6 Parameter Analysis
We investigated the importance of the ease threshold parameter in the complexity perception classification algorithm.
In order to determine the parameter value that produced the
FER best performance on both datasets, we employed
five-fold cross validation on the Fer2013 validation set and
the CK+ dataset. Figure 5 and Figure 6 show the recognition
accuracy of our algorithm versus different values of the ease
threshold on two datasets, and compares the recognition
accuracy with the baseline.
As can be seen in the above two figures, the recognition
Acknowledgments
This work was supported by the China National Science
Foundation (60973083, 61273363), the Science and Technology Planning Project of Guangdong Province
(2014A010103009, 2015A020217002), and the Guangdong
Science
and
Technology
Planning
Project
(201504291154480).
References
Figure 5: Recognition accuracy of the proposed algorithm versus
different values of on the Fer2013 validation set.
Figure 6: The recognition accuracy of the proposed algorithm versus different values of on the CK+ dataset.
accuracy of our algorithm exceeded the baseline for most
values of , which illustrates the robustness of our algorithm.
We also found that the random forest classifier was not stable
enough due to its randomness. In addition, the relationship
between recognition accuracy and different values of was
similar for the linear SVM and Softmax classifiers.
5 Conclusions
In this paper, we designed a new and efficient CNN model
for facial feature extraction, and proposed a complexity
perception classification(CPC) algorithm for facial expression recognition. Our algorithm distinguished easily classifiable subspace from difficultly classifiable subspace to
achieve discriminant learning of the facial feature distribution. Experimental results on the Fer2013 and CK+ datasets
demonstrated
that
our
algorithm
outperformed
state-of-the-art methods for facial expression recognition in
terms of mean recognition accuracy.
[Friensen et al., 1983] W. Friensen, P. Ekman, EMFACS-7:
Emotional Facial Action Coding System, Technical
Report, University of California, San Francisico, 1983.
[Sun et al., 2017] Wenyun Sun, Haitao Zhao, Zhong Jin. An
Efficient Unconstrained Facial Expression Recognition
Algorithm based on Stack Binarized Auto-encoders and
Binarized Neural Networks. Neurocomputing, 2017.
[Yan et al., 2017] Haibin Yan. Collaborative discriminative
multi-metric learning for facial expression recognition in
video. Pattern Recognition, 2017.
[Corneanu et al., 2016] CA Corneanu, MO Simón, JF Cohn,
et al. Survey on RGB, 3D, Thermal, and Multimodal
Approaches for Facial Expression Recognition: History,
Trends, and Affect-Related Applications. IEEE
Transactions on Pattern Analysis & Machine Intelligence,
2016.
[Le et al., 2011] Vuong Le, H. Tang, and T.S. Huang.
Expression recognition from 3D dynamic faces using
robust spatio-temporal shape features. IEEE International
Conference on Automatic Face & Gesture Recognition
and Workshops IEEE, pages 414-421, 2011.
[Walecki et al., 2015] R. Walecki, O. Rudovic, V. Pavlovic,
and M. Pantic, Variablestate latent conditional random
fields for facial expression recognition and action unit
detection, in Proc. IEEE Int. Conf. Autom. Face Gesture
Recog., 2015.
[Savran et al., 2015] A. Savran, H. Cao, A. Nenkova, and R.
Verma, Temporal Bayesian fusion for affect sensing:
Combining video, audio, and lexical modali-ties, IEEE
Trans. Cybern., pages 1927–1941, 2015.
[Dahmane et al., 2011] M Dahmane, J Meunier. Emotion
recognition using dynamic grid-based HoG features.
IEEE International Conference on Automatic Face &
Gesture Recognition and Workshops IEEE, 2011.
[Berretti et al., 2011] S Berretti, BB Amor, M Daoudi. 3D
facial expression recognition using SIFT descriptors of
automatically detected keypoints. Visual Computer
International Journal of Computer Graphics, 2011.
[Zhang et al., 2005] W Zhang, S Shan, X Chen. Local Gabor
Binary Pattern Histogram Sequence (LGBPHS): A Novel
Non-Statistical Model for Face Representation and
Recognition." Tenth IEEE International Conference on
Computer Vision IEEE, pages 786-791, 2005.
[Ranzato et al., 2011] M Ranzato, J Susskind, V Mnih G
Hinton. On deep generative models with applications to
recognition. Computer Vision and Pattern Recognition
IEEE, pages 2857-2864, 2011.
[Liu et al., 2014] Mengyi Liu,R Wang, X Chen. Learning
Expressionlets on Spatio-temporal Manifold for Dynamic
Facial Expression Recognition. IEEE Conference on
Computer Vision and Pattern Recognition IEEE
Computer Society, pages 1749-1756, 2014.
[Kim et al., 2015] BK Kim, H Lee, J Roh. Hierarchical
Committee of Deep CNNs with Exponentially-Weighted
Decision Fusion for Static Facial Expression Recognition.
ACM on International Conference on Multimodal
Interaction ACM, pages 427-434, 2015:.
[Lopes et al., 2017] AT Lopes, ED Aguiar, AFD Souza.
Facial expression recognition with Convolutional Neural
Networks: Coping with few data and the training sample
order. Pattern Recognition pages 610-628, 2017.
[He et al., 2016] K He, X Zhang, S Ren, J Sun. Deep Residual
Learning for Image Recognition. Computer Vision and
Pattern Recognition IEEE, pages 770-778, 2016.
[Huang et al., 2016] G Huang, Z Liu, LVD Maaten, KQ
Weinberger.
Densely
Connected
Convolutional
Networks, 2016.
[Goodfellow et al., 2013] IJ Goodfellow, D Erhan, PL Carrier.
Challenges in Representation Learning: A Report on
Three Machine Learning Contests. International
Conference on Neural Information Processing Springer,
Berlin, Heidelberg, pages 117-124, 2013.
[Lucey et al., 2010] P Lucey, JF Cohn, T Kanade, J Saragihet.
The Extended Cohn-Kanade Dataset (CK+): A complete
dataset for action unit and emotion-specified expression.
Computer Vision and Pattern Recognition Workshops
IEEE, pages 94-101, 2010.
[Mollahosseini et al., 2016] Ali Mollahosseini, D. Chan, and
MH Mahoor. Going deeper in facial expression
recognition using deep neural networks. Applications of
Computer Vision IEEE, pages 1-10, 2016.
[Tang et al., 2013] Y Tang. Deep Learning using Linear
Support Vector Machines. Computer Science, 2013.
[Liu et al., 2016] Kuang Liu, M. Zhang, and Z. Pan. Facial
Expression Recognition with CNN Ensemble.
International Conference on Cyberworlds IEEE, pages
163-166, 2016.
[Guo et al., 2016] Yanan Guo, J Yu, H Xiong. Deep Neural
Networks with Relativity Learning for facial expression
recognition. IEEE International Conference on
Multimedia & Expo Workshops IEEE Computer Society,
pages 1-6, 2016.
[Kim et al., 2016] BoKyeong Kim, J Roh, and SY Dong.
Hierarchical committee of deep convolutional neural
networks for robust facial expression recognition. Journal
on Multimodal User Interfaces, pages 1-17, 10.2, 2016.
[Wu et al., 2015] Chongliang Wu,S Wang, and Q Ji.
Multi-instance Hidden Markov Model for facial
expression recognition. IEEE International Conference
and Workshops on Automatic Face and Gesture
Recognition IEEE, pages 1-6, 2015.
[Happy et al., 2015] SL Happy, and A Routray. Automatic
facial expression recognition using features of salient
facial patches. IEEE Transactions on Affective
Computing , pages 1-12, 6.1, 2015.
[Sun et al., 2016] Yaxin Sun, and Wen G. Cognitive facial
expression recognition with constrained dimensionality
reduction. Neurocomputing, 2016.
[Barman et al., 2017] Asit Barman, and P Dutta. Facial
expression recognition using distance and shape signature
features. Pattern Recognition Letters, 2017.
[Sariyanidi et al., 2017] Evangelos Sariyanidi, H Gunes, and
A Cavallaro. Learning Bases of Activity for Facial
Expression Recognition. IEEE Transactions on Image
Processing. pages 1965-1978, 2017.
| 2cs.AI
|
1
Mobile Unmanned Aerial Vehicles (UAVs) for
Energy-Efficient Internet of Things
Communications
Mohammad Mozaffari1 , Walid Saad1 , Mehdi Bennis2 , and Mérouane Debbah3
arXiv:1703.05401v2 [cs.IT] 12 Sep 2017
1
Wireless@VT, Electrical and Computer Engineering Department, Virginia Tech, VA, USA,
Emails:{mmozaff,walids}@vt.edu.
2
3
CWC - Centre for Wireless Communications, Oulu, Finland, Email: [email protected].
Mathematical and Algorithmic Sciences Lab, Huawei France R & D, Paris, France, and CentraleSupelec,
Université Paris-Saclay, Gif-sur-Yvette, France, Email: [email protected].
Abstract
In this paper, the efficient deployment and mobility of multiple unmanned aerial vehicles (UAVs),
used as aerial base stations to collect data from ground Internet of Things (IoT) devices, is investigated.
In particular, to enable reliable uplink communications for IoT devices with a minimum total transmit
power, a novel framework is proposed for jointly optimizing the three-dimensional (3D) placement and
mobility of the UAVs, device-UAV association, and uplink power control. First, given the locations of
active IoT devices at each time instant, the optimal UAVs’ locations and associations are determined.
Next, to dynamically serve the IoT devices in a time-varying network, the optimal mobility patterns
of the UAVs are analyzed. To this end, based on the activation process of the IoT devices, the time
instances at which the UAVs must update their locations are derived. Moreover, the optimal 3D trajectory
of each UAV is obtained in a way that the total energy used for the mobility of the UAVs is minimized
while serving the IoT devices. Simulation results show that, using the proposed approach, the total
transmit power of the IoT devices is reduced by 45% compared to a case in which stationary aerial
base stations are deployed. In addition, the proposed approach can yield a maximum of 28% enhanced
system reliability compared to the stationary case. The results also reveal an inherent tradeoff between
the number of update times, the mobility of the UAVs, and the transmit power of the IoT devices. In
essence, a higher number of updates can lead to lower transmit powers for the IoT devices at the cost
of an increased mobility for the UAVs.
I. I NTRODUCTION
The use of unmanned aerial vehicles (UAVs) as flying wireless communication platforms
has received significant attention recently [1]–[8]. On the one hand, UAVs can be used as
wireless relays for improving connectivity and coverage of ground wireless devices. On the
2
other hand, UAVs can act as mobile aerial base stations to provide reliable downlink and uplink
communications for ground users and boost the capacity of wireless networks [1], [2], [6]–[12].
Compared to the terrestrial base stations, the advantage of using UAV-based aerial base stations
is their ability to provide on-the-fly communications. Furthermore, the high altitude of UAVs
enables them to effectively establish line-of-sight (LoS) communication links thus mitigating
signal blockage and shadowing. Due to their adjustable altitude and mobility, UAVs can move
towards potential ground users and establish reliable connections with a low transmit power.
Hence, they can provide a cost-effective and energy-efficient solution to collect data from ground
mobile users that are spread over a geographical area with limited terrestrial infrastructure.
Indeed, UAVs can play a key role in the Internet of Things (IoT) which is composed of small,
battery-limited devices such as sensors, and health monitors [13]. These devices are typically
unable to transmit over a long distance due to their energy constraints [13]. In such IoT scenarios,
UAVs can dynamically move towards IoT devices, collect the IoT data, and transmit it to other
devices which are out of the communication ranges of the transmitters [13]. In this case, the
UAVs play the role of moving aggregators or base stations for IoT networks [5]. However,
to effectively use UAVs for the IoT, several challenges must be addressed such as optimal
deployment, mobility and energy-efficient use of UAVs as outlined in [6] and [1].
In [3], the authors investigated the optimal trajectory of UAVs equipped with multiple antennas
for maximizing sum-rate in uplink communications. The work in [4] maximizes the throughput of
a relay-based UAV system by jointly optimizing the UAV’s trajectory as well as the source/relay
transmit power. However these works considered a single UAV in their models. In [1], we
investigated the optimal deployment and movement of a single UAV for supporting downlink
wireless communications. The work in [14] proposed a low-complexity algorithm for the optimal
deployment of multiple UAVs that provide coverage for ground users. The work in [10] provided
a comprehensive downlink coverage analysis for a network in which a finite number of UAVs
serve the ground users. In [15], the authors used UAVs to efficiently collect data and recharge the
clusters’ head in a wireless sensor network which is partitioned into multiple clusters. However,
this work is limited to a static sensor network, and does not investigate the optimal deployment
of the UAVs. While the energy efficiency of uplink data transmission in a machine-to-machine
(M2M) communication network was investigated [16], the presence of UAVs was not considered.
In fact, none of the prior studies in [1]–[16], addressed the problem of jointly optimizing the
3
deployment and mobility of UAVs, device association, and uplink power control for enabling
reliable and energy-efficient communications for IoT devices. To our best knowledge, this paper
is one of the first comprehensive studies on the joint optimal 3D deployment of aerial base
stations, device association, and uplink power control in an IoT ecosystem.
The main contribution of this paper is to introduce a novel framework for optimized deployment and mobility of multiple UAVs for the purpose of energy-efficient uplink data collection
from ground IoT devices. In particular, we consider an IoT network in which the IoT devices can
be active at different time instances. To minimize the total transmit power of these IoT devices,
given device-specific signal-to-interference-plus-noise-ratio (SINR) constraints, we propose an
efficient approach to jointly and dynamically find the UAVs’ locations, the association of devices
to UAVs, and the optimal uplink transmit power. Our proposed framework is composed of two
key steps. First, given the locations of the IoT devices, we propose a solution for optimizing
the deployment and association of the UAVs. In this case, we solve the formulated problem by
decomposing it into two subproblems which are solved iteratively. In the first subproblem, given
the fixed UAVs’ locations, we find the jointly optimal device-UAV association and the devices’
transmit power. In the second subproblem, given the fixed device association, we determine
the joint 3D UAVs’ locations. For this subproblem, we transform the non-convex continuous
location optimization problem to a convex form and provide tractable solutions. Next, following
our proposed algorithm, the results of solving the second subproblem are used as inputs to
the first subproblem for the next iteration. Here, we show that our proposed approach leads to
an efficient solution with a reasonable accuracy compared to the global optimal solution that
requires significant overhead. Clearly, the UAVs’ locations and the device association that we
obtain in this first step will depend on the locations of active IoT devices.
In the second step, we analyze the IoT network over a time period during which the set of
active devices changes. In this case, we present a framework for optimizing the UAVs’ mobility
by allowing them dynamically update their locations depending on the time-varying devices’
activation process. First, we derive the closed-form expressions for the time instances (update
times) at which the UAVs must move according to the activation process of the devices. Next,
using the update time results, we derive the optimal 3D UAVs’ trajectory such that the total
movement of the UAVs while updating their locations is minimized. Our simulation results
show that, using the proposed approach, the total transmit power of the IoT devices can be
4
significantly reduced compared to a case in which stationary aerial base stations are deployed.
The results also verify our analytical derivations for the update times and reveal an inherent
tradeoff between the number of updates, the mobility of the UAVs, and transmit power of the
IoT devices. In particular, it is shown that a higher number of updates leads to lower transmit
powers for the IoT devices at the cost of higher UAVs’ energy consumptions.
The rest of this paper is organized as follows. In Section II, we present the system model
and problem formulation. Section III presents the optimal deployment of UAVs and device
association. In Section IV, we address the mobility and update time of the UAVs. In Section V
we provide the simulation and analytical results, and Section VI draws some conclusions.
II. S YSTEM M ODEL AND P ROBLEM F ORMULATION
Consider an IoT system consisting of a set L = {1, 2, ..., L} of L IoT devices. Examples of
such devices include various types of sensors used for environmental monitoring, smart traffic
control, and smart parking devices. In this system, a set K = {1, 2, ..., K} of K rotary wing UAVs
must be deployed to collect the data from the ground IoT devices. These UAVs can dynamically
move, when needed, to effectively serve the IoT devices using uplink communication links. Here,
the term served by a UAV implies that the uplink SINR is above the threshold and, thus, the
UAV can successfully collect data from the ground IoT device. In our model, we assume that the
devices transmit their data to the UAVs in the uplink using frequency division multiple access
(FDMA) over R orthogonal channels. Let Emax be the maximum energy that each UAV can
spend on its movement. The locations of device i ∈ L and UAV j ∈ K are, respectively, given
uav
by (xi , yi ) and v j = (xuav
j , yj , hj ) as shown in Fig. 1. In our model, we consider a centralized
network in which the locations of the devices and UAVs are known to a control center located at
a central cloud server. The cloud server will determine the UAVs’ locations, device association,
and the transmit power of each IoT device.
We analyze the IoT network within a time interval [0, T ] during which the IoT devices can be
active at different time instances and must be served by the UAVs at some pre-defined time slots.
At the beginning of each slot, the positions of the UAVs as well as the device-UAV association
are updated based on the locations of currently active devices that are assumed to be known to the
5
cloud center1 . Hereinafter, the time instance at which the UAVs’ locations and associations are
jointly updated, is referred to as the update time. The update times are denoted by tn , 1 ≤ n ≤ N ,
with N being the number of updates. At each update time tn , based on the location of active
devices, the optimal UAVs’ locations and the corresponding association must be determined for
effectively serving the ground devices. Here, the IoT devices that become active during [tn−1 , tn )
are served by the UAVs during the time period [tn , tn+1 ). Note that, during [tn−1 , tn ), the UAVs’
locations and their device association do not change until the next update time, tn . Clearly,
since at different update times, a different subset of devices might be active, the locations of
the UAVs must dynamically change at each update time. Therefore, each UAV’s trajectory will
consist of N stop locations at which the UAV serves the ground devices. Note that, in our model,
the UAVs locations are not necessarily updated once the set of active devices changes. Instead,
we consider some specific time instances (update times) at which the UAVs locations device
associations, and devices’ transmit power are optimized. In particular, considering the fact that
the set of active devices may continuously change, continuously updating the UAVs’ locations,
the devices transmit powers, and the device-UAV associations may not be feasible as it can
lead to low reliability, high UAVs’ energy consumption, and a need to solve complex real-time
optimization processes. In our model, the update times are design parameters that depend on the
activity of the devices, and the energy of UAVs. Given this model, our objective is to find the
optimal joint UAVs’ locations and device association at each update time tn so as to minimize
the total transmit power of the active devices while meeting each device’s SINR requirement.
Moreover, we need to develop a framework for determining the update times as well as the
UAVs’ mobility to handle dynamic changes in the activity of the devices. To this end, first, we
present the ground-to-air channel model and the activation models for the IoT devices.
A. Ground-to-Air Path Loss Model
In our model, while optimizing the locations of the UAVs, the information available includes
the ground devices’ locations, and the type of environment (e.g. rural, suburban, urban, highrise
urban, etc.). Note that, in such practical scenarios, one will not have any additional information
about the exact locations, heights, and number of the obstacles. Therefore, one must consider the
1
We consider static IoT devices in delay-tolerant applications which their fixed locations and activation patterns are known
to the cloud center.
6
uav
move UAVj ( x uav
j , y j , hj )
move
d ij
hj
Devicei ( xi , yi )
Control center
IoT device
Fig. 1: System model.
randomness associated with the LoS and NLoS links while designing the UAV-based communication system. Therefore, for ground-to-air communications, each device will typically have a
LoS view towards a specific UAV with a given probability. This LoS probability depends on the
environment, location of the device and the UAV as well as the elevation angle [8]. One suitable
expression for the LoS probability is given by [1], [6], [8]:
ij
PLoS
=
1
,
1 + ψ exp(−β [θij − ψ])
(1)
where ψ and β are constant values which depend on the carrier frequency and type of environment
−1 hj
180
,
such as rural, urban, or dense urban, and θij is the elevation angle. Clearly, θ = π ×sin
dij
q
uav 2
2
2
where dij = (xi − xuav
j ) + (yi − yj ) + hj is the distance between device i and UAV j.
From (1), we can see that by increasing the elevation angle or increasing the UAV altitude,
the LoS probability increases. The path loss model for LoS and non-line-of-sight (NLoS) links
between device i and UAV j is given by [6] and [8]:
α
η1 4πfc dij ,
c α
Lij =
η2 4πfc dij ,
c
LoS link,
(2)
NLoS link,
where fc is the carrier frequency, α is the path loss exponent, η1 and η2 (η2 > η1 > 1) are the
excessive path loss coefficients in LoS and NLoS cases, and c is the speed of light. Note that,
ij
ij
the NLoS probability is PNLoS
= 1 − PLoS
. Typically, given only the locations of the UAVs and
devices, it is not possible to exactly determine which path loss type (LoS/NLoS) is experienced
by the device-UAV link. In this case, the path loss average considering both LoS and NLoS
links can be used for the device-UAV communications [6] and [8]. Now, using (1) and (2), the
average path loss between device i and UAV j can be expressed as:
α
α
ij
4πfc dij
4πfc dij
ij
ij
ij
L̄ij = PLoS η1
+ PNLoS η2
= PLoS
η1 + PNLoS
η2 (Ko dij )α ,
c
c
(3)
7
where Ko =
ḡij =
1
.
L̄ij
4πfc
.
c
Clearly, the average channel gain between the UAV and the device is
Note that, by using the average channel gain, there is no need to account for LoS
and NLoS links separately, and, hence, the SINR expressions become more tractable. Therefore,
we use the average channel gain to model the interference and desired links for all device-UAV
communications while computing the SINRs.
B. IoT Device Activation Model
Indeed, the activation of IoT devices depends on the services that they are supporting. For
instance, in some applications such as weather monitoring, smart grids, or home automation, the
IoT devices need to report their data periodically. However, the IoT devices can have random
activations in health monitoring, or smart traffic control applications. Therefore, the UAVs must
be properly deployed to collect the IoT devices data while dynamically adapting to the activity
patterns of these devices. Naturally, the optimal locations of the UAVs and their update times
depend on the activation process of the IoT devices. Here, we consider two activation models. In
the first model, the IoT devices are randomly activated, as in smart traffic control applications.
In this case, the concurrent transmissions of a massive number of devices within a short time
duration can lead to a bursty traffic as pointed out in [17] and [18]. In fact, when massive IoT
devices attempt to transmit within a short time period, the arrival patterns become more bursty
[19]. Thus, 3GPP suggests a beta distribution to capture this traffic characteristic of IoT devices
[20]. In this case, each IoT device will be active at time t ∈ [0, T ] following the beta distribution
with parameters κ and ω [18]–[20]:
f (t) =
tκ−1 (T − t)ω−1
,
T κ+ω−1 B(κ, ω)
(4)
where [0, T ] is the time interval within which the IoT devices can be active, and B(κ, ω) =
R 1 κ−1
t (1 − t)ω−1 dt is the beta function with parameters κ and ω [21].
0
In addition, IoT devices such as smart meters typically report their data periodically rather
than randomly. For such devices, the activation process is deterministic and assumed to be known
in advance. In such case, we assume that device i becomes active each τi seconds during [0, T ]
j k
time duration. Clearly, the number of activations for a device i during [0, T ] is τTi .
C. Channel Assignment Strategy
Here, given only the devices’ locations, a practical channel assignment approach is to assign
different channels to devices which are located in proximity of each other. This approach
8
significantly mitigates the possibility of having strong interference between two closely located
devices. For the channel assignment problem, we have adopted a constrained K-mean clustering
strategy [22], which is an efficient distance-based clustering approach in which a set of given
points are grouped into K clusters based on their proximity. In this case, given the number of
active devices, Ln , and the number of orthogonal channels, R ≤ Ln , we group the devices based
on proximity, and assign different channels to devices which are in the same group.
Now, we present our optimization problem to find the UAVs’ locations, device association,
and transmit power of IoT devices at each update time tn during [0, T ]:
(OP):
min
v j ,c,P
s.t. P
Ln
X
Pi ,
∀i ∈ Ln , ∀j ∈ K,
(5)
i=1
Pi ḡici (v ci )
≥ γ,
Pk ḡkci (v ci ) + σ 2
(6)
k∈Zi
0 < Pi ≤ Pmax ,
(7)
where Ln is the total number of active devices at update time tn , and Ln is the set of devices’
index. P is the transmit power vector with each element Pi being the transmit power of device
i. Also, v j is the 3D location of UAV j, and c is the device association vector with each element
ci being the index of the UAV that is assigned to device i. Pmax is the maximum transmit power
of each IoT device, and σ 2 is the noise power. Furthermore, ḡici (v ci ) is the average channel
gain between device i and UAV ci which is a function of the UAV’s location. Also, ḡkci (v ci ) is
the average channel gain between interfering device k and UAV ci . In (6), Zi is the set of all
other devices that use the same channel as device i and create interference. γ is the SINR target
which must be achieved by all the devices, (6) represents the SINR requirement, and (7) shows
the maximum transmit power constraint. Hereinafter, we call (OP) the original problem.
Note that, in (5), the transmit power of the IoT devices, the 3D locations of the UAVs, and the
UAV-device associations are unknowns. Clearly, the locations of the UAVs impact the channel
gain between the devices and UAVs, and, hence, they affect the transmit power of each device,
Pi . Furthermore, given (6), due to the mutual interference between the devices, the transmit
power of each device depends also on the transmit power of the interfering devices as well as
the device-UAV associations. In addition, the device-UAV associations depend on the UAVs’
locations which are also unknowns. Therefore, there is a mutual dependency between all the
9
Fig. 2: Block diagram for the proposed solution.
optimization variables in (OP). Moreover, considering (1) and constraint (6), we can see that,
this optimization problem is highly non-linear and non-convex. Indeed, solving (5) is significantly
challenging due to the mutual dependency of the optimization variables, non-linearity, and nonconvexity of the problem. Next, we propose a framework for solving this optimization problem.
In essence, our proposed framework for solving (OP) proceeds as follows. At each update
time tn , given the fixed UAVs’ locations, we find the optimal device-UAV association and the
transmit power of the devices. Next, given the fixed UAV association from the previous step,
we determine the sub-optimal locations of the UAVs and update the transmit power the devices
accordingly. This procedure is done iteratively until the 3D UAVs’ locations, device association,
and the transmit power of the devices are found. Clearly, at each step, the total transmit power
of the devices decreases, and, hence, the proposed algorithm converges. Fig. 2 shows a block
diagram that summarizes the main steps for solving (OP). Next, we discuss, in detail, each block
of the proposed solution in Fig. 2.
III. UAV D EPLOYMENT AND D EVICE A SSOCIATION WITH P OWER CONTROL
Here, given the locations of active IoT devices, we minimize the total transmit power of
the devices by solving (5). Clearly, the UAVs’ locations and the device association are mutually
dependent. In particular, to find the device association, the locations of the UAVs must be known.
Moreover, the UAVs’ locations cannot be optimized without knowing the device association.
Therefore, we decompose (OP) into two subproblems that will be solved iteratively. In the
first subproblem, given the locations of the UAVs, we find the optimal device association and
the transmit power of the devices such that the uplink SINR requirements of all active devices
are satisfied with a minimum total transmit power. In the second subproblem, given the device
association resulting from the first subproblem, we determine the sub-optimal locations of the
10
UAVs for which the transmit power of the devices is minimized. Note that, this is an iterative
process in which the results of each subproblem are used in the other subproblem for the next
iteration. These computations are performed by the control center until the 3D UAVs’ locations,
device association, and transmit power of the devices are obtained.
Note that, given the limited number of available orthogonal channels, the interference between
the devices will depend on the number of active devices at each update time. Clearly, there is no
interference when the number of active devices at time tn is less than the number of orthogonal
channels, or equivalently Ln ≤ R. Given that, in the interference-free scenario, one can provide
a more tractable analysis, here, for the deployment and association steps. Therefore, we will
investigate the interference and interference-free scenarios, separately.
A. Device Association and Power Control
Here, given initial locations of the UAVs, we aim to find the optimal device association as well
as the transmit power of each IoT device such that the total transmit power used for successful
uplink communications is minimized.
1) Interference scenario: In the presence of uplink interference when Ln > R, the power
minimization problem at update time tn will be given by:
(P1-a):
min
c,P
s.t. P
Ln
X
Pi ,
∀i ∈ Ln , ∀j ∈ K,
(8)
i=1
Pi ḡici
≥ γ,
Pk ḡkci + σ 2
(9)
k∈Zi
0 < Pi ≤ Pmax .
(10)
To solve (P1-a), we need to jointly find the optimal device association and the transmit power
of all active devices under the SINR constraints for the given UAVs’ locations. Clearly, given the
fixed UAVs’ locations, optimization variables are the device association and the transmit power
of the devices. Note that, satisfying the SINR requirement of each device significantly depends
on the distance and altitude of its serving UAV. Therefore, the feasibility of the optimization
problem in (8) depends on the locations of the UAVs. Next, we derive an upper bound and a
lower bound for the altitude of serving UAV j as a function of its distance from device i.
11
Proposition 1. The lower and upper bounds for the altitude of a UAV j needed to serve a
device i (meeting its SINR requirement), are given by:
1/α
ψQ
Pmax
1
ln
+ ψ ≤ hj ≤
,
dij sin
β
1−Q
γKoα σ 2 η1
where dij is the distance between UAV j and device i, and Q =
Pmax
α 2
γdα
ij Ko σ (η1 −η2 )
(11)
−
η2
.
η1 −η2
Proof: Let Ii be the cumulative interference from interfering devices on device i, then:
Pi ḡij
≥ γ,
Ii + σ 2
Pi
Pmax
≤
,
dαij ≤
ij
ij
ij
ij
α
2
α
2
γKo (Ii + σ ) η1 PLoS + η2 PNLoS
γKo σ η1 PLoS + η2 (1 − PLoS
)
Pmax
η2
ij
PLoS
≥ α α 2
−
,
γdij Ko σ (η1 − η2 ) η1 − η2
SINRi =
considering Q =
(a)
− η2 )
−
η2
, and using equation (1),
η1 − η2
ψQ
+ ψ,
1−Q
1
ψQ
hj ≥ dij sin
ln
+ψ ,
β
1−Q
θij ≥
1
ln
β
Pmax
α
α
γdij Ko σ 2 (η1
(12)
where (a) stems from (1). Also, we have:
dαij ≤
γKoα σ 2
(b)
Pmax
Pmax
≤
,
ij
ij
γKoα σ 2 η1
η1 PLoS + η2 (1 − PLoS )
where in (b) we consider PLoS = 1 which is equivalent to hj = dij . Finally,
1/α
Pmax
hj ≤
.
γKoα σ 2 η1
(13)
(14)
Clearly, (12) and (14) prove the proposition.
Proposition 1 provides the necessary conditions for the UAV’s altitude needed in order to be
able to serve the IoT device. From (11), the minimum altitude must increase as the distance
increases. In other words, the UAV’s altitude needs to be adjusted based on the distance such
ψQ
that the elevation angle between the device and the UAV exceeds β1 ln 1−Q
+ ψ. Furthermore,
as expected, the maximum altitude of the UAVs significantly depends on the maximum transmit
power of the devices as given in (14).
Now, given the fixed UAVs’ locations, problem (P1-a) corresponds to the problem of joint
user association and uplink power control in the terrestrial base station scenario. The algorithm
12
presented in [23] and [24] leads to the global optimal solution to the joint user association and
uplink power control under the SINR and maximum transmit power constraints. As a result, the
optimal transmit power of users and the base station association for which the total uplink transmit
power is globally minimized, is determined. In problem (P1-a), the IoT devices correspond to
the users, and fixed positioned UAVs correspond to the terrestrial base stations. For our case,
this algorithm, as given in Algorithm 1, will proceed as follows. We start with an initial value
(t)
for transmit power of all active devices in step 3. Then, in step 4 we compute ρij at iteration t.
(t)
In this case, ρij represents the minimum required transmit power of device i to reach an SINR
of 1 while connecting to UAV j, given the fixed transmit power of other devices. In step 5, we
find the minimum transmit power of device i if it connects to the best UAV. Then, the index of
the best UAV which is assigned to device i is given in step 6. In step 7 we update the transmit
power of device i in order to achieve an SINR of γ. Steps 4 to 7 must be repeated for all devices
to obtain the optimal transmit power and the device association vectors.
Algorithm 1 Iterative algorithm for joint power control and device-UAV association
1: Inputs: Locations of UAVs and IoT devices
2: Outputs: Device association vector (c), and transmit power of all devices (P ).
(0)
(0)
3: Set t = 0, and initialize P (0) = P1 , ..., PK
σ2 +
(t)
k∈Zi
4: Define ρij =
5:
6:
7:
P
.
(t)
Pk ḡkj
ḡij
.
(t)
Compute Si (P (t) ) = min ρij .
j∈K
(t)
Find ci (P (t) ) = arg min ρij .
j∈Kn
o
(t+1)
Update Pi
= min γSi (P (t) ), Pmax ,
8: Repeat steps 4 to 7 for all devices until P
∀i ∈ Ln .
(t)
converges.
h
i
9: P = P (t) , c = ci (P (t) ) , ∀i ∈ Ln .
As shown in [23], after several iterations this algorithm quickly converges to the global optimal
solution if the SINR of each device is equal to γ. Hence, by solving (P1-a), we are able to find
the optimal transmit power of the devices and the device association for any given fixed locations
of the UAVs. Then, the device association and transmit power of the devices will be used as
inputs for solving the second subproblem in which the UAVs’ locations need to be optimized
(in Subsection III-B).
13
2) Interference-free scenario: At each update time tn , if the number of active devices is lower
than the number of orthogonal channels or equivalently Ln ≤ R, there will be no interference
between the devices. Unlike in the interference scenario, here, the transmit power of each device
can be computed only based on the channel gain between the device and its serving UAV.
Therefore, considering (3), and (6) without interference, the minimum transmit power of device
i in order to connect to UAV j is Pi = γσ 2 L̄ij . In this case, given the locations of the UAVs
(fixed for all v j ), L̄ij is known for all devices and problem (P1-a) can be simplified. Hence, the
optimal association problem under minimum power in the interference-free scenario will be:
(P1-b):
K X
Ln
X
min
Aij L̄ij ,
(15)
Aij
s.t.
K
X
j=1 i=1
Aij = 1, ∀i ∈ Ln ,
(16)
j=1
Aij L̄ij ≤
Pmax
, Aij ∈ {0, 1}, ∀i ∈ Ln , j ∈ K,
γσ 2
(17)
where L̄ij is the average path loss between device i and UAV j, which is known, give the locations
of the UAV and the device. Aij is equal to 1 if device i is assigned to UAV j, otherwise Aij
will be equal to 0. Clearly, the optimization problem in (15) is an integer linear programming
(ILP). In general, this problem can be solved by using standard ILP solution methods such as the
cutting plane. However, these solutions might not be efficient as the size of the problem grows.
In particular, due to the potentially high number of IoT devices, a more efficient technique for
solving (15) is needed. Here, we transform problem (15) to a standard assignment problem [25]
which can be solved in polynomial time. In the assignment problem, the objective is to find the
optimal one-to-one assignment between two sets of nodes with a minimum cost. In our problem,
the devices and the UAVs can be considered as two sets of nodes that need to be assigned to each
other with an assignment cost of Lij between nodes i and j. However, compared to the classical
assignment problem, (P1-b) has an additional constraint in (17) which results from the transmit
maximum power constraint. This constraint indicates that device i cannot be assigned to UAV j if
L̄ij >
Pmax
.
γσ 2
Therefore, in the assignment problem we can consider Lij = +∞ to avoid assigning
device i to UAV j when L̄ij >
Pmax
γσ 2
that implies the constraint in (17) is violated. Subsequently,
using the updated assignment costs, Lij , problem (P1-b) will be transformed to the classical
assignment problem which can be solved using the Hungarian method with a time complexity
14
of O((Ln K)3 ) [26]. We note that, in absence of interference, problems (P1-a) and (P1-b) have
the same solution. Next, we present the second subproblem of the original optimization problem
(5) in order to optimize the UAVs’ locations.
B. Optimal Locations of the UAVs
In this section, given the optimal device association, our goal is to find the sub-optimal
locations of the UAVs for which the total transmit power of the devices is minimized. In other
words, considering the mobile nature of the UAVs, we intelligently update the location of each
UAV based on the location of its associated IoT devices.
1) Interference scenario: In this scenario, given the UAV-device associations, the optimization
problem to find the 3D locations of the UAVs and the transmit power of the devices will be:
(P2-a):
min
v j ,P
s.t. P
Ln
X
Pi ,
∀i ∈ Ln , ∀j ∈ K,
(18)
i=1
Pi ḡij (v j )
≥ γ,
Pk ḡkj (v j ) + σ 2
(19)
k∈Zi
0 < Pi ≤ Pmax ,
(20)
uav
where vj = (xuav
j , yj , hj ) indicates the 3D location of UAV j. Clearly, the channel gains used
in (19) depend on the locations of the UAVs. Note that, according to (1) and (3), ḡij (v j ) is
a non-convex function of v j . Consequently, constraint (19) is also non-linear and non-convex.
Furthermore, the transmit power of the devices and the UAVs’ locations are mutually dependent.
On the one hand, the location of each UAV must be determined such that its associated devices
can connect to it with a minimum transmit power. On the other hand, the UAV’s location
will impact the amount of interference received from other interfering devices. Indeed, solving
the optimization problem in (P2-a) is challenging as the problem is highly non-linear and nonconvex. In particular, the complexity of this problem stems from the mutual dependence between
the transmit power of the devices and the locations of the UAVs.
Our proposed approach to solve (P2-a) is based on optimizing the location of each UAV
separately. Note that, using the results of (P1-a), for each UAV, the associated and non-associated
devices and their transmit power, Pi∗ , are known. Our proposed solution proceeds as follows.
The cloud starts by considering a single UAV and then optimizing its location given the fixed
15
transmit power for the non-associated devices. Then, the cloud updates the transmit power of
the associated devices according to the new location of their serving UAV. Hence, at each step,
the location of a UAV and the transmit power of its associated devices are updated. At each
iteration, after finding Pi∗ , we set Pmax = Pi∗ for the next iteration. This ensures that the transmit
power of the devices does not increase during the iterative process. The entire process is repeated
by the cloud for all UAVs one-by-one, until the transmit power of the devices cannot be further
reduced by changing the UAVs’ locations. Note that, at each step, one must determine the optimal
location of each UAV such that the total transmit power of its associated devices is minimized.
Now, let Cj be the set of devices’ index associated to UAV j. Given (3), (18), and (19), the
optimal location of UAV j can be determined by solving the following problem:
X
min
Fi (v j ),
vj
(21)
i∈Cj
ij
ij
s.t. Fi (v j ) = γ η1 PLoS
+ η2 PNLoS
(Ko dij )α
k∈Zi
Fi (v j ) ≤ Pi∗ , ∀i ∈ Cj ,
Pk
X
kj
η1 PLoS
+
kj
η2 PNLoS
α
(Ko dkj )
+ σ 2 , (22)
(23)
ij
kj
Note that, PLoS
, PLoS
, dkj , and dij depend on the locations of UAVs (vj ). Also, (23) guarantees
that the transmit power of each device is reduced by updating the location of serving UAV.
Clearly, (21), (22), and (23) are non-linear and non-convex. Considering the fact that the
objective function and constraints are twice differentiable, we convert (21) to a quadratic form
which can be solved using efficient techniques. In particular, we adopt the sequential quadratic
programming (SQP) method as one of the most powerful algorithms for solving large scale and
constrained differentiable non-linear optimization problems [27]. Clearly, considering the high
non-linearity of (22) as well as the large number of constraints, the SQP is a suitable method for
solving our optimization problem. In the SQP method, the objective function is approximated by
a quadratic function, and the constraints are linearized. Subsequently, the optimization problem
is solved by solving multiple quadratic subproblems. In our optimization problem, to find the
optimal location of UAV j, v j,k , we start with an initial point v j,k (starting with k = 0). Then,
we use the first order necessary optimality or Karush-Kuhn-Tucker (KKT) conditions to find the
16
Lagrangian variables. In particular, we use:
X
∇L(v j,k , λk ) =
∇Fi (v j,k ) + ∇wi (v j,k )λk = 0,
(24)
i∈Cj
where L(v j,k , λk ) =
P
Fi (v j,k ) + λT w(v j,k ) is the Lagrangian function, λk is the vector of
i∈Cj
Lagrangian variables, and w(v j,k ) is a vector of functions with each element being wi (v j,k ) =
(Fi (v j,k ) − Pi∗ ). Then, given v j,k , we determine the Lagrange variables by [27]:
X
−1
∇Fi (v j,k ).
λk = − wi (v j,k )T ∇wi (v j,k )
∇wi (v j,k )T
(25)
i∈Cj
In the next step, we update v j,k+1 = v j,k + dk , where dk is the solution to the following
quadratic programming problem:
dk = arg min
d
X
i∈Cj
1
∇Fi (v j,k )T d + dT ∇2 [L(v j,k , λk )] d,
2
s.t. Fi (v j,k ) + ∇Fi (v j,k )T d − Pi∗ ≤ 0, ∀i ∈ Cj ,
(26)
(27)
where, ∇ and ∇2 indicate the gradient and Hessian operations. Clearly, (26) is an inequality constrained quadratic programming. Moreover, it can be shown that the Hessian matrix,
∇2 [L(v, λk )], is not positive semidefinite, and, hence, (26) is non-convex in general. In this
case, the two possible solution approaches are the active set, and the interior point methods.
Typically, the active set method is preferred when the Hessian matrix is moderate/small and
dense. The interior point, however, is a suitable approach when the Hessian matrix is large and
sparse [28]. In our problem, due to the potential possible high number of active devices, the
number of constraints can be high. Therefore, the Hessian matrix, ∇2 [L(v j,k , λk )], is large and
sparse, and, hence, the interior point method is used.
Finally, based on (24)-(27), the sub-optimal location of each UAV (v j ), given the fixed
device association, will be determined. Next, we address the UAVs’ location optimization in
an interference-free scenario.
2) Interference-free scenario: In the absence of interference, we are able to provide tractable
analysis on the UAVs’ locations optimization. Considering α = 2 for LoS ground-to-air propagation [8], the optimal locations of the UAVs will be given by:
17
3.52
Error (%)
3.5
3.48
3.46
3.44
3.42
50
100 150 200 250 300 350 400 450 500
UAV Altitude (m)
Fig. 3: Error in the objective function approximation.
(P2-b):
min
vj
X
2
ij
ij
Ko2 σ 2 γ η1 PLoS
+ η2 PNLoS
dij ,
(28)
i∈Cj
2
Pmax
ij
ij
s.t. η1 PLoS
+ η2 PNLoS
dij ≤ 2 2 , ∀i ∈ Cj .
Ko σ γ
(29)
uav
This optimization problem is non-convex over vj = (xuav
j , yj , hj ). However, given any altitude
hj , we can provide a tractable solution to this problem. First, given hj , we consider the following
function that is used in (28):
2
ij
ij
q(dij ) = Ko2 σ 2 γ η1 PLoS
+ η2 PNLoS
dij .
(30)
ij
ij
ij
Clearly, considering the fact that 0 ≤ PLoS
≤ 1, and PNLoS
= 1 − PLoS
, we have:
Ko2 σ 2 γη1 d2ij ≤ q(dij ) ≤ Ko2 σ 2 γη2 d2ij .
(31)
From (31), we can see that q(dij ) is bounded between two quadratic functions that each is linearly
proportional to d2ij . Now, using the least square estimation method, we find the coefficients α1 and
α2 such that, given any hj , q(dij ) is approximated by the following convex quadratic function:
q(dij ) ≈ α1 d2ij + α2 .
(32)
where α1 and α2 are altitude dependent coefficients. Note that, using the quadratic approximation,
the solution of (28) becomes more tractable.
Fig. 3 shows the error in the objective function (28) due the quadratic approximation. As we
can see from Fig. 3 which is obtained based on the parameters in Table I, the error is less than
4% for different UAVs’ altitudes.
2
ij
ij
Now, in constraint (29), we consider D = η1 PLoS
+ η2 PNLoS
dij . Clearly, D is an increasing
function of dij since η1 −η2 < 0, and for a fixed altitude, LoS probability is a decreasing function
2
uav 2
2
of distance. Therefore, using dij 2 = (xi − xuav
j ) + (yi − yj ) + hj , and (32), for any given hj
we can write the optimization problem (28) as:
X
2
2
2
uav
min
(xuav
j − xi ) + (yj − yi ) + hj ,
uav uav
xj ,yj
i∈Cj
(33)
18
2
uav
2
2
2
s.t. (xuav
j − xi ) + (yj − yi ) + hj − ≤ 0, ∀ i ∈ Cj , and j ∈ K.
(34)
where ε = {d|Ko2 σ 2 γ (η1 PLoS + η2 PNLoS ) d2 = Pmax }. Next, we derive the solution to problem
(33) that corresponds to finding the sub-optimal UAVs’ locations.
∗ uav ∗
, yj ) = −P (λ)−1 Q(λ), with the
Theorem 1. The solution to (33) is given by s∗ = (xuav
j
vector λ that maximizes the following concave function:
max
λ
1
Q(λ)T P (λ)−1 Q(λ) + r(λ),
2
(35)
s.t. λ ≥ 0,
where P (λ) = P o +
|C
j|
P
λi Pi , Q(λ) = Qo +
(36)
|C
j|
P
λi Qi and r(λ) = ro +
i=1
i=1
|C
j|
P
λi ri , with P o , Qo ,
i=1
ro , P i , Qi , and ri given in the proof.
Proof: As we can see from (33), the optimization problem is a quadratically constrained
quadratic program (QCQP) whose general form is given by [29]:
min
s
s.t.
1 T
s P o s + QTo s + ro ,
2
1 T
s P i s + QTi s + ri , i ∈ Cj .
2
(37)
(38)
Given (33) and (34), we have:
T
h
iT
|C
|C
j|
j|
2|Cj |
0
2 0
P
P
, Qi = −2xi −2yi .
Po =
, Pi =
, Qo = −2 xi −2 yi
i=1
i=1
0
2|Cj |
0 2
|C
|
|C
|
j
j
P
P
Also, ro =
x2i + yi2 , and ri = x2i + yi2 + h2j − 2 with ε = {d|Ko2 σ 2 γ (η1 PLoS + η2 PNLoS ) d2 =
i=1
i=1
Pmax }. Note that, P o and P i are positive semidefinite matrices, and, hence, the QCQP problem
in (37) is convex. Now, we write the Lagrange dual function as:
X 1
1 T
T
T
T
f (λ) = inf s P o s + Qo s + ro +
λi
s P i s + Q i s + ri
s 2
2
i
1 T
T
= inf s P (λ)s + Q(λ) s + r(λ) .
s
2
(39)
Clearly, by taking the gradient of the function inside the infimum with respect to s, we find
s∗ = −P (λ)−1 Q(λ). As a result, using s∗ , f (λ) = 21 Q(λ)T P (λ)−1 Q(λ) + r(λ). Finally, the
dual of problem (37) or (33) will be:
max f (λ), s.t. λ ≥ 0,
(40)
19
which proves Theorem 1.
Using Theorem 1, for a fixed altitude, we find the optimal 2D coordinates of the UAV, s∗ =
∗ uav ∗
(xuav
, yj ). Then, the optimal UAV’s altitude is the argument that minimizes the following
j
one-dimensional function as:
h
i
h∗j = arg min α1 h2j + ks∗ k2 + α2 .
(41)
hj
where α1 and α2 are the altitude dependent coefficients given in (32). Note that, given (41), the
sub-optimal altitude of the UAV is obtained via one dimensional search over a feasible range of
altitudes. Consequently, we can determine the optimal 3D location of each UAV.
Note that, the device association (presented in Section III), and UAVs’ locations optimization
(in Section IV) are applied iteratively until there is no change in the location update step. Clearly,
at each iteration, the total transmit power of the devices is reduced and the objective function
is monotonically decreasing. Hence, the solution converges after several iterations. Note that,
our proposed approach provides a suboptimal solution to the original problem. Nevertheless, our
solution has a reasonable accuracy but significantly fast compared to the global optimal solution
that can be achieved by the brute-force search, as will be further corroborated in the simulations.
Thus far, we considered the IoT network at one snapshot in the time duration [0, T ]. Next, we
analyze the IoT network considering the entire time duration [0, T ] in which the set of active
devices changes. In this case, to maintain the power-efficient and reliable uplink communications
of the devices, the UAVs must update their locations at different update times tn .
IV. U PDATE T IMES AND M OBILITY OF UAV S
Here, we find the optimal update time and trajectory of the UAVs to guarantee the reliable
uplink transmissions of the IoT devices. Clearly, the trajectory of the UAVs, as well their update
time depend on the activation process of the IoT devices. Furthermore, to move along the optimal
trajectories, the UAVs must spend a minimum total energy on mobility so as to remain operational
for a longer time. In the considered ground IoT network, the set of active IoT devices changes
over time. Consequently, the UAVs must frequently update their locations accordingly. Note that,
the UAVs do not continuously move as they must stop, serve the devices, and then update their
locations. Furthermore, the mobility of the UAVs is also limited due to their energy constraints.
Hence, the UAVs update their locations only at some specific times. In this case, during time
interval [0, T ], we need to find update times tn , 1 ≤ n ≤ N with N updates, and a framework for
20
optimizing the mobility of the UAVs at different update times. Here, for tractability, we assume
that the devices are synchronized at t = 0. In this case, the synchronization process needs to be
done only once during the entire activation period [0, T ]. Note that, our optimization problems
for jointly finding the optimal UAVs’ locations, the device association, and devices’ transmit
power at each update time do not depend on this synchronization assumption.
A. Update Time Analysis
First, we propose a framework to find the update times of the UAVs. As discussed in Section II,
each UAV’s trajectory consists of multiple stop locations (determined in update times) at which
each UAV serves its associated ground devices. Clearly, the update times depend on the activation
of the IoT devices during the given time period [0, T ]. Indeed, the number of update times, N ,
impacts the optimal location and trajectory of the UAVs as well as the power consumption of the
IoT devices. A higher number of updates leads to a shorter time interval between the consecutive
updates. Hence, a lower number of devices will be active during the shorter time interval. In
such a case, the active devices experience lower interference from each other while transmitting
their data to the UAVs. Therefore, the IoT devices can use lower transmit power to meet their
SINR constraint. However, a higher number of updates requires more mobility and higher energy
consumption for the UAVs. Next, we provide insightful analysis on the update time based on
the probabilistic and periodic activation models of the IoT devices.
1) Periodic IoT activation: In some applications such as weather monitoring, smart grids
(e.g. smart meters), and home automation, the IoT devices can report their data periodically.
Therefore, the devices are activated periodically. Let τi be the activation period of device i
during [0, T ]. Without loss of generality, assume τ1 ≤ τ2 ≤ ... ≤ τL . Due to the periodic nature
of devices’ activation, we can find the exact number of active devices at each update time tn .
Proposition 2. The exact number of active IoT devices at update time tn is given by:
−
L
X
tn−1
tn
>
, n > 1,
bn =
1
τ
τ
i
i
i=1
b1 = arg max {t1 > τi } ,
(42)
(43)
i
where 1(.) is the indicator function which can only be equal to 1 or 0, and tn − = lim+ (tn − ε).
ε→0
21
Proof: User i becomes active during [tn−1 , tn ) if there exists q ∈ N such that tn−1 ≤ qτi < tn .
Thus, the number of activations of device i before tn must be greater than the one until tn−1 .
j −k
j
k
Considering the fact that the number of activations before tn is tnτi , and until tn−1 is tn−1
,
τi
we must have:
tn −
τi
tn−1
>
.
τi
(44)
Hence, the total number of active devices which need to be served at tn is equal to:
−
L
X
tn
tn−1
bn =
1
>
.
τi
τi
i=1
(45)
Finally, considering t0 = 0, we can write b1 as: b1 = arg max {t1 > τi }.
i
Proposition 2 gives the exact number of devices that must be served by UAVs at each update
time. In this case, the update times can be adjusted according to the number of devices that
can be served by the UAVs. Indeed, knowing the exact number of active devices enables us to
determine the update times in a deterministic and efficient way based on system requirements.
2) Probabilistic IoT activation: Certain IoT devices can have probabilistic activations in
applications such as health monitoring, and smart traffic control. In this case, each IoT device
becomes active at time t ∈ [0, T ] following the beta distribution as given in (4). For this scenario,
we will next derive the specific update times as a function of the average number of active devices.
Theorem 2. The update times during which, on the average, a total of an devices must be
served by the UAVs, are given by:
a
+ I tn−1 (κ, ω) , κ, ω , n > 1,
T
L
a
1
t1 = T × I −1
, κ, ω ,
L
tn = T × I −1
n
(46)
(47)
where Ix (.) is the regularized incomplete beta function and I −1 (.) is its inverse function. L is the
total number of IoT devices, and [0, T ] is the time interval during which the devices can be active.
Proof: First, we find the probability that each device becomes active in order to send its
data to a UAV at update time tn . As discussed in the system model, a device needs to transmit at
time tn if it becomes active during time [tn−1 , tn ). Thus, the probability that each device needs
to be served at tn is:
Z
tn
pn =
tn−1
tκ−1 (T − t)ω−1
dt =
T κ+ω−1 B(κ, ω)
Z
tn
T
tn−1
T
tκ−1 (1 − t)ω−1
dt,
B(κ, ω)
22
B tn (κ, ω) − B tn−1 (κ, ω)
=
where Bx (κ, ω) =
Rx
0
T
= I tn (κ, ω) − I tn−1 (κ, ω),
T
B(κ, ω)
T
(48)
T
y κ−1 (1 − y)ω−1 dy is the incomplete beta function with parameters κ and
ω, and Ix (.) is the regularized incomplete beta function.
Now, the average number of active devices at tn is given by:
L
L
X
X
L
(L − 1)!
L−k
k
an =
pn (1 − pn )
= Lpn
pn k−1 (1 − pn )L−k ,
k
(k − 1)! (L − k)!
k=1
k=1
0
L
X
(L0 )!
L0 −k0
k0 −1
=
p
(1
−
p
)
= Lpn ,
n
n
(k 0 )! (L0 − k 0 )!
k0 =0
(49)
where in (a), we used L0 = L − 1 and k 0 = k − 1. Note that, (49) corresponds to the mean of a
binomial distribution. Then we have:
h
i
L I tn (κ, ω) − I tn−1 (κ, ω) = an .
T
(50)
T
This leads to:
tn = T × I
−1
a
n
L
+I
tn−1
T
Finally, considering I0 (.) = 0, we find t1 = T × I −1
(κ, ω) , κ, ω .
a1
, κ, ω
L
(51)
.
Clearly, the update times need to be determined based on the IoT devices’ activation patterns.
In fact, tn depends on the number of IoT devices, and their activation distribution. Furthermore,
according to (46), each tn depends also on the previous update time, tn−1 . This is due to the
fact that, the number of active devices that need to be served at tn , depends on the update time
difference tn − tn−1 . Using Theorem 2, the update times of the UAVs can be set based on the
average number of active devices. Typically, at each update time, the number of devices which
need to be served by the UAVs should not be high in order to avoid high interference. However,
considering the number of available resources (orthogonal channels and UAVs), it is preferable to
serve a maximum number of active devices at each update time. Hence, in this case, the number
of active devices at each update time must not be relatively low.Therefore, considering system
requirements and different parameters such as mutual interference between devices, acceptable
delay for serving the devices, and number of available channels, an appropriate tn must be
adopted. For instance, using Theorem 2, the update times can be set such that the average number
of active devices be lower than the number of channels, R, to avoid interference between the
devices. Next, we investigate the UAVs’ mobility during the update times.
23
B. UAVs’ Mobility
Thus far, we have determined the update times as well as the stop locations at each update
time. Here, we investigate how the UAVs should move between the stop locations at different
update times. In this case, considering the energy limitation of the UAVs, Emax , we find the
optimal trajectory of each UAV to guarantee reliable and energy-efficient uplink transmissions of
active IoT devices. The UAVs update their locations according to the activity of the IoT devices.
Therefore, the UAVs move from their initial locations at tn−1 to a new optimal locations at tn .
This mobility should be done in such a way that the UAVs spend a minimum total energy on the
mobility so as to remain operational for a longer time. In fact, given the optimal sets of UAVs’
locations at tn−1 and tn obtained from Section III, we determine how to move the UAVs between
the initial and the new sets of locations in order to minimize total mobility of the UAVs.
Now, let In−1 and In be two sets comprising the UAVs’ locations at two consecutive update
times tn−1 and tn . Our goal is to find the optimal mapping between these two sets in a way that
the energy used for transportations (between two sets) is minimized. Not that, in our model, the
total energy which each UAV can use for the mobility during [0, T ] is limited to Emax . Clearly, in
the multiple updates (mobilities) during [0, T ], the maximum energy consumption of each UAV
at each update is equal to the remaining energy of the UAV. Let Γn,k be the remaining energy
of the UAV at the location having index k ∈ In−1 at time tn . Then, we can write the following
UAVs’ mobility optimization problem:
X X
min
Ekl Zkl ,
Z
s.t.
X
l∈In
(52)
l∈In k∈In−1
Zkl = 1,
X
Zkl = 1,
(53)
k∈In−1
Elk ≤ Γn,k , Zkl ∈ {0, 1}, ∀k ∈ In−1 , ∀l ∈ In ,
(54)
where In−1 and In , are the initial and new sets of UAVs’ locations at times tn−1 and tn . Z
is the |In | × |In | assignment matrix with each element Zkl being 1 if UAV k is assigned to
location l, and 0 otherwise. Ekl is the energy used for moving a UAV from its initial location
with index k ∈ In−1 to a new location with index l ∈ In . Also, Γn,k is the remaining energy
for the UAVs at time tn . Note that, (54) guarantees that UAVs remain operational until the end
of the period T . The total energy consumption of the rotary wing UAV while moving between
24
two stop locations can be computed as done in [30]:
E=
D
(PV + PH ) ,
v
(55)
where D is the distance between two stop locations, D/v is the flight duration, PV is the power
consumption for vertical movement, and PH is the power consumption for horizontal movement.
Clearly, if the altitude difference between two stop locations is ∆h, the effective vertical and
. According to
horizontal velocities will be vv = v sin φ and vh = v cos φ, with φ = sin−1 ∆h
D
[31] and [32], PH is composed of parasitic power and induced power needed for overcoming the
parasitic drag and the lift-induced drag. The parasitic power, based on [32, equations (13.32),
(13.27), and (11.3)], can be given by:
v 2
π
1
h
3
3 4
PP = ρCDo Ae vh + Nb cb ρCDo ω R 1 + 3
,
2
4
ωR
(56)
where vh is the effective horizontal velocity, CDo is the drag coefficient, ρ is the air density, cb
is the blade chord, Nb is the number of blades, and Ae is the reference area (frontal area of the
UAV) [31] and [32]. We note that the second term in (56) represents the blade power profile.
Using [32, equations (13.19), (13.13), and (12.2)], the induced power (assuming zero tilt angle)
can be computed by:
PI = ωRW × λ,
(57)
where R is the rotor disk radius, W is the weight of the UAV, and ω is the angular velocity.
Also, given [32, (13.18), (13.13), and (12.1)], we can find λ by solving the following equation:
r
vh2
2 4
+ λ2 − W = 0.
(58)
g (λ) = 2ρπω R λ
ω 2 R2
The power consumption due to the vertical climbing and descending (assuming rapid descent)
can be given by [32, equations (12.35), (12.47), (12.50)]:
q
W vv + W vv 2 + 2W2 , climbing,
2
2
ρπR
PV =
q
W vv − W vv 2 − 2W2 , descending (in windmill state),
2
2
ρπR
(59)
where vv is the effective vertical velocity. Finally, the total mobility energy consumption is
computed using (55)-(59).
Clearly, the optimization problem in (52) is an integer linear programming (ILP). Following the
similar approach we used for solving (15), we transform problem (52) to a standard assignment
problem which can be solved using the Hungarian method in a polynomial time with a complexity
25
Table I: Simulation parameters.
Parameter
Description
Value
Pmax
Maximum transmit power of each device
200 mW
α
Path loss exponent for LoS links
2
σ2
Noise power
-130 dBm
γ
SINR threshold
5 dB
L
Total number of IoT devices
500
η1
Additional path loss to free space for LoS
3 dB
η2
Additional path loss to free space for NLoS
23 dB
of O(|In |3 ). To this end, we need to remove constraint (54) by considering Elk = +∞ when
the constraint is not satisfied. To determine when (54) is not satisfied, we use In−1 and In to
compute Elk , and compare it with the remaining energy of the UAVs, Γn,k . Then, in the objective
function (52), we replace each Elk corresponding to the unsatisfied constraint with Elk = +∞.
Consequently, (52) is transformed into a standard assignment problem. The result of solving (52)
will be the assignment matrix, Z, that optimally assigns the UAVs to the destinations. Therefore,
the locations of the UAVs are updated according to the new destinations. Then, having the destinations of each UAV at different update times, we can find the optimal trajectory of the UAVs.
V. S IMULATION R ESULTS AND A NALYSIS
For our simulations, the IoT devices are deployed within a geographical area of size 1 km ×
1 km. In this case, we consider a total number of 500 IoT devices which are uniformly distributed
on the area. Furthermore, we consider UAV-based communications in an urban environment with
ψ = 11.95 and β = 0.14 at 2 GHz carrier frequency [8]. Table I lists the simulation parameters.
Here, we analyze the transmit power of the IoT devices, the energy consumption of UAVs
on their mobility, and the update times. In our update time analysis, unless otherwise stated,
we consider the probabilistic activation model for the IoT devices with the beta distribution
parameters κ = 3, and ω = 4 [20]. When applicable, we compare our results with pre-deployed
stationary aerial base stations (UAVs) scenario while adopting the optimal device association
and power control technique of Subsection III-A. In the stationary case, the locations of UAVs
are assumed to be fixed over the target area and they are not updated according to the devices’
locations. All statistical results are averaged over a large number of independent runs.
Note that, in the given IoT network, serving all the active devices may not be possible due to
the limitations on the number of UAVs and the maximum transmit power of the devices. Thus,
26
0.9
0.8
Stationary aerial base stations
Proposed approach
Reliability
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
20
40
60
80
100
120
Pmax (mW)
Fig. 4: Reliability comparison between the proposed approach and stationary aerial base stations using 5 UAVs.
in Fig. 4, we show the achieved system reliability which, here, is defined as the probability that
all the active devices can be served by the UAVs. Clearly, the reliability depends on the locations
and transmit powers of the devices as well as the number of UAVs.
Fig. 4 shows the reliability as the maximum transmit power of the devices, Pmax , varies. In
this case, 5 UAVs are deployed to serve 100 active IoT devices. Clearly, as Pmax increases, the
reliability also increases. In fact, for higher Pmax values, the devices have higher a chance to
successfully connect to UAVs. From Fig. 4, we can see that, our proposed approach leads to a
significantly improved reliability compared to the case in which stationary aerial base stations are
used. In particular, the difference between the reliability of the stationary case and our proposed
approach is significant for lower Pmax . Indeed, a higher reliability is achieved by dynamically
optimizing the UAVs’ locations based on the locations of the IoT devices. As shown in Fig.
4, by increasing Pmax from 40 mW to 100 W, the reliability increases from 0.3 to 0.72 for the
stationary case, while it increases from 0.58 to 0.82 in our proposed approach. Furthermore, the
proposed approach yields a maximum of 28% improvement in the system reliability.
Fig. 5 shows a snapshot of the UAVs’ locations and their associated IoT devices (indicated
by the same color) resulting from the proposed approach. In this figure, 5 UAVs are efficiently
deployed to serve 100 active IoT devices which are uniformly distributed on the area. In this
case, all the devices are able to send their data to the associated UAVs by using a minimum
total transmit power. Here, the 3D locations of the UAVs as well as the device association are
determined based on the locations of the ground IoT devices and their transmit power.
In Fig. 6, we show the total transmit power needed by the IoT devices for reliable uplink
communications, versus the number of UAVs in the interference scenario. Clearly, the total
transmit power of the IoT devices can be reduced by deploying more UAVs. For instance,
27
UAV 1
UAV 2
UAV 3
UAV 4
UAV 5
Devices
Altitude (m)
400
300
200
100
0
1000
800
600
400
y−coordinate (m)
200
0
0
200
600
400
1000
800
x−coordinate (m)
Fig. 5: UAVs’ locations and associations for one illustrative snapshot.
5
Proposed approach
Stationary aerial base stations
4
3.5
3
2.5
2
1.5
1
0.5
0
0.5
Stationary aerial base stations
Proposed approach
0.45
Total transmit power (W)
Total transmit power (W)
4.5
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
5
6
7
8
9
10
Number of UAVs
0
5
6
7
8
9
10
Number of UAVs
Fig. 6: Total transmit power of devices vs. number of Fig. 7: Total transmit power of devices vs. number of
UAVs in the presence of interference.
UAVs in the interference-free scenario.
considering 100 active devices and 20 available channels, using our proposed approach, the total
transmit power decreases from 2.4 W to 0.2 W by increasing the number UAVs from 5 to 10.
Furthermore, using the proposed approach, the total transmit power of the devices decreases by
45% (on the average) compared to the stationary case. Clearly, for a lower number of UAVs, the
proposed approach leads to higher power reduction compare to the stationary case. In other words,
intelligently optimizing the locations of UAVs provides more power reduction gains when the
number of UAVs is low. In fact, for very dense networks with a high number of UAVs, updating
the UAVs’ locations is obviously no longer necessary compared to a case with a low number of
UAVs. For instance, as we can see from Fig. 6, the power reduction gain achieved by deploying
5 UAVs is around 7 times larger than the case with 10 UAVs.
Fig. 7 shows the total transmit power of the IoT devices as a function of the number of
UAVs in an interference-free scenario. Compared to the interference scenario, the devices can
obviously use a lower transmit power for sending their data to the UAVs. For instance, by
efficiently deploying only 5 UAVs, the devices can establish reliable uplink communications
with a total transmit power of 70 mW. Furthermore, Fig. 7 shows that, our proposed approach
28
Total transmit power (W)
1.8
Stationary aerial base stations
Proposed approach
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
25
30
35
40
45
50
Number of orthogonal channels (R)
Fig. 8: Total transmit power of devices vs. number of orthogonal channels.
leads to an average of 80% power reduction compared to the stationary case.
Fig. 8 shows the total transmit power of devices used for meeting the SINR requirement as
the number of available channels varies. The result in Fig. 8 corresponds to a case with 100
active devices which are served by 5 UAVs. Clearly, the total transmit power decreases as the
number of channels increases. This is due to the fact that, when more orthogonal resources
are available, the interference between the devices will decrease. As a result, each device can
reduce its transmit power while connecting to the serving UAV. From Fig. 8, we can see that,
by increasing the number of channels from 25 to 50, the total transmit power of devices can be
reduced by 68% in the proposed approach. In fact, the average number of interfering devices
decreases from 4 to 2 when we increase the number of channels from 25 to 50. Consequently,
less interference is generated by the devices while transmitting to the UAVs.
In Fig. 9, we show the average number of active devices that must be served by UAVs at
different update times tn which are normalized by T . Clearly, the number of active devices at
each update time depends the activation process of the devices and the number of update times
that indicates how frequently the UAVs serve the devices. In Fig. 9, due to the beta distributionbased activation pattern of the IoT devices, the number of active devices decreases when tn
exceeds 0.5 for N = 10. From Fig. 9, we can see that, for a higher number of update times
or equivalently shorter time period between consecutive updates, the average number of devices
that need to transmit their data decreases. For instance, considering tn = 0.6, the average number
of active devices decreases from 180 to 80 when the number of updates increases from 5 to 10.
We also note that, while a lower number of active devices leads to a lower interference between
the devices, it requires more updates and mobility for the UAVs. Fig. 9 also verifies that the
analytical results in Theorem 2 match the simulations. Furthermore, in Fig. 10, we show the
Average number of active devices
29
N=5 (Theory)
N=10 (Theory)
N=5 (Simulations)
N=10 (Simulations)
200
180
160
140
120
100
80
60
40
20
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Update time (normalized)
Fig. 9: Average number of active devices at update times for the probabilistic activation.
Exact number of active devices
300
250
200
10 updates
30 updates
150
100
50
0.1
0.2
0.3 0.4 0.5 0.6 0.7 0.8
Update time (normalized)
0.9
1
Fig. 10: Exact number of active devices at different update times for the periodic activation.
exact number of active devices for the periodic activation case obtained from Proposition 2. In
this case, each device becomes active with a certain activation period, τi . As expected, for a
higher number of updates, a lower number of active devices will need to be served by the UAVs.
For instance, by increasing the number of updates from 10 to 30, on the average, the number of
active devices decreases by 58%. Moreover, Fig. 10 shows that the maximum number of active
devices for 10 updates is about two times larger than the case with 30 updates. Therefore, in
order to avoid the interference between the devices, the number of orthogonal channels must be
increased by a two-fold factor when the number of updates decreases from 30 to 10.
Fig. 11 presents a direct result of Theorem 2 that computes the update times based on the
average number of active devices. Fig. 11 shows how to set update times in order to ensure that
the number of devices (which needs to be served) at each update time does not exceed a specified
number, a. As we can see from Fig. 11, to achieve a lower value of a, updates must occur more
frequently to reduce the time interval between the consecutive updates. For example, as can be
seen from this figure, to meet a = 100, 75, and 50, the 5th update must occur at tn = 0.41,
0.55, and 1. Moreover, Fig. 11 shows that, the number of updates increases as a decreases. For
example, in this case, to reduce a from 100 to 50, the number of updates needs to be doubled.
Fig. 12 shows the impact of the number of updates on the amount of energy that the UAVs use
30
Update time (normalized)
1
0.9
0.8
0.7
0.6
0.5
0.4
a = 100
a = 75
a = 50
0.3
0.2
0.1
1
2
3
4
5
6
7
8
9
10
Update number
Fig. 11: Update times for different average number of active devices.
to move. For our simulations, we have considered v = 10 m/s, ρ = 1.225 kg/m−3 , ω = 20 rad/s,
R = 0.5 m, cb = 10 cm, Nb = 4, and W = 50 N [32]. Intuitively, a higher number of updates
requires more mobility of the UAVs. Therefore, by increasing the number of updates, the total
energy consumption of the UAVs will also increase. As we can see from Fig. 12, by increasing
the number of updates from 3 to 6, the energy consumption of UAVs increases by a factor of
2.1 when the target area size is 1 km × 1 km. Note that, the mobility of the UAVs also depends
on the size of geographical area in which the devices are distributed. Clearly, on average, the
UAVs need to move further for covering a larger area.
Interestingly, there is an inherent tradeoff between the number of updates, mobility of the
UAVs, and transmit power of the IoT devices. In fact, considering Fig. 12, a higher number of
updates leads to a higher energy consumption of the UAVs due to the higher mobility. In addition,
as shown in Fig. 9, as the number of updates increases, a lower number of the IoT devices will
be active at each update time, and, hence, there will be lower interference between the devices.
As a result, the transmit power of the devices that is needed for satisfying the SINR requirement,
can be reduced. Note that, as we showed in Fig. 8, the devices’ transmit power decreases as the
interference decreases (by increasing the number of orthogonal channels). Hence, while a higher
number of updates leads to a lower devices’ transmit power, it requires more UAVs’ mobility.
Fig. 13 shows the overall convergence of the proposed power minimization algorithm that
is used for solving the original problem (5) considering 5 UAVs. As we can see from the
figure, in this case, the total transmit power of the IoT devices converges after 5 iterations.
In Fig. 13, each iteration corresponds to a joint solution to the device association and UAVs’
locations optimization problems. Clearly, after several iterations, updating the device association
Total energy consumption of UAVs (kJ)
31
500
Area size= 0.7km × 0.7km
Area size= 1km × 1km
400
300
200
100
0
2
4
6
8
10
Number of updates
25
1000
3
20
800
15
600
10
400
5
200
2
1
0
1
2
3
4
5
6
0
10
7
Number of iterations
12
14
16
Relative computational time
4
Optimality gap (%)
Total transmit power (W)
Fig. 12: Total UAV energy consumption vs. number of updates.
0
18
Number of active devices
Fig. 13: Overall convergence of the algorithm.
Fig. 14: Proposed approach vs. optimal solution.
and UAVs’ locations will no longer improve the solution.
In Fig. 14, we show an example to compare the accuracy and time complexity of our proposed
approach with the optimal solution obtained by an exhaustive search. Here, to perform an
exhaustive search over the continuous space, we have discretized the space with a resolution of
0.1 m. In this case, two UAVs are deployed to serve the devices. Clearly, the average gap between
the proposed solution and the optimal solution is around 11%. However, in this example, the
proposed solution is around 500 times, on the average, faster than the optimal solution.
VI. C ONCLUSION
In this paper, we have proposed a novel framework for efficiently deploying and moving UAVs
to collect data in the uplink from ground IoT devices. In particular, we have determined the
jointly optimal UAVs’ locations, device association, and uplink power control of the IoT devices
such that the total transmit power of the devices under their SINR constraints is minimized. In
addition, we have investigated the effective movement of the UAVs to collect the IoT data in a
time-varying IoT network. For this case, based on the devices activation process, we have derived
the update time instances at which the UAVs must update their locations. Furthermore, we have
obtained the optimal trajectories that are used by the UAVs to dynamically serve the IoT devices
with a minimum energy consumption. The results have shown that by intelligently moving and
32
deploying the UAVs, the total transmit power of the devices significantly decreases compared
to the case with pre-deployed stationary aerial base stations. Moreover, there is a fundamental
tradeoff between the number of updates, the UAVs’ mobility, and the devices’ transmit power.
R EFERENCES
[1] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Unmanned aerial vehicle with underlaid device-to-device communications: Performance and tradeoffs,” IEEE Transactions on Wireless Communications, vol. 15, no. 6, pp. 3949–3963, June
2016.
[2] Y. Zeng, R. Zhang, and T. J. Lim, “Wireless communications with unmanned aerial vehicles: opportunities and challenges,”
IEEE Communications Magazine, vol. 54, no. 5, pp. 36–42, May 2016.
[3] F. Jiang and A. L. Swindlehurst, “Optimization of UAV heading for the ground-to-air uplink,” IEEE Journal on Selected
Areas in Communications, vol. 30, no. 5, pp. 993–1005, June 2012.
[4] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for UAV-enabled mobile relaying systems,” IEEE Transactions
on Communications, vol. 64, no. 12, pp. 4983–4996, Dec. 2016.
[5] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Mobile Internet of Things: Can UAVs provide an energy-efficient
mobile architecture?” in Proc. of IEEE Global Communications Conference (GLOBECOM), Washington, DC, USA, Dec.
2016.
[6] R. Yaliniz, A. El-Keyi, and H. Yanikomeroglu, “Efficient 3-D placement of an aerial base station in next generation cellular
networks,” in Proc. of IEEE International Conference on Communications (ICC), Kuala Lumpur, Malaysia, May 2016.
[7] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Optimal transport theory for cell association in UAV-enabled cellular
networks,” IEEE Communications Letters, vol. 21, no. 9, pp. 2053–2056, Sept. 2017.
[8] A. Al-Hourani, S. Kandeepan, and S. Lardner, “Optimal LAP altitude for maximum coverage,” IEEE Wireless Communication Letters, vol. 3, no. 6, pp. 569–572, Dec. 2014.
[9] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Efficient deployment of multiple unmanned aerial vehicles for optimal
wireless coverage,” IEEE Communications Letters, vol. 20, no. 8, pp. 1647–1650, Aug. 2016.
[10] V. V. Chetlur and H. S. Dhillon, “Downlink coverage analysis for a finite 3D wireless network of unmanned aerial vehicles,”
IEEE Transactions on Communications, to appear, 2017.
[11] M. Chen, M. Mozaffari, W. Saad, C. Yin, M. Debbah, and C. S. Hong, “Caching in the sky: Proactive deployment of cacheenabled unmanned aerial vehicles for optimized quality-of-experience,” IEEE Journal on Selected Areas in Communications,
vol. 35, no. 5, pp. 1046–1061, May 2017.
[12] M. Mozaffari, W. Saad, M. Bennis, and M. Debbah, “Wireless communication using unmanned aerial vehicles (UAVs):
Optimal transport theory for hover time optimization,” available online: arxiv.org/abs/1704.04813, 2017.
[13] S.-Y. Lien, K.-C. Chen, and Y. Lin, “Toward ubiquitous massive accesses in 3GPP machine-to-machine communications,”
IEEE Communications Magazine, vol. 49, no. 4, pp. 66–74, Apr. 2011.
[14] J. Lyu, Y. Zeng, R. Zhang, and T. J. Lim, “Placement optimization of UAV-mounted mobile base stations,” IEEE
Communications Letters, vol. 21, no. 3, pp. 604–607, Mar. 2017.
[15] Y. Pang, Y. Zhang, Y. Gu, M. Pan, Z. Han, and P. Li, “Efficient data collection for wireless rechargeable sensor clusters
in harsh terrains using UAVs,” in Proc. of IEEE Global Communications Conference (GLOBECOM), Austin, TX, USA,
Dec. 2014.
33
[16] C.-Y. Tu, C.-Y. Ho, and C.-Y. Huang, “Energy-efficient algorithms and evaluations for massive access management in
cellular based machine to machine communications,” in Proc. of Vehicular Technology Conference (VTC), San Francisco,
CA, USA, Sep. 2011.
[17] K.-C. Chen and S.-Y. Lien, “Machine-to-machine communications: Technologies and challenges,” Ad Hoc Networks,
vol. 18, pp. 3–23, July 2014.
[18] M. Tavana, V. Shah-Mansouri, and V. W. S. Wong, “Congestion control for bursty M2M traffic in LTE networks,” in Proc.
of IEEE International Conference on Communications (ICC), London, UK, June 2015.
[19] X. Jian, X. Zeng, Y. Jia, L. Zhang, and Y. He, “Beta/M/1 model for machine type communication,” IEEE Communications
Letters, vol. 17, no. 3, pp. 584–587, Mar. 2013.
[20] 3GPP, “Study on RAN improvements for machine type communication,” TR 37.868, Sep. 2011.
[21] A. K. Gupta and S. Nadarajah, Handbook of beta distribution and its applications.
CRC press, 2004.
[22] S. Z. Selim and M. A. Ismail, “K-means-type algorithms: a generalized convergence theorem and characterization of local
optimality,” IEEE Transactions on Pattern Analysis and Machine Intelligence, no. 1, pp. 81–87, 1984.
[23] R. D. Yates, “A framework for uplink power control in cellular radio systems,” IEEE Journal on Selected Areas in
Communications, vol. 13, no. 7, pp. 1341–1347, Sep. 1995.
[24] R. Sun, M. Hong, and Z. Q. Luo, “Joint downlink base station association and power control for max-min fairness:
Computation and complexity,” IEEE Journal on Selected Areas in Communications, vol. 33, no. 6, pp. 1040–1054, June.
2015.
[25] R. E. Burkard, M. Dell’Amico, and S. Martello, Assignment Problems, Revised Reprint.
SIAM, 2009.
[26] H. W. Kuhn, “The hungarian method for the assignment problem,” Naval research logistics quarterly, vol. 2, no. 1-2, pp.
83–97, 1955.
[27] P. T. Boggs and J. W. Tolle, “Sequential quadratic programming,” Acta numerica, vol. 4, pp. 1–51, 1995.
[28] K. Scheinberg, “An efficient implementation of an active set method for SVMs,” Journal of Machine Learning Research,
vol. 7, no. Oct, pp. 2237–2257, 2006.
[29] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge university press, 2004.
[30] C. Di Franco and G. Buttazzo, “Energy-aware coverage path planning of UAVs,” in Proc. of IEEE International Conference
on Autonomous Robot Systems and Competitions (ICARSC), Vila Real, Portugal, Apr. 2015.
[31] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with trajectory optimization,” IEEE Transactions on Wireless
Communications, vol. 16, no. 6, pp. 3747–3760, June 2017.
[32] A. Filippone, Flight performance of fixed and rotary wing aircraft.
Elsevier, 2006.
| 7cs.IT
|
On the weak order ideal associated to linear codes
Mijail Borges-Quintana,
Miguel Ángel Borges-Trenard∗
arXiv:1705.06609v1 [cs.IT] 17 May 2017
Edgar Martı́nez-Moro
†
Abstract
In this work we study a weak order ideal associated with the coset
leaders of a non-binary linear code. This set allows the incrementally
computation of the coset leaders and the definitions of the set of leader
codewords. This set of codewords has some nice properties related to
the monotonicity of the weight compatible order on the generalized
support of a vector in Fnq which allows to describe a test set, a trial set
and the set of zero neighbours of a linear code in terms of the leader
codewords.
keywords Linear codes, order ideals, test set, trial set, zero neighbours,
correctable errors
1
Introduction
As it is pointed in [5] it is common folklore in the theory of binary linear
codes that there is an ordering on the coset leaders chosen as the lexicographically smallest minimum weight vectors that provides a monotone structure.
This is expressed as follows: if x is a coset leader and y ⊆ x (i.e. yi ≤ xi
for all i) then y is also a coset leader. This nice property has been proved
of great value, see for example [12], and it has been used for analyzing the
error-correction capability of binary linear codes [5]. In this last paper the
authors introduce the concept of a trial set of codewords and they provide
a gradient-like decoding algorithm based on this set.
Finding the weight distribution of cosets leaders for a code C is a classic
problem in coding theory. This problem is still unsolved for many family
∗
M. Borges-Quintana and M.A. Borges-Trenard are with the Department of Department of Mathematics, Faculty of Ciencias Naturales y Exactas, Universidad de Oriente,
Santiago de Cuba, Cuba. Emails: [email protected], [email protected]. M. BorgesQuintana has been partially supported by a Post-doctorate scholarship at the University
of Valladolid (09-2014 to 02-2015) by Erasmus Mundus Program, Mundus Lindo Project.
†
E. Martı́nez-Moro is with the Institute of Mathematics IMUVa, University of Valladolid. Valladolid, Castilla, Spain. Email: [email protected]. Partially supported by
Partially supported by the Spanish MINECO under grants MTM2015-65764-C3-1-P and
MTM2015-69138-REDT.
1
of linear codes even for first-order Reed-Muller codes (see [6]). The set
of coset leaders is also related with the minimum distance decoding and
bounded distance decoding problems as well with the set of minimal support
codewords. [1, 9, 10]
Despite their interest no generalization of these ideas is known by the
authors of this communication to non-binary case. In this paper we provide
such a (non straightforward) generalization.
The outline of the paper is as follows Section 2 introduces the idea of a
generalized support of a vector In Section 3 is is defined the weak order ideal
associated with the coset leaders O(C) and it is shown that can be computed
incrementally. Theorem 3.4 establishes that all the coset leaders of the code
belong to O(C). Subsection 3.2 is devoted to the study of the set of leader
codewords of a code as a zero neighbour set and their properties. Finally
in Section 4 we analyze the correctable and uncorrectable errors defining a
trial set for a linear code from the set of leader codewords.
The limitations from a practical point view of the results and properties
studied in this paper are clear because of the size and the complexity of computing the set of coset leaders. Anyway our main interest is the study and
characterizations of some objects related to the codes like zero neighbours,
trial set, and the set of correctable and uncorrectable errors.
2
Preliminaries
From now on we shall denote by Fq the finite field with with q = pm elements,
p a prime. A linear code C over Fq of length n and dimension k is a kdimensional subspace of Fnq . We will call the vectors v in Fnq words and
those v ∈ C, codewords. For every word v ∈ Fnq its support is defined
as supp(v) = {i | vi 6= 0} and its Hamming weight, denoted by wH (v) as
the cardinality of supp(v) and the Hamming distance dH (x, y) between two
words x, y ∈ Fnq is dH (x, y) = wH (x − y). The minimum distance d(C)
of a linear code C is defined as the minimum weight among all nonzero
codewords.
The words of minimal Hamming weight in the cosets of Fnq /C is the set of
coset leaders of the code C in Fnq and we will denote it by CL(C). CL(y) will
denote the subset of coset leaders corresponding to the coset y + C. Given
a coset y + C we define the weight of the coset wH (y + C) as the smallest
Hamming weight among all vectors in the coset, or equivalently the weight
⌋ where ⌊·⌋
of one of its leaders. It is well known that given t = ⌊ d(C)−1
2
denotes the greatest integer function then every coset of weight at most t
has a unique coset leader.
Let f (X) be an irreducible polynomial over Fp of degree m and β be a
root of f (X), then any element a ∈ Fq can be represented as a1 + a2 β + . . . +
am β m−1 with ai ∈ Fp for i ∈ {1, . . . , m}. For a word v = (v1 , . . . , vn ) ∈ Fnq ,
2
such that the i-th component of v is vi = vi,1 + vi,2 β + . . . + vi,m β m−1 we
define the generalized support of a vector v as the support of the nm-tuple
given by the concatenations of the p-adic expansion of each component vi
of v, i.e.
suppgen (v) = (supp((vi,1 , . . . , vi,m )) : i = 1 . . . n),
and suppgen (v)[i] = supp((vi,1 , . . . , vi,m )). We will say that (i, j) ∈ suppgen (v)
if the corresponding vi,j is not zero. From now on the set
eij = β j−1 ei : i = 1, . . . , n; j = 1, . . . m
will be denoted as Can(Fq , f ) and it represents the canonical basis of (Fnq , +),
the additive monoid Fnq with respect to the “+” operation, where f is the
irreducible polynomial used to define Fq .
We state the following connection between Fnq and Nnm :
∆ : Fnq → Nnm
v 7→ (ψ(vi,j ) : i = 1, . . . , n, j = 1, . . . , m), where
ψ : Fp → N
k · 1Fp 7→ k mod p.
On the other hand,
∇ : Nnm → Fnq
a 7→ ((am(i−1)+1 + am(i−1)+2 β + . . . + am(i−1)+m β m−1 ) : i = 1, . . . , n).
P
P
Given x, y ∈ (Fnq , +), x = i,j xij eij , y = i,j yij eij , we say x ⊂ y if
ψ(xij ) ≤ ψ(yij ) for all i ∈ [1, n], and j ∈ [1, m].
By using ∆ it is possible to relate orders on Fnq with orders on Nnm ,
and vice versa. An admissible order on (Nnm , +) is a total order < on Nnm
satisfying the following two conditions
1. 0 < x, for all x ∈ Nnm , x 6= 0.
2. If x < y, then x + z < y + z, for all z ∈ Nnm .
In particular, Any admissible order on (Nnm , +), like the Lexicographical,
Degree Lexicographical, Degree Reverse Lexicographical orders, induces an
order on (Fnq , +).
We will say that a representation of a word v as an nm-tuple over N is
in standard form if ∆(∇(v)) = v. We will denote the standard form of v as
SF(v, f ) (note that ∇(v) = ∇(SF(v, f ))). Therefore, v is in standard form
if v = SF(v, f ) (we will also say v ∈ SF(Fnq , f )).
Remark 1. From now on we will use Can(Fq ) and SF(Fnq ) instead of Can(Fq , f )
and SF(Fnq , f ) respectively, since it is clear that different elections of f or β
provide equivalent generalized supports.
3
Definition 1. A subset O of Nk is an order ideal if for all w ∈ O and v ∈ Nk
s.t. vi ≤ wi , i = 1, . . . , k, we have v ∈ O.
In the same fashion we say that a subset S of Fnq is an order ideal if ∆(S)
is an order ideal in Nnm . It is easy to check that an equivalent definition for
the order ideal would be that for all w ∈ S, and for all (i, j) ∈ suppgen (w),
and v ∈ Fnq s.t. w = v + eij we have v ∈ S. If instead of for all (i, j) ∈
suppgen (w) the condition is satisfied at least for one (i, j) ∈ suppgen (w) we
say that S is a weak order ideal.
Definition 2. A subset S of Fnq is a weak order ideal if for all w ∈ S \ 0
there exists (i, j) ∈ suppgen (w) s.t. for v ∈ Fnq s.t. w = v + eij we have
v ∈ S.
Definition 3. The Voronoi region of a codeword c ∈ C is the set
D(c) = y ∈ Fnq | dH (y, c) ≤ dH (y, c′ ), ∀c′ ∈ C \ {c} .
The set of all the Voronoi regions for a given linear code C covers the
space Fnq and D(0) = CL(C). However, some words in Fnq may be contained
in several regions. For any subset A ⊂ Fnq we define
X (A) = y ∈ Fnq | min {dH (y, a) : a ∈ A} = 1
as the set of words at Hamming distance 1 from A, i.e. The boundary of A
is defined as δ(A) = X (A) ∪ X (Fnq \ A).
Definition 4. A nonzero codeword z ∈ C is called a zero neighbour if its
Voronoi region shares a common boundary with the set of coset leaders, i.e.
δ(D(z)) ∩ δ(D(0)) 6= ∅. The set of all zero neighbours of C is denoted by
Z(C) = {z ∈ C \ {0} : δ(D(z)) ∩ δ(D(0)) 6= ∅} .
Definition 5. A test-set T for a given linear code C is a set of codewords
such that every word y
1. either y lies in D(0)
2. or there exists v ∈ T such that wH (y − v) < wH (y).
The set of zero neighbours is a test set, also from the set of zero neighbours can be obtained any minimal test set according to the cardinality of
the set [1].
3
The weak order ideal of the coset leaders
The first idea that allows us to compute incrementally the set of all coset
leaders for a linear code was introduced in [2]. In that paper we used the
additive structure of Fnq with the set of canonical generators Can(Fq ). Unfortunately in [2] most of the chosen coset representatives may not be coset
leaders if the weight of the coset is greater than t.
4
Theorem 3.1 ([7], Theorem 1.12.6.v). Assume that x is a coset leader of
C. If x′ ∈ Fnq and x′i = xi for all i ∈ supp(x′ ), then x′ is also a coset leader
of C.
In order to incrementally generate all coset leaders starting from 0 adding
elements in Can(Fq ), we must consider words with weight one more than the
previous chosen coset leader. Next result is a byproduct of Theorem 3.1, we
may characterize which vectors we need to generate with weight one more
than its coset leader in order to ensure all coset leaders are generated.
Theorem 3.2. Let x ∈ SF(Fnq ) be an element in CL(C), let i ∈ supp(x). If
x′ ∈ Fnq and x′j = xj for all j ∈ supp(x′ )\{i}, then wH (x′ ) ≤ wH (x′ + C)+ 1.
Proof. By Theorem 3.1, x − xi ∈ CL(C). The proof of the Theorem is
analogous to the proof of Theorem 1.12.6.v in [7]. Note that if we suppose
that wH (x′ ) ≥ wH (x′ + C) + 2 it would imply that x is not a coset leader,
which is a contradiction.
Let w ∈ SF(Fnq ) be an element in CL(C), and (i, j) ∈ suppgen (w). Let
y ∈ SF(Fnq ) s.t. w = y + eij then, as a consequence of the previous theorem
we have that
wH (y) ≤ wH (y + C) + 1.
(1)
In the situation above we will say that the coset leader w is an ancestor
of the word y, and that y is a descendant of w. In the binary case this
definitions behave as the ones in [7, §11.7] but in the case q 6= 2 there is a
subtle difference, a coset leader could be an ancestor of another coset leader
or an ancestor of a word at Hamming distance 1 to a coset leader (this last
case is not possible in the binary case).
3.1
The set O(C)
Given ≺1 an admissible order on (Nnm , +) we define the weight compatible
order ≺ on (Fnq , +) associated to ≺1 as the ordering given by
1. x ≺ y if wH (x) < wH (y) or
2. if wH (x) = wH (y) then ∆(x) ≺1 ∆(y).
I.e. the words are ordered according their weights and the order ≺1 break
ties. These class of orders is a subset of the class of monotone α-orderings
in [5]. In fact we will need a little more than monotonicity, for the purpose
of this work we will also need that for every pair v, w ∈ SF(Fnq )
if v ⊂ w, then v ≺ w.
5
(2)
Note that (2) is satisfied for a weight compatible order. In addition, for
any weight compatible order ≺ every strictly decreasing sequence terminates
(due to the finiteness of the set Fnq ). In the binary case the behavior of the
coset leaders can be translated to the fact that the set of coset leader is an
order ideal of Fn2 ; whereas, for non binary linear codes this is no longer true
even if we try to use the characterization of order ideals given in [4], where
order ideals do not need to be associated with admissible orders.
Definition 6. We define the weak order ideal of the coset leaders of C as
the set O(C) of elements in Fnq verifying one of the following items:
1. 0 ∈ O(C).
2. Criterion
1: If v ∈ O(C) and wH (v) = wH (v + C) then
v + eij | ∆(v) + ∆(eij ) ∈ SF(Fnq ) ⊂ O(C).
3. Criterion 2: If v ∈ O(C) and wH (v) = wnH (v + C) + 1 then
v + eij | i ∈ supp(v), ∆(v) + ∆(eij ) ∈ SF(Fq ) , v − vi ∈ CL(C)} ⊂ O(C).
Remark 2. It is clear by the Criteria 1 and 2 of the definition above that
O(C) is a weak order ideal.
Theorem 3.3. Let w ∈ Fnq . If there exists i ∈ 1, . . . , n s.t. w − wi ∈ CL(C)
then w ∈ O(C).
Proof. We will proceed by induction on Fnq with respect to the order ≺.
The statement is true for 0 ∈ Fnq . Now for the inductive step, we assume
that the desired property is true for any word u ∈ Fnq such that there exists
i ∈ 1, . . . , n s.t. u − ui ∈ CL(C) and also u is smaller than an arbitrary but
fixed w 6= 0 with respect to ≺ and w − wj ∈ CL(C), for some j ∈ 1, . . . , n,
i.e. if u − ui ∈ CL(C), for some i ∈ 1, . . . , n, and u ≺ w then u ∈ O(C).
We will show that the previous conditions imply that w is also in O(C).
Let w = v + eij , with (i, j) ∈ suppgen (w) then v ≺ w by (2). As
w − wi ∈ CL(C), the same is true for v, i.e, v − vi ∈ CL(C) then by the
induction hypothesis we have that v ∈ O(C). By Theorem 3.2, wH (v) ≤
wH (v + C) + 1; therefore, by Criteria 1 or 2 in Definition 6 it is guaranteed
that w ∈ O(C).
Theorem 3.4. Let w ∈ Fnq and w ∈ CL(C) then w ∈ O(C).
Proof. Let (i, j) ∈ suppgen (w), since w ∈ CL(C), by Theorem 3.2, w − wi ∈
CL(C); then, by Theorem 3.3, w ∈ O(C).
The previous theorem has been shown that O(C) contains the set of coset
leaders of the linear code C.
6
3.2
Zero neighbours and leader codewords
Definition 7. The set of leader codewords of a linear code C is defined as
v1 + eij − v2 ∈ C \ {0} | ∆(v1 ) + ∆(eij ) ∈ SF(Fnq ),
L(C) =
.
v2 ∈ CL(C) and v1 − v1i ∈ CL(C)
Note that the definition is a bit more complex that the one for binary
codes in [3] due to the fact that in the general case not all coset leaders need
to be ancestors of coset leaders. The name of leader codewords comes from
the fact that one could compute all coset leaders of a corresponding word
knowing the set L(C) adapting [3, Algorithm 3].
Remark 3. The algorithm for computing L(C) is based on the construction of
O(C). Theorem 3.3 guarantees that w ∈ O(C) provided that w−wi ∈ CL(C)
for some i, then the associated set of leader codewords may be computed as
{w − v : w ∈ O(C), w − wi ∈ CL(C), v ∈ CL(w) and v 6= w}.
Theorem 3.5 (Properties of L(C)). Let C be a linear code then
1. L(C) is a test set for C.
2. Let w be an element in L(C) then wH (w) ≤ 2ρ(C) + 1 where ρ(C) is
the covering radius of the code C.
3. If w ∈ L(C) then X (D(0)) ∩ (D(w) ∪ X (D(w))) 6= ∅.
4. If X (D(0)) ∩ D(w) 6= ∅ then w ∈ L(C).
Proof. 1) Let y ∈
/ CL(C) and supp(y)
= {ik : k = 1, . . . , l}, l ≤ n. Let
P
/
s be such that 1 ≤ s < l, v1 = sk=1 yik eik ∈ CL(C), v1 +PyiS+1 eiS+1 ∈
z
CL(C) and v2 = CL(v1 + yiS+1 eiS+1 ). Let yiS+1 eiS+1 =
t=1 eijt and
v1 ′ = v1 + yiS+1 eiS+1 − eijZ . Then v1 ′ − v1 ′i = v1 ∈ CL(C) implies that
w = v1 ′ + eijZ − v2 ∈ L(C). In addition,
wH (y − w) = wH (v2 + (y − v1 ′ − eijZ )) ≤ wH (v2 ) + wH (y − v1 ′ − eijZ ) and
wH (v2 ) + wH (y − v1 ′ − eijZ ) < wH (v1 ′ + eijZ ) + wH (y − v1 ′ − eijZ ).
Note that supp(v1 ′ + eijZ ) ∩ supp(y − v1 ′ − eijZ ) = ∅; consequently,
wH (y) = wH (v1 ′ + eijZ ) + wH (y − v1 ′ − eijZ ) and wH (y − w) < wH (y).
Thus, L(C) is a test set.
2) Let c ∈ L(C) then there exists v2 ∈ CL(C), v1 ∈ SF(Fnq ), 1 ≤ i ≤ n
and 1 ≤ j ≤ m such that v1 − v1i ∈ CL(C) and c = v1 + eij − v2 . Applying
the definition of covering radius we have that wH (v1 − v1i ), wH (v2 ) ≤ ρ,
thus wH (c) ≤ 2ρ + 1.
3) Let w ∈ L(C), then w = v1 + eij − v2 , where v1 , v2 are elements in
n
Fq such that v1 + eij ∈ SF(Fnq ), v2 ∈ CL(C) and v1 − v1i ∈ CL(C).
• If v1 + eij ∈
/ CL(C), then v1 + eij ∈ X (D(0)) and (v1 + eij ) − w =
v2 ∈ CL(C) implies that v1 + eij ∈ D(w).
7
• If v1 + eij ∈ CL(C) we define v1′ = v1′ (0) = v1 + eij . It is clear
be a number in the set
that v1′ , v2 ∈ CL(v2 ). Since w 6= 0 let l P
′
{1, . . . , n} such that v1l −v2l 6= 0. Let v2l = Tj=1 elij , for 1 ≤ h ≤ T ,
P
P
v1 ′ (h) = v1′ + hj=1 elij and v2 (h) = v2 − hj=1 elij .
/ CL(C) and
If there exists an h (1 ≤ h < T ) such that v1′ (h) ∈
′
v1 (h − 1) ∈ CL(C) then these two conditions imply that v1′ (h) ∈
X (D(0)). On the other hand, v2 (h) = v1′ (h) − w is either a coset
leader (v1′ (h) ∈ D(w)) or dH (v2 (h), v2 ) = 1 (v1′ (h) ∈ X (D(w))).
If there is no such and h (1 ≤ h < T ) satisfying the condition then
wH (v1′ (T )) = wH (v2 (T )) + 1, which means that v1′ (T ) is not a coset
leader and v1′ (T − 1) is a coset leader. Then using the same idea of
the previous paragraph we have that v1′ (T ) ∈ X (D(0)) and v1′ (T ) ∈
D(w) ∪ X (D(w)).
4) If X (D(0)) ∩ D(w) 6= ∅, let u ∈ X (D(0)) ∩ D(w). The first condition
u ∈ X (D(0)) implies u = v1 + eij for some v1 ∈ SF(Fnq ) , (i, j) ∈ suppgen (u)
and v1 − v1i ∈ CL(C). On the other hand, v1 + eij ∈ D(w) implies that
v2 = (v1 + eij ) − w ∈ CL(C). Therefore, w = v1 + eij − v2 ∈ L(C).
Remark 4. Note that item 3 in Theorem 3.5 implies that any leader codeword is a zero neighbour however, one of the differences with the binary
case is that it is not always true that for a leader codeword w we have that
X (D(0)) ∩ D(w) 6= ∅, although by item 4 we have that w is a leader codeword provided this condition is satisfied. Furthermore, item 4 guarantees
that the set of leader codewords contains all the minimal test set according to its cardinality (see [1]). As a consequence of all these properties in
Theorem 3.5 we could say that the the set of leader codewords is a “good
enough” subset of the set of zero neighbours.
4
Correctable and uncorrectable errors
We define the relation ⊂1 in the additive monoid which describe exactly the
relation ⊂ in the vector space Fnq . Given x, y ∈ (Fnq , +)
x ⊂1 y if x ⊂ y and supp(x) ∩ supp(y − x) = ∅.
(3)
Note that this definition translates to Fnq the binary case situation in
[5]. In this case given a y ∈ (Fnq , +) there are more words x ∈ (Fnq , +)
such that x ⊂ y than if we consider x, y as elements in the vector space Fnq .
Of course, any relation x ⊂ y in Fnq as a vector space it is also true in the
additive monoid, but it is not true the other way round.
The set E 0 (C) of correctable errors of a linear code C is the set of the
minimal elements with respect to ≺ in each coset. The elements of the set
8
E 1 (C) = Fnq \E 0 (C) will be called uncorrectable errors. A trial set T ⊂ C \{0}
of the code C is a set which has the following property
y ∈ E 0 (C) if and only if y y + c, for all c ∈ T.
Since ≺ is a monotone α-ordering on Fnq , the set of correctable and
uncorrectable errors form a monotone structure. Namely, if x ⊂1 y then
x ∈ E 1 (C) implies y ∈ E 1 (C) and y ∈ E 0 (C) implies x ∈ E 0 (C). In the
general case q 6= 2 there is a difference with respect to the binary case, there
may be words y′ ∈ Fnq s.t. suppgen (y′ ) = suppgen (y), y′ ⊂ y and y′ could
be either a correctable error or an uncorrectable error, so, the monotone
structure it is not sustained by ⊂ in the additive monoid (Fnq , +).
Let the set of minimal uncorrectable errors M 1 (C) be the set of y ∈ E 1 (C)
such that, if x ⊆1 y and x ∈ E 1 (C), then x = y. In a similar way, the set
of maximal correctable errors is the set M 0 (C) of elements x ∈ E 0 (C) such
that, if x ⊆1 y and y ∈ E 0 (C), then x = y.
For c ∈ C \ {0}, a larger half is defined as a minimal word u ⊆1 c in the
ordering such that u − c ≺ u. The weight of such a word u is such that
wH (c) ≤ 2wH (u) ≤ wH (c) + 2,
see [5] for more details. The set of larger halves of a codeword c is denoted
by LH (c), and for U ⊆ C \ {0} the set of larger halves for elements of U is
denoted by LH (U ). Note that LH (C) ⊆ E 1 (C).
For any y ∈ Fnq , let H(y) = {c ∈ C : y − c ≺ y}, and we have y ∈ E 0 (C)
if and only if H(y) = ∅, and y ∈ E 1 (C) if and only if H(y) 6= ∅.
In [5, Theorem 1] there is a characterization of the set M 1 (C) in terms of
H(·) and larger halves of the set of minimal codewords M (C) for the binary
case. It is easy to proof that this Theorem and [5, Corollary 3] are also true
for any linear code.
Proposition 4.1 (Corollary 3 in [5] ). Let C be a linear code and T ⊆ C\{0}.
The following statements are equivalent,:
1. T is a trial set for C.
2. If y ∈ M 1 (C), then T ∩ H(y) 6= ∅.
3. M 1 (C) ⊆ LH (T ).
Now we will formulate the result which relates the trial sets for a given
weight compatible order ≺ and the set of leader codewords.
Theorem 4.2. Let C be a linear code and L(C) the set of leaders codewords
for C, the following statements are satisfied.
1. L(C) is a trial set for any given ≺.
9
2. Algorithm 2 in [3] can be adapted to compute a set of leader codewords
which is a trial set T for a given ≺ such that satisfies the following
property
For any c ∈ T , there exists y ∈ M 1 (C) ∩ LH (c) s.t. y − c ∈ E 0 (C).
Proof.
Proof of 1) We will prove statement 2 of Proposition 4.1. Let y ∈ M 1 (C),
let i such that suppgen (y)[i] 6= ∅ and v1 = y − yi . Since y ∈ M 1 (C) we have
that v1 ∈ E 0 (C), thus it is a coset leader. On the other hand, let v2 ∈ E 0 (C)
such that v2 ∈ CL(y) and c = y − v2 . It is clear that c is a leader codeword
and y − c = v2 ≺ y. Therefore c ∈ H(y).
Proof of 2) In the Algorithm 2 in [3], as the first step, it is necessary to
add to the function InsertNext the Criteria 2 of the construction of O(C),
whose elements are stored in Listing. On the other hand, in the steps of
the construction of the leader codewords (Steps 11 - 13) it is enough to state
the condition t ∈ M 1 (C) for t and taking tk only equal to the coset leader
of CL(t), that is the corresponding correctable error and add the codeword
t − tk to the set L(C).
References
[1] A. Barg, Complexity issues in coding theory. In Handbook of coding
theory, Vol. I, North-Holland, Amsterdam, pp. 649–754, 1998.
[2] M. Borges-Quintana, M.A. Borges-Trenard, E. Martı́nez-Moro, On a
Gröbner bases structure associated to linear codes, J. Discret. Math.
Sci. Cryptogr. 10(2), pp. 151–191, 2007.
[3] M. Borges-Quintana, M.A. Borges-Trenard, I. Márquez-Corbella, E.
Martı́nez-Moro, Computing coset leaders and leader codewords of binary
codes, Journal of Algebra and its Applications 14, 2015.
[4] G. Braun, S. Pokutta. A polyhedral characterization of border bases.
SIAM J. Discrete Math 30(1), pp. 239–265, 2016.
[5] T. Helleseth, T. Kløve, I.L. Vladimir, Error-correction capability of binary linear codes, IEEE Transactions on Information Theory 51(4), pp.
1408–1423, 2005.
[6] R.P. Kurshan and N.J.A. Sloane. Coset analysis of reed muller codes
via translates of finite vector spaces, Information and Control 20(5), pp.
410 - 414, 1972.
[7] W.C. Huffman, V. Pless, Fundamentals of error-correcting codes, Cambridge University Press, Cambridge, 2003.
10
[8] F. J. MacWilliams, N. J. A. Sloane. The Theory of Error-Correcting
Codes, Elsevier/North Holland, 1997.
[9] I. Márquez-Corbella, E. Martı́nez-Moro. Algebraic structure of the minimal support codewords set of some linear codes. Adv. Math. Commun.
5(2), pp. 233–244, 2011.
[10] J. L. Massey. Minimal codewords and secret sharing, Proceedings of
the 6th Joint Swedish-Russian International Workshop on Information
Theory, pp. 246–249, 1993.
[11] T. Mora. Solving Polynomial Equation Systems II: Macaulay’s
Paradigm and Gröbner Technology. Cambridge University Press, 2005.
[12] G. Zémor, G. Cohen. The threshold probability of a code, IEEE Trans.
Inform. Theory 41(2), pp. 469–477, 1995.
11
| 7cs.IT
|
Deep learning for word-level handwritten
Indic script identification
Soumya Ukil1 , Swarnendu Ghosh1 , Sk Md Obaidullah2 , K. C. Santosh3 , Kaushik Roy4 ,
and Nibaran Das1
1
Dept. of Computer Science & Engineering, Jadavpur University, Kolkata 700032, WB, India
Dept. of Computer Science & Engineering, Aliah University, Kolkata 700156, WB, India
3
Dept. of Computer Science, The University of South Dakota, Vermillion, SD 57069, USA
Dept. of Computer Science & Engineering, West Bengal State University, 700126, WB, India
arXiv:1801.01627v1 [cs.CV] 5 Jan 2018
2
4
Corresponding authors: K. C. Santosh ([email protected]) & N. Das ([email protected])
Abstract. We propose a novel method that uses convolutional neural networks
(CNNs) for feature extraction. Not just limited to conventional spatial domain representation, we use multilevel 2D discrete Haar wavelet transform, where image representations are scaled to a variety of different sizes. These are then used to train
different CNNs to select features. To be precise, we use 10 different CNNs that select
a set of 10240 features, i.e. 1024/CNN. With this, 11 different handwritten scripts
are identified, where 1K words per script are used. In our test, we have achieved the
maximum script identification rate of 94.73% using multi-layer perceptron (MLP).
Our results outperform the state-of-the-art techniques.
Keywords: Convolutional neural network, deep learning, multi-layer perceptron,
discrete wavelet transform, Indic script identification
1
Introduction
Optical character recognition (OCR) has always been a challenging field in pattern recognition. OCR techniques are used to convert handwritten or machine printed scanned document
images to machine-encoded texts. These OCR techniques are script dependent. Therefore,
script identification is considered as a precursor to OCR. In particular, in case of a multilingual country like India script identification is the must since a single document, such as
postal documents and business forms, contains several different scripts (see Fig. 1).
Indic handwritten script identification has a rich state-of-the-art literature [1–4]. More
often, previous works have been focusing on word-level script identification [5]. Not stopping
there, in a recent work [6], authors introduced page-level script identification performance
to see whether we can expedite the processing time. In general, in their works, hand-crafted
features that are based on structural and/or visual appearances (morphology-based) were
used. The question is, are we just relying on what we see and use apply features accordingly
or can we just let machine to select features that are required for optimal identification rate?
This inspires to use deep learning, where CNNs can be used for extracting and/or selecting
features for identification task(s).
Needless to say, CNNs have stood well with their immense contribution in the field of
OCR. Their onset has ben marked by the ground-breaking performance of CNNs on MNIST
dataset [7]. Very recently, the use CNN for Indic script (Bangla character recognition) has
Fig. 1. Two multi-script postal document images, where Bangla, Roman and Devanagari scripts
are used.
been reported [8]. Not to be confused, the primary of goal of this paper is to use deep learning
concept to identify 11 different handwritten Indic scripts: Bangla, Devnagari, Gujarati,
Gurumukhi, Kannada, Malayalam, Oriya, Roman, Tamil, Telugu and Urdu. Inspired from
deep learning-based concept, we use CNNs to select features from handwritten document
images (scanned), where we use multilevel 2D discrete Haar wavelet transform (in addition
to conventional spatial domain representation) and image representations are scaled to a
variety of different sizes. With these representation, several different CNNs are used to
select features. In short, the primary idea behind this is to avoid using hand-crafted features
for identification. Using multi-layer perceptron (MLP), 11 different handwritten scripts (as
mentioned earlier) are identified with satisfactory performance.
The remainder of the paper can be summarized as follows. Section 2 provides a quick
overview of our contribution, where it includes CNN architecture and feature extraction
process. In Section 3, experimental results are provided. It also includes a quick comparison
study. Section 4 concludes our paper.
2
Contribution outline
As mentioned earlier, in stead of using hand-crafted features for document image representation, our goal is to let deep learning to select distinguishing features for optimal script
identification. For a few but recent works, where CNNs have used with successful classification, we refer to [7, 9, 10]. We observe that CNNs work especially when we have sufficient
data to train. This means, data redundancies will be helpful. In general, CNN takes raw
pixel data (image) and as training proceeds, the model learns distinguishing features that
can successfully contribute to identification/classification. Such a training process produces
a feature vector that summarize the important aspects of the studied image(s).
More precisely, our approach is twofold: first, we use a two- and three-layered CNNs for
three different scales of the input image; and secondly, we use exactly same CNNs for two
different scales of the transformed image (wavelet transform). We then merge those features
and make ready for script identification. In what follows, we explain our CNN architecture
including definitions, parameters for wavelet transform and the way we produce features.
Fig. 2. Schematic block diagram of handwritten Indic script identification showing different modules: feature extraction/selection and classification.
2.1
CNN architecture
In general, a CNN has a layered architecture consisting of three basic types of layers namely,
1. convolutional layer (CL),
2. pooling layer (PL) and
3. fully connected layer (FCL).
CLs consist of a set of kernels that produce parameters and help in convolution operation. In
CL, every kernel generates an activation map as an output. PLs do not have parameters but,
their major role is to avoid possible data redundancies (still preserving their significance). In
our approach, all CNNs have max-pooling operation at their corresponding PLs. In addition
to these two different types of layers, FCL is used, where MLP has been in place.
In Fig. 2, we provide a complete schematic block diagram of handwritten Indic script
identification showing different modules: feature extraction/selection and classification. In
our study, 10 different CNNs are used to select features from a variety of representations
of the studied image and we label each of them as CNNd,x,y . In every CNNd,x,y , d and x
respectively refer to domain representation and dimension of the studied image, and y refers
to the number of convolutional and pooling layers in that particular CNN. For example,
domain representation can be expressed as d = {s, f }, where s refers to spatial domain and
f , frequency. In our case, either of these is taken into account. Note that, with the use of
the Haar wavelet transform (HWT) (see Section 2.2), certain frequencies are removed. In
case of dimension (x), we have x = {[32 × 32], [48 × 48], [128 × 128]}, and for simplicity,
x = {32, 48, 128} is used. All of these dimensions signify resolution to which the input
images are scaled. In case of CL and PL, y = {2, 3}: one of the two is taken in CNN, and
y = 2 means that there are two pairs of convolutional and pooling layers in the CNNs.
In our model, two broad CNN architectures: CNNd,x,2 and CNNd,x,3 are used and can be
summarized as in Table 1.
Table 1. Architecture: CNNd,x,y , where y = {2, 3}.
Architecture
Parameter
CL1
PL1
CL2
Layer
PL2 CL3
CNNd,x,2
Channel
Filter size
Pad size
32
5×5
2
32
2×2
—
64
5×5
2
64
2×2
—
CNNd,x,3
Channel
Filter size
Pad size
32
7×7
3
32
2×2
—
64
5×5
2
64
2×2
—
PL3
FCL1
FCL2
Softmax
—
—
—
—
—
—
1024
—
—
512
—
—
11
—
—
128
3×3
1
128
2×2
—
1024
—
—
512
—
—
11
—
—
Index
CL = convolutional layer
PL = pooling layer
FCL = fully connected layer
1. CNNd,x,2 : We have six different layers, i.e. two CLs, two PLs and two FCLs. The first
two CLs are followed by PLs. In FCLs, the first layer takes image representation that
has been generated by CLs and PLs and reshapes them in the form of a vector, and the
second FCL produces a set of 1024 features.
2. CNNd,x,3 : Every CNN has eight different layers: three CLs, three PLs and two FCLs.
In general, such an architecture is very much similar to previously mentioned CNNd,x,2 .
The difference lies in additional pair of CL and PL that follows second pair of the same.
Like the CNNd,x,2 , these CNNs produce a set of 1024 features for any studied image.
Once again, the architectural details of aforementioned CNNs are summarized in Table 1
that follows schematic block diagram of the system (see Fig. 2). For more understanding,
in Fig. 3, we provide the activation maps for CNNs,128,3 . This means that it uses spatial
domain image representation with the dimensionality of 128 and three pairs of convolutional
and pooling layers in the CNNs.
2.2
Data representation
In general, since Fourier transform [11] may not work to successfully provide information
about what frequencies are present at what time, wavelet transform (WT) [12] and short
time Fourier transform [13] are typically used. Both of these help identify the frequency
components present in any signal at any given time. WT, on the other hand, can provide
dynamic resolution.
We consider an image a 2D time signal that can be resized. We then use multilevel 2D
discrete WT on a scaled/resized image (128×128) to generate frequency domain representation. To be precise, we use the Haar wavelet [14] with seven different level decomposition that
can generate approximated and detailed coefficients. Since the approximated coefficients are
equivalent to zero, we use the detailed coefficients in addition to modified approximated
coefficients to reconstruct the image. In the modified approximated coefficients, we consider
only high frequency components.
In our method, using a variety of different WTs, such as the Daubechies [15] and several
decomposition levels, the best results were observed with the Haar wavelet and a decomposition level of 7. This reconstructed image is further resized to 32 × 32 (x = 32) and 48 × 48
Fig. 3. Illustrating the activation maps for CNNs,128,3 : spatial domain image representation with the dimensionality of 128 and three-layered
convolutional and pooling layers.
(x = 48), and are fed into multiple CNNs as mentioned in Section 2.1. Like in the frequency
domain representation, spatial domain representations are resized/scaled to 32 × 32 (x =
32), 48 × 48 (x=48) and 128 × 128 (x =128) are fed into multiple CNNs.
3
3.1
Experiments
Dataset, and evaluation metrics and protocol.
To evaluate our proposed approach for word-level handwritten script identification, we have
considered a dataset named PHD Indic 11 [6]. It is composed of 11K scanned word images
(grayscale) from 11 different Indic script, i.e. 1K per script. A few samples are shown in
Fig. 4. For more information about dataset, we refer to recently reported work [6]. The
primary reason behind considering PHD Indic 11 dataset in our test is, no big size data has
been reported in the literature for research purpose, till this date.
Using the exact same CNN representation as before, Cd,x,y [i][j] represents the count
where an instance with label i is classified as j. Accuracy (acc) of the particular CNN can
then be computed as
P11
Cd,x,y [i][i]
accd,x,y = P11 i=1
.
P11
i=1
j=1 Cd,x,y [j][i]
Precision (prec) can be computed as
P11
i=1
precd,x,y =
precid,x,y
Cd,x,y [i][i]
and precid,x,y = P11
,
11
j=1 Cd,x,y [i][j]
where precid,x,y refers to precision for any i-th label. In a similar fashion, recall (rec) can be
computed as
P11
recd,x,y =
recid,x,y
Cd,x,y [i][i]
and recid,x,y = P11
,
11
j=1 Cd,x,y [i][j]
i=1
where recid,x,y refers to recall for any i-th label. Having both precision and recall, f-score can
be computed as
P11
f-scored,x,y =
i=1
precid,x,y × recid,x,y
f-scoreid,x,y
and f-scoreid,x,y = 2 ×
.
11
precid,x,y + recid,x,y
Following conventional 4:1, i.e. train:test evaluation protocol, we have separated 8.8K
images for training and the remaining 2.2K images for testing. We ran our experiments by
using a machine: GTX 730 with 384 CUDA cores and 4GB GPU RAM. Besides, it has Intel
Pentium Core2Quad Q6600 and 4GB RAM.
3.2
Experimental set up
As mentioned earlier, are are required to train the CNNs first before testing. In other words,
it is important to see how training and testing have been performed.
Fig. 4. Illustrating few samples from dataset named PHDIndic 11 used in our experiment.
Our CNNs in this study represented by CNNd,x,y are trained independently using the
training dataset (as mentioned earlier). To clarify once again, these CNNs has either two or
three pairs of consecutive convolutional and pooling layers. Besides, each of them has three
fully connected layers. The first of these three layers function as an input layer with number
of neurons that depends on the size of the input image specified by x. The second layer in
CNNs has 1024 neurons, and during training, we apply a dropout probability of 0.5. The
final layer has 11 neurons, whose outputs are used as input to an 11-way soft-max classifier
that provides us with classification/identification probabilities for each of 11 possible classes.
Our training set of 8.8K word images are split into multiple batches of 50 word images and
the CNNs are trained accordingly. For optimization, Adam optimizer [16] was used with
learning rate of 1 × 10−3 having default parameters: β1 = 0.9 and β2 = 0.999. This helps
apply gradients for loss to the weight parameters during back propagation. We computed
the accuracy of the CNN as training proceeds by taking ratio of the of images successfully
classified in the batch to the total number of images in the studied batch.
After training CNNs with 8.8K word images, we evaluated/tested each of them independently with the test set that is composed of 2.2K word images.
More specifically, for each input size specification, we have two CNNs, i.e. for domain (d)
and input size (x × x): CNNd,x,2 and CNNd,x,3 . Altogether, we have 10 different CNNs since
we have three different input sizes (x = {32, 48, 128}) for raw image and two different input
sizes (x = {32, 48}) for wavelet transformed image/data. For better understanding, we refer
readers to Fig. 2. Note that we have trained the CNNs to extract 1024 features from each
one, i.e. 120 × 1. These are then concatenated to form a single 10240 × 1 vector, where ten
different CNNs are employed. Like in the conventional machine learning classification, these
features are used for training and testing purpose using MLP classifier.
Fig. 5. Our results (in terms of accuracy, precision, recall and f-score) for all networks: CNNs and
their possible combinations.
3.3
Our results and comparative study
In this section, using dataset, and evaluation metrics (see Section 3.1), and experimental
setup (see Section 3.2), we summarize our results and comparative study as follows:
1. We provide results that have been produced from different architectures (CNNs), and
select the highest script identification rate from them; and
2. We then take highest script identification for a comparative study, where previous relevant works are considered.
Our results: Fig. 5 shows the comparison of the individual CNNs along with the effect of
combining them. The individual CNNd,x,y produced the maximum script identification rate
of 90% for CNNs,128,3 . As we ensemble two- and three-layered networks of corresponding
domain (d) and input size (x), we observe a positive correlation between input size and accuracy. CNNs,128 provides the maximum script identification rate of 91.41% in this category.
The primary reason behind the increase in accuracy is that the ensemble of two- and threelayered networks suggests that these networks complement each other. Further, to study an
effect of the spatial (s) and frequency (f) domain representation in CNNd,x,y , we ensemble
networks across all input sizes and depth of network. The spatial representation, CNNs ,
have produced script identification rate of 94.14% and the frequency domain representation,
CNNf , have escalated up to 90.27%. However, the frequency domain representations learning can be complimented and it has been clearly seen when we combined them. In their
combination, we have achieved the highest script identification rate of 94.73%. Since we
Fig. 6. Misclassified samples, where script names in the bracket are the actual scripts but, our
system identified them incorrectly. For example, in the first case, word image has been identified as
Gurumukhi and it actually is Bangla.
have not received 100% script identification rate, it is wise to provide a few samples where
our system failed to identify correctly (see Fig. 6).
Like we have mentioned in Section 3.1, we also provide precision, recall and f-score for
all architectures in Fig. 5. In what follows, the highest script identification rate, i.e. 94.73%
will be taken for a comparison.
Comparative study: For a fair comparison, widely used deep learning methods, such as
LeNet [7] and AlexNet [9] were taken. In addition, recently reported work on 11 handwritten
Indic script dataset [6] (including their baseline results) was considered. In Table 2, we
summarize the results. Our comparative study is focused on accuracy (not precision, recall
and f-score), since other methods reported accuracy, i.e. identification rate. Of course, in
Fig. 5, we are not just limited to accuracy.
In Table 2, our method outperforms all other methods. Precisely, it outperforms Obaidullah et al. [6] by 4.7%, LeNet [7] by 2.73% and AlexNet [9] by 2.61%.
4
Conclusion
In this paper, we have proposed a novel framework that uses convolutional neural networks
(CNNs) for feature extraction. In our method, in addition, to conventional spatial domain
representation, we have used multilevel 2D discrete Haar wavelet transform, where image
representations have been scaled to a variety of different sizes. Having these, several different
CNNs have been used to select features. With this, 11 different handwritten scripts: Bangla,
Devnagari, Gujarati, Gurumukhi, Kannada, Malayalam, Oriya, Roman, Tamil, Telugu and
Urdu, have been identified, where 1000 words per script are used. In our test, we have
achieved the maximum script identification rate of 94.73% using multi-layer perceptron
(MLP). To the best of our knowledge, this is the biggest data for Indic script identification
work. Considering the complexity and the size of the dataset, our method outperforms the
previously reported techniques.
Table 2. Comparative study.
Method
Obaidullah et al. [6]
(Hand-crafted features)
LeNet [7]
(CNN)
AlexNet [9]
(CNN)
Our method
(Multiscale CNN + WT)
Accuracy
91.00%
82.00%
92.14%
94.73%
References
1. Ghosh, D., Dube, T., Shivaprasad, A.: Script recognitiona review. IEEE Transactions on
pattern analysis and machine intelligence 32(12) (2010) 2142–2161
2. Pal, U., Jayadevan, R., Sharma, N.: Handwriting recognition in indian regional scripts: a survey
of offline techniques. ACM Transactions on Asian Language Information Processing (TALIP)
11(1) (2012) 1
3. Singh, P.K., Sarkar, R., Nasipuri, M., Doermann, D.: Word-level script identification for handwritten indic scripts. In: Document Analysis and Recognition (ICDAR), 2015 13th International
Conference on, IEEE (2015) 1106–1110
4. Hangarge, M., Santosh, K., Pardeshi, R.: Directional discrete cosine transform for handwritten
script identification. In: Document Analysis and Recognition (ICDAR), 2013 12th International
Conference on, IEEE (2013) 344–348
5. Pati, P.B., Ramakrishnan, A.: Word level multi-script identification. Pattern Recognition
Letters 29(9) (2008) 1218 – 1229
6. Obaidullah, S.M., Halder, C., Santosh, K., Das, N., Roy, K.: Phdindic 11: page-level handwritten document image dataset of 11 official indic scripts for script identification. Multimedia
Tools and Applications (2017) 1–36
7. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. Proceedings of the IEEE 86(11) (1998) 2278–2324
8. Roy, S., Das, N., Kundu, M., Nasipuri, M.: Handwritten isolated bangla compound character
recognition: A new benchmark using a novel deep learning approach. Pattern Recognition
Letters 90 (2017) 15–21
9. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: Advances in neural information processing systems. (2012) 1097–1105
10. Sarkhel, R., Das, N., Das, A., Kundu, M., Nasipuri, M.: A multi-scale deep quad tree based
feature extraction method for the recognition of isolated handwritten characters of popular
indic scripts. Pattern Recognition (2017)
11. Smith, S.W., et al.: The scientist and engineer’s guide to digital signal processing. (1997)
12. Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE
transactions on information theory 36(5) (1990) 961–1005
13. Portnoff, M.: Time-frequency representation of digital signals and systems based on short-time
fourier analysis. IEEE Transactions on Acoustics, Speech, and Signal Processing 28(1) (1980)
55–69
14. Sundararajan, D.: Fundamentals of the discrete haar wavelet transform. (2011)
15. Vonesch, C., Blu, T., Unser, M.: Generalized daubechies wavelet families. IEEE Transactions
on Signal Processing 55(9) (2007) 4415–4429
16. Kingma, D., Ba, J.:
arXiv:1412.6980 (2014)
Adam: A method for stochastic optimization.
arXiv preprint
| 1cs.CV
|
arXiv:1412.8054v2 [math.OC] 15 Nov 2016
Power Flow as an Algebraic System
Jakub Marecek, Timothy McCoy, and Martin Mevissen
∗
January 24, 2018
Abstract
Steady states of alternating-current (AC) circuits have been studied
in considerable detail. In 1982, Baillieul and Byrnes derived an upper
bound on the number of steady states in a loss-less AC circuit [IEEE
TCAS, 29(11): 724–737] and conjectured that this bound holds for AC
circuits in general. We prove this is indeed the case, among other results,
by studying a certain multi-homogeneous structure in an algebraisation.
1
Introduction
For more than 60 years [37], steady states of alternating-current (AC) circuits
have been studied in considerable detail. The key problem, sometimes known
as the power flow or load flow problem, considers complex voltages Vk at all
buses k as variables, except for one reference bus (k = 0), where the power
supplied. When one denotes complex admittance matrix Y , complex current
Ik , and complex power Sk at bus k, the steady-state equations are based on:
X
X
∗
∗
Sk = Vk Ik∗ = Vk
Yk,l
Vl∗ =
Yk,l
Vk Vl∗
(1)
l∈N
l∈N
where asterisk denotes complex conjugate. This captures the complex, nonconvex non-linear nature [16, 38, 11, 15, 18] of any problem in the AC model.
In order to obtain an algebraic system from (1), one needs to reformulate the
complex conjugate. In order to do so, one may replace all Vk∗ with independent
variables Uk , and filter for “real” solutions where Uk = Vk∗ = ℜVk − ℑVk ı once
the complex solutions are obtained. Thereby, we obtain a particular structure,
which allows us to prove a variety of results.
In particular, the main contributions of our paper are:
• a reformulation of the steady-state equations to a multi-homogeneous algebraic system
∗ J. Marecek and M. Mevissen are with IBM Research – Ireland, IBM Technology Campus
Damastown, Dublin D15, Ireland. e-mail: [email protected]. T. McCoy is with
Google.
1
• analytical results on the number and structure of feasible solutions considering losses, resolving a conjecture of Baillieul and Byrnes [2], which
has been open for over three decades
• empirical results for some well-known instances, including the numbers of
roots, conditions for non-uniqueness of optima, and tree-width.
Our analytical results rely on the work of Morgan and Sommese [32, 36] on multihomogeneous structures. Our empirical results rely on Bertini [6], a leading
implementation of homotopy-continuation methods. As we explain in Section
8, ours is not the first algebraisation of the system (1), cf. [37, 4, 3, 2], and
there is a long history [34, 14, 26, 23, 31, 29, 9, 39, 41] of the use of homotopycontinuation methods.
2
The Problem
In order to make the paper self-contained, we restate of the steady-state equations. Consider a circuit represented by an undirected graph (N, E), where
vertices n ∈ N are called buses and edges {l, m} ∈ E ⊆ N × N are called
branches, and an admittance matrix Y = G + Bı ∈ C|N |×|N |, where the real
part of an element is called conductance G = (glm ) and the imaginary part
susceptance B = (blm ). Each bus k ∈ N is associated with complex voltage
Vk = ℜVk + ℑVk ı, complex current Ik = ℜIk + ℑIk ı, and power Sk = Pk + Qk ı
demanded or generated. Let 0 ∈ N correspond to a reference bus with phase
ℑV0 = 0 and magnitude |V0 | fixed; powers at all other buses are fixed too. (In
a variety of extensions, there are other buses, denoted generators, where voltage magnitude, but not the phase and not the power is fixed.) Each branch
(l, m) ∈ E is associated with the complex power Slm = Plm + Qlm ı. The
key constraint linking the buses is Kirchhoff’s current law, which stipulates the
sum of the currents injected and withdrawn at each bus is 0. Considering the
2
relationship I = Y V , the steady state equations hence are:
Pkg = Pkd + ℜVk
n
X
(ℜyik ℜVi − ℑyik ℑVi )
i=1
+ ℑVk
n
X
(ℑyik ℜVi + ℜyik ℑVi )
(2)
i=1
Qgk = Qdk + ℜVk
n
X
(−ℑyik ℜVi − ℜyik ℑVi )
i=1
+ ℑVk
n
X
(ℜyik ℜVi − ℑyik ℑVi )
(3)
i=1
Plm = blm (ℜVl ℑVm − ℜVm ℑVl )
(4)
2
2
+ glm (ℜVl + ℑVm − ℑVl ℑVm − ℜVl ℜVm )
Qlm = blm (ℜVl ℑVm − ℑVl ℑVm − ℜVl 2 − ℑVl 2 )
+ glm (ℜVl ℑVm − ℜVm ℑVl − ℜVm ℑVl )
−
b̄lm
(ℜVl 2 + ℑVl 2 )
2
(5)
Additionally, one can optimise a variety of objectives over the steady states.
In one commonly used objective function, one approximates the costs of real
power P0 generated at the reference bus 0 by a quadratic function f0 :
cost := f0 (P0 ).
(6)
(In a variety of extensions, in which there are other buses where the power
is not fixed, there would be a quadratic function for each such bus and the
quadratic function of power would be summed across all of these buses.) In
the Lp -norm loss objective, one computes a norm of the vector D obtained by
summing apparent powers S(u, v) + S(v, u)∀(u, v) ∈ E for:
||D||p =
X
|S(u, v) + S(v, u)|
(u,v)∈E
p
1/p
.
(7)
The usual ||D||1 is denoted loss below. We consider these objectives only in
Section VII, while our results in Sections IV–VI apply independently of the use
of any objective function whatsoever.
3
Definitions from Algebraic Geometry
In order to state our results, we need some definitions from algebraic geometry.
While we refer the reader to [2, 27] for the basics, we present the concepts
introduced in the past three decades, not yet widely covered by textbooks. For
a more comprehensive treatment, please see [32, 36, 35].
3
Let n ≥ 0 be an integer and f (z) be a system of n polynomial equations in
z ∈ C n with support (A1 , . . . , An ):
P
α1 α2
αn
α∈A1 f1α z1 z2 · · · zn
f1 (z) =
..
(8)
. P
α1 α2
αn
fn (z) =
α∈An fnα z1 z2 · · · zn ,
where coefficients fiα are non-zero complex numbers. It is well known [12] that
the polynomials define n projective hypersurfaces in a projective space CPn .
Bézout theorem states that either hypersurfaces intersect in an infinite set
with some component of positive dimension, or the number of intersection
points, counted with multiplicity, is equal to the product d1 · · · dn , where di
is the degree of polynomial i. We call the product d1 · · · dn the usual Bézout
number.
The usual Bézout number can be improved by considering:
Definition 1 (Structure). Any partition of the index set {1, . . . , n} into k sets
I1 , . . . , Ik defines a structure. There, Zj = {zi : i ∈ Ij } is known as the group
of variables for each set Ij . The associated degree dij of a polynomial fi with
respect to group Zj is
X
def
αl .
(9)
dij = max
α∈Ai
l∈Ij
We say that fi has multi-degree (di1 , . . . , din ).
Whenever for some j, for all i, the same dij is attained for all α ∈ Ai , we
call the system (11) homogeneous in the group of variables Zj . The projective
space associated to the group of variables Zj in a structure has dimension
|Ij | − 1 if (11) is homogeneous in Zj , and
def
aj =
(10)
|Ij |
otherwise.
Pk
Definition 2 (Multi-homogeneous Bézout Number). Assuming n = j=1 aj ,
the multi-homogeneous Bézout number Béz(A1 , . . . , An ; I1 , . . . , Ik ) is defined as
Qk
a
the coefficient of the term j=1 ζj j , where aj is the associated dimension (10),
Q n Pk
within the polynomial i=1 j=1 dij ζj , in variables ζj , j = 1 . . . k with coefficients dij are the associated degrees (9); that is (d11 ζ1 + d12 ζ2 + . . . + d1k ζk )
(d21 ζ1 + d22 ζ2 + . . . + d2k ζk ) · · · (d2n ζ1 + d2n ζ2 + . . . + dnk ζk ).
Consider the example of Wampler
p1 (z) =
p2 (z) =
p3 (z) =
[42] in x ∈ C 3 :
x21 + x2 + 1,
x1 x3 + x2 + 2,
x2 x3 + x3 + 3,
(11)
with the usual Bézout number of 8. Considering the partition {x1 , x2 }, {x3 },
where d11 = 2, d12 = 0, d21 = d22 = d31 = d32 = 1. the monomial ζ12 ζ21 is
4
to be looked up in the polynomial 2ζ1 (ζ1 + ζ2 )2 . The corresponding multihomogeneous Bézout number is hence 4 and this is the minimum across all
possible structures.
In general, the multi-homogeneous Bézout number Béz(A1 , . . . , An ; I1 , . . . , Ik )
is an upper bound on the number of isolated roots of (11) in CPa1 × · · · × CPak ,
and thereby an upper bounds the number of isolated finite complex roots of (11).
There are a variety of additional methods for computing the multi-homogeneous
Bézout number, e.g., [42]. In the particular case where A = A1 = · · · = An , we
denote
Y
k
n
def
a
(12)
dj j ,
Béz(A1 , . . . , An ; I1 , . . . , Ik ) =
a1 a2 · · · ak
j=1
where dj = dij (equal for each i) and the multinomial coefficient
n!
n
def
=
a1 a2 · · · ak
a1 ! a2 ! · · · ak !
Qk
Pk
is the coefficient of j=1 ζjak in (ζ1 + · · · + ζk )n with n = j=1 aj , as above.
In summary, the multi-homogeneous Bézout number provides a sharper
bound on the numberQof isolated solutions of a system of equations than the
n
usual Bézout number i=1 di = d1 · · · dn . In the famous example of the eigenvalue problem [43], it is known that the Bézout number is 2n , whereas there
exists a structure with multi-homogeneous Bézout number of n. We hence
study the multi-homogeneous structure within the steady state equations of
alternating-current circuits.
4
The Multi-Homogeneous Structure
Notice that in order to obtain an algebraic system from the steady-state equations (2–5), one needs to reformulate the complex conjugate. In order to do so,
one may replace all vn∗ with independent variables un , and later filter for solutions where un = vn∗ once the complex solutions are obtained. We denote such
solution “real”. Let G be the set of slack generators for which |vn | is specified,
and assume 0 ∈ G corresponds to a reference node with phase 0. Notice that
the use of variables vn and un produces a multi-homogeneous structure with
variable groups {vn } and {un }:
vn
X
Yn,k uk + un
X
Yn,k uk − un
k
vn
k
X
∗
Yn,k
vk = 2pn
n∈N \G
X
∗
Yn,k
vk = 2qn
n∈N \G
k
k
vn un = |vn |2
v0 = |v0 |, u0 = |v0 |
n ∈ G − {0}
(13)
5
For example for the two-bus network, we obtain:
∗
∗
v1 (Y1,0 u0 + Y1,1 u1 ) + u1 (Y1,0
v0 + Y1,1
v1 ) = 2p1
∗
∗
v1 (Y1,0 u0 + Y1,1 u1 ) − u1 (Y1,0
v0 + Y1,1
v1 ) = 2q1
v0 = |v0 |, u0 = |v0 |
(14)
Using the algebraic system, one can formulate a number of structural results
concerning power flows.
5
An Analysis for s = 1
For the particular multi-homogeneous structure, which is the partition of the
variables into several groups in (13), we can bound the number of isolated
solutions:
Theorem 1. With exceptions on a parameter set of measure zero, the alternatingcurrent power flow (13) has a finite number of complex solutions, which is
bounded above by:
2n − 2
(15)
n−1
Proof. Each equation in the system (13) is linear in the V variables and also
in the V ∗ variables, giving rise to a natural multi-homogeneous structure of
mult-idegree (1, 1). Since the slack bus voltage is fixed at a reference value, the
system has 2n − 2 such equations in (n − 1, n − 1) variables. By the multihomogeneous form of Bézout’s Theorem (see e.g. Theorem 8.4.7 in [36]), the
total number of solutions in multi-projective space CPn−1 × CPn−1 is precisely
the stated bound, counting multiplicity. Some subset of these lie on the affine
patch C n−1 × C n−1 ⊂ CPn−1 × CPn−1 , giving the result.
Notice that this applies also to some well-known instances of alternatingcurrent optimal power flows (ACOPF). For example, the instances of Lesieutre
et al. [22] and Bukhsh et al. [8] have only a single “slack” bus, whose active
and reactive powers are not fixed, and hence the result applies. Notice that the
exception of measure-zero set is necessary; cf. Example 4.1 of [2].
As we will illustrate in the next section, this bound is tight in some cases.
Deciding whether the bound on the number of roots obtained using a particular
multi-homogeneous structure is tight for a particular instance is, nevertheless,
hard. This can be seen from:
Theorem 2 (Theorem 1 of Malajovich and Meer [27]). There does not exist
a polynomial time algorithm to approximate the minimal multi-homogeneous
Bézout number for a polynomial system (11) up to any fixed factor, unless P =
NP.
6
We can, however, show there exists a certain structure among these solutions:
Corollary 1. If there exists a feasible solution of the alternating-current power
flow, then the solution has even multiplicity greater or equal to 2 or another
solution exists.
Proof. The finite number of solutions to the power flows problem of Theorem 1 is
even. Observe that (U, Û ) is a solution of the system (13) if and only if (Û ∗ , U ∗ )
is a solution. This implies that the non-“real” solutions, that is solutions for
which U ∗ 6= Û , necessarily come in pairs. It follows that “real” feasible solutions
are also even in number, counting multiplicity. The result follows.
Note that a solution having multiplicity greater than 1 is a special case that
is highly unlikely in a real system. Moreover, it is easily detected, since the
Jacobian at a solution is nonsingular if and only if the solution has multiplicity
1.
6
Alternating-Current Optimal Power Flow
One may also make the following observations about the alternating-current
optimal power flows, i.e., the problem of optimising an objective over the steady
state:
Remark 1. For the alternating-current power flows, where powers are fixed at
all but the reference bus, whenever there exists a real feasible solution, except
for a parameter set of measure zero, one can enumerate all feasible solutions in
finite time.
Indeed: By Theorem 1, we know there exist a finite number of isolated
solutions to the system (13). By the homotopy-continuation method of Sommese
et al. [36, 6], we can enumerate the roots with probability 1, which allows us
to pick the global optimum, trivially. Notice that Bertini, the implementation
of the method of Sommese [6], makes it possible to check that all roots are
obtained. Notice that the addition of inequalities can be accommodated by
filtering the real roots.
Nevertheless, this method is not practical, as there may be too many isolated
solutions to enumerate. Generically, this is indeed the case, whenever there
are two or more generators with variable output, i.e., buses, whose active and
reactive power is not fixed:
Corollary 2. In the alternating-current optimal power flow problem, i.e., with
s > 1, where powers are variable outside of the reference bus and there are no additional inequalities, the complex solution set is empty or positive-dimensional,
except for a parameter set of measure zero. When the complex solution set
is positive-dimensional, if a smooth real feasible solution exists, then there are
infinitely many real feasible solutions.
7
Proof. For each slack bus after the first, the system has two variables but only
one equation. With s + 1 slack buses, the rank of the Jacobian and hence the
dimension of the complex solution set will be at least s by Lemma 13.4.1 in [36].
Furthermore, if a real feasible solution U exists and the solution set is smooth
at this point, then the local dimension of the complex and real solution sets are
equal at U . Therefore, since the complex solution set is positive-dimensional,
so is the real set at U , and so infinitely many real feasible solutions exist.
Although there are a variety of methods for studying positive-dimensional
systems, including the enumeration of a point within each connected component
[5, 33] and studying the critical points of the restriction to the variety of the
distance function to such points [1], we suggest the method of moments [20, 13]
may be more suitable for studying the feasible set of alternating-current optimal
power flows. It has been shown recently [13] that it allows for very small errors
on systems in dimension of over five thousand.
As is often the case in engineering applications, one may also be interested
in the distance of a point to the set of feasible solutions. Again, by considering
our algebraisation, one can bound the probability this distance is small using
the theorem of Lotz [25] on the zero-set V of multivariate polynomials. This
could be seen as the converse of the results of [24].
8
Method
No. |N | of buses
3
4
5
6
7
8
9
10
11
12
13
14
9
Bézout’s upper bound
A BKK-based upper bound
Theorem 1
16
8
6
64
40
20
256
192
70
1024
864
252
4096
3712
924
16384
15488
3432
65536
63488
12870
262144
257536
48620
1048576
1038336
184756
4194304
4171776
705432
16777216
16728064
2704156
67108864
67002368
10400600
Generic lower bound
6
20
70
252
924
3432
12870
48620
184756
705432
2704156
10400600
Table 1: The maximum number of steady states in a circuit with a fixed number of buses.
Treewidth tw(P )
No. |X| of solutions
minx∈X (cost(x))
No. of min. wrt. cost
avgx∈X (cost(x))
maxx∈X (cost(x))
minx∈X (loss(x))
No. of min. wrt. loss
avgx∈X (loss(x))
maxx∈X (loss(x))
Bukhsh et al. [8]
Klos and Wojcicka [18]
Lavaei and Low [21]
Bukhsh et al. [8]
McCoy et al.
McCoy et al.
Bukhsh et al. [8]
Grainger and Stevenson [45]
Lesieutre et al. [22]
McCoy et al.
McCoy et al.
McCoy et al.
Bukhsh et al. [8]
McCoy et al.
McCoy et al.
McCoy et al.
Bukhsh et al. [8]
Chow [45]
McCoy et al.
McCoy et al.
McCoy et al.
McCoy et al.
McCoy et al.
McCoy et al.
No. |E| of branches
10
case2w
case3KW
case3LL
case3w
case4
case4ac
case4cyc
case4gs
case5w
case6ac
case6ac2
case6b
case6cyc
case6cyc2
case6cyc3
case7
case8cyc
case9
case9g
caseK4
caseK4sym
caseK6
caseK6b
caseK6sym
Source
No. |N | of buses
Instance
2
3
3
3
4
4
4
4
5
6
6
6
6
6
6
7
8
9
9
4
4
6
6
6
1
3
3
2
3
4
4
4
6
6
6
6
6
6
6
7
8
9
9
6
6
15
15
15
1
2
2
1
1
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
3
5
5
5
2
6
2
2
4
6
4
10
2
23
22
30
30
12
12
2
60
16
2
8
6
36
40
48
8.42
-0.0
1502.07
5.88
1502.51
2005.25
3001.61
2316.37
2003.57
4005.44
4005.32
3005.7
3005.7
1506.87
4502.42
1667.31
6506.11
3504.44
2604.41
2008.62
2009.28
2018.2
6002.79
6003.01
1
1
1
1
1
2
1
1
1
1
1
5
4
1
1
1
1
1
1
1
2
1
1
1
9.04
1250.0
1502.19
5.94
1505.19
3337.5
3003.56
3681.07
2004.6
5574.23
5554.03
3606.3
3606.3
3131.55
4505.95
1668.36
8557.72
5692.02
4103.36
3006.32
3339.29
5075.47
6017.01
6017.75
9.66
1500.0
1502.31
6.01
1507.88
4005.51
3005.52
4595.8
2005.63
6013.26
6012.82
4508.28
4508.28
4508.24
4508.01
1669.41
9511.04
6505.02
5602.31
4005.51
4005.91
6030.0
6019.41
6019.86
0.71
-0.0
0.22
0.55
0.11
0.01
0.01
0.37
0.47
0.02
0.02
0.01
0.01
0.03
0.02
0.08
0.01
0.03
0.08
0.01
0.01
0.02
0.01
0.01
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1.02
-0.0
0.34
0.58
2.79
2.44
1.96
27.24
1.5
6.26
5.97
3.9
3.9
4.15
3.55
0.26
4.52
1.37
0.26
4.6
3.96
17.16
14.23
14.75
1.33
0.0
0.46
0.61
5.48
3.78
3.92
39.72
2.53
10.51
10.21
5.88
5.88
6.56
5.61
0.45
7.84
1.87
0.45
7.03
7.28
27.25
16.62
16.86
Table 2: Properties of the instances tested.
variable_group V0, V1; variable_group U0, U1;
I0 = V0*Yv0_0 + V1*Yv0_1; I1 = V0*Yv1_0 + V1*Yv1_1;
J0 = U0*Yu0_0 + U1*Yu0_1; J1 = U0*Yu1_0 + U1*Yu1_1;
fv0 = V0 - 1.0; fv1 = I1*U1 + J1*V1 + 7.0;
fu0 = U0 - 1.0; fu1 = -I1*U1 + J1*V1 - 7.0*I;
Figure 1: A Bertini encoding of ACPF on the two-bus instance of Bukhsh et al.
2012, where the impedance of a single branch is 0.04 + 0.2i.
7
Computational Illustrations
In order to illustrate Theorem 1, we first present the maximum number of
steady states in a circuit with a fixed number of buses in Table , and compare
it to the values of our upper bound, Bézout-based upper bound, and the BKKbased upper bound [10, 31]. Notice that the maximum number of steady states
in a circuit with a fixed number of buses is achieved when (N, E) is a clique.
Notice further that the generic lower bound, obtained as the number of solutions
found by tracing the paths, matches the upper bound of Theorem 1 throughout
Table 7.
To illustrate Proposition 1, we have enumerated the steady states using
Bertini, a versatile package for homotopy-continuation methods by Sommese
et al. [36]. See Figure 1 for an example of Bertini input corresponding to
the example above (14), with constants Yu i j representing Yi,j and Yv i j
∗
representing Yi,j
. The results are summarised in Table .
To illustrate Theorem 1 further, we present the values of our upper bound on
a collection of instances widely known in the power systems community. The
instances are mostly available from the Test Case Archive of Optimal Power
Flow (OPF) Problems with Local Optima of Bukhsh1 , while some have appeared
in well-known papers, e.g. [18], and some are available in recent distributions of
Matpower [45], a well-known benchmark. In particular, we present the numbers
of distinct roots of the instances. In all cases, where Theorem 1 applies, the
number of solutions found by tracing the paths matches the upper bound of
Theorem 1, certifying the completeness. In other cases, one could rely on Bertini
certificates of completeness of the search.
Empirically, we observe there exists a unique global optimum in all these instances tested with respect to the L1 -loss objective. For the generation cost objective, however, there are a number of instances (case4ac, caseK4sym, case6b,
case6cyc), where the global optima are not unique. The case of caseK4sym is a
particularly good illustration, where the symmetry between two generators and
two demand nodes in a complete graph results in multiple global optima.
In order to provide material for further study of structural properties, we
present tree-width of the instances in Table 7 in column tw(N ). Notice that
for the well-known small instances, tree-width is 1 or 2, e.g., 1 for the instance
1 http://www.maths.ed.ac.uk/OptEnergy/LocalOpt/ ,
11
accessed November 30th, 2014.
in Figure 1, and 2 for the instance of Lesieutre et al. As the instances grow,
however, this need not be the case: Kloks [17] shows treewidth is not bounded
even in sparse random graphs, with high probability. In complete graphs, such
as caseK4sym with tree-width 3 above, tree-width grows linearly in the number
of vertices.
8
Related Work
There is a long history of study of the number and structure of solutions of
power flows [37, 4, 3, 2]. [37] considered the Bézout bound. [10, 31] considered
a bound based on the work of Bernstein [7] and Kushnirenko [19]. [2, 4] derived
the same expression as in Theorem 1 using intersection theory, but in the lossless
AC model. They highlight that the number of solutions in an alternatingcurrent model with losses is an important open problem. We note that Theorem
1 subsumes Theorem 4.1 of [2] as a special case. Finally, we note that [18]
present a lower bound without a proof and recent papers [39, 44] bound certain
distinguished solutions, but not all solutions; cf. [28].
There is also a long history of applications of homotopy-continuation methods in power systems [34, 14, 26, 23], although often, e.g. in [26], the set-up
of the homotopy restricted the methods to a heuristic, which could not enumerate all the solutions of the power flow [30]. Recently, these have attracted
much interest [31, 29, 9] following the work of Trias [39, 40, 41]. See [28] for an
overview.
9
Conclusions
We hope that the structural results provided will aid the development of faster
solvers for the related non-linear problems [15]. Arguably, one could:
• By using Theorem 1 in the construction of start systems for homotopycontinuation methods [43], allow for larger zero-dimensional systems to be
studied.
• Extend Corollary 2 to finding at least one point in each connected component [33, 1].
• Extend the homotopy-continuation methods to consider inequalities within
the tracing, rather than only in the filtering phase, which could improve
their computational performance considerably.
• Develop methods for the optimal power flow problem, whose complexity
would be superpolynomial only in the tree-width and the number of buses.
The latter two may be some of the most important challenges within the analysis
of circuits and systems.
12
Acknowledgements Parts of this work have been done while Tim was visiting IBM Research. Jakub would like to thank Isaac Newton Institute for
Mathematical Sciences at the University of Cambridge for their generous support for his visits. Dhagash Mehta has kindly provided a variety of suggestions
as the related work.
References
[1] P. Aubry, F. Rouillier, and M. S. E. Din. Real solving for positive dimensional systems. Journal of Symbolic Computation, 34(6):543 – 560, 2002.
[2] J. Baillieul and C. Byrnes. Geometric critical point analysis of lossless power system models. IEEE Transactions on Circuits and Systems,
29(11):724–737, Nov 1982.
[3] J. Baillieul and C. I. Byrnes. Remarks on the number of solutions to the
load flow equations for a power system with electrical losses. In Decision
and Control, 1982 21st IEEE Conference on, pages 919–924, Dec 1982.
[4] J. Baillieul, C. I. Byrnes, and R. B. Washburn. An algebraic-geometric and
topological analysis of the solution to the load-flow equations for a power
system. In Decision and Control including the Symposium on Adaptive
Processes, 1981 20th IEEE Conference on, pages 1312–1320, Dec 1981.
[5] S. Basu, R. Pollack, and M.-F. Roy. A new algorithm to find a point in
every cell defined by a family of polynomials. In Quantifier elimination and
cylindrical algebraic decomposition, pages 341–350. Springer Vienna, 1998.
[6] D. Bates, J. Hauenstein, A. Sommese, and C. Wampler. Numerically Solving Polynomial Systems with Bertini:. Software, Environments, and Tools.
Society for Industrial and Applied Mathematics, 2013.
[7] D. N. Bernstein. The number of roots of a system of equations. Funkcional.
Anal. i Prilozen., 9(3):1–4, 1975.
[8] W. Bukhsh, A. Grothey, K. McKinnon, and P. Trodden. Local solutions
of the optimal power flow problem. IEEE Trans. Power Syst., 28(4):4780–
4788, 2013.
[9] S. Chandra, D. Mehta, and A. Chakrabortty. Equilibria analysis of power
systems using a numerical homotopy method. In Power Energy Society
General Meeting, 2015 IEEE, pages 1–5, July 2015.
[10] T. Chen and D. Mehta. On the Network Topology Dependent Solution
Count of the Algebraic Load Flow Equations. 2015. ArXiv e-prints
1512.04987.
13
[11] H.-D. Chiang, C.-W. Liu, P. P. Varaiya, F. F. Wu, and M. G. Lauby.
Chaos in a simple power system. IEEE Transactions on Power Systems,
8(4):1407–1417, Nov 1993.
[12] W. Fulton. Intersection Theory. Springer New York, Jan. 1998.
[13] B. Ghaddar, J. Marecek, and M. Mevissen. Optimal power flow as a
polynomial optimization problem. IEEE Transactions on Power Systems,
31(1):539–546, Jan 2016.
[14] S. X. Guo and F. M. A. Salam. The real homotopy-based method for
computing solutions of electric power systems. In Circuits and Systems,
1992. ISCAS ’92. Proceedings., 1992 IEEE International Symposium on,
volume 6, pages 2737–2740 vol.6, May 1992.
[15] I. A. Hiskens. Analysis tools for power systems-contending with nonlinearities. Proceedings of the IEEE, 83(11):1573–1587, Nov 1995.
[16] I. A. Hiskens and R. J. Davy. Exploring the power flow solution space
boundary. IEEE Transactions on Power Systems, 16(3):389–395, Aug 2001.
[17] T. Kloks. Only few graphs have bounded treewidth. In T. Kloks, editor,
Treewidth, volume 842 of Lecture Notes in Computer Science, pages 51–60.
Springer Berlin Heidelberg, 1994.
[18] A. Klos and J. Wojcicka. Physical aspects of the nonuniqueness of load flow
solutions. International Journal of Electrical Power and Energy Systems,
13(5):268–276, 1991.
[19] A. Kushnirenko. Newton polytopes and the Bézout theorem. Funct. Anal.
Appl., 10:233–235, 1976.
[20] J. B. Lasserre. Convergent sdp-relaxations in polynomial optimization with
sparsity. SIAM Journal on Optimization, 17(3):822–843, 2006.
[21] J. Lavaei and S. Low. Zero duality gap in optimal power flow problem.
Power Systems, IEEE Transactions on, 27(1):92 –107, feb. 2012.
[22] B. Lesieutre, D. Molzahn, A. Borden, and C. DeMarco. Examining the
limits of the application of semidefinite programming to power flow problems. In Communication, Control, and Computing (Allerton), 2011 49th
Annual Allerton Conference on, pages 1492–1499, 2011.
[23] C.-W. Liu, C.-S. Chang, J. A. Jiang, and G. H. Yeh. Toward a cpflow-based
algorithm to compute all the type-1 load-flow solutions in electric power
systems. IEEE Transactions on Circuits and Systems I: Regular Papers,
52(3):625–630, March 2005.
14
[24] C.-W. Liu and J. S. Thorp. A novel method to compute the closest unstable
equilibrium point for transient stability region estimate in power systems.
IEEE Transactions on Circuits and Systems I: Fundamental Theory and
Applications, 44(7):630–635, Jul 1997.
[25] M. Lotz. On the volume of tubular neighborhoods of real algebraic varieties.
Proc. Amer. Math. Soc., 143:1875–1889, 2015.
[26] W. Ma and J. S. Thorp. An efficient algorithm to locate all the load
flow solutions. IEEE Transactions on Power Systems, 8(3):1077–1083, Aug
1993.
[27] G. Malajovich and K. Meer. Computing minimal multi-homogeneous
Bézout numbers is hard. Theory of Computing Systems, 40(4):553–570,
2007.
[28] D. Mehta, D. K. Molzahn, and K. Turitsyn. Recent advances in computational methods for the power flow equations. In 2016 American Control
Conference (ACC), pages 1753–1765, July 2016.
[29] D. Mehta, H. D. Nguyen, and K. Turitsyn. Numerical polynomial homotopy
continuation method to locate all the power flow solutions. IET Generation,
Transmission Distribution, 10(12):2972–2980, 2016.
[30] D. K. Molzahn, B. C. Lesieutre, and H. Chen. Counterexample to a
continuation-based algorithm for finding all power flow solutions. IEEE
Transactions on Power Systems, 28(1):564–565, Feb 2013.
[31] D. K. Molzahn, D. Mehta, and M. Niemerg. Toward topologically based
upper bounds on the number of power flow solutions. In 2016 American
Control Conference (ACC), pages 5927–5932, July 2016.
[32] A. Morgan and A. Sommese. A homotopy for solving general polynomial
systems that respects m-homogeneous structures. Appl. Math. Comput.,
24(2):101–113, 1987.
[33] F. Rouillier, M.-F. Roy, and M. S. E. Din. Finding at least one point in each
connected component of a real algebraic set defined by a single equation.
Journal of Complexity, 16(4):716–750, 2000.
[34] F. M. A. Salam, L. Ni, S. Guo, and X. Sun. Parallel processing for the
load flow of power systems: the approach and applications. In Decision
and Control, 1989., Proceedings of the 28th IEEE Conference on, pages
2173–2178 vol.3, Dec 1989.
[35] I. R. Shafarevich. Basic algebraic geometry. Springer-Verlag, springer study
edition edition, 1977. Translated from the Russian by K. A. Hirsch; Revised printing of Grundlehren der mathematischen Wissenschaften, Vol.
213, 1974.
15
[36] A. J. Sommese and C. W. Wampler. The numerical solution of systems of
polynomials arising in Engineering and Science. World Scientific, 2005.
[37] C. J. Tavora and O. J. M. Smith. Equilibrium analysis of power systems.
IEEE Transactions on Power Apparatus and Systems, PAS-91(3):1131–
1137, May 1972.
[38] J. S. Thorp and S. A. Naqavi. Load flow fractals. In Decision and Control,
1989., Proceedings of the 28th IEEE Conference on, pages 1822–1827 vol.2,
Dec 1989.
[39] A. Trias. The holomorphic embedding load flow method. In Power and
Energy Society General Meeting, 2012 IEEE, pages 1–8, July 2012.
[40] A. Trias. System and method for monitoring and managing electrical power
transmission and distribution networks, US Patent 7519506 and 7979239,
2009 and 2010, number = US 7519506 and 7979239, type = Patent, version
= , location = US,.
[41] A. Trias and J. L. Marı́n. The holomorphic embedding loadflow method for
dc power systems and nonlinear dc circuits. IEEE Transactions on Circuits
and Systems I: Regular Papers, 63(2):322–333, Feb 2016.
[42] C. W. Wampler. Bezout number calculations for multi-homogeneous polynomial systems. Applied Mathematics and Computation, 51(2):143 – 157,
1992.
[43] C. W. Wampler. An efficient start system for multi-homogeneous polynomial continuation. Numerische Mathematik, 66(4):517–524, 1993/94.
[44] T. Wang and H. D. Chiang. On the number of system separations in
electric power systems. IEEE Transactions on Circuits and Systems I:
Regular Papers, 63(5):661–670, May 2016.
[45] R. Zimmerman, C. Murillo-Sánchez, and R. Thomas. Matpower: Steadystate operations, planning, and analysis tools for power systems research
and education. Power Systems, IEEE Transactions on, 26(1):12–19, 2011.
16
| 5cs.CE
|
FESTUNG: A MATLAB / GNU Octave toolbox for the discontinuous Galerkin
method. Part II: Advection operator and slope limiting
Balthasar Reutera , Vadym Aizingera,∗, Manuel Wielanda , Florian Frankb , Peter Knabnera
a Friedrich–Alexander
arXiv:1602.05754v1 [math.NA] 18 Feb 2016
b Rice
University of Erlangen–Nürnberg, Department of Mathematics, Cauerstraße 11, 91058 Erlangen, Germany
University, Department of Computational and Applied Mathematics, 6100 Main Street – MS 134, Houston, TX 77005-1892, USA
Abstract
This is the second in a series of papers on implementing a discontinuous Galerkin (DG) method as an open source
MATLAB / GNU Octave toolbox. The intention of this ongoing project is to offer a rapid prototyping package for
application development using DG methods. The implementation relies on fully vectorized matrix / vector operations
and is comprehensively documented. Particular attention was paid to maintaining a direct mapping between discretization terms and code routines as well as to supporting the full code functionality in GNU Octave. The present
work focuses on a two-dimensional time-dependent linear advection equation with space / time-varying coefficients,
and provides a general order implementation of several slope limiting schemes for the DG method.
Keywords: MATLAB, GNU Octave, discontinuous Galerkin method, slope limiting, vectorization, open source,
advection operator
1. Introduction
The development milestones for the MATLAB / GNU Octave toolbox FESTUNG (F inite E lement S imulation
T oolbox for UN structured G rids) available at [1, 2] run somewhat counter to the history of the development of the
discontinuous Galerkin (DG) methods. Thus, our first paper in series [3] introduced a local discontinuous Galerkin
discretization for a time-dependent diffusion equation using the numerical methods introduced in [4]. The current
work, however, enhances the package with the functionality for purely hyperbolic equations—namely the original
purpose of the DG method proposed by Reed and Hill in [5] and analyzed by Johnson and Pitkäranta in [6]. The
reason behind this time inversion is that the numerical and software development technology necessary to produce a
fully functional DG solver for hyperbolic equations has to include upwind fluxes and slope limiters—both tasks more
complicated to solve in a computationally efficient manner than those needed for a pure diffusion equation.
The continued development of this toolbox still adheres to the same design principles declared in [3]:
1. Design a general-purpose software package using the DG method for a range of standard applications and
provide this toolbox as a research and learning tool in the open source format (cf. [1]).
2. Supply a well-documented, intuitive user-interface to ease adoption by a wider community of application and
engineering professionals.
3. Relying on the vectorization capabilities of MATLAB / GNU Octave, optimize the computational performance
of the toolbox components and demonstrate these software development strategies.
4. Maintain throughout full compatibility with GNU Octave to support users of open source software.
∗ Corresponding
author
Email addresses: [email protected] (Balthasar Reuter), [email protected] (Vadym Aizinger),
[email protected] (Manuel Wieland), [email protected] (Florian Frank), [email protected]
(Peter Knabner)
Preprint submitted to Elsevier
September 15, 2017
We refer to [3] for a literature review on DG methods and open source packages offering a DG capability. The
present work expands the functionality of the numerical solver published in the first paper in series by adding linear
advection terms and vertex-based slope limiters of general order. The latter development is particularly interesting,
since, to the best of our knowledge, no closed form description of vertex-based slope limiters for general order discretizations are to be found in the literature, even less so implementations of such limiters. In addition to hierarchical
vertex-based limiters of Kuzmin [7], this publication and the accompanying code includes an extension of the standard
linear vertex-based slope limiter to general order discretizations and a new scheme based on the hierarchical vertexbased limiter but using a stricter limiting strategy. Further additions in this work include a selection of TVD (total
variation diminishing) Runge–Kutta methods of orders one, two, and three employed for time discretization instead
of a simple implicit Euler method used in the first paper.
The rest of this paper is organized as follows: We introduce the model problem in the remainder of this section
and describe its discretization using the DG method in Sec. 2. Section 3 introduces slope limiting algorithms, first, for
linear DG discretizations followed by the general order case. Implementation specific details such as reformulation
and assembly of matrix blocks as well as numerical results are given in Sec. 4. All routines mentioned in this work
are listed and documented in Sec. 5. Section 6 concludes the work and gives future perspectives.
1.1. Model problem
Let J ≔ (0, tend ) be a finite time interval and Ω ⊂ R2 a polygonally bounded domain with boundary ∂Ω. We
consider the advection equation in conservative form
∂t c(t, x) + ∇ · u(t, x) c(t, x) = f (t, x)
in J × Ω
(1a)
with time / space-varying coefficients u : J × Ω → R2 and f : J × Ω → R. A prototype application of (1a) is the
advective transport in fluids, i. e., the movement of a solute due to the bulk movement of the fluid, in which case the
primary unknown c denotes the solute concentration, u the velocity of the fluid, and f accounts for generation or
degradation of c, e. g., by chemical reactions. Equation (1a) is complemented by the following boundary and initial
conditions:
on J × ∂Ωin (t) ,
c = cD
c = c
0
on {0} × Ω
(1b)
(1c)
with inflow boundary ∂Ωin (t) ≔ {x ∈ ∂Ω | u(t, x) · ν(x) < 0} and ν(x) denoting the outward unit normal. The outflow
boundary ∂Ωout (t) is defined as ∂Ωout (t) ≔ ∂Ω \ ∂Ωin (t); c0 : Ω → R+0 and cD : J × ∂Ωin (t) → R+0 are the given initial
and Dirichlet boundary data, respectively.
2. Discretization
2.1. Notation
Before describing the DG scheme for (1) we introduce some notation; an overview can be found in the Section “Index of notation”. Let Th = {T } be a regular family of non-overlapping partitions of Ω into K closed triangles T of
characteristic size h such that Ω = ∪T . For T ∈ Th , let νT denote the unit normal on ∂T exterior to T . Let EΩ denote
the set of interior edges, E∂Ω the set of boundary edges, and E ≔ EΩ ∪ E∂Ω = {E} the set of all edges (the subscript h
is suppressed here). For an interior edge E ∈ EΩ shared by triangles T − and T + , we define the one-sided values of
a scalar quantity w = w(x) on E by
w− (x) ≔ lim+ w(x − ε νT − )
ε→0
and
w+ (x) ≔ lim+ w(x − ε νT + ) ,
ε→0
respectively. For a boundary edge E ∈ E∂Ω , only the definition on the left is meaningful.
2
2.2. Variational formulation
Because of the local nature of the DG method, we can formulate the variational system of equations on a triangleby-triangle basis. To do that, we multiply (1a) by a smooth test function w : T → R and integrate by parts over
element T ∈ Th . This gives us
Z
Z
Z
Z
w f (t) dx .
w u(t) c(t) · νT ds =
∇w · u(t) c(t) dx +
w ∂t c(t) dx −
T
∂T
T
T
2.3. Semi-discrete formulation
We denote by P p (T ) the space of complete polynomials of degree at most p on T ∈ Th . Let
n
o
P p (Th ) ≔ wh : Ω → R ; ∀T ∈ Th , wh |T ∈ P p (T )
denote the broken polynomial space on the triangulation Th . For the semi-discrete formulation, we assume that the
coefficient functions (for t ∈ J fixed) are approximated as: uh ∈ [P p (Th )]2 and fh (t), c0h ∈ P p (Th ). A specific way to
compute these approximations was given in the first paper of the series [3]; here we use the standard L2 -projection into
P p (T ), therefore the accuracy of this approximation improves with increasing polynomial order p. Choosing the same
polynomial space for all functions simplifies the implementation and is done in preparation for later applications, in
which uh might be part of the solution of a coupled system. Incorporating the boundary condition (1b), the semidiscrete formulation reads:
Seek ch (t) ∈ P p (Th ) such that the following holds for t ∈ J and ∀T − ∈ Th , ∀wh ∈ P p (Th ) :
Z
Z
Z
Z
wh ∂t ch (t) dx −
∇wh · uh (t) ch (t) dx +
w−h u(t) · νT − ĉh (t) ds =
wh fh (t) dx ,
(2)
T−
T−
∂T −
T−
where the boundary integral is calculated using the upwind-sided value
ĉh (t, x) ∂T −
−
ch (t, x) if
+
ch (t, x) if
=
c (t, x) if
D
u(t, x) · νT − ≥ 0
u(t, x) · νT − < 0 ∧ x < ∂Ωin
x ∈ ∂Ωin
(outflow from T − )
(inflow into T − from T + )
(inflow into T − over ∂Ωin )
.
Note that we did not use the approximate representation of the velocity uh in the boundary integral. This is due to
the fact that the L2 -projection on elements may have poor approximation quality on edges and generally produces
different values on both sides of the edge ultimately leading to different upwind-sided values and inconsistent flux
approximations. Instead we evaluate the normal velocity u · νT in each quadrature point analytically and use the result
for both the numerical integration and the determination of the upwind direction as will be demonstrated in Sec. 2.3.4.
Thus far, we used an algebraic indexing style. In the remainder, we switch to a mixture of algebraic and numerical
style: for instance, Ekn ∈ ∂T k ∩ EΩ means all possible combinations of element indices k ∈ {1, . . . , K} and local edge
indices n ∈ {1, 2, 3} such that Ekn lies in ∂T k ∩ EΩ . This implicitly fixes the numerical indices which accordingly can
be used to index matrices or arrays.
We use a bracket notation followed by a subscript to index matrices and multidimensional arrays. Thus, for an
n-dimensional array X, the symbol [X]i1 ,...,in stands for the component of X with index il in the l-th dimension. As in
MATLAB / GNU Octave, a colon is used to abbreviate all indices within a single dimension. For example, [X]:,:,i3,...,in
is a two-dimensional array / matrix.
2.3.1. Local basis representation
In contrast to globally continuous basis functions mostly used by the continuous finite element method, the DG basis functions have no continuity constraints across triangle boundaries. Thus a standard DG basis function ϕki : Ω → R
is only supported on the triangle T k ∈ Th (i. e., ϕki = 0 on Ω r T k ) and can be defined arbitrarily while ensuring
!
(p + 1)(p + 2)
p+2
=
(3)
where N p ≔
∀k ∈ {1, . . . , K} , P p (T k ) = span ϕki i∈{1,...,N p } ,
p
2
3
is the number of local degrees of freedom. Note that N p may in general vary from triangle to triangle, but, for
simplicity, we assume here a uniform polynomial degree p for every triangle and abbreviate N ≔ N p . Clearly, the
number of global degrees of freedom equals KN. Closed-form expressions for orthonormal basis functions on the
reference triangle T̂ (cf. Sec. 4.1) employed in our implementation up to order two can be found in our first paper [3].
The basis functions up to order four are provided in the routine phi and their gradients in gradPhi. Bases of even
higher order can be constructed, e. g., with the Gram–Schmidt algorithm or by using a three-term recursion relation—
the latter is unfortunately not trivial to derive in the case of triangles. Note that these so-called modal basis functions ϕ̂i
do not posses interpolation properties at nodes unlike Lagrangian / nodal basis functions, which are often used by the
continuous finite element or nodal DG methods.
The local concentration ch and the local velocity uh on T k ∈ Th can be represented in terms of the local basis {ϕki }i∈{1,...,N} :
ch (t, x) T ≕
k
N
X
Ck j (t) ϕk j (x) ,
uh (t, x) T ≕
k
j=1
N X
2
X
Ukmj (t) em ϕk j (x) ,
j=1 m=1
where em denotes the m-th unit vector in R2 . We condense the coefficients associated with unknowns into twodimensional arrays C(t) such that Ck j (t) ≔ [C(t)]k, j , etc. The symbol [C]k,: is called local representation matrix
of ch on T k with respect to the basis ϕki i∈{1,...,N} . In a similar way, we express the coefficient functions as linear
combinations of the basis functions: On T k ∈ Th , we use the local representation matrices [C0 ]k,: for c0h and [F]k,:
for fh .
2.3.2. System of equations
Testing (2) with wh = ϕki for i ∈ {1, . . . , N} yields a time-dependent system of equations whose contribution
from T k (identified with T k− in boundary integrals) reads
N
X
∂tCk j (t)
Tk
j=1
|
+
Z
ϕki ϕk j dx −
{z
∂T k−
|
Ck j (t)
j=1
N X
2
X
Z
Uklm (t)
∂ xm ϕki ϕkl ϕk j dx
Tk
l=1 m=1
} |
I
Z
N
X
{z
II
N
X
Ck− j (t) ϕk− j if u(t) · νk− ≥ 0
j=1
X
N
ϕk− i u(t) · νk−
Ck+ j (t) ϕk+ j if u(t) · νk− < 0 ∧ x < ∂Ωin
j=1
cD (t) if x ∈ ∂Ωin
{z
III
}
Z
N
X
ϕki ϕkl dx ,
ds =
Fkl (t)
Tk
l=1
|
{z
}
IV
}
where we abbreviated νT k by νk . Written in matrix form, system (4) is then given by
M ∂t C + −G1 − G2 + R C = L − KD
| {z }
|
{z
}
≕ A(t)
(4)
(5)
≕ V(t)
with the representation vector
h
C(t) ≔ C11 (t)
· · · C1N (t)
· · · · · · C K1 (t)
· · · C KN (t)
iT
.
The block matrices and the right-hand side vectors of (5) are described in Sections 2.3.3 and 2.3.4. Note that all blocks
except for the mass matrix M are time-dependent (we have suppressed the time arguments here).
4
n−
Ek −
Tk −
ν k − n−
Tk +
Figure 1: Two triangles adjacent to edge Ek− n− . It holds: Ek− n− = Ek+ n+ and νk− n− = −νk+ n+ .
2.3.3. Contributions from area terms I, II, IV
The matrices in the remainder of this section have sparse block structure; by giving definitions for non-zero blocks
we tacitly assume a zero fill-in. The mass matrix M ∈ RKN×KN in term I is defined component-wise as
Z
ϕki ϕk j .
[M](k−1)N+i,(k−1)N+ j ≔
Tk
Since the basis functions ϕki , i ∈ {1, . . . , N} are supported only on T k , M has a block-diagonal structure
MT 1
Z ϕk1 ϕk1 · · · ϕk1 ϕkN
..
.. dx ,
..
..
M =
with
MT k ≔
.
.
.
.
Tk
MT K
ϕkN ϕk1 · · · ϕkN ϕkN
(6)
i. e., it consists of K local mass matrices MT k ∈ RN×N . Henceforth, we write M = diag MT 1 , . . . , MT K .
m
KN×KN
The block matrices G ∈ R
, m ∈ {1, 2} from term II are given by
[Gm ](k−1)N+i,(k−1)N+ j ≔
N
X
Uklm (t)
l=1
Z
∂ xm ϕki ϕkl ϕk j dx .
Tk
m
Similarly to M, the matrices Gm = diag Gm
T 1 , . . . , GT K are block-diagonal with local matrices
Gm
Tk ≔
N
X
l=1
Z ∂ xm ϕk1 ϕkl ϕk1
..
Uklm (t)
.
Tk
∂ xm ϕkN ϕkl ϕk1
· · · ∂ xm ϕk1 ϕkl ϕkN
..
..
dx .
.
.
· · · ∂ xm ϕkN ϕkl ϕkN
(7)
Vector L(t) resulting from IV is obtained by multiplication of the representation vector of fh (t) to the global mass
matrix:
h
iT
L(t) = M F11 (t) · · · F1N (t) · · · · · · F K1 (t) · · · F KN (t) .
2.3.4. Contributions from edge term III
Interior Edges EΩ . In this section, we consider a fixed triangle T k = T k− with an interior edge Ek− n− ∈ ∂T k− ∩ EΩ =
∂T k− ∩∂T k+ shared by an adjacent triangle T k+ and associated with fixed local edge indices n− , n+ ∈ {1, 2, 3} (cf. Fig. 1).
For a fixed index i ∈ {1, . . . , N}, we have a contribution for ϕk− i in a block matrix RΩ ∈ RKN×KN
Z
(PN Ck− j (t) ϕk− j if u · νk− n− ≥ 0)
PNj=1
−
−
−
ds .
ϕk i u(t) · νk n
E k − n−
j=1 Ck+ j (t) ϕk+ j if u · νk− n− < 0
This means that, depending on the direction of the velocity field u(t, x), we obtain entries in the diagonal or offdiagonal blocks of RΩ . Entries in diagonal blocks are then component-wise given by
[RΩ ](k−1)N+i,(k−1)N+ j ≔
X
Ekn ∈∂T k ∩EΩ
Z
Ekn
ϕki ϕk j u·νkn δu·νkn ≥0 ds with
5
δu·νkn ≥0 (t, x) ≔
(
1
0
)
if u(t, x) · νkn ≥ 0
.
if u(t, x) · νkn < 0
(8a)
Entries in off-diagonal blocks in R are possibly non-zero only for pairs of triangles T k− , T k+ with ∂T k− ∩ ∂T k+ , ∅ and
read
Z
(8b)
ϕk− i ϕk+ j u · νk− n− δu·νk− n− <0 ds with δu·νk− n− <0 ≔ 1 − δu·νk− n− ≥0 .
[RΩ ](k− −1)N+i,(k+ −1)N+ j ≔
E k − n−
Boundary Edges E∂Ω . Similarly to interior edges we have contributions for a boundary edge Ekn ∈ ∂T k ∩ E∂Ω
Z
(PN Ck j (t) ϕk j if u(t) · νkn ≥ 0)
j=1
ds .
ϕki u(t) · νkn
cD (t)
if u(t) · νkn < 0
Ekn
These consist of entries in the block diagonal matrix R∂Ω ∈ RKN×KN
X Z
ϕki ϕk j u · νkn δu·νkn ≥0 ds ,
[R∂Ω ](k−1)N+i,(k−1)N+ j ≔
Ekn ∈∂T k ∩E∂Ω
(9)
Ekn
and in the right-hand side vector KD ∈ RKN
X Z
ϕki cD (t) u · νkn δu·νkn <0 ds .
[KD ](k−1)N+i ≔
Ekn ∈∂T k ∩E∂Ω
(10)
Ekn
We combine the block matrices RΩ , R∂Ω into a block matrix R ∈ RKN×KN
R ≔ RΩ + R∂Ω .
(11)
Since the definition of entries in the diagonal blocks in Eqns. (8a), (9) is the same for both matrices differing only in
the set of edges included in the sum, we can disregard the fact whether they are interior or boundary edges and simply
assemble the entries for all Ekn ∈ ∂T k .
2.4. Time discretization
The system (5) is equivalent to
M∂t C(t) = V(t) − A(t) C(t) ≕ S C(t), t
(12)
with A(t) ∈ RKN×KN and right-hand-side vector V(t) ∈ RKN as defined in (5).
We discretize system (12) in time using TVD (total variation diminishing) Runge–Kutta methods [8] of orders one,
two, and three, which are representatives of the class of SSP (strong stability preserving) Runge–Kutta methods [9].
The advantage of using a time stepping algorithm of such type lies in the guaranteed preservation of the monotonicity
of the solution if the DG discretization is also post processed by a slope limiting method.
Let 0 = t1 < t2 < . . . < tend be a not necessarily equidistant decomposition of the time interval J and let
∆tn ≔ tn+1 − tn denote the time step size. The update scheme of the s-step Runge–Kutta method is given by
C(0)
C(i)
Cn+1
= Cn ,
= ωi Cn + (1 − ωi ) C(i−1) + ∆tn M−1 Sn+δi ,
= C(s) ,
for i = 1, . . . , s ,
(13)
where we abbreviated Cn ≔ C(tn ) and Sn+δi ≔ S(C(i−1) , tn + δi ∆tn ) with coefficients
s=1 :
ω1 = 0 ,
δ1 = 0 .
s=2 :
s=3 :
ω1 = 0 , ω2 = 1/2 ,
ω1 = 0 , ω2 = 3/4 , ω3 = 1/3 ,
δ1 = 0 , δ2 = 1 .
δ1 = 0 , δ2 = 1 , δ3 = 1/2 .
When possible, we choose the order of the time-discretization to be p+1, with p being the spatial approximation order,
in order to avoid the temporal discretization error dominating the spatial one. Unfortunately, no optimal SSP Runge–
Kutta methods higher than order three are known [9], which is why we restrict ourselves to orders one to three for the
time discretization.
6
3. Slope limiting
Slope limiters are a technique to prevent the onset of spurious oscillations that violate the monotonicity preserving
property of the piecewise constant part of a DG solution by means of restricting some of the degrees of freedom (generally linear and superlinear) to certain bounds and thus eliminating over- and undershoots. All limiting procedures
utilize the fact that the lowest order (piecewise constant) part of a DG solution in explicit TVD time stepping schemes
is guaranteed to preserve the monotonicity of the solution and produce no spurious extrema. Using this physically
consistent but numerically not very accurate solution part, all slope limiters attempt to modify the full higher order
DG solution in a suitable way—on the one hand, to prevent any oscillations and, on the other hand, to preserve as
much of the accuracy as possible. The key differences in slope limiters affect the limiting stencil used (edge neighbors, node neighbors, neighbors of the neighbors, etc.), presence of ad hoc parameters, the amount of the introduced
numerical diffusion, a strict or less strict preservation of the monotonicity, and the degree of solution degradation in
smooth extrema.
Whereas a large literature on slope limiting for piecewise linear DG discretizations exists [10, 11, 12], the limiting
of DG solutions with p ≥ 2 is a much less explored area. The traditional approach to dealing with superlinear
DG solutions [13] has been based on ignoring all higher order degrees of freedom on elements on which linear limiting
is active. Other methods require a much larger stencil [14, 15] to provide enough information for the reconstruction of
higher order derivatives. The hierarchical vertex-based limiters of Kuzmin [16, 7] represent a computationally efficient
scheme easily extendable to any discretization order and supporting fully unstructured meshes. These limiters do not
guarantee the strict monotonicity of the DG solution, but the violations are small and may be further reduced by simple
modifications described in Sec. 3.4.
3.1. Taylor basis representation
Many limiting procedures rely on some fundamental properties of a certain choice of basis, in our case the 2D Taylor basis, which we introduce in a way similar to Kuzmin [16]. Consider the 2D Taylor series expansion of a local
solution ch ∈ P p (T k ),
ch (x) =
X
∂ a ch (xkc )
0≤|a|≤p
(x − xkc ) a
a!
on T k ∈ Th
(14)
T
T
about the centroid xkc = [x1kc , x2kc ] of T k ∈ Th with a two-dimensional multi-index a = [a1 , a2 ] ∈ N20 , where we use
some standard notation for multi-indices a, b ∈ N20 and x ∈ R2 :
T
a ± b = [a1 ± b1 , a2 ± b2 ] ,
1
2
x a ≔ (x1 )a (x2 )a ,
For v : T k → R, let v ≔
form [17, 18, 19]
ch (x) = ch +
R
1
|T k | T k
|a| ≔ a1 + a2 ,
1
2
∂ a ≔ ∂|a| ∂(x1 )a ∂(x2 )a .
a! ≔ a1 !a2 ! ,
v(x) dx denote the integral mean of v on T k . We express (14) in the equivalent
X
∂ch
∂ch
(x − xkc ) a − (x − xkc ) a
1
1
2
2
a
(x
)(x
−
x
)
+
(x
)(x
−
x
)
+
∂
c
(x
)
kc
kc
h
kc
kc
kc
a!
∂x1
∂x2
2≤|a|≤p
on T k ∈ Th . (15)
Note that varying any terms in (15) except ch does not affect the mean of ch (x) over T k .
To be able to identify each term in expansions (14), (15) by a consecutive index, we introduce a linear index
mapping I : N20 → N corresponding to any two-dimensional multi-index a ∈ N20 as
I(a) = N|a|−1 + a2 + 1 =
|a|(|a| + 1)
+ a2 + 1
2
(16)
with N p = dim P p (T ) as defined in (3). We implicitly define a j such that ∀ j ∈ N, I(a j ) = j. The linear indices,
polynomial degrees, and corresponding multi-indices up to order four are listed in Table 1. This leads to the following
7
p
0
1
2
3
4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
I(ai )
ai T [0, 0] [1, 0] [0, 1] [2, 0] [1, 1] [0, 2] [3, 0] [2, 1] [1, 2] [0, 3] [4, 0] [3, 1] [2, 2] [1, 3] [0, 4]
Table 1: Multi-indices (bottom), linear indices (middle), and corresponding polynomial degrees in the Taylor basis (17)
(top).
definition of the local Taylor basis [17]:
φk1 = 1 ,
φk2 =
x1k − x1kc
∆(x1k )
,
φk3 =
x2k − x2kc
∆(x2k )
,
φki =
(x − xkc ) ai − (x − xkc ) ai
ai ! (∆xk ) ai
for i ≥ 4 .
(17)
As opposed to the DG basis {ϕk j } (cf. [3], Sec. 2.4.1), the basis {φk j } cannot be defined on a reference element T̂ .
T
j
j
j
j
j
j
The scaling by ∆xk = ∆(x1k ), ∆(x2k ) with ∆(xk ) ≔ (xk,max − xk,min )/2, where xk,max ≔ maxi∈{1,2,3} xki and xk,min ≔
j
mini∈{1,2,3} xki are the minimum and maximum values of the corresponding spatial coordinates on T k , is introduced to
obtain a better conditioned operator [17]. The Taylor degrees of freedom are now proportional to the cell mean values
ch and derivatives of ch at the centroid xkc
ch (x) = ch φk1 +
Np
X
!
!
∂ch
∂ch
1
2
φ
(x)+
φ
(x)+
(x
)
∆
x
(x
)
∆
x
∂ ai ch (xkc ) (∆xk ) ai φki (x) on T k ∈ Th . (18)
k2
k3
kc
kc
k
k
1
2
∂x
∂x
i=4
Note that the Taylor basis is non-orthogonal on triangular meshes [16], but the cell means are still decoupled from
the other degrees of freedom since
Z
Z
φk1 φk j dx = 0 for j > 1 .
φ2k1 dx = |T k | ,
Tk
Tk
To transform a function ch from the modal basis representation with representation matrix C(t) ∈ RK×N —as
described in Sec. 2.3.1—into a Taylor basis representation with representation matrix CTaylor (t) ∈ RK×N , we employ
the L2 -projection defined locally for T k ∈ Th by
N
N
Z
Z
X
X Taylor
wh Ck j (t) ϕk j dx =
∀wh ∈ P p (T k ) ,
wh Ck j (t) φk j dx .
Tk
Tk
j=1
j=1
Choosing wh = ϕki for i ∈ {1, . . . , N} we obtain
N
X
j=1
Ck j (t)
Z
ϕki ϕk j dx =
Tk
N
X
j=1
Taylor
Ck j
(t)
Z
ϕki φk j dx
Tk
⇔
DG,Taylor
MT k [C]k,: = MT k
h
CTaylor
i
k,:
with the local mass matrix MT k as defined in Eq. 6 and the local basis transformation matrix
Z ϕk1 φk1 · · · ϕk1 φkN
..
DG,Taylor
.. dx .
..
MT k
≔
(19)
.
.
.
Tk
ϕkN φk1 · · · ϕkN φkN
DG,Taylor
DG,Taylor
and representation vectors C, CTaylor ∈ RKN we obtain a linear
Using MDG,Taylor ≔ diag MT 1
, . . . , MT K
system of equations
M C = MDG,Taylor CTaylor ,
(20)
which can be employed to transform back-and-forth between both bases.
8
Figure 2: The neighborhood of a vertex xki ∈ T k (red circle) considered in Eq. (21) consists of the patch of elements
max
containing xki (red area). The bounds cmin
ki , cki from Eq. (22) are determined from all centroid values (green squares)
within this neighborhood.
3.2. Linear vertex-based limiter
Kuzmin [7, 16] and Aizinger [20] described the vertex-based limiter, which is based on the Barth–Jespersen
limiter [21] and improved it further by taking the bounds from all elements containing the vertex instead of taking
only edge neighbors of the cell. The goal is to determine the maximum admissible slope for a linear reconstruction of
the form
ch (x) = ckc + αke ∇ch (xkc ) · (x − xkc ) ,
0 ≤ αke ≤ 1 ,
x ∈ Tk ,
where we abbreviated the function value ckc ≔ ch (xkc ) in the centroid xkc . The correction factor αke is chosen such
that above reconstruction is bounded in all vertices xki ∈ T k by the minimum and maximum centroid values of all
elements containing xki , that is
∀T k ∈ Th , ∀i ∈ {1, 2, 3} ,
max
cmin
ki ≤ ch (xki ) ≤ cki
(21)
with
cmin
ki ≔
min
{T l ∈Th | xki ∈T l }
clc ,
cmax
ki ≔
max
{T l ∈Th | xki ∈T l }
clc
(22)
(cf. Fig. 2 for an illustration). To enforce (21), the correction factor αke is defined as [16]
max
(c − ckc ) (cki − ckc ) if
cki > cmax
ki
ki
min
max
1
if cki ≤ cki ≤ cki
,
∀T k ∈ Th , αke ≔ min
i ∈ {1,2,3}
(cmin − c ) (c − c ) if
c < cmin
ki
kc
ki
kc
ki
(23)
ki
where cki ≔ ckc + ∇ch (xkc ) · (xki − xkc ) is the unconstrained linear reconstruction in xki . The limited counterpart of the
DG solution (18) becomes then
#
"
!
!
∂ch
∂ch
1
2
(x
)
∆
x
(x
)
∆
x
φ
(x)
+
φ
(x)
on T k ∈ Th ,
ch (x) = ch φk1 + αke
kc
kc
k2
k3
k
k
∂x1
∂x2
i. e., the linear degrees of freedom are scaled by αke , and any degrees of freedom associated with higher polynomial
degrees are set to zero. In the particular case of a linear DG approximation, the limiting can be performed using any
hierarchical basis as opposed to higher-order DG solutions that require a Taylor basis representation (see Sec. 3.3). In
case of αke = 1, both linear and superlinear degrees of freedom remain unchanged.
3.3. Hierarchical vertex-based limiter
Further improvements by Kuzmin combine the vertex-based or standard Barth–Jespersen limiter with the higher
order limiting scheme of Yang and Wang [17], who limit the numerical solution by multiplying all derivatives of
9
(q)
order q by a common correction factor αke instead of applying the correction only to the linear terms and dropping
all higher degrees of freedom. Kuzmin [16] described this scheme in detail for quadratic representations, and here we
offer a closed form expression of this limiting procedure for DG discretizations of arbitrary orders.
Let Aq ≔ {a ∈ N20 |a| = q} be the set of all two-dimensional multi-indices of order q. We determine the correction
(q)
factor αke
for each order q ≤ p by computing correction factors (23) using the linear vertex-based limiter for all linear
reconstructions of derivatives of order q − 1,
∀ a ∈ Aq−1 ,
Taylor
Taylor
Taylor
ck,a,i ≔ Ck,I(a) φk1 (xki ) + Ck,I(a+[1,0]T ) φk2 (xki ) + Ck,I(a+[0,1]T ) φk3 (xki ) on T k ∈ Th ,
(24)
where the indices of the corresponding degrees of freedom are given by I(a) and the first x1 -derivatives (identified by
(q)
I(a + [1, 0]T )) and x2 -derivatives (identified by I(a + [0, 1]T )). Formally, the correction factor αke is defined as
(q)
(q)
αke
= min αka
,
a∈Aq−1
with
α(q)
ka
max
(ck,a,i − ck,a,c ) (ck,a,i − ck,a,c )
1
≔ min
i∈{1,2,3}
(cmin − ck,a,c )(ck,a,i − ck,a,c )
k,a,i
if
c a,i > cmax
a,i
min
if ck,a,i ≤ ck,a,i ≤ cmax
k,a,i
if
ck,a,i < cmin
k,a,i
, (25)
max
where cmin
k,a,i , ck,a,i are defined as in (22). To avoid the loss of accuracy at smooth extrema, the lower order derivatives
should be limited by a factor not exceeding that of the higher order derivatives, since lower orders are typically
smoother. Beginning with the highest-order degrees of freedom, we compute the correction factors
∀q ≥ 1
(d)
α(q)
ke ≔ max αke .
(26)
q≤d≤p
Once the correction factor α(q)
ke becomes equal to one for some d > 1, no further limiting on this element is necessary.
The limited solution becomes
N
X
!
!
∂ch
(1) ∂ch
1
2
i |)
φ
(x)
+
α
φ
(x)
+
ch (x) = ch φk1 + α(1)
α(|a
∂ ai ch (xkc ) (∆xk ) ai φki (x) .
(x
)
∆
x
(x
)
∆
x
k2
k3
kc
kc
k
k
ke
ke
ke
1
2
∂x
∂x
i=4
3.4. Stricter form of the vertex-based limiter
Our numerical experiments showed that implicitly assuming that higher order derivatives are always smoother
than lower order derivatives results in limiting procedures that do not guarantee strict fulfillment of condition (21),
especially at discontinuities in the solution. We modified two key components of the limiter presented in the previous
section and obtained a limiter that exhibited slightly stronger peak clipping but turned out to be always effective:
1. Instead of employing only the linear reconstruction as given in Eq. (24), we replace ck,a,i in the computation of
the correction factor in Eq. (25) by the full reconstruction
X
Taylor
∀ a ∈ Aq−1 , ck,a,i ≔
Ck,I(a+b) φk,I(b) (xki ) ,
0≤|b|<p−q
where p is the polynomial degree of the DG solution ch (x).
2. Again, we begin with the highest-order derivative but drop the hierarchical limiting condition (26) and instead
(q)
apply each correction coefficient αke immediately to all coefficients corresponding to polynomial degree q or
(q−1)
higher. These limited coefficients are then used to compute the next correction coefficient αke .
The result of our stricter limiter is
ch (x) =
ch φk1 +α(1)
ke
N
X
!
!
∂ch
(1) ∂ch
(1)
(|ai |)
1
2
ai
ai
)
(∆x
φ
(x)+α
φ
(x)+
(x
)
∆
x
(x
)
∆
x
α
·
·
·
α
φki (x) .
∂
c
(x
)
k2
k3
kc
kc
h
kc
k
k
k
ke
ke
ke
∂x1
∂x2
i=4
10
3.5. Slope limiting in time-dependent problems
For time-dependent problems, the slope limiting procedure is applied to each intermediate solution C(i) in the
update scheme (13). However, due to the fact that the Taylor basis is non-orthogonal on triangles—as discussed in
Sec. 3.1—an implicit coupling between the spatial derivatives is present and leads to non-smooth spatial variations
in the time derivatives of ch . For that reason, Kuzmin [7] applied the slope limiter not only to the solution of each
Runge–Kutta stage but also to the time derivative ċ ≔ ∂t c and used in addition a filtering procedure that can be
interpreted as selective mass lumping.
We describe this technique first for a discretization in Taylor basis and extend it then to arbitrary basis representations. Let ΦTaylor denote the slope limiting operator that applies any of the above slope limiting procedures to a global
representation vector CTaylor (t) of a solution ch (t) in Taylor basis representation. The semi-discrete system (12) written
in a Taylor basis
MC ∂t CTaylor (t) = STaylor CTaylor (t), t
is replaced by
ML ∂t CTaylor (t) = STaylor ΦTaylor CTaylor (t), t + (ML − MC ) ΦTaylor ∂t CTaylor (t) ,
where MC = {mi j }, ML ≔ diag{mii } denote the full and the lumped mass matrices in Taylor basis (to improve readability, we drop the superscript ‘Taylor’ here). Note that for the case ΦTaylor = I both formulations are identical.
Consequently, update scheme (13) is modified replacing
Taylor,n+δi
CTaylor,(i) = ωi CTaylor,n + (1 − ωi ) CTaylor,(i−1) + ∆tn ĊTaylor,(i)
with ĊTaylor,(i) ≔ M−1
C S
by the selectively lumped and limited update
i
h
CTaylor,(i) = ΦTaylor ωi CTaylor,n + (1 − ωi ) CTaylor,(i−1) + ∆tn C̃Taylor,(i)
with
i
h
Taylor Taylor,(i)
Ċ
+ STaylor,n+δi .
C̃Taylor,(i) = M−1
L (ML − MC ) Φ
Although the mass matrix M in the modal basis is diagonal, the implicit coupling between the spatial x1 - and x2 derivatives is still present. Only the Taylor basis has vectors coinciding with the coordinate directions, hence, the
lumping technique cannot be directly applied to representations in other bases.
To get rid of this implicit coupling of the spatial derivatives in the time derivative of the modal DG basis, we
reformulate the lumped time derivative in Taylor basis as
−1 Taylor,n+δi
Taylor Taylor,(i)
Taylor Taylor,(i)
Ċ
+ M−1
Ċ
− M−1
C̃Taylor,(i) = M−1
L MC MC S
L MC Φ
L ML Φ
| {z }
=I
Taylor Taylor,(i)
−1 Taylor,n+δi
Taylor Taylor,(i)
−1
−Φ
Ċ
=Φ
Ċ
+ ML MC MC S
|
{z
}
=ĊTaylor,(i)
Taylor,(i)
− ΦTaylor ĊTaylor,(i) .
= ΦTaylor ĊTaylor,(i) + M−1
L MC Ċ
Using the time derivative in modal basis Ċ(i) = M−1 Sn+δi and transformation (20), we obtain
i
h
Taylor,(i)
− ΦTaylor ĊTaylor,(i) ,
C̃(i) = M−1 MDG,Taylor C̃Taylor,(i) = M−1 MDG,Taylor ΦTaylor ĊTaylor,(i) + M−1
L MC Ċ
−1
where ĊTaylor,(i) ≔ MDG,Taylor MĊ(i) . Thus, the fully modified version of update scheme (13) reads as
C(0)
C(i)
Cn+1
= Cnh,
i
= Φ ωi Cn + (1 − ωi ) C(i−1) + ∆tn C̃(i)
= C(s) ,
for ∀i = 1, . . . , s ,
with C̃(i) as given in (27) and the slope limiting operator Φ formally defined as
−1
Φ ≔ M−1 MDG,Taylor ΦTaylor MDG,Taylor M .
11
(27)
(28)
(29)
x̂2
xk3
Fk
x̂3
1
Ek1
Ek2
Ê1
x2
Ê2
T̂
x̂1
0
0
xk1
Tk
xk2
Ek3
x̂2
Ê3
1
1 x̂
x1
Figure 3: The affine mapping Fk transforms the reference triangle T̂ with vertices x̂1 = [0, 0]T , x̂2 = [1, 0]T , x̂3 =
[0, 1]T to the physical triangle T k with counter-clockwise-ordered vertices xki , i ∈ {1, 2, 3}.
3.6. Boundary conditions
max
Problems can occur when computing the bounds cmin
ki , cki in Eq. (21) for control points xki ∈ ∂Ωin on the Dirichlet
boundary. To account for the boundary data in the limiting procedure we include on those control points the boundary
value cD (tn + δi ∆tn , xki ) in the minimum/maximum-operation in Eq. (22) for α(1)
ke when applying the slope limiting
operator Φ in Eq. (28).
4. Implementation
An extensive documentation on our data structures, grid, etc. can be found in the first paper of the series [3]. These
explanations are not reproduced here; greater detail is provided for routines first introduced in the present work.
4.1. Backtransformation to the reference triangle
We are using a back transformation to the reference triangle T̂ = {(0, 0), (1, 0), (0, 1)} defined by an affine mapping
h
i
Fk : T̂ ∋ x̂ 7→ Bk x̂ + xk1 = x ∈ T k ,
with R2×2 ∋ Bk ≔ xk2 − xk1 xk3 − xk1 ,
(30)
for any triangle T k = {xk1 , xk2 , xk3 } ∈ Th (see Fig. 3). It holds 0 < det Bk = 2|T k |. Any function w : T k → R implies
ŵ : T̂ → R by ŵ = w ◦ Fk , i. e., w(x) = ŵ( x̂) . The transformation of the gradient is obtained by the chain rule:
ˆ k −T ∇
ˆ,
∇ = ∇F
(31)
ˆ = [∂ x̂1 , ∂ x̂2 ]T . This results in transformation formulas for integrals over an element T k or
where we abbreviated ∇
an edge Ekn ⊂ T k for a function w : Ω → R
Z
Z
Z
Z
|T k |
(32a)
w(x) dx =
w ◦ Fk ( x̂) d x̂ = 2|T k | w ◦ Fk ( x̂) d x̂ = 2|T k | ŵ( x̂) d x̂ ,
|T̂ | T̂
T̂
T̂
Tk
Z
Z
Z
|Ekn |
|Ekn |
w(x) dx =
w ◦ Fk ( x̂) d x̂ =
ŵ( x̂) d x̂ .
(32b)
|Ên | Ên
|Ên | Ên
Ekn
4.2. Numerical integration
As an alternative to the symbolic integration functions provided by MATLAB, we implemented a quadrature
integration functionality for triangle and edge integrals.
Since we transform all integrals on T k ∈ Th to the reference triangle T̂ (cf. Sec. 4.1), it is sufficient to define the
quadrature rules on T̂ (which, of course, can be rewritten to apply for every physical triangle T = FT (T̂ )):
Z
T̂
ĝ( x̂) d x̂ ≈
R
X
ωr ĝ( q̂r )
(33)
r=1
12
with R quadrature points q̂r ∈ T̂ and quadrature weights ωr ∈ R. The order of a quadrature rule is the largest integer s
such that (33) is exact for polynomials g ∈ P s (T̂ ). Note that we exclusively rely on quadrature rules with positive
weights and quadrature points located strictly in the interior of T̂ and not on ∂T̂ . The rules used in the implementation
are found in the routine quadRule2D. An overview of quadrature rules on triangles can be found in the “Encyclopaedia
of Cubature Formulas” [22]. For edge integration, we rely on standard Gauss quadrature rules of required order.
The integrals in (4) contain integrands that are polynomials of maximum order 3p−1 on triangles and of maximum
order 3p on edges. Using quadrature integration, one could choose rules that integrate all such terms exactly; however,
sufficient accuracy can be achieved with quadrature rules that are exact for polynomials of order 2p on triangles and
2p + 1 on edges (cf. [23]).
4.3. Assembly
The aim of this section is to transform the terms required to build the block matrices in (5) to the reference
triangle T̂ and then to evaluate those either via numerical quadrature or analytically. The assembly of block matrices
from local contributions is then performed in vectorized operations.
For the implementation, we need the explicit form for the components of the mappings Fk : T̂ → T k and their
inverses F−1
k : T k → T̂ as defined in (30). Recalling that 0 < det Bk = 2|T k | (cf. Sec. 4.1) we obtain
"
#
#
" 11 1
1 B22
(x1 − a1k1 ) − B12
(x2 − a2k1 )
Bk x̂ + B12
x̂2 + a1k1
−1
k
k
k
.
and
Fk (x) =
Fk ( x̂) =
2
2
21 1
1
1
22 2
2
B21
2 |T k | B11
k (x − ak1 ) − Bk (x − ak1 )
k x̂ + Bk x̂ + ak1
From (31) we obtain the component-wise rule for the gradient in x ∈ T k :
#
"
" #
21
1 B22
∂ x1
k ∂ x̂1 − Bk ∂ x̂2 .
=
12
∂ x2
2 |T k | B11
k ∂ x̂2 − Bk ∂ x̂1
(34)
Similarly to the first paper in series [3], we make extensive use of the Kronecker product A ⊗ B of two matrices
A = [ai j ] ∈ Rma ×na , B = [bkl ] ∈ Rmb ×nb defined as
h
i
A ⊗ B ≔ ai j B ∈ Rma mb ×na nb .
(35)
In the following, we present the necessary transformation for all blocks of system (5) and name the corresponding
MATLAB / GNU Octave routines that can be found in Sec. 5.
4.3.1. Assembly of M
Using the transformation rule (32a), the following holds for the local mass matrix MT k as defined in (6):
MT k = 2|T k | M̂
with
Z ϕ̂1 ϕ̂1
.
M̂ ≔
..
T̂
ϕ̂N ϕ̂1
· · · ϕ̂1 ϕ̂N
.. ,
..
.
.
· · · ϕ̂N ϕ̂N
(36)
where M̂ ∈ RN×N is the representation of the local mass matrix on the reference triangle T̂ . With (6) we see that the
global mass matrix M can be expressed as a Kronecker product of a matrix containing the areas |T k | and the local
matrix M̂:
|T 1 |
MT 1
..
..
=
2
M =
⊗ M̂ .
.
.
|T K |
MT K
In the corresponding assembly routine assembleMatElemPhiPhi , the sparse block-diagonal matrix is generated using
the command spdiags with the list g.areaT (cf. [3]).
13
4.3.2. Assembly of Gm
Application of the product rule, (32a), and (34) gives us
Z
Z
11
22
21
∂ x2 ϕki ϕkl ϕk j = −B12
∂ x1 ϕki ϕkl ϕk j = Bk [Ĝ]i, j,l,1 − Bk [Ĝ]i, j,l,2 ,
k [Ĝ]i, j,l,1 + Bk [Ĝ]i, j,l,2
Tk
Tk
with a multidimensional array Ĝ ∈ RN×N×N×2 representing the transformed integral on the reference triangle T̂ :
Z
∂ x̂m ϕ̂i ϕ̂ j ϕ̂l , for m ∈ {1, 2} .
(37)
[Ĝ]i, j,l,m ≔
T̂
Now we can express the local matrix G1T k from (7) as
G1T k =
N
X
l=1
=
N
X
l=1
Z
22
1
Ukl (t) Bk
T̂
∂ x̂1 ϕ̂1 ϕ̂1 ϕ̂l
..
.
∂ x̂1 ϕ̂N ϕ̂1 ϕ̂l
· · · ∂ x̂1 ϕ̂1 ϕ̂N ϕ̂l
Z
..
..
− B21
.
k
.
T̂
· · · ∂ x̂1 ϕ̂N ϕ̂N ϕ̂l
21
Ukl1 (t) B22
k [Ĝ]:,:,l,1 − Bk [Ĝ]:,:,l,2
∂ x̂2 ϕ̂1 ϕ̂1 ϕ̂l
..
.
∂ x̂2 ϕ̂N ϕ̂1 ϕ̂l
· · · ∂ x̂2 ϕ̂1 ϕ̂N ϕ̂l
..
..
.
.
· · · ∂ x̂2 ϕ̂N ϕ̂N ϕ̂l
m
and analogously G2T k . With Gm = diag(Gm
T 1 , . . . , GT K ) we can vectorize over all triangles using the Kronecker product
as done in the routine assembleMatElemDphiPhiFuncDiscVec . We would like to point out that this is identical to the
assembly of Gm in our first paper [3] except for the vectorial coefficient function; however, the respective section
had some typos, above is the corrected version. Additionally, in the assembly routine the wrong component of the
normal vector was used for the assembly of Gm . Corrected and up-to-date versions of the code can be found in our
Github-repository [2].
4.3.3. Assembly of R
To ease the assembly of R we split the global matrix as given in (11) into a block-diagonal part and a remainder so
that R = Rdiag + Roffdiag holds. We first consider the block-diagonal entries of R given in Eqns. (8a), (9) and transform
the integral terms to the n-th edge of the reference triangle Ên :
Z
Z
|Ekn |
ϕki ϕk j u · νkn δu·νkn ≥0 dx =
ϕ̂i ϕ̂ j (u ◦ Fk ( x̂)) · νkn δu·νkn ≥0 d x̂
|Ên | Ên
Ekn
Z
|Ekn | 1
ϕ̂i ◦ γ̂n (s) ϕ̂ j ◦ γ̂n (s) (u ◦ Fk ◦ γ̂n (s)) ·νkn δûkn ·νkn ≥0 |γ̂′n (s)| ds
=
|
{z
}
|Ên | 0
≕ ûkn (s)
≈ |Ekn |
R
X
ωr ϕ̂i ◦ γ̂n (qr ) ϕ̂ j ◦ γ̂n (qr ) uknˆ(qr ) · νkn δûkn ·νkn ≥0 ,
|
{z
}
r=1
≕[R̂diag ]i, j,n,r
(38)
where we used transformation rule (32b), quadrature rule (33), and |γ̂′n (s)| = |Ên |. The explicit forms of the mappings γ̂n : [0, 1] → Ên can be easily derived:
"
#
"
#
" #
1−s
0
s
γ̂1 (s) ≔
,
γ̂2 (s) ≔
,
γ̂3 (s) ≔
.
(39)
s
1−s
0
Thus, we can assemble the global matrix using the Kronecker product
Rdiag ≔
3 X
R
X
n=1 r=1
|E1n |
ωr
..
.
(û1n (qr ) · ν1n ) δû1n ·ν1n ≥0
◦
|E Kn |
14
..
.
(ûKn (qr ) · νKn ) δûKn ·νKn ≥0
h diag i
⊗ R̂
,
:,:,n,r
where ‘◦’ denotes the Hadamard product.
Next, we consider the off-diagonal blocks of R stored in Roffdiag . For an interior edge Ek− n− = Ek+ n+ ∈ ∂T k− ∩
∂T k+ , n− , n+ ∈ {1, 2, 3}, we obtain analogously:
Z
ϕk− i ϕk+ j u · νk− n δu·νk− n <0 dx
E k − n−
=I
Z
z }| {
|Ek− n− |
=
ϕ̂i ϕk+ j ◦ Fk+ ◦ F−1
k+ ◦Fk− ( x̂) (u ◦ Fk− ( x̂)) · νk− n− δu·νk− n− <0 d x̂
|Ên− | Ên−
Z
|Ek− n− |
ϕ̂i ϕ̂ j ◦ F−1
=
k+ ◦ Fk− ( x̂) (u ◦ Fk− ( x̂)) · νk− n− δu·νk− n− <0 d x̂
|Ên− | Ên−
Z 1
−
−
ϕ̂i ◦ γ̂n− (s) ϕ̂ j ◦ F−1
= |Ek n |
k+ ◦ Fk− ◦ γ̂n− (s) ûk− n− (s) · νk− n− δûk− n− ·νk− n− <0 ds .
0
−
Note that F−1
k+ ◦ Fk maps from T̂ to T̂ . Since we compute a line integral, the integration domain is further restricted to
an edge Ên− , n− ∈ {1, 2, 3} and its co-domain to an edge Ên+ , n+ ∈ {1, 2, 3}. As a result, this integration can be boiled
down to nine possible maps between the sides of the reference triangle expressed as
Ên− ∋ x̂ 7→ ϑ̂n− n+ ( x̂) = F−1
k+ ◦ Fk− ( x̂) ∈ Ê n+
ϑ̂n− n+ :
for an arbitrary index pair {k− , k+ } as described above. The closed-form expressions of the nine cases are given in our
first paper [3]. We apply quadrature rule (33) and define R̂offdiag ∈ RN×N×3×3×R by
h
i
R̂offdiag
≔ ϕ̂i ◦ γ̂n− (qr ) ϕ̂ j ◦ ϑ̂n− n+ ◦ γ̂n− (qr )
(40)
− +
i, j,n ,n ,r
and thus arrive at
Roffdiag
0
δE1n− =E2n+ . . .
...
δE1n− =EKn+
..
..
δE − =E +
.
.
0
3 X
3 X
R
2n 1n
X
..
.
.
.
.
..
..
..
..
≔
ωr
.
n− =1 n+ =1 r=1
..
..
.
0
δE(K−1)n− =EKn+
.
δEKn− =E1n+
...
. . . δEKn− =E(K−1)n+
0
|E1n− | (û1n− · ν1n− ) δû1n− ·ν1n− <0 · · · |E1n− | (û1n− · ν1n− ) δû1n− ·ν1n− <0
..
..
⊗ [R̂offdiag ]:,:,n− ,n+ ,r .
◦
.
.
|E Kn− | (ûKn− · νKn− ) δûKn− ·νKn− <0 · · · |E Kn− | (ûKn− · νKn− ) δûKn− ·νKn− <0
The sparsity structure for off-diagonal blocks depends on the numbering of mesh entities and is given for each combination of n− and n+ by the list markE0TE0T . Due to the upwind flux in the edge integrals, it is not possible here to
include the quadrature rule directly in the element-blocks R̂offdiag as opposed to the assembly of element integrals and
the assembly routines for edges in the first paper [3]. Implementing the assembly with above formulation is possible
but expensive, since the global matrix would have to to be built for every quadrature point. Instead we make use of
the fact that the sparsity structure is the same for every quadrature point as we solely rely on the element numbering
and do not account for the upwind direction when determining the structure of Roffdiag . We define a tensor of block
vectors R̃offdiag ∈ RKN×N×3×3
|E1n− | (û1n− · ν1n− ) δû1n− ·ν1n− <0
R
X
h
i
..
⊗ [R̂offdiag ]:,:,n− ,n+ ,r .
R̃offdiag
≔
ωr
.
(k−1)N+1:kN,:,n−,n+
r=1
|E Kn− | (ûKn− · νKn− ) δûKn− ·νKn− <0
For mb = rma , r ∈ N, let
· ⊗V · :
i
h
Rma ×na × Rmb ×nb ∋ (A, B) 7→ A ⊗V B ≔ [A]i, j [B](i−1)r : ir,: ∈ Rmb ×na nb
15
(41)
be an operator which can be interpreted as a Kronecker product (cf. Eq. (35)) that takes a different right-hand side for
every row of the left-hand side and is implemented in the routine kronVec. This allows us to write
0
δE1n− =E2n+ . . .
...
δE1n− =EKn+
..
..
δE − =E +
.
0
.
3 X
3
2n 1n
X
h
i
..
..
offdiag
.
.
.
offdiag
.
.
.
R
=
,
⊗V R̃
.
.
.
.
.
:,:,n−,n+
n− =1 n+ =1
.
.
..
..
0
δE(K−1)n− =EKn+
δEKn− =E1n+
...
. . . δEKn− =E(K−1)n+
0
which omits the expensive assembly of the global matrix in every quadrature point.
The routine assembleMatEdgePhiPhiValUpwind assembles the matrices Rdiag and Roffdiag directly into R with
a code very similar to the formulation above. To avoid repeated computation of the normal velocity u·νkn , we evaluate
it once in all quadrature points on each edge using the globally continuous function u(t, x) and store it in a dedicated
variable vNormalOnQuadEdge , which is then employed in the decision of the upwind direction using δu·νkn ≥0 and δu·νkn <0
as well as the normal velocity in the assembly of R and KD .
4.3.4. Assembly of KD
The entries of KD in (10) are transformed using transformation rule (32b) and mapping (39)
X Z
m
ϕki cD (t) u · νm
[KD ](k−1)N+i =
kn δu·νkn <0 dx
Ekn ∈∂T k ∩E∂Ω
=
X
Ekn ∈∂T k ∩E∂Ω
=
X
Ekn ∈∂T k ∩E∂Ω
Ekn
|Ekn |
|Ên |
|Ekn |
Z
Z
Ên
1
0
m
ϕ̂i cD (t, Fk ( x̂)) (u ◦ Fk ( x̂)) · νm
kn δu·νkn <0 d x̂
m
ϕ̂i ◦ γ̂n (s) cD (t, Fk ◦ γ̂n (s)) ûkn (s) · νm
kn δûkn ·νkn <0 ds ,
where we again implicitly assumed the application of Fk and γ̂n to δu·νkn ≥0 . This integral is then approximated using
a 1D quadrature rule (33) on the reference interval (0, 1)
[KD ](k−1)N+i ≈
X
Ekn ∈∂T k ∩E∂Ω
|Ekn |
R
X
r=1
m
ωr ϕ̂i ◦ γ̂n (qr ) cD (t, Fk ◦ γ̂n (qr )) ûkn (qr ) · νm
kn δûkn ·νkn <0
allowing to vectorize the computation over all triangles and resulting in the routine assembleVecEdgePhiIntFuncContVal .
4.4. Slope limiters
For the implementation of the slope limiters described in Sec. 3, three parts must be considered:
1. The assembly of the transformation matrix MDG,Taylor from Eq. (19);
2. the slope limiters themselves;
3. the selective mass lumping in the limiting of time-derivatives, as explained in Sec. 3.5.
4.4.1. Assembly of MDG,Taylor
The entries in MDG,Taylor are transformed using transformation rule (32a), and the integral is then approximated
using a 2D quadrature rule (33) on the reference triangle T̂
Z
Z
R
X
i
h
ϕki φk j dx = 2|T k | ϕ̂i φk j ◦ Fk ( x̂) d x̂ ≈ 2|T k |
ϕ̂i (qr ) φk j ◦ Fk (qr ) .
MDG,Taylor
=
(k−1)N+i,(k−1)N+ j
T̂
Tk
r=1
Recall that we cannot define the Taylor basis (17) on the reference triangle T̂ since the basis functions depend directly
on the physical coordinates of the element. The assembly is vectorized over all triangles resulting in the routine
assembleMatElemPhiDiscPhiTaylor . Matrices MDG,Taylor and M assembled in Sec. 4.3.1 are then used in the routines
projectDataDisc2DataTaylor and projectDataTaylor2DataDisc to transform the representation matrix between the
modal and Taylor basis—as explained in Sec. 3.1—by solving system (20).
16
4.4.2. Slope limiting operators
The slope limiters themselves are implemented in a generic manner according to Sec. 3. Beginning with the
highest-order derivatives, we evaluate the linear or the full reconstruction of these derivatives at all control points xki .
max
These values along with the centroid values are used to determine the minimum and maximum values cmin
ki and cki
(see Eq. (21)) for each control point and, from these, the element-wise correction factors αe according to Eq. (23)
or (25) are calculated. This is implemented in routine computeVertexBasedLimiter . Depending on the limiter type,
this value is then applied to a certain subset of degrees of freedom, and the computation is repeated for the next lowerorder derivatives. The full slope limiting operators Φ, ΦTaylor are provided by the functions applySlopeLimiterDisc
or applySlopeLimiterTaylor , respectively.
min
To overcome numerical problems for cases where cki ≈ cmax
ki or cki ≈ cki , we modify Eqs. (23) and (25) so that
max
min
the condition for cases 1 and 3 are modified to cki > cki − ǫ and cki < cki + ǫ for a small 0 < ǫ ∈ R. Additionally,
we increase the absolute value of the denominators in cases 1 and 3 by ǫ, i. e., add or subtract ǫ, respectively. This
makes our limiter slightly more strict than the one given by condition (21); ǫ = 10−8 has been found to be a suitable
value for double precision computations. Moreover, we enforce 0 ≤ αke ≤ 1 to overcome cases where the division by
a number close to zero might lead to values αke > 1. To reduce execution time, we perform the necessary evaluation
of the Taylor basis functions in all vertices only once and store the result in a global variable that is used in the slope
limiting routines. This is implemented in computeTaylorBasesV0T .
4.4.3. Limiting time-derivatives
To obtain the selectively lumped time-derivative C̃(i) required for update scheme (28), we compute the discrete
time derivative Ċ(i) as given in Sec. 3.5 and transform it to a Taylor basis representation using Eq. (20) (implemented
in routine projectDataDisc2DataTaylor ). We compute the stationary matrix
Mcorr ≔ M−1
L M
in the beginning (which is computationally cheap, since ML is a diagonal matrix) and calculate the selectively lumped
time-derivative C̃Taylor,(i) from Equation (27). Backtransformation to the modal DG basis is again cheap, since it
requires only the inverse of the diagonal matrix M (implemented in projectDataTaylor2DataDisc ) to produce the
selectively lumped time-derivative C̃(i) . Each intermediate solution C(i) in update scheme (28) is then obtained by
applying the slope limiting operator Φ implemented in applySlopeLimiterDisc .
4.5. Numerical results
4.5.1. Analytical convergence test
The code is verified by showing that the numerically estimated orders of convergences match the analytically
predicted ones for prescribed smooth solutions. To verify the spatial discretization, we restrict ourselves to the stationary version of (1) and investigate the impact of different slope limiting schemes on the convergence order. The
effectiveness of the slope limiters is verified in the next section.
We choose the exact solution c(x) ≔ cos(7x1 ) cos(7x2 ) and velocity field u(x) ≔ [exp((x1 + x2 )/2), exp((x1 −
2
x )/2)]T on the domain Ω ≔ (0, 1)2. The data cD and f are derived analytically by inserting c and u into (1). We
then compute the solution ch j for a sequence of increasingly finer meshes with element sizes h j , where the coarsest
grid Th0 covering Ω is an irregular grid, and each finer grid is obtained by regular refinement of its predecessor. The
discretization error kch (t) − c(t)kL2 (Ω) at time t ∈ J is computed as the L2 -norm of the difference between the discrete
solution ch (t) and the analytical solution c(t), as it was described in detail in our first paper [3]. Table 2 contains the
results demonstrating the experimental order of convergence α estimated using
!,
!
kch j−1 − ckL2 (Ω)
h j−1
α ≔ ln
ln
.
kch j − ckL2 (Ω)
hj
The interesting points to compare in Table 2 are the errors for different limiting schemes. Whereas the linear approximation does not get limited at all on finer grids—indicating that the vertex-based limiters are able to distinguish
between smooth and non-smooth solutions given sufficient mesh resolution, the case of higher order approximation
warrants a closer look. First of all, the linear limiter takes a heavy toll on the higher order degrees of freedom:
17
strict
hier. vert. based
linear
none
limiter
p
j
0
kch − ck
0
α
1
kch − ck
1
α
2
kch − ck
2
α
3
kch − ck
3
α
4
kch − ck
4
α
0
1
2
3
4
5
6
3.64e–1
5.64e–1
3.49e–1
2.49e–1
1.73e–1
1.09e–1
6.35e–2
—
–0.63
0.69
0.49
0.53
0.66
0.78
5.44e–1
4.25e–1
1.28e–1
3.65e–2
8.29e–3
1.93e–3
4.71e–4
—
0.36
1.73
1.81
2.14
2.11
2.03
6.98e–1
1.32e–1
2.82e–2
3.15e–3
3.68e–4
4.50e–5
5.59e–6
—
2.40
2.23
3.16
3.10
3.03
3.01
2.73e–1
8.51e–2
6.10e–3
3.87e–4
2.39e–5
1.50e–6
9.36e–8
—
1.68
3.80
3.98
4.02
4.00
4.00
2.21e–1
1.46e–2
1.25e–3
4.58e–5
1.51e–6
4.80e–8
1.51e–9
—
3.92
3.54
4.77
4.92
4.97
4.99
0
1
2
3
4
5
6
—
—
—
—
—
—
—
—
—
—
—
—
—
—
5.15e–1
4.53e–1
1.64e–1
4.53e–2
8.76e–3
1.98e–3
4.74e–4
—
0.19
1.47
1.85
2.37
2.14
2.07
7.21e–1
3.55e–1
1.34e–1
3.12e–2
4.48e–3
6.20e–4
7.63e–5
—
1.02
1.41
2.10
2.80
2.85
3.02
4.60e–1
3.54e–1
1.51e–1
3.12e–2
4.46e–3
6.19e–4
7.61e–5
—
0.38
1.23
2.27
2.81
2.85
3.02
5.27e–1
3.59e–1
1.47e–1
3.12e–2
4.46e–3
6.19e–4
7.61e–5
—
0.55
1.29
2.23
2.81
2.85
3.02
0
1
2
3
4
5
6
—
—
—
—
—
—
—
—
—
—
—
—
—
—
5.15e–1
4.53e–1
1.64e–1
4.53e–2
8.76e–3
1.98e–3
4.74e–4
—
0.19
1.47
1.85
2.37
2.14
2.07
6.98e–1
2.47e–1
1.06e–1
7.87e–3
6.39e–4
5.32e–5
5.94e–6
—
1.50
1.22
3.76
3.62
3.59
3.16
2.73e–1
2.57e–1
5.89e–2
4.02e–3
6.16e–5
2.62e–6
1.17e–7
—
0.09
2.12
3.87
6.03
4.56
4.48
2.21e–1
1.33e–1
5.10e–2
3.17e–3
6.53e–6
1.08e–7
5.48e–9
—
0.73
1.38
4.01
8.92
5.92
4.30
0
1
2
3
4
5
6
—
—
—
—
—
—
—
—
—
—
—
—
—
—
5.15e–1
4.53e–1
1.64e–1
4.53e–2
8.76e–3
1.98e–3
4.74e–4
—
0.19
1.47
1.85
2.37
2.14
2.07
6.84e–1
3.79e–1
1.14e–1
1.23e–2
1.15e–3
1.19e–4
6.64e–6
—
0.85
1.74
3.21
3.41
3.27
4.17
3.28e–1
3.69e–1
6.74e–2
1.20e–2
8.99e–4
1.33e–4
1.09e–6
—
–0.17
2.45
2.49
3.74
2.76
6.93
3.56e–1
3.54e–1
6.15e–2
1.16e–2
8.85e–4
1.33e–4
1.13e–6
—
0.00
2.53
2.40
3.72
2.74
6.88
Table 2: Discretization errors measured in L2 (Ω) and estimated orders of convergences for different polynomial degrees and limiter types. We have h j = 3·21 j and K = 36 · 4 j triangles in the jth refinement level.
quadratic, cubic, and quartic solutions produce virtually the same error, thus negating the effect of a more accurate
DG solution. The strict limiter seems to flatten out after the cubic approximation, thus gaining one order of convergence on the linear limiter. The hierarchical limiter of Kuzmin, however, appears to perform very well for all tested
meshes and approximation orders having a very small effect on the errors of the analytical test case.
4.5.2. Solid body rotation
As a benchmark problem, we use solid body rotation test proposed by LeVeque [24], which is often used to
investigate limiter performance [7, 13]. It consists of a slotted cylinder, a sharp cone, and a smooth hump (see
Figure 4a) that are placed in a square domain Ω = [0, 1]2 and transported by a time-independent velocity field u(x) =
T
1
[0.5 − x2 , x1 − 0.5] in a counterclockwise rotation over J = (0, 2π). With r = 0.0225 and G(x, x0) ≔ 0.15
kx − x0 k2 ,
we choose initial data satisfying
(x1 − 0.5)2 + (x2 − 0.75)2 ≤ r
1
if
(slotted
cylinder)
1
1
2
∧
(x
≤
0.475
∨
x
≥
0.525
∨
x
≥
0.85)
T
1
2
2
2
0
1
−
G(x,
[0.5,
0.25]
)
if
(x
−
0.5)
+
(x
−
0.25)
≤
r
(sharp
cone)
c (x) =
T
1
1
2
2
2
(1
+
cos(πG(x,
[0.25,
0.5]
))
if
(x
−
0.25)
+
(x
−
0.5)
≤
r
(smooth
hump)
4
0
otherwise
and zero boundary cD = 0 and right-hand side f = 0.
18
(a) Projected and limited initial data.
(b) Linear limiter, no lumped time-derivative.
(c) Hier. vert.-based lim., no lumped time-derivative.
(d) Linear limiter, lumped time derivative.
(e) Hier. vert.-based limiter, lumped time derivative.
(f) Strict limiter, lumped time derivative.
Figure 4: Visualization of the DG solutions with p = 2 at end time tend = 2π.
First, we would like to emphasize the huge improvement of the solution when applying Kuzmin’s selectively
lumped time-stepping scheme (cf. Sec. 3.5), which reduces numerical diffusion and peak clipping, as visible in the
intersection lines in Fig. 5 and the 3D visualization (Figs. 4b–4e) and results in a smaller error (see Tab. 3). Without
this technique, the implicit coupling of the derivatives renders the higher order limiter inferior to the linear limiter.
19
1
0.8
y = 0.25
yy==0.75
0.75
The results for the different slope limiters (see Fig. 4d–4f) when applying the lumped time-stepping scheme are
similar with all limiters producing errors in the same range (see Tab. 3). The linear vertex based limiter exhibits
significantly stronger peak clipping than the higher order limiters in the intersection lines of Fig. 6. When looking at
the slotted cylinder the strict limiter outperforms the others, producing less fill-in in the slot and better preserving the
shape of the cylinder. Also, the total error is lowest for the strict limiter.
0.6
11
x = 0.5
1
0.4
0.8
0.8
y = 0.75
y = 0.25
0.8
1
0.6
0.6
x = 0.25
0.4
0.4
0.2
0.21
0.2
0
0
0.6
0.6
0.4
0.4
0
0.2
0.4
0.6
x = 0.25
0.8
1
0
0.5
1
0.4
0.8
0.8
0
0.3
0.6
0.2
0.4
0.1
0.2
0
0.6
x = 0.5
0.8
0
initial
0
0.2
0.4
linear non-lumped
0.6
0.8
0.2
1
linear lumped
0.2
0
0
0.2
0.4
0.6
x = 0.5
0.4
0.4
0.6
hier. vert. based non-lump.
0.8
0.8
1
1
0.6
0.8
hier. vert. based lump.
0.2
0.2
Figure 5: Intersection lines of the DG solution with p = 2 at end time tend = 2π with and without the selectively
lumped time-derivative, as described in Section. 3.5. The large deviation of the solutions from the initial data in the
lower right plot are due to fill-in of the slot in the cylinder.
0.4
For polynomial degrees p ≥ 2 the solution might still violate the bounds along the edges when using only control
points in the vertices of the element for the slope limiting procedure. When evaluating the numerical fluxes, these
values are transported to the neighboring elements and can lead to cell averages lying outside of the initial bounds in
the next time step. Table 3 shows that all limiter types suffer from this, however these small violations of the bounds
usually smooth out over time due to the effects of numerical diffusion and do not introduce numerical problems. The
linear limiter is the only that preserves the upper bound but comes at the price of stronger peak-clipping. To effectively
restrict the solution to these bounds for p = 2, additional control points on the edges are necessary, which leads to
a significant increase in numerical diffusion, hence this is not a suitable technique.
00
0
0.2
0.2
00
5. Register of Routines
0.4
0.6
0.8
0.2
0.2 0.4
0.4 0.6
0.6
1
0.8
0.8
We list here all routines of our implementation that were added since the first paper [3] in alphabetic order. For
the reason of compactness, we waive the check for correct function arguments, e. g., by means of routines as assert.
However, it is strongly recommended to catch exceptions if the code is to be extended. The argument g is always
0.2
0
0
0.4
0.2
0
0.4
0.6
20
0.2
0.6
0.8
0.4
0.8
1
0.6
1
0.
y = 0.25
yy==0.75
0.75
0.6
y = 0.75
y = 0.25
11
1
0.4
1
x = 0.5
0.8
0.6
0.6
x = 0.25
0.4
0.8
0.8
0.4
0.2
0.21
0.2
0
0
0.6
0.6
0
0.2
0.4
0.6
x = 0.25
0.8
1
0
0.5
1
0.4
0
0.8
0.8
0.3
0.6
0.2
0.4
0.1
0.2
0.4
0.4
0
0
0.6
x = 0.5
0.8
0
0.2
0.2
0.4
0.6
0.8
initial
linear
0.4
0.6
x = 0.5
0.8
1
0.4
0
0
1
0.2
0.2
hier. vert. based
0.4
0.6
0.6
0.8
0.8
1
strict
Figure 6: Intersection lines of the DG solution with p = 2 at end time tend = 2π when applying the lumped timestepping scheme (cf. Sec. 3.5). Compared are solutions with the different limiters described in Section 3.
0.2
0.2
0.4
0
mink ch (xkc )
mink,i ch (xki )
mink, j ch (xe,k j )
maxk ch (xkc )
maxk,i ch (xki )
maxk, j ch (xe,k j )
kch (0) − c0 kL2 (Ω)
kch (tend ) − c0 kL2 (Ω)
linear
non-lumped
hier. vertex-based
0.00
–2.93e–2
–3.32e–8
1.00000
1.00000
1.00000
3.73e–2
8.18e–2
–1.11e–5
–2.47e–2
–4.28e–3
1.00000
1.00915
1.00014
3.66e–2
1.15e–1
linear
lumped
hier. vertex-based
strict
–3.95e–10
–9.53e–3
–1.18e–7
1.00000
1.00000
1.00000
3.73e–2
7.38e–2
–6.18e–5
–9.69e–3
–4.28e–3
1.00000
1.01760
1.00297
3.66e–2
7.40e–2
–6.42e–4
–9.79e–16
–5.35e–3
1.00142
1.00000
1.00296
3.66e–2
7.07e–2
00
0.2 0.4
0.2
0.6
0.8
1
00
0.2
0.2 0.4
0.4 0.6
0.6 0.8
0.8
Table 3: Minimum and maximum values in centroids xkc , nodes xki , and edge midpoints xe,k j with p = 2 over the
entire simulation time J = (0, 2π). Two last rows give the L2 -errors of the slope-limited projection of initial data and
the end solution when compared to the analytical initial data.
0
0.4
a struct representing the triangulation Th , the argument N is always the number of local basis functions N. A script
that demonstrates the application of the presented routines is given in mainAdvection.m.
Additionally, we recommend using Laurent Sorbers slightly faster implementation of the Kronecker product [25]
for sparse matrices from MATLAB File Exchange1 to speed up the computation. Note that this implementation is
only suitable for logical and double matrices.
0.2
0
0
0.4
0.6
0.2
0.6
1 http://www.mathworks.com/matlabcentral/fileexchange/
0.2
21
0.8
0.4
0.8
1
0.6
1
0.
dataDisc = applySlopeLimiterDisc(g, dataDisc, markV0TbdrD, dataV0T, globM, globMDiscTaylor, type)
implements the slope limiting operator Φ, as given in Eq. (29) with the chosen limiter type given as a string in type.
Parameter dataDisc is the representation matrix in modal DG basis, markV0TbdrD is a logical matrix marking all
Dirichlet boundary nodes for which boundary data is given in dataV0T.
function dataDisc = a p p l y S l o p e L i m i t e r D i s c(g , dataDisc , markV0TbdrD , dataV0T , globM , globMDiscTaylor , type )
d a t a T a y l o r = p r o j e c t D a t a D i s c 2 D a t a T a y l o r( dataDisc , globM , g l o b M D i s c T a y l o r) ;
d a t a T a y l o r = a p p l y S l o p e L i m i t e r T a y l o r(g , dataTaylor , markV0TbdrD , dataV0T , type ) ;
dataDisc = p r o j e c t D a t a T a y l o r 2 D a t a D i s c( dataTaylor , globM , g l o b M D i s c T a y l o r) ;
end
dataTaylor = applySlopeLimiterTaylor(g, dataTaylor, markV0TbdrD, dataV0T, type) implements the slope
limiting operator ΦTaylor , as described in Sec. 3.5. Parameter dataTaylor is the representation matrix in Taylor basis,
the other parameters are the same as for applySlopeLimiterDisc .
function d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r(g , dataTaylor , markV0TbdrD , dataV0T , type )
switch type
case ' linear '
d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r L i n e a r(g , dataTaylor , markV0TbdrD , dataV0T ) ;
case ' h i e r a r c h _ v e r t'
d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r H i e r a r c h i c a l V e r t e x(g , dataTaylor , markV0TbdrD , dataV0T ) ;
case ' strict '
d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r S t r i c t(g , dataTaylor , markV0TbdrD , dataV0T ) ;
otherwise
error ( ' Unknown limiter type ') ;
end
end
dataTaylorLim = applySlopeLimiterTaylorHierarchicalVertex(g, dataTaylor, markV0TbdrD, dataV0T)
applies the hierarchical vertex-based limiter as described in Sec. 3.3, with input parameters as for
applySlopeLimiterTaylor .
function d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r H i e r a r c h i c a l V e r t e x(g , dataTaylor , markV0TbdrD , dataV0T )
global g P h i T a y l o r V 0 T
[K , N ] = size ( d a t a T a y l o r) ; p = ( sqrt (8* N +1) -3) /2;
d a t a T a y l o r L i m = zeros ( size ( d a t a T a y l o r) ) ; d a t a T a y l o r L i m(: , 1) = d a t a T a y l o r(: , 1) ;
alpha = zeros (K , 1) ;
for ord = p : -1 : 1
alphaOrd = ones ( K , 1) ;
indDOF = ord *( ord +1) /2 + 1 : ( ord +1) *( ord +2) /2;
for i = 1 : ord
mult = bsxfun ( @plus , [ ord - i , i - 1] , m u l t i i n d e x (1) ) ;
ind = mult2ind ( mult ) ;
valV0T = c o m p u t e F u n c D i s c A t P o i n t s( d a t a T a y l o r(: , ind ) , g P h i T a y l o r V 0 T(: , : , 1:3) ) ;
if ord > 1
alphaTmp = c o m p u t e V e r t e x B a s e d L i m i t e r(g , d a t a T a y l o r(: , ind (1) ) , valV0T , markV0TbdrD , valV0T ) ;
else
alphaTmp = c o m p u t e V e r t e x B a s e d L i m i t e r(g , d a t a T a y l o r(: , ind (1) ) , valV0T , markV0TbdrD , dataV0T ) ;
end
alphaOrd = min ( alphaOrd , alphaTmp ) ;
end
alpha = max ( alpha , alphaOrd ) ;
d a t a T a y l o r L i m(: , indDOF ) = bsxfun ( @times , alpha , d a t a T a y l o r(: , indDOF ) ) ;
end
end % f u n c t i o n
dataTaylorLim = applySlopeLimiterTaylorLinear(g, dataTaylor, markV0TbdrD, dataV0T) applies the linear
vertex-based limiter as described in Sec. 3.2, with input parameters as for applySlopeLimiterTaylor .
function d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r L i n e a r(g , dataTaylor , markV0TbdrD , dataV0T )
global g P h i T a y l o r V 0 T
alphaE = c o m p u t e V e r t e x B a s e d L i m i t e r(g , d a t a T a y l o r(: , 1) , c o m p u t e F u n c D i s c A t P o i n t s( d a t a T a y l o r(: , 1:3) ,
֒→g P h i T a y l o r V 0 T(: ,: ,1:3)) , markV0TbdrD , dataV0T ) ;
d a t a T a y l o r L i m = [ d a t a T a y l o r(: ,1) , bsxfun ( @times , alphaE , d a t a T a y l o r(: , 2:3) ) , bsxfun ( @times , alphaE ==
֒→1 , d a t a T a y l o r(: , 4: end ) ) ];
end % f u n c t i o n
22
dataTaylorLim = applySlopeLimiterTaylorStrict(g,dataTaylor, markV0TbdrD, dataV0T) applies the stricter
form of the hierarchical vertex-based limiter as described in Sec. 3.4, with input parameters as for
applySlopeLimiterTaylor .
function d a t a T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r S t r i c t(g , dataTaylor , markV0TbdrD , dataV0T )
global g P h i T a y l o r V 0 T
[K , N ] = size ( d a t a T a y l o r) ; p = ( sqrt (8* N +1) -3) /2;
d a t a T a y l o r L i m = d a t a T a y l o r;
for ord = p : -1 : 1
alphaOrd = ones ( K , 1) ;
indDOF = ord * ( ord + 1) / 2 + 1 : N ;
for i = 1 : ord
p R e c o n s t r u c t i o n = p - ord + 1;
mult = bsxfun ( @plus , [ ord - i , i - 1] , m u l t i i n d e x( p R e c o n s t r u c t i o n) ) ; ind = mult2ind ( mult ) ;
valV0T = c o m p u t e F u n c D i s c A t P o i n t s( d a t a T a y l o r L i m(: , ind ) , g P h i T a y l o r V 0 T(: , : , 1 : size ( ind , 1) ) ) ;
if ord > 1
alphaTmp = c o m p u t e V e r t e x B a s e d L i m i t e r(g , d a t a T a y l o r L i m(: , ind (1) ) , valV0T , markV0TbdrD , valV0T ) ;
else
alphaTmp = c o m p u t e V e r t e x B a s e d L i m i t e r(g , d a t a T a y l o r L i m(: , ind (1) ) , valV0T , markV0TbdrD , dataV0T ) ;
end
alphaOrd = min ( alphaOrd , alphaTmp ) ;
end
d a t a T a y l o r L i m(: , indDOF ) = bsxfun ( @times , alphaOrd , d a t a T a y l o r L i m(: , indDOF ) ) ;
end
end % f u n c t i o n
ret = assembleMatEdgePhiPhiValUpwind(g, refEdgePhiIntPhiInt, refEdgePhiIntPhiExt, valOnQuad) assem-
bles the matrix R according to Sec. 4.3.3, containing edge integrals of products of two basis functions multiplied with
for each quadrature point specified values, stored in valOnQuad, and where the upwind-sided value w. r. t. valOnQuad
is chosen. The input arguments refEdgePhiIntPhiInt and refEdgePhiIntPhiExt provide the local matrices R̂diag
and R̂offdiag (multidimensional arrays), respectively.
function ret = a s s e m b l e M a t E d g e P h i P h i V a l U p w i n d(g , r e f E d g e P h i I n t P h iI nt O nQ ua d , r e f E d g e P h i I n t P h i Ex tO n Qu ad ,
֒→v a l O n Q u a d)
K = g . numT ; N = size ( r e f E d g e P h i I n t P h i In t On Qu a d , 1) ;
ret = sparse ( K *N , K * N ) ;
p = ( sqrt (8* N +1) -3) /2;
qOrd = 2* p +1; [~ , W ] = q u a d R u l e 1 D( qOrd ) ;
for nn = 1 : 3
Rkn = g . areaE0T (: , nn ) ;
for r = 1 : length ( W ) % D i a g o n a l b l o c k s
ret = ret + kron ( spdiags ( W ( r ) .* Rkn .* v a l O n Q u a d(: , nn , r ) .* ( v a l O n Q u a d(: , nn , r ) > 0) , 0 , K , K ) ,
֒→r e f E d g e P h i I n t P h i I n t O n Q u a d(: , : , nn , r ) ) ;
end
for np = 1 : 3 % Off - d i a g o n a l b l o c k s
R k n T i m e s V a l = sparse ( K *N , N ) ;
for r = 1 : length ( W )
R k n T i m e s V a l = R k n T i m e s V a l + kron ( W ( r ) .* Rkn .* v a l O n Q u a d(: , nn , r ) .* sparse ( v a l O n Q u a d(: , nn , r )
֒→< 0) , r e f E d g e P h i I n t P h i E x t O n Q u a d(: , : , nn , np , r ) ) ;
end
ret = ret + kronVec ( g . m a r k E 0 T E 0 T{ nn , np } , R k n T i m e s V a l) ;
end % for
end % for
end % f u n c t i o n
assembles two
matrices, each containing integrals of products of a basis function with a (spatial) derivative of a basis function
and with a component of a discontinuous coefficient function whose coefficients are specified in dataDisc1
and dataDisc2, respectively. The matrices are returned in a 2 × 1 cell variable. This corresponds to the matrices Gm ,
m ∈ {1, 2} according to Sec. 4.3.2. The input argument refElemDphiPhiPhi stores the local matrices Ĝ (multidimensional array) as defined in (37) and can be computed by integrateRefElemDphiPhiPhi [3]. The coefficients
of the projection of the algebraic diffusion coefficient d into the broken polynomial space are stored in the input
arguments dataDisc1 and dataDisc2 and can be computed by projectFuncCont2dataDisc .
ret = assembleMatElemDphiPhiFuncDiscVec(g, refElemDphiPhiPhi, dataDisc1, dataDisc2)
function ret = a s s e m b l e M a t E l e m D p h i P h i F u n c D i s c V e c(g , r e f E l e m D p h iPh iP hi , dataDisc1 , d a t a D i s c 2)
[K , N ] = size ( d a t a D i s c 1) ;
ret = cell (2 , 1) ; ret {1} = sparse ( K *N , K * N ) ; ret {2} = sparse ( K *N , K * N ) ;
for l = 1 : N
ret {1} = ret {1} + kron ( spdiags ( d a t a D i s c 1(: , l ) .* g . B (: ,2 ,2) , 0 ,K , K ) , r e f E l e m D p h i P h i P h i(: ,: ,l ,1) ) ...
23
- kron ( spdiags ( d a t a D i s c 1(: , l ) .* g . B (: ,2 ,1) , 0 ,K , K ) , r e f E l e m D p h i P h i P h i(: ,: ,l ,2) ) ;
ret {2} = ret {2} - kron ( spdiags ( d a t a D i s c 2(: , l ) .* g . B (: ,1 ,2) , 0 ,K , K ) , r e f E l e m D p h i P h i P h i(: ,: ,l ,1) ) ...
+ kron ( spdiags ( d a t a D i s c 2(: , l ) .* g . B (: ,1 ,1) , 0 ,K , K ) , r e f E l e m D p h i P h i P h i(: ,: ,l ,2) ) ;
end % for
end % f u n c t i o n
ret = assembleMatElemPhiDiscPhiTaylor(g, N) assembles the matrix MDG,Taylor according to Sec. 3.1, which cor-
responds to the basis transformation matrix with one basis function from each, modal and Taylor basis. It is required for the transformation between modal and Taylor basis in the routines projectDataDisc2DataTaylor and
projectDataTaylor2DataDisc .
function ret = a s s e m b l e M a t E l e m P h i D i s c P h i T a y l o r(g , N )
global gPhi2D
p = ( sqrt (8* N +1) -3) /2; qOrd = max (2* p +1 , 1) ; [ Q1 , Q2 , W ] = q u a d R u l e 2 D( qOrd ) ;
K = g . numT ; ret = sparse ( K *N , K * N ) ;
for i = 1 : N
for j = 1 : N
i n t P h i I P h i J = ( repmat ( gPhi2D { qOrd }(: , i ) ', [ K 1]) .* p h i T a y l o r R e f(g , j , Q1 , Q2 ) ) * W ';
ret = ret + sparse ( i : N : K *N , j : N : K *N , 2 * g . areaT .* intPhiIPhiJ , K *N , K * N ) ;
end % for
end % for
end % f u n c t i o n
ret = assembleMatElemPhiTaylorPhiTaylor(g, N) assembles the mass matrix in Taylor basis MTaylor .
function ret = a s s e m b l e M a t E l e m P h i T a y l o r P h i T a y l o r(g , N )
p = ( sqrt (8* N +1) -3) /2; qOrd = max (2*p , 1) ;
[ Q1 , Q2 , W ] = q u a d R u l e 2 D( qOrd ) ;
K = g . numT ;
ret = sparse ( K *N , K * N ) ;
for i = 1 : N
for j = 1 : N
i n t P h i I P h i J = ( p h i T a y l o r R e f(g , i , Q1 , Q2 ) .* p h i T a y l o r R e f(g , j , Q1 , Q2 ) ) * W ';
ret = ret + sparse ( i : N : K *N , j : N : K *N , 2 * g . areaT .* intPhiIPhiJ , K *N , K * N ) ;
end % for
end % for
end % f u n c t i o n
ret = assembleVecEdgePhiIntFuncContVal(g, markE0Tbdr, funcCont, valOnQuad, N) assembles a vector con-
taining integrals of products of a basis function with a continuous function and a given value that is provided in
each quadrature point on each edge for all triangles. This corresponds to the contributions of Dirichlet boundaries KD
to the right-hand side of (4) according to Sec. 4.3.4. markE0Tbdr marks the boundary edges on which the vector
should be assembled, funcCont is a function handle and valOnQuad is the value provided in each quadrature point, as
computed by computeFuncContNuOnQuadEdge .
function ret = a s s e m b l e V e c E d g e P h i I n t F u n c C o n t V a l(g , markE0Tbdr , funcCont , valOnQuad , N )
global gPhi1D
K = g . numT ; p = ( sqrt (8* N +1) -3) /2; qOrd = 2* p +1; [Q , W ] = q u a d R u l e 1 D( qOrd ) ;
Q2X1 = @ ( X1 , X2 ) g . B (: ,1 ,1) * X1 + g . B (: ,1 ,2) * X2 + g . coordV0T (: ,1 ,1) * ones ( size ( X1 ) ) ;
Q2X2 = @ ( X1 , X2 ) g . B (: ,2 ,1) * X1 + g . B (: ,2 ,2) * X2 + g . coordV0T (: ,1 ,2) * ones ( size ( X1 ) ) ;
ret = zeros (K , N ) ;
for n = 1 : 3
[ Q1 , Q2 ] = gammaMap (n , Q ) ; f u n c O n Q u a d = funcCont ( Q2X1 ( Q1 , Q2 ) , Q2X2 ( Q1 , Q2 ) ) ;
Kkn = m a r k E 0 T b d r(: , n ) .* g . areaE0T (: , n ) ;
for i = 1 : N
integral = ( f u n c O n Q u a d .* squeeze (( v a l O n Q u a d(: , n , :) < 0) .* v a l O n Q u a d(: , n , :) ) ) * ( W ' .* gPhi1D {
֒→qOrd }(: , i , n ) ) ;
ret (: , i ) = ret (: , i ) + Kkn .* integral ;
end % for
end % for
ret = reshape ( ret ' , K *N ,1) ;
end % f u n c t i o n
ret = computeFuncContNuOnQuadEdge(g, funcCont1, funcCont2, qOrd) assembles a three-dimensional array with
the normal velocity u · νki evaluated in all quadrature points of all edges of each triangle.
function ret = c o m p u t e F u n c C o n t N u O n Q u a d E d g e(g , funcCont1 , funcCont2 , qOrd )
K = g . numT ; [Q , W ] = q u a d R u l e 1 D( qOrd ) ;
24
ret = zeros (K , 3 , length ( W ) ) ;
for n = 1 : 3
[ Q1 , Q2 ] = gammaMap (n , Q ) ;
ret (: ,n ,:) = bsxfun ( @times , g . nuE0T (: , n ,1) , f u n c C o n t 1( g . m a p R e f 2 P h y(1 , Q1 , Q2 ) ,g . m a p R e f 2 P h y(2 , Q1 , Q2 ) ) ) +...
bsxfun ( @times , g . nuE0T (: , n ,2) , f u n c C o n t 2( g . m a p R e f 2 P h y(1 , Q1 , Q2 ) ,g . m a p R e f 2 P h y(2 , Q1 , Q2 ) ) ) ;
end % for
end
valV0T = computeFuncContV0T(g, funcCont) assembles a matrix containing the function funcCont evaluated in
each node of each triangle.
function valV0T = c o m p u t e F u n c C o n t V 0 T(g , funcCont )
valV0T = zeros ( g . numT ,3) ;
for n = 1 : 3
valV0T (: , n ) = funcCont ( g . coordV0T (: , n , 1) , g . coordV0T (: , n , 2) ) ;
end
end
ret = computeFuncDiscAtPoints(funcDisc, phiAtPoints) assembles a matrix containing the values of a discrete
function with representation matrix stored in funcDisc evaluated in all points, for which the evaluated basis functions
are given in phiAtPoints.
function ret = c o m p u t e F u n c D i s c A t P o i n t s( funcDisc , p h i A t P o i n t s)
nPoints = size ( phiAtPoints , 2) ; K = size ( funcDisc , 1) ;
ret = zeros (K , nPoints ) ;
for i = 1 : nPoints
ret (: , i ) = sum ( funcDisc .* squeeze ( p h i A t P o i n t s(: , i , :) ) , 2) ;
end % for
end % f u n c t i o n
minMaxV0T = computeMinMaxV0TElementPatch(g, valCentroid, markV0TbdrD, dataV0T) determines a matrix
min
with bounds cmax
ki , cki of Eq. (22) for each vertex of each triangle, as required by computeVertexBasedLimiter .
function m i n M a x V 0 T = c o m p u t e M i n M a x V 0 T E l e m e n t P a t c h(g , valCentroid , markV0TbdrD , dataV0T )
m i n M a x V 0 T = cell (2 ,1) ;
m i n M a x V 0 T{1} = zeros ( g . numT , 3) ; m i n M a x V 0 T{2} = zeros ( g . numT , 3) ;
s h i f t C e n t r o i d P o s = abs ( min ( v a l C e n t r o i d) ) + 1;
s h i f t C e n t r o i d N e g = abs ( max ( v a l C e n t r o i d) ) + 1;
v a l C e n t r o i d P o s = v a l C e n t r o i d + s h i f t C e n t r o i d P o s;
v a l C e n t r o i d N e g = v a l C e n t r o i d - s h i f t C e n t r o i d N e g;
valD = NaN ( g . numT , 3) ; valD ( m a r k V 0 T b d r D) = dataV0T ( m a r k V 0 T b d r D) ;
for i = 1 : 3
m a r k N b V 0 T = g . m a r k V 0 T V 0 T{i , 1} | g . m a r k V 0 T V 0 T{i , 2} | g . m a r k V 0 T V 0 T{i , 3};
if exist ( ' O C T A V E _ V E R S I O N' , ' builtin ')
m a r k N b V 0 T = m a r k N b V 0 T + 0 * speye ( size ( markNbV0T , 1) , size ( markNbV0T , 2) ) ;
end
m i n M a x V 0 T {1}(: , i ) = min ( min ( bsxfun ( @times , markNbV0T , valCentroidNeg ') ,[] ,2) + shiftCentroidNeg , valD (: , i ) ) ;
m i n M a x V 0 T {2}(: , i ) = max ( max ( bsxfun ( @times , markNbV0T , valCentroidPos ') ,[] ,2) - shiftCentroidPos , valD (: , i ) ) ;
end % for
end % f u n c t i o n
computeTaylorBasesV0T(g, N) evaluates the Taylor basis functions Φk j in all vertices of all triangles xki and stores
them in a global multidimensional array.
function c o m p u t e T a y l o r B a s e s V 0 T(g , N )
global g P h i T a y l o r V 0 T
g P h i T a y l o r V 0 T = zeros ( g . numT , 3 , N ) ;
for n = 1 : 3
for i = 1 : N
g P h i T a y l o r V 0 T(: , n , i ) = p h i T a y l o r P h y(g , i , g . coordV0T (: , n , 1) , g . coordV0T (: , n , 2) ) ;
end
end
end
alphaE = computeVertexBasedLimiter(g, valCentroid, valV0T, markV0TbdrD, dataV0T) computes a vector
with correction factors αke (cf. Eq. (23)) for all elements. Centroid values ckc are given in valCentroid , values of
the unconstrained reconstruction cki are specified in valV0T and markV0TbdrD is a logical matrix marking all Dirichlet
boundary nodes for which boundary data is given in dataV0T.
25
function alphaE = c o m p u t e V e r t e x B a s e d L i m i t e r(g , valCentroid , valV0T , markV0TbdrD , dataV0T )
m i n M a x V 0 T = c o m p u t e M i n M a x V 0 T E l e m e n t P a t c h(g , valCentroid , markV0TbdrD , dataV0T ) ;
d i f f V 0 T C e n t r o i d = valV0T - repmat ( valCentroid , [1 3]) ;
d i f f M i n C e n t r o i d = m i n M a x V 0 T {1} - repmat ( valCentroid , [1 3]) ;
d i f f M a x C e n t r o i d = m i n M a x V 0 T {2} - repmat ( valCentroid , [1 3]) ;
tol = 1. e -8;
markNeg = d i f f V 0 T C e n t r o i d < d i f f M i n C e n t r o i d + tol ;
markPos = d i f f V 0 T C e n t r o i d > d i f f M a x C e n t r o i d - tol ;
a l p h a E V 0 T = ones ( g . numT ,3) ;
a l p h a E V 0 T( markNeg ) = max (0 , min (1 , d i f f M i n C e n t r o i d( markNeg ) ./ ( d i f f V 0 T C e n t r o i d( markNeg ) - tol ) ) ) ;
a l p h a E V 0 T( markPos ) = max (0 , min (1 , d i f f M a x C e n t r o i d( markPos ) ./ ( d i f f V 0 T C e n t r o i d( markPos ) + tol ) ) ) ;
alphaE = min ( alphaEV0T ,[] ,2) ;
end % f u n c t i o n
ret = integrateRefEdgePhiIntPhiExtPerQuad(N) computes a multidimensional array of functionals in the quadra-
ture points on the edges of the reference triangle T̂ that consist of all permutations of two basis functions of which one
belongs to a neighboring element that is transformed using ϑ̂. This corresponds to the local matrix R̂offdiag as given
in (40).
function ret = i n t e g r a t e R e f E d g e P h i I n t P h i E x t P e r Q u a d( N )
global gPhi1D g T h e t a P h i 1 D
p = ( sqrt (8* N +1) -3) /2; qOrd = 2* p +1; [~ , W ] = q u a d R u l e 1 D( qOrd ) ;
ret = zeros (N ,N ,3 ,3 , length ( W ) ) ; % [ N x N x N x 3 x 3]
for nn = 1 : 3 % 3 edges
for np = 1 : 3
for i = 1 : N
for j = 1 : N
ret (i , j , nn , np , :) = gPhi1D { qOrd }(: , i , nn ) .* g T h e t a P h i 1 D{ qOrd }(: , j , nn , np ) ;
end % for
end % for
end % for
end % for
end
ret = integrateRefEdgePhiIntPhiIntPerQuad(N) computes a multidimensional array of functionals in the quadra-
ture points on the edges of the reference triangle T̂ that consist of all permutations of two basis functions. This
corresponds to the local matrix R̂diag as given in (38).
function ret = i n t e g r a t e R e f E d g e P h i I n t P h i I n t P e r Q u a d( N )
global gPhi1D
p = ( sqrt (8* N +1) -3) /2; qOrd = max (2* p +1 ,1) ; [~ , W ] = q u a d R u l e 1 D( qOrd ) ;
ret = zeros (N , N , 3 , length ( W ) ) ; % [ N x N x 3 x R ]
for n = 1 : 3 % 3 edges
for i = 1 : N
for j = 1 : N
ret (i ,j ,n ,:) = gPhi1D { qOrd }(: , i , n ) .* gPhi1D { qOrd }(: , j , n ) ;
end % for
end % for
end % for
end % f u n c t i o n
K = kronVec(A,B) computes the result of Rmb ×na nb ∋ K = A ⊗V B as given in Eq. (41).
function K = kronVec (A , B )
[ ma , na ] = size ( A ) ;
[ mb , nb ] = size ( B ) ;
mc = mb / ma ;
if ~ issparse ( A ) && ~ issparse ( B ) % Both i n p u t s full , r e s u l t is full .
A = reshape (A , [1 ma 1 na ]) ;
B = reshape (B , [ mc ma nb 1]) ;
K = reshape ( bsxfun ( @times , A , B ) , [ mb na * nb ]) ;
else
[ i2 , j2 , v2 ] = find ( kron (A , ones ( mc , 1) ) ) ;
ik = repmat ( i2 , [1 nb ]) ;
jk = bsxfun ( @plus , nb * ( j2 - 1) , 1 : nb ) ;
sk = bsxfun ( @times , v2 , B ( i2 , :) ) ;
K = sparse ( ik , jk , sk , mb , na * nb ) ;
end % if
end % f u n c t i o n
26
mainAdvection.m This is the main script to solve (1) which can be used as a template for further modifications.
Modifiable parameters are found in Lines 5–16, the problem data (initial condition, velocity, right-hand side and
boundary data) is specified in Lines 35–42.
function m a i n A d v e c t i o n()
more off % d i s a b l e p a g i n g of o u t p u t
tic % Start time m e a s u r e m e n t
% % P a r a m e t e r s.
hmax
= 2^ -6;
% m a x i m u m edge l e n g t h of t r i a n g l e
p
= 2;
% local p o l y n o m i a l d e g r e e
ordRK
= min ( p +1 ,3) ;
% order of Runge Kutta time s t e p p e r.
tEnd
= 2* pi ;
% end time
numSteps
= 3142;
% n u m b e r of time steps
isVisGrid
= false ;
% v i s u a l i z a t i o n of grid
isVisSol
= true ;
% v i s u a l i z a t i o n of s o l u t i o n
i s S l o p e L i m = true ;
% slope l i m i t i n g
t y p e S l o p e L i m = ' h i e r a r c h _ v e r t'; % Type of slope l i m i t e r ( linear , h i e r a r c h_ver t , s t r i c t)
o u t p u t F r e q u e n c y = 100;
% no v i s u a l i z a t i o n of every t i m e s t e p
o u t p u t B a s e n a m e = [ ' s o l u t i o n _' t y p e S l o p e L i m]; % B a s e n a m e of o u t p u t files
outputTypes
= cellstr ([ ' vtk '; ' tec ' ]) ;
% % P a r a m e t e r check .
diary ([ o u t p u t B a s e n a m e '. log ' ])
assert ( p >= 0 && p <= 4
, ' P o l y n o m i a l order must be zero to four . '
)
assert ( ordRK >= 1 && ordRK <= 3 , ' Order of Runge Kutta must be zero to three . ')
assert ( hmax > 0
, ' Maximum edge length must be positive . '
)
assert ( numSteps > 0
, ' Number of time steps must be positive . '
)
assert (~ i s S l o p e L i m || p > 0
, ' Slope limiting only a v a i l a b l e for p > 0. '
)
% % T r i a n g u l a t i o n.
g = d o m a i n S q u a r e( hmax ) ; % A l t e r n a t i v e: g = d o m a i n P o l y g o n([0 1 1 0] , [0 0 1 1] , hmax ) ;
if isVisGrid , v i s u a l i z e G r i d( g ) ; end
% % G l o b a l l y c o n s t a n t p a r a m e t e r s.
K
= g . numT ;
% n u m b e r of t r i a n g l e s
N
= nchoosek ( p + 2 , p ) ;
% n u m b e r of local DOFs
tau
= tEnd / numSteps ;
% time step size
m a r k E 0 T i n t = g . idE0T == 0;
% [ K x 3] mark local edges that are i n t e r i o r
m a r k E 0 T b d r D = ~ m a r k E 0 T i n t;
% [ K x 3] mark local edges on the D i r i c h l e t b o u n d a r y
m a r k V 0 T b d r D = ismember ( g . V0T , g . V0E ( g . E0T ( m a r k E 0 T b d r D) ,:) ) ; % [ K x 3] mark v e r t i c e s on D i r i c h l e t bdr
% % C o e f f i c i e n t s and b o u n d a r y data ( LeVeque ' s solid body r o t a t i o n) .
G = @ ( x1 , x2 , x1_0 , x2_0 ) (1/0.15) * sqrt (( x1 - x1_0 ) .^2 + ( x2 - x2_0 ) .^2) ;
c0Cont = @ ( x1 , x2 ) (( x1 -0.5) .^2+( x2 -0.75) .^2 <= 0.0225 & ( x1 <=0.475| x1 >=0.525| x2 >=0.85) ) + ...
(1 - G ( x1 , x2 , 0.5 , 0.25) ) .* (( x1 - 0.5) .^2 + ( x2 - 0.25) .^2 <= 0.0225) + ...
0.25*(1+ cos ( pi * G ( x1 , x2 , 0.25 , 0.5) ) ) .*(( x1 - 0.25) .^2 + ( x2 - 0.5) .^2 <= 0.0225) ;
fCont = @ (t , x1 , x2 ) zeros ( size ( x1 ) ) ;
u1Cont = @ (t , x1 , x2 ) 0.5 - x2 ;
u2Cont = @ (t , x1 , x2 ) x1 - 0.5;
cDCont = @ (t , x1 , x2 ) zeros ( size ( x1 ) ) ;
% % L o o k u p table for basis f u n c t i o n.
c o m p u t e B a s e s O n Q u a d( N ) ;
if isSlopeLim , c o m p u t e T a y l o r B a s e s V 0 T(g , N ) ; end
% % C o m p u t a t i o n of m a t r i c e s on the r e f e r e n c e t r i a n g l e.
hatM
= i n t e g r a t e R e f E l e m P h i P h i( N ) ;
hatG
= i n t e g r a t e R e f E l e m D p h i P h i P h i( N ) ;
hatRdiagOnQuad
= i n t e g r a t e R e f E d g e P h i I n t P h i I n t P e r Q u a d( N ) ;
h a t R o f f d i a g O n Q u a d = i n t e g r a t e R e f E d g e P h i I n t P h i E x t P e r Q u a d( N ) ;
% % A s s e m b l y of time - i n d e p e n d e n t g l o b a l m a t r i c e s.
globM = a s s e m b l e M a t E l e m P h i P h i(g , hatM ) ;
if i s S l o p e L i m
globMTaylor
= a s s e m b l e M a t E l e m P h i T a y l o r P h i T a y l o r(g , N ) ;
g l o b M D i s c T a y l o r = a s s e m b l e M a t E l e m P h i D i s c P h i T a y l o r(g , N ) ;
globMCorr
= spdiags (1./ diag ( g l o b M T a y l o r) , 0 , K *N , K * N ) * g l o b M T a y l o r;
end % if
% % I n i t i a l data .
cDisc = p r o j e c t F u n c C o n t 2 D a t a D i s c(g , c0Cont , 2* p +1 , hatM ) ;
if i s S l o p e L i m
cDV0T = c o m p u t e F u n c C o n t V 0 T(g , @ ( x1 , x2 ) cDCont (0 , x1 , x2 ) ) ;
cDisc = a p p l y S l o p e L i m i t e r D i s c(g , cDisc , markV0TbdrD , cDV0T , globM , globMDiscTaylor , t y p e S l o p e L i m) ;
end % if
fprintf ( ' L2 error w . r . t . the initial c o n d i t i o n: % g \ n ' , c o m p u t e L 2 E r r o r(g , cDisc , c0Cont , 2* p ) ) ;
% % v i s u a l i z a t i o n of i n i t a l c o n d i t i o n.
if isVisSol
c L a g r a n g e = p r o j e c t D a t a D i s c 2 D a t a L a g r( cDisc ) ;
v i s u a l i z e D a t a L a g r(g , cLagrange , ' u_h ' , outputBasename , 0 , o u t p u t T y p e s)
end
% % Time s t e p p i n g.
fprintf ( ' Starting time i n t e g r a t i o n from 0 to % g using time step size % g (% d steps ) .\ n ' , tEnd , tau ,
֒→numSteps )
for nStep = 1 : numSteps
27
[t , omega ] = r u n g e K u t t a S S P( ordRK , tau , ( nStep - 1) * tau ) ;
cDiscRK = cell ( length ( omega ) +1 , 1) ; cDiscRK {1} = reshape ( cDisc ' , [ K * N 1]) ;
% % P e r f o r m Runge - Kutta steps
for rkStep = 1 : length ( omega )
% L2 p r o j e c t i o n s of C o n t e b r a i c c o e f f i c i e n t s
fDisc = p r o j e c t F u n c C o n t 2 D a t a D i s c(g , @ ( x1 , x2 ) fCont ( t ( rkStep ) ,x1 , x2 ) , 2* p , hatM ) ;
u1Disc = p r o j e c t F u n c C o n t 2 D a t a D i s c(g , @ ( x1 , x2 ) u1Cont ( t ( rkStep ) ,x1 , x2 ) , 2* p , hatM ) ;
u2Disc = p r o j e c t F u n c C o n t 2 D a t a D i s c(g , @ ( x1 , x2 ) u2Cont ( t ( rkStep ) ,x1 , x2 ) , 2* p , hatM ) ;
% E v a l u a t e n o r m a l v e l o c i t y in q u a d r a t u r e p o i n t s of edges
v N o r m a l O n Q u a d E d g e = c o m p u t e F u n c C o n t N u O n Q u a d E d g e(g , @ ( x1 , x2 ) u1Cont ( t ( rkStep ) ,x1 , x2 ) , @ ( x1 , x2 ) u2Cont
֒→( t ( rkStep ) ,x1 , x2 ) , 2* p +1) ; % veloc \ dot \ nu on q u a d r a t u r p o i n t s on edges
% A s s e m b l y of time - d e p e n d e n t g l o b a l m a t r i c e s
globG = a s s e m b l e M a t E l e m D p h i P h i F u n c D i s c V e c(g , hatG , u1Disc , u2Disc ) ;
globR = a s s e m b l e M a t E d g e P h i P h i V a l U p w i n d(g , hatRdiagOnQuad , h a t R o f f d i a g OnQ ua d , v N o r m a l O n Q u a d E d g e) ;
% A s s e m b l y of D i r i c h l e t b o u n d a r y c o n t r i b u t i o n s
globKD = a s s e m b l e V e c E d g e P h i I n t F u n c C o n t V a l(g , markE0TbdrD , @ ( x1 , x2 ) cDCont ( t ( rkStep ) ,x1 , x2 ) ,
֒→v N o r m a l O n Q ua dEd ge , N ) ;
% A s s e m b l y of the s o u r c e c o n t r i b u t i o n
globL = globM * reshape ( fDisc ' , K *N , 1) ;
% B u i l d i n g the s y s t e m
sysA = - globG {1} - globG {2} + globR ;
sysV = globL - globKD ;
% C o m p u t i n g the d i s c r e t e time d e r i v a t i v e
cDiscDot = globM \ ( sysV - sysA * cDiscRK { rkStep }) ;
% Apply slope l i m i t i n g to time d e r i v a t i v e
if i s S l o p e L i m
c D i s c D o t T a y l o r = p r o j e c t D a t a D i s c 2 D a t a T a y l o r( reshape ( cDiscDot , [ N K ]) ', globM , g l o b M D i s c T a y l o r) ;
c D i s c D o t T a y l o r L i m = a p p l y S l o p e L i m i t e r T a y l o r(g , cDiscDotTaylor , markV0TbdrD , NaN (K ,3) , t y p e S l o p e L i m) ;
c D i s c D o t T a y l o r = reshape ( c D i s c D o t T a ylo rL im ' , [ K * N 1]) + g l o b M C o r r * reshape (( c D i s c D o t T a y l o r ֒→c D i s c D o t T a y l o r L i m) ', [ K * N 1]) ;
cDiscDot = reshape ( p r o j e c t D a t a T a y l o r 2 D a t a D i s c( reshape ( cDiscDotTaylor , [ N K ]) ', globM ,
֒→g l o b M D i s c T a y l o r) ', [ K * N 1]) ;
end
% C o m p u t e next step
cDiscRK { rkStep + 1} = omega ( rkStep ) * cDiscRK {1} + (1 - omega ( rkStep ) ) * ( cDiscRK { rkStep } + tau *
֒→cDiscDot ) ;
% L i m i t i n g the s o l u t i o n
if i s S l o p e L i m
cDV0T = c o m p u t e F u n c C o n t V 0 T(g , @ ( x1 , x2 ) cDCont ( t ( rkStep ) , x1 , x2 ) ) ;
cDiscRK { rkStep + 1} = reshape ( a p p l y S l o p e L i m i t e r D i s c(g , reshape ( cDiscRK { rkStep + 1} , [ N K ]) ',
֒→markV0TbdrD , cDV0T , globM , globMDiscTaylor , t y p e S l o p e L i m) ', [ K * N 1]) ;
end % if
end % for
cDisc = reshape ( cDiscRK { end } , N , K ) ';
%% visualization
if isVisSol && mod ( nStep , o u t p u t F r e q u e n c y) == 0
c L a g r a n g e = p r o j e c t D a t a D i s c 2 D a t a L a g r( cDisc ) ;
v i s u a l i z e D a t a L a g r(g , cLagrange , ' u_h ' , outputBasename , nStep , o u t p u t T y p e s) ;
end
end % for
if isVisSol
c L a g r a n g e = p r o j e c t D a t a D i s c 2 D a t a L a g r( cDisc ) ;
v i s u a l i z e D a t a L a g r(g , cLagrange , ' u_h ' , outputBasename , nStep , o u t p u t T y p e s) ;
end
fprintf ( ' L2 error w . r . t . the initial c o n d i t i o n: % g \ n ' , c o m p u t e L 2 E r r o r(g , cDisc , c0Cont , 2* p ) ) ;
fprintf ( ' Total c o m p u t a t i o n time : % g seconds .\ n ' , toc ) ;
diary off
end % f u n c t i o n
ind = mult2ind(a) computes the linear index I(a) corresponding to a two-dimensional multi-index a as defined
in (16).
function ind = mult2ind ( a )
p = sum (a , 2) ; N = p .* ( p + 1) / 2;
ind = N + 1 + a (: , 2) ;
end
mult = multiindex(p) computes all two-dimensional multi-indices involved in the representation of a polynomial
solution of degree p and returns them in a N × 2 array.
function mult = m u l t i i n d e x( p )
mult = zeros ( p * ( p +1) / 2 , 2) ;
mult (1 ,:) = [0 , 0];
for ord = 1 : p
28
offset = ord * ( ord +1) / 2;
for i = 1 : ord + 1
mult ( offset + i , :) = mult (1 , :) + [ ord - i + 1 , i - 1];
end
end
end
ret = phiTaylorPhy(g, i, X1, X2) evaluates the ith basis function φi on each triangle T ∈ Th (cf. Sec. 3.1) at points
specified by a list of n x1 coordinates X1 ∈ RK×n and x2 coordinates X2 ∈ RK×n .
function ret = p h i T a y l o r P h y(g , i , X1 , X2 )
qOrd = ceil (( sqrt (8* i +1) -3) /2) ;
[ Q1 , Q2 , W ] = q u a d R u l e 2 D( qOrd ) ;
Q2X1 = @ ( X1 , X2 ) g . B (: , 1 , 1) * X1 + g . B (: , 1 , 2) * X2 + g . coordV0T (: , 1 , 1) * ones ( size ( X1 ) ) ;
Q2X2 = @ ( X1 , X2 ) g . B (: , 2 , 1) * X1 + g . B (: , 2 , 2) * X2 + g . coordV0T (: , 1 , 2) * ones ( size ( X1 ) ) ;
R = length ( W ) ; K = g . numT ; numP = size ( X1 , 2) ;
dX1 = repmat (2 ./ ( max ( g . coordV0T (: ,: ,1) ,[] ,2) - min ( g . coordV0T (: ,: ,1) ,[] ,2) ) , [1 numP ]) ;
dX2 = repmat (2 ./ ( max ( g . coordV0T (: ,: ,2) ,[] ,2) - min ( g . coordV0T (: ,: ,2) ,[] ,2) ) , [1 numP ]) ;
switch i
case 1 % (0 ,0)
ret = ones (K , numP ) ;
case 2 % (1 ,0)
ret = ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) .* dX1 ;
case 3 % (0 ,1)
ret = ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) .* dX2 ;
case 4 % (2 ,0)
ret = ( 0.5*( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^2 -...
repmat ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^2 * W ' , [1 numP ]) ) .*( dX1 .* dX1 ) ;
case 5 % (1 ,1)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) -...
repmat ( 2 * ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) ) * W ' , [1 numP ]) ) .*( dX1 .* dX2 ) ;
case 6 % (0 ,2)
ret = ( 0.5*( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^2 -...
repmat ( ( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^2 * W ' , [1 numP ]) ) .*( dX2 .* dX2 ) ;
case 7 % (3 ,0)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^3 / 6 -...
repmat ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^3 * W ' / 3 , [1 numP ]) ) .* dX1 .^3;
case 8 % (2 ,1)
ret = ( 0.5*( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^2 .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) -...
repmat ( ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^2 .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) ) * W ' , [1 numP ]) ) .*( dX1 .^2.* dX2 ) ;
case 9 % (1 ,2)
ret = ( 0.5*( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^2 -...
repmat ( ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^2 ) * W ' , [1 numP ]) ) .*( dX1 .* dX2 .^2) ;
case 10 % (0 ,3)
ret = ( ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^3 / 6 -...
repmat ( ( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^3 * W ' / 3 , [1 numP ]) ) .* dX2 .^3;
case 11 % (4 ,0)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^4 / 24 -...
repmat ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^4 * W ' / 12 , [1 numP ]) ) .* dX1 .^4;
case 12 % (3 ,1)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^3 .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) / 6 -...
repmat ( ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^3 .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) ) * W ' / 3 , [1 numP ]) ) .*( dX1 .^3.* dX2 ) ;
case 13 % (2 ,2)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .^2 .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^2 / 4 -...
repmat ( ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .^2 .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^2 ) * W '/2 , [1 numP ]) ) .*( dX1 .* dX2 ) .^2;
case 14 % (1 ,3)
ret = ( ( X1 - repmat ( g . baryT (: ,1) , [1 numP ]) ) .* ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^3 / 6 -...
repmat ( ( ( Q2X1 ( Q1 , Q2 ) - repmat ( g . baryT (: ,1) , [1 R ]) ) .* ...
( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^3) * W '/3 , [1 numP ]) ) .*( dX1 .* dX2 .^3) ;
case 15 % (0 ,4)
ret = ( ( X2 - repmat ( g . baryT (: ,2) , [1 numP ]) ) .^4 / 24 -...
repmat ( ( Q2X2 ( Q1 , Q2 ) - repmat ( g . baryT (: ,2) , [1 R ]) ) .^4 * W ' / 12 , [1 numP ]) ) .* dX2 .^4;
end % s w i t c h
end
ret = phiTaylorRef(g, i, hatX1, hatX2) evaluates the ith basis function φi on each triangle T ∈ Th (cf. Sec. 3.1)
at points specified by a list of x̂1 coordinates hatX1 and x̂2 coordinates hatX2.
function ret = p h i T a y l o r R e f(g , i , hatX1 , hatX2 )
29
Q2X1 = @ ( X1 , X2 ) g . B (: , 1 , 1) * X1 + g . B (: , 1 , 2) * X2 + g . coordV0T (: , 1 , 1) * ones ( size ( X1 ) ) ;
Q2X2 = @ ( X1 , X2 ) g . B (: , 2 , 1) * X1 + g . B (: , 2 , 2) * X2 + g . coordV0T (: , 1 , 2) * ones ( size ( X1 ) ) ;
ret = p h i T a y l o r P h y(g , i , Q2X1 ( hatX1 , hatX2 ) , Q2X2 ( hatX1 , hatX2 ) ) ;
end
dataTaylor = projectDataDisc2DataTaylor(dataDisc, globMDisc, globMDiscTaylor) converts the representa-
tion matrix in the DG / modal basis to the respective representation matrix in a Taylor basis, both of size K × N.
It solves Eq. (20) for CTaylor .
function d a t a T a y l o r = p r o j e c t D a t a D i s c 2 D a t a T a y l o r( dataDisc , globMDisc , g l o b M D i s c T a y l o r)
[K , N ] = size ( dataDisc ) ;
d a t a T a y l o r = reshape ( g l o b M D i s c T a y l o r \ ( g l o b M D i s c * reshape ( dataDisc ' , [ K * N 1]) ) , [ N K ]) ';
end % f u n c t i o n
dataDisc = projectDataTaylor2DataDisc(dataTaylor, globMDisc, globMDiscTaylor) converts the representa-
tion matrix in the Taylor basis to the respective representation matrix in a DG / modal basis, both of size K × N.
It solves Eq. (20) for CDG .
function dataDisc = p r o j e c t D a t a T a y l o r 2 D a t a D i s c( dataTaylor , globMDisc , g l o b M D i s c T a y l o r)
[K , N ] = size ( d a t a T a y l o r) ;
dataDisc = reshape ( g l o b M D i s c \ ( g l o b M D i s c T a y l o r * reshape ( dataTaylor ' , [ K * N 1]) ) , [ N K ]) ';
end % f u n c t i o n
[t, omega] = rungeKuttaSSP(ord, tau, nStep) provides a list of time levels t(i) = tn + δi ∆tn and weights ωi for the
n-th time step nStep with time step size tau according to Sec. 2.4. The order of the Runge-Kutta method is given as
parameter ord.
function [t , omega ] = r u n g e K u t t a S S P( ord , tau , t0 )
switch ord
case 1
omega = 0;
t = t0 ;
case 2
omega = [0 , 0.5];
t = t0 + [0 , 1] * tau ;
case 3
omega = [0 , 3/4 , 1/3]; t = t0 + [0 , 1 , 0.5] * tau ;
end
end % f u n c t i o n
6. Conclusion and Outlook
The second installment in the present paper series on implementing a MATLAB / GNU Octave toolbox introduced
performance optimized techniques for dealing with linear advection operators, higher order Runge–Kutta time discretizations, and a range of slope limiters designed to support general order DG discretizations. Our future work
plans include nonlinear advection operators and coupled systems of PDEs as well as multi-physics applications with
corresponding coupling mechanisms.
Acknowledgments
The work of B. Reuter was supported by the German Research Foundation (DFG) under grant AI 117/1-1.
30
Index of notation
Symbol
Definition
·
R
Integral mean, v ≔ |T1 | T v(x) dx, where v : T → R.
"
#
A
≔
, block-diagonal matrix with blocks A, B.
diag(A, B)
#M
a·b
∇
◦
⊗
c
c0
cD
C
CTaylor
δ[condition]
em
Ekn , Ên
V, E, T
EΩ , E∂Ω
f
Fk
h
hT
J
K
ν
νT
νk
N = Np
ωr
Ω, ∂Ω
∂Ωin , ∂Ωout
p
ϕki , ϕ̂i
φki
P p (T )
P p (Th )
Φ
ΦTaylor
q̂r
R
R+ , R+0
t
tn
tend
ϑ̂n− n+
∆tn
T k , ∂T k
T̂
u
x
x̂
xkc
xki
B
Cardinality of a set M.
P
≔ 2m=1 am bm , Euclidean scalar product in R2 .
≔ [∂ x1 , ∂ x2 ]T , spatial gradient in the physical domain Ω.
Composition of functions or Hadamard product.
Kronecker product.
Concentration (scalar-valued unknown).
Concentration prescribed at initial time t = 0.
Concentration prescribed on the inflow boundary.
∈ RKN , representation vector of ch ∈ P p (Th ) with respect to {ϕk j }.
∈ RKN , representation vector of ch ∈ P p (Th ) with respect to {φk j }.
≔ {1 if condition is true, 0 otherwise}, Kronecker delta.
mth unit vector.
nth edge of the physical triangle T k , nth edge of the reference triangle T̂ .
Sets of vertices, edges, and triangles.
Set of interior edges, set of boundary edges.
Source / sink (scalar-valued coefficient function).
Affine mapping from T̂ to T k .
Mesh fineness of Th .
≔ diam(T ), diameter of triangle T ∈ Th .
≔ (0, tend ), open time interval.
≔ #Th , number of triangles.
Unit normal on ∂Ω pointing outward of Ω.
Unit normal on ∂T pointing outward of T .
≔ νT k .
≔ (p + 1)(p + 2)/2, number of local degrees of freedom of P p (T ).
Quadrature weight associated with q̂r .
spatial domain in two dimensions, boundary of Ω.
inflow√ and outflow boundaries, ∂Ω = ∂Ωin ∪ ∂Ωout .
= ( 8N + 1 − 3)/2, polynomial degree.
ith hierarchical basis function on T k , ith hierarchical basis function on T̂ .
ith Taylor basis function on T k .
Space of polynomials on T ∈ Th of degree at most p.
≔ {wh : Ω → R ; ∀T ∈ Th , wh |T ∈ P p (T )}.
Slope limiting operator with respect to {ϕk j }.
Slope limiting operator with respect to {φk j }.
rth quadrature point in T̂ .
Number of quadrature points.
Set of (strictly) positive real numbers, set of non-negative real numbers.
Time variable.
nth time level.
End time.
Mapping from Ên− to Ên+ .
≔ tn+1 − tn , time step size.
kth physical triangle, boundary of T k .
Bi-unit reference triangle.
Velocity (vector-valued coefficient function).
T
= [x1 , x2 ] , space variable in the physical domain Ω.
T
1
2
= [ x̂ , x̂ ] , space variable in the reference triangle T̂ .
Centroid of the element T k ∈ Th .
ith vertex of the physical triangle T k .
References
[1] F. Frank, B. Reuter, V. Aizinger, FESTUNG—The Finite Element Simulation Toolbox for UNstructured Grids (2016).
doi:10.5281/zenodo.46069.
URL http://www.math.fau.de/FESTUNG
31
[2] F. Frank, B. Reuter, FESTUNG: The Finite Element Simulation Toolbox for UNstructured Grids (2016).
URL https://github.com/FESTUNG
[3] F. Frank, B. Reuter, V. Aizinger, P. Knabner, FESTUNG: A MATLAB / GNU Octave toolbox for the discontinuous Galerkin method, Part I:
Diffusion operator, Computers & Mathematics with Applications 70 (1) (2015) 11–46. doi:10.1016/j.camwa.2015.04.013.
[4] B. Cockburn, C. Shu, The local discontinuous Galerkin method for time-dependent convection–diffusion systems, SIAM Journal on Numerical Analysis 35 (6) (1998) 2440–2463. doi:10.1137/S0036142997316712.
[5] H. Reed, T. R. Hill, Triangular mesh methods for the neutron transport equation, Tech. Rep. LA-UR-73-479, Los Alamos Scientific Laboratory, NM (1973).
[6] C. Johnson, J. Pitkranta, An Analysis of the Discontinuous Galerkin Method for a Sacalar Hyperbolic Equation, Mathematics of Computation
46 (173) (1986) 1–26.
[7] D. Kuzmin, Slope limiting for discontinuous Galerkin approximations with a possibly non-orthogonal Taylor basis, International Journal for
Numerical Methods in Fluids 71 (9) (2013) 1178–1190. doi:10.1002/fld.3707.
[8] S. Gottlieb, C.-W. Shu, Strong stability-preserving high-order time discretization methods, Math. Comp. 67 (221) (1998) 73–85.
doi:10.1090/S0025-5718-98-00913-2.
[9] S. Gottlieb, C.-W. Shu, E. Tadmor, Strong stability-preserving high-order time discretization methods, SIAM Review 43 (1) (2001) 89–112.
[10] B. Cockburn, C.-W. Shu, TVB Runge-Kutta local projection discontinuous Galerkin finite element method for conservation laws. II. General
framework, Mathematics of computation 52 (186) (1989) 411–435.
[11] L. Krivodonova, J. Xin, J.-F. Remacle, N. Chevaugeon, J. E. Flaherty, Shock detection and limiting with discontinuous galerkin methods for
hyperbolic conservation laws, Applied Numerical Mathematics 48 (3) (2004) 323–338.
[12] S. Tu, S. Aliabadi, A slope limiting procedure in discontinuous galerkin finite element method for gasdynamics applications, International
Journal of Numerical Analysis and Modeling 2 (2) (2005) 163–178.
[13] C. Michoski, C. Mirabito, C. Dawson, D. Wirasaet, E. Kubatko, J. Westerink, Adaptive hierarchic transformations for dynamically p-enriched
slope-limiting over discontinuous galerkin systems of generalized equations, Journal of Computational Physics 230 (22) (2011) 8028 – 8056.
doi:10.1016/j.jcp.2011.07.009.
[14] M. Yang, Z.-J. Wang, A parameter-free generalized moment limiter for high-order methods on unstructured grids, Adv. Appl. Math. Mech
1 (4) (2009) 451–480.
[15] X. Zhang, Y. Xia, C.-W. Shu, Maximum-principle-satisfying and positivity-preserving high order discontinuous Galerkin schemes for conservation laws on triangular meshes, J. Sci. Comput. 50 (1) (2012) 29–62. doi:10.1007/s10915-011-9472-8.
[16] D. Kuzmin, A vertex-based hierarchical slope limiter for adaptive discontinuous Galerkin methods, Journal of Computational
and Applied Mathematics 233 (12) (2010) 3077–3085, Finite Element Methods in Engineering and Science (FEMTEC 2009).
doi:10.1016/j.cam.2009.05.028.
[17] M. Yang, Z.-J. Wang, A parameter-free generalized moment limiter for high-order methods on unstructured grids, in: 47th AIAA Aerospace
Sciences Meeting Including The New Horizons Forum and Aerospace Exposition, AIAA-2009-605. doi:10.2514/6.2009-605.
[18] K. Michalak, C. Ollivier-Gooch, Limiters for unstructured higher-order accurate solutions of the euler equations, in: 46th AIAA Aerospace
Sciences Meeting and Exhibit, AIAA-2008-776. doi:10.2514/6.2008-776.
[19] H. Luo, J. D. Baum, R. Lhner, A discontinuous Galerkin method based on a Taylor basis for the compressible flows on arbitrary grids, Journal
of Computational Physics 227 (20) (2008) 8875 – 8893. doi:10.1016/j.jcp.2008.06.035.
[20] V. Aizinger, A geometry independent slope limiter for the discontinuous Galerkin method, in: E. Krause, Y. Shokin, M. Resch, D. Krner,
N. Shokina (Eds.), Computational Science and High Performance Computing IV, Vol. 115 of Notes on Numerical Fluid Mechanics and
Multidisciplinary Design, Springer Berlin Heidelberg, 2011, pp. 207–217. doi:10.1007/978-3-642-17770-5_16.
[21] T. Barth, D. Jespersen, The design and application of upwind schemes on unstructuredmeshes, in: Proc. AIAA 27th Aerospace Sciences
Meeting, Reno, 1989.
[22] R. Cools, An encyclopaedia of cubature formulas, Journal of Complexity 19 (3) (2003) 445–453. doi:10.1016/S0885-064X(03)00011-6.
[23] B. Cockburn, C.-W. Shu, The Runge–Kutta discontinuous Galerkin method for conservation laws V: multidimensional systems, J. Comput.
Phys. 141 (2) (1998) 199–224. doi:10.1006/jcph.1998.5892.
[24] R. J. Leveque, High-resolution conservative algorithms for advection in incompressible flow, SIAM Journal on Numerical Analysis 33 (2)
(1996) 627–665. doi:10.2307/2158391.
[25] L. Sorber, Kronecker product, MATLAB Central File Exchange. Retrieved October 30, 2015 (2010).
URL http://de.mathworks.com/matlabcentral/fileexchange/28889-kronecker-product
32
| 5cs.CE
|
Additive Approximation for Near-Perfect
Phylogeny Construction ⋆
arXiv:1206.3334v1 [cs.DS] 14 Jun 2012
Pranjal Awasthi, Avrim Blum, Jamie Morgenstern, and Or Sheffet
Carnegie Mellon University, Pittsburgh,
5000 Forbes Ave., Pittsburgh PA 15213, USA,
{pawasthi,avrim,jamiemmt,osheffet}@cs.cmu.edu
Abstract. We study the problem of constructing phylogenetic trees for
a given set of species. The problem is formulated as that of finding a
minimum Steiner tree on n points over the Boolean hypercube of dimension d. It is known that an optimal tree can be found in linear time [1]
if the given dataset has a perfect phylogeny, i.e. cost of the optimal phylogeny is exactly d. Moreover, if the data has a near-perfect phylogeny,
i.e. the cost of the optimal Steiner tree is d + q, it is known [2] that an
exact solution can be found in running time which is polynomial in the
number of species and d, yet exponential in q. In this work, we give a
polynomial-time algorithm (in both d and q) that finds a phylogenetic
tree of cost d+O(q 2 ). This provides the best guarantees known—namely,
√
a (1 + o(1))-approximation—for the case log(d) ≪ q ≪ d, broadening
the range of settings for which near-optimal solutions can be efficiently
found. We also discuss the motivation and reasoning for studying such
additive approximations.
1
Introduction
Phylogenetics, a subfield of computational biology, aims to construct simple
and accurate descriptions of evolutionary history. These descriptions are represented as evolutionary trees for a given set of species, each of which is represented by some set of features ([3,4]). A typical choice for these features are
single nucleotide polymorphisms (SNPs), binary indicator variables for common
mutations found in DNA[5,6]; see, for example, [2,1,7,8,9,10]. This challenging
problem has attracted much attention in recent years, with progress in studying
various computational formulations of this problem ([3,11,2,1,12,7]). The problem is often posed as that of constructing the most parsimonious tree induced
by the set of species.
Formally, a phylogeny or a phylogenetic tree for a set C of n species, each
represented by a string (called taxa) of length d over a finite alphabet Σ, is an
unrooted tree T = (V, E) such that C ⊆ V ⊆ Σ d . Given a distance metric µ over
⋆
This work was supported in part by the National Science Foundation under grant
CCF-1116892, by an NSF Graduate Fellowship, and by the MSR-CMU Center for
Computational Thinking.
P
Σ d , we define the cost of T as (u,v)∈E µ(u, v). The tree of maximum parsimony
for a dataset is the tree which minimizes this cost with respect to the Hamming
metric; i.e., it is the optimum Steiner tree for the set C under this metric.
The Steiner tree problem is known to be NP-hard in general [13], and remains NP-hard even in the case of a binary alphabet with the metric induced
by the Hamming distance [14]. Extensive recent work, both experimental and
theoretical, has focused on the binary character set with the Hamming metric
([3,2,1,12,7,4,15,16]). This version of the phylogeny problem will also be the focus
of this paper.
A phylogeny is called perfect if each coordinate i ∈ [d] flips exactly once in the
tree (representing a single mutation of i amongst the set of species)1 . If a dataset
admits a perfect phylogeny, an optimal tree can be constructed in polynomial
time [17] (even linear time, in the case where the alphabet is binary [3]). In
this work, we investigate near perfect phylogenies – instances whose optimal
phylogenetic tree has cost d + q, where q ≪ d. Near perfect phylogenies have
been studied in theoretical ([11,2,12,16]) and experimental settings ([15]). The
work of [11,2,12,16] has given a series of randomized algorithms which find the
optimal phylogeny in running time polynomial in n and d but exponential in q.
Clearly, when q = ω(log d), these algorithms are not tractable.
An alternative approach for finding a phylogenetic tree of low cost is to use a
generic Steiner tree approximation algorithm. The best current such algorithm
yields a tree of cost at most 1.39(d+q) [18] (we comment that the exponential size
of the explicit hypercube with respect to its small representation size requires
one implement such an algorithm using techniques devised especially for the
hypercube, e.g. Alon et al. [7].) However, notice that for moderate q (e.g., for
q = polylog(d)), the excess of this tree—meaning the difference between its cost
and d—may be extremely large compared to the excess q of the optimal tree. In
such cases, one would much prefer an algorithm whose excess could be written
as a function of q only.
In this work, we present a randomized poly(n, d, q)-time algorithm that finds
a phylogenetic tree of cost d + O(q 2 ).
Theorem 1. Given a set C ⊆ {0, 1}d of n terminals, such that the optimal phylogeny of C has cost d + q, there exists a randomized poly(n, d, q)-time algorithm
that finds a phylogenetic tree of cost d + O(q 2 ) w.p. ≥ 1/2.
Note that Theorem 1 provides
√ a substantial improvement over prior work
for the case that log d ≪ q ≪ d. In this range, the exact algorithms are no
longer tractable, and the multiplicative approximations yield significantly worse
bounds. Alternatively, viewed as a multiplicative guarantee, in this range our tree
is within a 1 + o(1) factor of optimal. To the best of our knowledge, this is the
first work to give an additive poly-time approximation to either the phylogeny
problem or any (non-trivial) setting the Steiner tree problem. One immediate
1
Without loss of generality, we may assume each coordinate flips at least once, since
all coordinates on which all species agree may be discarded up front.
question, which remains open, is whether our results can be improved to d+o(q 2 )
or perhaps even to d + O(q).
The rest of the paper is organized as follows. After surveying related work in
Section 1.1, we detail notation and preliminaries in Section 2. The presentation
of our algorithm is partitioned into two parts. In Section 3, we present the
algorithm for the case where no pair of coordinates is identical over all terminals
(formal definition there). In Section 4, we alter the algorithm for the simple case,
in a nontrivial way, so that the modified algorithm finds a low-cost phylogeny
for any dataset. We conclude in Section 5 with a discussion, motivating the
problem of near-perfect phylogeny tree from a different perspective, and present
open problems for future research.
1.1
Related Work
As mentioned in the introduction, the problem of constructing an optimal phylogeny is NP-complete even when restricted to binary alphabets [14]. Schwartz
et al. [11] give an algorithm based on an Integer Linear Programming (ILP)
formulation to solve the multi-state problem optimally, and show experimentally the algorithm is efficient on small instances. Perfect phylogenies (datasets
which admit a tree in which any coordinate changes exactly once) have optimal
parsimony trees which can be constructed in linear time in the binary case [1]
and in polynomial time for a fixed alphabet [12]. Unfortunately, finding the perfect phylogeny for arbitrary alphabets is NP-hard [19]. Recent work [2] gives an
algorithm to construct optimal phylogenetic trees for binary, near-perfect phylogenies (where only a small number of coordinates mutate more than once in
the optimal tree). However, the running time of the algorithm presented in their
work [2] is exponential in the number of additional mutations.
There has also been a lot of work on computing multiplicative approximations
to the Steiner tree problem. A Minimum Spanning Tree (MST) over the set of
terminals achieves an approximation ratio of 2 and a long line of work has led to
the current best bound of 1.39 [20,21,22,23,24,8,25,9,10,18]. The more recent of
these papers use a result due to Borchers and Du [26] showing that an optimal
Steiner tree can be approximated to arbitrary precision using k-restricted Steiner
trees.
Some of these approximations to the Steiner tree problem are not immediately extendable to the problem of constructing phylogenetic trees. This is
because the size of the vertex set for the phylogeny problem is exponential in
d (there are 2d vertices in the hypercube). If an algorithm works on an explicit
representation of the graph G defined by the hypercube, then it does not solve
the phylogeny problem in polynomial time. However, the line of work started by
Robins and Zelikovsky [9,10] used the notion of k-restricted Steiner trees, which
can be efficiently implemented on the hypercube. In particular, Alon et al. [7]
showed that in finding the optimal k-restricted component for a given set of k
terminals, it is sufficient to only consider topologies with the given k terminals
at the leaves. Using this, they were able to extend that work to achieve a 1.55
approximation ratio for the maximum parsimony problem, and a 16/9 approximation for maximum likelihood. Byrka et al. [18] considered a new LP relaxation
to the k-restricted Steiner tree problem and achieved an approximation ratio of
1.39, which can be combined with the topological argument from Alon et al. [7]
to achieve the same ratio for phylogenies.
2
Notation and Preliminaries
Our dataset C ⊆ {0, 1}d consists of n terminals over d binary coordinates. A
Steiner tree (or phylogeny) over C consists of a tree T on the hypercube that
spans C (plus possibly additional Steiner nodes), where we label each edge e in
T with the index i ∈ {1, . . . , d} of the coordinate flipped on edge e. The cost
of such a Steiner tree is the number of edges in the tree. Given a collection of
datasets P = {P1 , P2 , . . . , Pk } ⊆ C we define the Steiner forest problem as the
problem of finding a minimal Steiner tree on every P ∈ P separately. We refer to
such collection as a partition from now on, even though it may contain a subset
of the original terminal set C.
In this work, we consider
instances C whose minimum Steiner tree has cost
√
d+q, and think of q = o( d) (otherwise, any off-the-shelf constant approximation
algorithm for the Steiner problem gives a solution of cost ≤ d + O(q 2 )). We fix T
to be some optimal Steiner tree. By optimality, all leaves in T must be terminals,
whereas the internal nodes of T may be either terminals or non-terminals (nonterminals are called Steiner nodes). We define a coordinate i to be good if exactly
one edge in T is labeled i, and bad if two or more edges in T are labeled with i.
We may assume all d coordinates appear in the tree, otherwise, some coordinates
in C are fixed and so the dimensionality of the problem is less than d. Therefore,
at most q coordinates are bad (each bad coordinate flips at least twice and thus
adds a cost of at least 2 to the tree).
Given a coordinate i of a set of terminals P , we define an i-cut as the partition
P0 = {x ∈ P : xi = 0} and P1 = {x ∈ P : xi = 1}. We call two coordinates
i 6= j interchangeable if they define the same cut. We now present the following
basic facts which are easy to verify (see [2] for proofs).
Fact 1
1. Let S be a set of interchangeable coordinates. Then all coordinates in S
appear together in the optimal tree T , adjacent to one another. That is, in T
there are paths s.t. for each path: all of its edges are labeled by some i ∈ S,
all coordinates in S have an edge on the path, and all internal nodes on the
paths aren’t terminals and have degree 2. On these paths, any reordering the
S-labeled edges yields an equivalent optimal tree.
2. For any two good coordinates, i 6= j, one side of the i-cut is contained within
one side of the j-cut. Equivalently, there exist values bj such that all terminals on one side of the i-cut have their jth coordinate set to bj .
3. Fix any good coordinate i and let j be a good coordinate such that all terminals on one side of the i-cut have their j coordinate set to bj . Then both
endpoints of the edge labeled i have their jth coordinate set to bj .
4. A good coordinate i and a bad coordinate i′ cannot define the same cut.
It immediately follows from Fact 1 that for a given good coordinate i one can
efficiently reconstruct the endpoints of the edge on which i mutates, except for at
most q coordinates. This leads us to the following definition. Given i, we denote
Di as the set of all coordinates that are fixed to a constant value vi on at least one
side of the i-cut (different coordinates may be fixed on different sides), and we
denote bi as the vector of the corresponding values, i.e. vi ’s, of the coordinates in
Di . The pair (Di , bi ) is called the pattern of coordinate i. That set of terminals
that match the pattern of i is the set Pbi = {x ∈ P : ∀j ∈ Di , xj = bij }.
3
A Simple Case: Each Coordinate Determines a Distinct
Cut
To show the main ideas behind our algorithm, we first discuss a special case
in which no two coordinates i and j define the same cut on the terminal set
C. Algorithms for constructing phylogenetic trees often make this assumption
as they preprocess C by contracting any pair of interchangeable coordinates.
However, in our case such contractions are problematic, as we discuss in the
next section. So in Section 4, when we deal with the general case, we deal with
interchangeable coordinates in a non-trivial fashion.
3.1
Basic Building Blocks
We now turn to the description of our algorithm. On a high level it is motivated
by the notion of maintaining a proper partition of the terminals.
Definition 1. Call a partition P proper if the forest produced by restricting the
optimal tree T to the components P ∈ P is composed of edge disjoint trees.
Equivalently, the path in T between two nodes x and y in the same component
P of P does not pass through any node x′ in any different component P ′ of P.
Clearly, our initial partition, P = {C}, is proper. Our goal is to maintain a
proper partition of the current terminals while decreasing the dimensionality of
the problem in each step. This is implemented by the two subroutines we now
detail.
Pluck a Leaf and Paste a Leaf. The first subroutine works by building the
optimal phylogeny bottom-up, finding a good coordinate i adjacent to a leaf
terminal t in the tree, and replacing t with its parent (t with i flipped) in the set
of terminals. Observe that if i is a good coordinate, then this removes the only
occurrence of i, leaving all terminals in our new dataset with a fixed i coordinate,
thus reducing the dimensionality of the problem by 1.
The matching subroutine to Pluck-a-leaf is Paste-a-leaf: if Pluck-a-leaf
succeeds and returns some (x, P ′ ), and we have found a Steiner forest for the terminals in P ′ . Then Paste-a-leaf merely connects x with x̄i by an edge labeled
i, then returns the resulting forest. (We omit formal description.)
Pluck-a-leaf
input: A partition P of current terminals.
if there exists P ∈ P and x ∈ P s.t. some coordinate i is non-constant on P , but only
the terminal x has xi = 0 (or xi = 1), then:
– Set P ′ = P \ {x} ∪ {x̄i }, where x̄i is identical to x except for flipping i.
– Return x and P ′ = P \ {P } ∪ {P ′ }.
else fail.
Lemma 1. If P is a proper partition and Pluck-a-leaf succeeds, then P ′ is a
proper partition.
Proof (Sketch). Let T [P ] be the subtree in which x resides. We claim that x is
a leaf in T [P ], attached by an edge labeled i to the rest of the terminals. If this
indeed is the case, then removing i means removing a leaf-adjacent edge from
T [P ] which clearly leaves all components in the forest edge-disjoint.
Wlog x lies on the i = 0 side of the cut. If x isn’t a leaf, then at least two
disjoint paths connect x to two other terminals. Since P is proper, both these
terminals are in P . This means T [P ] crosses the i-cut twice, but then we can
replace T [P ] with an even less costly tree in which i is flipped once, by projecting
the path between the two occurrences of i onto the i = 1 side.
⊓
⊔
Observe that lemma 1 holds only when the underlying alphabet of the problem is binary. In particular, for a non-binary alphabet, such x can be a non-leaf.
Split and Merge. When Pluck-a-leaf can no longer find leaves to pluck,
we switch to the second subroutine, one that works by splitting the set of terminals into two disjoint sets, based on the value of the i-th coordinate. We would
like to split our set of terminals according to the i-cut, and recurse on each side
separately. But, in order to properly reconnect the two subproblems, we need
to introduce the two endpoints of the i-labeled edge to their respective sides of
the i-cut. Our Split subroutine deals with one particular case in which these
endpoints are easily identified.
Split(i)
input: A partition P of current terminals, a coordinate i that is not constant on every
component of P.
– Find a component P on which i isn’t constant. Denote the i-cut of P as (P0 , P1 ).
– Find Pbi , the set of terminals that match the pattern of i.
– if exists some x which is the unique terminal that matches the pattern of i in one
side of the cut (that is, if for some x we have Pbi ∩ P0 = {x} or Pbi ∩ P1 = {x})
• Flip the i-th coordinate of x, and let x̄i be the resulting node.
• Add x to its side of the i-cut, add x̄i to the other side of the cut.
• Return x, x̄i and P ′ = P \ {P } ∪ {P0 , P1 }.
else fail.
The matching subroutine to Split is Merge: Assume Split succeeds and
returns some (x, x̄i , P ′ ), and assume we have found a Steiner forest for the terminals in P ′ . Then Merge merely connects x with x̄i by an edge labeled i, then
returns the resulting forest. (Again, formal description is omitted.)
Lemma 2. Assume P is a proper partition. Assume Split is called on a good
coordinate i s.t. the edge labeled i in T has at least one endpoint which is a
terminal. Then the returned partition P ′ is proper.
Proof (Sketch). Since P is proper, then the induced tree T [P ] is the only tree in
the forest that contains the i-labeled edge. The lemma then follows from showing
that x and x̄i are the two endpoints of i-labeled edge in T [P ]. This follows from
the observation that the endpoints of the i-labeled edge must both match the
pattern of i. Let u be an endpoint and wlog u belongs to the (i = 0)-side of the
cut. On all coordinates that are fixed on the (i = 0)-side, u obviously has the
right values. All coordinates that are fixed on the (i = 1)-side can only flip on
the (i = 0)-side, but only after traversing u, so u has them set to the value fixed
on the (i = 1)-side.
⊓
⊔
3.2
The Algorithm
We can now introduce our algorithm.
input: A partition P of current terminals. Initially, P is the singleton set P = {C}.
1. if Pluck-a-leaf succeeds and returns (x, P ′ )
– recurse on P’, then Paste-a-leaf x back and return the resulting forest.
2. else-if the number of non-constant coordinates on P is at least 40q 2
– Pick a non-constant coordinate i u.a.r and invoke Split(i) .
– if Split succeeds: recurse on P ′ , then Merge x and x̄i , and return the resulting forest; otherwise fail.
3. else
– For every P ∈ P find its MST, T (P ), and return the forest {T (P )}.
Fig. 1. Algorithm for the simple case
Theorem 2. With probability ≥ 1/2, the algorithm in Figure 1 returns a tree
whose cost is at most d + O(q 2 ).
In order to prove Theorem 2, fix an optimal phylogeny T over our initial set
of terminals, and for any partition P our algorithm creates, denote T [P] as the
forest induced by T on this partition. The proof of the theorem relies on the
following lemma.
Lemma 3. If P is a proper partition, then with probability ≥ 1 − (8q)−1 , Split
is called on a good coordinate and succeeds. Furthermore, Split is executed at
most 4q times.
Proof (of Theorem 2). The proof follows from lemmas 1 and 3. Since we start
with a proper partition, then with probability at least 1 − (4q)(8q)−1 ≥ 1/2 we
keep recursing on proper partitions, until reaching the base of the recursion. By
the time the algorithm reaches the base of the recursion, the dimensionality of
the problem was reduced to d′ ≤ 40q 2 , so the cost of the optimal Steiner forest
is at most d′ + q. As MSTs give a 2-approximation to the optimal Steiner tree
problem, our forest is of cost ≤ 2(d′ + q). Then, the algorithm reconnects the
forest, adding the coordinates (edges) the algorithm as removed in the first two
steps of the algorithm. Since the algorithm removed at most d − d′ edges, the
tree it outputs is of overall cost at most d − d′ + 2(d′ + q) = d + 40q 2 + 2q.
⊓
⊔
Proof (of Lemma 3). Let P be the partition in the first iteration of the algorithm
for which Split was invoked, and assume P is proper. Thus, the forest T [P]
contains disjoint components. We call any vertex in this forest of degree ≥ 3
an internal split. Suppose we replace each internal split v with deg(v) many
new vertices, each adjacent to one edge. This breaks the forest into a collections
of paths we call the path decomposition of the tree. In addition, remove from
this path decomposition all edges that are labeled with a bad coordinate to
obtain the good path decomposition. Denote the number of paths in the good
path decomposition as t.
First, we claim that any call to Split (on P or any partition succeeding
P), on a coordinate i which lies on a path of length ≥ 2 in the abovementioned
decomposition, does not fail.
Assume Split was called on i and denote its adjacent coordinate on the
path as j (choose one arbitrarily if i has two adjacent coordinates on its path),
and both are non-constant on P ∈ P. Observe that our decomposition leaves
only good coordinates, so both i and j are good. Therefore, j is fixed on one
side of the i-cut and i is fixed on one side of the j-cut. It follows that there exist
binary values bi , bj s.t. for every x ∈ P , if xi = bi then xj = bj ; and if xj = 1 − bj
then xi = 1 − bi . In fact, the only node on the entire tree for which xi = 1 − bi
and xj = bj is the node connecting the i-edge and the j-edge. Recall that we
assume for the special case i and j do not define the same cut. It follows that
the node between i and j has to be a terminal, so now we can use Lemma 2 and
deduce Split succeeds.
So, Split can either fail or return a non-proper partition only if it was
invoked either on a bad coordinate or on a good coordinate that lies on a path
of length 1 in our path decomposition. There are at most q bad coordinates and
q+t
at most t paths of length 1, so each call to Split fails w.p. ≤ 40q
2 . Furthermore,
calling Split on a good edge i lying on a path of length at least 2 results in
both i’s endpoints as new leaves in their respective sides of the i-cut. As a result,
Pluck-a-leaf then completely unravels the path on which i lies. Therefore, in
a successful run of the algorithm, Split is called no more than t times. All that
remains is to bound t.
t is the number of paths on the path decomposition of P, a partition for which
Pluck-a-leaf failed to execute. Observe that if the forest T [P] had even a single
leaf connected to the rest of its tree by a good coordinate, then Pluck-a-leaf
would continue – such a leaf, by definition, is the only terminal on which the
good coordinate takes a certain value. It follows that l, the number of leaves in
T [P] is bounded by 2q, the number of bad edges in T . Removing the internal
splits then leaves us with at most 2l paths; removing the bad coordinates’ edges
adds at most 2q − l new paths (for every bad coordinate k adjacent to a leaf,
removing k does not create a new path). All in all, t ≤ 2l+2q −l ≤ 4q. Therefore,
1
each call to Split has success probability ≥ 1 − 4q+q
40q2 = 1 − 8q , and Split is
called at most 4q times.
⊓
⊔
4
The General Case: Interchangeable Coordinates May
Exist
Before describing the general case, let us briefly discuss why the conventional
way of initially contracting all interchangeable coordinates and applying the
algorithm from the Section 3 might result in a tree of cost d+ω(q 2 ). The analysis
of the first two steps of the algorithm still holds. The problem lies in the base
of the recursion, where the algorithm runs the MST-based 2-approximation.
Indeed, the MST algorithm is invoked on < 40q 2 contracted coordinates, but
they correspond to d˜ original coordinates, and it is possible that d˜ ≫ q 2 . So by
using any constant approximation on this entire forest, we may end with a tree
of cost d + 2d˜ which isn’t d + O(q 2 ).
Our revised algorithm does not contract edges initially. Instead, let us define a
simple coordinate as one for which Split(i) succeeds. So, the first alteration we
make to the algorithm is to call Split as long as the set of simple coordinates
is sufficiently big. However, most alterations lie in the base of the recursion.
Below we detail the algorithm and analyze its correctness. In the algorithm’s
description, for any coordinate i we denote the set of coordinates interchangeable
with i by Wi , and their number as w(i) = |W (i)|.
Theorem 3. With probability ≥ 1/2, the algorithm in Figure 2 returns a tree
of cost d + O(q 2 ).
The proof of Theorem 3 follows the same outline as the proof of Theorem 2.
Observe that Lemmas 1 and 3 still hold2 . Therefore, with probability ≥ 1/2, the
algorithm enters the base of the recursion with a proper partition. Thus, by the
following lemma, the algorithm outputs a tree of cost d + O(q 2 ).
Lemma 4. Assume that the base of the recursion (i.e., Step 3) is called on a
proper partition P of the terminals over d′ non-constant coordinates. Then the
algorithm returns a forest of cost d′ + O(q 2 ).
The full proof of Lemma 4 is deferred to the appendix. However, let us sketch
the main outline of the proof. Recall the good path decomposition we used in
the proof of Lemma 3. We partition its paths in the following way.
2
Clearly, Split cannot abort now, but it might be the case that the algorithm picks
i which is a bad coordinate. This can happen with probability ≤ q/8q 2 = 1/8q.
input: A partition P of current terminals. Initially, P is the singleton set P = {C}.
1. if Pluck-a-leaf succeeds and returns (x, P ′ )
– recurse on P’, then Paste-a-leaf x and return the resulting forest.
2. else-if the number of simple coordinates on P is at least 8q 2
– Pick a simple coordinate i u.a.r and invoke Split(i) .
– if Split succeeds: recurse on P ′ , then Merge x and x̄i , and return the resulting forest; otherwise fail.
3. else
– Contract all Wi into ī
– For every ī with w(i) > q and the (unique) component P in which the i-cut
resides,
• Apply pattern matching to (P, i). Let (Di , bi ) be the pattern of i.
• if i is simple, split P into P0 ∪ {xi } and P1 ∪ {xi }.
• else
∗ Define the node y(i) as the node where yi = 0, every coordinate j ∈ Di
is set to bij , and every coordinate j ∈
/ Di is set to 0.
∗ Define y(i) to be y(i) with coordinate i flipped.
∗ P = P \ {P } ∪ {P0 ∪ {y(i)}} ∪ {P1 ∪ {y(i)}}.
– For every P ∈ P find its MST T (P ), and retrieve the forest {T (P )}.
– For every i with w(i) > q:
if i was simple, add an edge labeled i between xi and xi
else add an edge labeled i between y(i) and y(i).
– Expand all contracted coordinates to their original set of coordinates by replacing i with a path of length w(i). Return the resulting forest.
Fig. 2. Algorithm for the general case
– Paths with at least one terminal on them. On such paths, because all interchangeable coordinates may appear in T in any order, then all coordinates
on such paths are simple. So, when we enter the base of the recursion, there
are at most 8q 2 edges on such paths.
– Paths with no terminal on them, with length > q. Such paths are composed
of interchangeable coordinates, and since there are more than q of those, we
deduce all of them are good. Therefore, the endpoints of such paths are fixed
up to at most q (bad) coordinates. We therefore contract these edges, split
on them, and introduce into each side of the cut an arbitrary endpoint, by
replacing non-fixed coordinates with zeros. So on each side of the cut the
cost of the subtree increases by at most q, and since there are at most 4q such
paths, our overall cost for introducing these artificial endpoints is O(q 2 ).
– Paths with no terminal on them, with length ≤ q. Such paths are composed
of interchangeable coordinates, but we do not contract them. Since there are
at most 4q ·q edges on such paths, we run the MST approximation, and incur
a cost of O(q 2 ) for edges on such paths.
Runtime analysis: Pluck-a-leaf can be implemented in time linear in the
size of the dataset, i.e. O(nd). Counting the number of simple coordinates takes
time O(nd2 ), and Split takes time O(nd). A naive implementation of the base
case of the recursion takes time O(nd2 ) for contracting coordinates, and the rest
can be implemented in time O(nd). Hence the time to process each node in the
recursion tree is at most O(nd2 ). Since there are at most O(q) nodes in the
recursion tree, the total runtime is O(qnd2 ).
5
Discussion and Open Problems
This paper presents a randomized approximation algorithm for constructing
near-perfect phylogenies. In order to achieve this, we obtain a Steiner tree of
low additive error. However, from the biological perspective, the goal is to find
a good evolutionary tree, one that will give correct answers to questions like
“what is the common ancestor of the following species?” or “which of the two
gene-mutations happened earlier?”. Such questions, we hope, can be answered
by finding the most-parsimonious phylogenetic tree over the given taxa. Hence,
it is also desirable that any low-cost tree which we output also captures a lot of
the structure of the optimal tree.
We would like to point out that our algorithm in fact has this valuable
property. Notice that until the base case of the recursion, both Pluck-a-leaf
and Split subroutines construct the optimal tree, and correctly identify the
endpoints of the edges they remove. Even when the algorithm reaches the base
case of the recursion – we can declare every edge of weight > q to be good,
and we know its endpoints up to at most q coordinates. In total, our algorithm
gives the structure of the optimal tree up to O(q 2 ) edges, and those edges can
be marked as “unsure”.
Several open problems remain for this work. The most straight-forward one
is whether one can devise an algorithm outputting a phylogenetic tree of cost
d + O(q)? Alternatively, one may try to design exact algorithms that are efficient even for q = ω(log d). We suspect that even the case of q = O((log d)2 )
poses quite a challenge. Finally, extending our results to non-binary alphabets
is intriguing. Note however that even the case of perfect phylogenies is NP-hard,
and tractable only for moderately sized alphabets. Furthermore, our bottom-up
approach completely breaks down for non-binary alphabets (see comment past
Lemma 1), so devising an additive-approximation algorithm for the phylogeny
problem with non-binary alphabets requires a different approach altogether.
References
1. Ding, Z., Filkov, V., Gusfield, D.: A linear-time algorithm for the perfect phylogeny
haplotyping (pph) problem. In: RECOMB, Springer (2005) 585–600
2. Blelloch, G.E., Dhamdhere, K., Halperin, E., Ravi, R., Sridhar, S.: Fixed parameter
tractability of binary near-perfect phylogenetic tree reconstruction. In: ICALP,
Springer (2006) 667–678
3. Gusfield, D.: Algorithms on strings, trees, and sequences: computer science and
computational biology. Cambridge University Press (1997)
4. Semple, C., Steel, M.: Phylogenetics. Oxford lecture series in mathematics and its
applications. Oxford University Press (2003)
5. Hinds, D.A., Stuve, L.L., Nilsen, G.B., Halperin, E., Eskin, E., Ballinger, D.G.,
Frazer, K.A., Cox, D.R.: Whole-genome patterns of common dna variation in
three human populations. Science 307(5712) (2005) 1072–1079
6. : The international hapmap project. Nature 426(6968) (2003) 789–96
7. Alon, N., Chor, B., Pardi, F., Rapoport, A.: Approximate maximum parsimony and
ancestral maximum likelihood. IEEE/ACM Trans. Comput. Biol. Bioinformatics
7 (Jan 2010) 183–187
8. Robins, G., Zelikovsky, A.: Improved steiner tree approximation in graphs. In:
SODA, Society for Industrial and Applied Mathematics (2000) 770–779
9. Robins, G., Zelikovsky, A.: Improved steiner tree approximation in graphs (2000)
10. Robins, G., Zelikovsky, A.: Tighter bounds for graph steiner tree approximation.
SIAM Journal on Discrete Mathematics 19 (2005) 122–134
11. Misra, N., Blelloch, G.E., Ravi, R., Schwartz, R.: Generalized buneman pruning
for inferring the most parsimonious multi-state phylogeny. In Berger, B., ed.:
RECOMB. Volume 6044 of Lecture Notes in Computer Science., Springer (Apr.
2010) 369–383
12. Fernández-Baca, D., Lagergren, J.: A polynomial-time algorithm for near-perfect
phylogeny. SIAM J. Comput. 32 (May 2003) 1115–1127
13. Karp, R.M.: Reducibility among combinatorial problems. In Miller, R.E.,
Thatcher, J.W., eds.: Complexity of Computer Computations. Plenum, New York
(1972) 85–103
14. Foulds, L.R., Graham, R.L.: The Steiner problem in phylogeny is NP-complete.
Adv. Appl. Math. 3 (1982)
15. Sridhar, S., Dhamdhere, K., Blelloch, G., Halperin, E., Ravi, R., Schwartz, R.:
Algorithms for efficient near-perfect phylogenetic tree reconstruction in theory and
practice. IEEE/ACM Trans. Comput. Biol. Bioinformatics 4 (Oct 2007) 561–571
16. Damaschke, P.: Parameterized enumeration, transversals, and imperfect phylogeny
reconstruction. Theor. Comput. Sci. 351 (Feb 2006) 337–350
17. Agarwala, R., Fernandez-Baca, D.: A polynomial-time algorithm for the perfect
phylogeny problem when the number of character states is fixed. In: SFCS. (Nov
1993) 140–147
18. Byrka, J., Grandoni, F., Rothvoß, T., Sanità, L.: An improved lp-based approximation for steiner tree. In: STOC, ACM (2010)
19. Bodlaender, H.L., Fellows, M.R., Warnow, T.: Two strikes against perfect phylogeny. In: ICALP’92. (1992) 273–283
20. Takahashi, H., Matsuyama, A.: An approximate solution for the steiner problem
in graphs. Mathematica Japonica 24 (1980) 573–577
21. Berman, P., Ramaiyer, V.: Improved approximations for the steiner tree problem.
In: SODA. (1992) 325–334
22. Pramel, H., Steger, A.: Rnc-approximation algorithms for the steiner problem. In
Reischuk, R., Morvan, M., eds.: STACS 97. Volume 1200 of Lecture Notes in Computer Science. Springer Berlin / Heidelberg (1997) 559–570 10.1007/BFb0023489.
23. Karpinski, M., Zelikovsky, A.: New approximation algorithms for the steiner tree
problems. Journal of Combinatorial Optimization 1 (1995) 47–65
24. Zelikovsky, A.: Better approximation bounds for the network and euclidean steiner
tree problems. Technical report (1996)
25. Hougardy, S., Promel, H.J.: A 1.598 approximation algorithm for the steiner problem in graphs. In: SODA. (1999) 448–453
26. Borchers, A., Du, D.Z.: The k-steiner ratio in graphs. In: STOC, ACM (1995)
641–649
A
Appendix
Here, we give short proofs of the basic properties presented in Fact 1. Many of
these results are also found in other work [2].
1. Let S be a set of coordinates that define the same cut over the terminals (up
to renaming of P0 and P1 ). Then whenever any edge in T is labeled with
some i ∈ S, then the edge lies on a path on which each edge is labeled with
a unique coordinate from S, and no node of the |S|-long path is a terminal.
Proof. Assume i, j ∈ S define the same cut over C. Given some edge ei where
i flips, consider the shortest path from a terminal t1 through ei to another
terminal t2 . If every node on this path is of degree 2, there is no terminal
before t2 , and j flips before t2 . Now, suppose some node on the path between
t1 and t2 has degree more than 2. Either j flips on each outgoing path before
any terminals, or j does not define the same cut (since each outgoing path
has a terminal on it). But if j flips on all outgoing paths before any terminal
occurs on those paths, relabel the Steiner nodes on each path so that j is
constant along those outgoing paths. Then, add one Steiner node and an
edge from the endpoint of ei which flips j. This tree has cost strictly less
than the tree which flipped j on each outgoing path, since there were at least
two paths, each of which flipped j.
⊓
⊔
2. For any two good coordinates, i 6= j, one side of the i-cut is contained
within one side of the j-cut. Alternatively, there exist values bj such that all
terminals on one side of the i-cut have their jth coordinate set to bj .
Proof. Suppose i is good. Then, consider the i-cut in T . Since j is good, j
may flip only once in the tree. If j flips in C0 , then j is constant in C1 . If j
flips in C1 , then j is constant in C0 .
⊓
⊔
3. Fix any good coordinate i and let j be a good coordinate such that all
terminals on one side of the i-cut have their j coordinate set to bj . Then
both endpoints of the edge labeled i have their j th coordinate set to bj .
Proof. Suppose that some endpoint of the edge on which i flips has j (which
is constant on C0 ) set to b̄j . Then, since the edge allows only one coordinate
to flip across it, both endpoints are labelled with j = b̄j . Then, the side
where j is constant (say C0 ) has to pay for j.
If j is constant and set to bj on C1 , j has to flip twice, a contradiction since
j is good. If j is constant and set to b̄j on C0 , then the labels on i’s edge
are labelled by some constant setting of j. If j is non-constant Assuming
j is non-constant for C1 , j flips somewhere in C1 . But then, j flips twice,
contradicting the fact that j is good.
⊓
⊔
4. A good coordinate i and a bad coordinate i′ cannot define the same cut.
Proof. This follows directly from Fact 2.1, since i and i′ will occur the same
number of times in an optimal tree and a good coordinate occurs exactly
once, while a bad coordinate occurs at least twice.
⊓
⊔
B
Proof of Lemma 4
Here we give the full proof of Lemma 4. For convenience, we reiterate the lemma.
Lemma 5. Assume that when step 3 is entered, P is a proper partition of the
terminals over d′ non-constant coordinates. Then step 3 of the algorithm returns
a forest of cost d′ + O(q 2 ).
Proof. Recall that for any coordinate i, we define the weight of i as the total
number of coordinates interchangeable with i. Let P post denote the instance
which results after splitting on all coordinates with weight greater than q. The
proof of the lemma reduces to showing that there exists a forest over P post of
cost O(q 2 ). If such a forest exists, the MST-based 2-approximation for each component returns a forest of cost O(q 2 ), and reconstructing the tree with plucked
and split edges has weight at most d′ . So, the forest over P we get has cost
≤ d′ + O(q 2 ).
The set of all terminals in all components of P post is composed of terminals
that belong to the original P, and the new terminals, added by coordinates of
weight more than q which are not simple. We construct a forest of low cost over
P post by (i) taking an optimal Steiner forest T over P, (ii) for every simple coordinate which is split upon, remove the path of corresponding coordinates from
T , and (iii) for every non-simple coordinate split upon, remove the path of corresponding coordinates and introduce the two vertices y(i) and y(i), connecting
each vertex to the end-point of the path that resides in the same side of the cut.
We denote the resulting forest by T post , and show that T post contains at most
O(q 2 ) edges.
First, observe that since there are at most q bad coordinates, by splitting
only on coordinates with weight more than q, we are guaranteed to split only
on good coordinates (we know from Fact 1 that the good and bad coordinates
cannot define the same cut). Therefore, since P is proper, the coordinate i resides
in a single component. Second, observe that the base of the recursion is executed
only when Pluck-a-leaf fails. Therefore, as in the proof of Lemma 3, T is a
forest with at most 2q leaves, so its path decomposition contains at most 4q
paths. Our next observation is the following proposition.
Proposition 1. Fix a path in the path decomposition of the forest. If there exists
some non-endpoint terminal on this path, all coordinates on this path are simple.
Proof. Let i be a coordinate associated with an edge on such a path. Let xi
be a terminal on the path which is the closest to i. Recall the observation from
before, that any two coordinates that yield the same terminal cut must appear in
the optimal tree adjacent to one another, and furthermore, their order does not
matter (see Fact 1). Therefore, we may assume i is adjacent to xi . Furthermore,
as xi is not an endpoint, there exists a good coordinate j 6= i adjacent to xi
on this path. It follows that i and j determine two different cuts over the set of
terminals. Then, just as in the proof of Lemma 3, i is simple, where xi is the
unique terminal on one side of its cut matching i’s pattern.
⊓
⊔
Proposition 1 means that any non-simple coordinate corresponds to an entire
path in the path decomposition (because all edges on a path with no terminals
on it determine the same cut). We deduce that (a) the number of non-simple
contracted coordinates of weight more than q is at most 8q. Furthermore, the
number of edges on any path of length no more than q with no terminal on them
is at most 8q 2 . Finally, since the base of the recursion is executed only when
Split fails to execute, the number of edges on paths that do have a terminal on
them is at most 16q 2 . It follows that the path decomposition of T contains at
most 16q 2 edges, in addition to the edges on paths of length more than q with
no terminal on them (and each such path causes the algorithm to split exactly
once).
We can now upper bound the number of edges in T post . The forest T post contains at most 2q ≤ 2q 2 bad edges; at most 16q 2 edges from T ; and all the paths
between the vertices we introduce by adding the y(i)-vertices and connecting
them to T . Let i be a non-simple contracted coordinate which was split upon,
and let u → v be the corresponding path on T which was removed. Assume y(i)
resides on the same side of the cut as u. Clearly, u and y(i) identify on all coordinates on the u → v path, but they also identify on all other good coordinates:
a good coordinate j appears most once in T , so u, v and all terminals on at least
one side of the u − v cut has the same value on their jth coordinate. It follows
that the path between y(i) and u is of length at most q, and the same applies to
the path between y(i) and v. Therefore, the number of edges on paths we have
yet to upper bound, sum to no more than 8q · 2 · q = 16q 2 . Thus T post contains
no more than 34q 2 edges. This completes the proof.
⊓
⊔
| 5cs.CE
|
1
Tracking Gaze and Visual Focus of Attention of
People Involved in Social Interaction
arXiv:1703.04727v2 [cs.CV] 21 Nov 2017
Benoit Massé, Silèye Ba, and Radu Horaud
Abstract—The visual focus of attention (VFOA) has been
recognized as a prominent conversational cue. We are interested
in estimating and tracking the VFOAs associated with multiparty social interactions. We note that in this type of situations
the participants either look at each other or at an object of
interest; therefore their eyes are not always visible. Consequently
both gaze and VFOA estimation cannot be based on eye detection
and tracking. We propose a method that exploits the correlation
between eye gaze and head movements. Both VFOA and gaze
are modeled as latent variables in a Bayesian switching statespace model. The proposed formulation leads to a tractable
learning procedure and to an efficient algorithm that simultaneously tracks gaze and visual focus. The method is tested and
benchmarked using two publicly available datasets that contain
typical multi-party human-robot and human-human interactions.
Index Terms—Visual focus of attention, eye gaze, head pose,
dynamic Bayesian model, switching Kalman filter, multi-party
dialog, human-robot interaction.
I. I NTRODUCTION
n this paper we are interested in the computational analysis
of social interactions. In addition to speech, people communicate via a large variety of non-verbal cues, e.g. prosody,
hand gestures, body movements, head nodding, eye gaze, and
facial expressions. For example, in a multi-party conversation,
a common behavior consists in looking either at a person,
e.g. the speaker, or at an object of current interest, e.g. a
computer screen, a painting on a wall, or an object lying on
a table. We are particularly interested in estimating the visual
focus of attention (VFOA), or who is looking at whom or at
what, which has been recognized as one of the most prominent
social cues. It is used in multi-party dialog to establish faceto-face communication, to respect social etiquette, to attract
someone’s attention, or to signify speech-turn taking, thus
complementing speech communication.
The VFOA characterizes a perceiver/target pair. It may be
defined either by the line from the perceiver’s face to the
perceived target, or by the perceiver’s direction of sight or gaze
direction (which is often referred to as eye gaze or simply
gaze). Indeed, one may state that the VFOA of person i is
target j if the perceiver’s gaze is aligned with the perceiverto-target line. From a physiological point of view, eye gaze
depends on both eyeball orientation and head orientation. Both
the eye and the head are rigid bodies with three and six degrees
of freedom respectively. The head position (three coordinates)
and the head orientation (three angles) are jointly referred to
I
B. Massé and R. Horaud are with INRIA Grenoble Rhône-Alpes, Montbonnot Saint-Martin, France.
S. Ba is with Dailymotion, Paris, France
This work is supported by ERC Advanced Grant VHIA #340113.
as the head pose. With proper choices for the head- and eyecentered coordinate frames, one can assume that gaze is a
combination of head pose and of eyeball orientation,1 and the
VFOA depends on head pose, eyeball orientation, and target
location.
In this paper we are interested into estimating and tracking
jointly the VFOAs of a group of people that communicate with
each other and with a robot, or multi-party HRI (human-robot
interaction), which may well be viewed as a generalization
of single-user HRI. From a methodological point of view the
former is more complex than the latter. Indeed, in single-user
HRI the person and the robot face each other and hence a
camera mounted onto the robot head provides high-resolution
frontal images of the user’s face such that head pose and
eye orientation can both be easily and robustly estimated.
In the case of multi-party HRI the eyes are barely detected
since the participants often turn their faces away from the
camera. Consequently, VFOA estimation methods based on
eye detection and eye tracking are ineffective and one has to
estimate the VFOA, indirectly, without explicit eye detection.
We propose a Bayesian switching dynamic model for the
estimation and tracking gaze directions and VFOAs of several
persons involved in social interaction. While it is assumed
that head poses (location and orientation) and target locations
can be directly detected from the data, the unknown gaze
directions and VFOAs are treated as latent random variables.
The proposed temporal graphical model, that incorporates gaze
dynamics and VFOA transitions, yields (i) a tractable learning algorithm and (ii) an efficient gaze-and-VFOA tracking
method.2 The proposed method may well be viewed as a
computational model of [1], [2].
The method is evaluated using two publicly available
datasets, Vernissage [3] and LAEO [4]. These datasets consist
of several hours of video containing situated dialog between
two persons and a robot (Vernissage) and human-human
interactions (LAEO). We are particularly interested in finding
participants that either gaze to each other, gaze to the robot,
or gaze to an object. Vernissage is recorded with a motion
capture system (a network of infrared cameras) and with a
camera placed onto the robot head. LAEO is collected from
TV shows.
1 Note that orientation generally refers to the pan, tilt and roll angles of
a rigid-body pose, while direction refers to the polar and azimuth angles or,
equivalently, a unit vector. Since the contribution of the roll angle to gaze is
generally marginal, in this paper we make no distinction between orientation
and direction.
2 Supplementary materials, that include a software package and examples
of results, are available at https://team.inria.fr/perception/research/eye-gaze/.
2
The remainder of this paper is organized as follows. Section II provides an overview of related work in gaze, VFOA
and head-pose estimation. Section III introduces the paper’s
mathematical notations and definitions, states the problem
formulation and describes the proposed model. Section IV
presents in detail the model inference and Section V derives
the learning algorithm. Section VI provides implementation
details and Section VII describes the experiments and reports
the results.
II. R ELATED W ORK
As already mentioned, the VFOA is correlated with gaze.
Several methods proceed in two steps, in which the gaze
direction is estimated first, and then used to estimate VFOA.
In scenarios that rely on precise estimation of gaze [5], [6]
a head-mounted camera, like the one in [7], can be used
to detect the iris with high accuracy. Head-mounted eye
trackers provide extremely accurate gaze measurements and in
some circumstances eye-tracking data can be used to estimate
objects of interest in videos [8]. Nevertheless, they are invasive
instruments and hence not appropriate for analyzing social
interactions.
Gaze estimation is relevant for a number of scenarios, such
as car driving [9] or interaction with smartphones [10]. In these
situations, either the field of view is limited, hence the range of
gaze directions is constrained (car driving), or active human
participation ensures that the device yields frontal views of
the user’s face, thus providing accurate eye measurements [7],
[9], [11], [12]. In some scenarios the user is even asked to
limit head movements [13], or to proceed through a calibration
phase [12], [14]. Even if no specific constraints are imposed,
single-user scenarios inherently facilitate the task of eye measurement [11]. At the best of our knowledge, there is no gaze
estimation method that can deal with unconstrained scenarios,
e.g. participants not facing the cameras, partially or totally
occluded eyes, etc. In general, eye analysis is inaccurate when
participants are faraway from the camera.
An alternative is to approximate gaze direction with head
pose [15]. Unlike eye-based methods, head pose can be
estimated from low-resolution images, i.e. distant cameras
[16], [17], [18], [19], [20]. These methods estimate gaze
only approximatively since eyeball orientation can differ from
head orientation by ±35° [21]. Gaze estimation from head
orientation can benefit from the observation that gaze shifts
are often achieved by synchronously moving the head and the
eyes [22], [1], [2]. The correlation between head pose and gaze
has also been exploited in [23]. More recently, [24] combined
head and eye features to estimate the gaze direction using an
RGB-D camera. The method still requires that both eyes are
visible.
Several methods were proposed to infer VFOAs either from
gaze directions [25], or from head poses [4], [26], [27], [28].
For example, in [4] it is proposed to build a gaze cone around
the head orientation and targets lying inside this cone are used
to estimate the VFOA. While this method was successfully
applied to movies, its limitation resides in its vagueness: the
VFOA information is limited to whether there are two people
looking at each other or not.
An interesting application of VFOA estimation it the analysis of social behavior of participants engaged in meetings,
e.g. [23], [26], [29], [30]. Meetings are characterized by
interactions between seated people that interact based on
speech and on head movements. Some methods estimate the
most likely VFOA associated with a head orientation [23],
[29]. The drawback of these approaches is that they must be
purposively trained for each particular meeting layout. The
correlation between VFOA and head pose was also investigated in [26] where an HMM is proposed to infer VFOAs
from head and body orientations. This work was extended
to deal with more complex scenarios, such as participants
interacting with a robot [27], [31]. An input-output HMM is
proposed in [31] to enable to model the following contextual
information: participants tend to look to the speaker, to the
robot, or to an object which is referred to by the speaker or
by the robot. The results of [31] show that this improves the
performance of VFOA estimation. Nevertheless, this method
requires additional information, such as speaker identification
or speech recognition.
The problem of joint estimation of gaze and of VFOA
was addressed in a human-robot cooperation task [28]. In
such a scenario the user doesn’t necessarily face the camera
and robot-mounted cameras have low-resolution, hence the
estimation of gaze from direct analysis of eye regions is not
feasible. [28] proposes to learn a regression between the space
of head poses and the space of gaze directions and then
to predict an unknown gaze from an observed head pose.
The head pose itself is estimated by fitting a 3D elliptical
cylinder to a detected face, while the associated gaze direction
corresponds to the 3D line joining the head center to the
target center. This implies that during the learning stage,
the user is instructed to gaze at targets lying on a table
in order to provide training data. The regression parameters
thus estimated correspond to a discrete set of head-pose/gazedirection pairs (one for each target); an erroneous gaze may be
predicted when the latter is not in the range of gaze directions
used for training.
A summary of the proposed Bayesian dynamic model and
experiments with the Vernissage [3] motion capture dataset
were presented in [32]. In this article we provide a detailed
and comprehensive description and analysis of the proposed
model, of the model inference, of the learning methodology,
and of the associated algorithms. We show results with both
motion capture and RGB data from Vernissage. Additionally,
we show results with the LAEO dataset [4].
III. P ROPOSED M ODEL
The proposed mathematical model model is inspired from
psychophysics [1], [2]. In unconstrained scenarios a person
switches his/her gaze from one target to another target,
possibly using both head and eye movements. Quick eye
movements towards a desired object of interest are called
saccades. Eye movements can also be caused by the vestibuleocular reflex that compensates for head movements such that
one can maintain his/her gaze in the direction of the target of
interest. Therefore, in the general case, gazing to an object is
achieved by a combination of eye and head movements.
3
Figure 1. This figure illustrates the principle of our method and displays the observed and latent variables associated with a person (left-person indexed by
i). The two images were grabbed with a camera mounted onto the head of a robot and they correspond to frames t − n (left image) and t (right image),
respectively. The following variables are displayed: head orientation (red arrow), Hit−n , Hit (observed variables), as well as the latent variables estimated
i
with the proposed method, namely gaze direction (green arrow), Git−n , Git , VFOA, Vt−n
, Vti , and head reference orientation (black arrow), Rit−n , Rit
(that coincides with upper-body orientation). In this example left-person gazes towards the robot at t − n, then turns her head to eventually gaze towards
i
right-person at t, hence her VFOA switches from Vt−n
= robot to Vti = right-person.
In the case of small gaze shifts, e.g. reading or watching
TV, eye movements are predominant. In the case of large
gaze shifts, often needed in social scenarios, head movements
are necessary since eyeball movements have limited range,
namely ±35° [21]. Therefore, the proposed model considers
that gaze shifts are produced by head movements that occur
simultaneously with eye movements.
A. Problem Formulation
We consider a scenario composed of N active targets and
M passive targets. An active target is likely to move and/or to
have a leading role in an interaction. Active targets are persons
and robots.3 Passive targets are objects, e.g. wall paintings.
The set of all targets is indexed from 0 to N + M , where
the index 0 designates “no target". Let i be an active target (a
person or a robot), 1 ≤ i ≤ N , and j be a passive target (an
object), N + 1 ≤ j ≤ N + M . A VFOA is a discrete random
variable defined as follows: Vti = j means person (or robot) i
looks at target j at time t. The VFOA of a person (or robot)
i that looks at none of the known targets is Vti = 0. The case
Vti = i is excluded. The set of all VFOAs at time t is denoted
by Vt = Vt1 , . . . , VtN .
Two continuous variables are now defined: head orientation
and gaze direction. The head orientation of person i at t
i
is denoted with Hit = [φiH,t , θH,t
]> , i.e. the pan and tilt
angles of the head with respect to some fixed coordinate
frame. The gaze direction of person i is denoted with Git
and is also parameterized by pan and tilt with respect to the
i
same coordinate frame, namely Git = [φiG,t , θG,t
]> . Although
eyeball orientation is neither needed nor used, it is worth
noticing that it is the difference between Git and Hit . These
variables are illustrated on Fig. 1.
3 Note that in case of a robot, the gaze direction and the head orientation
are identical and that the latter can be easily estimated from the head motors.
Finally, to establish a link between VFOAs and gaze directions, the target locations must be defined as well. Let
Xit = [xit , yti , zti ]> be the location of target i. In the case
of a person, this location corresponds to the head center
while in the case of a passive target, it corresponds to the
target center. These locations are defined in the same coordinate frame as above. Also notice that the direction from
the active target i to target j is defined by the unit vector
j
j
i
i
Xij
t = (Xt −Xt )/kXt −Xt k which can also be parameterized
ij
i,j
i,j >
by two angles, Xt = [φX,t , θX,t
] .
As already mentioned, target locations and head orientations
are observed random variables, while VFOAs and gaze directions are latent random variables. The problem to be solved
can now be formulated as a maximum a posteriori (MAP)
problem:
V̂t , Ĝt = argmax P (Vt , Gt |H1:t , X1:t )
(1)
Vt ,Gt
Since there is no deterministic relationship between head
orientation and gaze direction, we propose to model it probabilistically. For this purpose, we introduce an additional
latent random variable, namely the head reference orientation,
i
Rit = [φiR,t , θR,t
]> , which we choose to coincide with
the upper-body orientation. We use the following generative
model, initially introduced in [26], linking gaze direction, head
orientation, and head reference orientation:
P (Hit |Git , Rit ; α, ΣH ) = N (Hit ; µiH,t , ΣH ),
with
µiH,t
=
αGit
+ (I2 −
α)Rit ,
(2)
(3)
where ΣH ∈ R2×2 is a covariance matrix, I2 ∈ R2×2 is the
identity matrix and α = Diag (α1 , α2 ) is a diagonal matrix
of mixing coefficients, 0 < α1 , α2 < 1. Also it is assumed
that the covariance matrix is the same for all the persons and
over time. Therefore, head orientation is an observed random
variable normally distributed around a convex combination
4
Xt−1
associated with different people are independent, given the
VFOAs. By combining the above equations with this independency assumption, we obtain:
Y
P (Ht |Gt , Rt ) =
N (Hit ; µiH,t , ΣH )
(9)
Xt
Vt−1
Vt
i
P (Gt |Gt−1 , Ġt−1 , Vt , Xt ) =
Gt−1
Gt
Y
i
δj (Vt )
N (Git ; µij
G,t , ΓG )
i,j
(10)
Rt−1
P (Rt |Rt−1 , Ṙt−1 ) =
Rt
Y
N (Rit ; µiR,t , ΓR )
(11)
i
Ht−1
Ht
Figure 2. Graphical representation showing the dependencies between the
variables of the proposed Bayesian dynamic model. The discrete latent
variables (visual focus of attention) are shown with squares while continuous
variables are shown with circles: observed variables (head pose and target
locations) are shown with shaded circles and latent variables (gaze and
reference) are shown with white circles.
between two latent variables: gaze direction and head reference
orientation.
B. Gaze Dynamics
where the dependencies between variables are embedded in
the variable means, i.e. (3) and (6). The covariance matrices
will be estimated via training. While gaze directions can vary
a lot, we assume that head reference orientations are almost
constant over time, which can be enforced by imposing that
the total variance of gaze is much larger than the total variance
of head reference orientation, namely:
Tr(ΓG ) Tr(ΓR ),
(12)
C. VFOA Dynamics
Using a first-order Markov approximation, the VFOA transition probabilities can be written as:
P (Vt |V1:t−1 ) = P (Vt |Vt−1 ),
The following model is proposed:
(13)
N
P (Git |Git−1 Ġit−1 , Vti = j, Xt ) = N (Git ; µij
G,t , ΓG ),
(4)
P (Ġit |Ġit−1 ) = N (Ġit ; Ġit−1 , ΓĠ ),
(5)
with:
(
µij
G,t
=
Git−1 + Ġit−1 dt,
if j = 0,
(6)
i
βGit−1 + (I2 − β)Xij
+
Ġ
dt,
if
j 6= 0,
t
t−1
where Ġit = dGit /dt is the gaze velocity, ΓG , ΓĠ ∈ R2×2
are covariance matrices, and β = Diag (β1 , β2 ) is a diagonal
matrix of mixing coefficients, 0 < β1 , β2 < 1. Therefore, if a
person looks at one of the targets, then his/her gaze dynamics
depends on the person-to-target direction Xij
t at a rate equal
to β, and if a person doesn’t look at one of the targets, then
his/her gaze dynamics follows a random walk.
The head reference orientation dynamics can be defined in
a similar way:
P (Rit |Rit−1 , Ṙit−1 ) = N (Rit ; µiR,t , ΓR ),
P (Ṙit |Ṙit−1 )
with µiR,t
=
=
N (Ṙit ; Ṙit−1 , ΓṘ ),
Rit−1 + Ṙit−1 dt,
(7)
(8)
where Ṙit = dRit /dt is the head reference orientation velocity
and ΓR , ΓṘ ∈ R2×2 are covariance matrices. The dependencies between all the model variables are shown as a graphical
representation in Figure 2.
It is assumed that the gaze directions associated with different people are independent, given the VFOAs V1:t . The crossdependency between different people is taken into account
by the VFOA dynamics as detailed in section III-C below.
Similarly, head orientations, and head reference orientations
Notice that matrix P (Vt |Vt−1 ) is of size (N + M ) × (N +
M )N . Indeed, there are N persons (active targets), and N +
M +1 targets (one "no" target, N active targets and M passive
targets) and the case of a person that looks to him/herself is
excluded. For example, if N = 2 and M = 4, matrix (13)
has (2 + 4)2×2 = 1296 entries. The estimation of this matrix
would require, in principle, a large amount of training data, in
particular in the presence of many symmetries. We show that,
in practice, only 15 different transitions are possible. This can
be seen on the following grounds.
We start by assuming conditional independence between the
VFOAs at t:
Y
P (Vt |Vt−1 ) =
P (Vti |Vt−1 ).
(14)
i
Vti ,
Let’s consider
the VFOA of person i at t, given Vt−1 ,
the VFOAs at t − 1. One can distinguish two cases:
i
• Vt−1 = k where k is either a passive target, N < k ≤
N + M , or it is none of the targets, k = 0; in this case
i
Vti depends only on Vt−1
, and
i
• Vt−1 = k, where k 6= i is a person 1 ≤ k ≤ N ; in this
i
k
.
case Vti depends on the both Vt−1
and Vt−1
To summarize, we can write that:
P (Vti = j|Vt−1 ) =
(
i
k
P (Vti = j|Vt−1
= k, Vt−1
= l) if 1 ≤ k ≤ N,
(15)
i
i
P (Vt = j|Vt−1 = k)
otherwise.
Based on this it is now possible to count the total number
of possible VFOA transitions. With the same notations as in
(15), we have the following possibilities:
5
k = 0 (no target): there are two possible transitions, j = 0
and j 6= 0.
• N < k ≤ N +M (passive target): there are three possible
transitions, j = 0, j = k, and j 6= k.
• 1 ≤ k ≤ N, l = 0 (active target k looks at no target):
there are three possible transitions, j = 0, j = k, and
j 6= k.
• 1 ≤ k ≤ N, l = i (active target k looks at person i): there
are three possible transitions, j = 0, j = k, and j 6= k.
• 1 ≤ k ≤ N, l 6= 0, i (active target k looks at active target l
different than i): there are four possible transitions, j = 0,
j = k, j = l and j 6= k, l.
Therefore, there are 15 different possibilities for P (Vti =
j|Vt−1 ), i.e. appendix A. Moreover, by assuming that the
VFOA transitions don’t depend on i, we conclude that the
transition matrix may have up to 15 different entries. Moreover, the number of possible transitions is even smaller if there
is no passive target (M = 0), or if the number of active targets
is small, e.g. N < 3. This considerably simplifies the task of
estimating this matrix and makes the task of learning tractable.
•
IV. I NFERENCE
We start by simplifying the notation, namely Lt =
[Gt ; Ġt ; Rt ; Ṙt ] where [·; ·] denotes vertical concatenation.
The emission probabilities (9) become:
Y
P (Ht |Lt ) =
N (Hit ; µiH,t , ΣH ),
(16)
i
with µiH,t = CLit ,
(17)
where matrix C is obtained from the definition of
and from (3):
α1 0 0 0 1 − α1
0
0
C=
0 α2 0 0
0
1 − α2 0
Lt above
0
.
0
j
ij i
ij
with µij
L,t = At Lt−1 + bt
ΓG
ΓĠ
and ΓL =
ΓR
P (Lt ,Vt |H1:t−1 , X1:t )
XZ
=
P (Lt , Vt , Lt−1 , Vt−1 |H1:t−1 , X1:t )dLt−1
Vt−1
=
XZ
× P (Lt−1 , Vt−1 |H1:t−1 , X1:t−1 )dLt−1 ,
(19)
(23)
where unnecessary dependencies were removed. Combining (22) and (23) we obtain a recursive formulation in
P (Vt , Lt |H1:t , X1:t ). However, this model is still intractable
without further assumptions. The main approximation used
in this work consists of assuming local independence for the
posteriors:
Y
P (Lt , Vt |H1:t , X1:t ) '
P (Lit , Vti |H1:t , X1:t ).
(24)
i
A. Switching Kalman Filter Approximation
Several strategies are possible, depending upon the structure of P (Lt , Vt |H1:t , X1:t ). Commonly used strategies to
evaluate this distribution include variational Bayes or MonteCarlo. Alternatively, we propose to cast the problem into the
framework of switching Kalman filters (SKF) [33]. We assume
the filtering distribution to be Gaussian,
P (Lt , Vt |H1:t , X1:t ) ∝ N (Lt ; µt , Σt ).
(25)
From (24) and (25) we obtain the following factorization:
YY
ij δj (Vti )
P (Lt , Vt |H1:t , X1:t ) ∝
N (Lit ; µij
. (26)
t , Σt )
j
Thus, (23) can be split into N components, one for each active
target i:
P (Lit ,Vti = j|H1:t , X1:t ) ∝ P (Hit |Lit )
XZ
ij
i
i
×
N (Lit ; Aij
t Lt−1 + bt )P (Vt |Vt−1 )
Vt−1
,
P (Lt |Vt , Lt−1 , Xt )P (Vt |Vt−1 )
Vt−1
i
The transition probabilities can be obtained by combining (10) and (11) with (5) and (8):
YY
δj (Vti )
P (Lt |Vt , Lt−1 , Xt ) =
N (Lit ; µij
, (18)
L,t , ΓL )
i
where c is the normalization evidence. Now we can introduce
Vt−1 and Lt−1 using the sum rule:
×
(20)
ΓṘ
ij
where Aij
t is an 8 × 8 matrix and bt is an 8 × 1 vector.
The indices i,j and t cannot be dropped since the transitions
depend on Xij
t from (6).
The MAP problem (1) can now be derived in a Bayesian
framework for the VFOA variables:
Z
P (Vt |H1:t , X1:t ) = P (Vt , Lt |H1:t , X1:t )dLt .
(21)
We propose to study the filtering distribution of the joint latent
variables, namely P (Vt , Lt |H1:t , X1:t ). Indeed, Bayes rule
yields:
1
P (Vt , Lt |H1:t , X1:t ) = P (Ht |Lt )P (Lt , Vt |H1:t−1 , X1:t ).
c
(22)
Y
i
ik
δk (Vt−1 )
N (Lit−1 ; µik
dLit−1 ,
t−1 , Σt−1 )
(27)
k
or, after several algebraic manipulations:
X ijk
ijk
P (Lit , Vti = j|H1:t , X1:t ) ∝
wt−1,t N (Lit ; µijk
t , Σt ).
k
(28)
In this expression, µijk
and Σijk
are obtained by performing
t
t
ik
constrained Kalman filtering on µik
t−1 , Σt−1 with transition
ij
ij
dynamics defined by At and bt , emission dynamics defined
ijk
by C, and observation Hit , i.e. [34]. The weights wt−1,t
are
i
i
defined as P (Vt−1 = k | Vt = j, H1:t , X1:t ). The constraint
comes from the fact that ||Git − Hit || < 35° and is achieved
by projecting the mean (refer to [34] for more details).
This can be rephrased as follows: from the filtering distribution at time t − 1, there are N + M possible dynamics
for Lit . The normal distribution at time t − 1 then becomes
6
a mixture of N + M normal distributions at time t as
shown in (28). However, we expect a single Gaussian such
ij
as P (Lit , Vti = j|H1:t , X1:t ) ∝ N (Lit ; µij
t , Σt ). This can be
done by moment matching:
X ijk
µij
wt−1,t µijk
(29)
t =
t
Σij
t =
k
X
ijk
ijk
>
(Σijk
+ (µijk
− µij
− µij
wt−1,t
t
t
t )(µt
t ) ) (30)
k
ijk
. Let’s introduce
Finally, it is necessary to evaluate wt−1,t
the following notations:
i
i
cijk
t−1,t = P (Vt = j, Vt−1 = k|H1:t , X1:t ),
cij
t
=
P (Vti
= j|H1:t , X1:t ).
(31)
A. Learning the VFOA Transition Probabilities
The VFOA transitions are drawn from the generalized
Bernoulli distribution. Therefore, the transition probabilities
can be estimated with P (Vti = j|Vt−1 ) = Et−1 [δj (Vti )],
where δj (i) is the Kronecker delta function. In Section III-C
we showed that there are up to 15 different possibilities for
the VFOA transition probability. This enables us to derive an
explicit formula for each case, see appendix B. Consider for
i
example one of these cases, namely p14 = P (Vti = l|Vt−1
=
k
k, Vt−1 = l), which is the conditional probability that at t
person i looks at target l, given that at t − 1 person i looked at
person k and that person k looked at target l. This probability
can be estimated with the following formula:
(32)
Nq Tq Nq
Q X
X
XX X
It follows that
cij
t =
X
cijk
t−1,t
ijk
and wt−1,t
=
cijk
t−1,t
k
cij
t
p̂14 =
.
By applying Bayes formula to cijk
t−1,t , yields:
q,i
q,k
δl (Vtq,i )δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
q,i
q,k
δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
i
i
cijk
t−1,t ∝ P (Ht |Vt = j, Vt−1 = k, H1:t−1 , X1:t )
i
i
×cik
t−1 P (Vt = j|Vt−1 = k, H1:t−1 , X1:t−1 )
(33)
ijk
Then, cik
t−1 is obtained from ct−2,t−1 calculated at previous
step. The last factor in (33) is either equal
P time
i
k
kl
i
to
c
P
l t−1 (Vt = j|Vt−1 = k, Vt−1 = l) if k is
i
i
an active target, or P (Vt = j|Vt−1 = k) otherwise.
Both cases are straightforward to compute. Finally, the
first factor in (33), the observation component, can be
i
= k, H1:t−1 , X1:t ) ×
factorized
as P (Hit |Vti = j, Vt−1
Q
P P
n
n
n
=
m,
V
=
p, H1:t−1 , X1:t ). By
|V
P
(H
t
t
t−1
n6=i
m
p
introducing the latent variable L, we obtain:
n
P (Hnt |Vtn = m, Vt−1
= p, H1:t−1 , X1:t )
Z
= P (Hnt |Lnt ) P (Lnt |Lnt−1 , Vtn = m, Xt )
n
× P (Lnt−1 |Vt−1
= p, H1:t−1 , X1:t−1 )dLnt−1 dLnt .
(34)
All the factors (34) are normal distributions, hence it integrates in closed-form. In summary, we devised a procedure
to estimate an online approximation of the joint filtering
distribution of the VFOAs, Vt , and of the gaze and head
reference directions, Lt .
V. L EARNING
The proposed model has two sets of parameters that must
be estimated: the transition probabilities associated with the
discrete VFOA variables, and the parameters associated with
the Gaussian distributions. Learning is carried out using Q
recordings with annotated VFOAs. Each recording is composed of Tq frames, 1 ≤ q ≤ Q and contains Nq active
targets (the robot is the active target 1 and the persons are
indexed from 2 to Nq ) and Mq passive targets. In addition to
target locations and head poses, it is worth noticing that the
learning algorithm requires VFOA ground-truth annotations,
while gaze directions are still treated as latent variables.
B. Learning the Gaussian Parameters
In Section IV we described the derivation of the proposed
model that is based on SKF. This model requires the parameters (means and covariances) of the Gaussian distributions
defined in (16) and (18). Notice however that the mean (17) of
(16) is parameterized by α. Similarly, the mean (19) of (18) is
parameterized by β. Consequently, the model parameters are:
θ = (α, β, ΓL , ΣH ),
(35)
and we remind that α and β are 2 × 2 diagonal matrices, ΓL
is a 8 × 8 covariance and and ΣH is a 2 × 2 covariance, and
that we assumed that these matrices are common to all the
active targets. Hence the total number of parameters is equal
to 2 + 2 + 36 + 3 = 43.
In the general case of SKF models, the discrete variables
are unobserved both for learning and for inference. Here
we propose a learning algorithm that takes advantage of the
fact that the discrete variables, i.e. VFOAs, are observed
during the learning process, namely the VFOAs are annotated.
We propose an EM algorithm adapted from [35]. In the
case of a standard Kalman filter, an EM iteration alternates
between a forward-backward pass to compute the expected
latent variables (E-step), and between the maximization of the
expected complete-data log-likelihood (M-step).
We start by describing the M-step. The complete-data loglikelihood is:
ln P (L1 , H1 , . . . , LQ , HQ |θ)
=
Nq Tq
Q X
X
X
q,i
ln P (Lq,i
t |Lt−1 , β, ΓL )
q=1 i=2 t=2
+
Nq Tq
Q X
X
X
q=1 i=2 t=1
q,i
ln P (Hq,i
t |Lt , α, ΣH ).
(36)
7
By taking the expectation w.r.t. the posterior distribution
P (L1 , . . . , LQ |H1 , . . . , HQ , θ), we obtain:
Q(θ, θ old ) = EL1 ,...,LK |θold ln P (L1 , H1 , . . . , LQ , HQ |θ) ,
(37)
which can be maximized w.r.t. to the parameters θ, which
yields closed-form expressions for the covariance matrices:
Nq Tq
Q X
X
X
ΓL =
Q
X
(38)
Kt = Pt,t−1 C (CPt,t−1 C + ΣH )
The backward pass estimates
N (Lt ; µ̂t , P̂t ) and leads to
Q
X
(Nq − 1)Tq
−1
.
P (Lt |H1 , . . . , HT )
µ̂t = µt + Jt (µ̂t+1 − (At+1 µt + bt+1 )),
Pt+1,t )J>
t ,
(45)
=
(46)
(47)
where:
(48)
The expectations are estimated by performing a forwardbackward pass over all the persons and all the recordings of
the training data. This yields the following formulas:
q,i
E[Lq,i
t ] = µ̂t
q,i
q,i
q,i >
E[(Hq,i
t − µH,t )(Ht − µH,t ) ]
q=1 i=2 t=1
(44)
>
−1
Jt = P t A >
.
t+1 (Pt+1,t )
(Nq − 1)(Tq − 1)
q,ij q,i
q,ij
where µq,ij
L,t = At Lt−1 + bt , i.e. (19), and:
, (39)
q=1
µq,i
H,t
>
P̂t = Pt + Jt (P̂t+1 −
q=1
ΣH =
Pt,t−1 = At Pt−1 A>
t + ΓL ,
q,ij
q,ij >
i
E[(Lq,i
t − µL,t )(Lt − µL,t ) ]
q=1 i=2 t=2
Nq Tq
Q X
X
X
where:
q,i >
E[Lq,i
]
t Lt
>
q,i
E[Lq,i
t Lt−1 ]
=
=
q,i >
+ µ̂q,i
t µ̂t
q,i >
q,i q,i >
P̂q,i
t Jt−1 + µ̂t µ̂t−1
P̂q,i
t
(49)
(50)
(51)
VI. I MPLEMENTATION D ETAILS
CLq,i
t ,
i.e. (17).
where
=
The estimation of α and of β is carried out in the following
way. ∂Q(θ, θ old )/∂β1 = 0 and ∂Q(θ, θ old )/∂β2 = 0 yield a
set of two linear equations in the two unknowns:
Nq Tq
Q X
X
X q,i
q,i
q,ij
> −1 ∂
E (Lt − µq,ij
)
Γ
(L
−
µ
)
= 0,
L
L,t
L,t
∂β1 t
q=1 i=2 t=2
Nq Tq
Q X
X q,i
X
q,ij > −1 ∂
q,i
q,ij
E (Lt − µL,t ) ΓL
(L − µL,t ) = 0,
∂β2 t
q=1 i=2 t=2
The proposed method was evaluated on the Vernissage
dataset [3] and on the Looking At Each Other (LAEO)
dataset [4]. We describe in detail these datasets and their annotations. We provide implementation details and we analyse
the complexity of the proposed algorithm.
A. The Vernissage Dataset
The Vernissage scenario can be briefly described as follows,
e.g. Fig. 3: there are three wall paintings, namely the passive
targets denoted with o1 , o2 , and o3 (M = 3); two persons,
(40) denoted left person (left-p) and right person (right-p), interact
with the robot, hence N = 3. The robot plays the role of
and similarly:
an art guide, describing the paintings and asking questions
Nq Tq
Q X
X
X q,i
to the two persons in front of him. Each recording is split
q,i
> −1 ∂
E (Ht − µq,i
(Hq,i
t − µH,t ) = 0, into two roughly equal parts. The first part is dedicated to
H,t ) ΣH
∂α1
q=1 i=2 t=1
painting explanation, with a one-way interaction. The second
Nq Tq
Q X
X
X
∂
part consists of a quiz, thus illustrating a dialog between the
q,i > −1
q,i
(Hq,i
E (Hq,i
t − µH,t ) = 0, two participants and the robot, most of the time concerning
t − µH,t ) ΣH
∂α2
q=1 i=2 t=1
(41) the paintings.
where as above, the expectation is taken w.r.t. to the posterior
distribution. Once the formulas above are expanded and once
q,i
the means µq,ij
L,t and µH,t are substituted with their expressions, the following terms remain to be estimated: E[Lq,i
t ],
q,i >
q,i q,i >
E[Lq,i
L
]
and
E[L
L
].
t
t
t
t−1
The E-step provides estimates of these expectations via
a forward-backward algorithm. For the sake of clarity, we
drop the superscripts i (active target index) and q (recording
index) up to equation (48) below. Introducing the notation
P (Lt |H1 , . . . , Ht ) = N (Lt ; µt , Pt ), the forward-pass equations are:
µt = At µt−1 + bt + Kt (Ht − C(At µt−1 + bt ))
(42)
Pt = (I − Kt C)Pt,t−1 ,
(43)
Figure 3. The Vernissage setup. Left: Global view of an “exhibition" showing
wall paintings, two participants, i.e. left-p and right-p, and the NAO robot.
Right: Top view of the room showing the Vernissage layout.
The scene was recorded with a camera embedded into
the robot head and with a VICON motion capture system
consisting of a network of infrared cameras, placed onto the
8
walls, and of optical markers, placed onto the robot and people
heads. Both were recorded at 25 frames per second (fps).
There is a total of ten recordings, each lasting ten minutes. The
VICON system provided accurate estimates of head positions,
X1:T and head orientations, H1:T . Head positions and head
orientations were also estimated using from the RGB images
gathered with the camera embedded into the robot head. The
RGB images are processed as follows. We use the OpenCV
version of [36] to detect faces and their bounding boxes which
are then tracked over time using [37]. Next, we extract HOG
descriptors from each bounding box and apply a head-pose
estimator, e.g. [38]. This yields H̃1:T . The 3D head positions,
X̃1:T , can be estimated using the line of sight through the
face center and the bounding-box size, which provides a rough
estimate of the depth along the line of sight.
In the remaining of this paper, X1:T and H1:T are referred
to as Vicon Data; X̃1:T and H̃1:T as RGB Data. Because the
whole setup was carefully calibrated, both Vicon and RGB
Data are represented in the same coordinate frame.
In all our experiments we assumed that the passive targets
are static and their positions are provided in advance. The
position of the robot itself is also known in advance and one
can easily estimate the orientation of the robot head from
motor readings. Finally, the VFOAs of the participants were
manually annotated in all the frames of all the recordings.
B. The LAEO Dataset
The LAEO dataset [4] is an extension of the TVHID (TV
Human Interaction Dataset) [39]. It consists of 300 videos
extracted from TV shows. At least two actors appear in each
video engaged in four human-human interactions: handshake,
highfive, hug, and kiss. There are 50 videos for each interaction and 100 videos with no interaction. The videos have been
grabbed at 25 fps and each video lasts from five seconds to
twenty-five seconds. LAEO is further annotated, namely some
of these videos are split into shots which are separated by cuts.
There are 443 shots in total which are manually annotated
whenever two persons look at each other, [4].
While there is no passive target in this dataset (M = 0),
the number of active targets (N ) corresponds to the number
of persons in each shot. In practice N varies from one to
eight persons. All the faces in the dataset are annotated with
a bounding box and with a coarse head-orientation label:
frontal-right, frontal-left, profile-right, profile-left, backward.
As with Vernissage, we use the bounding-box center and size
to estimate the 3D coordinates of the heads, X1:T . We also
assigned a yaw value to each one of the five coarse head
orientations, H1:T . We also computed finer head orientations,
H̃1:T , using [38].
C. Algorithmic Details
The inference procedure is summarized in Algorithm 1.
This is basically an iterative filtering procedure. The update
step consists of applying the recursive relationship, derived in
ij
ij
ijk
ijk
Section IV, to µij
and cijk
t , Σt and ct , using µt , Σt
t−1,t as
intermediate variables. The VFOA is chosen using MAP, given
observations up to the current frame, and the gaze direction is
the mean of the filtered distribution (the first two components
of µij
t are indeed the mean for the pan and tilt gaze angles).
Algorithm 1 Inference
1: procedure G AZE A NDVFOA
2:
X1 , H1 ← G ET O BSERVATIONS(time = 1)
3:
c1 , µ1 , Σ1 ← I NITIALIZATION(H1 , X1 )
4:
V1i ← argmaxj cij
1
5:
Gi1 ← µij
1 [1..2]
6:
for t = 2..T do
7:
Xt , Ht ← G ET O BSERVATIONS(time = t)
8:
ct , µt , Σt ← U PDATE(Ht , Xt , ct−1 , µt−1 , Σt−1 )
9:
Vti ← argmaxj cij
t
10:
Git ← µij
[1..2]
t
11:
return V1:T , G1:T
Let’s now describe the initialization procedure used by
Algorithm 2. In a probabilistic framework, parameter intialization is generally addressed by defining an initial distribution, e.g. P (L1 |V1 ). Here, we did not explicitly define
such a distribution. Initialization is based on the fact that,
with repeated similar observation inputs, the algorithm reaches
a steady-state. The initialization algorithm uses a repeated
update method with initial observation to provide an estimate
of gaze and of reference directions. Consequently, the initial
filtering distribution P (L1 , V1 |H1 , X1 ) is implicitly defined
as the expected stationary state.
Algorithm 2 Initialization
1: procedure I NITIALIZATION (H1 , X1 )
2:
µin ← [H1 ; 0; H1 ; 0]
3:
Σin ← I
1
4:
cin ← N +M
(Uniform)
5:
while Not Convergence do
6:
cin , µin , Σin ← U PDATE(H1 , X1 , cin , µin , Σin )
7:
return cin , µin , Σin
D. Algorithm Complexity
The computational complexity of Algorithm 1 is
C = T (CO + CU ) + TI CU ,
(52)
where T is the number of frames in a test video, TI is the
number of iterations needed by the Algorithm 2 (initialization) to converge, CO is the computational complexity of
G ET O BSERVATION and CU is the computational complexity
of U PDATE. The complexity of one iteration of Algorithm 1
is CO + CU . CO depends on face detection and head pose
estimation algorithms. Hence we concentrate on CU . From
Section IV one sees that the following values need to be
i
computed: P (Hit |Vti = j, Vt−1
= k, H1:t−1 , X1:t−1 ), cijk
t−1,t ,
ijk
ijk
ij
ij
ij
µt , Σt , and then ct , µt and Σt , for each active target
i, and for each combination of targets j and k different from
i. There are N possible values for i and (N + M ) possible
values for j and k. Then,
CU = K × N (N + M )2 ,
(53)
9
where K is a factor whose complexity can be estimated as
follows. The most time-consuming part is the Kalman Filter
ik
algorithm used to estimate µijk
and Σijk
from µik
t
t
t and Σt .
These calculations are dominated by several 8×8 and 2×8
matrix inversions and multiplications. By neglecting scalar
multiplications and matrix additions, and by denoting with
CKF the complexity of the Kalman filter, we obtain that
K ≈ CKF and hence CU ≈ CKF × N (N + M )2 .
VII. E XPERIMENTAL RESULTS
A. Vernissage Dataset
We applied the same experimental protocol to the Vicon and
RGB data. We used a leave-one-video-out strategy for training.
The test is performed on the left out video. We used the frame
recognition rate (FRR) metrics to quantitatively evaluate the
methods. FRR computes the percentage of frames for which
the VFOA is correctly estimated. One should note however
that the ground-truth VFOAs were obtained by manually
annotating each frame in the data. This is subject to errors
since the annotator has to associate a target with each person.
The VFOA transition probabilities and the model parameters
were estimated using the learning method described in Section V. Appendix B provides the formulas used for estimating
the VFOA transition probabilities given the annotated data.
Notice that the fifteen transitions probabilities thus estimated
are identical for both data, Vicon and RGB.
The Gaussian parameters, i.e. (35), were estimated using
the EM algorithm of Section V-B. This learning procedure
requires head-pose estimates as well as the targets locations,
estimated as just explained. Since these estimates are different
for the two kinds of data (Vicon and RGB) we carried out
the learning process twice, with the Vicon data and with the
RGB data. The EM algorithm needs initialization. The initial
parameter values for α and β are α0 = β 0 = Diag (0.5, 0.5).
Matrices ΣH and ΓL defined in (20) are initialized with
isotropic covariances: Σ0H = σI2 , Γ0G = Γ0Ġ = γI2 , and
Γ0R = Γ0Ṙ = ηI2 with σ = 15, γ = 5, and η = 0.5. In
particular, this initialization is consistent with (12). In practice
we noticed that the covariances estimated by training remain
consistent with (12).
B. Results with Vicon Data
The FRR of the estimated VFOAs for the Vicon data are
summarized in Table I. A few examples are shown in Figure 5.
The FRR score varies between 28.3% and 74.4% for [26]
and between 43.1% and 79.8% for the proposed method.
Notice that high scores are obtained by both methods for
recording #27. Similarly, low scores are obtained for recording
#26. Since both methods assume that head motions and gaze
shifts occur synchronously, an explanation could be that this
hypothesis is only valid for some of the participants. The
confusion matrices for VFOA classification using Vicon data
are given in Figure 4. There are a few similarities between
the results obtained with the two methods. In particular, wall
painting #o2 stands just behind Nao and both methods don’t
always discriminate between these two targets. In addition,
the head of one of the persons is often aligned with painting
Table I
FRR SCORES OF THE ESTIMATED VFOA S FOR THE V ICON DATA FOR THE
LEFT AND RIGHT PERSONS ( LEFT- P AND RIGHT- P ).
Recording
09
10
12
15
18
19
24
26
27
30
Mean
Ba & Odobez [26]
left-p
right-p
51.6
65.1
64.3
74.4
53.5
67.6
67.1
46.2
37.5
28.3
56.7
45.4
44.9
49.0
40.3
32.9
65.8
72.0
69.1
49.1
54.5
Proposed
left-p right-p
59.8
61.4
76.5
65.0
61.6
63.2
64.8
67.6
62.0
53.7
54.5
60.4
59.7
54.7
43.6
43.1
79.8
78.3
72.0
63.9
62.6
Figure 4. Confusion matrices for the Vicon data. Left: [26]. Right: Proposed
algorithm. Row-wise: ground-truth VFOAs. Column-wise: estimated VFOAs.
Diagonal terms represent the recall.
#o1 from the viewpoint of the other person. A similar remark
holds for painting #o3 . As a consequence both methods often
confuse the VFOA in these cases. This can be seen in the third
image of Figure 5. Indeed, it is difficult to estimate whether
the left person (left-p) looks at #o1 or at right-p.
Finally, both methods have problems with recognizing the
VFOA “nothing” or gaze aversion (Vti = 0). We propose
the following explanation: the targets are widespread in the
scene, hence it is likely that an acceptable target lies in most
of the gaze directions. Moreover, Nao is centrally positioned,
therefore the head orientation used to look at Nao is similar to
the resting head orientation used for gaze aversion. However,
in [26] the reference head orientation is fixed and poorly suited
for dynamic head-to-gaze mapping, hence the high error rate
on painting #o3 . Our method favors the selection of a target,
either active or passive, over the no target (nothing) case.
C. Results with RGB Data
The RGB images were processed as described in section VI-A above in order to obtain head orientations, H̃1:T ,
and 3D head positions, X̃1:T . Table II shows the accuracy
of these measurements (in degrees and in centimeters), when
compared with the ground truth provided by the Vicon motion
capture system. As it can be seen, while the head orientation
estimates are quite accurate, the error in estimating the head
positions can be as large as 0.8 m for participants lying in
between 1.5 m and 2.5 m in front of a robot, e.g. recordings
#19 and #24. In particular this error increases as a participant
is farther away from the robot. In these cases, the bounding
10
Figure 5. Results obtained with the proposed method on Vicon data. Gaze directions are shown with green arrows, head reference directions with dark-grey
arrows and observed head directions with red arrows. The ground-truth VFOA is shown with a black circle. The top row displays the image of the robot-head
camera. Top views of the room show results obtained for the left-p (middle row) and for the right-p (bottom row). In the last example the left-p gazes at
“nothing".
box is larger than it should be and hence the head position is,
on an average, one meter closer than the true position. These
relatively large errors in 3D head position affect the overall
behavior of the algorithm.
The FRR scores obtained with the RGB data are shown in
Table III. As expected the loss in accuracy is correlated with
the head position error: the results obtained with recordings
#09 and #30 are close to the ones obtained with the Vicon
data, whereas there is a significant loss in accuracy for the
other recordings. The loss is notable for [26] in the case of
the right person (right-p) for the recordings #12, #18 and #27.
The confusion matrices obtained with the RGB data are shown
Table II
M EAN ERROR FOR HEAD POSE ESTIMATIONS FROM RGB DATA , FOR THE
LEFT PERSON ( LEFT- P ) AND THE RIGHT PERSON ( RIGHT- P ). T HE ERRORS
IN HEAD POSITION ( CENTIMETERS ) AND ORIENTATION ( DEGREES ) ARE
COMPUTED WITH RESPECT TO VALUES PROVIDED BY THE MOTION
CAPTURE SYSTEM .
Video
09
12
18
19
24
26
27
30
Mean
Position error (cm)
left-p
right-p
18.1
20.8
35.7
41.5
36.9
12.8
86.0
87.4
86.5
73.9
50.2
56.9
64.5
58.3
16.7
13.3
46.4
Pan error
left-p
right-p
4.4°
4.8°
4.8°
5.5°
6.8°
3.7°
4.0°
5.8°
3.3°
3.5°
7.4°
9.0°
4.1°
5.8°
2.8°
2.9°
5.0°
Tilt error
left-p right-p
3.7°
3.2°
2.6°
3.8°
5.8°
2.5°
2.7°
3.7°
2.8°
2.7°
4.1°
5.2°
3.2°
4.4°
1.8°
2.7°
3.3°
on Fig. 6.
In the case of RGB data, the comparison between our
method and the method of [31] is biased by the use of different
head orientation and 3D head position estimators. Indeed,
the RGB data results reported in [31] were obtained with
unpublished methods for estimating head orientations and 3D
head positions, and for head tracking. Moreover, [31] uses
cross-modal information, namely the speaker identity based
on the audio track (one of the participants or the robot) as
well as the identity of the object of interest. We also note
that [31] reports mean FRR values obtained over all the
test recordings, instead of an FRR value for each recording.
Table IV summarizes a comparison between the average FRR
obtained with our method, with [26], and with [31]. Our
method yields a similar FRR score as [31] using the Vicon
data (first row) in which case the same head pose inputs are
Table III
FRR SCORES OF THE ESTIMATED VFOA S OBTAINED WITH [26] AND
WITH THE PROPOSED METHOD FOR THE RGB DATA . T HE LAST TWO
COLUMNS SHOW THE 3D HEAD POSITION ERRORS OF TABLE II.
Video
09
12
18
27
30
Mean
Ba & Odobez [26]
left-p
right-p
50.3
59.8
54.2
14.8
39.0
16.1
38.2
17.1
61.6
44.6
39.0
Proposed
left-p right-p
58.1
55.9
59.0
46.5
64.2
33.1
53.3
55.1
54.7
66.6
54.7
Head pos. error
left-p
right-p
18.1
20.8
35.7
41.5
36.9
12.8
64.5
58.3
16.7
13.3
11
comparable with those of [4] on this dataset. This is quite
remarkable knowing that we estimated the model parameters
with the Vernissage training data.
Table V
AVERAGE SHOT RECOGNITION RATE (SRR) OBTAINED WITH [26] AND
WITH THE PROPOSED METHOD .
Coarse head orientation
Fine head orientation
Figure 6. Confusion matrices for the RGB data. Left: [26]. Right: Proposed
algorithm. Row-wise: ground-truth VFOAs. Column-wise: estimated VFOAs.
Diagonal terms represent the recall.
Table IV
M EAN FRR SCORES OBTAINED WITH [26], WITH [31] AND WITH THE
PROPOSED METHOD . R ECORDING #26 WAS EXCLUDED FROM THE FRR
MEANS AS REPORTED IN [31]. M OREOVER , [31] USES ADDITIONAL
CONTEXTUAL INFORMATION .
Vicon data
RGB data
Ba & Odobez [26]
56.5
39.0
Sheikhi [31]
66.6
62.4
Proposed
64.7
54.7
used.
D. LAEO Dataset
As already mentioned in Section VI-B above, the LAEO
annotations are incomplete to estimate the person-wise VFOA
at each frame. Indeed, the only VFOA-related annotation is
whether two people are looking at each other during the shot.
This is not sufficient to know in which frames they are actually
looking at each other. Moreover, when more than two people
appear in a shot, the annotations don’t specify who are the
people that look at each other. For these reasons, we decided
to estimate the parameters using Vicon data of the whole
Vernissage dataset, i.e. cross-validation.
We used the same pipeline as with the Vernissage RGB
data to estimate 3D head positions, X̃1:T , from the face
bounding boxes. Concerning head orientation, there are two
cases: coarse head orientations (manually annotated) and fine
head orientations (estimated). Coarse head orientations were
obtained in the following way: pan and tilt values were
associated with each head orientation label, namely the pan
angles −20°, 20°, −80°, 80°, and 180° were assigned to
labels frontal-left, frontal-right, profile-left, profile-right, and
backwards respectively, while a tilt anble of 0° was assigned
to all five labels. Fine head orientations were estimated using
the same procedure as in the case of the Vernissage RGB
data, namely face detection, face tracking, and head orientation
estimation using [38]. Algorithm 1 was used to compute the
VFOA for each frame and for each person thus allowing to
determine who looks at whom, e.g. Fig. 8.
We used two shot-wise, not frame-wise, metrics since the
LAEO annotations are for each shot: the shot recognition rate
(SRR), e.g. Table V, and the average precision (AP), e.g.
Table VI. We note that [4] only provides AP scores. It is
interesting to note that the proposed method yields results
Ba & Odobez [26]
0.535
0.363
Proposed
0.727
0.479
Table VI
AVERAGE PRECISION (AP) OBTAINED WITH [4], WITH BA & O DOBEZ
[26] AND WITH THE PROPOSED METHOD .
Coarse head orientation
Fine head orientation
Marin-Jimenez et al. [4]
0.925
0.896
[26]
0.916
0.838
Proposed
0.923
0.890
VIII. C ONCLUSIONS
In this paper we addressed the problem of estimating and
tracking gaze and visual focus of attention of a group of
participants involved in social interaction. We proposed a
Bayesian dynamic formulation that exploits the correlation
between head movements and eye gaze on one side, and
between visual focus of attention and eye gaze on the other
side. We described in detail the proposed model. In particular
we showed that the entries of the large-sized matrix of VFOA
transition probabilities have a very small number of different
possibilities for which we provided closed-form formulae. The
immediate consequence of this simplified transition matrix is
that the associated learning doesn’t require a large training
dataset. We showed that the problem of simultaneously inferring VFOAs and gaze directions over time can be cast in
the framework of a switching state-space model, which yields
tractable learning and inference algorithms.
We applied the proposed method to two datasets, Vernissage
and LAEO. Vernissage contains several recordings of a humanrobot interaction scenario. We experimented both with motion
capture data gathered with a Vicon system and with RGB
data gathered with a camera mounted onto a robot head.
We also experimented with the LAEO dataset that contains
several hundreds of video shots extracted from TV shows.
A quite remarkable result is that the parameters obtained
by training the model with the Vernissage data have been
successfully used for testing the method with the LAEO data,
i.e. cross-validation. This can be explained by the fact that
social interactions, even in different contexts, share a lot of
characteristics. We compared our method with three other
methods, based on HMMs [26], on input-output HMMs [31],
and on a geometric model [4]. The interest of these methods
(including ours) resides in the fact that eye detection, unlike
many existing gaze estimation methods, is not needed. This
feature makes the above methods practical and effective in a
very large number of situations, e.g. social interaction.
We note that gaze inference from head orientation is an
ill-posed problem. Indeed, the correlation between gaze and
head movements is person dependent as well as context
12
Figure 7. Results obtained with the proposed method on RGB data. Gaze directions are shown with green arrows, head reference directions with dark-grey
arrows and observed head directions with red arrows. The ground-truth VFOA is shown with a black circle. The top row displays the image of the robot-head
camera. Top views of the room show results obtained for the left person (left-p, middle row) and the right person (right-p, bottom row).
Figure 8. This figure shows some results obtained with the LAEO dataset. The top row shows results obtained with coarse head orientation and the bottom
row shows results obtained with fine head orientation. Head orientations are shown with red arrows. The algorithm infers gaze directions (green arrows) and
VFOAs (blue circles). People looking at each others are shown with a dashed blue line.
dependent. It is however important to detect gaze whenever
the eyes cannot be reliably extracted from images and properly
analyzed. We proposed to solve the problem based on the
fact that alignments often occur between gaze directions and
a finite number of targets, which is a sensible assumption in
practice.
Contextual information could considerably improve the results. Indeed, additional information such as speaker recognition (as in [31]), speaker localization [40], speech recognition,
or speech-turn detection [41] may be used to learn VFOA
transitions in multi-party multimodal dialog systems.
In the future we plan to investigate discriminative methods
based on neural network architectures for inferring gaze direc-
tions from head orientations and from contextual information.
For example one could train a deep learning network from
input-output pairs of head pose and visual focus of attention.
For this purpose, one can combine a multiple-camera system,
to accurately detect the eyes of several participants and to estimate their head poses, with a microphone-array and associated
algorithms to infer speaker and speech information.
A PPENDIX A
VFOA T RANSITION P ROBABILITIES
Using the notations introduced in section III-C, let i, 1 ≤
i ≤ N , be an active target. In section III-C we showed that in
13
Nq Tq Nq +Mq
Q X
X
X X
practice the probability transition matrix has up to 15 different
entries. For completeness, these entries are listed below.
• The VFOA of i at t − 1 is neither an active nor a passive
target (k = 0):
p̂4 =
i
p1 =P (Vti = 0|Vt−1
= 0)
q,i
δk (Vtq,i )δk (Vt−1
)
q=1 i=2 t=2 k=Nq +1
Nq Tq Nq +Mq
Q X
X
X X
q,i
)
δk (Vt−1
q=1 i=2 t=2 k=Nq +1
i
p2 =P (Vti = j|Vt−1
= 0)
•
The VFOA of i at t − 1 is a passive target (N < k ≤
N + M ):
Nq Tq Nq +Mq
Q X
X X X
X
p̂5 =
q=1 i=2 t=2 k=Nq +1 j6=i,k
Nq Tq Nq +Mq
Q X
X
X X
i
p3 =P (Vti = 0|Vt−1
= k)
p4 =P (Vti
p5 =P (Vti
•
=
=
i
k|Vt−1
i
j|Vt−1
= k)
p6
p7
p8
p9
p10
p11
p12
p13
p14
p15
=
=
=
=
=
=
=
=
=
=
= k)
i
k
0|Vt−1
= k, Vt−1
= 0)
i
k
k|Vt−1 = k, Vt−1 = 0)
i
k
j|Vt−1
= k, Vt−1
= 0)
i
k
0|Vt−1 = k, Vt−1 = i)
i
k
k|Vt−1
= k, Vt−1
= i)
i
k
j|Vt−1 = k, Vt−1 = i)
i
k
0|Vt−1
= k, Vt−1
= l)
i
k
k|Vt−1 = k, Vt−1 = l)
i
k
l|Vt−1
= k, Vt−1
= l)
i
k
j|Vt−1 = k, Vt−1 = l)
Nq Tq Nq
Q X
X
XX
p̂6 =
This appendix provides the formulae allowing to estimate
the 15 transitions probabilities as explained in section V-A.
p̂1 =
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
X
XX
Nq Tq Nq
Q X
X
XX
p̂7 =
q,i
q,k
δk (Vt−1
)δ0 (Vt−1
)
q,i
q,k
)δ0 (Vt−1
)
δk (Vtq,i )δk (Vt−1
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
XX
X
q,i
q,k
δk (Vt−1
)δ0 (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
X
XX X
p̂8 =
q,i
q,k
)δ0 (Vt−1
)
δj (Vtq,i )δk (Vt−1
q=1 i=2 t=2 k=1 j6=i,k
k6=i
Nq Tq Nq
Q X
X
XX
q,i
δ0 (Vtq,i )δ0 (Vt−1
)
q,i
q,k
δk (Vt−1
)δ0 (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
q=1 i=2 t=2
Nq Tq
Q X
X
X
q,i
δ0 (Vt−1
)
Nq Tq Nq
Q X
X
XX
q=1 i=2 t=2
Nq Tq
Q X
X
XX
p̂2 =
q,i
q,k
δ0 (Vtq,i )δk (Vt−1
)δ0 (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
A PPENDIX B
VFOA L EARNING
Nq Tq
Q X
X
X
q,i
δk (Vt−1
)
q=1 i=2 t=2 k=Nq +1
The VFOA of i at t − 1 is an active target (1 ≤ k ≤
N, k 6= i):
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
=P (Vti
q,i
)
δj (Vtq,i )δk (Vt−1
q,i
δj (Vtq,i )δ0 (Vt−1
)
p̂9 =
q,i
q,k
δ0 (Vtq,i )δk (Vt−1
)δi (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
X
XX
q=1 i=2 t=2 j6=i
Nq Tq
Q X
X
X
q,i
q,k
δk (Vt−1
)δi (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
q,i
δ0 (Vt−1
)
q=1 i=2 t=2
Nq Tq Nq +Mq
Q X
X
X X
p̂3 =
Nq Tq Nq
Q X
X
XX
q,i
δ0 (Vtq,i )δk (Vt−1
)
p̂10 =
q=1 i=2 t=2 k=Nq +1
Nq Tq Nq +Mq
Q X
X
X X
q=1 i=2 t=2 k=Nq +1
q,i
δk (Vt−1
)
q,i
q,k
δk (Vtq,i )δk (Vt−1
)δi (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
X
XX
q=1 i=2 t=2 k=1
k6=i
q,i
q,k
δk (Vt−1
)δi (Vt−1
)
14
Nq Tq Nq
Q X
X
XX X
p̂11 =
q,i
q,k
δj (Vtq,i )δk (Vt−1
)δi (Vt−1
)
q=1 i=2 t=2 k=1 j6=i,k
k6=i
Nq Tq Nq
Q X
X
XX
q,i
q,k
δk (Vt−1
)δi (Vt−1
)
q=1 i=2 t=2 k=1
k6=i
Nq Tq Nq
Q X
X
XX X
p̂12 =
q,i
q,k
δ0 (Vtq,i )δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
q,i
q,k
δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
p̂13 =
q,i
q,k
δk (Vtq,i )δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
q,i
q,k
δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
p̂14 =
q,i
q,k
δl (Vtq,i )δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X
q,i
q,k
δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
Nq Tq Nq
Q X
X
XX X X
p̂15 =
q,i
q,k
δj (Vtq,i )δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k j6=i,k,l
k6=i
Nq Tq Nq
Q X
X
XX X
q,i
q,k
δk (Vt−1
)δl (Vt−1
)
q=1 i=2 t=2 k=1 l6=i,k
k6=i
ACKNOWLEDGMENTS
The authors would like to thank Vincent Drouard for his
valuable expertise in head pose estimation and tracking.
R EFERENCES
[1] E. G. Freedman and D. L. Sparks, “Eye-head coordination during headunrestrained gaze shifts in rhesus monkeys,” Journal of Neurophysiology, 1997.
[2] E. G. Freedman, “Coordination of the eyes and head during visual
orienting,” Experimental Brain Research, vol. 190, 2008.
[3] D. B. Jayagopi et al., “The vernissage corpus: A multimodal humanrobot-interaction dataset,” IDIAP, Tech. Rep., 2012.
[4] M. J. Marin-Jimenez, A. Zisserman, M. Eichner, and V. Ferrari, “Detecting people looking at each other in videos,” International Journal of
Computer Vision, vol. 106, 2014.
[5] L. H. Yu and M. Eizenman, “A new methodology for determining pointof-gaze in head-mounted eye tracking systems,” IEEE Transactions on
Biomedical Engineering, vol. 51, Oct 2004.
[6] T. Toyama, T. Kieninger, F. Shafait, and A. Dengel, “Gaze guided object
recognition using a head-mounted eye tracker,” in Proceedings of the
ETRA Symposium, 2012.
[7] A. K. A. Hong, J. Pelz, and J. Cockburn, “Lightweight, low-cost, sidemounted mobile eye tracking system,” in IEEE WNYIPW, 2012.
[8] K. Kurzhals, M. Hlawatsch, C. Seeger, and D. Weiskopf, “Visual
analytics for mobile eye tracking,” IEEE Transactions on Visualization
and Computer Graphics, vol. 23, no. 1, pp. 301–310, Jan 2017.
[9] P. Smith, M. Shah, and N. Da Vitoria Lobo, “Determining driver
visual attention with one camera,” IEEE Transactions on Intelligent
Transportation Systems, vol. 4, 2003.
[10] K. Krafka, A. Khosla, P. Kellnhofer, H. Kannan, S. Bhandarkar, W. Matusik, and A. Torralba, “Eye tracking for everyone,” in IEEE CVPR, June
2016.
[11] Y. Matsumoto, T. Ogasawara, and A. Zelinsky, “Behavior recognition
based on head pose and gaze direction measurement,” in IEEE IROS,
vol. 3, 2000.
[12] T. Ohno and N. Mukawa, “A free-head, simple calibration, gaze tracking
system that enables gaze-based interaction,” in Proceedings of the ETRA
Symposium. ACM, 2004.
[13] F. Lu, Y. Sugano, T. Okabe, and Y. Sato, “Adaptive linear regression
for appearance-based gaze estimation,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 36, Oct 2014.
[14] F. Lu, T. Okabe, Y. Sugano, and Y. Sato, “Learning gaze biases with
head motion for head pose-free gaze estimation,” Image and Vision
Computing, vol. 32, 2014.
[15] E. Murphy-Chutorian and M. Trivedi, “Head pose estimation in computer vision: A survey,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 31, 2009.
[16] X. Zabulis, T. Sarmis, and A. A. Argyros, “3D head pose estimation
from multiple distant views,” in BMVC, 2009.
[17] I. Chamveha, Y. Sugano, D. Sugimura, T. Siriteerakul, T. Okabe, Y. Sato,
and A. Sugimoto, “Head direction estimation from low resolution images
with scene adaptation,” Computer Vision and Image Understanding, vol.
117, 2013.
[18] A. K. Rajagopal, R. Subramanian, E. Ricci, R. L. Vieriu, O. Lanz, and
N. Sebe, “Exploring transfer learning approaches for head pose classification from multi-view surveillance images,” International Journal of
Computer Vision, vol. 109, 2014.
[19] Y. Yan, E. Ricci, R. Subramanian, G. Liu, O. Lanz, and N. Sebe, “A
multi-task learning framework for head pose estimation under target motion,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 38, 2016.
[20] Z. Qin and C. R. Shelton, “Social grouping for multi-target tracking and
head pose estimation in video,” IEEE Transactions on Pattern Analysis
and Machine Intelligence, vol. 38, 2016.
[21] J. S. Stahl, “Amplitude of human head movements associated with
horizontal saccades,” Experimental Brain Research, vol. 126, 1999.
[22] H. H. Goossens and A. Van Opstal, “Human eye-head coordination in
two dimensions under different sensorimotor conditions,” Experimental
Brain Research, vol. 114, 1997.
[23] R. Stiefelhagen and J. Zhu, “Head orientation and gaze direction in
meetings,” in Human Factors in Computing Systems, 2002.
[24] P. Lanillos, J. F. Ferreira, and J. Dias, “A bayesian hierarchy for robust
gaze estimation in human-robot interaction,” International Journal of
Approximate Reasoning, vol. 87, 05 2017.
[25] S. Asteriadis, K. Karpouzis, and S. Kollias, “Visual focus of attention
in non-calibrated environments using gaze estimation,” International
Journal of Computer Vision, vol. 107, 2014.
[26] S. Ba and J.-M. Odobez, “Recognizing visual focus of attention from
head pose in natural meetings,” IEEE Transactions on System Men and
Cybernetics. Part B., 2009.
[27] S. Sheikhi and J.-M. Odobez, “Recognizing the visual focus of attention for human robot interaction,” in Human Behavior Understanding
Workshop, 2012.
[28] Z. Yucel, A. A. Salah, C. Mericli, T. Mericli, R. Valenti, and T. Gevers,
“Joint attention by gaze interpolation and saliency,” IEEE Transactions
on System Men and Cybernetics. Part B., 2013.
[29] K. Otsuka, J. Yamato, and Y. Takemae, “Conversation scene analysis
with dynamic bayesian network based on visual head tracking,” in IEEE
ICME, 2006.
[30] S. Duffner and C. Garcia, “Visual focus of attention estimation with
unsupervised incremental learning,” IEEE Transactions on Circuits and
Systems for Video Technology, 2015.
[31] S. Sheikhi and J.-M. Odobez, “Combining dynamic head pose–gaze
mapping with the robot conversational state for attention recognition in
human–robot interactions,” Pattern Recognition Letters, vol. 66, 2015.
[32] B. Massé, S. Ba, and R. Horaud, “Simultaneous estimation of gaze direction and visual focus of attention for multi-person-to-robot interaction,”
in IEEE ICME, Seattle, WA, Jul. 2016.
15
[33] K. P. Murphy, “Switching Kalman filters,” UC Berkeley, Tech. Rep.,
1998.
[34] D. Simon, “Kalman filtering with state constraints: a survey of linear
and nonlinear algorithms,” Control Theory Applications, IET, 2010.
[35] C. M. Bishop, Pattern Recognition and Machine Learning. SpringerVerlag, 2006.
[36] P. Viola and M. Jones, “Rapid object detection using a boosted cascade
of simple features,” in IEEE CVPR, vol. 1, 2001.
[37] S.-H. Bae and K.-J. Yoon, “Robust online multi-object tracking based
on tracklet confidence and online discriminative appearance learning,”
in IEEE CVPR, 2014.
[38] V. Drouard, R. Horaud, A. Deleforge, S. Ba, and G. Evangelidis,
“Robust head-pose estimation based on partially-latent mixture of linear
regressions,” IEEE Transactions on Image Processing, vol. 26, Jan.
2017.
[39] A. Patron-Perez, M. Marszałek, A. Zisserman, and I. D. Reid, “High
five: Recognising human interactions in TV shows,” in British Machine
Vision Conference, 2010.
[40] X. Li, L. Girin, R. Horaud, and S. Gannot, “Multiple-speaker localization
based on direct-path features and likelihood maximization with spatial
sparsity regularization,” IEEE/ACM Transactions on Audio, Speech, and
Language Processing, vol. 25, no. 10, pp. 1997–2012, Oct 2017.
[41] I. Gebru, S. Ba, X. Li, and R. Horaud, “Audio-visual speaker diarization
based on spatiotemporal bayesian fusion,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2017.
Benoit Massé received the M.Eng. degree in applied
mathematics and computer science from ENSIMAG,
Institut National Polytechnique de Grenoble, France,
in 2013, and the M.Sc. degree in graphics, vision and
robotics from Université Joseph Fourier, Grenoble,
France, in 2014. Currently he is a PhD student in
the PERCEPTION team at INRIA Grenoble RhoneAlpes. His research interests include scene understanding, machine learning and computer vision,
with special emphasis on attention recognition for
human-robot interaction.
Silèye Ba received the M.Sc. (2000) in applied
mathematics and signal processing from University
of Dakar, Dakar, Senegal, and the M.Sc. (2002) in
mathematics, computer vision, and machine learning
from Ecole Normale Supérieure de Cachan, Paris,
France. From 2003 to 2009 he was a PhD student and then a post-doctoral researcher at IDIAP
Research Institute, Martigny, Switzerland, where he
worked on probabilistic models for object tracking
and human activity recognition. From 2009 to 2013,
he was a researcher at Telecom Bretagne, Brest,
France working on variational models for multi-modal geophysical data
processing. From 2013 to 2014 he worked at RN3D Innovation Lab, Marseille,
France, as a research engineer, where he used computer vision and machine
learning principles and methods to develop human-computer interaction
software tools. From 2014 to 2016 he was a researcher in the PERCEPTION
team at INRIA Grenoble Rhône-Alpes, working on machine learning and
computer vision models for human-robot interaction. Since May 2016 he is
a computer vision scientist with VideoStitch, Paris.
Radu Horaud received the B.Sc. degree in Electrical Engineering, the M.Sc. degree in Control Engineering, and the Ph.D. degree in Computer Science
from the Institut National Polytechnique de Grenoble, France. In 1982-1984 he was a post-doctoral
fellow with the Artificial Intelligence Center, SRI
International, Menlo Park, CA. Currently he holds a
position of director of research with INRIA Grenoble Rhône-Alpes, where he is the founder and head
of the PERCEPTION team. His research interests
include computer vision, machine learning, audio
signal processing, audiovisual analysis, and robotics. Radu Horaud and his
collaborators received numerous best paper awards. He is an area editor of
the Elsevier Computer Vision and Image Understanding, a member of the
advisory board of the Sage International Journal of Robotics Research, and
an associate editor of the Kluwer International Journal of Computer Vision.
He was program co-chair of IEEE ICCV’01 and of ACM ICMI’15. In 2013,
Radu Horaud was awarded an ERC Advanced Grant for his project Vision
and Hearing in Action (VHIA) and in 2017 he was awarded an ERC Proof
of Concept Grant for this project VHIALab.
| 1cs.CV
|
An Atomistic Fingerprint Algorithm for Learning Ab Initio
Molecular Force Fields
arXiv:1709.09235v3 [cs.CE] 14 Dec 2017
Yu-Hang Tang , Dongkun Zhang , and George Em Karniadakis
Division of Applied Mathematics, Brown University, Rhode Island, USA 02912
Abstract Molecular fingerprints, i.e. feature vectors describing atomistic neighborhood configurations,
is an important abstraction and a key ingredient for data-driven modeling of potential energy surface
and interatomic force. In this paper, we present the Density-Encoded Canonically Aligned Fingerprint
(DECAF) fingerprint algorithm, which is robust and efficient, for fitting per-atom scalar and vector
quantities. The fingerprint is essentially a continuous density field formed through the superimposition
of smoothing kernels centered on the atoms. Rotational invariance of the fingerprint is achieved by
aligning, for each fingerprint instance, the neighboring atoms onto a local canonical coordinate frame
computed from a kernel minisum optimization procedure. We show that this approach is superior over
PCA-based methods especially when the atomistic neighborhood is sparse and/or contains symmetry.
We propose that the ‘distance’ between the density fields be measured using a volume integral of their
pointwise difference. This can be efficiently computed using optimal quadrature rules, which only require
discrete sampling at a small number of grid points. We also experiment on the choice of weight functions
for constructing the density fields, and characterize their performance for fitting interatomic potentials.
The applicability of the fingerprint is demonstrated through a set of benchmark problems.
Keywords: active learning, Gaussian process regression, quantum mechanics, molecular dynamics,
next generation force fields
1
Introduction
Molecular Dynamics (MD) simulations have been widely used for studying atomistic systems, e.g. proteins
and catalysts, due to their ability to precisely capture transient events and to predict macroscopic properties
from microscopic details [1, 2]. In its most prevalent implementation, the trajectory of an atomistic system
is integrated in time according to the Newton’s law of motion using forces calculated as the negative
gradient of a Hamiltonian, whose functional form and parameters are collectively referred to as a force
field [3–5]. Traditionally, the pairwise and many-body terms that comprise a force field are derived
empirically by fitting to quantum mechanical calculations and experimental data.
Three properties directly relate to the applicability of a force field: accuracy, transferrability, and
complexity [6, 7]. Over the years, a large number of force fields have been developed, each carrying
1
Tang et al.
Molecular fingerprints
Figure 1: In the pipeline of machine learning-driven molecular computations, atomistic neighborhood
configurations are transformed into feature vectors, called fingerprints, and used to train non-linear
regression models.
a particular emphasis over these three properties. However, the combinatorial complexity of atomistic
systems can easily outpace force field development efforts, the difficulty of which explodes following the
curse of dimensionality [8]. A deceptively simple system that can demonstrate the situation is water,
a triatomic molecule with a well-characterized molecular structure. In fact, all common water models,
such as SPC-E, TIP3P, and TIP4P, have only succeeded in reproducing a small number of structural and
dynamical properties of water due to the difficulty in modeling strong intermolecular many-body effects
such as hydrogen bonding and polarization [9, 10].
In lieu of a force field, quantum mechanical (QM) calculations can be employed straightforwardly to
drive molecular dynamics simulations. The method achieves significantly better accuracy and transferrability by solving for the electronic structure of the system. However, the computational complexity of
QM methods is at least cubic in the number of electrons, and consequently the time and length scales
accessible by QM-driven molecular dynamics are severely constrained.
Assuming that there is smoothness in the potential energy surface of the atomistic system, one possible
strategy to accelerate QM-driven molecular dynamics is to use QM calculations on only a subset of the
time steps, and to interpolate for similar atomic configurations [11–14]. A schematic overview of the
process is given in Figure 1, which is enabled by the recent development of high-dimensional nonlinear
statistical learning and regression techniques such as Gaussian process regression [15] and artificial neural
networks [16].
This paper focuses on a particular aspect of the machine-learning-driven molecular computation
pipeline, i.e. fingerprint algorithms, whose importance arises naturally from the aforementioned regression protocol. A fingerprint is an encoding of an atomistic configuration that can facilitate regression tasks
such as similarity comparison across structures consisting of variable numbers of atoms and elements. As
has been pointed out previously [12], a good fingerprint should possess the following properties:
1. It can be encoded as a fixed-length vector so as to facilitate regression (particularly for artificial
neural networks).
2. It is complete, i.e. different atomistic neighborhood configurations lead to different fingerprints
2
Tang et al.
Molecular fingerprints
and vice versa, and the ‘distance’ between the fingerprints should be proportional to the intrinsic
difference between the atomistic neighborhood configurations.
3. It is continuous with regard to atomistic coordinates, and the change in fingerprint should be approximately proportional to the structural variation as characterized by, for example, some internal
coordinates.
4. It is invariant under permutation, rotation, and translation.
5. It is computationally feasible and straightforward to implement.
Before we proceed to the details of our work, we will first briefly review several fingerprints that are
closely related to our work, i.e. the Smooth Overlap of Atomic Positions (SOAP) kernel [12], the Coulomb
matrix [17], and the Graph Approximated Energy (GRAPE) kernel [18].
Smooth Overlap of Atomic Positions (SOAP): The SOAP kernel is built on the idea of representing
atomistic neighborhoods as smoothed density fields using Gaussian kernels each centered at a neighbor
atom. Similarity is measured as the inner product between density fields, while rotational invariance
is achieved by integrating over all possible 3D rotations, which can be performed analytically using the
power spectrum of the density field. In fact, our fingerprint algorithm is inspired by this idea of treating
atoms as smoothed density fields. However, we take a different approach to endorse the fingerprint with
rotational invariance, and use the Euclidean distance instead of inner product as a distance metric.
Coulomb Matrix: The practice of using graphs to represent atomistic neighbor configurations was first
implied by the Coulomb matrix, and later further formulated in the GRAPE kernel, where the diffusion
distance was proposed as a similarity measure between different local chemical environments [19]. The idea
is to construct an undirected, unlabeled graph G = ( V , E ) with atoms serving as the vertices and pairwise
interactions weighting the edges. For example, the Coulomb matrix can be treated as a physically-inspired
Laplacian matrix [20]
M=D−A
2.4 if I = J
0.5 ZI
DIJ =
0
if I , J
if I = J
0
AIJ =
Z Z
− kR I−RJ k if I , J
I
I
(1)
(2)
(3)
where the degree matrix D encodes a polynomial fit of atomic energies to the nuclear charge, while the
adjacency matrix A corresponds to the Coulombic interactions between all pairs of atoms. Due to the use
of only relative positions between atoms in the adjacency matrix, the Coulomb matrix is automatically
invariant under translation and rotation. However, the matrix itself is not invariant under permutation,
as swapping the order of two atoms will result in an exchange of the corresponding columns and the
rows. To address this, the sorted list of eigenvalues of the Coulomb matrix can be used instead as a feature
vector, while an `p norm can be used as a distance metric. In practice, due to the fact that the number of
3
Tang et al.
Molecular fingerprints
neighbor atoms may change, the shorter eigenvalue list is padded with zeros in a distance computation.
Graph Approximated Energy (GRAPE): The GRAPE kernel evaluates the simultaneous random walks
on the direct product of the two graphs representing two atomistic neighborhood configurations. Permutational invariance is achieved by choosing a uniform starting and stopping distribution across nodes
of both graphs. However, the cost of distance computation between two graphs scales as O(N2 ) with a
one-time per-graph diagonalization cost of O(N3 ).
In the sections below, we present our new fingerprint algorithm, namely the Density-Encoded Canonically Aligned Fingerprint (DECAF). The paper is organized as follows: in Section 2, we introduce a robust
algorithm that can determine canonical coordinate frames for obtaining symmetry-invariant projections;
in Section 3, we present numerical recipes to use smoothed atomistic density fields as a fingerprint for
molecular configuration; in Section 4, we demonstrate the capability of the fingerprint via examples
involving the regression of atomistic potential energy surfaces; in Section 5, we discuss the connection
between our algorithm and previously proposed ones; we conclude with a discussion in Section 6.
2
2.1
Localized Canonical Coordinate Frame for Rotationally Invariant Description of Atomistic Neighborhood
Kernel Minisum Approach
To improve model generalization while minimizing data redundancy, a fingerprint algorithm should be
able to recognize atomistic structures that differ only by a rigid-body transformation or a permutation
of atoms of the same element, and to extract feature vectors invariant under these transformations. As
summarized in Table 1, a variety of strategies have been successfully employed by common fingerprint
algorithms to achieve rotational invariance.
However, these approaches do not provide a means for the acquisition of vector-valued quantities in
a rotational invariant form. One approach is to only acquire and interpolate the potential energy, a scalar
quantity, and then take the derivative of the regression model. This approach, however, triggers the need
for methods to decompose the total energy among the atoms, which is a property of the entire system
rather than individual atoms [21].
Another approach proposed by Li et al. [13, 22] is to project vector quantities onto a potentially
overcomplete set of non-orthogonal basis vectors obtained from a weighted sum of the atomic coordinate
vectors:
"
#
X
kxi k pk
.
(4)
Vk =
xi exp −
Rc
i
However, the approach may suffer from robustness issues. For example, all of the Vk generated with
different pk will point in the same direction if the radial distance of the atoms are all equal. Further, the
4
Tang et al.
Molecular fingerprints
configuration with 4 atoms at (r cos ε, r sin ε), (0, r), (−r, 0), (0, −r) leads to
Vk = c · [(r cos ε, r sin ε) + (0, r) + (−r, 0) + (0, −r)]
= c · r · (1 − cos ε, sin ε).
(5)
(6)
Thus, if ε gets close to zero, Vk will always point toward either (0, 1) or (0, −1), even if the vector quantity
of interest may point in other directions.
Table 1: Comparison of strategies used by fingerprint algorithms to obtain feature vectors which are
invariant under translation, permutation, and rotation.
Invariance
Fingerprint
Coulomb matrix [17]
Behler [11]
SOAP [12]
GRAPE [18]
Translation
Permutation
Rotation
relative distance
relative distance
relative distance
relative distance
sorting eigenvalues
summation
summation
uniform distribution
all vs. all graph
ignoring angular information
integrating over all rotations
uniform distribution
Here, we present a robust kernel PCA-inspired algorithm for the explicit determination of a canonical
coordinate frame, within which the projection of the atomistic neighborhood is invariant under rigidbody rotation. Furthermore, the canonical coordinate frame can be directly used to capture vector-valued
quantities in a rotational-invariant form. Given N atoms with position x1 , . . . , xN ∈ Rd , we first formulate
the Lp PCA algorithm as an optimization problem where we seek a unit vector w∗ that maximizes the sum
of the projections:
w = argmaxkwk=1
∗
= argmaxkwk=1
N
X
i=1
N
X
wT xi
p
(7)
p
|ri |p wT ei ,
(8)
i=1
where ri = kxi k is the distance from the origin to atom i, ei = xi /ri is the unit vector pointing toward atom i,
respectively. The optimization process can only uniquely determine the orientation of a projection vector
up to a line, because wT e ≡ −wT e . As a consequence, further heuristics are needed to identify a specific
direction for the PCA vectors.
To overcome this difficulty, we generalize the |ri |p term into a weight function g(r) of radial distance
p
and the wT ei term into a bivariate kernel function κ(w, e) between two vectors. We then attempt to seek
a unit vector w∗ that minimizes the kernel summation:
w = argminkwk=1
∗
N
X
g(ri ) κ(w, ei ).
(9)
i=1
In particular, we have found a square angle (SA) kernel and an exponentiated cosine (EC) kernel that
5
Tang et al.
Molecular fingerprints
perform well in practice:
1
arccos2 (wT e),
2
κEC (w, e) exp −wT e .
κSA (w, e)
(10)
(11)
As shown in Figure 2, both kernels are minimal when w and e are parallel, and monotonically reach
maximum when w and e are antiparallel. Intuitively, optimizing the minisum objective function generated
by the SA kernel will yield a vector that, loosely speaking, bisects the sector occupied by the atoms. The
EC kernel exhibits very similar behavior but leads to a smoother objective function. As shown in Figure 2,
this allows for the determination of a projection vector without ambiguity, even if the atom configuration
contains perfect symmetry.
Square Angle
Exponentiated Cosine
PCA
π
0
0
-π/2
π
θ
Aw
0
-π
π
0
0
-π/2
π/2
π
-π
θ
0
-π/2
π
θ
π/2
π
-π
π
Aw
-π
π/2
Aw
Aw
π/2
0
π
0
-π/2
θ
Figure 2: Shown here is an illustration of the minisum algorithm that determines projection vectors for
rotational invariant description of atomistic neighbor configurations. Black dots represent atoms which
all carry equal importance. The vectors that point from the origin to the atoms are used as the input
to bivariate kernels to compute the minisum objective function, which are drawn in solid lines. For
reference, the negated values of the PCA objective function are drawn in dashed lines. Projection vectors
are obtained by finding the unit vector w∗ that minimizes the objective function.
A major advantage of the kernel minisum approach versus Lp norm-based PCA, lies in its 1) robustness
in the presence of structural symmetry; and 2) continuity of the resulting principal axes with respect to
angular movement of the input data. As shown in Figure 3, kernel minisum is particularly suitable for
atomistic systems where strong symmetries are common and the continuity against angular movement
is desired. The minisum framework can also be used with other customized kernels to suit for the
characteristics of specific application scenarios.
6
Tang et al.
1.0
Molecular fingerprints
(A) Rotational invariance
(B) Angular continuity
(C) Cluster angular continuity
0.8
0.6
0.4
0.2
0.0
0.0
PCA − L2
0.2
PCA − L1
0.4
Kernel Minisum
0.6
0.8
1.0
Figure 3: A comparison of the orthogonal bases obtained using principal components analysis (PCA) and
kernel minisum with the square-angle (MSA) kernel. (A) The PCA algorithm, used in conjunction with
the L2 norm, fails to extract a principal axis that rotates with a system that exhibits planar C4 symmetry.
Both MSA and PCA with the L1 norm can accommodate this scenario. (B) Both L1 and L2 principal axes
change orientation abruptly when the system undergoes a slight angular motion. In contrast, the MSA
output is continuous with regard to this movement as it always bisects the angle formed by the atoms
and the origin. (C) Loosely speaking, the MSA axis points at the majority direction of the atoms if only a
single cluster is present within the cutoff distance, but bisects the angle between two atom clusters. This
is different from PCA-based results and should deliver robust rotational invariance as well as continuity.
Arrows are drawn at different lengths to improve visual clarity in situations of overlapping. They are to
be understood as unit vectors. De-trending is not performed because the radial distances of the atoms
carry physically significant information.
2.2
Solving the Kernel Minisum Optimization Problems
The optimization problem can be solved very efficiently using a gradient descent algorithm as detailed
below.
Square Angle: The objective function of the minisum problem using the square angle (SA) kernel is
N
ASA (w)
1X
g(ri ) arccos2 (wT ei ).
2
i=1
7
(12)
Tang et al.
Molecular fingerprints
The gradient of ASA with respect to w is
∇w ASA =
N
X
i=1
Te )
i
arccos(w
Note that √
T
1−(w ei )2
arccos(wT ei )
ei .
−g(ri ) p
1 − (wT ei )2
(13)
is singular when w k ei . This can be treated numerically by replacing the removable
Te )
i
arccos(w
singularities at wT ei = 1 with the left-limit limwT ei →1− √
T
1−(w ei )2
= 1, while truncating the gradient at a
finite threshold near the poles at wT ei = −1.
A local minimum can be iteratively searched for with gradient descent while renormalizing w after
each iteration. Moreover, due to the locally quadratic nature of the objective function, we have found that
the Barzilai-Borwein algorithm [23] can significantly accelerate the convergence at a minimal cost. The
algorithm is presented in Alg. 1.
Algorithm 1 Gradient descent for solving the square angle minisum problem.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
function MinSquareAngle(E = [e1 , e2 , . . . , eN ], G = [g1 , g2 , . . . , gN ], w0 )
w ← w0
repeat
Compute gradient ∇w using Eq. 13.
T
Obtain tangential component of gradient ∇⊥
w ← (I − ww )∇w
if at step 0 then
α ← 0.01
. Small initial step size for bootstrapping
else
⊥
(w−w−1 )T (∇⊥
w −∇w−1 )
. Adaptive subsequent steps by Barzilai-Borwein
α←
⊥
⊥
k∇ −∇
k2
w
w−1
end if
⊥
Save w as w−1 , save ∇⊥
w as ∇w−1
⊥
Update w ← w − α ∇w and normalize to unit length
until kw − w−1 k < ε
return w
end function
Exponentiated Cosine:
(EC) kernel is:
The objective function of the minisum problem using the exponentiated cosine
AEC (w)
N
X
g(ri ) exp −wT ei .
(14)
−g(ri ) exp −wT ei ei .
(15)
i=1
The gradient of AEC with respect to w is
∇w AEC =
N
X
i=1
8
Tang et al.
Molecular fingerprints
The gradient contains no singularity. However, it is not always locally quadratic or convex. This can
cause the Barzilai-Borwein algorithm to generate negative step sizes and consequently divert the search
towards a maximum. Luckily, this can be easily overcome by always using the absolute value of the step
size generated by the Barzilai-Borwein algorithm. Such enforcement prevents the minimization algorithm
from going uphill. The complete algorithm is given in Alg. 2.
Algorithm 2 Gradient descent for solving the exponentiated cosine minisum problem.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
function MinExpCosine(E = [e1 , e2 , . . . , eN ], G = [g1 , g2 , . . . , gN ], w0 )
w ← w0
repeat
Compute gradient ∇w using Eq. 15.
T
Obtain tangential component of gradient ∇⊥
w ← (I − ww )∇w
if at step 0 then
α ← 0.01
. Small initial step size for bootstrapping
else
⊥
(w−w−1 )T (∇⊥
w −∇w−1 )
α←
. Adaptive subsequent steps by Barzilai-Borwein
k∇⊥ −∇⊥ k2
w
w−1
end if
⊥
Save w as w−1 , save ∇⊥
w as ∇w−1
⊥
Update w ← w − α ∇w and normalize to unit length
until kw − w−1 k < ε
return w
end function
As shown in Table 2, both Alg. 1 and Alg. 2 converge quickly and consistently across a wide range
of representative point configurations commonly found in molecular systems. However, the gradient
descent method can only find local optima. Thus, multiple trials should be performed using different
initial guesses to ensure that a global minimum can be located.
2.3
Complete Set of Orthogonal Projection Vectors as A Canonical Coordinate Frame
In 3D, a complete set of orthogonal bases can be found greedily using the protocol as described in Alg. 3.
Specifically, we use the globally optimal solution of the minisum optimization problem as the first basis
bα , and the constrained optimal solution in a plane orthogonal to bα as the second basis bβ . Special care
must be taken for determining the third basis bγ , as the only degree of freedom now is its sign due to the
orthogonality constraint. The straightforward approach of choosing the direction that gives the smaller
objective function value may fail, for example, when the system contains improper rotational symmetry. In
that case, bα and bβ are interchangeable and both perpendicular to the rotation-reflection axis. As a result,
the two candidates of bγ will both align with the rotation-reflection axis and are thus indistinguishable
by kernel minisum. However, the projection of the atoms into the two seemingly equivalent coordinate
frames are not identical, but rather mirror images of each other. Fortunately, this can be addressed by
choosing the direction of the half-space, as created by the plane bα -bα , that yields the smaller kernel
objective function between the bisector of bα and bβ versus the points lies in that half-space. This rule can
also handle general situations with/without symmetry.
9
Tang et al.
Molecular fingerprints
Table 2: Listed here is the number of iterations and initial guesses used by the gradient descent algorithm
to find an local optimum of the kernel minisum problems. The numbers are averaged over 500 repetitions,
and the convergence criterion is 10−14 . In cases where the iterative algorithm does not converge within 64
iterations, the optimization will be restarted with a new guess.
Kernel
Square Angle
Exponentiated Cosine
Configuration
Itrs./Guess
Guesses
Itrs./Guess
Guesses
Single point
Two points, angle < π/2
Two points, angle ≥ π/2
Two points, angle = π
Planar C3
Planar C4
Tetrahedra
Octahedra
Improper S4
Improper S6
2D random 10 points
3D random 50 points
8.8
7.1
6.1
6.0
6.2
5.6
10.6
11.7
17.9
14.7
7.2
16.9
1
1
1
1
1
1
1
1
1
1
1.1
1.2
7.9
7.7
6.3
5.6
6.3
6.9
7.7
9.4
23.1
18.0
8.9
14.4
1
1
1
1
1
1
1
1
1
1
1
1
Algorithm 3 The procedure for determining a canonical coordinate frame using kernel minisum.
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
function GetCanonicalProjection3D(E = [e1 , e2 , . . . , eN ], G = [g1 , g2 , . . . , gN ] )
bα ← global minimum of Minisum(E, G) using multiple runs of Alg. 1 or Alg. 2
if E contains only 1 point then
construct arbitrary bβ ⊥ bα
bγ ← bα × bβ
else
bβ ← global minimum of Minisum(E, G) using multiple runs of Alg. 1 or Alg. 2, subjecting to
the constraint that the √
probe vector and the gradient are all ⊥ bα
d←
P (bα + bβ )/ 2
P
if ∀ei ; eT (bα ×bβ )≥0 gi κ(d, ei ) ≤ ∀ei ; eT (bα ×bβ )<0 gi κ(d, ei ) then
i
i
bγ ← bα × bβ
else
bγ ← −bα × bβ
end if
end if
return bα , bβ , bγ
end function
10
Tang et al.
Molecular fingerprints
It is difficult to prove global uniqueness of the kernel minisum solution given the non-convex nature
of the exponentiated cosine and square angle kernels. In fact, it seems that the only kernel that allows
analytical proof of solution uniqueness is κCOM (w, e) −wT e, whose solution simply corresponds to the
weighted center of mass of the neighbor atoms. Unfortunately, this simple kernel is not robust against
reflectional and rotational symmetry. Luckily, the rare cases where two global optimal solutions do
coexist can be safely captured by the repeated searching procedure starting from different seeds. Thus, a
fingerprint can be extracted using each of the resulting coordidate frame. This may mildly increase the
size of the training set, which could even be advantagenous when training data is scarce.
3
3.1
Density-Encoded Canonically Aligned Fingerprint
Density Field and Approximation of Volume Integral
(A)
(B)
Atoms
Density
ρ1
1/2
d(ρ1 , ρ2 ) = [∫ (ρ1 − ρ2 )2 dV]
ρ2
Figure 4: (A) Two 1D density profiles, ρ1 and ρ2 , are generated from two different atomistic configurations
using atom-centered smoothing kernel functions. The ‘distance’ between them is measured as the L2 norm
of their pointwise difference, which corresponds to the orange area in the middle plot. (B) Shown here is
a 2D density field using smoothing kernels whose widths depend on the distances of the atoms from the
origin. Darker shades indicate higher density.
The local density field ρs (r) around a point s is formulated as a superimposition of smoothing kernel
functions each centered at a neighbor atom i = 1, 2, . . . , N with relative displacement xi with regard to s
and within a cutoff distance Rc :
ρs (r) =
N
X
Wc (xi − s) Wd (xi − r)
(16)
i,kxi −sk<Rc
This density field, as has been pointed out previously [12], may be used as a fingerprint of the local
atomistic environment. Here, we assume that the smoothing kernel Wd (r) takes the form of a stationary
h
i
Gaussian σ−1 exp − 21 krk2 /σ2 . We also assume that the density scaling function Wc (r), which ensures the
continuity of the density field when atoms enter or exit the cutoff distance, is a bell-shaped function with
compact support. Further discussion on both Wc (r) and Wd (r) can be found in Section 3.2 and Section 3.3,
respectively.
11
Tang et al.
Molecular fingerprints
To achieve rotational invariance, we project the atom coordinates into the canonical coordinate frame
R [bα , bβ , bγ ] as determined by the kernel minisum algorithm, when generating the density field:
N
X
ρs (r) =
Wc (RT xi − s) Wd (RT xi − r).
(17)
i,kxi −sk<Rc
Depending on the specific application, s may not necessarily overlap with any of the xi . Scalar properties
can be acquired directly from the target atom, while vector-valued properties, such as force, can be
acquired and interpolated in the local orthogonal coordinates as ỹ = RT y.
We define the distance between two density fields ρi and ρ j as a weighted volume integral of their
pointwise difference:
∆ρs1 s2 (r) ρs1 (r) − ρs2 (r),
!1/2
Z
2
d(ρs1 , ρs2 )
w(r) ∆ρs1 s2 (r) dV (r)
.
(18)
(19)
R3
The weight function w(r) provides additional flexibility for emphasizing particular regions of the atomistic
neighborhood. It could play an important role for fitting properties with steep gradients, e.g. the repulsive
part of the Lennard-Jones potential.
We now introduce an optimal quadrature rule to approximate the integral in Eq. 19 in a computationally
R
P
tractable manner. A quadrature rule is a numerical recipe in the form f (x) = N
i=0 wi f (xi ), which
numerically approximates a definite integral using only discrete evaluations of the integrand. To determine
the quadrature nodes and weights, we decompose the volume integral in Eq. 19 into a surface integral
over spherical shells and a 1D integral along the radial direction:
Z
Z
2
R3
w(r) ∆ρs1 s2 (r) dV (r) =
∞
r=0
Z
2π
Z
ϕ=0
π
θ=0
2
!
w(r, θ, ϕ) ∆ρs1 s2 (r, θ, ϕ) sin θ dθ dϕ r2 dr
(20)
The surface integral can be optimally approximated using the Lebedev quadrature rule [24]:
Z
∆Ps1 s2 (r)
2π
ϕ=0
Z
π
θ=0
2
w(r, θ, ϕ) ∆ρs1 s2 (r, θ, ϕ) sin θ dθ dϕ ≈ 4π
N
a (r)
X
2
w(r · qm ) βm ∆ρs1 s2 (r · qm ) ,
(21)
m=1
where βm , qm (xm , ym , zm ), and Na are the weights, positional unit vectors, and number of the Lebedev
nodes, respectively. The radial integral fits well into the generalized Laguerre-Gauss quadrature formula
with weight function r2 e−r [25]:
∞
Z
0
∆Ps1 s2 (r) r2 dr ≈
Nr
X
αn ern ∆Ps1 s2 (rn ),
(22)
n=1
where αn , rn , and Nr are the weights, coordinates, and number of the Laguerre nodes, respectively.
Combining Eq. 20–22, a composite quadrature rule can be generated consisting of several spherical layers
12
Tang et al.
Molecular fingerprints
of nodes. As shown in Figure 5, the radial coordinates of the quadrature nodes are determined by the
Laguerre quadrature nodes, while the angular positions are determined by the Lebedev quadrature nodes,
respectively. This composite quadrature formula translates the 3D volume integral into a summation over
discrete grid points:
Z
2
R3
w(r) ∆ρs1 s2 (r) dV (r) ≈ 4π
Nr N
a (n)
X
X
2
αn βm w(rn · qm ) ern ∆ρs1 s2 (rn · qm ) .
(23)
n=1 m=1
Using the right hand side of Eq. 23 to replace the integral in Eq. 19, and use the multi-index notation
k = (n, m); 1 ≤ n ≤ Nr , 1 ≤ m ≤ Na (n) to enumerate over the quadrature nodes located at rk = rn · qm with
weights wk = 4π αn βm w(rn · qm ) ern , we obtain the final discretized formula for computing the distance
between the fingerprints:
1/2
N
X
2
d(ρs1 , ρs2 ) ≈
wk ρs1 (rk ) − ρs2 (rk ) .
(24)
k=1
For quick reference, we tabulated in Appendix the values for rn and αn in the Laguerre quadrature of up
to 6 points, and the values for qm , βm in the Lebedev quadrature of up to 50 points.
In addition, the quadrature nodes could be radially scaled such that the outer most nodes lie at a
radius R∗ within some cutoff distance Rc . This allows us to fit a Laguerre quadrature of any order within
an arbitrary cutoff distance. The scaled quadrature rule is given by:
τ = R∗ / max (rn ) ,
n
1/2
N
X
2
τ3 wk ρs1 (τ rk ) − ρs2 (τ rk ) .
d∗ (ρs1 , ρs2 ) ≈
(25)
(26)
k=1
Since the scaling is simply constant among all nodes, it can be safely omitted in many regression tasks
where only the relative distance between the fingerprints are of significance.
3.2
Radial Weight Functions
In this section, we examine two radial weight functions that can be used to fine-tune the density field
fingerprint: the density scaling function Wc (r) as appears in Eq. 17 and the weight of integral w(r) as
appears in Eq. 19.
Driven by the interest of reducing computational cost, we would like to use a cutoff distance to select
atoms involved in constructing the density field. However, it is important to ensure that atoms will enter
and exit the neighborhood smoothly. This naturally requests that the contribution of an atom to be zero
outside of the cutoff, and to increase continuously and smoothly when the atom approaching entrance.
Correspondingly, Wc (r) should: 1) become unity at the origin; 2) smoothly approach zero at the cutoff
distance; and 3) be twice-differentiable to ensure the differentiability of regression models based on the
13
Tang et al.
Molecular fingerprints
(A) Laguerre
(B) Lebedev
(C) Fingerprint by Quadrature
5
4
3
a1
a2
a3
bk
2
1
0
1
Shell 0
Shell 1
Shell 2
Figure 5: (A) Laguerre quadrature nodes with order 1 to 5, normalized by the reciprocal of the largest node
onto the unit interval. (B) The a1 , a2 , a3 , and bk class of Lebedev grid points on a unit sphere. (C) The DECAF
molecular fingerprint essentially comprises of a grid of quadrature nodes that optimally samples the
density field induced by the neighbor atoms. Shown here is an example of one such composite quadrature
grid which combines a 3-node Laguerre quadrature rule with three layers of Lebedev quadrature nodes
of 3rd, 5th, and 7th order, respectively.
fingerprint. Candidates satisfying the above conditions include, for example, a tent-like kernel
Wc (r) = (1 − krk/Rc )t , t > 2
(27)
and a bell-shaped polynomial kernel with compact support
Wc (r) =
−b (1 − krk/Rc )a + a (1 − krk/Rc )b
, a>b>2
a−b
(28)
as detailed in Appendix.
The approximation of the radial integral using a Laguerre quadrature requires that the integrand,
i.e. the pointwise difference between the atomistic density fields, decays sufficiently fast beyond the
outermost quadrature nodes in order to achieve acceptable convergence. In addition, the steeply repulsive
yet flat attractive interatomic short-range interactions prompt that the sensitivity of fingerprint be adjusted
correspondingly in order to avoid numerical difficulties in training a regression model. The weight of
the integral, w(r), provides a convenient means for achieving the purpose. Different from Wc (r), w(r)
R
should instead satisfy the following conditions: 1) is normalized such that w(r)dV(r) = 1; 2) decays
sufficiently fast, but not necessarily to 0, beyond the outer most quadrature node; and 3) be sufficiently
smooth beyond the outermost quadrature node. Candidates for w(r) includes the tent-like kernel and the
bell-shaped kernel for Wc (r), albeit with a different normalization factor. The Laplacian kernel
w(r) = exp(− |r| /l)/(8πl3 ), l > 0
(29)
with a properly sized length scale l also appears to be a candidate due to its similarity with e−r part of the
14
Tang et al.
Molecular fingerprints
Bell
Weight of Integral w(r)
Unity
Fingerprint Distance
1
0
Density Scaling Function Wc (r)
Tent
(A)
6.5
Rc
1.0
Unity
(B)
6.5
Rc
1.0
1.0
6.5
Rc
(C)
Rc 6.5
1.0
6.5
Rc
1.0
(D)
Rc 6.5
1.0
Rc 6.5
1.0
1.0
Rc 6.5
Figure 6: Shown here are examples of the distance matrices between fingerprints sampled from a biatomic
system. As manifested by the difference between (A) against (B), a bell-shaped weight of integral helps
to emphasize the near field. Meanwhile, the obvious discontinuities in the second row of matrices
demonstrate the importance of the density scaling function when the fingerprint algorithm uses only
atoms within a finite cutoff distance.
weight function of the Laguerre quadrature. Note that the constant kernel
w(r) = 3/(4πR3c )
(30)
may also be a choice as long as the density field already decays fast enough due to the density scaling
function Wc (r).
In Figure 6, we demonstrate the effect of the density scaling function and the weight of integral on
the distance matrices between fingerprint obtained from a pair of atoms. The comparison between panel
A and B shows that a bell-shaped integration weight allows the distance between fingerprints to change
15
Tang et al.
Molecular fingerprints
more rapidly when the atoms are closer but more slowly when the atoms are farther apart. The visible
discontinuity in the second row clearly demonstrates the importance of the damping function when only
atoms within a finite cutoff distance are used to compute the fingerprint.
We further examine the impact of the weight functions on the performance of Gaussian process
regression (GPR) using the fingerprint of the interatomic force of a minimal system containing two
nitrogen atoms. Despite the simplicity of the system, this case is of fundamental importance because of
its ubiquity, and because the fast-growing repulsive regime of the Lennard-Jones potential could cause
difficulty as a small change in the system configuration can trigger large changes in the regression target
function. In Figure 7, we compare the performance among the combination of four weights of the integral
and two density scaling functions. The initial training set consists of two samples collected at rN-N = 1.0
and 6.0. The regression is then refined using a greedy strategy that consecutively learns the point with the
largest posterior variance until the largest uncertainty, defined as twice the posterior standard deviation,
is less than 0.1 eV/Å. The active learning scheme is able to delivery a GPR model, for each and every
combination of the weight functions, that closely fits the target function. However, the numbers of
refinement iterations and the achieved accuracy do vary. Therefore, it is important to evaluate and choose
the weight functions in the context of specific application scenarios.
3.3
Quadrature Resolution and Density Kernel
Despite the formal convergence of the composite quadrature in DECAF, a cost of O(NM) distance calculations and kernel evaluations are needed to sample a density field generated by N atoms using M
quadrature nodes. A less prominent cost is associated with the L2 distance calculation, which comes at a
cost of O(M) floating point operations. Thus, in practice it is often desirable to use as few nodes as possible
to capture only information of the density field within a certain band limit [26]. Accordingly, the integral
cutoff Rc , the number of quadrature nodes, and the width of the density kernel need to be tuned to obtain
an optimal balance between resolution and computational speed.
When designing the composite quadrature rule, we chose the Laguerre quadrature for the radial
direction because its nodes are denser near the origin but sparser farther away. This is consistent with our
physical intuition that the near field generally has a stronger influence than the far field in an atomistic
neighborhood. For example, the Van de Waals potential grows rapidly when atoms are in direct contact,
but flattens out of the first coordinate shell. Accordingly, it may be possible for us to use sparser outerlayer grids to reduce the total number of quadrature nodes, while still keeping enough nodes in the inner
layers to maintain the sensitivity of the quadrature toward close neighbors. Cooperatively, we can also
use non-stationary Gaussian density kernels whose width dependent on the distance from the atom to
the origin. In this way, the sparser nodes should still sufficiently sample the smoother far field. Wider
kernels at remote atoms also reduce the total difference between the far fields of two fingerprints in a
statistical sense. Thus, the contribution of the far field in the integral can be effectively tuned even though
the weights on the quadrature nodes remain the same.
In Figure 8, we demonstrate how a variable-resolution quadrature can be combined with a widening
smoothing density kernel to simultaneously reduce the computational complexity and preserve the quality
16
Tang et al.
Molecular fingerprints
GPR
Truth
Scaling of Density
seed
refine
30
Bell
Tent
17 points, RMSE 0.05 eV/Å
11 points, RMSE 0.15 eV/Å
f (eV/Å)
Bell
20
10
10
0
5
7
2
−10
14
1
3
11
13
0
8
4
12
0
6
9
4
5
1
−20
2
16 points, RMSE 0.06 eV/Å
30
6
7
8
3
10 points, RMSE 0.21 eV/Å
f (eV/Å)
Tent
10
10
0
5
7
2
−10
13
3
−20
9
1
0
8
4
12
11
3
6
7
0
6
4
5
1
2
20 points, RMSE 0.04 eV/Å
30
11 points, RMSE 0.14 eV/Å
20
f (eV/Å)
Laplace
Weight of Integral
20
10
11
0
5
7
−10
12
3 14 5 17
−20
1
9
6
16
2
13
8
4
15
10
0
0
4
1
2
21 points, RMSE 0.03 eV/Å
30
6
7
8
3
12 points, RMSE 0.11 eV/Å
f (eV/Å)
Unity
20
10
5
11
0
−10
6
16
−20
1
3 14
7
18
1
10
4
15 2
9
5
12
0
8
17
13
4
3
7
6
8
1
2
9
0
4
3
N-N distance (Å)
5
6
1
2
2
4
3
N-N distance (Å)
5
6
Figure 7: Gaussian process regression of the force between two nitrogen atoms as a function of interatomic
distance using different combinations of radial weight functions. Inset figures are plots of the regression
function using distances from the feature space.
17
Tang et al.
Molecular fingerprints
of the fingerprint. In column A, a dense grid is used to sample density fields generated by a wide
smoothing length. By examining the distance matrices of fingerprints sampled during bond stretching
and angular stretching movements, we note that the radial similarity decreases monotonically while the
angular similarity changes nearly constantly. In column B, the number of quadrature nodes is kept the
same, but the smoothing length is reduced as an attempt to increase fingerprint sensitivity. Better response
in the near field of the radial direction is obtained, but the linearity in the far field in the angular direction
is compromised. In column C, the fingerprint performs even worse due to the combination of a sparser
quadrature grid and a small smoothing length. In column D, the performance recovered because we let
the smoothing length parameter σ of the Gaussian density kernels Wd (r) depend on the distance from the
origin to each atom, and simultaneously adjust the quadrature node density according to this pattern.
(A)
50 nodes × 4 layers
σ = 1.4
(B)
50 nodes × 4 layers
σ = 1.0
(C)
26 nodes × 4 layers
σ = 1.0
(D)
6, 14, 26, 38 nodes
σ = 0.7 + 0.3r
6.0
Radial
1.0
1.0
6.0 1.0
6.0 1.0
6.0 1.0
6.0
π
π/6
π
Angular
π/6
r=1.5
r=3.0
r=1.5
r=3.0
r=1.5
r=3.0
r=1.5
r=3.0
Fingerprint Distance
0
1
Figure 8: A comparison of fingerprint distance matrices corresponding to bond stretching and angular
stretching movements. (A) Dense grid + large smoothing length: radial similarity decreases monotonically
while angular similarity changes nearly constantly. (B) Dense grid + smaller smoothing length: better
fingerprint sensitivity in the near field for bond stretching, compromised linearity in the far field for
angular stretching. (C) Sparser grid + smaller smoothing length: compromised far-field performance for
both bond and angular movements. (D) Variable-resolution grid + radially dependent smoothing length:
good resolution and linearity in both near and far fields.
18
Tang et al.
4
4.1
Molecular fingerprints
Demonstration
Method
Regression tasks throughout this work are performed using Gaussian process regression (GPR), a nonlinear
kernel method that treats training data points as discrete observations from an instantiation of a Gaussian
process. Predictions are made using the posterior mean and variance of the joint distribution between
the test data and the training data. One particular interesting property about Gaussian process is that
the posterior variance may be interpreted as a measure of prediction uncertainty, which can be exploited
to design active learning algorithms for sampling expensive functions. The actual computation used our
own software implementation which was made publicly available on Zenodo [27]. We use the square
exponential covariance kernel to compute the covariance, i.e. correlation, between the samples:
1
kSE (x, x0 ) = σ2 exp[− d(x, x0 )2 /l2 ]
2
where x and x0 are DECAF fingerprints, and d(x, x0 ) the distance between norm as computed by Eq. 24 or
Eq. 26. The kernel is stationary, meaning that the covariance depends only on the relative distance between
two samples but not their absolute position. The training process searches for the hyperparameters, i.e.
the output variance σ and the length scale l, that maximizes the likelihood of the training data. A detailed
tutorial on GPR can be found in Ref. [15]. An illustration on the complete workflow of using the density
field fingerprint to perform regression tasks is given in Figure 9.
4.2
Potential Energy Surface
First, we attempt to fit the potential energy surface of a protonated water dimer system, in a head-to-head
configuration, as a function of the oxygen-oxygen distance rO-O and the dihedral angle ϕ between the two
planes each formed by a water molecule. As shown in Figure 10A, the system contains an improperly
rotational symmetry, which we wish to capture with the kernel minisum algorithm. A GPR model was
seeded with 8 training points corresponding to the combinations of rO-O = 2.2, 2.4, 2.6, 2.8 and ϕ = 0, π/2.
Subsequently, an active learning protocol was used to greedily absorb points with the highest uncertainty
into the training set. Despite that we restricted all training data to be within the subdomain ϕ <= π/2,
as shown by Figure 10B and 10C, we are able to accurately reproduce the target function over the entire
parameter space after a few active learning steps.
The DECAF fingerprint used here is constructed with 3 spherical layers within a cutoff distance
Rc of 6.0 Å, each consisting of 14, 26, and 38 Lebedev quadrature nodes, respectively. The weight of
integral was chosen as w(r) = W 6,4 (1 − r/Rc ), where W 6,4 is the bell-shaped polynomial as defined in
Appendix Eq. 35. The density scaling function Wc (r) = (1 − krk/Rc )3 , where r is the vector from the
atom to the fingerprint center, is the tent-like kernel as defined in Eq. 27 with t = 3. The density
kernel that sits on the oxygen atoms assumes the form of a non-stationary Gaussian as discussed in
h
i
Section 3.3: WdO (r, r0 ) = σO (r)−1 exp − 21 kr0 k2 /σO (r)2 , σO (r) = 1.5 + 0.25krk with r being the vector from
the atom to the fingerprint center and r0 being the vector from the atom to the quadrature node. The
19
Tang et al.
Molecular fingerprints
Regression
Configuration
sample
prediction uncertainty
Kernel Minisum
Aligned Configuration
prediction
Neural Network
Gaussian Process
Distance/Similarity Matrix
Distance Computation
Density Field Generation
Density Field
Quadrature
Fingerprint (Feature Vector)
Figure 9: Shown here is the workflow of regression using the density field fingerprint. The key stages, i.e.
kernel minisum optimization, density field generation, and quadrature-based distance computation, are
covered in detail in Sec. 2 and Sec. 3. The fingerprints can also be readily used to train artificial neural
networks.
20
Tang et al.
Molecular fingerprints
density kernel for the hydrogen atoms has a different weight and width to ensure discriminability:
i
h
WdH (r, r0 ) = 0.75 σH (r)−1 exp − 12 kr0 k2 /σH (r)2 , σH (r) = 0.9 + 0.15krk.
φ
rO-O
Oxygen
Hydrogen
Figure 10: Gaussian process regression is carried out to fit the potential energy surface of a protonated
water dimer as a function of two internal variables, i.e. the oxygen-oxygen distance rO-O and the dihedral
angle ϕ between the plane determined by the two water molecules. This system contains an improperly
rotational symmetry, which can be correctly recognized by the kernel minisum-based algorithm given in
Alg. 3. Only a quarter of the domain was used to train the GP, yet the model can accurately predict the
energy of the entire parameter space thanks to symmetry detection.
4.3
Geometry Optimization and Vibrational Analysis
Next, we demonstrate the usability of fingerprint for fitting vector-valued quantities by performing
geometry optimization and vibrational analysis on a single water molecule. The process involves the
simultaneous regression of: 1) energy, a molecular scalar quantity; 2) force, a per-atom vector quantity;
and 3) dipole, a molecular vector quantity. Correspondingly, we performed GPR of energy and dipole
using fingerprints extracted from the center of mass of the molecule, and GPR of force using fingerprints
extracted from each atom. Each component of the vector properties is modeled independently as a
scalar Gaussian process. The training set consists of 45 configurations uniformly covering the range
rO-H = 0.93, 0.95, 0.97, 0.99, 1.05Å and θH-O-H = 101◦ , 105.5◦ , 111◦ . As shown in Table 3, the GPR model can
successful drive calculations of the infrared spectrum of the molecule from randomly perturbed initial
structures in arbitrary orientation. The fingerprint configuration is the same as in the previous section.
4.4
Molecular Dyanmics Trajectory
As shown in Figure 11, here we attempt to fit for the forces felt by the atoms in a benzene molecule along
the MD Trajectories as obtained from a sibling database of QM7 [28, 29]. The density kernel for the carbon
atoms assumes the same functional form with that of the oxygen atoms, but uses a different smoothing
length function σC (r) = 1.2 + 0.2krk. The rest of the parameters are inherited from the previous examples.
The training configurations were chosen adaptively in an iterative process using the sum of the GPR
posterior variance and the prediction error as the acquisition function.
21
Tang et al.
Molecular fingerprints
Table 3: Geometry optimization and vibrational analysis of a single water molecule using GPR and
our proposed fingerprint algorithm. 256 independent trials were performed using coordinates of water
perturbed from equilibrium by a Gaussian noise N(0, 0.15I) followed by a randomly chosen rigid-body
rotation.
Zero-point energy
Static dipole
Residual Force
Mode
0
1
2
GPR
DFT
0.591 ± 0.003 eV
2.1580 ± 0.0001 D
0.0016 ± 0.0005 eV/Å
0.583 eV
2.159 D
0.0003 eV/Å
Frequency (cmˆ-1)
Intensity (D/A)ˆ2 amuˆ-1
GPR
DFT
GPR
DFT
1576.5 ± 1.4
3819.3 ± 0.9
3916.7 ± 1.6
1602.4
3817.5
3922.6
1.5726 ± 0.0005
0.2516 ± 0.0005
1.3349 ± 0.0028
1.5767
0.2159
1.3401
⋅ mol−1 )
−1
⋅ mol−1 )
0
fz (eV ⋅ Å
−1
40
fy (eV ⋅ Å
fx (eV ⋅ Å
−1
⋅ mol−1 )
80
−40
−80
0
25
75
50
100
0
25
Time (ns)
MAE = 0.17 (eV ⋅ Å
−1
75
50
100
0
25
Time (ns)
⋅ mol−1 )
MAE = 0.12 (eV ⋅ Å
−1
75
50
100
Time (ns)
⋅ mol−1 )
MAE = 0.13 (eV ⋅ Å
−1
⋅ mol−1 )
80
fit
40
0
−40
−80
−80
−40
0
40
80
−80
−40
0
truth
truth
40
80
−80
−40
0
40
80
truth
Figure 11: The force exerted on the carbon atoms in a benzene molecule was fit using 100 out of the 1200
configurations in a 100-ns molecular dynamics trajectory at 0.5 ns time step size. Color coding corresponds
to carbon atom id, while crosses indicate the atomistic configurations used in the GPR training.
22
Tang et al.
5
Molecular fingerprints
Connection to Other Fingerprint Algorithms
In Figure 12, we compare the ability to distinguish atomistic configurations of our fingerprint as well
as SOAP and the Coulomb matrix. Our work is inspired by the SOAP descriptor [12], which proposes
the use of smoothed densities to represent atomistic neighborhoods. However, instead of converting the
density field into the frequency domain using spherical harmonics, we perform density field sampling and
comparison directly in the real space. This is enabled thanks to the available of canonical coordinate frame
as computed through the kernel minisum optimization. We have mainly used the L2 norm to compute
the distance between atomistic neighborhoods. However, our fingerprint exhibits very similar behavior
to SOAP when used together with an inner product formula
PN
d(ρs1 , ρs2 ) ≈ q
k=1 wk ρs1 (rk ) ρs2 (rk )
PN
k=1 wk ρs1 (rk ) ρs1 (rk )
(31)
q
PN
k=1 wk ρs2 (rk ) ρs2 (rk )
as demonstrated in Figure 12A. Thus, our fingerprint could be used in conjunction with a wide variety of
covariance functions based on either the Euclidean distance or the inner product similarity.
(A)
(B)
1.0
1.0
Δφ
Δφ
DECAF-L2
SOAP
DECAF-Dot
Coulomb
0.0
random
rigid-body
rotation
radial
stretch
SOAP
DECAF − Dot
Coulomb
0.0
1.0
11.0
Interatomic Distance
random
rigid-body
rotation
i.i.d.
noise
perturbation
0.0
1.0
Δφ(DECAF-L2 )
Figure 12: A comparison between the distance measure used DECAF, SOAP, and the Coulomb matrix.
Fingerprint distances ∆ϕ shown in the plots are measured against ri j = 1.0 in (A) and an randomly chosen
initial state in (B).
At first sight, DECAF is very different from the Coulomb matrix fingerprint and GRAPE, which are
both graph-based algorithms [17, 18]. However, instead of trying to capture the overall density field, if
we measure the contribution from each individual atom on the quadrature nodes at z1 , z2 , . . . , zM as a row
vector, and stacked up the results to yield the matrix
EN×M
= k(xi , z j ).
ij
(32)
Then E can be regarded as an incidence matrix [30] between atoms and the quadrature nodes. This is
similar to the graph-based abstraction as seen in the Coulomb matrix and the GRAPE kernel. However,
in both cases the vertices in the graph represent atoms while the edges represent pairwise interatomic
interactions. Here, the density-based incidence matrix adopts the opposite pattern and constructs a graph
with the quadrature nodes being vertices and atoms being edges. The adjacency matrix in this case is
23
Tang et al.
Molecular fingerprints
Molecule
Graph
x3
H1
Coulomb Matrix
O H1 H2
x2
H2
x1
O
O
O
H1
H1
H2
H2
Density-based
Sensors
z1 z2 z3 z4
O
z2
Adjacency Matrix
z1
z1 z2 z3 z4
z2
H1
z3
H2
z3
H2
Density-based
Incidence Matrix
z1
z4
GRAPE
O
H1
z4
Figure 13: A comparison between graph-based molecular fingerprints. The Coulomb Matrix and the
GRAPE kernel construct graphs where nodes corresponds to atoms while the weights on the edges are
determined by some pairwise inter-atomic interactions. In contrast, in the density-based incidence matrix
we construct a graph on a set of quadrature nodes whose connectivity is weighted by a sum of contributions
from individual atoms.
computed as the inner product ET E:
AM×M
= (ET E)i j =
ij
N
X
k(xk , zi ) k(xk , z j ).
(33)
k=1
The weight on the edges, as represented by the elements of the adjacency matrix A, can be interpreted as
the total flux as contributed by all paths each bridged by an atom k. We have numerically found that the
smallest N eigenvalues (except for the 0 eigenvalue) of the symmetric normalized Laplacian
L = I − D−1/2 AD−1/2 , where Dii = δi j
X
Ai j
(34)
j
is invariant under rotation up to a certain noise level, even if the quadrature nodes do not rotate with the
atoms. Nonetheless, this detour appears to represent a pure theoretical interest rather than any practical
value.
6
Conclusion
In this paper, we presented the Density-Encoded Canonically Aligned Fingerprint (DECAF) by exploring
the idea of using smoothed density fields to represent and compare atomistic neighborhoods. One of
24
Tang et al.
Molecular fingerprints
the key enabling technique in DECAF is a kernel minisum algorithm, which allows the unambiguous
identification of a canonically aligned coordinate frame that can be used for rotationally invariant projection of the density field as well as any associated vector quantities. We have performed detailed analysis
to study the behavior of the fingerprint by changing various parameter, such as resolution, smoothing
length, and the choice of weight functions. We demonstrate that the fingerprint algorithm can be used to
implement highly accurate regressions of both scalar and vector properties of atomistic systems including
energy, force and dipole moment, and could be a useful building block for constructing data-driven next
generation force fields to accelerate molecular mechanics calculations with an accuracy comparable to
those driven by quantum mechanical theories and calculators.
Acknowledgment
This work was supported by the Department of Energy (DOE) Collaboratory on Mathematics for Mesoscopic Modeling of Materials (CM4). This work was also supported by the Army Research Laboratory
under Cooperative Agreement Number W911NF-12-2-0023.
References
[1] G. Zhao, J. R. Perilla, E. L. Yufenyuy, X. Meng, B. Chen, J. Ning, J. Ahn, A. M. Gronenborn, K. Schulten,
C. Aiken, and Others. Mature HIV-1 capsid structure by cryo-electron microscopy and all-atom
molecular dynamics. Nature, 497(7451):643–646, 2013.
[2] K. Lindorff-Larsen, P. Maragakis, S. Piana, and D. E. Shaw. Picosecond to Millisecond Structural
Dynamics in Human Ubiquitin. The Journal of Physical Chemistry B, 120(33):8313–8320, 2016.
[3] A. K. Rappe, C. J. Casewit, K. S. Colwell, W. A. Goddard, and W. M. Skiff. UFF, a full periodic table
force field for molecular mechanics and molecular dynamics simulations. Journal of the American
Chemical Society, 114(25):10024–10035, 1992.
[4] W. D. Cornell, P. Cieplak, C. I. Bayly, I. R. Gould, K. M. Merz, D. M. Ferguson, D. C. Spellmeyer, T. Fox,
J. W. Caldwell, and P. A. Kollman. A second generation force field for the simulation of proteins,
nucleic acids, and organic molecules. Journal of the American Chemical Society, 117(19):5179–5197, 1995.
[5] W. L. Jorgensen, D. S. Maxwell, and J. Tirado-Rives. Development and Testing of the OPLS All-Atom
Force Field on Conformational Energetics and Properties of Organic Liquids. Journal of the American
Chemical Society, 118(15):11225–11236, 1996.
[6] D. Frenkel and B. Smit. Understanding molecular simulation: from algorithms to applications, volume 1.
Academic press, 2001.
[7] A. R. Leach. Molecular modelling: principles and applications. Pearson education, 2001.
25
Tang et al.
Molecular fingerprints
[8] T. Cheng, A. Jaramillo-Botero, W. A. Goddard, and H. Sun. Adaptive accelerated ReaxFF reactive
dynamics with validation from simulating hydrogen combustion. Journal of the American Chemical
Society, 136(26):9434–9442, 2014.
[9] D. Braun, S. Boresch, and O. Steinhauser. Transport and dielectric properties of water and the
influence of coarse-graining: Comparing BMW, SPC/E, and TIP3P models. The Journal of Chemical
Physics, 140(6):064107, 2014.
[10] S. Boonstra, P. R. Onck, and E. van der Giessen. CHARMM TIP3P Water Model Suppresses Peptide
Folding by Solvating the Unfolded State. The Journal of Physical Chemistry B, 120(15):3692–3698, 2016.
[11] J. Behler. Atom-centered symmetry functions for constructing high-dimensional neural network
potentials. The Journal of Chemical Physics, 134(134), 2011.
[12] A. P. Bartók, R. Kondor, and G. Csányi. On representing chemical environments. Physical Review B,
87(18):184115, 2013.
[13] Z. Li, J. R. Kermode, and A. De Vita. Molecular Dynamics with On-the-Fly Machine Learning of
Quantum-Mechanical Forces. Physical Review Letters, 114(9):096405, 2015.
[14] A. Khorshidi and A. A. Peterson. Amp: A modular approach to machine learning in atomistic
simulations. Computer Physics Communications, 207:310–324, 2016.
[15] C. E. Rasmussen and C. K. I. Williams. Gaussian Processes for Machine Learning. 2006.
[16] D. Specht. A general regression neural network. IEEE Transactions on Neural Networks, 2(6):568–576,
1991.
[17] M. Rupp, A. Tkatchenko, K.-R. Muller, O. A. von Lilienfeld, K.-R. R. Müller, O. Anatole Von Lilienfeld,
and O. A. von Lilienfeld. Fast and accurate modeling of molecular atomization energies with machine
learning. Physical Review Letters, 108(5):58301, 2012.
[18] G. Ferré, T. Haut, and K. Barros. Learning molecular energies using localized graph kernels. The
Journal of Chemical Physics, 146(11):114107, 2017.
[19] H. Y. Sun. Learning over Molecules : Representations and Kernels. PhD thesis, Harvard University, 2014.
[20] R. R. Coifman, S. Lafon, A. B. Lee, M. Maggioni, B. Nadler, F. Warner, and S. W. Zucker. Geometric
diffusions as a tool for harmonic analysis and structure definition of data: diffusion maps. Proceedings
of the National Academy of Sciences of the United States of America, 102(21):7426–31, 2005.
[21] A. P. Bartók, M. C. Payne, R. Kondor, and G. Csányi. Gaussian Approximation Potentials: The
Accuracy of Quantum Mechanics, without the Electrons. Physical Review Letters, 104(13):136403,
2010.
[22] V. Botu and R. Ramprasad. Learning scheme to predict atomic forces and accelerate materials
simulations. Physical Review B, 92(9):094306, 2015.
26
Tang et al.
Molecular fingerprints
[23] J. Barzilai and J. M. Borwein. Two-Point Step Size Gradient Methods. IMA Journal of Numerical
Analysis, 8(1):141–148, 1988.
[24] D. Lebedev, VI and Laikov. A quadrature formula for the sphere of the 131st algebraic order of
accuracy. Doklady. Mathematics, 59(3):477–481, 1999.
[25] P. Rabinowitz and G. Weiss. Tables of Abscissas and Weights for Numerical Evaluation of Integrals of
the Form \int 0ˆ\infty eˆ{-x} xˆn f(x) dx. Mathematical Tables and Other Aids to Computation, 13(68):285–
294, 1959.
[26] B. Leistedt and J. D. McEwen. Exact Wavelets on the Ball. IEEE Transactions on Signal Processing,
60(12):6257–6269, 2012.
[27] Y.-H. Tang. Reference implementation for the algorithms presented in ”An Atomistic Fingerprint
Algorithm for Learning Ab Initio Molecular Force Fields”, 2017. DOI: 10.5281/ZENODO.1054550.
[28] K. T. Schütt, F. Arbabzadah, S. Chmiela, K. R. Müller, and A. Tkatchenko. Quantum-chemical insights
from deep tensor neural networks. Nature Communications, 8:13890, 2017.
[29] S. Chmiela, A. Tkatchenko, H. E. Sauceda, I. Poltavsky, K. T. Schütt, and K.-R. Müller. Machine
learning of accurate energy-conserving molecular force fields. Science Advances, 3(5):e1603015, 2017.
[30] J. L. Gross and J. Yellen. Graph theory and its applications. Chapman & Hall/CRC, 2005.
[31] M. Liu, G. Liu, and K. Lam. Constructing smoothing functions in smoothed particle hydrodynamics
with applications. Journal of Computational and Applied Mathematics, 155(2):263–284, 2003.
[32] L. B. Lucy. A numerical approach to the testing of the fission hypothesis. The Astronomical Journal,
82(12), 1977.
Appendix
Polynomial Smoothing Functions with Compact Support
As candidates for the weight of integral and density scaling functions (Section 3.2), a class of compact
polynomials that satisfy the criteria [31]:
1. is compactly supported,
2. is strictly positive within some cutoff distance rc ,
3. decreases monotonically,
4. is at least twice continuously differentiable
with minimal number of non-zero terms are:
W a,b (s) =
−b sa + a sb
, a > b > 2,
σ
27
(35)
Tang et al.
Molecular fingerprints
where s = 1 − r/h is the normalized complementary coordinate within the span h of the kernel, and
σ = 8πh
3
a
b
− 3
3
2
2
b + 6b + 11b + 6 a + 6a + 11a + 6
!
(36)
is an optional normalization factor to ensure that the integral of the kernel in a 3D ball of radius h is
unity. The parameters a and b are free parameters that can be used to adjust the smoothness and width
of the kernel, and can take any real numbers satisfying the condition a > b > 2. Note that the kernel W 4,3
is equivalent to the Lucy kernel commonly used in Smoothed Particle Hydrodynamics simulations [32].
The kernel can be evaluated very efficiently using only multiplication and addition when both a and b are
integers.
2.5
(A) Kernel
0.0
0.1
−1.0
2.0
0.0
0.8
0.9
1.0
−20
0.0
−3.0
−5.0
(B) Second Derivative
5
−40
−4.0
0.5
20
0
−2.0
1.5
1.0
(B) First Derivative
−60
−0.5
0.9
1.0
0
0.9
1.0
0.0
−6.0
−80
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0 0.2 0.4 0.6 0.8 1.0
r/h
r/h
Gaussian
Cosine
W 8,6
W 6,5
W 6,3
W 5,4
W 5,3
W 4,3
7 5
W 2,2
r/h
Figure 14: Visualization of the polynomial kernels as given in Eq. 35 with a unit support radius. The
kernels are bell-shaped with a derivative of 0 at the origin. Both the first and second derivatives of the
kernels transition smoothly to 0 at its support radius. In contrast, the Gaussian kernel and its derivatives
does not decay to zero at any finite distance, while the second derivative of the Cosine kernel as mentioned
in previous work [11, 12] is not zero at the cutoff distance.
Table of Quadrature Nodes and Weights
In Table 4, we list the nodes and weights of the Laguerre quadrature rules up to Nr = 6, using notations
from Eq. 22. In Table 5, we list the nodes and weights of the Lebedev quadrature rules up to Nr = 6, using
notations from Eq. 21. The Laguerre and Lebedev quadrature nodes can be combined using Eq. 23-26 into
composite grids for sampling the atomistic density field.
28
Tang et al.
Molecular fingerprints
Table 4: Laguerre quadrature nodes and weights up to 6 points.
================================================================================
n = 0
n = 1
n = 2
n = 3
n = 4
n = 5
-------------------------------------------------------------------------------Nr = 2
r_n
2.0000
6.0000
a_n
1.5000
0.5000
-------------------------------------------------------------------------------Nr = 3
r_n
1.5174
4.3116
9.1710
a_n
1.0375
0.9058
0.0568
-------------------------------------------------------------------------------Nr = 4
r_n
1.2268
3.4125
6.9027
12.4580
a_n
0.7255
1.0634
0.2067
0.0044
-------------------------------------------------------------------------------Nr = 5
r_n
1.0311
2.8372
5.6203
9.6829
15.8285
a_n
0.5209
1.0667
0.3835
0.0286
0.0003
-------------------------------------------------------------------------------Nr = 6
r_n
0.8899
2.4331
4.7662
8.0483
12.6004
19.2620
a_n
0.3844
0.9971
0.5361
0.0795
0.0029
0.0000
================================================================================
29
Tang et al.
Molecular fingerprints
Table 5: Lebedev quadrature nodes and weights up to 50 points.
================================================================================
Na = 6
x_m
y_m
z_m
b_m
m = 0,1
+ -1.00000
0.00000
0.00000
0.16667
m = 2,3
0.00000
+ -1.00000
0.00000
0.16667
m = 4,5
0.00000
0.00000
+ -1.00000
0.16667
-------------------------------------------------------------------------------Na = 14
x_m
y_m
z_m
b_m
m = 0,1
+ -1.00000
0.00000
0.00000
0.06667
m = 2,3
0.00000
+ -1.00000
0.00000
0.06667
m = 4,5
0.00000
0.00000
+ -1.00000
0.06667
m = 6 -13
+ -0.57735
+ -0.57735
+ -0.57735
0.07500
-------------------------------------------------------------------------------Na = 26
x_m
y_m
z_m
b_m
m = 0,1
+ -1.00000
0.00000
0.00000
0.04762
m = 2,3
0.00000
+ -1.00000
0.00000
0.04762
m = 4,5
0.00000
0.00000
+ -1.00000
0.04762
m = 6-9
0.00000
+ -0.70711
+ -0.70711
0.03810
m = 10 -13
+ -0.70711
0.00000
+ -0.70711
0.03810
m = 14 -17
+ -0.70711
+ -0.70711
0.00000
0.03810
m = 18 -25
+ -0.57735
+ -0.57735
+ -0.57735
0.03214
-------------------------------------------------------------------------------Na = 38
x_m
y_m
z_m
b_m
m = 0,1
+ -1.00000
0.00000
0.00000
0.00952
m = 2,3
0.00000
+ -1.00000
0.00000
0.00952
m = 4,5
0.00000
0.00000
+ -1.00000
0.00952
m = 6 -13
+ -0.57735
+ -0.57735
+ -0.57735
0.03214
m = 14 -17
+ -0.45970
+ -0.88807
0.00000
0.02857
m = 18 -21
+ -0.88807
+ -0.45970
0.00000
0.02857
m = 22 -25
+ -0.45970
0.00000
+ -0.88807
0.02857
m = 26 -29
+ -0.88807
0.00000
+ -0.45970
0.02857
m = 30 -33
0.00000
+ -0.45970
+ -0.88807
0.02857
m = 34 -37
0.00000
+ -0.88807
+ -0.45970
0.02857
-------------------------------------------------------------------------------Na = 50
x_m
y_m
z_m
b_m
m = 0,1
+ -1.00000
0.00000
0.00000
0.01270
m = 2,3
0.00000
+ -1.00000
0.00000
0.01270
m = 4,5
0.00000
0.00000
+ -1.00000
0.01270
m = 6-9
0.00000
+ -0.70711
+ -0.70711
0.02257
m = 10 -13
+ -0.70711
0.00000
+ -0.70711
0.02257
m = 14 -17
+ -0.70711
+ -0.70711
0.00000
0.02257
m = 18 -25
+ -0.57735
+ -0.57735
+ -0.57735
0.02109
m = 26 -33
+ -0.30151
+ -0.30151
+ -0.90453
0.02017
m = 34 -41
+ -0.30151
+ -0.90453
+ -0.30151
0.02017
m = 42 -49
+ -0.90453
+ -0.30151
+ -0.30151
0.02017
================================================================================
30
| 5cs.CE
|
arXiv:1711.03147v1 [cs.AI] 8 Nov 2017
On the incorporation of interval-valued fuzzy sets into
the Bousi-Prolog system: declarative semantics,
implementation and applications
Clemente Rubio-Manzanoa , Martı́n Pereira-Fariñab
b
a
Dep. of Information Systems, University of the Bı́o-Bı́o, Chile.
Centre for Argument Technology, University of Dundee, UK. Departamento de Filosofı́a
e Antropoloxı́a, Universidade de Santiago de Compostela,Spain.
Abstract
In this paper we analyse the benefits of incorporating interval-valued fuzzy
sets into the Bousi-Prolog system. A syntax, declarative semantics and implementation for this extension is presented and formalised. We show, by
using potential applications, that fuzzy logic programming frameworks enhanced with them can correctly work together with lexical resources and
ontologies in order to improve their capabilities for knowledge representation
and reasoning.
Keywords: Interval-valued fuzzy sets, Approximate Reasoning, Lexical Knowledge Resources, Fuzzy Logic Programming, Fuzzy Prolog.
1. Introduction and Motivation
Nowadays, lexical knowledge resources as well as ontologies of concepts
are widely employed for modelling domain independent knowledge [1, 2] or
Email address: [email protected] (Clemente Rubio-Manzano)
Preprint submitted to Summited to Studies of Computational Intelligent SeriesNovember 16, 2017
by automated reasoners [3]. In the case of approximate reasoning, this makes
possible to incorporate general knowledge into any system, which is independent of the programmer’s background [4].
Inside the former and current frameworks of fuzzy logic programming [5,
6, 7, 8, 9, 10], we argue that lexical reasoning might be an appropriate way for
tackling this challenge, because of this type of knowledge is usually expressed
linguistically. However, from a computational point of view, this source of
information involves vagueness and uncertainty and, consequently, it must
be specifically addressed. Fuzzy set theory (FS) is a good candidate, but
it shows some particular limitations to this aim: i) sometimes, words mean
different things to different people and this generates and additional layer
of uncertainty that cannot be adequately handled by FS; ii) the definition
of membership functions for word meaning is also a debatable question and,
therefore, achieving an agreement by means of a standard fuzzy set it is
difficult; and, iii) with respect to semantic similarity measures used in this
proposal, there is not a dominant one and, therefore, for two given words,
different degrees of resemblance can be obtained with the resulting additional
level of uncertainty.
In the specific field of fuzzy logic programming and fuzzy Prolog systems,
little attention has been paid to the impact of this type of high degree of
uncertainty and vagueness inherent to lexical knowledge, which is used in the
definition of knowledge bases and inference processes. Next, a very simple
example is introduced in order to illustrate i) and ii) in the building of a
2
Prolog knowledge base.
Example 1. Suppose that we extract from Internet two people’s opinions
about a particular football player. The first one says “a is a normal player”
and the second one says “a is a bad player”. If we consider the label for
qualifying the highest quality (e.g., “good”) as a basic component, this lexical knowledge could be modelled by using two annotated facts as: “football player(a,good):-0.8.” and “football player(a,good):-0.6.”, respectively. In
this case, we use “football player(a,good):-0.6.” given the infimum is usually
employed. However, as it can be observed, the information of the first person
is lost.
Case iii) deserves a special attention, given it involves the use of independent linguistic resources (such as WordNet Similarity [11]). As we said, this
tool provide us different measures according to alternative criteria for assessing the degree of similarity between two words. In Example 2, we illustrate
this situation by means of a simple case.
Example 2. Suppose we have the fact “loves(a,b)” and we extract the closeness between “loves” and “desires” by using two different semantics measures
obtaining 0.8 and 0.6. Therefore, in order to represent this semantic knowledge we could employ two facts either “desires(a,b):-0.8” or “desires(a,b):0.6”.
In order to address both Examples 1 and 2 inside the same frame, we
propose to enhance the Bousi-Prolog system with interval-valued fuzzy sets
(IVFSs), since they allow us to capture the uncertainty associated to lexical
knowledge better than FS. Several advantages have pointed out for dealing
with environments with high uncertainty or imprecision using IVFSs, such
as [12]; other authors have also shown that IVFSs can generate better results
than standard FSs [13]. Additionally, the use of intervals for describing
3
uncertain information has been successfully applied in the realms of decision
making, risk analysis, engineering design, or scheduling [14].
Both Example 1 and Example 2 can be easily modelled by means of
IVFSs, using and interval for combining information of the different sources
into a single fact such as “football player(a,good):-[0.6,0.8]” or “desires(a,b):[0.6,0.8]”, respectively.
The main contribution of this paper is to design and implement an intervalvalued fuzzy logic language, and to incorporate it into the Bousi-Prolog system [15]. This task involves different challenges both from theoretical and
implementation points of view. The former entails adding a IVFSs arithmetic into the Warren Abstract Machine based on Similarity (SWAM) [16];
the latter, means to establish a (model-theoretic) declarative semantics for
the language in the classical way, formalising the notion of least interval
valued fuzzy Herbrand model for interval-valued fuzzy definite programs.
This paper is divided into the following sections: section 2 introduces the
concepts that support our approach; section 3 describes the details of the
syntax, semantics and implementation of the proposed language; section 4
analyses different realms where this programming language can be applied;
in section 5, the main differences between this proposal an others that are
described in the literature are discussed; and, finally, section 6 summarizes
our main conclusions and some ideas for future work.
4
2. Preliminary Concepts
2.1. Interval-Valued Fuzzy Sets
IVFSs are a fuzzy formalism based on two membership mappings instead
of a single one, like in standard FSs. Each one of these membership functions
are called, lower membership function and upper membership function. Both
are established on a universe of discourse X , and they map each element from
X to a real number in the [0, 1] interval, where the elements of X belongs to
A according to an interval.
Definition 2.1. An interval-valued fuzzy set A in X is a (crisp) set of ordered triples: A = {(x, µA (x), µA (x)) : x ∈ X ; µA (x), µA (x) : X → [0, 1]}
where: µ, µ are the lower and the upper membership functions, respectively,
satisfying the following condition: 0 ≤ µA (x) ≤ µA (x) ≤ 1 ∀x ∈ X
As can be observed in Definition 2.1, those intervals are included in [0, 1]
and closed at both ends. On the other hand, some arithmetic operations
on interval-numbers have been recalled since they are useful in operating on
cardinalities of IVFSs. Let a=[a, a], b=[b, b] be intervals in R, and r ∈ R+.
The arithmetic operations ’+’, ’-’, ’·’ and power are defined as follows:
[a, a] + [b, b] = [a + b, a + b];
(1)
[a, a] − [b, b] = [a − b, a − b];
(2)
[a, a] · [b, b] = [min(a · b, a · b, a · b, a · b), max(a · b, a · b, a · b, a · b)];
([a, a])r = [ar , ar ] for non-negative a, a
5
(3)
(4)
The operations of union and intersection for IVFSs are defined by triangular norms. Let A, B be IVFSs in X , t a t-norm and s a t-conorm. The union of
A and B is the interval-valued fuzzy set A∪B with the membership function:
µA∪B (x) = [s(µA (x), µB (x)), s(µA (x), µB (x))]. The intersection of A and B
is the IVFSs A∩B in which µA∩B (x) = [t(µA (x), µB (x)), t(µA (x), (µB (x))].
Thus, de Morgan’s laws for IVFSs A,B in X are: (A ∪ B)c = Ac ∩ B c and
(A ∩ B)c = Ac ∪ B c .
Let L be a lattice of intervals in [0, 1] that satisfies:
L = [x1 , x2 ] ∈ [0, 1]2 with x1 ≤ x2 ;
(5)
[x1 , x2 ] ≤L [y1 , y2 ] if f x1 ≤ y1 and x2 ≤ y2 .
(6)
Also by definition:
[x1 , x2 ] <L [y1 , y2] ⇔ x1 < y1 , x2 ≤ y2 or x1 ≤ y1 , x2 < y2 ;
(7)
[x1 , x2 ] =L [y1 , y2 ] ⇔ x1 = y1 , x2 = y2 .
(8)
Hence, 0L = [0, 0] and 1L = [1, 1] are the smallest and the greatest elements in L.
2.2. Approximate Deductive Reasoning
When we consider a collection of imprecise premises and a possible imprecise conclusion inferred from them in a Prolog program, we are applying
a process of approximate deductive reasoning. These set of statements can
6
be interpreted under two different frames [17] in a Prolog program: conditional and set-based interpretations. If the former is assumed, an imprecise
premise is an assertion qualified by a degree of truth; e.g. “John is tall with
[0.2,0.5]” means that the degree of truthfulness of this sentence using and
IVFS is [0.2, 0.5]. On the other hand, if the latter is adopted, the interval
that qualifies the sentence means the degree of membership of an element to
a specific set; e.g., “John is tall with [0.2,0.5]” means that the membership
of John to the set of tall people is [0.2, 0.5]. The conclusion inferred from
an imprecise premise must be also qualified by the same type of degree; e.g.
“John is a good player with [0.2,0.5]”.
In order to preserver the coherence with classical Prolog, we adopt the
propositional interpretation (the interval indicates the degree of truth of the
assertion) and, consequently, approximate deductive reasoning is based on
multi-valued modus ponens [18]:
Q, [α, α]
(9)
A ← Q, [β, β]
(10)
A, T ([α, α], [β, β]]
(11)
If we have (9) and (10), we can deduce (11) with T a t-norm defined on the
lattice L([0, 1]).
7
3. Simple Interval-valued fuzzy prolog: syntax, semantics and implementation
The design of a programming language involves three main steps. Firstly,
the definition of the syntax; secondly, the elaboration of a formal study of its
semantics; and thirdly, an implementation of the system. In order to address
the tasks related with syntax and semantics, we will follow the guidelines
established in [19] and [20]1 ; for the implementation task, we will follow the
guidelines detailed in [16].
3.1. Sintax
An Interval-valued fuzzy program conveys a classical Prolog knowledge
base and a set of IVFSs, which are used for annotating the facts by means
of an interval-valued fuzzy degree: p(t1 , . . . , tn )[α, α].
Definition 3.1. An interval-valued fuzzy definite clause is a Horn clause of
the form A[α, α] or A ← B1 , . . . , Bn [β, β], where A is called the head, and
B1 , . . . , Bn denote a conjunction which is called the body (variables in a clause
are assumed to be universally quantified).
Definition 3.2. An interval-valued fuzzy definite program is a finite set of
interval-valued fuzzy clauses.
Example 3. Let Π = {p(X) ← q(X), q(a)[0.8, 0.9], q(b)[0.7, 0.8]} be an intervalvalued fuzzy definite program, Π generates a first order language, L, whose
alphabet is comprised of the set of variable symbols, X , constant symbols, C,
function symbols, F and predicate symbols, P, which appear in the clauses
of Π. We assume that the first order language L has, at least one constant
symbol; i.e., an assertion. If there are not constants available in the alphabet, an artificial constant “a” must be added to it. The first order language
L generated by Π is: X = {x}, C = {a, b, c}, F = ∅ and P = {p, q, r}.
1
We assume familiarity with the theory and practice of logic programming.
8
3.2. Declarative Semantics
In logic programming, the declarative semantics for a program is traditionally formulated on the basis of the least Herbrand model (conceived as
the infimum of a set of interpretations). In this section, we formally introduce the semantic notions of Herbrand interpretation, Herbrand model and
least Herbrand model for an interval-valued fuzzy program Π, in order to
characterise it.
In our framework, truth-values of the facts are modelled in terms of
interval-valued degrees [α, α] with 0 ≤ α ≤ α ≤ 1. An interval-valued fuzzy
interpretation I is a pair hD, J i where D is the domain of the interpretation
and J is a mapping which assigns meaning to the symbols of L: specifically
n-ary relation symbols are interpreted as mappings D n −→ L([0, 1]). In order to evaluate open formulas, we have to introduce the notion of variable
assignment. A variable assignment, ϑ, w.r.t. an interpretation I = hD, J i,
is a mapping ϑ : V −→ D, from the set of variables V of L to the elements of
the interpretation domain D. This notion can be extended to the set of terms
of L by structural induction as usual. The following definition formalises the
notion of valuation of a formula in our framework.
Definition 3.3. Given an interval-valued fuzzy interpretation I and a variable assignment ϑ in I, the valuation of a formula w.r.t. I and ϑ is:
1. (a) I(p(t1 , . . . , tn ))[ϑ] = p̄(t1 ϑ, . . . , tn ϑ), where J (p) = p̄;
(b) I(A1 , . . . , An ))[ϑ] = inf {I(A1 )[ϑ], . . . , I(An )[ϑ]};
2. I(A ← Q)[ϑ] = 1 if I(A) >= I(Q); I(A ← Q)[ϑ] = I(A)[ϑ] if
I(A) < I(Q);
9
3. I((∀x)C)[ϑ] = inf{I(C)[ϑ′ ] | ϑ′ x–equivalent to ϑ} where p is a predicate symbol, A and Ai atomic formulas and Q any body, C any clause,
T is any left-continuous t-norm defined on L([0, 1]). An assignment ϑ′
is x–equivalent to ϑ when zϑ′ = zϑ for all variables z 6= x in V.
Definition 3.4. Let L be a first order language. The Herbrand universe UL
for L, is the set of all ground terms, which can be formed out of the constants
and function symbols appearing in L.
Definition 3.5. Let L be a first order language. The Herbrand base BL
for L is the set of all ground atoms which can be formed by using predicate
symbols from L with ground terms from the Herbrand universe as arguments.
Example 4. Let us consider again the language L generated by the program
Π of Example 3, the Herbrand universe UL = {a, b, c} and the Herbrand base:
BL = {p(a), p(b), p(c), q(a), q(b), q(c), r(a), r(b), r(c)}.
It is well-known that, in the classical case, it is possible to identify a
Herbrand interpretation with a subset of the Herbrand base. Therefore, a
convenient generalization of the notion of Herbrand interpretation to the
interval-valued fuzzy case consists in establishing an interval-valued fuzzy
Herbrand interpretation as an interval-valued fuzzy subset of the Herbrand
base.
Definition 3.6 (Interval-valued fuzzy interpretation). Given, a first order language L, an interval-valued fuzzy Herbrand interpretation for L is a
mapping I : BL −→ L([0, 1]).
Hence, the truth value of a ground atom A ∈ BL is I(A). Sometimes we
will represent an interval-valued fuzzy Herbrand interpretation I extensively:
as a set of pairs {hA, [α, α]i | A ∈ BL and [α, α] = I(A)}.
Now, we introduce the notion of Interval-valued Fuzzy Herbrand Model,
which is formalised in Definitions 3.7 and 3.8. We employ a declarative
10
semantics based on a threshold [21, 20]. Intuitively, a threshold [λ, λ] is
delimiting truth degrees equal o greater that [λ, λ] as true. Therefore, we
are going to speak of Interval-valued Fuzzy Herbrand Model at level [λ, λ] or
simply [λ, λ]-model.
Definition 3.7. An Interval-valued fuzzy Herbrand Interpretation is a [λ, λ]model of an interval-valued fuzzy clause C[α, α] if and only if I(C) ≥ [α, α] ≥
[λ, λ].
Definition 3.8. An Interval-valued fuzzy Herbrand Interpretation is a [λ, λ]model of an interval-valued fuzzy program Π if and only if I is a [λ, λ]-model
for each clause C[α, α] ∈ Π.
Theorem 3.9. Let Π be an Interval-valued fuzzy program and suppose Π has
a [λ, λ]-model. Then Π has a Herbrand [λ, λ]-model.
Proof 1. Suppose that M is a [λ, λ]-model of Π. Let M′ be an Intervalvalued fuzzy Herbrand interpretation: M′ = {A ∈ BΠ | M(A) ≥ [λ, λ]}. We
are going to prove that this interpretation is a [λ, λ]-model for all clauses of
Π. Let C any clause, by initial supposition and by definition of [λ, λ]-model
for an interval-valued fuzzy program, we have that:
C ≡ ∀x1 , . . . , xn (p(x1 , . . . , xn ) ← q1 (x1 , . . . , xn ) ∧ . . . ∧ qm (x1 , . . . , xn ))[β, β]
M is a [λ, λ]-model of C iff ∀a1 , . . . , an ∈ UL , M(C) ≥ [β, β] ≥ [λ, λ].
Let a1 , . . . , an ∈ UL then we have that M(p(a1 , . . . , an )) = [β, β] ≥ [λ, λ]
what implies that M′ (p(a1 , . . . , an )) ≥ [λ, λ]
Definition 3.10. Let Π be an interval-valued fuzzy program. Let A be an
interval-valued fuzzy clause of Π. Then A is a logical consequence of Π at
level [λ, λ] if and only if for each interval-valued fuzzy interpretation I, if I
is a [λ, λ]-model for Π then I is a [λ, λ]-model for A.
Proposition 3.11. A is a logical consequence of an interval-valued fuzzy
program Π at level [λ, λ] if and only if for every interval-valued fuzzy Herbrand
interpretation I for Π, if I is a [λ, λ]-model for Π, it is an interval-valued
fuzzy Herbrand [λ, λ]-model for A.
11
Proof 2. First, let us suppose that A is a logical consequence for Π at level
[λ, λ], then, by definition, for any interval-valued fuzzy interpretation I if I is
[λ, λ]-model for Π, it is a [λ, λ]-model for A. Moreover, by the Theorem 3.9,
there must exist I’ which being an interval-valued fuzzy Herbrand model for
Π at level [λ, λ], it is a [λ, λ]-model for A. This establishes the first side
of the argument. Now, we have that for every interpretation I, if I is a
Herbrand model for Π at level [λ, λ], it is a Herbrand [λ, λ]-model for A. Let
M be an interpretation, not necessarily Herbrand, which is a [λ, λ]-model
for Π. We have that: M′ = {p(t1 , . . . , tn )[α, α] with p(t1 , . . . , tn ) ∈ BL |
M(p(t1 , . . . , tn )) ≥ [α, α] ≥ [λ, λ]} and by the Theorem 3.9 M’ is a [λ, λ]model for Π. And so it is for A. So, M is a [λ, λ]-model for all ground
instances A’ of A. As result M is a [λ, λ]-model for A’, hence for A and A’.
This establishes the other side of the argument.
The ordering ≤ in the lattice L([0, 1]) can be extended to the set of
interval-valued fuzzy interpretation as follows: I1 ⊑ I2 iff I1 (A) ≤ I2 (A) for
all interval-valued fuzzy atom A ∈ BL . It is important note that the pair
hH IV F , ⊑i is a complete lattice. Then it comes equipped with t-norms and
t-conorms, that is, T (I1 , I2 ) is an interval-valued fuzzy interpretation for all
A ∈ BL , and t(I1 , I2 ) an interval-valued fuzzy interpretation for all A ∈ BL .
Therefore, the top element of this lattice is hA, [1, 1]i with A ∈ BL and the
bottom element is hA, [0, 0]i with A ∈ BL .
Interval-valued fuzzy interpretations have an important property which
allow us to characterize the semantics of an interval-valued fuzzy program
Π.
Definition 3.12. If M1 is a model of Π at level [λ1 , λ1 ] and M2 is a model
of Π at level [λ2 , λ2 ], then M1 ∩ M2 contains the interval-valued fuzzy atom
in both M1 and M2 but to degree min([λ1 , λ1 ], [λ2 , λ2 ]).
Proposition 3.13 (Intersection Property of Models: Min-Model). Let
Π be an interval-valued fuzzy program. Let M1 , . . . , Mn be a non-empty set of
12
model for Π at levels [λ1 , λ1 ] . . . [λn , λn ], respectively. Then
min([λ1 , λ1 ] . . . [λn , λn ]) is a min-model for Π.
T
(M1 , . . . , Mn ) ≥
Proof 3. We prove this proposition by induction on the number of interpretations i:
1. Base Case (i=2). Let M1 and M2 be models for Π at levels [λ1 , λ1 ]
and [λ2 , λ2 ]. Then for all interval-valued fuzzy clause C, M1 (C) ≥
[λ1 , λ1 ] and M2 (C) ≥ [λ2 , λ2 ], so M1 ∩ M2 is a min([λ1 , λ1 ], [λ2 , λ2 ])model for Π;
2. Inductive Case (i=n). Let M1 , M2 , . . . Mn be models for Π at levels [λ1 , λ1 ] . . . [λn , λn ]. Then for all interval-valued fuzzy clause C,
Mi (C) ≥ min([λi , λi ]), so by the properties of the minimum.
Definition 3.14. Let Π be an interval-valued
fuzzy program. The least model
T
for Π is defined as follows: M = {I(A) ≥ [λ, λ] | A ∈ BL }. We call it a
min-interval-valued fuzzy degree [λ, λ]min .
Theorem 3.15. Let Π an interval-valued fuzzy program. Let M be the least
model of Π. Let A ∈ BL a ground atom of the interval-valued fuzzy Herbrand
base. M(A) ≥ [λ, λ]min if and only if A is logical consequence of Π at level
[λ, λ]min .
T
Proof 4. First, by definition M = {I(A) ≥ [λ, λ] | A ∈ BL }. Hence,
for all model I of Π, I(A) ≥ M(A) ≥ [λ, λ]min . That is, A is a logical
consequence for Π at level [λ, λ]min . This establishes the first side of the
argument. Now, If A is a logical consequence of Π by definition all model I
for Π, I is a [λ, λ]min -model
T for A. That is, I(A) ≥ [λ, λ]min . Therefore, by
definition of least model, (I(A)) ≥ [λ, λ] what implies that M ≥ [λ, λ]min .
This establishes the another side of the argument.
3.3. Fixpoint Semantics
In this section, we give a deeper characterisation of the least Herbrand
model for an interval-valued fuzzy program Π using fixpoint concepts.
This is possible because of each interval-valued fuzzy program has associated a complete lattice of interval-valued fuzzy Herbrand interpretations and
13
we can define a continuous operator on that lattice. This allows us to provide
a constructive vision of the meaning of a program by defining an immediate
consequences operator and to construct the least Herbrand model by means
of successive applications.
Definition 3.16 (Fixpoint Characterization of the least Herbrand model).
Let Π be an interval-valued fuzzy program, the mapping O : 2BL → 2BL is
defined as follows. Let I be an interval-valued fuzzy Herbrand interpretation,
then:
O = {A ∈ BL : A ← B1 , . . . , Bn [α, α] is a ground instance of a clause in Π
and I(Bi ) ≥ [α, α] ≥ [λ, λ] where I(A) ≥ inf (I(B1 , . . . , Bn ))}
As in the case of classical logic programming, interval-valued fuzzy Herbrand interpretations which are models can be characterised in terms of the
operator O.
Theorem 3.17. Let Π be an interval-valued fuzzy program. Let I be an
interval-valued fuzzy Herbrand interpretation of Π. I is [λ, λ]-model for Π if
and only if O(I) ⊆ I.
Proof 5. I is a [λ, λ]-model for Π if and only if for all clause C in Π then
I(C) ≥ [λ, λ]. Therefore, it is fulfilled if and only if for every variable assignment ϑ, I(Cϑ) ≥ [λ, λ]. Therefore, supposing without loss of generality that
C ≡ A ← B1 , . . . , Bn [α, α] then I(A ← B1 , . . . , Bn ϑ) ≥ [λ, λ], by the properties of the t-norm minimun I(B1 , . . . , Bn ϑ) ≥ [α, α] ≥ [λ, λ]min what implies
that I(B1 , . . . , Bn ϑ) ⊆ O and hence I(Aϑ) ⊆ O(I), again by the properties
of the t-norm minimun I(Aϑ) ≥ inf (I(B1 ϑ), . . . , I(Bn ϑ)) what implies that
O(I) ⊆ I
Now we are ready to demonstrate the main theorem of this subsection,
but first we recall the following results from fixpoint theory.
14
Theorem 3.18 (FixPoint Theorem). Let hL, ≤i be a complete lattice and
O : L → L be a monotonic mapping. Then O has a least fixpoint lf p(O) =
inf {x | O(x) = x} = inf {x | T (x) ≤ x}.
Proposition 3.19. Let hL, ≤i be a complete lattice and O : L → L be a
continuous mapping. Then lf p(O) = O ↑ ω.
Proof 6. See [19]
Theorem 3.20. Let Π be an interval-valued fuzzy definite program. Then
M = lf p(OΠT ) = O ↑ ω.
Proof 7. M is the least model which is the intersection of any [λ, λ]-model
for Π. As the lattice of interval-valued fuzzy Herbrand models is a complete
one, then we can use the Theorem 3.18, the Proposition 3.19 and the Theorem 3.17. Applying them and the continuity of O establishes the theorem.
Example 5. Given the program Π of Example 3, the least Herbrand model
for Π:
O ↑ 0 = I⊥ ;
O ↑ 1 = O(O ↑ 0) = {hp(a), [0, 0]i, hp(b), [0, 0]i, hq(a), [0.8, 0.9]i, hq(a), [0.7, 0.8]i}
O ↑ 2 = O(O ↑ 1) = {hp(a), [0.8, 0.9]i, hp(b), [0.7, 0.8]i, hq(a), [0.8, 0.9]i, hq(a), [0.7, 0.8]i}
O↑3=O↑2
Therefore, as the fixpoint is reached at the next item: M = O ↑ 2
3.4. Operational Semantics
We begin by providing definitions of an interval-valued SLD-derivation
and an interval-valued fuzzy SLD-refutation that will be used later for showing the soundness and the completeness of the system.
15
Definition 3.21. Let G be ← A1 , . . . , Am , . . . , Ak and C be either A[α, α] or
A ← B1 , . . . , Bq [β, β]. Then G′ is derived from G and C using mgu θ if the
following conditions hold (G’ is the interval-fuzzy resolvent of G and C): i)
Am is an atom called the selected atom in G; ii) θ is a mgu of Am and A;
iii) G′ is the interval-valued fuzzy goal ← (A1 , . . . , B1 , . . . , Bq , , . . . , Ak )θ with
[αG′ , αG′ ] = min([αC , αC ], [αG , αG ])
Definition 3.22. An interval-valued fuzzy SLD-derivation of Π∪G is a successful interval-valued SLD-derivation of Π ∪ G which has the empty clause
as the last goal in the derivation. If Gn is the empty clause, we say that the
derivation has length n. The empty clause is derived from ← (A1 , . . . , Am , . . . , Ak )[αG , αG ]
and A(t1 , . . . , tq )[αA , αA ] ← with [αGn , αGn ] = min([αA , αA ], [αG , αG ])
Definition 3.23. Let Π be an interval-valued fuzzy program and G be an
interval-valued fuzzy goal. An interval-valued fuzzy computed answer hθ, [β, β]i
for Π∪G is the substitution obtained by restricting the composition θ1 , . . . , θn
to the variables of G, where θ1 , . . . , θn is the sequence of mgu’s employed in the
finite interval-valued fuzzy SLD-derivation of Π ∪ G with an interval-valued
approximation degree [β, β]
Definition 3.24. Let Π be an interval-valued fuzzy program, G be an intervalvalued fuzzy goal ← (A1 , . . . , Ak ) and hθ, [β, β]i be an answer for Π ∪ G. We
say that hθ, [β, β]i is an interval-valued fuzzy correct answer if ∀(A1 , . . . , Ak )θ
is a logical consequence of Π at level [λ, λ]min , that is, [β, β] ≥ [λ, λ]min .
3.5. Implementation
In this section, we briefly explain how interval-valued fuzzy sets are incorporated into the Bousi-Prolog system2 . Here, we describe the structure
and main features of its abstract machine. It was created as extension of the
SWAM for the execution of Bousi-Prolog programs. We have appropriately
modified the compiler, some machine instructions and SWAM structures in
2
A
beta
version
can
be
http://www.face.ubiobio.cl/∼clrubio/bousiTools/
16
founded
at
the
URL:
order to trigger the interval-valued fuzzy resolution. It is worth noting that,
to the best of our knowledge, this is the first SWAM implementation that
supports interval-valued fuzzy resolution.
A mandatory step to achieve this result is to include a new data structure
into the architecture for computing with interval-valued fuzzy sets. This data
structure has been implemented by using a class called IntervalFS which is
formed by two private attributes of double type: upper limit, lower limit. We
define the public method constructor IntervalFS(double ll,double lu)
and the four methods (sets and gets): double getUpperLimit(); double
getLowerLimit(); void setUpperLimit(double v); void setLowerLimit(double
v).
Additionally, we overwrite both the toString and the equals meth-
ods in the usual way. Finally, methods for adding, substracting and computing minimum of interval valued fuzzy set are implemented: IntervalFS
add(IntervalFS a, IntervalFS b); IntervalFS substract(IntervalFS a, IntervalFS b); IntervalFS min(IntervalFS a, IntervalFS b).
The following example illustrates the new features of the SWAM enhanced
with IVFSs.
Example 6. Let us suppose that we want to represent the following knowledge: a football player is good when he is fast, tall and coordinated. We
know a particular player that is fast, quite tall but he is not very coordinated.
Thus, is he a good player? Answering this question and in this scenario, the
linguistic expression “is not very coordinate” could be represented by the fact
“coordinate(a) [0.2,0.4]”, the linguistic term “fast” could be represented by
the fact “fast(a) [0.9,1.0]” and “quite tall” could be represented by the fact
“tall(a)[0.8,0.9]”. A possible solution by employing a Bousi-Prolog program
is described as follows:
17
%% FACTS
coordinate(a) [0.2,0.4]
fast(a) [0.9,1.0]
tall(a )[0.8,0.9]
%% RULES
good_player(X):-tall(X), fast(X), coordinate(X)
The SWAM enhanced with IVFSs allows us to obtain the answer: “X=a
with [0.2,0.4]”. The SWAM code generated for this program is as follows:
00:good_player:trust_me [1.0,1.0]
01:
allocate
02:
get_variable Y0 A0
03:
put_value Y0 A0
04:
call coordinate (11)
05:
put_value Y0 A0
06:
call fast
(14)
07:
put_value Y0 A0
08:
call tall
(17)
09:
deallocate
10:
proceed
11:coordinate:trust_me
[0.2,0.4]
12:
get_constant a A0
13:
proceed
14:
fast: trust_me [0.9,1.0]
15:
get_constant a A0
16:
proceed
17:tall:
trust_me
[0.8,0.9]
18:
get_constant a A0
19:
proceed
20:query:
trust_me
21:
create_variable Q0 X
22:
put_value Q0 A0
23:
call good_player (00)
24:
halt
The first instruction to be executed is the one labelled with the key
“query”, hence the execution starts at the position 20 with a degree D =
[1.0, 1.0] (which is fixed in the instruction trust me). After that, from line 20
to line 23 the query is launched and the variable X is created (create variable
instruction). After that from line 00 to line 04 the first subgoal (coordinate(X)) is launched, then the execution goes to line 11 and the unification
with the term ”coordinate(a) is produced (from line 11 to 13) (put value
18
and get constant instructions), a new approximation degree is established
D = min([1.0, 1.0], [0.2, 0.4]) (trust me instruction), as these terms unify the
following subgoal (f ast(X), line 05 and from line 14 to line 16) is launched
with an approximation degree D = min([0.2, 0.4], [0.9, 1.0]); as the terms
unify, then the following subgoal (tall(X), line 08 and from line 17 to 19)
is launched with an approximation degree D = min([0.2, 0.4], [0.9, 1.0]). Finally, the assignation X = a with [0.2, 0.4] is produced.
We have implemented a limit to the expansion of the search space in a
computation by what we called a “λ-cut for IVFSs”. When the LambdaCutIVFS flag is set to a value different than [0.0, 0.0], the weak unification
process fails if the computed approximation degree goes below the stored
LambdaCutIVFS value. Therefore, the computation also fails and all possible branches starting from that choice point are discarded. By default the
LambdaCutIVFS value is [0.0, 0.0]. However, the lambda cut flag can be set
to a different value by means of a λ-cut directive: “:-lambdaCutIVFS(N).”,
where N is an interval between [0.0, 0.0] and [1.0, 1.0]. For example, a λcut of [0.5, 0.5] could be established by using the following directive: “:lambdaCutIVFS([0.5,0.5])”.
4. Applications
The main realms for the application of the IVFSs programming language
described in this paper are those which involve natural language semantics
processing. In this section, we will discuss two of them: linguistic knowledge
19
modelling and proximity-based logic programming using linguistic resources.
4.1. Linguistic Knowledge Modelling
Linguistic knowledge modelling handles the computational representation
of knowledge that is embedded in natural language. This framework can be
enhanced by combining multiadjoint paradigm with interval-valued fuzzy sets
[14]. For example, we can define interval-valued annotated atoms. Let us
assume the same definition of suitable journal given in [22], that is, a journal
with a high impact factor, a medium immediacy index, a relatively big halflife and with a not bad position in the listing of the category. Now, we
introduce in the program the following inference rule:
suitable_journal(X):-impact_factor(X)[0.8,0.9],
immediacy_index(X)[0.4,0.6],
cited_half_life(X)[0.6,0.7],
best_position(X)[0.4,0.6].
Now, let us suppose the IEEE Transactions of Fuzzy System journal has
the following properties: “high” impact factor, “small” immediacy index,
“relatively small” cited half life and the “best position”. Regarding the
linguistic variables: “high”, “medium”, “relatively big” and “not bad”, which
can be related to the following truth-values: [0.8, 0.9], [0.4, 0.6], [0.6, 0.7] and
[0.4, 0.6], respectively, considering the variables “medium” and “not a bad”
with a similar meaning. This knowledge could be model in an interval-valued
fuzzy logic language as follows:
20
%% high impact factor
impact_factor(ieee_fs)[0.8,0.9]
%% small immediacy index
immediacy_index(ieee_fs) [0.3,0.5]
%% relatively small
cited_half_life(ieee_fs)[0.3,0.5]
%% best position
best_position(ieee_fs) [1,1]
When the query “suitable journal(X)” is launched, then the system answers: “X = ieee f s′′ with[0.3, 0.6].
4.2. Proximity-based Logic Programming based on WordNet
Proximity-based Logic Programming is a framework that provides us with
the capability of enriching semantically classical logic programming languages
by using Proximity Equations (PEs). A limitation of this approach is that
PEs are mostly defined for a specific domain [6, 23], being the designer who
manually fixes the values of these equations. This fact makes harder to use
PLP systems in real applications.
A possible solution consists in obtaining the proximity equations from
WordNet which requires to employ interval-valued fuzzy sets in order to
deal with the high uncertainty generated by the possibility of using several
21
different semantic similarity metrics. Let us assume a fragment of a deductive
database that stores information about people and their preferences. The
proximity equations can be generated from WordNet, we only put here some
of them (see [4] for more detail).
%% m loves mountaineering
loves(mary,mountaineering).
%% j likes football
likes(john,football).
%% peter
plays basketball
plays(peter,basketball).
%% if a person practises sports
%% the he/she is a healthy person
healthy(X):- practices(X,sport).
%%automatically generated from wordnet
love~passion=[0.25,0.8].
basketball~hoops=[1,1].
play~act=[0.25,0.7].
practice~rehearse=[1,1].
sport~variation=[0.1,0.5].
sport~fun=[0.3,0.8].
22
5. Related Work
In the literature, other proposals that address our same goal can be
found [24, 25]. One of the most relevant ones is Ciao-Prolog [25] and, for that
reason, we will analyse in detail the differences between it and Bousi-Prolog
in order to clarify and reinforce the novelty of our proposal:
• From the point of view of its implementation. In Ciao-Prolog,
IVFSs are included by means of constrains and hence a translator must
be implemented. As a result, the programmer must code the variables
in order to manage the truth values and get the answers from the system based on those constraints. In Bousi-Prolog, on the other hand,
IVFSs are included in a different way, where the compiler and the warren abstract machine are enhanced by using a IVFSs data structure
which has been created and adapted for this architecture. As a result, intervals work as a standard data structure in the code of the
program instead of a particular set of variables defined ad hoc by the
programmer. This feature allows us to include IVFSs in both fuzzy
unification (see [4]) and fuzzy resolution. In addition, this framework
also allows other possible extensions, such as the incorporation of a
reasoning module using WordNet (see [4]).
• From the point of view of its syntax. Ciao-Prolog and BousiProlog, although both are Prolog languages, they have a well differentiated syntax. The former only allows the annotation of facts, rules
23
cannot be annotated because these only allow the use of an aggregator
operator for the computing of the annotated IVFSs. The latter, on the
other hand, allows the user both the annotation the fact and rules by
means of IVFSs. In addition, if we focus on the inference engine, while
Ciao-Prolog only extends the resolution mechanism, Bousi-Prolog uses
interval-valued proximity equations (e.g., young teenager=[0.6,0.8]),
which extends both the resolution and unification process.
• From the point of view of its semantics. Ciao-Prolog and BousiProlog have relevant differences at the semantic levels as well. Firstly,
Bousi-Prolog implements the concept of cut-level, which allows to the
user imposes a threshold in the system, and according to it you can be
as precise as you want in your answer. This is a substantial change due
to the introduction of a threshold operational semantics. Therefore,
our operational mechanism behaves very much as the one of a Prolog
system (obtaining correct answers one by one), while this option is not
available in Ciao semantics. As we mentioned in section 3.5, a λ-cut for
IVFS approximation degrees has been implemented. The concepts of
interpretation, least model semantics, model, so on, are presented and
defined in a different way, in Bousi-Prolog the operational semantics is
based an extension of SLD Resolution. In [25] the type of resolution is
based on the classical SLD Resolution of Prolog Systems.
24
6. Conclusions and future work
We have formally defined and efficiently implemented a simple intervalvalued fuzzy programming language using interval-valued fuzzy sets for modelling the uncertainty and imprecision of the knowledge associated to lexical
resources. As future work, we propose to extend our language and to provide
results of soundness and completeness. Additionally, we want to develop a
fully integrated framework in which interval-valued fuzzy sets and intervalvalued fuzzy relations can be combined in a same framework.
Acknowledgements
The authors gratefully acknowledges the comments made by reviewers. This work has
been partially supported by FEDER and the State Research Agency (AEI) of the Spanish
Ministry of Economy and Competition under grants TIN2016-76843-C4-2-R (AEI/FEDER,
UE) and TIN2014-56633-C3-1-R, the Consellerı́a de Cultura, Educación e Ordenación Universitaria (the Postdoctoral Training Grants 2016 and Centro singular de investigación
de Galicia accreditation 2016-2019, ED431G/08) and European Regional Development
Fund (ERDF). This work has been done in collaboration with the research group SOMOS
(SOftware-MOdelling-Science) funded by the Research Agency and the Graduate School
of Management of the Bı́o-Bı́o University.
References
[1] Miller, G.A.: Wordnet: A lexical database for english. Communications
of the ACM 38(11) (1995) 39–41
25
[2] Liu, H., Singh, P.: Commonsense reasoning in and over natural language. In Negoita, M.G., Howlett, R.J., Jain, L.C., eds.: Proceedings of
the 8th International Conference on Knowledge-Based Intelligent Information and Engineering Systems (KES 2004), Wellington, New Zealand,
September 20-25, Berlin, Heidelberg, Springer (2004) 293–306
[3] León-Aráuz, P., Gómez-Romero, J., Bobillo, F.: A fuzzy ontology extension of wordnet and eurowordnet for specialized knowledge. In: Proceedings of the TKE 2012. (2012)
[4] Rubio-Manzano, C., Pereira-Fariña, M.: Towards fuzzy lexical reasoning. Journal of Intelligent & Fuzzy Systems 32(3) (Feb 2017) 2425–2436
[5] Medina, J., Ojeda-Aciego, M.: Multi-adjoint logic programming. In:
Proceedings of the 10th International Conference on Information Processing and Managment of Uncertainty in Knowledge-Based Systems
(IPMU 2004). (2004) 823–830
[6] Julián-Iranzo, P., Moreno, G., Penabad, J., Vázquez, C. In: A Declarative Semantics for a Fuzzy Logic Language Managing Similarities and
Truth Degrees. Springer International Publishing, Cham (2016) 68–82
[7] Julián-Iranzo, P., Moreno, G., Penabad, J., Vázquez, C.: A fuzzy logic
programming environment for managing similarity and truth degrees.
(2015)
26
[8] Muñoz-Hernandez, S., Pablos-Ceruelo, V., Strass, H.: Rfuzzy: Syntax,
semantics and implementation details of a simple and expressive fuzzy
tool over prolog. Information Sciences 181(10) (2011) 1951 – 1970 Special Issue on Information Engineering Applications Based on Lattices.
[9] Straccia, U. In: Managing Uncertainty and Vagueness in Description
Logics, Logic Programs and Description Logic Programs. Springer Berlin
Heidelberg, Berlin, Heidelberg (2008) 54–103
[10] Le, V.H., Liu, F., Tran, D.K.: Fuzzy linguistic logic programming and
its applications. Theory and Practice of Logic Programming 9(03) (may
2009) 309–341
[11] Pedersen, T., Patwardhan, S., Michelizzi, J.:
Measuring the relatedness of concepts.
Wordnet::similarity:
In: Demonstration Papers
at HLT-NAACL 2004. HLT-NAACL–Demonstrations ’04, Stroudsburg,
PA, USA, Association for Computational Linguistics (2004) 38–41
[12] Turksen, I.:
Four methods of approximate reasoning with interval-
valued fuzzy sets. International Journal of Approximate Reasoning 3(2)
(mar 1989) 121–142
[13] Bustince, H.: Interval-valued fuzzy sets in soft computing. International
Journal of Computational Intelligence Systems 3(2) (2010) 215–222
[14] Medina, J.: Adjoint pairs on interval-valued fuzzy sets. In: Proceedings of the 13th International Conference on Information Processing and
27
Managment of Uncertainty in Knowledge-Based Systems (IPMU 2010),
Springer Berlin Heidelberg (2010) 430–439
[15] Rubio-Manzano, C., Julián-Iranzo, P.: A fuzzy linguistic prolog and
its applications. Journal of Intelligent & Fuzzy Systems 26(3) (2014)
1503–1516
[16] Julián-Iranzo, P., Rubio-Manzano, C. In: A Similarity-Based WAM
for Bousi˜Prolog. Springer Berlin Heidelberg, Berlin, Heidelberg (2009)
245–252
[17] Pereira-Fariña, M.:
Some reflections on the set-based and the
conditional-based interpretations of statements in syllogistic reasoning.
Archives for the Philosophy and History of Soft Computing 1/2014(1)
(2014) 1–16
[18] Hajek, P.: Metamathematics of fuzzy logic. Springer Science and Business Media (1998)
[19] Lloyd, J.: Foundations of Logic Programming. Springer-Verlag, Berlin
(1987)
[20] Julián-Iranzo, P., Rubio-Manzano, C.:
A declarative semantics for
bousi prolog. In: Proceedings of the 11th ACM SIGPLAN Conference
on Principles and Practice of Declarative Programming. PPDP ’09, New
York, NY, USA, ACM (2009) 149–160
28
[21] Ebrahim, R.: Fuzzy logic programming. Fuzzy Sets and Systems 117(2)
(2001) 215 – 230
[22] Medina, J., Ojeda-Aciego, M., Ruı́z-Calviño, J.: Formal concept analysis via multi-adjoint concept lattices. Fuzzy Sets and Systems 160(2)
(2009) 130 – 144
[23] Rodrı́guez-Artalejo, M., Romero-Dı́az, C.A.: A declarative semantics
for clp with qualification and proximity. Theory and Practice of Logic
Programming 10(4-6) (2010) 627642
[24] Atanassov, K., Georgiev, C.: Intuitionistic fuzzy prolog. Fuzzy Sets and
Systems 53(2) (1993) 121 – 128
[25] Guadarrama, S., Muoz, S., Vaucheret, C.: Fuzzy prolog: a new approach using soft constraints propagation. Fuzzy Sets and Systems
144(1) (2004) 127 – 150
29
| 2cs.AI
|
TROPICAL LIMITS OF PROBABILITY SPACES, PART I
arXiv:1704.00297v2 [math.MG] 16 Jun 2017
THE INTRINSIC KOLMOGOROV-SINAI DISTANCE AND THE
ASYMPTOTIC EQUIPARTITION PROPERTY FOR CONFIGURATIONS
R. MATVEEV AND J. W. PORTEGIES
Abstract. The entropy of a finite probability space X measures the observable cardinality of large independent products X ⊗n of the probability
space. If two probability spaces X and Y have the same entropy, there is an
almost measure-preserving bijection between large parts of X ⊗n and Y ⊗n .
In this way, X and Y are asymptotically equivalent.
It turns out to be challenging to generalize this notion of asymptotic
equivalence to configurations of probability spaces, which are collections of
probability spaces with measure-preserving maps between some of them.
In this article we introduce the intrinsic Kolmogorov-Sinai distance on
the space of configurations of probability spaces. Concentrating on the
large-scale geometry we pass to the asymptotic Kolmogorov-Sinai distance.
It induces an asymptotic equivalence relation on sequences of configurations
of probability spaces. We will call the equivalence classes tropical probability
spaces.
In this context we prove an Asymptotic Equipartition Property for configurations. It states that tropical configurations can always be approximated by homogeneous configurations. In addition, we show that the
solutions to certain Information-Optimization problems are Lipschitz-continuous with respect to the asymptotic Kolmogorov-Sinai distance. It follows from these two statements that in order to solve an InformationOptimization problem, it suffices to consider homogeneous configurations.
Finally, we show that spaces of trajectories of length n of certain stochastic processes, in particular stationary Markov chains, have a tropical
limit.
0. Introduction
The aim of the present article is to develop a theory of tropical probability spaces, which are asymptotic classes of finite probability spaces. Together
with the accompanying techniques, we expect them to be relevant to problems arising in information theory, causal inference, artificial intelligence and
neuroscience.
As a matter of introduction and motivation of the research presented in the
article, we start by considering a few simple examples.
0.1. Single probability spaces. We consider a finite probability space X =
(S, p), where S is a finite set, and p is a probability measure on S. For
simplicity, assume for now that the measure p has full support. Next, we
1
2
0. Introduction
consider the, so-called, Bernoulli sequence of probability spaces
X ⊗n = (S n , p⊗n )
where S n denotes the n-fold Cartesian product of S, and p⊗n is the n-fold
product measure.
This situation arises in several contexts. For example, in physics, X ⊗n would
encode the state of the system comprised of many identical non-interacting
(weakly interacting) subsystems with state space X. In information theory,
X ⊗n would describe the output of an i.i.d. random source. In dynamical systems or stochastic processes the setting corresponds to Bernoulli shifts and
Bernoulli processes.
The entropy of X is the exponential growth rate of the observable cardinality
of tensor powers of X. The observable cardinality, loosely speaking, is the
cardinality of the set X ⊗n after (the biggest possible) set of small measure
of elements, each with negligible measure, has been removed. It turns out
that the observable cardinality of X ⊗n might be much smaller than ∣S∣n , the
cardinality of the whole of X ⊗n , in the following sense.
The Asymptotic Equipartition Property states that for every ε > 0 and suf(n)
ficiently large n one can find a, so-called, typical subset Aε ⊂ S n , such that
it takes up almost all of the mass of X ⊗n and the probability distribution on
(n)
Aε is almost uniform on the normalized logarithmic scale,
(n)
(i) p⊗n (Aε ) ≥ 1 − ε
(n)
(ii) For any a, a′ ∈ Aε holds ∣ n1 ln p(a) − n1 ln p(a′ )∣ ≤ ε
(n)
The cardinality ∣Aε ∣ may be much smaller than ∣S∣n , but it will still grow
(n)
exponentially with n. Even though there are many choices for such a set Aε ,
the exponential growth rate with respect to n is well-defined upto 2ε. In fact,
(n)
there exists a number hX such that for any choice of the typical subset Aε
holds
(n)
en⋅hX −ε ≤ ∣Aε ∣ ≤ en⋅hX +ε
The limit of the growth rate as ε → 0+ is called the entropy of X, as explained
in more detail in Section 1.3
1
(n)
Ent(X) ∶= lim lim ln ∣Aε ∣
ε↓0 n→∞ n
By the law of large numbers
Ent(X) = hX = − ∑ p(x) ln p(x)
x∈S
which is the formula by which the Shannon entropy is usually introduced.
Entropy is especially easy to evaluate if the space is uniform, since for any
finite probability space with the uniform distribution holds
(0.1)
Ent(X) = ln ∣X∣
This point of view on entropy goes back to the original idea of Boltzmann,
according to which entropy is the logarithm of the number of equiprobable
3
0. Introduction
states, that a system, comprised of many identical weakly interacting subsystems, may take on.
0.1.1. Asymptotic equivalence. The Asymptotic Equipartion Property implies
that the sequence X ⊗n is asymptotically equivalent to a sequence of uniform
spaces in the following sense. Let us denote by pU the probability distribution
(n)
that is supported on Aε , and is uniform on its support. Then a sequence
from independent samples according to pU is very hard to discriminate from a
sequence of independent samples from p⊗n
X .
Similarly, when X and Y are probability spaces with the same entropy,
the sets X and Y are asymptotically equivalent in the sense that there is a
bijection between the typical sets, which can be seen as a change of code. This
is essentially the content of Shannon’s source coding theorem.
In [Gro12], Gromov proposed this existence of an “almost-bijection” as a
basis of an asymptotic equivalence relation on sequences of probability spaces.
Even though we were greatly influenced by ideas in [Gro12], we found that
Gromov’s definition does not extend easily to configurations of probability
spaces.
By a configuration of probability spaces we mean a collection of probability
spaces with measure-preserving maps between some of them. We will give
a precise definition in Section 1.2, but will consider some particular examples below. Formalizing and studying a notion of asymptotic equivalence for
configurations of probability spaces is the main topic of the present article.
0.2. Configurations of probability spaces. Suppose that now instead of a
single probability space, we consider a pair of probability spaces X = (X, pX )
and Y = (Y , pY ) with a joint distribution, that is a probability measure on
X × Y that pushes forward to pX and pY under coordinate projections. In
other words, we consider a triple of probability spaces X, Y and U with a pair
of measure-preserving maps U → X and U → Y . This is what we later call a
minimal two-fan of probability spaces
U
X
Y
and is a particular instance of a configuration of probability spaces.
0.2.1. Three examples. Three examples of such an object are shown on Figure
1, which is to be interpreted in the following way. Each of the spaces Xi
and Yi , i = 1, 2, 3, have cardinality six and a uniform distribution, where the
weight of each atom is 61 . The spaces Ui , i = 1, 2, 3, have cardinality 12 and
1
the distribution is also uniform with all weights being 12
. The support of the
measure on Ui ’s is colored grey on the pictures. The maps from Ui to Xi and
Yi are coordinate projections.
4
0. Introduction
U1
U2
Y1
U3
Y2
X1
Y3
X2
X3
Figure 1. Examples of pairs of probability spaces together with
a joint distribution.
In view of equation (0.1) we have for each i = 1, 2, 3,
Ent(Xi ) = ln 6
Ent(Yi ) = ln 6
Ent(Ui ) = ln 12
Now we would like to ask the following.
Question.
Is it possible to find an almost-bijection between sufficiently
high powers of (Xi⊗n ← Zi⊗n → Yi⊗n ) by (Xj⊗n ← Zj⊗n → Yj⊗n )
for i ≠ j with an arbitrary given precision as in Shannon’s
coding theorem as described at the end of the previous subsection 0.1? More generally, what is the proper generalization of
an asymptotic equivalence relation as discussed in the previous
subsection to sequences of tensor powers of two-fans?
We would like to argue that even though the entropies of the constituent
spaces are all (pairwise) the same, all three examples above should be pairwise
asymptotically different.
To establish that the examples in Figure 1 are different, that is, not isomorphic (see also Section 1.2) is relatively easy, since they have non-isomorphic
symmetry groups. However, we present a different argument, that lends itself
for generalization to prove that the examples at hand are not asymptotically
equivalent and that also gives a quantitative difference between them.
To distinguish Example 1 from both 2 and 3, one could argue along the
following lines. We could try to add a third space Z = (Z, pZ ) to the pair X
and Y and provide a joint distribution pQ on
Q = (X × Y × Z, pQ )
such that the projection of pQ on the first two factors is pU and on the third
factor is pZ .
Once we do that, we could evaluate entropies of various push-forwards of
pQ . Denote by V = (X × Z, pV ) and W = (Y × Z, pW ), where pV and pW
are push-forwards of pQ under corresponding coordinate projections. All the
5
0. Introduction
probability spaces now fit into a commutative diagram
Q
U
V
W
X
Y
Z
where each arrow is a reduction, which is simply a measure-preserving map
between probability spaces.
We consider the set of all possible extensions of the above form and denote
it by Ext(X, Y, U ). For any extension E = (X, Y, Z, U, V, W, Q) in Ext(X, Y, U )
we have four “new” entropies
Ent(Q),
(0.2)
Ent(V ),
Ent(W ),
Ent(Z)
in addition to the “known” entropies of X, Y and U . The vector
Ent∗ (E) ∶= ( Ent(X), Ent(Y ), Ent(Z), Ent(U ), Ent(V ), Ent(W ), Ent(Q))
is the entropy vector of the extension E.
The set of all possible values of the entropy vector for all extensions of
(X, Y, U )
Γ○ (X, Y, U ) ∶=
{Ent∗ (E) ∈ R7 ∣ E is an extension of (X, Y, U )} ⊂ R7
is what we call the unstabilized relative entropic set of the two-fan (X ← U →
Y ).
0.2.2. The unstabilized relative entropic sets for the examples. It turns out that
these unstabilized relative entropic sets of (X1 ← U1 → Y1 ) and (X2 ← U2 → Y2 )
are different
Γ○ (X1 , Y1 , U1 ) ≠ Γ○ (X2 , Y2 , U2 )
To see this, let us calculate some particular points in the unstabilized relative
entropic sets of the Examples 1–3. We consider the constrained InformationOptimization problem, of finding an extension E = (X, Y, Z, U, V, W, Q) of
(X, Y, U ) such that
(i) the space Z is a reduction of U , that is
Ent(Q) = Ent(U )
(ii) the spaces X and Y are independent conditioned on Z,
Ent(Q) + Ent(Z) = Ent(V ) + Ent(W )
(iii) the sum
Ent(X∣ Z) + Ent(Y ∣ Z)
is maximal, subject to conditions (i) and (ii).
6
0. Introduction
It is very easy to read the solutions Ê1 , Ê2 and Ê3 of this optimization
problem for Examples 1, 2 and 3 right from the pictures in Figure 1. Indeed,
condition (i) says that Zi must be a partition of Ui . Condition (ii) says that
each set in the partition must be “rectangular”, that is it must be a Cartesian
product of a subset of X i and a subset of Y i . The quantity to be maximized
is the average log-area of the sets in the partition.
Optima are very easy to find “by hand”. For Example 1 it is a partition of
U1 into three 2 × 2 squares. In Examples 2 and 3, one of the solutions is the
partition of U2 , (resp. U3 ) into 1 × 2 rectangles. Thus, the optimal values are
(2 ln 2) for Example 1, and (ln 2) for Examples 2 and 3.
0.2.3. The stabilized relative entropic set. We have just seen that Examples
1 and 2 can be told apart by determining the unstabilized relative entropic
set. However, this is not really what we are interested in. Rather, we wonder
whether high tensor powers can be distinguished this way.
This relates to why we used the adjective “unstabilized”: because the relative entropic set usually grows (is not stable) under taking tensor powers.
That is, for every n, k ∈ N it holds that
(0.3)
k ⋅ Γ○ (X ⊗n , Y ⊗n , U ⊗n ) ⊂ Γ○ (X ⊗(k⋅n) , Y ⊗(k⋅n) , U ⊗(k⋅n) )
but in general the set on the right-hand side can be strictly larger than the set
on the left.
In view of the inclusion (0.3) we may define the stabilized relative entropic
set
1
Γ(X, Y, U ) = Closure ( lim Γ○ (X ⊗n , Y ⊗n , U ⊗n ))
n→∞ n
This set turns out to be convex.
0.2.4. The stabilized relative entropic set for the examples. In fact, the stabilized relative entropic set also differentiates between Examples 1 and 2
Γ(X1 , Y1 , U1 ) ≠ Γ(X2 , Y2 , U2 )
The proof of this fact follows the same lines as in Section 0.2.2, but the stabilization makes the argument much more technical.
We expect that the stabilized relative entropic set cannot differentiate between Examples 2 and 3. However, there are other types of relative entropic
sets, and other Information-Optimization problems that can differentiate between Examples 2 and 3.
The relative entropic sets are discussed in Section 7.
0.3. Information-Optimization problems and relative entropic sets.
In Section 0.2.2 we used an Information-Optimization problem to find particular points in the (unstabilized) relative entropic set. This is no coincidence,
and the link between stabilized Information-Optimization problems and the
stable relative entropic set can be made very explicit. Because the stable relative entropic set is convex, it can be completely characterized by InformationOptimization problems and vice versa.
0. Introduction
7
Such Information-Optimization problems play a very important role in information theory [Yeu12], causal inference [SA15], artificial intelligence [VDP13],
information decomposition [BRO+ 14], robotics [ABD+ 08], and neuroscience
[Fri09]. The techniques developed in the article allow one to address this type
of problems easily and efficiently.
0.4. The intrinsic Kolmogorov-Sinai distance. As we mentioned at the
end of Section 0.1, one is tempted to define asymptotically equivalent configurations along the lines of Shannon’s source coding theorem following [Gro12].
Two configurations would be asymptotically equivalent if there is an almost
measure-preserving bijection between subspaces of almost full measure in their
high tensor powers.
However, we found this approach inconvenient. Instead of finding an almost
measure-preserving bijection between large parts of the two spaces, we consider
a stochastic coupling (transportation plan, joint distribution) between a pair
of spaces and measure its deviation from being an isomorphism of probability
spaces, that is a measure-preserving bijection. Such a measure of deviation
from being an isomorphism then leads to the notion of intrinsic KolmogorovSinai distance, and its stable version – the asymptotic Kolmogorov-Sinai distance, as explained in Section 4.
In the case of single probability spaces we define the intrinsic KolmogorovSinai distance between two probability spaces X = (X, pX ) and Y = (Y , pY )
by
k(X, Y ) ∶= inf {[ Ent(Z) − Ent(X)] + [ Ent(Z) − Ent(Y )]}
where the infimum is taken over all choices of the joint distribution Z = (X ×
Y , pZ ). Note that each of the summands is nonnegative and vanishes if and
only if the corresponding marginalization Z → X or Z → Y is an isomorphism
of probability spaces. In this sense the distance measures the deviation from
the existence of a measure-preserving bijection between X and Y .
Furthermore, we define the asymptotic Kolmogorov-Sinai distance between
two probability spaces X and Y by
1
κ(X, Y ) = lim k(X ⊗n , Y ⊗n ).
n→∞ n
This definition could be generalized to configurations of probability spaces
and we will say that two configurations are asymptotically equivalent if the
asymptotic Kolmogorov-Sinai distance between them vanishes.
0.5. Asymptotic Equipartition Property. Examples 1, 2, and 3 above
have the property that the symmetry group acts transitively on the support
of the measure on Ui and they are particular instances of what we call homogeneous configurations.
In Section 6, we show an Asymptotic Equipartion Property for configurations: Theorem 6.1 states that every sequence of tensor powers of a configuration can be approximated in the asymptotic Kolmogorov-Sinai distance by
a sequence of homogeneous configurations.
8
0. Introduction
This Asymptotic Equipartition Property allows one to substitute configurations of probability spaces by homogeneous approximations. Homogeneous
probability spaces are just uniform probability spaces, and as a first simple consequence of the Asymptotic Equipartition Property, the asymptotic
Kolmogorov-Sinai distance between probability spaces X and Y can be computed and equals
κ(X, Y ) = ∣ Ent(X) − Ent(Y )∣.
Homogeneous configurations are, unlike homogeneous probability spaces,
rather complex objects. Nonetheless, they seem to be simpler than arbitrary
configurations of probability spaces for the types of problems that we would
like to address.
More specifically, we show in Section 7 that the optimal values in (stabilized)
Information-Optimization problems only depend on the asymptotic class of a
configuration and that they are continuous with respect to the asymptotic
Kolmogorov-Sinai distance; in many cases, the optimizers are continuous as
well. The Asymptotic Equipartition Property implies that for the purposes of
calculating optimal values and approximate optimizers, one only needs to consider homogeneous configurations and this can greatly simplify computations.
Summarizing, the Asymptotic Equipartition Property and the continuity of
Information-Optimization problems are important justifications for the choice
of asymptotic equivalence relation and the introduction of the intrinsic and
asymptotic Kolmogorov-Sinai distances.
0.6. The article. The article has the following structure. Section 1 is devoted
to the basic setup used throughout the text. In Section 2 we explain what we
mean by configurations of probability spaces, give examples, describe simple
properties and operations. Further, in Section 3 we generalize the notion of
probability distribution to that on configurations and discuss the theory of
types for configurations. In Section 4 the intrinsic Kolmogorov-Sinai distance
and the asymptotic Kolmogorov-Sinai distance are introduced and some technical tools for the estimation of Kolmogorov distance are developed. Section 5
contains estimates on the distances between types. We use these estimates in
the proof of the Asymptotic Equipartition Property for configurations in Section 6. Section 7 deals with extensions of configurations. We prove there the
Extension Lemma, which is used to show continuity of extensions and implies,
in particular, that solutions of the constrained optimization problem for the
entropies of extensions are Lipschitz-continuous with respect to the asymptotic
Kolmogorov-Sinai distance, thus they only depend on the asymptotic classes of
configurations. In Section 8 we briefly discuss a special type of configurations
called mixtures, which will play an important role in the construction of tropical probability spaces. Finally, in Section 9 we introduce the notion of tropical
probability spaces and configurations thereof, and list some of their properties.
We will continue our study of tropical probability spaces and configurations in
subsequent articles.
1. Category
9
Some technical and not very illuminating proofs are deferred to Section T
Technical Proofs. In the electronic version one can move between the proof in
the technical section and the statement in the main text by following the link
(arrow up or down).
Acknowledgments. We would like to thank Tobias Fritz, František Matúš,
Misha Movshev and Johannes Rauh for inspiring discussions. We are grateful
to the participants of the Wednesday Morning Session at the CASA group at
the Eindhoven University of Technology for valuable feedback on the introduction of the article. Finally, we thank the Max Planck Institute for Mathematics
in the Sciences, Leipzig, for its hospitality.
1. Category of probability spaces and configurations
This section is devoted to the basic setup used throughout the present article. We introduce a category of probability spaces and reductions, similar
to categories introduced in [BFL11] and [Gro12], and define configurations of
probability spaces and the corresponding category. The last subsection recalls
the notion of entropy and its elementary properties.
1.1. Probability spaces and reductions. Below we will consider probability spaces such that the support of the probability measure is finite. Any such
space contains a full-measure subspace isomorphic to a finite space, thus we
call such objects finite probability spaces. For a probability space X = (S, pX )
denote by X = supp pX the support of the measure and by ∣X∣ its cardinality.
Slightly abusing the language, we call this quantity the cardinality of X.
For a pair of probability spaces a reduction X → Y is a class of measurepreserving maps, with two maps being equivalent if they coincide on a set of
full measure. The composition of two reductions is itself a reduction. Two
probability spaces are isomorphic if there is a measure-preserving bijection
between the supports of the probability measures. Such a bijection defines an
invertible reduction from one space into another. Clearly the cardinality ∣X∣
is an isomorphism invariant. The automorphism group Aut(X) is the group
of all self-isomorphisms of X.
A probability space X is called homogeneous if the automorphism group
Aut(X) acts transitively on the support X of the measure. The property of
being homogeneous is an isomorphism invariant. In the isomorphism class of
a homogeneous space there is a representative with uniform measure.
The finite probability spaces and reductions form a category, that we denote
by Prob. The subcategory of homogeneous spaces will be denoted by Probh .
The isomorphism in the category coincides with the notion of isomorphism
above.
The category Prob is not a small category. However it has a small full
subcategory, that contains an object for every isomorphism class in Prob and
for every pair of objects in it, it contains all the available morphisms between
10
1. Category
them. From now on we imagine that such a subcategory was chosen and fixed
and replaces Prob in all considerations below.
There is a product in Prob given by the Cartesian product of probability
spaces, that we will denote by X ⊗ Y ∶= (X × Y , pX ⊗ pY ). There are canonical
reductions X ⊗ Y → X and X ⊗ Y → Y given by projections to factors. For
a pair of reductions fi ∶ Xi → Yi , i = 1, 2 their tensor product is the reduction
f1 ⊗ f2 ∶ X1 ⊗ X2 → Y1 ⊗ Y2 , which is equal to the class of the Cartesian product
of maps representing fi ’s. The tensor product is not however a categorical
product. The product leaves the subcategory of homogeneous spaces invariant.
The probability measure on X will be usually denoted by pX or simply p,
when the risk of confusion is low.
1.2. Configurations of probability spaces. Essentially, a configuration
X = {Xi ; fij } is a commutative diagram consisting of a finite number of probability spaces and reductions between some of them, that is transitively closed,
while a morphism ρ ∶ X → Y between two configurations X = {Xi ; fij } and
Y = {Yi ; gij } of the same combinatorial type is a collection of reductions between corresponding individual objects ρi ∶ Xi → Yi , that commute with the
reductions within each configuration, ρj ○ fij = gij ○ ρi .
We need to keep track of the combinatorial structure of the collection of
reductions within a configuration. There are several possibilities for doing so:
● the reductions form a directed transitively closed graph without loops;
● the spaces in the configuration form a poset;
● the underlying combinatorial structure could be recorded as a finite
category.
The last option seems to be most convenient since it has many operations
necessary for our analysis already built-in.
A diagram category G is a finite category such that for each pair of objects
O1 , O2 in G the morphism space between them
HomG (O1 , O2 ) ∪ HomG (O2 , O1 )
contains at most one element.
For a diagram category G a configuration of probability spaces modeled on G
is a functor X ∶ G → Prob. The collection of all configurations of probability
spaces modeled on a fixed diagram category G forms the category of functors
Prob ⟨G⟩ ∶= [G, Prob]. The objects of Prob ⟨G⟩ are configurations, that is
functors from G to Prob, while morphisms in Prob ⟨G⟩ are natural transformations between them. For a configuration X ∈ Prob ⟨G⟩, the diagram
category G will be called the combinatorial type of X .
For a diagram category G or a configuration X ∈ Prob ⟨G⟩ we denote by
[[G]] = [[X ]] the number of objects in the category G.
An object O in a diagram category G will be called initial, if it is not a
target of any morphism except for the identity. Likewise a terminal object is
not a source of any morphism, except for the identity morphism. Note that this
11
1. Category
terminology is somewhat unconventional from the point of view of category
theory.
A diagram category is called complete if it has a unique initial object. Thus
a configuration modeled on a complete category includes a space that reduces
to all other spaces in the configuration.
The above terminology transfers to configurations modeled on G: An initial
space in X ∈ Prob ⟨G⟩ is one that is not a target space of any reduction within
the configuration, a terminal space is not a source of any non-trivial reduction
and X is complete if G is, that is there is a unique initial space.
The tensor product of probability spaces extends to a tensor product of
configurations. For X , Y ∈ Prob ⟨G⟩, such that X = {Xi ; fij } and Y = {Yi ; gij }
define
X ⊗ Y ∶= {Xi ⊗ Yi ; fij ⊗ gij }
Occasionally we will also talk about configuration of sets. Denote by Set the
category of finite sets and surjective maps. Then all of the above constructions
could be repeated for sets instead of probability spaces. Thus we could talk
about the category of configurations of sets Set ⟨G⟩.
Given a reduction f ∶ X → Y between two probability spaces, the restriction
f ∶ X → Y is a well-defined surjective map. Given a configuration X = {Xi ; fij }
of probability spaces, there is an underlying configuration of sets, obtained by
taking the supports of measures on each level and restricting reductions on
these supports. We will denote it by X = {X i ; fij }, where X i ∶= supp pXi .
Thus we have a forgetful functor
⋅ ∶ Prob ⟨G⟩ → Set ⟨G⟩
For now we will consider two important examples of diagram categories and
configurations modeled on them. We give further examples in Section 2.1.
1.2.1. Two-fans: A two-fan is a configuration modeled on the category Λ with
three objects, one initial and two terminal.
Λ = (O1 ← O12 → O2 )
There is a special significance to two-fans, since these are the simplest nontrivial configurations, as we will see later.
Essentially, a two-fan X ← Z → Y is a triple of probability spaces and a pair
of reductions between them.
A reduction of a two-fan X ← Z → Y to another two-fan X ′ ← Z ′ → Y ′ is
a triple of reductions Z → Z ′ , Y → Y ′ and X → X ′ that commute with the
reductions within each fan, that is, the following diagram is commutative
X
Z
Y
X′
Z′
Y′
12
1. Category
Isomorphisms and the automorphism group Aut(⋅) are defined accordingly.
Note that terminal spaces in a two-fan are labeled and reductions preserve the
labeling.
A two-fan X ← Z → Y is called minimal if for a.e. x ∈ X and y ∈ Y there
is a unique z ∈ Z, that reduces to y and to x. Given a two-fan X ← Z → Y ,
there is always a reduction to a minimal two-fan X ← Z ′ → Y . Such minimal
reduction is unique up to isomorphism. Explicitly, take Z ′ ∶= X × Y as a set
and consider a probability distribution on Z ′ induced by a map Z → Z ′ which
is the Cartesian product of the reductions X ← Z → Y in the original two-fan.
The notion of being minimal is in fact a categorical notion. It could be
equivalently defined by saying that a two-fan X = (X1 ← X12 → X2 ) is minimal
if for any reduction λ ∶ X → X ′ holds: if both λ1 and λ2 are isomorphisms, then
λ12 is also an isomorphism. Consequently, if one specifies reductions from the
terminal spaces of a minimal two-fan to another two-fan, then there exists at
most one extension to the reduction of the whole fan.
The inclusion of a pair of probability spaces X and Y as terminal vertices
in a minimal two-fan is equivalent to specifying a joint distribution on X × Y .
An arbitrary configuration X will be called minimal if with every two-fan,
it also contains a minimal two-fan with the same terminal spaces. We will
denote the space of minimal configurations modelled on a diagram category G
by Prob ⟨G⟩m .
Given a two-fan
F = (X ← Z → Y )
with terminal spaces X and Y , and a point x ∈ X with pX (x) > 0, one may
construct a conditional probability distribution pY ( ⋅ ∣ x) on Y . We denote the
corresponding space Y ∣ x ∶= (Y , pY ( ⋅ ∣ x)). The usual bar “∣”, that is normally
used for conditioning, interferes with our notations for cardinality of spaces.
We will give more details in Section 2.7.
1.2.2. A diamond configuration. A “diamond” configuration is modeled on a
diamond category ◇, that consists of a two-fan and a “co-fan”:
⎛
⎜
⎜
⎜
◇ = ⎜ O1
⎜
⎜
⎜
⎝
O12
O●
⎞
⎟
⎟
⎟
O2 ⎟
⎟
⎟
⎟
⎠
Of course, there is also a morphism O12 → O● , which lies in the transitive
closure of the given four morphisms. As a rule, we will skip writing morphisms,
that are implied by the transitive closure.
A diamond configuration is minimal if the top two-fan in it is minimal.
1.3. Entropy. Our working definition of entropy will be based on the following version of the asymptotic equipartition theorem for Bernoulli process, see
[CT91].
13
2. Configurations
Theorem 1.1. Suppose X is a finite probability space, then for any ε > 0 and
(n)
any n >> 0 there exists a subset Aε ⊂ X ⊗n such that
(n)
(i) p(Aε ) ≥ 1 − ε
(n)
(ii) For any a, a′ ∈ Aε holds
ln p(a) ln p(a′ )
∣
−
∣≤ε
n
n
(n)
(n)
Moreover, if Aε and Bε are two subsets of X ⊗n satisfying two conditions
above, then their cardinalities satisfy
RRR
(n) R
(n)
RRR ln ∣Aε ∣ − ln ∣Bε ∣ RRRRR ≤ 2ε
(1.1)
RRR n
n RRRR
RR
R
⊠
Then we define
1
(n)
ln ∣Aε ∣
ε→0 n→∞ n
Clearly, in view of the property (1.1) in the Theorem 1.1 the limit above is
(n)
well-defined and is independent of the choice of the typical subsets Aε .
Entropy satisfies the so-called Shannon inequality, see for example [CT91],
namely for any minimal diamond configuration
Ent(X) ∶= lim lim
X12
X1
X2
X●
the following inequality holds,
(1.2)
Ent(X1 ) + Ent(X2 ) ≥ Ent(X12 ) + Ent(X● )
Furthermore, entropy is additive with respect to the tensor product, that is,
for a pair of probability spaces X, Y ∈ Prob holds
(1.3)
Ent(X ⊗ Y ) = Ent(X) + Ent(Y )
Further, for a pair X, Y of probability spaces included in a minimal two-fan
(X ← Z → Y ) we define the conditional entropy
Ent(X∣ Y ) ∶= Ent(Z) − Ent(X)
The above quantity is always non-negative in view of Shannon inequality (1.2). Moreover, the following identity holds, see [CT91]
(1.4)
Ent(X∣ Y ) = ∫ Ent(X∣ y) d pY (y)
Y
2. Configurations
In this section we will look at configurations in more detail. We start by
considering some important examples.
14
2. Configurations
2.1. Examples of configurations.
2.1.1. Singleton. We denote by ● a diagram category with a single object.
Clearly configurations modeled on ● are just probability spaces and we have
Prob ≡ Prob ⟨●⟩.
2.1.2. Chains. The chain Cn of length n ∈ N is a category with n objects
n
{Oi }i=1 and morphisms from Oi to Oj whenever i ≤ j. A configuration X ∈
Prob ⟨Cn ⟩ is a chain of reductions
X = (X1 → X2 → ⋯ → Xn )
2.1.3. Two-fan. The two-fan Λ is a category with three objects {O1 , O12 , O2 }
and two non-identity morphisms O12 → O1 and O12 → O2 . See also Section 1.2.1. Two-fans are the simplest configurations for which asymptotic
equivalence classes contain more information than just entropies of the entries.
Recall that a fan (X ← Z → Y ) is called minimal, if for any pair of points
x ∈ X and y ∈ Y with positive weights there exists at most one z ∈ Z, that
reduces to x and to y. Equivalently, for any super-configuration
Z
Z′
X
Y
the reduction Z → Z ′ must be an isomorphism.
2.1.4. Full configuration. The full category Λn on n objects is a category with
objects {OI }I∈2{1,...,n} ∖{∅} indexed by all non-empty subsets I ∈ 2{1,...,n} and a
morphism from OI to OJ , whenever J ⊆ I.
For a collection of random variables X1 , . . . , Xn one may construct a minimal
full configuration X ∈ Prob ⟨Λn ⟩ by considering all joint distributions and
“marginalization” reductions. We denote such a configuration by ⟨X1 , . . . , Xn ⟩.
On the other hand, the terminal vertices of a full configuration can be viewed
as random variables on the domain of definition given by the (unique) initial
space.
Suppose X ∈ Prob ⟨Λn ⟩ is a minimal full configuration with terminal vertices
X1 , . . . , Xn . It is convenient to view X as a distribution on the Cartesian
product of the underlying sets of the terminal vertices:
pX ∈ ∆(X 1 × ⋯ × X n )
Once the underlying sets of the terminal spaces are fixed, there is a one-to-one
correspondence between the full minimal configurations and distributions as
above.
15
2. Configurations
2.1.5. “Two-tents” configuration. The “two-tents” category M2 consists of five
objects, of which two are initial and three are terminal, and morphisms are as
follows
⎛ O
O23 ⎞
⎜ 12
⎟
⎟
M2 = ⎜
⎜
⎟
⎜
⎟
⎝ O1
O2
O3 ⎠
Thus, a typical two-tents configuration consists of five probability spaces
and reduction as in
X = (X ← U → Y ← V → Z)
The probability spaces U and V are initial and X, Y and Z are terminal.
2.1.6. “Many-tents” configuration. The previous example could be generalized
to a “many-tents” category
Mn = (O1 ← O12 → O2 ← ⋯ → On−1 ← On−1,n → On )
2.1.7. “Fence” configuration. The “fence” category W3 consists of six objects
and the morphisms are
⎛
⎜ O12
W3 = ⎜
⎜
⎝ O1
O13
O2
⎞
O23 ⎟
⎟
⎟
O3 ⎠
2.1.8. Co-fan. A co-fan V is a category with three objects and morphisms as
in the diagram
⎛
⎞
O2 ⎟
⎜ O1
⎟
V=⎜
⎜
⎟
O●
⎝
⎠
2.1.9. “Diamond” configurations. A “diamond” configuration ◇ is modeled on
a diamond category that consists of a fan and a co-fan
⎛
⎜
⎜
⎜
◇ = ⎜ O1
⎜
⎜
⎜
⎝
O12
O●
⎞
⎟
⎟
⎟
O2 ⎟
⎟
⎟
⎟
⎠
See also Section 1.2.2.
Examples 2.1.1, 2.1.2, 2.1.3, 2.1.4 and 2.1.9 are complete. Examples 2.1.1,
2.1.2 and 2.1.8 do not contain a two-fan. Tropical limits of such configurations
are very simple. Essentially, such tropical limits correspond to the tuple of
numbers corresponding to the entropies of the constituent spaces. Therefore,
we call configurations not containing a two-fan simple.
16
2. Configurations
2.2. Constant configurations. Suppose X is a probability space and G is
a diagram category. One may form a constant G-configuration by considering
a functor that maps all objects in G to X and all the morphisms to the
identity morphism X → X. We denote such a constant configuration by X G
or simply by X, when G is clear from the context. Any constant configuration
is automatically minimal.
If Y = {Yi ; fij } is another G-configuration, then a reduction ρ ∶ Y → X G
(which we write sometimes simply as ρ ∶ Y → X) is a collection of reductions
ρi ∶ Yi → X, such that
fij ○ ρi = ρj
2.3. Configurations of configurations. Of course, the operation of “configuration” could be iterated, so given a pair G1 , G2 of diagram categories
we could form a G2 -configuration of G1 -configurations, so we could speak, for
example, about a two-fan of configurations of the same type.
Prob ⟨G1 , G2 ⟩ ∶= Prob ⟨G1 ⟩ ⟨G2 ⟩ = Prob ⟨G1 ◻ G2 ⟩
where G1 ◻ G2 is the “Cartesian product of graphs” (as every diagram category could be considered as a transitively closed directed graph). This operation is commutative, thus, for example, a two-fan of G-configurations is a
G-configuration of two-fans.
We will rarely need anything beyond a two-fan of configurations.
A two-fan F = (X ← Z → Y) of G-configurations is called minimal if in any
extension of F of the form
Z
f
Z′
X
Y
the reduction f ∶ Z → Z ′ must be an isomorphism of G-configurations.
Recall that a two-fan of G-configurations could also be viewed as a Gconfiguration of two-fans of probability spaces. In the following lemma we
show that in order to verify the minimality of a two-fan of configurations it is
sufficient to check the minimality of all the constituent two-fans.
Lemma 2.1. ↓ Let G be a diagram category. Then
(i) A two-fan F = (X ← Z → Y) of G-configurations is minimal, if and
only if the constituent two-fans of probability spaces Fi = (Xi ← Zi →
Yi ) are all minimal.
(ii) For any two-fan F = (X ← Z → Y) of G-configurations its minimal
reduction exists, that is, there exists a minimal two-fan F ′ = (X ←
17
2. Configurations
Z ′ → Y) included in the following diagram
Z
Z′
X
Y
⊠
Even though this lemma is rather elementary, there are many similar statements that are not true. Thus we are compelled to provide a proof, which can
be found in Section T on page 59.
Similarly, a full configuration F of G-configurations is called minimal if for
every two-fan of G-configurations in F there is a minimal two-fan in F of
G-configurations with the same terminal configurations.
Lemma 2.1 has the following corollary and counterpart for full configurations
of G-configurations.
Corollary 2.2. Let G be a diagram category. Then
(i) A full configuration F of G-configurations is minimal, if and only if the
constituent full configurations of probability spaces Fi are all minimal.
(ii) For any full configuration F of G-configurations its minimal reduction
exists.
⊠
2.4. Restrictions and extensions. Suppose R ∶ G1 → G2 is a functor between two diagram categories. For a configuration X ∶ G2 → Prob, the pullback configuration Y = R∗ X ∈ Prob ⟨G1 ⟩ defined as the composition
Y ∶= X ○ R
is called an R-restriction of X to G1 and X is the extension of Y. If the functor
R is injective then we call Y = R∗ X a sub-configuration of X and write Y ⊂ X ,
likewise X will be called a super-configuration of Y.
The restriction operation is functorial in the sense that given two configurations X , X ′ ∈ Prob ⟨G1 ⟩ and a reduction f ∶ X → X ′ , there is a canonical
reduction R∗ f ∶ R∗ X → R∗ X ′ . Thus R∗ can be considered as a functor
R∗ ∶ Prob ⟨G2 ⟩ → Prob ⟨G1 ⟩
Some important examples of restrictions and extensions are below.
2.4.1. Restriction of a full configuration to a smaller full configuration. Recall
that, as explained in Section 2.1.4, the terminal vertices of a full configuration
could be considered as random variables and any collection of random variables
“generates” a full configuration.
For a full configuration X = ⟨Xi ⟩ni=1 and a subset I ⊂ {1, . . . , n} we denote
by RI∗ X = ⟨Xi ⟩i∈I the restriction of X to a full configuration generated by Xi ,
18
2. Configurations
∗
for the restriction operator
i ∈ I. We will also make use of the notation Rk,l
∗
R{1,...,k} ∶ Prob ⟨Λl ⟩ → Prob ⟨Λk ⟩.
2.4.2. Restriction of a Λ3 -configuration to an M2 -configuration. Given a full
configuration X ∈ Prob ⟨Λ3 ⟩ we may “forget” part of the data. If we, for
example, forget the top space and the relation between a pair out of three
terminal spaces we end up with the two-tents configuration. This operation
corresponds to the inclusion functor
⎛ O
⎜ 12
M ∶⎜
⎜
⎜
⎝ O1
O23
O2
O3
⎛
⎜
⎜
⎞
⎜
⎜
⎟
⎟ Ð→ ⎜
⎜ Q12
⎟
⎜
⎟
⎜
⎜
⎠
⎜
⎜
⎝ Q1
Q123
Q31
Q23
Q2
Q3
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
that preserves the sub-indices.
We show in Section 2.5 below that the corresponding restriction operator
M ∗ ∶ Prob ⟨Λ3 ⟩ → Prob ⟨M3 ⟩
is surjective, both on objects and all morphisms. Thus, as a map of collections
of objects it has a right inverse. However, no natural right inverse exists.
2.4.3. Restriction of a Λ3 -configuration to a W3 -configuration. Starting with
a full configuration we might choose to forget the initial space (and reductions,
for which it was the domain). The remaining configuration has the combinatorial type of a fence. This operation corresponds to the functor
⎛ O
⎜ 12
W ∶⎜
⎜
⎜
⎝ O1
O31
O2
O23
O3
⎛
⎜
⎜
⎞
⎜
⎜
⎟
⎟ Ð→ ⎜
⎜ Q12
⎟
⎜
⎟
⎜
⎜
⎠
⎜
⎜
⎝ Q1
Q123
Q31
Q23
Q2
Q3
⎞
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎠
The corresponding operator
W ∗ ∶ Prob ⟨Λ3 ⟩ → Prob ⟨W3 ⟩
is not surjective. To find out when a W3 -configuration is extendable to a
Λ3 -configuration is an interesting problem, see for example, [ABK+ 15] and
references therein. It is our hope that the methods developed in this article
might be useful to address these questions.
19
2. Configurations
2.4.4. Doubling. This will be the first example of an interesting functor between diagram categories, which is not injective. In this situation the term
“restriction” does not really reflect the operation of pull-back well, however
we did not come up with a better terminology.
The doubling operation is the restriction of a two-fan to a two-tents configuration. Consider the two-fan category Λ = (O1 ← O12 → O2 ) and a two-tents
category M2 = (Q1 ← Q12 → Q2 ← Q23 → Q3 ). Define the functor D ∶ M2 → Λ
by setting
⎧
⎪
Q1 ↦ O1
⎪
⎪
⎪
⎪
⎪
⎪
Q2 ↦ O2
⎪
⎪
⎪
D ∶ ⎨Q3 ↦ O1
⎪
⎪
⎪
⎪
Q12 ↦ O12
⎪
⎪
⎪
⎪
⎪
⎪
⎩Q23 ↦ O12
Note that D extends uniquely to the spaces of morphisms, since each morphism space is either empty or a one-point set.
Thus D∗ (X ← Z → Y ) = (X ← Z → Y ← Z → X), where the “left” and
“right” two-fans are isomorphic.
This operation along with a particular Information-Optimization problem
is related to the so-called copy operation, that was used to find many nonShannon information inequalities, as described, for example, in [DFZ11].
2.5. Adhesion. Given a minimal two-tents configuration X = (X ← U → Y ←
V → Z) ∈ Prob ⟨M2 ⟩m one could always construct an extension of X to a
full configuration ad(X ) ∈ Prob ⟨Λ3 ⟩m in the following way: As explained in
Section 2.1.4, to construct a minimal full configuration with terminal vertices
X, Y and Z it is sufficient to provide a distribution on X × Y × Z with the
correct marginals. We do this by setting
p(x, y, z) ∶=
pU (x, y) ⋅ pV (y, z)
pY (y)
It is straightforward to check that the appropriate restriction of the full configuration defined in the above manner is indeed the original two-tents configuration. Essentially, to extend we need to provide a relationship (coupling)
between spaces X and Z and we do it by declaring X and Z independent
relative to Y . This is an instance of operation called adhesion, see [Mat07].
If we call the top vertex in the full configuration W , the entropies achieve
equality in the Shannon inequality, that is
Ent(U ) + Ent(V ) − Ent(W ) − Ent(Y ) = 0.
Adhesion provides a right inverse ad to the restriction functor M ∗ described
in Section 2.4.2
∗
Prob ⟨Λ3 ⟩m M Prob ⟨M2 ⟩m
ad
20
2. Configurations
Figure 2. Examples of homogeneous configurations
It is important to note though, that the map ad is not functorial and, in
fact, no functorial inverse of M ∗ exists.
2.6. Homogeneous configurations. A configuration X ∈ Prob ⟨G⟩ modeled on some diagram category G is called homogeneous if its automorphism
group Aut(X ) acts transitively on every probability space in X . Three examples of homogeneous configurations were given in the introduction. Other examples of a homogeneous configurations (of combinatorial type Λ3 ) are shown
in Figure 2. The subcategory of all homogeneous configurations modeled on
G will be denoted Prob ⟨G⟩h .
In fact, for X to be homogeneous it is sufficient that the Aut(X ) acts transitively on every initial space in X . Thus, if X is complete with initial space
X0 , to check homogeneity it is sufficient to check the transitivity of the action
of the symmetries of X on X0 .
By functoriality of the restriction operator, any restriction of a homogeneous
configuration is also homogeneous. In other words, if R ∶ G → G′ is a functor
21
2. Configurations
and
R∗ ∶ Prob ⟨G′ ⟩ → Prob ⟨G⟩
is the associated restriction operator, then
R∗ ( Prob ⟨G′ ⟩h ) ⊂ Prob ⟨G⟩h
In particular, all the individual spaces of a homogeneous configuration are
homogeneous
Prob ⟨G⟩h ⊂ Probh ⟨G⟩
However homogeneity of the whole of the configuration is a stronger property
than homogeneity of the individual spaces in the configuration, thus in general
Prob ⟨G⟩h ⊊ Probh ⟨G⟩
A single probability space is homogeneous if and only if there is a representative in its isomorphism class with uniform measure and the same holds true
for chain configurations, for the co-fan or any other configuration that does
not contain a two-fan. However, for more complex configurations, for example
for two-fans, no such simple description is available.
2.6.1. Universal construction of homogeneous configurations. Examples of homogeneous configurations could be constructed in the following manner. Suppose Γ is a finite group and {Hi } is a collection of subgroups. Consider a
collection of sets X i ∶= Γ/Hi and consider a natural surjection fij ∶ X i → X j
whenever Hi is a subgroup of Hj . Equipping each X i with the uniform distribution one can turn the configuration of sets {X i ; fij } into a homogeneous
configuration. It will be complete if there is a smallest subgroup (under inclusion) among Hi ’s.
Such a configuration will be complete and minimal, if together with any pair
of groups Hi and Hj in the collection, their intersection Hi ∩ Hj also belongs
to the collection {Hi }.
In fact, any homogeneous configuration arises this way. Suppose configuration X = {Xi ; fij } is homogeneous, then we set Γ = Aut(X ) and choose a collection of points xi ∈ Xi such that fij (xi ) = xj and denote by Hi ∶= Stab(xi ) ⊂ Γ.
Then, if one applies the construction of the previous paragraph to Γ, with the
collection of subgroups {Hi }, one recovers the original configuration X .
2.7. Conditioning. Suppose a configuration X contains a fan
f
g
F = (X ←Ð Z Ð→ Y )
Given a point x ∈ X with a non-zero weight one may consider conditional
probability distributions pZ ( ⋅ ∣ x) on Z, and pY ( ⋅ ∣ x) on Y . The distribution
pZ ( ⋅ ∣ x) is supported on f −1 (x) and is given by
pZ (z∣ x) =
pZ (z)
pX (x)
22
3. Distributions and types
The distribution pY ( ⋅ ∣ x) is the pushforward of pZ ( ⋅ ∣ x) under g
pY ( ⋅ ∣ x) = g∗ pZ ( ⋅ ∣ x)
Recall that if F is minimal, the underlying set of Z can be assumed to be the
product X × Y . In that case
pY (y∣ x) =
pZ (x, y)
pX (x)
We denote the corresponding space Y ∣ x ∶= (Y , pY ( ⋅ ∣ x)), as discussed at
the end of Section 1.2.1.
Under some assumptions it is possible to condition a whole sub-configuration
of X . More specifically, if a configuration X contains a sub-configuration Y
and a probability space X satisfying the condition that
for every Y in Y there is a fan in X with terminal vertices X
and Y ,
then we may condition the whole of Y on x ∈ X given that pX (x) > 0.
For x ∈ X with positive weight we denote by Y∣ x the configuration of spaces
in Y conditioned on x ∈ X. The configuration Y∣ x has the same combinatorial
type as Y and will be called the slice of Y over x ∈ X. Note that the space X
itself may or may not belong to Y. The conditioning Y∣ x may depend on the
choice of a fan between Y and X, however when X is complete the conditioning
Y∣ x is well-defined and is independent of the choice of fans.
Suppose now that there are two subconfiguration Y and Z in X and in
′
addition Z is a constant configuration, Z = Z G for some diagram category
G′ . Let z ∈ Z, then Y∣ z is well defined and is independent of the choice of the
space in Z, the element of which z is to be considered.
If X is homogeneous, then Y∣ x is also homogeneous and its isomorphism
class does not depend on the choice of x ∈ X.
2.8. Entropy. For a G-configuration X = {Xi , fij } define the entropy function
Ent∗ ∶ Prob ⟨G⟩ → R[[G]] ,
Ent∗ ∶ X = {Xi , fij } ↦ ( Ent(Xi )) ∈ R[[G]]
It will be convenient for us to equip the target R[[G]] with the `1 -norm. Thus
[[G]]
∣ Ent∗ (X )∣1 = ∑ Ent(Xi )
i=1
If X is a complete G-configuration with initial space X0 , then by Shannon
inequality (1.2) there is an obvious estimate
Ent(X0 ) ≤ ∣ Ent∗ (X )∣1 ≤ [[X ]] ⋅ Ent(X0 )
3. Distributions and types
In this section we recall some elementary inequalities for (relative) entropies
and the total variation distance for distributions on finite sets. Furthermore,
we generalize the notion of a probability distribution on a set to a distribution
23
3. Distributions and types
on a configuration of sets. Finally, we give a perspective on the theory of types,
and also introduce types in the context of complete configurations.
3.1. Distributions.
3.1.1. Single probability spaces. For a finite set S we denote by ∆S the collection of all probability distributions on S. It is a unit simplex in the real
vector space RS . We often use the fact that it is a compact, convex set, whose
interior points correspond to fully supported probability measures on S.
For π1 , π2 ∈ ∆S denote by ∣π1 − π2 ∣1 the total variation of the signed measure
(π1 − π2 ) and define the entropy of the distribution π1 by
(3.1)
h(π1 ) ∶= − ∑ π1 (x) ln π1 (x)
x∈X
If, in addition, π2 lies in the interior of ∆S define the relative entropy by
π1 (x)
(3.2)
D(π1 ∣∣ π2 ) ∶= ∑ π1 (x) ln
π2 (x)
x∈X
The entropy of a probability space is often defined through formula (3.1). It
is a standard fact, and can be verified with the help of Lemma 3.2 below, that
for π ∈ ∆S holds
(3.3)
h(π) = Ent(S, π)
which justifies the name “entropy” for the function h ∶ ∆S → R.
Define a divergence ball of radius ε > 0 centered at π ∈ Interior ∆S as
(3.4)
Bε (π) ∶= {π ′ ∈ ∆S ∣ D(π ′ ∣∣ π) ≤ ε}
For a fixed π and ε << 1 the ball Bε (π) also lies in the interior of ∆S.
Lemma 3.1. Let S be a finite set, then
(i) For any π1 , π2 ∈ ∆S, Pinsker’s inequality holds
√
∣π1 − π2 ∣1 ≤ 2D(π1 ∣∣ π2 )
(ii) For any π2 ∈ Interior ∆S there exists a positive constant C = Cπ2 such
that for any π1 ∈ ∆S, holds
√
∣π1 − π2 ∣1 ≥ C D(π1 ∣∣ π2 )
(iii) Suppose π is a point in the interior of ∆S and r > 0 is such that Br (π)
also lies in the interior of ∆S. There exist a constant C = Cπ,r such
that for any ε ≤ r holds
√
max {∣h(π1 ) − h(π2 )∣ ∣ π1 , π2 ∈ Bε (π)} ≤ C ε
⊠
The first claim of the Lemma, Pinsker’s inequality, is a well-known inequality
in for instance information theory, and a proof can be found in [CT91].
The second claim follows from the fact that for the fixed π2 ∈ Interior ∆S
the relative entropy as the function of the first argument is bounded, smooth
on the interior of the simplex and has a minimum at π2 .
24
3. Distributions and types
To prove the last claim, note that the entropy function h is smooth in the
interior of the simplex. Then this last claim follows from the first claim.
3.1.2. Distributions on configurations. A map f ∶ S → S ′ between two finite
sets induces an affine map f∗ ∶ ∆S → ∆S ′ .
For a configuration of sets S = {Si ; fij } we define the space of distributions
on the configuration S by
∆S ∶= {(πi ) ∈ ∏ ∆Si ∣∣ (fij )∗ πi = πj }
i
Essentially, an element of ∆S is a collection of distributions on the sets Si in
S that is consistent with respect to the maps fij . The consistency conditions
(fij )∗ πi = πj form a collection of linear equations with integer coefficients with
respect to the standard convex coordinates in ∏ ∆Si . Thus, ∆S is a rational
affine subspace in the product of simplices. In particular, ∆S has a convex
structure.
If S is complete with initial set S0 , then specifying a distribution π0 ∈ ∆S0
uniquely determines distributions on all of the Si ’s by setting πi ∶= (f0i )∗ π0 . In
such a situation we have
(3.5)
∆S ≅ ∆S0
If S is not complete and S0 , . . . , Sk is a collection of its initial sets, then ∆S
is isomorphic to an affine subspace of the product ∆S0 × ⋅ ⋅ ⋅ × ∆Sk cut out by
linear equations with integer coefficients corresponding to co-fans in S with
initial sets among S0 , . . . , Sk .
To simplify notation, for a probability space X or a configuration X we will
write
∆X ∶= ∆X
∆X ∶= ∆X
We now discuss briefly the theory of types. Types are special subspaces of
tensor powers that consist of seqences with the same “empirical distribution”
as explained in details below. For a more detailed discussion the reader is
referred to [CT91] and [Csi98]. We generalize the theory of types to complete
configurations of sets and complete configurations of probability spaces.
The theory of types for configurations, that are not complete, is more complex and will be addressed in a subsequent article.
3.2. Types for single probability spaces. Let S be a finite set. For n ∈ N
denote also
1
(n)
∆ S ∶= ∆S ∩ ZS
n
a collection of rational points in ∆S with denominator n. (We say that a
rational number r ∈ Q has denominator n ∈ N if r ⋅ n ∈ Z)
25
3. Distributions and types
Define the empirical distribution map q ∶ S n → ∆S, that sends (si )ni=1 = s ∈
S n to the empirical distribution q(s) ∈ ∆S given by
1
q(s)(a) = ⋅ ∣ {i ∣ si = a} ∣ for any a ∈ S
n
(n)
Clearly the image of q lies in ∆ S.
(n)
(n)
For π ∈ ∆ S, the space Tπ S ∶= q−1 (π) equipped with the uniform measure
is called a type over π. The symmetric group Sn acts on S ⊗n by permuting
the coordinates. This action leaves the empirical distribution invariant and
therefore could be restricted to each type, where it acts transitively. Thus, for
(n)
(n)
π ∈ ∆ S the probability space (Tπ S, u) with u being a uniform (Sn -invariant)
distribution, is a homogeneous space.
Suppose X = (X, p) is a probability space. Let τn be the pushforward of p⊗n
(n)
under the empirical distribution map q ∶ X n → ∆X . Clearly supp τn ⊂ ∆ X,
thus (∆X, τn ) is a finite probability space. Therefore we have a reduction
q ∶ X ⊗n → (∆X, τn )
(n)
which we call the empirical reduction. If π ∈ ∆ X is such that τn (π) > 0, then
(3.6)
(n)
Tπ X = X ⊗n ∣ π
In particular, it follows that the right-hand side does not depend on the probability p on X as long as π is “compatible” to it.
The following lemma records some standard facts about types, which can
be checked by elementary combinatorics and found in [CT91].
26
3. Distributions and types
Lemma 3.2. Let X be a probability space and x ∈ X ⊗n , then
(i)
(n)
n + ∣X∣
∣∆ X∣ = (
) ≤ e∣X∣⋅ln(n+1)
∣X∣
(ii)
[ (
) (
)]
p⊗n (x) = e−n h q(x) +D q(x) ∣∣ p
(iii)
(n)
en⋅h(π)−∣X∣⋅ln(n+1) ≤ ∣Tπ X∣ ≤ en⋅h(π)
(iv)
(n)
e−n⋅D(π ∣∣ p)−∣X∣⋅ln(n+1) ≤ τn (π) = p⊗n (Tπ X) ≤ e−n⋅D(π ∣∣ p)
⊠
If X = (X, pX ) is a probability space with rational probability distribution
with denominator n, then the type over pX will be called the true type of X
(n)
T (n) X ∶= TpX X
As a corollary to Lemma 3.2 and equation (3.3) we obtain the following.
(n)
Corollary 3.3. For a finite set S and π ∈ ∆ S holds
(n)
n ⋅ h(π) − ∣S∣ ⋅ ln(n + 1) ≤ Ent(Tπ S) ≤ n ⋅ h(π)
In particular, for a finite probability space X = (S, p) with a rational distribution p with denominator n holds
n ⋅ Ent(X) − ∣S∣ ⋅ ln(n + 1) ≤ Ent(T (n) X) ≤ n ⋅ Ent(X)
⊠
The following important theorem is known as Sanov’s theorem. It can be
derived from Lemma 3.2 and found in [CT91].
Theorem 3.4. (Sanov’s Theorem) Let X be a finite probability space and let
q ∶ X ⊗n → (∆X, τn ) be the empirical reduction. Then for every r > 0,
τn (∆X/Br (p)) ≤ e−n⋅r+∣X∣⋅ln(n+1)
where Br (p) is the divergence ball (relative entropy ball) defined in (3.4).
⊠
3.3. Types for complete configurations. In this subsection we generalize
the theory of types for configurations modeled on a complete category. The
theory for a non-complete configurations is more complex and will be addressed
in our future work. We will give three equivalent definitions of a type for a
complete configuration, each of which will be useful in its own way. Before we
describe the three approaches we need some preparatory material.
27
3. Distributions and types
Lemma 3.5. Given a diamond configuration of probability spaces
⎛ X
D=⎜
⎜ ρ1
⎝ A
Y ⎞
⎟
ρ2 ⎟
B ⎠
f
g
the following two conditions are equivalent
(i) The minimal reduction of the diamond D is isomorphic to the adhesion
of its co-fan or equivalently the following independence condition holds
A Y ∣B
(ii) For any a, a′ ∈ A such that f (a) = f (a′ ) = b ∈ B holds
Y ∣ a = Y ∣ a′ = Y ∣ b
⊠
Suppose D is the diamond as in the Lemma 3.5. The top row X → Y is
a two-chain subconfiguration of D and we can consider a conditioning by an
element a ∈ A
X∣ a → Y ∣ a
If D satisfies any of two conditions in Lemma 3.5, then Y ∣ a = Y ∣ b for b = g(a).
Thus, we constructed a reduction
X∣ a → Y ∣ b
Suppose we have a reduction f ∶ X → Y between a pair of probability spaces.
Then for any n ∈ N there is an induced reduction f∗ ∶ (∆X, τn ) → (∆Y, τn )
that can be included in the following diamond configuration
f ⊗n
X ⊗n
Y ⊗n
q
q
(∆X, τn )
f∗
(∆Y, τn )
that satisfies conditions in Lemma 3.5. It means that there is a reduction
(n)
(n)
T f ∶ Tπ X → Tπ′ Y
(n)
(n)
for π ∈ ∆ X and π ′ = f∗ π ∈ ∆ Y .
Now we are ready to give the definitions of types. Let X ∈ Prob ⟨G⟩ be a
(n)
complete configuration, X = {Xi ; fij } with initial space X0 and let π ∈ ∆ X .
3.3.1. Type of a configuration as the configuration of types. Define the type
(n)
Tπ X as the G-configuration, whose individual spaces are types of the individual spaces of X over the corresponding push-forwards of π
(n)
(n)
Tπ X ∶= {Tπi X i ; T fij }
28
3. Distributions and types
3.3.2. Types as Sn -orbits in the tensor power. By a section in X we mean a
consistent collection of points
[[X ]]−1
x = (x0 , . . . , x[[X ]]−1 ) ∈ ∏ X i
i=0
such that fij xi = xj , whenever fij is defined. For any j define the projection
ρj ∶ ∏ X i → X j , so that ρj (x) = xj .
The symmetric group Sn acts on the collection of sections in the tensor
power X ⊗n , by permuting the coordinates. Let T ⊂ ∏ X i be an orbit of the
(n)
action such that ρ0 (T ) = Tπ X 0 . Suppose that the pair (i, j) is such, that fij
is defined. Since fij⊗n ∶ Xi⊗n → Xj⊗n is Sn -equivariant, we have a map
T f ∶ ρi (T ) → ρj (T )
We can turn T into a G-configuration, which we will call a type of X
(n)
T̂π X ∶= {ρi (T ), T fij }
where π is the value of the impirical distribution on ρ0 (T ).
(n)
(n)
Since the initial space in T̂π X coincides with the initial space in Tπ X and
all the reductions coincide, we conclude that
(n)
(n)
T̂π X = Tπ X
3.3.3. Type as conditionining of the tensor power. We can extend X ⊗n to a configuration Xˆ by adding (∆X , τn ) and the empirical reduction X0⊗n → (∆X , τn ).
Let π ∈ ∆X with τn (π) > 0 and recall ∆X0 ≅ ∆X . We may now define
(n)
⊗n
X ∣ π as in Section 2.7. Define a type of X over π ∈ ∆ X by
(n)
Ťπ X ∶= X ⊗n ∣ π
By definition, it holds that
(n)
X0⊗n ∣ π = Tπ X0
Let πi = (f0i )∗ π. Using Lemma 3.5 and discussion thereafter we conclude that
Xi⊗n ∣ π = Xi⊗n ∣ πi
and therefore
(n)
(n)
Ťπ X = Tπ X
3.3.4. The empirical two-fan. We construct a two-fan of G-configurations with
terminal vertices X ⊗n and (∆X , τn )G (the constant G-configuration in which
every probability space is (∆X , τn ), and every reduction is an identity)
(3.7)
⎛
⎜
⎜
Rn (X ) = ⎜
⎜
⎜
⎝ X ⊗n
X˜ (n)
(n)
(∆ X , τn )G
⎞
⎟
⎟
⎟
⎟
⎟
⎠
29
4. The Kolmogorov-Sinai distance
With the help of Lemma 2.1, we construct Rn (X ) as the minimal reduction
of the two-fan of G-configurations
(X0⊗n )G
n
f0∗
qG
X ⊗n
(∆X , τn )G
Let π ∈ (∆X , τn ) with τn (π) > 0. Then within Rn (X ) holds
(n)
X ⊗n ∣ π = Tπ X
(n)
(n)
For every n ∈ N and π ∈ ∆ X 0 the type Tπ X is a homogeneous configuration. Suppose that a complete configuration X is such that the probability
distribution p0 on the initial set is rational with the denominator n, then we
(n)
call Tp X the true type of X and denote
T (n) X ∶= Tpn0 X
4. The Kolmogorov-Sinai distance
We turn the space of configurations into a pseudo-metric space by introducing the intrinsic Kolmogorov-Sinai distance and asymptotic Kolmogorov-Sinai
distance. For brevity, we will usually call it the Kolmogorov distance and
asymptotic Kolmogorov distance. The intrinsic Kolmogorov-Sinai distance is
obtained by taking an infimum of the shared information distance over all possible joint distributions on two probability spaces. The name is justified by
the fact that the shared information distance (not under this name) appears
in the proof of the theorem about generating partitions for ergodic systems
by Kolmogorov and Sinai, see for example [Sin76]. Note that the Kolmogorov
distance in statistics refers to a different notion.
4.1. Kolmogorov distance and asymptotic Kolmogorov distance.
4.1.1. Kolmogorov distance in the case of single probability spaces. For a twofan F = (X ← Z → Y ) define a “distance” kd(F) between probability spaces
X and Y with respect to F by
kd(F) ∶= Ent(Z∣ Y ) + Ent(Z∣ X)
= 2 Ent(Z) − Ent(X) − Ent(Y )
Essentially kd(F) measures the deviation of the statistical map defined by
F from being a deterministic bijection between X and Y .
The minimal reduction F ′ of F satisfies
(4.1)
kd(F ′ ) ≤ kd(F)
If the two-fan F is minimal the “distance” kd(F) can also be calculated by
kd(F) = h(pX ) + h(pY ) − 2D(pZ ∣∣ pX ⊗ pY ),
30
4. The Kolmogorov-Sinai distance
where h and D are respectively the entropy and relative entropy functions
defined in (3.1) and (3.2).
For a pair of probability spaces X, Y define the intrinsic Kolmogorov-Sinai
distance as
k(X, Y ) ∶= inf {kd(F) ∣ F = (X ← Z → Y ) is a two-fan}
The optimization takes place over all two-fans with terminal spaces X and
Y . In view of inequality (4.1) one could as well optimize over the space of
minimal two-fans, which we will also refer to as couplings between X and Y .
The tensor product of X and Y trivially provides a coupling and the set of
couplings is compact, therefore an optimum is always achieved and it is finite.
The bivariate function k ∶ Prob × Prob → R≥0 defines a notion of pseudodistance and it vanishes exactly on pairs of isomorphic probability spaces.
This follows directly from the Shannon inequality (1.2), and a more general
statement will be proven in Proposition 4.1 below.
4.1.2. Kolmogorov distance for complete configurations. The definition of Kolmogorov distance for complete configurations repeats almost literally the definition for single spaces. We fix a complete diagram category G and will be
considering configurations from Prob ⟨G⟩.
Consider three configurations X = {Xi , fij }, Y = {Yi , gij } and Z = {Zi , hij }
from Prob ⟨G⟩. Recall that a two-fan F = (X ← Z → Y) is a G-configuration
of two-fans
Fi = (Xi ← Zi → Yi )
Define
kd(F) ∶= ∑ kd(Fi )
i
= ∑ (2 Ent(Zi ) − Ent(Xi ) − Ent(Yi ))
i
The quantity kd(F) vanishes if and only if the fan F provides isomorphisms
between all individual spaces in X and Y that commute with the inner structure of the configurations, that is, it provides an isomorphism between X and
Y in Prob ⟨G⟩.
The intrinsic Kolmogorov-Sinai distance between configurations is defined in
analogy with the case of single probability spaces
k(X , Y) ∶= inf {kd(F) ∣ F = (X ← Z → Y)}
where the infimum is over all two-fans of G-configurations with terminal vertices X and Y.
The following proposition records that the intrinsic Kolmogorov distance
is in fact a pseudo-distance on Prob ⟨G⟩, provided G is a complete diagram
category (that is when G has a unique initial space).
Proposition 4.1. ↓ Let G be a complete diagram category. Then the bivariate
function
k ∶ Prob ⟨G⟩ × Prob ⟨G⟩ → R
4. The Kolmogorov-Sinai distance
31
is a pseudo-distance on Prob ⟨G⟩.
Moreover, two configurations X , Y ∈ Prob ⟨G⟩ satisfy k(X , Y) = 0 if and only
if X is isomorphic to Y in Prob ⟨G⟩.
⊠
The idea of the proof is very simple. In the case of single probability spaces
X, Y, Z a coupling between X and Z can be constructed from a coupling between X and Y and a coupling between Y and Z by adhesion on Y , see Section
2.5. The triangle inequality then follows from a Shannon inequality. However,
since we are dealing with configurations the combinatorial structure requires
careful treatment. Therefore, we provide a detailed proof on page 61.
It is important to note, that the proof uses the fact that G is complete. In
fact, even though the definition of k could be easily extended to some bivariate
function on the space of configurations of any fixed combinatorial type, it
fails to satisfy the triangle inequality in general, because the composition of
couplings requires completeness of G.
4.1.3. The asymptotic Kolmogorov-Sinai distance. Let G be a complete diagram category. We define the asymptotic Kolmogorov-Sinai distance between
two configurations X , Y ∈ Prob ⟨G⟩ by
1
(4.2)
κ(X , Y) = lim k(X ⊗n , Y ⊗n ).
n→∞ n
We will show in Corollary 4.5, that the sequence
n ↦ k(X ⊗n , Y ⊗n )
is subadditive, and therefore the limit in the definition (4.2) of κ(X , Y) always
exists and for all n ∈ N holds
1
(4.3)
κ(X , Y) ≤ ⋅ k(X ⊗n , Y ⊗n ).
n
As a corollary of Proposition 4.1 and definition (4.2) we immediately obtain
that also the asymptotic Kolmogorov-Sinai distance is a pseudo-distance on
Prob ⟨G⟩.
Corollary 4.2. Let G be a complete diagram category. Then the bivariate
function
κ ∶ Prob ⟨G⟩ × Prob ⟨G⟩ → R
is a pseudo-distance on Prob ⟨G⟩ satisfying the following homogeneity property. For any pair of configurations X , Y ∈ Prob ⟨G⟩ and any n ∈ N0 holds
κ(X ⊗n , Y ⊗n ) = n ⋅ κ(X , Y)
⊠
We will show in a later section, however, that there are probability spaces
X and Y for which κ(X, Y ) = 0 that are not isomorphic.
In the rest of this section we derive some elementary properties of the intrinsic Kolmogorov distance and the asymptotic Kolmogorov distance.
32
4. The Kolmogorov-Sinai distance
4.2. Lipschitz property for operations. In this section we show that certain natural operations on configurations, namely the tensor product, entropy
function and restriction operator, are Lipschitz continuous. In Section 7 we
will show Lipschitz continuity of certain extension operations.
4.2.1. Tensor product. We show that the tensor product on the space of configurations is 1-Lipschitz. Later this will allow us to give a simple description
of tropical configurations, that is of points in the asymptotic cone of Prob ⟨G⟩,
as limits of certain sequences of “classical” configurations.
Proposition 4.3. ↓ Let G be a complete diagram category. Then with respect
to the Kolmogorov distance on Prob ⟨G⟩ the tensor product
⊗ ∶ (Prob ⟨G⟩ , k)2 → (Prob ⟨G⟩ , k)
is 1-Lipschitz in each variable, that is, for every triple X , Y, Y ′ ∈ Prob ⟨G⟩ the
following bound holds
k(X ⊗ Y, X ⊗ Y ′ ) ≤ k(Y, Y ′ )
⊠
This statement is a direct consequence of additivity of entropy with respect
to the tensor product. Details can be found on page 63.
It follows directly from definition (4.2) and Proposition 4.3, that the asymptotic Kolmogorov distance enjoys a similar property.
Corollary 4.4. Let G be a complete diagram category. Then with respect to
the Kolmogorov distance on Prob ⟨G⟩ the tensor product
⊗ ∶ (Prob ⟨G⟩ , κ)2 → (Prob ⟨G⟩ , κ)
⊠
is 1-Lipschitz in each variable.
As another corollary we obtain the subadditivity properties of the intrinsic
Kolmogorov distance and asymptotic Kolmogorov distance.
Corollary 4.5. Let G be a complete diagram category and let X , Y, U, V ∈
Prob ⟨G⟩, then
k(X ⊗ U, Y ⊗ V) ≤ k(X , Y) + k(U, V).
and
κ(X ⊗ U, Y ⊗ V) ≤ κ(X , Y) + κ(U, V).
⊠
It implies in particular that shifts are non-expanding maps in (Prob ⟨G⟩ , k)
or (Prob ⟨G⟩ , κ).
Corollary 4.6. Let G be a complete diagram category and δ = k, κ be either
Kolmogorov distance or asymptotic Kolmogorov distance on Prob ⟨G⟩. Let
U ∈ Prob ⟨G⟩. Then the shift map
U ⊗ ⋅ ∶ (Prob ⟨G⟩ , δ) → (Prob ⟨G⟩ , δ),
X ↦U ⊗X
33
4. The Kolmogorov-Sinai distance
is a non-expanding map with respect to either Kolmogorov distance or asymptotic Kolmogorov distance.
⊠
4.2.2. Entropy. Recall that we defined the entropy function
Ent∗ ∶ Prob ⟨G⟩ → R[[G]]
by evaluating the entropy of all individual spaces in a G-configuration. The
target space R[[G]] will be endowed with the `1 -norm with respect to the natural
coordinate system. With such a choice, the entropy function is 1-Lipschitz with
respect to the Kolmogorov distance on Prob ⟨G⟩.
Proposition 4.7. ↓ Suppose G is a complete diagram category and δ = k, κ is
either Kolmogorov distance or asymptotic Kolmogorov distance on Prob ⟨G⟩.
Then the entropy function
Ent∗ ∶ (Prob ⟨G⟩ , δ) → (R[[G]] , ∣ ⋅ ∣1 ),
X = {Xi , fij } ↦ (EntXi )i ∈ R[[G]]
⊠
is 1-Lipschitz.
Again, the proof of the proposition above is an application of Shannon’s
inequality, see page 64 for details.
4.2.3. Restrictions. The restriction operators are also Lipschitz, as shown in
the next proposition.
Proposition 4.8. ↓ Suppose R ∶ G′ → G is a functor between two complete diagram categories and δ stands for either Kolmogorov or asymptotic Kolmogorov
distance. Then the restriction operator
R∗ ∶ (Prob ⟨G⟩ , δ) → (Prob ⟨G′ ⟩ , δ),
X ↦X ○R
⊠
is Lipschitz.
As can be seen from the proof on page 64, the Lipschitz constant in the
proposition above can be bounded by [[G′ ]]. In fact, a more careful analysis
provides a better bound by the maximal number of objects in G′ that are
mapped by R to a single object in G.
4.3. The Slicing Lemma. The Slicing Lemma, Proposition 4.9 below, allows to estimate the Kolmogorov distance between two configurations with
the integrated Kolmogorov distance between “slices”, which are configurations
obtained by conditioning on another probability space.
The Slicing Lemma, along with the local estimate in Section 4.4, turned out
to be a very powerful tool for estimation of the Kolmogorov distance and will
be used below on many occasions.
As described in Section 2.2, by a reduction of a configuration X = {Xi , fij }
to a single space U we mean a collection of reductions {ρi ∶ Xi → U } from the
individual spaces in X to U , that commute with the reductions within X
ρj ○ fij = ρi
34
4. The Kolmogorov-Sinai distance
Alternatively, whenever a single probability space appears together with a Gconfiguration in a commutative diagram, it should be replaced by a constant
G-configuration.
Proposition 4.9. ↓ (Slicing Lemma) Suppose G is a complete diagram category and we are given X , Xˆ , Y, Ŷ ∈ Prob ⟨G⟩ – four G-configurations and
U, V, W ∈ Prob – probability spaces, that are included into the following threetents configuration
Xˆ
X
Ŷ
W
U
V
Y
such that the two-fan (U ← W → V ) is minimal. Then the following estimate
holds
k(X , Y) ≤ ∫ k(X ∣ u, Y∣ v) d pW (u, v)
W
+ [[G]] ⋅ kd(U ← W → V )
+ ∑ [ Ent(U ∣ Xi ) + Ent(V ∣ Yi )]
i
⊠
The idea of the proof of the Slicing Lemma (page 65) is as follows. For every
pair (u, v) ∈ W we consider an optimal two-fan Guv coupling X ∣ u and Y∣ v.
These fans have the same underlying configuration of sets. Then we construct
a coupling between X and Y as a convex combination of distributions of Guv ’s
weighted by pW (u, v). The estimates on the resulting two-fan then imply the
proposition.
Various implications of the Slicing Lemma are summarized in the next corollary.
Corollary 4.10. Let G be a complete diagram category, X , Y ∈ Prob ⟨G⟩ and
U ∈ Prob.
(i) Given a “two-tents” configuration
X ← Xˆ → U ← Ŷ → Y
the following inequality holds
k(X , Y) ≤ ∫ k(X ∣ u, Y∣ u) d pU (u) + 2 ⋅ [[G]] ⋅ Ent(U )
U
(ii) Given a fan
X ← Xˆ → U
the following inequality holds
k(X , Y) ≤ ∫
UV
k(X ∣ u, Y) d pU (u) + 2 ⋅ [[G]] ⋅ Ent(U )
4. The Kolmogorov-Sinai distance
35
(iii) Let X → U be a reduction, then
k(X , Y) ≤ ∫ k(X ∣ u, Y) d pU (u) + [[G]] ⋅ Ent(U )
U
(iv) For a co-fan X → U ← Y holds
k(X , Y) ≤ ∫ k(X ∣ u, Y∣ u) d pU (u)
U
⊠
4.4. Local estimate. Fix a complete diagram category G and consider a Gconfiguration of sets S ∈ Set ⟨G⟩ with S0 being an initial set in S. As discussed
in Section 3.1.2, the space of distributions on S could be identified with the
space of distributions on the initial set
∆S ≅ ∆S0
Therefore, all G-configurations of probability spaces with the underlying configuration of sets equal to S are in one-to-one correspondence with the interior
points of ∆S0 . The set Interior ∆S0 consists of fully supported measures on
the set S0 and carries a total variation distance, which is just an `1 -distance
with respect to the convex coordinates on the simplex ∆S0 . Our task presently
is to compare the total variation distance with the Kolmogorov distance on
the space of configurations with the fixed underlying configuration of sets.
The upper bound on Kolmogorov distance, that we derive below, has two
summands. One is linear in the total variation distance with the slope proportional to the log-cardinality of S0 . The second one is super-linear in the total
variation distance, but it does not depend on S. So we have the following interesting observation: of course, the super-linear summand always dominates
the linear one locally. However as the cardinality of S becomes large it is the
linear summand that starts playing the main role.
4.4.1. The estimate. Suppose we are given a configuration of sets S = {Si , fij } ∈
Set ⟨G⟩ modeled on a complete diagram category G with the initial set S0 .
We use once again the isomorphism
≅
∆S → ∆S0
that sends p ∈ ∆S to its component in the initial space p0 ∈ ∆S0 , while its
inverse is given by p = {(f0i )∗ p0 }. For a pair of distributions p0 , q0 ∈ ∆S0
denote by ∣p0 − q0 ∣ the total variation of the difference.
For α ∈ [0, 1] consider a binary probability space with the weight of one of
the atoms equal to α
Λα ∶= ( {◻, ∎} ; pΛα (◻) = 1 − α, pΛα (∎) = α)
36
5. Distance between types
Proposition 4.11. ↓ Let S = {Si , fij } ∈ Set ⟨G⟩ be a configuration of sets
modeled on a complete diagram category G with the initial set S0 . Let p, q ∈
∆S be two probability distributions. Denote X ∶= (S, p), Y ∶= (S, q) and α =
1
2 ∣p0 − q0 ∣1 . Then
k(X , Y) ≤ 2 ⋅ [[G]] ⋅ (α ⋅ ln ∣S0 ∣ + Ent(Λα ))
⊠
To prove the local estimate we decompose both p and q into a convex combination of a common part p̂ and rests p+ and q + . The coupling between the
common parts gives no contribution to the distance, and the worst possible
estimate on the other parts is still enough to get the bound in the lemma, by
using Corollary 4.10 part (i). Details of the proof can be found on page 67.
In fact, the lower bound also holds, more specifically given a complete Gconfiguration of sets S and p ∈ ∆S there is a constant C > 0 such that for any
q ∈ ∆S with ∣p − q∣1 << 1 holds
C ⋅ Ent(Λα ) ≤ k(X , Y)
where X = (S, p), Y = (S, q) and α = 21 ∣p − q∣1 . We will not use this fact and
therefore do not include a proof.
Once the G-configuration of sets S is fixed, there is a map from ∆S to
Prob ⟨G⟩. As can be seen from the discussion above, even though the map is
continuous it is not Lipschitz.
5. Distance between types
As explained in Section 3.3, given a complete G-configuration S of sets and a
(n)
rational distribution π ∈ ∆S we construct a homogeneous configuration Tπ S,
which is called the type of S over π. Our goal in this section is to estimate
the Kolmogorov distance between two types over two different distributions
(n)
π1 , π2 ∈ ∆ S in terms of the total variation distance ∣π1 − π2 ∣1 .
For this purpose we use a “lagging” technique which is explained below.
5.1. The lagging trick. Let Λα be a binary probability space,
Λα = ( {◻, ∎} ; pΛα (∎) = α)
and let X = {(X i , pi ); fij }, Z = {(Z i , qi ); gij } be two configurations modeled on
a complete diagram category G and included in a minimal two-fan
λ
ρ
Λα ←Ð Z Ð→ X
Recall that the left terminal vertex in this two-fan should be interpreted as a
constant G-configuration ΛG
α.
Assume further that the distribution q on Z is rational with denominator
(n)
n ∈ N, that is q ∈ ∆ Z. It follows that p and pΛα are also rational with the
same denominator n.
37
5. Distance between types
We construct a lagging two-fan
(5.1)
Tρ
l
L ∶= (T ((1−α)n) (X ∣ ◻) ←Ð T (n) Z Ð→ T (n) X )
as follows. The right leg T ρ of L is induced by the right leg ρ of the original
two-fan. The left leg
l ∶ T (n) Z → T ((1−α)n) (X ∣ ◻)
is obtained by erasing symbols that reduce to ∎ and applying ρ to the remaining
symbols. The target space for the reduction l is the true type of X ∣ ◻ which is
“lagging” behind T (n) Z by a factor of (1−α). More specifically, the reduction l
is constructed as follows. Let λj ∶ Zj → Λα be the components of the reduction
λ ∶ Z → Λα .
Given z = (zi )ni=1 ∈ T (n) Zj define the subset of indexes
Iz ∶= {i ∣ λj (zi ) = ◻}
and define the j th component of l by
lj ((zi )ni=1 ) ∶= (ρ(zi ))i∈Iz
By equivariance each lj is a reduction of homogeneous spaces, since the
inverse image of any point has the same cardinality. Moreover the reductions lj
commute with the reductions in T (n) Z as explained in Section 3.3 and therefore
l is a reduction of configurations.
The next lemma uses the lagging two-fan to estimate the Kolmogorov distance between its terminal configurations.
Lemma 5.1. Let X , Z ∈ Prob ⟨G⟩ be two configurations modeled on a complete diagram category G and included in a minimal two-fan
λ
ρ
Λα ←Ð Z Ð→ X
where distribution on Z is rational with denominator n ∈ N. Then
k (T ((1−α)n) (X ∣ ◻) , T (n) X )
≤ n ⋅ [[G]] ⋅ [2 Ent(Λα ) + α ⋅ ln ∣X0 ∣] + 2 ⋅ [[G]] ⋅ ∣X0 ∣ ⋅ ln(n + 1)
= n ⋅ [[G]] ⋅ [2 Ent(Λα ) + α ⋅ ln ∣X0 ∣] + O (∣X0 ∣ ⋅ ln n)
⊠
The Lemma (and Proposition 5.2 below) are closely related to the local
estimate, Proposition 4.11. It is an immediate consequence of the Slicing
Lemma, in particular Corollary 4.10 part (ii) that
k (X ∣ ◻ , X ) ≤ [[G]] ⋅ [2 Ent(Λα ) + α ⋅ ln ∣X0 ∣]
This is a tacit ingredient in the proof of the local estimate. By the subadditivity
of the Kolmogorov distance,
k ((X ∣ ◻)⊗n , X ⊗n ) ≤ n ⋅ [[G]] ⋅ [2 Ent(Λα ) + α ⋅ ln ∣X0 ∣]
38
5. Distance between types
This bound is almost the estimate in Lemma 5.1, except Lemma 5.1 estimates the distance between types rather than tensor powers. We will soon
see that tensor powers and types are very close in the Kolmogorov distance.
However, for the purpose of the proof of Lemma 5.1, it suffices to know that
their entropies are close, an estimate that is provided by Corollary 3.3.
Proof (of Lemma 5.1): We will use the lagging two-fan constructed in Equation (5.1), namely
l
Tρ
L ∶= (T ((1−α)n) (X ∣ ◻) ←Ð T (n) Z Ð→ T (n) X )
as a coupling to estimate the Kolmogorov distance
k (T ((1−α)n) (X ∣ ◻) , T (n) X ) ≤ kd(L)
Recall that by Corollary 3.3 for a probability space X with a rational distribution we have
n ⋅ Ent(X) − ∣X∣ ⋅ ln(n + 1) ≤ Ent(T (n) X) ≤ n ⋅ Ent(X)
Thus we can estimate kd(L) as follows
kd(L) = ∑ [( Ent(T (n) Zi ) − Ent(T (n) Xi ))
i
+ ( Ent(T (n) Zi ) − Ent(T ((1−α)n) (Xi ∣ ◻)))]
≤ n ⋅ ∑ [( Ent(Zi ) − Ent(Xi )) + ( Ent(Zi ) − (1 − α) Ent(Xi ∣ ◻))]
i
+ 2 ⋅ [[G]] ⋅ ∣X0 ∣ ⋅ ln(n + 1)
By minimality of the original two-fan and Shannon inequality (1.2) we have
a bound
Ent(Zi ) − Ent(Xi ) ≤ Ent(Λα )
The second part in the sum can be estimated using relation (1.4) as follows
Ent(Zi ) − (1 − α) Ent(Xi ∣ ◻) = Ent(Λα ) + Ent(Xi ∣ Λα ) − (1 − α) Ent(Xi ∣ ◻)
= Ent(Λα ) + (1 − α) Ent(Xi ∣ ◻) + α Ent(Xi ∣ ∎)−
− (1 − α) Ent(Xi ∣ ◻)
≤ Ent(Λα ) + α ⋅ ln ∣Xi ∣
Combining all of the above we obtain the estimate in the conclusion of the
lemma.
⊠
5.2. Distance between types. In this section we use the lagging trick as
described above to estimate the distance between types over two different
distributions in ∆S where S is a complete configuration of sets.
39
5. Distance between types
Proposition 5.2. Suppose S is a complete G-configuration of sets with initial
(n)
set S0 . Suppose p, q ∈ ∆ S and let α = 12 ∣p0 − q0 ∣1 . Then
(n)
(n)
k(Tp S, Tq S) ≤ 2n ⋅ [[G]] ⋅ [α ⋅ ln ∣S0 ∣ + 2 Ent(Λα )] + 4[[G]] ⋅ ∣X0 ∣ ⋅ ln(n + 1)
= 2n ⋅ [[G]] ⋅ [α ⋅ ln ∣S0 ∣ + 2 Ent(Λα )] + O(∣X0 ∣ ⋅ ln n)
⊠
As in the local estimate, the idea of the proof is to write p and q as a convex
combination of a common distribution p̂ and “small amounts” of p+ and q + ,
respectively. Then we use the lagging trick to estimate distances between types
over p and p̂, as well as between types over q and p̂. We now present details
of the proof.
Proof (of Proposition 5.2): Recall that for a complete configuration S with
initial set S0 we have
∆S ≅ ∆S0
(5.2)
Our goal now is to write p and q as the convex combination of three other
distributions p̂, p+ and q + as in
p = (1 − α) ⋅ p̂ + α ⋅ p+
q = (1 − α) ⋅ p̂ + α ⋅ q +
We could do it the following way. Let α ∶= 21 ∣p0 − q0 ∣1 . If α = 1 then the
proposition follows trivially by constructing a tensor-product fan, so from now
on we assume that α < 1. Define three probability distributions p̂0 , p+0 and q0+
on S0 by setting for every x ∈ S0
1
min {p0 (x), q0 (x)}
1−α
1
p+0 ∶= (p0 − (1 − α)p̂0 )
α
1
+
q0 ∶= (q0 − (1 − α)p̂0 )
α
p̂0 (x) ∶=
Denote by p̂, p+ , q + ∈ ∆S the distributions corresponding to p̂0 , p+0 , q0+ ∈ ∆S0
under the affine isomorphism (5.2). Thus we have
p = (1 − α)p̂ + α ⋅ p+
q = (1 − α)p̂ + α ⋅ q +
Now we construct a pair of two-fans of G-configurations
(5.3)
Λα ← X˜ → X
Λα ← Ỹ → Y
40
6. AEP for configurations
by setting
X ∶= (S, p)
Y ∶= (S, q)
X̃i ∶= (Si × Λα ; p̃i (s, ◻) = (1 − α)p̂i (s), p̃i (s, ∎) = α ⋅ p+i (s))
Ỹi ∶= (Si × Λα ; q̃i (s, ◻) = (1 − α)p̂i (s), q̃i (s, ∎) = α ⋅ qi+ (s))
and
X˜ ∶= {X̃i ; fij × Id}
Ỹ ∶= {Ỹi ; fij × Id}
The reductions in (5.3) are given by coordinate projections. We have the
following isomorphisms
X ∣ ◻ ≅ Y∣ ◻ ≅ (S, p̂)
To estimate the distance between types we now apply Lemma 5.1 to the
fans in (5.3)
(n)
(n)
k(Tp S, Tq S) = k(T (n) X , T (n) Y)
≤ k (T (n) X , T ((1−α)n) (X ∣ ◻)) + k (T ((1−α)n) (Y∣ ◻), T (n) Y)
≤ 2n ⋅ [[G]] ⋅ [α ⋅ ln ∣S0 ∣ + 2 Ent(Λα )] + 4[[G]] ⋅ ∣X0 ∣ ⋅ ln(n + 1)
⊠
The reason for the similarity between the local estimate and the distance
estimate between types will become clear in the next section, when we establish the asymptotic equivalence between the Bernoulli sequence of probability
spaces and sequence of types over rational distributions approximating the
true distribution.
6. Asymptotic equipartition property for configurations
Below we prove that any Bernoulli sequence can be approximated by a
sequence of homogeneous configurations. This is essentially the Asymptotic
Equipartition Theorem for configurations.
Theorem 6.1. Suppose X ∈ Prob ⟨G⟩ is a complete configuration of probab
ility spaces. Then there exists a sequence H = (Hn )∞
n=0 of homogeneous configurations of the same c ombinatorial type as X such that
√
⎛
1
ln3 n ⎞
k(X ⊗n , Hn ) = O
n
n ⎠
⎝
41
6. AEP for configurations
More precisely, the sequence H may be chosen such that for all n ≥ ∣X0 ∣
√
1
ln3 n
(6.1)
k(X ⊗n , Hn ) ≤ C(∣X0 ∣, [[G]]) ⋅
n
n
where C(∣X0 ∣, [[G]]) is a constant only depending on ∣X0 ∣ and [[G]].
⊠
Proof: Denote by S = X the underlying configuration of sets and by pX the
true distribution on S, such that
X = (S, pX )
We will construct the approximating homogeneous sequence by taking types
over rational approximations of pX in ∆S, that converge sufficiently fast to the
true distribution pX .
(n)
More specifically, we select rational distributions pn ∈ ∆ S such that
∣pn − pX ∣1 ≤
∣S0 ∣
n
(n)
As homogeneous spaces Hn we set Hn = Tpn S. We will show that the
Kolmogorov distance between Hn and X ⊗n satisfies the required estimate (6.1).
First we apply slicing along the empirical two-fan
Rn (X ) = (X ⊗n ← X˜ (n) → (∆S, τn )G )
defined in Section 3.3, Equation (3.7) on page 28.
For the estimate below we use the fact that
(n)
Ent(∆S, τn ) ≤ ln ∣∆ S∣ ≤ ∣S0 ∣ ⋅ ln(n + 1)
By slicing (see Corollary 4.10(ii)) along the empirical two-fan we have
(n)
k(Tpn S, X ⊗n ) ≤ 2 ⋅ [[G]] ⋅ Ent(∆S, τn ) + ∫
(n)
∆S
≤ 2 ⋅ [[G]] ⋅ ∣S0 ∣ ⋅ ln(n + 1) + ∫
(n)
k(Tpn S, Tπ S) d τn (π)
(n)
∆S
(n)
k(Tpn S, Tπ S) d τn (π)
To estimate the integral we split the domain into a small divergence ball Bεn =
Bεn (pX ) around the “true” distribution and its complement
(n)
(n)
(n)
(n)
∫∆S k(Tpn S, Tπ S) d τn (π) = ∫∆S∖B k(Tpn S, Tπ S) d τn (π)
εn
+∫
(6.2)
(n)
Bεn
(n)
k(Tpn S, Tπ S) d τn (π)
and we set the radius εn equal to
ln(n + 1)
n
To estimate the first integral on the right-hand side of equality (6.2) note
that the distance between two types over the same configuration of sets can
always be crudely estimated by
εn ∶= (∣S0 ∣ + 1)
2 ⋅ ln ∣S0 ∣ ⋅ [[G]] ⋅ n
42
7. Extensions
Moreover, by Sanov’s theorem, Theorem 3.1, we can estimate the empirical
measure of the complement of the divergence ball
τn (∆S ∖ Bεn ) ≤ e−n⋅εn +∣S0 ∣⋅ln(n+1) ≤
1
n
where we used the definition of εn to conclude the last inequality. Therefore
we obtain
(n)
(n)
∫∆S∖B k(Tpn S, Tπ S) d τn (π) ≤ 2 ⋅ ln ∣S0 ∣ ⋅ [[G]] ⋅ n ⋅ τn (∆S ∖ Bεn )
εn
≤ 2 ⋅ ln ∣S0 ∣ ⋅ [[G]]
Define
αn =
∣S0 ∣ √
+ 2εn
n
if the right-hand side is smaller than 1 and set αn = 1 otherwise. Then every
π ∈ Bεn (pX ) satisfies ∣pn − π∣ ≤ 2αn by Pinsker’s inequality (Lemma 3.1, (i)),
and the triangle inequality. Consequently, by the estimate on the distance
between types in Proposition 5.2
(n)
(n)
∫B k(Tpn S, Tπ S) d τn (π)
εn
≤ 2n ⋅ [[G]] ⋅ (αn ln ∣S0 ∣ + 2 Ent(Λαn )) + 4 ⋅ [[G]] ⋅ ∣S0 ∣ ⋅ ln(n + 1)
Using the definition of αn and εn we find that
∫B
εn
√
(n)
(n)
k(Tpn S, Tπ S) d τn (π) = O ( n ⋅ ln3 n)
and hence combining the above estimates
⎛
1
(n)
k(Tpn S, X ⊗n ) = O
n
⎝
√
ln3 n ⎞
.
n ⎠
A more precise check shows that for n ≥ ∣S0 ∣, the constants appearing in O
only depend on ∣S0 ∣ and [[G]].
⊠
It is worth noting that each type considered as a subspace of the tensor
power takes up only small probability. In fact its probability converges to
zero with growing n. But as the calculation above shows, most (in terms of
probability) of the configuration X ⊗n consists of polynomially many types that
are k-similar to each other. Relative to the exponential growth of sizes of all
the parts, “polynomially many” is as good as one. This is the difference with
the setup used in Gromov’s [Gro12]
43
7. Extensions
7. Extensions
In the introduction we have already emphasized the close relationship between relative entropic sets and Information-Optimization problems. There,
our definitions were restricted to extensions of two-fans to full configurations
corresponding to three random variables. We will now generalize these definitions, and make the relationship between relative entropic sets and InformationOptimization problems explicit.
Further we will prove the Extension Lemma and use it to show that the
relative entropic set associated to a full configuration depends continuously on
the configuration.
7.1. Information-Optimization and the relative entropic set. In Section 2.4.1 we introduced (for k ≤ l) the restriction operator
∗
Rk,l
∶ Prob ⟨Λl ⟩ → Prob ⟨Λk ⟩
as follows. For a minimal full configuration Y = ⟨Yi ⟩li=1 we denote by
∗
Y = ⟨Yi ⟩i∈{1,...,k}
Rk,l
the restriction of Y to a minimal full configuration generated by Yi , i ∈ {1, . . . , k}.
We call a minimal configuration Y ∈ Prob ⟨Λk ⟩ an l-extension of a configuration X if
∗
Y =X
Rk,l
and we denote the class of all l-extensions by Extl (X ).
Recall that for a full configuration Y ∈ Prob ⟨Λl ⟩, we record the entropies of
{1,...,l} ∖{∅}
all its probability spaces in a vector in R2
that we denote by
Ent∗ (Y) ∶= ( Ent(YI ))I∈2{1,...,l} ∖{∅}
The entries in this vector are all nonnegative. To simplify notations we set
El ∶= (R2
{1,...,l} ∖{∅}
, ∣ ⋅ ∣1 )
and denote by E∗l its dual vector-space.
As in the introduction, we introduce the unstabilized relative entropic set
Γ○l (X ) ∶= {Ent∗ (Y) ∣ Y ∈ Extl (X )}
By the additivity property of the entropy with respect to tensor powers, there
is the inclusion
(7.1)
Γ○l (X ⊗m ) + Γ○l (X ⊗n ) ⊂ Γ○l (X ⊗(m+n) )
where the sum on the left hand side is the Minkowski sum. This allows us to
define the limit
1
1
lim Γ○l (X ⊗n ) ∶= ⋃ Γ○l (X ⊗n )
n→∞ n
n∈N n
and we define the stabilized relative entropic set by
1
Γl (X ) ∶= Closure ( lim Γ○l (X ⊗n ))
n→∞ n
44
7. Extensions
which is a closed convex subset of El by property (7.1).
For a vector c ∈ E∗l , we define the Information-Optimization problem
IOc (X ) ∶=
inf
Y∈Extl (X )
⟨ c , Ent∗ (Y) ⟩ =
inf
Y∈Extl (X )
∑
cI ⋅ Ent(YI )
I⊂{1,...,l}
where cI ’s are the coordinates of the vector c with respect to the basis in E∗l
dual to the standard basis in El . Note that, equation (7.1) implies that the
sequence
n ↦ IOc (X ⊗n )
is subadditive. Hence, the limit
1
lim IOc (X ⊗n )
n→∞ n
always exists (but may be equal to −∞). If for all n ∈ N,
1
IOc (X ⊗n ) = IOc (X )
n
we call the optimization problem associated to c stable. In general, we define
the stabilized optimization problem
1
IOsc (X ) ∶= lim IOc (Xn⊗n ).
n→∞ n
As the stabilized relative entropic set is convex, it is the intersection of
half-spaces that are defined by linear inequalities on entropies
Γk (X ) = ⋂ {x ∈ El ∣ ⟨c, x⟩ ≥ IOsc (X )}
c∈E∗l
In other words, the stabilized information optimization problems, that occur
so often in practice, identify supporting hyper-planes of the convex set. The
solution of all such linear problems determine the shape of the relative entropic
set and vice versa.
7.2. The entropic set and the entropic cone. The definitions of relative
entropic sets are motivated by the more classical notion of the entropic cone,
which we will briefly discuss now. For l ∈ N, the entropic set is defined as
Γ○l ∶= {Ent∗ (Y) ∣ Y ∈ Prob ⟨Λl ⟩ , Y is minimal}
Its closure is usually referred to as the entropic cone
Γl ∶= Closure(Γ○l )
Indeed, the entropic cone Γl is a closed, convex cone in R2 −1 [Yeu12]. For
l ≤ 3, the entropic cone Γl is polyhedral and completely described by Shannon
inequalities. However, for l ≥ 4, the situation is much more complicated. It is
known that Γl is not polyhedral for l ≥ 4 [Mat07]. The shape of the entropic
cone is not known as of the time of writing this article. It is an important
open problem in information theory to find tight bounds on the entropic cone
for l ≥ 4. We hope that the techniques developed in this article will eventually
lend itself to finding a useful characterization.
l
45
7. Extensions
In fact, the entropic cone can be considered as the relative entropic set of
/ ∈ Prob ⟨∅⟩, that corresponds to the empty diagram
an empty configuration O
category ∅ = Λ0
/
Γl = Γl (O)
For a diagram category G let us denote by {●} = {●} the constant GΛ
configuration of one-point spaces. Given an l-extension Y ∈ Prob ⟨Λl ⟩ of {●} k
the restriction to the last l − k terminal spaces induces a linear isomorphism
G
(7.2)
Γl ({●} k ) ≅ Γl−k
Λ
7.3. Extension lemma. The Lipschitz continuity of relative entropic sets will
follow from the following important proposition, which we will refer to as the
Extension Lemma.
Proposition 7.1. ↓ (Extension Lemma) Let k, l ∈ N, k ≤ l and let X , X ′ ∈
Prob ⟨Λk ⟩ be minimal full configurations. For every Y ∈ Extl X there exists a
Y ′ ∈ Extl X ′ such that
k(Y ′ , Y) ≤ 2l−k k(X ′ , X )
⊠
The key behind the proof of the Extension Lemma, is that there is a full
configuration Z that extends both Y and the optimal coupling between X and
X ′ . The configuration Y ′ can be chosen to be the restriction of Z to the full
configuration generated by X ′ and the terminal spaces in Y which are not in X .
The estimate directly follows from Shannon inequalities. We present details
at page 69.
It follows immediately from the Extension Lemma and the Lipschitz property of the entropy function Ent∗ that asymptotically equivalent configurations
have the same solutions to all Information-Optimization problems and, consequently, they have the same stabilized relative entropic set.
In fact, we have a much stronger statement. Both the unstabilized and stabilized relative entropic sets have a Lipschitz dependence on the configuration,
if the distance between sets is measured by the Hausdorff distance.
Let us endow the collection of subsets of El with the Hausdorff metric with
respect to the `1 -distance. For two subsets S1 , S2 of El , define the Hausdorff
distance between them by
dH (S1 , S2 ) = inf {ε > 0 ∣ S1 ⊂ S2 + Bε and S2 ⊂ S1 + Bε }
where Bε is the `1 -ball of size ε around the origin in El .
In fact, at this point the Hausdorff distance is only an extended pseudometric, in the sense, that it may take infinite values and it may vanish on pairs
of non-identical points.
Suppose now that we are given two minimal full configurations X , X ′ ∈
Prob ⟨Λk ⟩, and suppose a point y ∈ El lies in the unstabilized relative entropic
set of X , that is
y ∈ Γ○l (X )
46
8. Mixtures
This means that there is an extension Y ∈ Extl (X ) such that
Ent∗ (Y) = y
By the Extension Lemma, there exists a configuration Y ′ ∈ Extl (X ′ ) such that
k(Y, Y ′ ) ≤ 2l−k k(X , X ′ )
and by the 1-Lipschitz property of the entropy function the point y ′ ∶= Ent∗ (Y ′ )
is close to the point y, that is
∣y − y ′ ∣1 = ∣ Ent∗ (Y) − Ent∗ (Y ′ )∣1 ≤ 2l−k k(X , X ′ )
We have thus obtained the following corollary to the Extension Lemma.
Corollary 7.2. Let k ∈ N and X , X ′ ∈ Prob ⟨Λk ⟩. Then the Hausdorff distance
between their unstabilized relative entropic sets satisfies the following Lipschitz
estimate
dH (Γ○l (X ), Γ○l (X ′ )) ≤ 2l−k k(X , X ′ )
⊠
Note that in particular, the distance between unstabilized relative entropic
sets is always finite and
dH (Γ○l (X ), Γ○l ({●} k )) ≤ 2l−k k(X , {●} k ) = 2l−k ∣ Ent∗ (X )∣1 .
Λ
Λ
Let us denote by Kk,l the metric space of closed convex sets K in El such that
dH (K, Γl ({●} k )) < ∞
Λ
endowed with the Hausdorff distance.
Theorem 7.3. ↓ Let k ∈ N and X , X ′ ∈ Prob ⟨Λk ⟩. Then for all l ∈ N, the
Hausdorff distance between their stabilized relative entropic sets satisfies the
Lipschitz estimate
dH (Γl (X ), Γl (X ′ )) ≤ 2l−k κ(X , X ′ )
In other words, the map Γl from minimal full configurations in Prob ⟨Λk ⟩ to
Kk,l is 2l−k -Lipschitz.
⊠
Finally, as a primer to Section 9, note that for any set K ∈ Kk,l the sequence
n↦
1
K
n
converges in the Hausdorff distance to Γl ({●} k ). The set K ⊂ El can be
viewed as a metric space itself, by just restricting the `1 -metric to it. The
above convergence can then be expressed by saying that the asymptotic cone
Λ
of K equals Γl ({●} k ) and is isomorphic to Γl−k .
Λ
47
8. Mixtures
8. Mixtures
Mixtures provide some technical tools, which we will use in Section 9.
The input data for the mixture operation is a family of G-configurations,
parametrized by a probability space. As result one obtains another G-configuration with the pre-specified conditionals. One particular instance of a mixture
G
is when one mixes two configurations X and {●} , the latter being a constant
G-configuration of one-point probability spaces. This operation will be used
as a substitute for taking radicals “X ⊗(1/n) ” in Section 9 below.
8.1. Definition and elementary properties. Let G be a complete diagram category and Θ be a probability space. Let {Xθ }θ∈Θ be a family of
G-configurations parametrized by Θ. The mixture of the family {Xθ } is the
reduction
Mix {Xθ } = (Y Ð→ ΘG )
such that
Y∣ θ ≅ Xθ
(8.1)
The mixture exists and is uniquely defined by property (8.1) up to an isomorphism which is identity on ΘG .
We denote the top configuration of the mixture
Y = ⊕ Xθ
θ∈Θ
and also call it the mixture of the family {Xθ }.
When
Θ = Λα = ( {◻, ∎} ; p(∎) = α)
is a binary space we write simply
X∎ ⊕Λα X◻
for the mixture. The configuration subindexed by the ∎ will always be the
first summand.
The entropy of the mixture can be evaluated by the following formula
Ent∗ (⊕ Xθ ) = ∫ Ent∗ (Xθ ) d p(θ) + Ent∗ (ΘG )
θ∈Θ
Θ
Mixtures satisfy the distributive law with respect to the tensor product
Mix ({Xθ }θ∈Θ ) ⊗ Mix ({Yθ′ }θ′ ∈Θ′ ) ≅ Mix ({Xθ ⊗ Yθ′ }(θ,θ′ )∈Θ⊗Θ′ )
(⊕ Xθ ) ⊗ ( ⊕ Yθ′ ) ≅
θ∈Θ
θ′ ∈Θ′
⊕
(θ,θ ′ )∈Θ⊗Θ′
(Xθ ⊗ Yθ′ )
48
9. Tropical probability
8.2. The distance estimates. Recall that for a diagram category G we deG
note by {●} = {●} the constant G-configuration of one-point spaces.
G
The mixture of a G-configuration with {●} may serve as an ersatz of taking
radicals of the configuration. The following lemma provides a justification
of this by some distance estimates related to mixtures and will be used in
Section 9.
Lemma 8.1. ↓ Let G be a complete diagram category and X , Y ∈ Prob ⟨G⟩.
Then
(i) κ(X , X ⊗n ⊕Λ1/n {●}) ≤ Ent(Λ1/n )
(ii) κ (X , (X ⊕Λ1/n {●})⊗n ) ≤ n ⋅ Ent(Λ1/n )
(iii) κ ((X ⊗ Y) ⊕Λ1/n {●} , (X ⊕Λ1/n {●}) ⊗ (Y ⊕Λ1/n {●})) ≤ 3 Ent(Λ1/n )
1
(iv) κ ((X ⊕Λ1/n {●}), (Y ⊕Λ1/n {●})) ≤ κ(X , Y)
n
⊠
The proof can be found on page 71. Note, that the distance estimates in the
lemma above are with respect to the asymptotic Kolmogorov distance. This
is essential, since from the perspective of the intrinsic Kolmogorov distance
mixtures are very badly behaved.
9. Tropical probability spaces and their configurations
In this section we introduce the notion of tropical probability spaces and
their configurations. Configurations of tropical probability spaces are points
in the asymptotic cone of the space Prob ⟨G⟩, that is they are “limits” of
certain divergent sequences of “normal” configurations. We will first give the
construction of an asymptotic cone in an abstract context. Next, we will apply
the construction to the particular case of configurations of probability spaces.
For some background on asymptotic cones, see for instance [BBI01].
9.1. Asymptotic cones of metric spaces. The asymptotic cone captures
large-scale geometry of a metric space. Abstractly, the asymptotic cone of a
pointed metric space is the pointed Gromov-Hausdorff limit of the sequence
of spaces obtained from the given one by scaling down the metric. Of course,
convergence is in general by no means assured. Sometimes a weaker type of
convergence (using ultrafilters) is considered. Since, in our case, the asymptotic
cone can be evaluated relatively explicitly we do not give the definition of
Gromov-Hausdorff convergence or convergence with respect to an ultrafilter
here, but instead give a construction.
We would like to understand asymptotic cones of the space of configurations
of probability spaces, considered as a metric space with the pseudo-metric k or
κ. For a fixed complete diagram category G the space Prob ⟨G⟩ is a monoid
with operation ⊗. It has the additional property that shifts are non-expanding
maps. This simplifies the construction and analysis of its asymptotic cone. In
fact, as we will see later the metric κ is already asymptotic relative to k. The
49
9. Tropical probability
application of the asymptotic cone construction to the metric κ allows us to
obtain a complete metric space with a simple description of points in it.
Note that even though the monoid (Prob ⟨G⟩ , ⊗) is not Abelian it has the
property that for any X1 , X2 ∈ Prob ⟨G⟩ one has
k(X1 ⊗ X2 , X2 ⊗ X1 ) = 0
Thus, from a metric perspective it is as good as being Abelian.
9.1.1. Metrics versus pseudo-metrics. A pseudo-metric δ on a set X is a bivariate function satisfying all the axioms of a distance function except it is nonnegative definite rather than positive definite. That is, the pseudo-distance
function is allowed to vanish on pairs of non-identical points. A set equipped
with a pseudo-metric will be called a pseudo-metric space. An isometry of
such spaces is a distance-preserving map, such that for any point in the target
space there is a point in the image, which is distance zero from it. Given such
an pseudo-metric space (X, δ) one could always construct an isometric metric
space (X/δ=0 , δ) by identifying all pairs of points that are distance zero apart.
Any property formulated in terms of the pseudo-metric holds simultaneously
for a pseudo-metric space and its metric quotient. It will be convenient for
us to construct pseudo-metrics on spaces instead of passing to the quotient
spaces.
9.1.2. Asymptotic cone of a metric Abelian monoid. Let (Γ, ⊗, δ) be a monoid
with a pseudo-metric δ, which satisfies the following properties
(i) The shifts
⋅ ⊗ γ ′ ∶ Γ → Γ, γ ↦ γ ⊗ γ ′
are non-expanding for any γ ′ ∈ Γ
(ii) For any γ, γ ′ ∈ Γ holds
δ(γ ⊗ γ ′ , γ ′ ⊗ γ) = 0
We will call a monoid with pseudo-metric that satisfies these conditions a
metric Abelian monoid. It follows from the shift-invariance property that for
any γ1 , γ2 , γ3 ∈ Γ holds
(9.1)
δ(γ1 ⊗ γ3 , γ2 ⊗ γ3 ) ≤ δ(γ1 , γ2 )
and for any quadruple γ1 , γ2 , γ3 , γ4 ∈ Γ holds
(9.2)
δ(γ1 ⊗ γ2 , γ3 ⊗ γ4 ) ≤ δ(γ1 , γ3 ) + δ(γ2 , γ4 )
and, in particular, the monoid operation is 1-Lipschitz with respect to each
argument.
As a direct consequence, for every n ∈ N, and γ1 , γ2 ∈ Γ also holds
(9.3)
δ(γ1⊗n , γ2⊗n ) ≤ n δ(γ1 , γ2 )
For a sequence γ = {γ(i)} ∈ ΓN0 define its defect with respect to the distance
function δ by
Defectδ (γ) = sup δ (γ(i + j), γ(i) ⊗ γ(j))
i,j∈N0
50
9. Tropical probability
The sequence γ will be called δ-linear if Defectδ (γ) = 0, and δ-quasi-linear
if Defectδ (γ) < ∞. Denote by Lδ (Γ) and QLδ (Γ) the sets of all linear and,
respectively, quasi-linear sequences in Γ with respect to the distance δ.
For two elements γ 1 , γ 2 ∈ QLd (Γ), define an asymptotic distance between
them by
1
δ̂(γ 1 , γ 2 ) ∶= lim δ (γ1 (n), γ2 (n))
n→∞ n
Lemma 9.1. ↓ For a pair γ 1 , γ 2 ∈ QLδ (Γ) the limit
1
δ (γ1 (n), γ2 (n))
n→∞ n
lim
⊠
exists and is finite.
We provide the proof in Section T on page 72.
The bivariate function δ̂ is a pseudo-distance on the set QLδ (Γ). We call
two sequences γ 1 , γ 2 ∈ QLδ (Γ) asymptotically equivalent if δ̂(γ 1 , γ 2 ) = 0 and
write
δ̂
γ1 = γ2
We will call a sequence γ weakly quasi-linear, if it is asymptotically equivalent to a quasi-linear sequence. Note that the space of all weakly quasi-linear
sequences can also be endowed with the asymptotic distance and it is isometric to the space of quasi-linear sequences. As we will see later all the natural
operations we consider are δ̂-Lipschitz and therefore coincide for the asymptotically equivalent sequences. Thus given a weakly quasi-linear sequence we
could always replace it by an equivalent quasi-linear sequence without any
visible effect. Thus, we take the liberty to omit the adverb “weakly”. Whenever we say quasi-linear sequence, we mean a weakly quasi-linear sequence,
that is silently replaced by an asymptotically equivalent genuine quasi-linear
sequence, if necessary.
The validity of the following constructions is very easy to verify, so we omit
the proofs.
The set QLδ (Γ) admits an action of the multiplicative semigroup (R≥0 , ⋅ )
defined in the following way. Let λ ∈ R≥0 and γ = {γ(n)} ∈ QLδ (Γ). Then
define the action of λ on γ by
(9.4)
γ λ ∶= {γ(⌊λ ⋅ n⌋)}n∈N0
This is only an action up to asymptotic equivalence. Similarly, in the constructions that follow we are tacitly assuming they are valid up to asymptotic
equivalence.
The action
⋅ ∶ R≥0 × (QLδ (Γ), δ̂) → (QLδ (Γ), δ̂)
is continuous with respect to δ̂ and, moreover it is a homothety (dilation), that
is
δ̂(γ λ1 , γ λ2 ) = λ ⋅ δ̂(γ 1 , γ 2 )
51
9. Tropical probability
The group operation ⊗ on Γ induces a δ̂-continuous (in fact, 1-Lipschitz)
group operation on QLδ (Γ) by multiplying sequences element-wise. The semigroup structure on QLδ (Γ) is distributive with respect to the R≥0 -action
(γ 1 ⊗ γ 2 )λ = γ λ1 ⊗ γ λ2
γ λ1 +λ2 = γ λ1 ⊗ γ λ2
δ̂
In particular for n ∈ N
γ n = γ ⊗n
δ̂
The path
⊗ γ λ2
[0, 1] ∋ λ ↦ γ 1−λ
1
will be called a convex interpolation and is a constant-speed δ̂-geodesic between
γ 1 and γ 2 , that is for λ ∈ [0, 1],
(1−λ)
⊗ γ λ2 , γ 1 ) = λ δ̂(γ 1 , γ 2 )
(1−λ)
⊗ γ λ2 , γ 2 ) = (1 − λ) δ̂(γ 1 , γ 2 )
δ̂(γ 1
δ̂(γ 1
9.1.3. Conditions for completeness. We would like to call
(∞)
Γδ
∶= (QLδ (Γ), ⊗, ⋅, δ̂)
the asymptotic cone of (Γ, ⊗, δ). However it is not clear in general, whether
(∞)
Γδ is a complete space.
We can simply consider the metric completion, and call it the asymptotic
cone of (Γ, ⊗, δ). We feel, however, that it adds just another level of obscurity
(∞)
as to what the points of Γδ are.
Under some circumstances, however, the completeness of the space of quasilinear sequences comes for free. This is the subject of the proposition below.
Suppose the metric Abelian monoid (Γ, ⊗, δ) has an additional property:
There exists a constant C > 0, such that for any quasi-linear sequence γ ∈
QLδ (Γ), there exists an asymptotically equivalent quasi-linear sequence γ ′ with
defect bounded by C. If this is the case, we say that the metric monoid (Γ, ⊗, δ)
has the (C-)uniformly bounded defect property.
Proposition 9.2. ↓ Suppose (Γ, ⊗, δ) is a metric Abelian monoid such that
(i) the distance function δ is homogeneous, that is for any γ1 , γ2 ∈ Γ and
n ∈ N0
δ(γ1⊗n , γ2⊗n ) = n ⋅ δ(γ1 , γ2 )
(ii) (Γ, ⊗, δ) has the uniformly bounded defect property.
Then the space (QLδ (Γ), δ̂) is complete.
The proof of the proposition can be found on page 72.
⊠
52
9. Tropical probability
9.1.4. On the density of linear sequences. In Section 6 we have shown that
Bernoulli sequences of configurations can be approximated by sequences of
homogeneous configurations. The proposition below will allow us to extend
this statement to a wider class of sequences. It gives a sufficient condition
under which the linear sequences are dense in the quasi-linear sequences.
Proposition 9.3. ↓ Suppose (Γ, ⊗, δ) has the ε-uniformly bounded defect property for every ε > 0. Then Lδ (Γ) is dense in QLδ (Γ)
⊠
See page 74 for the proof.
9.1.5. Asymptotic metric on original semigroup. Starting with an element γ ∈
→
Γ one can construct a linear sequence γ = {γ ⊗i }i∈N0 . In view of inequality (9.3),
this map is a contraction
(9.5)
(Γ, δ ) → ( Lδ (Γ), δ̂ )
By the inclusions in (9.5) we have an induced metric δ̂ on Γ, satisfying for
any γ1 , γ2 ∈ Γ
(9.6)
δ̂(γ1 , γ2 ) ≤ δ(γ1 , γ2 )
and the following scale-invariance condition is gained
(9.7)
δ̂(γ1⊗n , γ2⊗n ) = n ⋅ δ̂(γ1 , γ2 )
for all n ∈ N0 .
Note moreover that if δ was scale-invariant to begin with, then δ̂ coincides
with δ on Γ.
9.1.6. Iteration of construction. We may now iterate the constructions above,
that is, we may apply them to (Γ, δ̂) instead of (Γ, δ). One may wonder what
is the purpose. However, we have already observed that δ̂ satisfies the scaleinvariance condition (9.7), which is one of the conditions going into a proof
of completeness in Proposition 9.2. Moreover, when we will later apply the
theory in this section to the particular case of Γ = Prob ⟨G⟩, we will see that
(Γ, δ̂) = (Prob ⟨G⟩ , κ)
and we will show that the latter space has the ε-uniformly bounded defect
property for every ε > 0.
By virtue of the bound δ̂ ≤ δ, sequences that are quasi-linear with respect
to δ̂, are also quasi-linear with respect to δ. Since δ̂ is scale-invariant, the
ˆ
associated asymptotic distance δ̂ coincides with δ̂ on Γ. We will show (in
ˆ
Lemma 9.4 below) that δ̂ also corresponds to δ̂ on δ-quasi-linear sequences.
In order to organize all these statements, and to be more precise, let us
include the spaces in the following commutative diagram.
53
9. Tropical probability
( Lδ (Γ), δ̂)
f
(9.8)
(Γ, δ̂)
1
( QLδ (Γ), δ̂)
ı1
ı2
ϕ
ˆ
( Lδ̂ (Γ), δ̂)
2
ˆ
( QLδ̂ (Γ), δ̂)
The maps f, ϕ and ı1 are isometries. The maps 1 and 2 are isometric
embeddings. The next lemmas show that ı2 is also an isometric embedding,
and it has dense image.
Lemma 9.4. ↓ The natural inclusion
ˆ
ı2 ∶ (QLδ (Γ), δ̂) ↪ (QLδ̂ (Γ), δ̂)
is an isometric embedding.
⊠
Lemma 9.5. ↓ The image of the isometric embedding
ˆ
ı2 ∶ (QLδ (Γ), δ̂) ↪ (QLδ̂ (Γ), δ̂)
ˆ
is dense in (QLδ̂ (Γ), δ̂)
⊠
The proofs of the two lemmas above are to be found on page 75.
9.2. Tropical probability spaces and configurations. Now we apply the
above construction to the space of complete configurations with fixed combinatorial type G.
Fix a complete diagram category G and consider the space Prob ⟨G⟩ of
configurations modeled on G. It carries the following structures:
(i) A pseudo-metric k or κ.
(ii) A 1-Lipschitz tensor product ⊗.
(iii) A 1-Lipschitz entropy function Ent∗ ∶ Prob ⟨G⟩ → R[[G]] .
The tensor product of configurations is commutative from a metric perspective. Recall that in Corollary 4.5 the subadditivity of both k and κ was
established, namely for any X , Y, U, V ∈ Prob ⟨G⟩ holds
k(X ⊗ U, Y ⊗ V) ≤ k(X , Y) + k(U, V).
and
κ(X ⊗ U, Y ⊗ V) ≤ κ(X , Y) + κ(U, V).
The space (Prob ⟨G⟩ , ⊗, k) is a metric Abelian monoid. Note also that
k̂ = κ on Prob ⟨G⟩, along the lines of Section 9.1.5.
However, the metric k is not scale-invariant. Moreover, it is unclear whether
the metric semigroup (Prob ⟨G⟩ , ⊗, k) has the uniformly bounded defect property. This is why we iterate the construction, as announced in Section 9.1.6,
and consider the space of κ-quasi-linear sequences instead.
54
9. Tropical probability
Lemma 9.6. ↓ For a complete diagram category, and for every ε > 0, the
space (Prob ⟨G⟩ , ⊗, κ) has the ε-uniformly bounded defect property, that is for
any κ-quasi-linear sequence X ∈ QLκ (Prob ⟨G⟩) there exists an asymptotically
equivalent sequence Y with defect not exceeding ε.
⊠
By applying the general setup in the previous section to the metric semigroups (Prob ⟨G⟩ , ⊗, k) and (Prob ⟨G⟩ , ⊗, κ) and as a corollary to Lemma
9.6 we obtain the following theorem.
Theorem 9.7. Consider the commutative diagram
(9.9)
( Lk (Prob ⟨G⟩), k̂)
f
(Prob ⟨G⟩ , κ)
1
( QLk (Prob ⟨G⟩), k̂)
ı1
ϕ
ı2
( Lκ (Prob ⟨G⟩), κ̂)
2
( QLκ (Prob ⟨G⟩), κ̂)
Then the following statements hold:
(i) The maps f, ϕ, ı1 are isometries.
(ii) The maps ı2 , 1 , 2 are isometric embeddings and each map has a dense
image in the corresponding target space.
(iii) The space in the lower-right corner, ( QLκ (Prob ⟨G⟩), κ̂), is complete.
⊠
We may finally define the space of tropical G-configurations, as the space in
the lower-right corner of the diagram
(∞)
Prob ⟨G⟩
∶= ( QLκ (Prob ⟨G⟩), ⊗, ⋅, κ̂)
By the Theorem 9.7 above, this space is complete.
The entropy function Ent∗ ∶ Prob ⟨G⟩ → R[[G]] extends to a linear functional
(∞)
Ent∗ ∶ Prob ⟨G⟩
→ (R[[G]] , ∣ ⋅ ∣1 )
of norm one, defined by
1
Ent∗ (X (n))
n→∞ n
Ent∗ (X ) = lim
9.2.1. Sequences of homogeneous configurations are dense. Let L̃k (Prob ⟨G⟩h )
stand for the weakly linear sequences of homogeneous configurations, that is
those sequences, that are asymptotically equivalent to a linear sequence (not
necessarily of homogeneous spaces).
For a sequence of homogeneous spaces H ∈ L̃k (Prob ⟨G⟩h ) define aep(H)
to be a k-linear sequence asymptotically equivalent to H.
55
9. Tropical probability
Now we can extend the commutative diagram (9.9) as follows
L̃k (Prob ⟨G⟩h )
(9.10)
(Prob ⟨G⟩ , κ)
aep
f
Lk (Prob ⟨G⟩)
1
QLk (Prob ⟨G⟩)
ı2
ı1
ϕ
Lκ (Prob ⟨G⟩)
2
(∞)
Prob ⟨G⟩
By the Asymptotic Equipartition Property for configurations, Theorem 6.1,
the map aep is an isometry, hence we have the following theorem.
Theorem 9.8. The map
(∞)
2 ○ ı1 ○ aep ∶ L̃k (Prob ⟨G⟩h ) → Prob ⟨G⟩
⊠
is an isometric embedding with dense image.
(∞)
(∞)
Let Prob ⟨G⟩h ⊂ Prob ⟨G⟩
denote the space of weakly quasi-linear
(∞)
sequences of configurations H ∈ Prob ⟨G⟩ , such that for every n ∈ N0 , the
(∞)
configuration H(n) is homogeneous. We will refer to Prob ⟨G⟩h as the space
of homogeneous tropical configurations.
Denote by aep the embedding
(∞)
aep ∶ Prob ⟨G⟩h
(∞)
↪ Prob ⟨G⟩
Theorem 9.9. (Asymptotic Equipartition Theorem for tropical configurations) Let G be a complete diagram category. Then the map
(∞)
aep ∶ Prob ⟨G⟩h
(∞)
↪ Prob ⟨G⟩
⊠
is an isometry.
(∞)
Proof: We need to show that for every tropical configuration X ∈ Prob ⟨G⟩ ,
(∞)
there exists a homogeneous tropical configuration H ∈ Prob ⟨G⟩h such that
κ̂(H, X ) = 0
By Lemma 9.6 and Proposition 9.3, for every j ∈ N there exists a sequence
Y j ∈ Lκ (Prob ⟨G⟩) ≅ Lk (Prob ⟨G⟩) such that
κ̂(Y j , X ) ≤
1
j
By the Asymptotic Equipartition Property for configurations, Theorem 6.1,
there are sequences of homogeneous configurations Hj such that
κ̂(Y j , Hj ) = 0
Define i(j) such that for all k ≥ i(j)
1
1
κ(Yj (k), Hj (k)) ≤
k
j
56
9. Tropical probability
and moreover
1
2
κ(Yj (k), X (k)) ≤
k
j
The function i can be chosen monotonically increasing. For every i ∈ N0 there
is a unique j(i) ∈ N0 such that
i(j(i)) ≤ i < i(j(i) + 1)
Define then
H(k) = Hj(k) (k)
It follows that for k > i(1),
3
1
κ(H(k), X (k)) ≤
k
j(k)
Since j is a non-decreasing, divergent sequence, the theorem follows.
⊠
Thus we have shown that all arrows in diagram (9.10) are isometric embeddings with dense images. We would like to conjecture, that, in fact, they are
all isometries. In any case, the difference between metrics κ̂ and κ is so small
(κ is defined on the dense subset of the domain of definition of κ̂ and they
coincide whenever both are defined), that we will not write the hat anymore
(∞)
and just use notation κ for the metric on Prob ⟨G⟩ .
9.3. Tropical probability spaces and tropical chains. In this section we
(∞)
evaluate the spaces Prob(∞) and Prob ⟨Cn ⟩ , where Cn is a chain, which is
the diagram category introduced in 2.1.2 on page 14.
Recall that a finite probability space U is homogeneous if Aut(U ) acts transitively on the support of the measure. The property of being homogeneous is
invariant under isomorphism and every homogeneous space is isomorphic to a
probability space with the uniform distribution.
Homogeneous chains also have a very simple description. A chain of reductions is homogeneous, if and only if all the individual spaces are homogeneous.
This simple description allows us to evaluate explicitly the Kolmogorov distance on the spaces of weakly linear sequences of homogeneous chains and
consequently the space of tropical chains.
Theorem 9.10.
(i) Prob(∞) ≅ (R≥0 , ∣ ⋅ − ⋅ ∣, +, ⋅)
(∞)
(ii) Prob ⟨Cn ⟩
⎧
⎫
RRRRRR
x
⎪
⎪
⎪
⎪
RR
⎪⎛ 1 ⎞
⎪
n RRRR
≅ ⎨⎜ ⋮ ⎟ ∈ R RRRRRR 0 ≤ xn ≤ ⋯ ≤ x1 ⎬
⎪
⎪
RRRRRR
⎪ ⎝ xn ⎠
⎪
⎪
⎪
⎩
⎭
RR
where the right-hand side is a cone in (Rn , ∣ ⋅ ∣1 ).
⊠
57
9. Tropical probability
To prove the Theorem 9.10 we evaluate first the isometry class of the space
of weakly linear sequences of homogeneous spaces (or chains). We will only
present an argument for single spaces, since the argument for chains is very
similar.
Lemma 9.11.
L̃k (Probh ) ≅ (R≥0 , ∣ ⋅ − ⋅ ∣, +, ⋅)
⊠
Note that the right-hand side is a complete metric space, thus the Asymptotic Equipartition Theorem for tropical configurations, Theorem 9.8, together
with Lemma 9.11, imply Theorem 9.10.
To prove Lemma 9.11 we need to evaluate the Kolmogorov distance between
two homogeneous spaces, or chains of homogeneous spaces. This is the subject
of the next lemma, from which Lemma 9.11 follows immediately.
Lemma 9.12. ↓ Denote by Un a finite uniform probability space of cardinality
n, then
(i)
n
k(Un , Um ) ≤ 2 ln 2 + ∣ln ∣
m
(ii)
κ(Un , Um ) = ∣ Ent(Un ) − Ent(Um )∣
⊠
9.4. Stochastic processes. Often, stochastic processes naturally give rise to
κ-quasi-linear sequences. We include this last subsection as an indication that
our statements, together with the construction of the tropical cone, have a
much larger reach than sequences of independent random variables. We will
be brief, and come back to the topic in a subsequent article.
For a minimal diamond configuration
C
A
B
D
we define the conditional mutual information between A and B given D by
I(A; B ∣ D) ∶= Ent(A) + Ent(B) − Ent(C) − Ent(D)
Shannon’s inequality (1.2) says that the conditional mutual information is
always non-negative. Any minimal two-fan A ← C → B can be completed to a
diamond with the one-point probability space {●} as the terminal vertex, and
the mutual information between A and B is defined as
I(A; B) = I(A; B ∣ {●}) = Ent(A) + Ent(B) − Ent(C)
58
T. Technical proofs
Let
. . . , X−1 , X0 , X1 , . . .
be a stationary stochastic process with finite state space X. Thus, for any I =
{k, k + 1, . . . , l} ⊂ Z we have jointly distributed random variables Xk , . . . , Xl ,
that generate a full configuration
XI = ⟨Xk , . . . , Xl ⟩ = {XJ }J⊂I
as explained in Section 2.1.4. This collection of full configurations is consistent
in the sense that for k ≤ k ′ ≤ l′ ≤ l there are canonical isomorphisms
XI ′ ≅ RI∗′ ,I XI
where I ∶= {k, . . . , l}, I ′ ∶= {k ′ , . . . , l′ } and RI∗′ ,I is the restriction operator
introduced in Section 2.4.1.
The property of being stationary means that there are canonical isomorphisms for any finite subset I ⊂ Z and l ∈ Z
≅
XI Ð→ XI+l
For I = {k, k + 1, . . . , l} we call the initial space XI of the configuration XI
the space of trajectories of the process over I and denote it Xkl .
Note that by stationarity, for every m ∈ Z, k ∈ N and l ∈ N0 ,
0
m+k−1
m+k+l−1
; X1l )
) = I(X−k+1
I(Xm
; Xm+k
Moreover the right-hand side is an increasing function of both k and l. We
make the following important observation. The defect of the sequence n ↦ X1n
is equal to
Defectκ ({X1n }) = sup κ (X1m+n , X1m ⊗ X1n )
m,n∈N0
= sup ∣Ent(X1m ⊗ X1n ) − Ent(X1m+n )∣
m,n∈N0
0
⊗ X1n ) − Ent(X1m+n )∣
= sup ∣Ent(X−m+1
m,n∈N0
0
= sup I(X−m+1
, X1n )
m,n∈N0
Therefore, the sequence n ↦ X1n is κ-quasi-linear if and only if
(9.11)
0
lim I(X−k+1
, X1l ) < ∞
k,l→∞
Once condition (9.11) is satisfied for a stochastic process, it defines a tropical
probability space X ∈ Prob.
Note that condition (9.11) is satisfied for any stationary, finite-state Markov
chains.
T. Technical proofs
This section contains some proofs that did not make it into the main text.
59
T. Technical proofs
T.2. Statements from the section “Configurations”.
Lemma 2.1. ↑ Let G be a diagram category. Then
(i) A two-fan F = (X ← Z → Y) of G-configurations is minimal, if and
only if the constituent two-fans of probability spaces Fi = (Xi ← Zi →
Yi ) are all minimal.
(ii) For any two-fan F = (X ← Z → Y) of G-configurations its minimal
reduction exists, that is, there exists a minimal two-fan F ′ = (X ←
Z ′ → Y) included in the following diagram
Z
Z′
X
Y
⊠
Proof: We will need the following lemma
Lemma T.1. Suppose we are given two two-fans of probability spaces
β
α
F = (X ←Ð Z Ð→ Y )
β ′′
α′′
F ′′ = (X ′′ ←Ð Z ′′ Ð→ Y ′′ )
such that F ′′ is minimal. Let
β′
α′
µ
F Ð→ F ′=(X ←ÐZ ′ Ð→Y )
be a minimal reduction of F. Then for any reduction ρ ∶ F → F ′′ , there exists
a reduction ρ′ ∶ F ′ → F ′′ such that ρ = ρ′ ○ µ
⊠
Proof: We define ρ′ on the terminal spaces of F ′ to coincide with ρ.
To prove the lemma we just need to provide a dashed arrow that makes the
following diagram commutative
Z
µ
α
α′
ρ
β
ρ′
Z′
Z ′′
β′
X
α′′
Y
ρ=ρ′
X ′′
β ′′
Y ′′
ρ=ρ′
The reduction ρ′ is constructed by simple diagram chasing and by using the
minimality of F ′′ . Suppose z ′ ∈ Z ′ and z1 , z2 ∈ Z are such that z ′ = µ(z1 ) =
µ(z2 ). By commutativity of the solid arrows in the diagram above, we have
α′′ ○ ρ(z1 ) = ρ ○ α′ ○ µ(z1 ) = ρ ○ α′ ○ µ(z2 ) = α′′ ○ ρ(z2 )
60
T. Technical proofs
Similarly
β ′′ ○ ρ(z1 ) = β ′′ ○ ρ(z2 )
Thus by minimality of F ′′ it follows that ρ(z1 ) = ρ(z2 ). Hence, ρ′ can be
constructed by setting ρ′ (z ′ ) = ρ(z1 ).
⊠
Now we proceed to prove claim (i) of Lemma 2.1. Let G = {Oi ; mij } be
a diagram category, X , Y, Z ∈ Prob ⟨G⟩ be three G-configurations and F =
(X ← Z → Y) be a two-fan. Recall that it can also be considered as a Gconfiguration of two-fans
F = {Fi ; fij }
Any minimizing reduction
F =(X ←Z →Y) Ð→ F ′ =(X ←Z ′ →Y)
induces reductions
Fi =(Xi ←Zi →Yi ) Ð→ Fi =(Xi ←Zi′ →Yi )
for all i in the index set I. It follows that if all Fi ’s are minimal, then so is F.
Now we prove the implication in the other direction. Suppose F is minimal.
We have to show that all Fi are minimal as well. Suppose there exist a nonminimal fan among Fi ’s. For an index i ∈ I let
ˇ ∶= {j ∈ I ∣ HomG (Oj , Oi ) ≠ ∅}
J(i)
ˆ ∶= {j ∈ I ∣ HomG (Oi , Oj ) ≠ ∅}
J(i)
Choose an index i0 such that
(i) Fi0 is not minimal
ˆ 0 )/{i0 } the two-fan Fj is minimal.
(ii) for any j ∈ J(i
Consider now the minimal reduction µ ∶ Fi0 → Fi′0 and construct a two-fan
G = {Gi ; gij } of G-configurations by setting
Gi ∶= {
and
where fi′0 j
Fi′
Fi
if i = i0
otherwise
⎧
ˇ 0)
⎪
µ ○ fij if j = i0 and i ∈ J(i
⎪
⎪
⎪ ′
ˆ 0)
gij ∶= ⎨fij
if i = i0 and j ∈ J(i
⎪
⎪
⎪
⎪
otherwise
⎩fij
is the reduction provided by the Lemma T.1 applied to the diagram
F i0
fi 0 j
µ
Fi′0
fi′
0j
Fj
We thus constructed a non-trivial reduction F → G which is identity on the
terminal G-configurations X and Y. This contradicts the minimality of F.
61
T. Technical proofs
To address the second assertion of the Lemma 2.1 observe that the argument
above gives an algorithm for the construction of a minimal reduction of any
two-fan of G-configurations.
⊠
T.4. Statements from the section “Kolmogorov distance”.
Proposition 4.1. ↑ Let G be a complete diagram category. Then the bivariate
function
k ∶ Prob ⟨G⟩ × Prob ⟨G⟩ → R
is a pseudo-distance on Prob ⟨G⟩.
Moreover, two configurations X , Y ∈ Prob ⟨G⟩ satisfy k(X , Y) = 0 if and only
if X is isomorphic to Y in Prob ⟨G⟩.
⊠
Proof: The symmetry of k is immediate. The non-negativity of k follows
from the fact that entropy of the target space of a reduction is not greater
then the entropy of the domain, which is a particular instance of the Shannon
inequality (1.2).
We proceed to prove the triangle inequality. We will make use of the following lemma
Lemma T.2. For a minimal full configuration of probability spaces
⎛
⎜
⎜
⎜
⟨X, Y, Z⟩ = ⎜ XY
⎜
⎜
⎜
⎝ X
XY Z
XZ
Y
⎞
⎟
⎟
⎟
YZ ⎟
⎟
⎟
⎟
Z ⎠
holds
kd(X ← XZ → Z) ≤ kd(X ← XY → Y ) + kd(Y ← Y Z → Z)
⊠
Proof: By Shannon inequality (1.2) on page 13 we have
Ent(X∣ Z) ≤ Ent(XY ∣ Z) ≤ Ent(X∣ Y ) + Ent(Y ∣ Z)
Similarly,
Ent(Z∣ X) ≤ Ent(Z∣ Y ) + Ent(Y ∣ X)
and therefore
kd(X ← XZ → Z) ≤ kd(X ← XY → Y ) + kd(Y ← Y Z → Z)
⊠
Now we continue with the proof of Proposition 4.1.
Let G be an arbitrary complete reduction category. Suppose X = {Xi ; fij },
Y = {Yi ; gij } and Z = {Zi ; hij } are G-configurations, with initial spaces being
X0 , Y0 and Z0 , respectively. Let
F̂ = (X ← F → Y)
Ĝ = (Y ← G → Z)
62
T. Technical proofs
be two optimal minimal two-fans satisfying
k(X , Y) = kd(F̂)
k(Y, Z) = kd(Ĝ)
Recall that each two-fan of G-configurations is a G-configuration of two-fans
between the individual spaces, that is
F = {Fi = (Xi ← Fi → Yi )}
G = {Gi = (Yi ← Gi → Zi )}
We construct a coupling H between X and Z in the following manner.
Starting with the two-tents configuration between the initial spaces, we use
adhesion to extend it to a full configuration, thus constructing a coupling
between X0 and Z0 . This full configuration could then be “pushed down”
and provides full extensions of two-tents on all lower levels. Thus we could
“compose” couplings F and G and use a Shannon inequality to establish the
triangle inequality for the Kolmogorov distance. Details are as follows.
Consider a two-tents configuration
X0 ← F0 → Y0 ← G0 → Z0
and extend it by adhesion, as described in Section 2.5 to a Λ3 -configuration
A0
F0
H0
G0
X0
Y0
Z0
Together with the reductions
(X0 )G → X
(Y0 )G → Y
(Z0 )G → Z
it gives rise to a Λ3 -configuration of G-configurations
(A0 )G
(T.1)
(F0 )G
(H0 )G
(G0 )G
X
Y
Z
Note that the minimal reductions of the two-fan subconfigurations of (T.1)
X ← (F0 )G → Y
Y ← (G0 )G → Z
are the two-fans F̂ and Ĝ, respectively, by Lemma 2.1.
63
T. Technical proofs
Now consider the “minimization” of the above configuration
A
(T.2)
F
H
G
X
Y
Z
It could also be viewed as a G-configuration of Λ3 configurations,
Ai
Fi
Hi
Gi
Xi
Yi
Zi
each of which is minimal by Corollary 2.2.
Now we can apply Lemma T.2 to each level to conclude that
k(X , Z) ≤ kd(X ← H → Z)
≤ kd(X ← F → Y) + kd(X ← G → Y)
= k(X , Y) + k(Y, Z)
Finally, if k(X , Y) = 0, then there is a two-fan F of G-configurations between
X and Y with kd(F) = 0, from which it follows that X and Y are isomorphic.⊠
Proposition 4.3. ↑ Let G be a complete diagram category. Then with respect
to the Kolmogorov distance on Prob ⟨G⟩ the tensor product
⊗ ∶ (Prob ⟨G⟩ , k)2 → (Prob ⟨G⟩ , k)
is 1-Lipschitz in each variable, that is, for every triple X , Y, Y ′ ∈ Prob ⟨G⟩ the
following bound holds
k(X ⊗ Y, X ⊗ Y ′ ) ≤ k(Y, Y ′ )
⊠
Proof: The claim follows easily from the additivity of entropy in equation
(1.3). Suppose that X = {Xi ; fij }, Y = {Yi ; gij } and Y ′ = {Yi′ ; gij′ } are three
G-configurations and
F = (Y ← Z → Y ′ )
is an optimal fan, so that
k(Y, Y ′ ) = ∑ [2 Ent(Zi ) − Ent(Yi ) − Ent(Yi′ )]
i
Consider the fan
G = (X ⊗ Y ← X ⊗ Z → X ⊗ Y ′ )
64
T. Technical proofs
Then, by additivity of entropy, in equation (1.3), we have
kd(F) = ∑ [2 Ent(Xi ⊗ Zi ) − Ent(Xi ⊗ Yi ) − Ent(Xi ⊗ Yi′ )]
i
= ∑ [2 Ent(Zi ) − Ent(Yi ) − Ent(Yi′ )]
i
= kd(G)
and, therefore,
k(X ⊗ Y, X ⊗ Y ′ ) ≤ kd(G) = kd(F) = k(Y, Y ′ )
Thus, the tensor product of probability spaces is 1-Lipschitz with respect to
each argument.
⊠
Proposition 4.7. ↑ Suppose G is a complete diagram category and δ = k, κ is
either Kolmogorov distance or asymptotic Kolmogorov distance on Prob ⟨G⟩.
Then the entropy function
Ent∗ ∶ (Prob ⟨G⟩ , δ) → (R[[G]] , ∣ ⋅ ∣1 ),
X = {Xi , fij } ↦ (EntXi )i ∈ R[[G]]
⊠
is 1-Lipschitz.
Proof:
Let X , Y ∈ Prob ⟨G⟩ and let
G = (X ← Z → Y)
be an optimal fan with components
Gi = (Xi ← Zi → Yi )
For a fixed index i we can estimate the difference of entropies
Ent(Xi ) − Ent(Yi ) = 2( Ent(Xi ) − Ent(Zi )) + kd(Gi ) ≤ kd(Gi )
By symmetry we then have
∣ Ent(Xi ) − Ent(Yi )∣ ≤ kd(Gi )
Adding above inequalities for all i we have
∣ Ent∗ (X ) − Ent∗ (Y)∣1 ≤ kd(G) = k(X , Y)
By the additivity of entropy we also obtain the 1-Lipschitz property of the
entropy function with respect to the asymptotic Kolmogorov distance κ. ⊠
Proposition 4.8. ↑ Suppose R ∶ G′ → G is a functor between two complete diagram categories and δ stands for either Kolmogorov or asymptotic Kolmogorov
distance. Then the restriction operator
R∗ ∶ (Prob ⟨G⟩ , δ) → (Prob ⟨G′ ⟩ , δ),
is Lipschitz.
X ↦X ○R
⊠
65
T. Technical proofs
Proof: The claim follows from the functoriality of the restriction operator. We
argue as follows.
Suppose that R ∶ G′ → G is a functor and R∗ is the corresponding restriction
operator. For X1 , X2 ∈ Prob ⟨G⟩ let
F = (X1 ← Y → X2 )
be an optimal fan. Then
F ′ ∶= (R∗ X1 ← R∗ Y → R∗ X2 )
is a fan with the terminal vertices being the restrictions of X1 and X2 . It can
be considered as a G′ -configuration of two-fans over individual spaces in R∗ X1
and R∗ X2 each of which also appears as a fan in F.
Thus, we obtain the rough estimate
k(R∗ X1 , R∗ X2 ) ≤ [[G′ ]] ⋅ k(X1 , X2 )
Since the restriction operator commutes with tensor powers, the same estimate
also holds for the asymptotic Kolmogorov distance κ.
⊠
Proposition 4.9. ↑ (Slicing Lemma) Suppose G is a complete diagram category and we are given X , Xˆ , Y, Ŷ ∈ Prob ⟨G⟩ – four G-configurations and
U, V, W ∈ Prob – probability spaces, that are included into the following threetents configuration
Xˆ
X
Ŷ
W
U
V
Y
such that the two-fan (U ← W → V ) is minimal. Then the following estimate
holds
k(X , Y) ≤ ∫ k(X ∣ u, Y∣ v) d pW (u, v)
W
+ [[G]] ⋅ kd(U ← W → V )
+ ∑ [ Ent(U ∣ Xi ) + Ent(V ∣ Yi )]
i
⊠
Proof: Since the two-fan (U ← W → V ) is minimal the probability space
W could be considered having underlying set to be a subset of the Cartesian
product of the underlying sets of U and V . For any pair (u, v) ∈ W with a
positive weight consider an optimal two-fan
(T.3)
πX
πY
Guv = (X ∣ u ←Ð Zuv Ð→ Y∣ v)
where Zuv = {Zuv,i ; ρij }. Let puv,i be the probability distributions on Zuv,i
– the individual spaces in the configuration Zuv . The next step is to take
a convex combination of distributions puv,i weighted by pW to construct a
coupling X ← Z → Y.
66
T. Technical proofs
First we extend the 7-vertex configuration to a full Λ4 -configuration of Gconfigurations, such that the top vertex has the distribution
pi (x, y, u, v) ∶= puv,i (x, y)pW (u, v)
(T.4)
as described in the Section 2.1.4.
If we integrate over y, we obtain
∑ pi (x, u, v, y) = ((πX ,i )∗ puv,i )(x)pW (u, v)
y
Then we use that by (T.3) it holds that (πX ,i )∗ puv,i = pXi ( ⋅ ∣ u) and therefore
∑ pi (x, y, u, v) = pXi (x∣ u)pW (u, v).
y
In the same way,
∑ pi (x, y, u, v) = pYi (y∣ v)pW (u, v).
x
Note that this exactly corresponds to adhesion, as described in Section 2.5. It
follows that
X ∣ uv = X ∣ u and Y∣ uv = Y∣ v
(T.5)
and
(T.6)
Ent(Xi ∣ U V ) = Ent(Xi ∣ U ) and
Ent(Yi ∣ U V ) = Ent(Yi ∣ V )
The extended configuration contains a two-fan of configurations F = (X ←
Z → Y) with terminal vertices X and Y. We call its initial vertex Z =
{XYi , fij }.
The following estimates conclude the proof the the Slicing Lemma. First we
use the definitions of intrinsic Kolmogorov distance k and of kd(F) to estimate
k(X , Y) ≤ kd(F)
= ∑ kd(Fi )
i
= ∑ [2 Ent(XYi ) − Ent(Xi ) − Ent(Yi )]
i
Next, we apply the definition of the conditional entropy to rewrite the righthand side
k(X , Y) ≤ ∑ [2 Ent(XYi ∣ U V ) + 2 Ent(U V ) − 2 Ent(U V ∣ XYi )
i
− Ent(Xi ∣ U ) − Ent(U ) + Ent(U ∣ Xi )
− Ent(Yi ∣ V ) − Ent(V ) + Ent(V ∣ Yi )]
We now use (T.6) and rearrange terms to obtain
k(X , Y) ≤ ∑ [2 Ent(XYi ∣ U V ) − Ent(Xi ∣ U V ) − Ent(Yi ∣ U V )
i
+ 2 Ent(U V ) − Ent(U ) − Ent(V )
− 2 Ent(U V ∣ XYi ) + Ent(U ∣ Xi ) + Ent(V ∣ Yi )]
67
T. Technical proofs
By the integral formula for conditional entropy (1.4) applied to the first three
terms we get
∑ [2 Ent(XYi ∣ U V ) − Ent(Xi ∣ U V ) − Ent(Yi ∣ U V )]
i
=∫
UV
k(X ∣ uv, Y∣ uv) d pW (u, v)
However, because of (T.5) this simplifies to
∫U V k(X ∣ uv, Y∣ uv) d pW (u, v) = ∫U V k(X ∣ u, Y∣ v) d pW (u, v)
Therefore,
k(X , Y) ≤ ∫
UV
k(X ∣ u, Y∣ v) d pW (u, v) + [[G]] ⋅ kd(U ← W → V )
+ ∑ [ Ent(U ∣ Xi ) + Ent(V ∣ Yi )]
i
⊠
Proposition 4.11. ↑ Let S = {Si , fij } ∈ Set ⟨G⟩ be a configuration of sets
modeled on a complete diagram category G with the initial set S0 . Let p, q ∈
∆S be two probability distributions. Denote X ∶= (S, p), Y ∶= (S, q) and α =
1
2 ∣p0 − q0 ∣1 . Then
k(X , Y) ≤ 2 ⋅ [[G]] ⋅ (α ⋅ ln ∣S0 ∣ + Ent(Λα ))
⊠
Proof: We will need the following obvious rough estimate of the Kolmogorov
distance that holds for any p, q ∈ ∆S:
(T.7)
2 ⋅ k(X , Y) ≤ 2[[G]] ⋅ ln ∣S0 ∣
It can be obtained by taking a tensor product for the coupling between X and
Y.
Our goal now is to write p and q as the convex combination of three other
distributions p̂, p+ and q + as in
p = (1 − α) ⋅ p̂ + α ⋅ p+
q = (1 − α) ⋅ p̂ + α ⋅ q +
with the smallest possible α ∈ [0, 1].
We could do it the following way. Let α ∶= 21 ∣p0 − q0 ∣. If α = 1 then the
proposition follows from the rough estimate (T.7), so from now on we assume
that α < 1. Define three probability distributions p̂0 , p+0 and q0+ on S0 by setting
68
T. Technical proofs
for every x ∈ S0
1
min {p0 (x), q0 (x)}
1−α
1
p+0 ∶= (p0 − (1 − α)p̂0 )
α
1
q0+ ∶= (q0 − (1 − α)p̂0 )
α
Denote by p̂, p+ , q + ∈ ∆S the distributions corresponding to p̂0 , p+0 , q0+ ∈ ∆S0
under isomorphism (3.5). Thus we have
p̂0 (x) ∶=
p = (1 − α)p̂ + α ⋅ p+
q = (1 − α)p̂ + α ⋅ q +
Now we construct a “two-tents” configuration of G-configurations
(T.8)
X ← X˜ → Λα ← Ỹ → Y
by setting
X̃i ∶= (Si × Λα ; p̃i (s, ◻) = (1 − α)p̂i (s), p̃i (s, ∎) = α ⋅ p+i (s))
Ỹi ∶= (Si × Λα ; q̃i (s, ◻) = (1 − α)p̂i (s), q̃i (s, ∎) = α ⋅ qi+ (s))
and
X˜ ∶= {X̃i ; fij × Id}
Ỹ ∶= {Ỹi ; fij × Id}
The reductions in the “two-tents” sub-configurations of (T.8) are given by
coordinate projections. Note that the following isomorphisms hold
X ∣ ◻ ≅ (S, p̂)
X ∣ ∎ ≅ (S, p+ )
Y∣ ◻ ≅ (S, p̂) ≅ X ∣ ◻
Y∣ ∎ ≅ (S, q + )
Now we apply part (i) of Corollary 4.10 to obtain the desired inequality
k(X , Y) ≤ (1 − α) k(X ∣ ◻, Y∣ ◻) + α ⋅ k(X ∣ ∎, Y∣ ∎)
+ ∑ [ Ent(Λα ∣ Xi ) + Ent(Λα ∣ Yi )]
i
≤ 2 ⋅ [[G]] ⋅ (α ⋅ ln ∣S0 ∣ + Ent(Λα ))
⊠
69
T. Technical proofs
T.7. Statements from the section “Extensions”.
Proposition 7.1. ↑ (Extension Lemma) Let k, l ∈ N, k ≤ l and let X , X ′ ∈
Prob ⟨Λk ⟩ be minimal full configurations. For every Y ∈ Extl X there exists a
Y ′ ∈ Extl X ′ such that
k(Y ′ , Y) ≤ 2l−k k(X ′ , X )
⊠
Proof: Denote by X0 , X0′ the initial spaces in the configurations X , X ′ , respectively. Let Y0 be the initial space of the full sub-configuration of Y generated
by Yk+1 , . . . , Yl . Let K0 be the initial space in the optimal coupling between X
and X ′
F = (X ′ ← K → X )
Recall that X0 could be considered as the Cartesian product of the underlying
sets of spaces generating X with some distribution on it. A similar view holds
for X0′ , Y0 and K0 . Thus we have in particular K 0 = X ′0 × X 0
Define a full minimal configuration Z ∈ Prob ⟨Λ2k+l ⟩ by providing a distribution on
X ′0 × X 0 × Y 0 = K 0 × Y 0
as explained in the Section 2.1.4. The distribution will be defined by
p(x′ , x, y) ∶= pF (x′ , x) ⋅ pY (x, y)/pX ′ (x)
It is clear that Z contains both the coupling F and configuration Y as
restrictions. It also contains the minimal full configuration
Y ′ = ⟨X1′ , . . . , Xk′ , Yk+1 , . . . , Yl ⟩
and a coupling G between Y and Y ′ .
For a pair of spaces A and B in Z we denote by AB the initial space of a
minimal fan in Z with the terminal spaces A and B. The two-fan of Λk+l -configurations G can be considered as a Λk+l -configuration of two-fans
GIJ ∶= (XI YJ ←Ð GIJ Ð→ XI′ YJ )
Using this notation we estimate for I ⊂ {1, . . . , k} and J ⊂ {k + 1, . . . , l}
k(Y, Y ′ ) ≤ kd G = ∑ kd(GIJ )
I,J
= ∑ [2 Ent(GIJ ) − Ent(XI YJ ) − Ent(XI′ YJ )]
I,J
≤ ∑ [2 Ent(XI XI′ ) − Ent(XI ) − Ent(XI′ )+
I,J
+ (2 Ent(YJ ∣ XI XI′ ) − Ent(YJ ∣ XI ) − Ent(YJ ∣ XI′ ))]
≤ 2k−l ∑ kd(FI )
I
≤2
k−l
kd(F) = 2k−l k(X , X ′ )
⊠
70
T. Technical proofs
Theorem 7.3. ↑ Let k ∈ N and X , X ′ ∈ Prob ⟨Λk ⟩. Then for all l ∈ N, the
Hausdorff distance between their stabilized relative entropic sets satisfies the
Lipschitz estimate
dH (Γl (X ), Γl (X ′ )) ≤ 2l−k κ(X , X ′ )
In other words, the map Γl from minimal full configurations in Prob ⟨Λk ⟩ to
Kk,l is 2l−k -Lipschitz.
⊠
Proof: Note that by Corollary 7.2, or more directly by Proposition 7.1, for
n∈N
1
1
dH (Γ○l ((X )⊗n ), Γ○l ((X ′ )⊗n )) ≤ 2l−k k ((X )⊗n , (X ′ )⊗n )
n
n
Hence, by the scaling properties of the Hausdorff distance
1
1
1
dH ( Γ○l ((X )⊗n ), Γ○l ((X ′ )⊗n )) ≤ 2l−k k ((X )⊗n , (X ′ )⊗n )
n
n
n
For convenience, we introduce the notation
1
Kn = Closure ( Γ○l ((X )⊗n ))
n
1
Kn′ = Closure ( Γ○l ((X ′ )⊗n ))
n
K = Γ○l (X )
K ′ = Γl (X ′ )
Recall that by definition,
K = Closure ( ⋃ Kn )
n∈N
K ′ = Closure ( ⋃ Kn′ )
n∈N
Note that by the superadditivity property of the unstabilized relative en′
tropic sets (see inclusion (7.1)) the sequences n ↦ Kn! and n ↦ Kn!
are monotonically increasing sequences of sets, and
n
⋃ Ki ⊂ Kn!
i=1
n
′
⋃ Ki′ ⊂ Kn!
i=1
Now select a large radius R > 0. Let BR (0) denote the ball of radius R
{1,...,l}
around the origin in R2
. By compactness and the definition of the stabilized relative entropic set
dH (Kn! ∩ BR (0), K ∩ BR (0)) → 0
′
dH (Kn!
∩ BR (0), K ′ ∩ BR (0)) → 0
as n → ∞. Therefore also
dH (K ∩ BR (0), K ′ ∩ BR (0)) ≤ 2l−k κ (X , X ′ )
Because this inequality holds for every R > 0, the estimate in the lemma
follows.
⊠
71
T. Technical proofs
T.8. Statements from the section “Mixtures”.
Lemma 8.1. ↑ Let G be a complete diagram category and X , Y ∈ Prob ⟨G⟩.
Then
(i) κ(X , X ⊗n ⊕Λ1/n {●}) ≤ Ent(Λ1/n )
(ii) κ (X , (X ⊕Λ1/n {●})⊗n ) ≤ n ⋅ Ent(Λ1/n )
(iii) κ ((X ⊗ Y) ⊕Λ1/n {●} , (X ⊕Λ1/n {●}) ⊗ (Y ⊕Λ1/n {●})) ≤ 3 Ent(Λ1/n )
1
(iv) κ ((X ⊕Λ1/n {●}), (Y ⊕Λ1/n {●})) ≤ κ(X , Y)
n
⊠
Proof: Recall that for the empirical reduction
q ∶ Λ⊗N
1/n → ∆Λ1/n
the quantity N ⋅ q(λ)(∎) counts the number of black squares in the sequence
λ. It is a binomially distributed random variable with the mean N /n and
variance Nn (1 − n1 ).
The first claim is then proven by the following calculation
κ(X , X ⊗n ⊕Λ1/n {●})
1
= lim
k (X ⊗N , (X ⊗n ⊕Λ1/n {●})⊗N )
N →∞ N
⎞
1 ⎛ ⊗N
k ⎜X , ⊕ X ⊗n⋅N ⋅q(λ)(∎) ⎟
= lim
N →∞ N
⎠
⎝
λ∈Λ⊗N
1/n
1
k(X ⊗N , X ⊗(N ⋅n⋅q(λ)(∎)) ) d p(λ)
∫
⊗n
N →∞ N λ∈Λ
1/n
n
⋅ ∫ ⊗N ∣N /n − N ⋅ q(λ)(∎)∣ d p(λ)
≤ Ent(Λ1/n ) + ∣ Ent∗ (X )∣1 ⋅ lim
N →∞ N
λ∈Λ1/n
√
n
1
1
≤ Ent(Λ1/n ) + ∣ Ent∗ (X )∣1 ⋅ lim
⋅ N ⋅ (1 − )
N →∞ N
n
n
= Ent(Λ1/n )
≤ Ent(Λ1/n ) + lim
The second claim is proven similarly and the third follows from the second
and the 1-Lipschitz property of the tensor product. Finally, the fourth follows
from Corollary 4.10(iv), by slicing both arguments along Λ1/n .
⊠
T.9. Statements from the section “Tropical Probability”.
Lemma T.3. Suppose the sequence {a(i)}i∈N0 of real numbers is bounded from
below and is quasi-subadditive, that is there is a constant C ∈ R such that for
any i, j ∈ N0 holds
a(i + j) ≤ a(i) + a(j) + C
Then the limit
1
lim a(i)
i→∞ i
72
T. Technical proofs
⊠
exists and is finite.
Proof: The lemma is standard and is sometimes refered to as Fekete’s subadditive lemma. We include a proof for the convenience of the reader. Assume
first that C = 0. Then the sequence satisfies a(k ⋅ i) ≤ k ⋅ a(i) and in particular
a(i) ≤ i ⋅ a(1). Let l ∶= lim inf 1i a(i) ∈ [0, ∞). Choose ε > 0. Then we can
find k ∈ N such that k1 a(k) ≤ l + ε. For n ∈ N let q, r be the quotient and the
reminder of the integer division of n by k, that is
n = q ⋅ k + r,
0≤r<k
Then
1
1
1
1
a(n) ≤ (q ⋅ a(k) + a(r)) ≤
(q ⋅ a(k)) + a(r) ≤ l + ε + ε = l + 2ε
n
n
q⋅k+r
n
The last inequality holds once n is sufficiently large, specifically when
1
n ≥ max a(i)
ε 0≤i≤k
Therefore
1
lim a(i) = l
i→∞ i
Now if C > 0 then the sequence b(i) ∶= a(i) + C is subadditive and 1i b(i)
converges by the previous argument. Thus we have
1
1
1
lim b(i) = lim (a(i) + C) = lim a(i)
i→∞ i
i→∞ i
i→∞ i
⊠
Lemma 9.1. ↑ For a pair γ 1 , γ 2 ∈ QLδ (Γ) the limit
1
lim δ (γ1 (n), γ2 (n))
n→∞ n
exists and is finite.
⊠
Proof: Suppose γ 1 and γ 2 are two quasi-linear sequences of elements of Γ, then
for any i, j ∈ N0
δ (γ1 (i + j),γ2 (i + j))
≤ δ (γ1 (i + j), γ1 (i) ⊗ γ1 (j)) + δ (γ2 (i + j), γ2 (i) ⊗ γ2 (j))
+ δ (γ1 (i) ⊗ γ1 (j), γ2 (i) ⊗ γ2 (j))
≤ Defectδ (γ 1 ) + Defectδ (γ 2 ) + δ (γ1 (i), γ2 (i)) + δ (γ1 (j), γ2 (j))
Thus the sequence δ(γ1 (i), γ2 (i)) is quasi-subadditive and by Lemma T.3
the limit
1
lim δ (γ1 (i), γ2 (i))
i→∞ i
exists and is finite.
⊠
Proposition 9.2. ↑ Suppose (Γ, ⊗, δ) is a metric Abelian monoid such that
73
T. Technical proofs
(i) the distance function δ is homogeneous, that is for any γ1 , γ2 ∈ Γ and
n ∈ N0
δ(γ1⊗n , γ2⊗n ) = n ⋅ δ(γ1 , γ2 )
(ii) (Γ, ⊗, δ) has the uniformly bounded defect property.
Then the space (QLδ (Γ), δ̂) is complete.
⊠
Proof: Given a Cauchy sequence {γ i } of elements in (QLδ (Γ), δ̂) we need
to find a limiting element ϕ ∈ QLδ (Γ). We will do that by a version of the
diagonal process, that is we define ϕ(n) to have value γi (n) for i sufficiently
large depending on n. The quasi-linearity of ϕ would follow from the fact that
for a fixed n and all sufficiently large i the set {γi (n)} is uniformly bounded.
Now we give the detailed argument. First we replace each element of the
sequence {γ i } by an asymptotically equivalent element with defect bounded
by the constant C according to assumption (ii) of the lemma. We will still call
the new sequence {γ i }. The Cauchy sequence {γ i } satisfies
sup δ̂(γ i , γ j ) → 0 as i → ∞
i,j≥i
By assumption (ii) of the lemma for any n, k ∈ N0 holds
k ⋅ δ (γi (n), γj (n)) = δ (γi (n)⊗k , γj (n)⊗k )
≤ δ (γi (kn), γj (kn)) + 2k ⋅ C
Dividing by k we obtain
δ(γi (n), γj (n)) ≤
1
δ(γi (kn), γj (kn)) + 2C
k
Now we pass to the limit sending k to infinity, while keeping n fixed:
δ(γi (n), γj (n)) ≤ n ⋅ δ̂(γ i , γ j ) + 2C
Given n let i(n) be a number such that for any i, j ≥ i(n) holds
δ̂(γ i , γ j ) ≤
1
n
We may assume that i(n) is nondecreasing as a function of n. Then for any
i, j, n ∈ N with i, j ≥ i(n) we have the following bound
(T.9)
δ (γi (n), γj (n)) ≤ 2C + 1
Now we are ready to define the limiting sequence ϕ by setting
ϕ(n) ∶= γi(n) (n)
74
T. Technical proofs
First we verify that ϕ is quasi-linear
δ(ϕ(n + m), ϕ(n) ⊗ ϕ(m)) = δ (γi(n+m) (n + m), γi(n) (n) ⊗ γi(m) (m))
≤ δ (γi(n+m) (n + m), γi(n+m) (n) ⊗ γi(n+m) (m))
+ δ (γi(n+m) (n) ⊗ γi(n+m) (m), γi(n) (n) ⊗ γi(m) (m))
≤ C + δ (γi(n+m) (n), γi(n) (n)) + δ (γi(n+m) (m), γi(m) (m))
≤ C + 2(2C + 1) = 5C + 2 =∶ C ′
The convergence of γ i to ϕ is shown as follows. For n, k ∈ N let q, r ∈ N0 be
the quotient and the remainder of the division of n by k, that is n = q ⋅ k + r
and 0 ≤ r < k. Fix k ∈ N and let i ≥ i(k), then
1
δ (γi (n), ϕ(n))
n→∞ n
1
= lim δ (γi (q ⋅ k + r), γi(n) (q ⋅ k + r))
n→∞ n
1
≤ lim (q ⋅ δ (γi (k), γi(n) (k)) + δ (γi (r), γi(n) (r)) + 2qC ′ + 2C ′ )
n→∞ n
1
≤ lim ((3q + 3) ⋅ C ′ )
n→∞ n
3C ′
=
k
δ̂(γ i , ϕ) = lim
Since k ∈ N is arbitrary we have
lim δ̂(γ i , ϕ) = 0
i→∞
⊠
Proposition 9.3. ↑ Suppose (Γ, ⊗, δ) has the ε-uniformly bounded defect property for every ε > 0. Then Lδ (Γ) is dense in QLδ (Γ)
⊠
Proof: Let γ = {γ(n)} be a δ-quasi-linear sequence. We need to approximate it
with linear sequences. For i ∈ N, let γ i be a sequence asymptotically equivalent
to γ and satisfying
Defectδ γ i ≤ 1/i
as provided by the 1/i-uniformly bounded defect property.
Define a δ-linear sequence η i by
ηi (n) ∶= γi (1)⊗n
T. Technical proofs
75
Then
δ̂(γ, η i ) = δ̂(γ i , η i )
1
= lim δ(γi (n), ηi (n))
n→∞ n
1
= lim δ(γi (n), γi (1)⊗n )
n→∞ n
1
≤ lim ⋅ n ⋅ Defectδ (γ i )
n→∞ n
1
≤
i
Thus lim η i = γ.
Lemma 9.4. ↑ The natural inclusion
⊠
ˆ
ı2 ∶ (QLδ (Γ), δ̂) ↪ (QLδ̂ (Γ), δ̂)
is an isometric embedding.
⊠
Proof: Let γ 1 , γ 2 ∈ QLδ (Γ) be two sequences of δ-quasi-linear sequences. We
have to show that the two numbers
1
δ̂(γ 1 , γ 2 ) = lim δ (γ1 (n), γ2 (n))
n→∞ n
and
1
ˆ
δ̂(γ 1 , γ 2 ) = lim δ̂ (γ1 (n), γ2 (n))
n→∞ n
are equal. Since shifts are non-expanding maps, we have δ̂ ≤ δ and it follows
immediately that
ˆ
δ̂(γ 1 , γ 2 ) ≤ δ̂(γ 1 , γ 2 )
and we are left to show the opposite inequality. We will do it as follows. Fix
n > 0, then
1
δ̂(γ 1 , γ 2 ) = lim
δ (γ1 (kn), γ2 (kn))
k→∞ kn
1
≤ lim
( δ (γ1 (n)⊗k , γ2 (n)⊗k ) + k ⋅ ( Defectδ (γ 1 ) + Defectδ (γ 2 )))
k→∞ kn
1
1
≤ δ̂ (γ1 (n), γ2 (n)) + ( Defectδ (γ 1 ) + Defectδ (γ 2 ))
n
n
Passing to the limit with respect to n gives required inequality
ˆ
δ̂(γ 1 , γ 2 ) ≤ δ̂(γ 1 , γ 2 )
Lemma 9.5. ↑ The image of the isometric embedding
ˆ
ı2 ∶ (QLδ (Γ), δ̂) ↪ (QLδ̂ (Γ), δ̂)
ˆ
is dense in (QLδ̂ (Γ), δ̂)
⊠
⊠
76
T. Technical proofs
ˆ
Proof: Given an element γ = {γ(n)} in QLδ̂ (Γ) we have to find a δ̂-approximating sequence γ i = {γi (n)} in QLδ (Γ). Define
γi (n) ∶= γ(i)⊗⌊ i ⌋
n
i→∞
ˆ
We have to show that each γ i is δ-quasi-linear and that δ̂(γ i , γ) Ð→ 0. These
follow from
δ̂ (γi (m + n), γi (m) ⊗ γi (n)) = δ̂ (γ(i)⊗⌊
m+n
⌋
i
, γ(i)⊗⌊ i ⌋ ⊗ γ(i)⊗⌊ i ⌋ )
m
n
≤ δ̂ (γ(i), 1)
and
1
ˆ
δ̂(γ i , γ) = lim δ̂ (γi (n), γ(n))
n→∞ n
1
n
= lim δ̂ (γ(i)⊗⌊ i ⌋ , γ(n))
n→∞ n
1
1
≤ lim [ δ̂ (γ (i⌊ ni ⌋) , γ(n)) + ⌊ ni ⌋ Defectδ̂ (γ)]
n→∞ n
n
i
1
1
≤ lim [
max δ̂ (1, γ(k)) + Defectδ̂ (γ)] + Defectδ̂ (γ)
n→∞ n k=0,...,i−1
n
i
1
≤ Defectδ̂ (γ)
i
It is worth noting, that the defect of γ i need not to be uniformly bounded with
respect to i.
⊠
Lemma 9.6. ↑ For a complete diagram category, and for every ε > 0, the
space (Prob ⟨G⟩ , ⊗, κ) has the ε-uniformly bounded defect property, that is for
any κ-quasi-linear sequence X ∈ QLκ (Prob ⟨G⟩) there exists an asymptotically
⊠
equivalent sequence Y with defect not exceeding ε.
Proof: Let X = {X (i)} be a quasi-linear sequence and let ε > 0. We will find
an asymptotically equivalent sequence with defect less than ε.
Define a new sequence Y = {Y(i)} by
Y(i) ∶= [X (k ⋅ i)] ⊕Λ1/k {●}
where the number k ∈ N will be chosen later. First we verify that the sequences
X and Y are asymptotically equivalent, that is
1
κ̂(X , Y) ∶= lim κ (X (i), Y(i)) = 0
i→∞ i
We estimate the asymptotic distance between individual members of sequences
X and Y using Lemma 8.1 as follows
κ(X (i), Y(i)) = κ (X (i), X (k ⋅ i) ⊕Λ1/k {●} )
≤ κ (X (i), X (i)⊗k ⊕Λ1/k {●}) + κ (X (i)⊗k ⊕Λ1/k {●} , X (k ⋅ i) ⊕Λ1/k {●})
≤ Ent(Λ1/k ) + Defectκ (X )
77
T. Technical proofs
Thus κ̂(X , Y) = 0 and the two sequences are asymptotically equivalent.
Next we show that the sequence Y is κ-quasi-linear and evaluate its defect
using Lemma 8.1. Let i, j ∈ N, then
κ(Y(i + j), Y(i) ⊗ Y(j))
= κ (X (k ⋅ i + k ⋅ j) ⊕Λ1/k {●} , [X (k ⋅ i) ⊕Λ1/k {●} ] ⊗ [X (k ⋅ j) ⊕Λ1/k {●} ])
≤ κ ([X (k ⋅ i)⊗X (k ⋅ j)] ⊕Λ1/k {●} , [X (k ⋅ i) ⊕Λ1/k {●} ] ⊗ [X (k ⋅ j) ⊕Λ1/k {●} ])
1
Defect(X )
k
1
≤ 3 Ent(Λ1/k ) + Defectκ (X )
k
Thus, by choosing k to be a solution to the inequality
1
3 Ent(Λ1/k ) + Defectκ (X ) ≤ ε
k
we can make sure that
Defectκ (Y) ≤ ε
+
⊠
Lemma 9.12. ↑ Denote by Un a finite uniform probability space of cardinality
n, then
(i)
n
k(Un , Um ) ≤ 2 ln 2 + ∣ln ∣
m
(ii)
κ(Un , Um ) = ∣ Ent(Un ) − Ent(Um )∣
⊠
f
g
Proof: Consider a two-fan Un ← Unm → Um . To construct specific reductions f
and g we identify Unm , Un and Um with the cyclic groups of the corresponding
order
Unm ↔ Znm
Un ↔ Zn
Um ↔ Zm
Consider the short exact sequences
×m
mod m
×n
mod n
{0} Ð→ Zn Ð→ Znm Ð→ Zm Ð→ {0}
{0} Ð→ Zm Ð→ Znm Ð→ Zn Ð→ {0}
Choose for f the left splitting in the first exact sequence, and for g the left
splitting in the second exact sequence.
78
10. Technical proofs
f
g
Now that we constructed a two-fan Un ← Unm → Um , let Un ← Z → Um be
its minimal reduction. Now we estimate ∣Z∣ ≤ n + m, which implies that
k(Un , Um ) ≤ 2 Ent(Z) − Ent(Un ) − Ent(Um )
≤ 2 ln(n + m) − ln n − ln m
≤ 2 ln 2 + 2 ln max {n, m} − ln n − ln m
n
= 2 ln 2 + ∣ln ∣
m
To prove the second assertion note that entropy is a k-1-Lipschitz function.
Therefore we have
∣ Ent(Un ) − Ent(Um )∣ ≤ k(Un , Um ) ≤ ∣ Ent(Un ) − Ent(Um )∣ + 2 ln 2
Substituting in the definition of asymptotic Kolmogorov distance we obtain
the required equality.
⊠
References
[ABD+ 08] Nihat Ay, Nils Bertschinger, Ralf Der, Frank Güttler, and Eckehard Olbrich.
Predictive information and explorative behavior of autonomous robots. The European Physical Journal B, 63(3):329–339, 2008.
[ABK+ 15] Samson Abramsky, Rui Soares Barbosa, Kohei Kishida, Raymond Lal, and
Shane Mansfield. Contextuality, cohomology and paradox. arXiv preprint
arXiv:1502.03097, 2015.
[BBI01] Dmitri Burago, Yuri Burago, and Sergei Ivanov. A course in metric geometry,
volume 33 of Graduate Studies in Mathematics. American Mathematical Society,
Providence, RI, 2001.
[BFL11] John C. Baez, Tobias Fritz, and Tom Leinster. A characterization of entropy in
terms of information loss. Entropy, 13(11):1945–1957, 2011.
[BRO+ 14] Nils Bertschinger, Johannes Rauh, Eckehard Olbrich, Jürgen Jost, and Nihat Ay.
Quantifying unique information. Entropy, 16(4):2161–2183, 2014.
[Csi98]
Imre Csiszár. The method of types. IEEE Trans. Inform. Theory, 44(6):2505–
2523, 1998. Information theory: 1948–1998.
[CT91]
Thomas M. Cover and Joy A. Thomas. Elements of information theory. Wiley
Series in Telecommunications. John Wiley & Sons, Inc., New York, 1991. A
Wiley-Interscience Publication.
[DFZ11] Randall Dougherty, Chris Freiling, and Kenneth Zeger. Non-shannon information
inequalities in four random variables. arXiv preprint arXiv:1104.3602, 2011.
[Fri09]
Karl Friston. The free-energy principle: a rough guide to the brain? Trends in
cognitive sciences, 13(7):293–301, 2009.
[Gro12]
Misha Gromov. In a search for a structure, part 1: On entropy. Preprint available
at http://www.ihes.fr/gromov, 2012.
[Mat07] Frantisek Matus. Infinitely many information inequalities. In Information Theory,
2007. ISIT 2007. IEEE International Symposium on, pages 41–44. IEEE, 2007.
[SA15]
Bastian Steudel and Nihat Ay. Information-theoretic inference of common ancestors. Entropy, 17(4):2304–2327, 2015.
[Sin76]
Ya. G. Sinai. Introduction to ergodic theory. Princeton University Press, Princeton, N.J., 1976. Translated by V. Scheffer, Mathematical Notes, 18.
10. Technical proofs
[VDP13]
[Yeu12]
79
Sander G Van Dijk and Daniel Polani. Informational constraints-driven organization in goal-directed behavior. Advances in Complex Systems, 16(02n03):1350016,
2013.
Raymond W Yeung. A first course in information theory. Springer Science &
Business Media, 2012.
| 7cs.IT
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:1605.01352v1 [cs.PL] 4 May 2016
Default Rules for Curry∗
SERGIO ANTOY
Computer Science Dept., Portland State University, Oregon, U.S.A.
(e-mail: [email protected])
MICHAEL HANUS
Institut für Informatik, CAU Kiel, D-24098 Kiel, Germany.
(e-mail: [email protected])
submitted January 31, 2016; revised April 13, 2016; accepted May 2, 2016
Abstract
In functional logic programs, rules are applicable independently of textual order, i.e., any
rule can potentially be used to evaluate an expression. This is similar to logic languages
and contrary to functional languages, e.g., Haskell enforces a strict sequential interpretation of rules. However, in some situations it is convenient to express alternatives by means
of compact default rules. Although default rules are often used in functional programs, the
non-deterministic nature of functional logic programs does not allow to directly transfer
this concept from functional to functional logic languages in a meaningful way. In this paper we propose a new concept of default rules for Curry that supports a programming style
similar to functional programming while preserving the core properties of functional logic
programming, i.e., completeness, non-determinism, and logic-oriented use of functions. We
discuss the basic concept and propose an implementation which exploits advanced features
of functional logic languages.
To appear in Theory and Practice of Logic Programming (TPLP)
KEYWORDS: functional logic programming, semantics, program transformation
1 Motivation
Functional logic languages combine the most important features of functional and
logic programming in a single language (see (Antoy and Hanus 2010; Hanus 2013)
for recent surveys). In particular, the functional logic language Curry (Hanus (ed.) 2016)
conceptually extends Haskell with common features of logic programming, i.e., nondeterminism, free variables, and constraint solving. Moreover, the amalgamated features of Curry support new programming techniques, like deep pattern matching
through the use of functional patterns, i.e., evaluable functions at pattern positions
(Antoy and Hanus 2005).
∗ This is an extended version of a paper presented at the international symposium on Practical
Aspects of Declarative Languages (PADL 2016), invited as a rapid communication in TPLP.
The authors acknowledge the assistance of the conference program chairs Marco Gavanelli and
John Reppy.
2
Sergio Antoy and Michael Hanus
For example, suppose that we want to compute two elements x and y in a list l
with the property that the distance between the two elements is n, i.e., in l there
are n − 1 elements between x and y. We will use this condition in the n-queens
program discussed later. Of course, there may be many pairs of elements in a list
satisfying the given condition (“++” denotes the concatenation of lists):
dist n (_++[x]++zs++[y]++_) | n == length zs + 1 = (x,y)
Defining functions by case distinction through pattern matching is a very useful feature. Functional patterns make this feature even more convenient. However,
in functional logic languages, this feature is slightly more delicate because of the
possibility of functional patterns, which typically stand for an infinite number of
standard patterns, and because there is no textual order among the rules defining
an operation. The variables in a functional pattern are bound like the variables in
ordinary patterns.
As a simple example, consider an operation isSet intended to check whether a
given list represents a set, i.e., does not contain duplicates. In Curry, we might
think to implement it as follows:
isSet (_++[x]++_++[x]++_) = False
isSet _
= True
The first rule uses a functional pattern: it returns False if the argument matches a
list where two identical elements occur. The intent of the second rule is to return
True if no identical elements occur in the argument. However, according to the
semantics of Curry, which ensures completeness w.r.t. finding solutions or values,
all rules are tried to evaluate an expression. Therefore, the second rule is always
applicable to calls of isSet so that the expression isSet [1,1] will be evaluated to
False and True.
The unintended application of the second rule can be avoided by the additional
requirement that this rule should be applied only if no other rule is applicable.
We call such a rule a default rule and mark it by adding the suffix ’default to
the function’s name (in order to avoid a syntactic extension of the base language).
Thus, if we define isSet with the rules
isSet (_++[x]++_++[x]++_) = False
isSet ’default _
= True
then isSet [1,1] evaluates only to False and isSet [0,1] only to True.
In this paper we propose a concept for default rules for Curry, define its precise
semantics, and discuss implementation options. In the next section, we review the
main concepts of functional logic programming and Curry. Our intended concept
of default rules is informally introduced in Sect. 3. Some examples showing the
convenience of default rules for programming are presented in Sect. 4. In order to
avoid the introduction of a new semantics specific to default rules, we define the
precise meaning of default rules by transforming them into already known concepts
in Sect. 5. Options to implement default rules efficiently are discussed and evaluated
Default Rules for Curry
3
in Sect. 6. Some benchmarking of alternative implementations of default rules are
shown in Sect. 7 before we relate our proposal to other work and conclude.
2 Functional Logic Programming and Curry
Before presenting the concept and implementation of default rules in more detail,
we briefly review those elements of functional logic languages and Curry that are
necessary to understand the contents of this paper. More details can be found in recent surveys on functional logic programming (Antoy and Hanus 2010; Hanus 2013)
and in the language report (Hanus (ed.) 2016).
Curry is a declarative multi-paradigm language combining in a seamless way features from functional, logic, and concurrent programming (concurrency is irrelevant
as our work goes, hence it is ignored in this paper). The syntax of Curry is close to
Haskell (Peyton Jones 2003), i.e., type variables and names of defined operations
usually start with lowercase letters and the names of type and data constructors
start with an uppercase letter. α → β denotes the type of all functions mapping
elements of type α into elements of type β (where β can also be a functional type,
i.e., functional types are “curried”), and the application of an operation f to an argument e is denoted by juxtaposition (“f e”). In addition to Haskell, Curry allows
free (logic) variables in conditions and right-hand sides of rules and expressions
evaluated by an interpreter. Moreover, the patterns of a defining rule can be nonlinear, i.e., they might contain multiple occurrences of some variable, which is an
abbreviation for equalities between these occurrences.
Example 1
The following simple program shows the functional and logic features of Curry. It
defines an operation “++” to concatenate two lists, which is identical to the Haskell
encoding. The second operation, dup, returns some list element having at least two
occurrences:1
(++) :: [a] → [a] → [a]
[]
++ ys = ys
(x:xs) ++ ys = x : (xs ++ ys)
dup :: [a] → a
dup xs | xs == _ ++ [x] ++ _ ++ [x] ++ _
= x
where x free
Operation applications can contain free variables. They are evaluated lazily where
free variables as demanded arguments are non-deterministically instantiated. Hence,
the condition of the rule defining dup is solved by instantiating x and the anonymous
free variables “-”. This evaluation method corresponds to narrowing (Slagle 1974;
Reddy 1985), but Curry narrows with possibly non-most-general unifiers to ensure
the optimality of computations (Antoy et al. 2000).
1
Note that Curry requires the explicit declaration of free variables, as x in the rule of dup, to
ensure checkable redundancy.
4
Sergio Antoy and Michael Hanus
Note that dup is a non-deterministic operation since it might deliver more than
one result for a given argument, e.g., the evaluation of dup [1,2,2,1] yields the
values 1 and 2. Non-deterministic operations, which are interpreted as mappings
from values into sets of values (González-Moreno et al. 1999), are an important
feature of contemporary functional logic languages. Hence, there is also a predefined
choice operation:
x ? _
_ ? y
=
=
x
y
Thus, the expression “0?1” evaluates to 0 and 1 with the value non-deterministically
chosen.
Some operations can be defined more easily and directly using functional patterns (Antoy and Hanus 2005). A functional pattern is a pattern occurring in an
argument of the left-hand side of a rule containing defined operations (and not only
data constructors and variables). Such a pattern abbreviates the set of all standard patterns to which the functional pattern can be evaluated (by narrowing). For
instance, we can rewrite the definition of dup as
dup (_++[x]++_++[x]++_) = x
Functional patterns are a powerful feature to express arbitrary selections in tree
structures, e.g., in XML documents (Hanus 2011). Details about their semantics
and a constructive implementation of functional patterns by a demand-driven unification procedure can be found in (Antoy and Hanus 2005).
Set functions (Antoy and Hanus 2009) allow the encapsulation of non-deterministic computations in a strategy-independent manner. For each defined operation
f , fS denotes the corresponding set function. fS encapsulates the non-determinism
caused by evaluating f except for the non-determinism caused by evaluating the
arguments to which f is applied. For instance, consider the operation decOrInc
defined by
decOrInc x = (x-1) ? (x+1)
Then “decOrIncS 3” evaluates to (an abstract representation of) the set {2, 4},
i.e., the non-determinism caused by decOrInc is encapsulated into a set. However,
“decOrIncS (2 ? 5)” evaluates to two different sets {1, 3} and {4, 6} due to its nondeterministic argument, i.e., the non-determinism caused by the argument is not
encapsulated. This property is desirable and essential to define and implement
default rules by a transformational approach, as shown in Sect. 5. In the following
section, we discuss default rules and their intended semantics.
3 Default Rules: Concept and Informal Semantics
Default rules are often used in both functional and logic programming. In languages
in which rules are applied in textual order, such as Haskell and Prolog, loosely speaking every rule is a default rule of all the preceding rules. For instance, the following
standard Haskell function takes two lists and returns the list of corresponding pairs,
where excess elements of a longer list are discarded:
Default Rules for Curry
5
zip (x:xs) (y:ys) = (x,y) : zip xs ys
zip _
_
= []
The second rule is applied only if the first rule is not applicable, i.e., if one of
the argument lists is empty. We can avoid the consideration of rule orderings by
replacing the second rule with rules for the patterns not matching the first rule:
zip (x:xs) (y:ys) = (x,y) : zip xs ys
zip (_:_) []
= []
zip []
_
= []
In general, this coding is cumbersome since the number of additional rules increases
if the patterns of the first rule are more complex (e.g., we need three additional
rules for the operation zip3 combining three lists). Moreover, this coding might
be impossible in conjunction with some functional patterns, as in the first rule
of isSet above. Some functional patterns conceptually denote an infinite set of
standard patterns (e.g., [x,x], [x,-,x], [-,x,-,x], . . . ) and the complement of this
set is infinite too.
In Prolog, one often uses the “cut” operator to implement the behavior of default
rules. For instance, zip can be defined as a Prolog predicate as follows:
zip([X|Xs],[Y|Ys],[(X,Y)|Zs]) :- !, zip(Xs,Ys,Zs).
zip(_,_,[]).
Although this definition behaves as intended for instantiated lists, the completeness of logic programming is destroyed by the cut operator. For instance, the goal
zip([],[],[]) is provable, but Prolog does not compute the answer {Xs=[],Ys=[],
Zs=[]} for the goal zip(Xs,Ys,Zs) .
These examples show that neither the functional style nor the logic style of default
rules is suitable for functional logic programming. The functional style, based on
textual order, curtails non-determinism. The logic style, based on the cut operator,
destroys the completeness of some computations. Thus, a new concept of default
rules is required for functional logic programming if we want to keep the strong
properties of the base language, in particular, a simple-to-use non-determinism and
the completeness of logic-oriented evaluations. Before presenting the exact definition
of default rules, we introduce them informally and discuss their intended semantics.
We intend to extend a “standard” operation definition by one default rule. Hence,
an operation definition with a default rule has the following form (ok denotes a
sequence of objects o1 . . . ok ):2
f tk1 | c1 = e1
.
.
.
f tkn | cn = en
f ’default tkn+1 | cn+1 = en+1
2
We consider only conditional rules since an unconditional rule can be regarded as a conditional
rule with condition True.
6
Sergio Antoy and Michael Hanus
We call the first n rules standard rules and the final rule the default rule of f .
Informally, the default rule is applied only if no standard rule is applicable, where
a rule is applicable if the pattern matches and the condition is satisfied. Hence, an
expression e = f sk , where sk are expressions, is evaluated as follows:
1. The arguments sk are evaluated enough to determine whether a standard rule
of f is applicable, i.e., whether there exists a standard rule whose left-hand
side matches the evaluated e and the condition is satisfied (i.e., evaluable to
True).
2. If a standard rule is applicable, it is applied; otherwise the default rule is
applied.
3. If some argument is non-deterministic, the previous points apply independently for each non-deterministic choice of the combination of arguments. In
particular, if an argument is a free variable, it is non-deterministically instantiated so that every potentially applicable rule can be used.
As usual in a non-strict language like Curry, arguments of an operation application are evaluated as they are demanded by the operation’s pattern matching and
condition. However, any non-determinism or failure during argument evaluation is
not passed inside the condition evaluation. A precise definition of “inside” is in
(Antoy and Hanus 2009, Def. 3). This behavior is quite similar to set functions to
encapsulate internal non-determinism. Therefore, we will exploit set functions to
implement default rules.
Before discussing the advantages and implementation of default rules, we explain
and motivate the intended semantics of our proposal. First, it should be noted
that this concept distinguishes non-determinism outside and inside a rule application. This difference is irrelevant in purely functional programming but essential in
functional logic programming.
Example 2
Consider the operation zip defined with a default rule:
zip (x:xs) (y:ys) = (x,y) : zip xs ys
zip ’default _ _
= []
Since the standard rule is applicable to zip [1] [2], the default rule is ignored so
that this expression is solely reduced to (1,2):zip [] []. Since the standard rule
is not applicable to zip [] [], the default rule is applied and yields the value [].
Altogether, the only value of zip [1] [2] is [(1,2)]. However, if some argument has
more than one value, we use the evaluation principle above for each combination.
Thus, the call zip ([1] ? []) [2] yields the two values [(1,2)] and [].
These considerations are even more relevant if the evaluation of the condition might
be non-deterministic, as the following example shows.
Example 3
Consider an operation to look up values for keys in an association list:
lookup key assoc
| assoc == (_ ++ [(key ,val)] ++ _)
= Just val
Default Rules for Curry
7
where val free
lookup ’default _ _ = Nothing
Note that the condition of the standard rule can be evaluated in various ways. In
particular, it can be evaluated (non-deterministically) to True and False for a fixed
association list and key. Therefore, using if-then-else (or an otherwise branch as in
Haskell) instead of the default rule might lead to unintended results.
If we evaluate lookup 2 [(2,14),(3,17),(2,18)] , the condition of the standard
rule is satisfiable so that the default rule is ignored. Since the condition has the two
solutions {val 7→ 14} and {val 7→ 18}, we yield the values Just 14 and Just 18. If
we evaluate lookup 2 [(3,17)], the condition of the standard rule is not satisfiable
but the default rule is applicable so that we obtain the result Nothing.
On the other hand, non-deterministic arguments might trigger different rules to
be applied. Consider the expression lookup (2 ? 3) [(3,17)]. Since the non-determinism in the arguments leads to independent evaluations of the expressions lookup 2
[(3,17)] and lookup 3 [(3,17)], we obtain the results Nothing and Just 17.
Similarly, free variables as arguments might lead to independent results since
free variables are equivalent to non-deterministic values (Antoy and Hanus 2006).
For instance, the expression lookup 2 xs yields the value Just v with the binding
{xs 7→ (2,v): }, but also the value Nothing with the binding {xs 7→ []} (as well as
many other solutions).
The latter desirable property also has implications for the handling of failures
occurring when arguments are evaluated. For instance, consider the expression
lookup 2 failed (where failed is a predefined operation which always fails whenever it is evaluated). Because the evaluation of the condition of the standard rule
demands the evaluation of failed and the subsequent failure comes from “outside”
the condition, the entire expression evaluation fails instead of returning the value
Nothing. This is motivated by the fact that we need the value of the association
list in order to check the satisfiability of the condition and, thus, to decide the
applicability of the standard rule, but this value is not available.
Example 4
To see why our design decision is reasonable, consider the following contrived definition of an operation that checks whether its argument is the unit value () (which
is the only value of the unit type):
isUnit x | x == () = True
isUnit ’default _
= False
In our proposal, the evaluation of “isUnit failed” fails. In an alternative design (like
Prolog’s if-then-else construct), one might skip any failure during condition checking
and proceed with the next rule. In this case, we would return the value False for
the expression isUnit failed. This is quite disturbing since the (deterministic!)
operation isUnit, which has only one possible input value, could return two values:
True for the call isUnit () and False for the call isUnit failed. Moreover, if we
call this operation with a free variable, like isUnit x, we obtain the single binding
{x 7→ ()} and value True (since free variables are never bound to failures). Thus,
8
Sergio Antoy and Michael Hanus
either our semantics would be incomplete for logic computations or we compute
too many values. In order to get a consistent behavior, we require that failures of
arguments demanded for condition checking lead to failures of evaluations.
4 Examples
To show the applicability and convenience of default rules for functional logic programming, we sketch a few more examples in this section.
Example 5
Default rules are important in combination with functional patterns, since functional patterns denote an infinite set of standard patterns which often has no finite
complement. Consider again the operation lookup as introduced in Example 3. With
functional patterns and default rules, this operation can be conveniently defined:
lookup key (_ ++ [(key ,val)] ++ _) = Just val
lookup ’default _ _
= Nothing
Example 6
Functional patterns are also useful to check the deep structure of arguments. In
this case, default rules are useful to express in an easy manner that the check is
not successful. For instance, consider an operation that checks whether a string
contains a float number (without an exponent but with an optional minus sign).
With functional patterns and default rules, the definition of this predicate is easy:
isFloat (("-" ? "") ++ n1 ++ "." ++ n2)
| (all isDigit n1 && all isDigit n2) = True
isFloat ’default _ = False
Example 7
In the classical n-queens puzzle, one must place n queens on a chess board so
that no queen can attack another queen. This can be solved by computing some
permutation of the list [1..n], where the i-th element denotes the row of the
queen placed in column i, and check whether this permutation is a safe placement
so that no queen can attack another in a diagonal. The latter property can easily
be expressed with functional patterns and default rules where the non-default rule
fails on a non-safe placement:
safeDiag (_++[x]++zs++[y]++_) | abs (x-y) == length zs + 1 = failed
safeDiag ’default xs = xs
Hence, a solution can be obtained by computing a safe permutation:
queens n = safeDiag (permute [1..n])
This example shows that default rules are a convenient way to express negation-asfailure from logic programming.
Default Rules for Curry
9
Example 8
This programming pattern can also be applied to solve the map coloring problem.
Our map consists of the states of the Pacific Northwest and a list of adjacent states:
data State = WA | OR | ID | BC
adjacent = [(WA,OR),(WA,ID),(WA,BC),(OR,ID),(ID,BC)]
Furthermore, we define the available colors and an operation that associates (nondeterministically) some color to a state:
data Color = Red | Green | Blue
color x = (x, Red ? Green ? Blue)
A map coloring can be computed by an operation solve that takes the information
about potential colorings and adjacent states as arguments, i.e., we compute correct
colorings by evaluating the initial expression
solve (map color [WA,OR,ID,BC]) adjacent
The operation solve fails on a coloring where two states have an identical color and
are adjacent, otherwise it returns the coloring:
solve (_++[(s1,c)]++_++[(s2,c)]++_) (_++[(s1,s2)]++_) = failed
solve ’default cs _ = cs
Note that the compact definition of the standard rule of solve exploits the ordering
in the definition of adjacent. For arbitrarily ordered adjacency lists, we have to
extend the standard rule as follows:
solve (_++[(s1,c)]++_++[(s2,c)]++_) (_++[(s1,s2) ? (s2,s1)]++_)
= failed
5 Transformational Semantics
In order to define a precise semantics of default rules, one could extend an existing
logic foundation of functional logic programming (e.g., (González-Moreno et al. 1999))
to include a meaning of default rules. This approach has been partially done in
(López-Fraguas and Sánchez-Hernández 2004) but without considering the different sources of non-determinism (inside vs. outside) which is important for our
intended semantics, as discussed in Sect. 3. Fortunately, the semantic aspects of
these issues have already been discussed in the context of encapsulated search
(Antoy and Hanus 2009; Christiansen et al. 2013) so that we can put our proposal
on these foundations. Hence, we do not develop a new logic foundation of functional
logic programming with default rules, but we provide a transformational semantics, i.e., we specify the meaning of default rules by a transformation into existing
constructs of functional logic programming.
We start the description of our transformational approach by explaining the
translation of the default rule for zip. A default rule is applied only if no standard
rule is applicable (because the rule’s pattern does not match the argument or the
rule’s condition is not satisfiable). Hence, we translate a default rule into a regular
10
Sergio Antoy and Michael Hanus
rule by adding the condition that no other rule is applicable. For this purpose, we
generate from the original standard rules a set of “test applicability only” rules
where the right-hand side is replaced by a constant (here: the unit value “()”).
Thus, the single standard rule of zip produces the following new rule:
zip ’TEST (x:xs) (y:ys) = ()
Now we have to add to the default rule the condition that zip’TEST is not applicable. Since we are interested in the failure of attempts to apply zip’TEST to the
actual argument, we have to check that this application has no value. Furthermore, non-determinism and failures in the evaluation of actual arguments must be
distinguished from similar outcomes caused by the evaluation of the condition.
All these requirements call for the encapsulation of a search for values of zip’TEST
where “inside” and “outside” non-determinism are distinguished and handled differently. Fortunately, set functions (Antoy and Hanus 2009) (as sketched in Sect. 2)
provide an appropriate solution to this problem. Since set functions have a strategyindependent denotational semantics (Christiansen et al. 2013), we will use them to
specify and implement default rules. Using set functions, one could translate the
default rule into
zip xs ys | isEmpty (zip ’TEST S xs ys) = []
Hence, this rule can be applied only if all attempts to apply the standard rule fail.
To complete our example, we add this translated default rule as a further alternative
to the standard rule so that we obtain the transformed program
zip ’TEST (x:xs) (y:ys) = ()
zip (x:xs) (y:ys) = (x,y) : zip xs ys
zip xs ys | isEmpty (zip ’TEST S xs ys) = []
Thanks to the logic features of Curry, one can also use this definition to generate appropriate argument values for zip. For instance, if we evaluate the equation zip xs ys == [] with the Curry implementation KiCS2 (Braßel et al. 2011), the
search space is finite and computes, among others, the solution {xs=[]}.
Unfortunately, this scheme does not yield the best code to ensure optimal computations. To understand the potential problem, consider the following operation:
f 0 1 = 1
f _ 2 = 2
Intuitively, the best strategy to evaluate a call to f starts with a case distinction
on the second argument, since its value determines which rule to apply. If the value
is 1, and only in this case, the strategy checks the first argument, since its value
determines whether to apply the first rule. A formal characterization of operations
that allow this strategy (Antoy 1992) and a discussion of the strategy itself will be
presented in Sect. 6.2. In this example, the pattern matching strategy is as follows:
1. Evaluate the second argument (to head normal form).
2. If its value is 2, apply the second rule.
3. If its value is 1, evaluate the first argument and try to apply the first rule.
Default Rules for Curry
11
4. Otherwise, no rule is applicable.
In particular, if loop denotes a non-terminating operation, the call f loop 2 evaluates
to 2. This is in contrast to Haskell (Peyton Jones 2003) which performs pattern
matching from left to right so that Haskell loops on this call. This strategy, which is
optimal for the class of programs referred to as inductively sequential (Antoy 1992)
for which it is intended, has been extended to functional logic computations (needed
narrowing (Antoy et al. 2000)) and to overlapping rules (Antoy 1997) in order to
cover general functional logic programs.
Now consider the following default rule for f:
f’default _ x = x
If we apply our transformation scheme sketched above, we obtain the following
Curry program:
f’TEST 0 1 = ()
f’TEST _ 2 = ()
f 0 1 = 1
f _ 2 = 2
f x y | isEmpty (f’TEST S x y) = y
As a result, the definition of f is no longer inductively sequential since the left-hand
sides of the first and third rule overlap. Since there is no argument demanded by all
rules of f, the rules could be applied independently. In fact, the Curry implementation KiCS2 (Braßel et al. 2011) loops on the call f loop 2 (since it tries to evaluate
the first argument in order to apply the first rule), whereas it yields the result 2
without the default rule.
To avoid this undesirable behavior when adding default rules, we could try to
use the same strategy for the standard rules and the test in the default rule. This
can be done by translating the original standard rules into an auxiliary operation
and redefining the original operation into one that either applies the standard rules
or the default rules. For our example, we transform the definition of f (with the
default rule) into the following functions:
f’TEST 0 1 = ()
f’TEST _ 2 = ()
f’INIT 0 1 = 1
f’INIT _ 2 = 2
f’DFLT x y | isEmpty (f’TEST S x y) = y
f x y = f’INIT x y ? f’DFLT x y
Now, both f’TEST and f’INIT are inductively sequential so that the optimal needed
narrowing strategy can be applied, and f simply denotes a choice (without an argument evaluation) between two expressions that are evaluated optimally. Observe
that at most one of these expressions is reducible. As a result, the Curry implementation KiCS2 evaluates f loop 2 to 2 and does not run into a loop.
The overall transformation of default rules can be described by the following
12
Sergio Antoy and Michael Hanus
scheme (its simplicity is advantageous to obtain a comprehensible definition of the
semantics of default rules). The operation definition
f tk1 | c1 = e1
.
.
.
f tkn | cn = en
f ’default tkn+1 | cn+1 = en+1
is transformed into (where f ’TEST, f ’INIT, f ’DFLT are new operation identifiers):
f ’TEST tk1 | c1 = ()
.
.
.
f ’TEST tkn | cn = ()
f ’INIT tk1 | c1 = e1
.
.
.
f ’INIT tkn | cn = en
f ’DFLT tkn+1 | isEmpty (f ’TEST S tkn+1 ) && cn+1 = en+1
f xk = f ’INIT xk ? f ’DFLT xk
Note that the patterns and conditions of the original rules are not changed. Hence,
this transformation is also compatible with other advanced features of Curry, like
functional patterns, “as” patterns, non-linear patterns, local declarations, etc. Furthermore, if an efficient strategy exists for the original standard rules, the same
strategy can be applied in the presence of default rules. This property can be formally stated as follows:
Proposition 1
Let R be a program without default rules, and R′ be the same program except that
default rules are added to some operations of R. If R is overlapping inductively
sequential, so is R′ .
Proof
Let f be an operation of R. The only interesting case is when a default rule of f is
in R′ . Operation f of R produces four different operations of R′ : f , f ’DFLT, f ’INIT,
and f ’TEST. The first two are overlapping inductively sequential since they are
defined by a single rule. The last two are overlapping inductively sequential when
f of R is overlapping inductively sequential since they have the same definitional
tree as f modulo a renaming of symbols.
The above proposition could be tightened a little when operation f is non-overlapping. In this case three of the four operations produced by the transformation are
non-overlapping as well. Prop. 1 is important for the efficiency of computations. In
overlapping inductively sequential systems, needed redexes exist and can be easily
and efficiently computed (Antoy 1997). If the original system has a strategy that
reduces only needed redexes, the transformed system has a strategy that reduces
Default Rules for Curry
13
only needed redexes. This ensures that optimal computations are preserved by the
transformation regardless of non-determinism.
This result is in contrast to Haskell (or Prolog), where the concept of default
rules is based on a sequential testing of rules, which might inhibit optimal evaluation and prevent or limit non-determinism. Hence, our concept of default rules is
more powerful than existing concepts in functional or logic programming (see also
Sect. 8).
We now relate values computed in the original system to those computed in the
transformed system and vice versa. As expected, extending an operation with a
default rule preserves the values computed without the default rule.
Proposition 2
Let R be a program without default rules, and R′ be the same program except that
default rules are added to some operations of R. If e is an expression of R that
evaluates to the value t w.r.t. R, then e evaluates to t w.r.t. R′ .
Proof
Let f tk → u w.r.t. R, for some expression u, a step of the evaluation of e. The
only interesting case is when a default rule of f is in R′ . By the definitions of f and
f ’INIT in R′ , f tk → f ’INIT tk → u w.r.t. R′ . A trivial induction on the length of
the evaluation of e completes the proof.
The converse of Prop. 2 does not hold because R′ typically computes more values
than R—that is the reason why there are default rules. The following statement
relates values computed in R′ to values computed in R.
Proposition 3
Let R be a program without default rules, and R′ be the same program except
that default rules are added to some operations of R. If e is an expression of R
that evaluates to the value t w.r.t. R′ , then either e evaluates to t w.r.t. R or some
∗
default rule of R′ is applied in e → t in R′ .
Proof
∗
Let A denote an evaluation e → t in R′ that never applies default rules. For any
operation f of R, the steps of A are of two kinds: (1) f tk → f ’INIT tk or (2)
f ’INIT tk → t ′ , for some expressions tk and t ′ . If we remove from A the steps of
kind (1) and replace f ’INIT with f , we obtain an evaluation of e to t in R.
In Curry, by design, the textual order of the rules is irrelevant. A default rule is
a constructive alternative to a certain kind of failure. For these reasons, a single
default rule, as opposed to multiple default rules without any order, is conceptually
simpler and adequate in practical situations. Nevertheless, a default rule of an
operation f may invoke an auxiliary operation with multiple ordinary rules, thus,
producing the same behavior of multiple default rules of f .
14
Sergio Antoy and Michael Hanus
6 Implementation
The implementation of default rules for Curry based on the transformational approach is available as a preprocessor. The preprocessor is integrated into the compilation chain of the Curry systems PAKCS (Hanus et al. 2016) and KiCS2 (Braßel et al. 2011).
In some future version of Curry, one could also add a specific syntax for default
rules and transform them in the front end of the Curry system.
The transformation scheme shown in the previous section is mainly intended to
specify the precise meaning of default rules (similarly to the specification of the
meaning of guards in Haskell (Peyton Jones 2003)). Although this transformation
scheme leads to a reasonably efficient implementation, the actual implementation
can be improved in various ways. In the following, we present two approaches to
improve the implementation of default rules.
6.1 Avoiding Duplicated Condition Checking
Our transformation scheme for default rules generates from a set of standard rules
the auxiliary operations f ’TEST and f ’INIT. f ’TEST is used in the condition of the
translated default rule to check the applicability of a standard rule, whereas f ’INIT
actually applies a standard rule. Since both alternatives (standard rules or default
rule) are eventually tried for application, the pattern matching and condition checking of some standard rule might be duplicated. For instance, if a standard rule is
applicable to some call and the same call matches the pattern of the default rule, it
might be tried twice: (1) the standard rule is applied by f ’INIT, and (2) its pattern
and condition is tested by f ’TEST in order to test the (non-)emptiness of the set of
all results. Although the amount of duplicated work is difficult to assess accurately
due to Curry’s lazy evaluation strategy (e.g., to check the non-emptiness in the
condition of f ’DFLT, it suffices to compute at most one element of the set), there is
some risk for operationally complex conditions or patterns, e.g., functional patterns.
This kind of duplicated work can be avoided by a more sophisticated transformation scheme where the common parts of the definitions of f ’TEST and f ’INIT
are joined into a single operation. This operation first tests the application of a
standard rule and, in case of a successful test, returns a continuation to proceed
with the corresponding rule. For instance, consider the rules for zip presented in
Example 2. The operations zip’TEST and zip’INIT generated by our first transformation scheme can be joined into a single operation zip’TESTC by the following
transformation:
zip ’TESTC (x:xs) (y:ys) = \_ → (x,y) : zip xs ys
zip ’DFLT _ _ = []
zip xs ys = let cs = zip ’TESTC S xs ys
in if isEmpty cs then zip ’DFLT xs ys
else (chooseValue cs) ()
Now, the standard rule is translated into a rule for the new operation zip’TESTC
where the rule’s right-hand side is encapsulated into a lambda abstraction to avoid
Default Rules for Curry
15
its immediate evaluation if this rule is applied. The actual implementation of zip
first checks whether the set of all such lambda abstractions is empty. If this is
the case, the standard rule is not applicable so that the default rule is applied.
Otherwise, we continue with the right-hand sides of all applicable standard rules
collected as lambda abstractions in the set cs.3
The general transformation scheme to obtain this behavior is defined as follows.
An operation definition of the form
f tk1 | c1 = e1
.
.
.
f tkn | cn = en
f ’default tkn+1 | cn+1 = en+1
is transformed into:
f ’TESTC tk1 | c1 = \_ → e1
.
.
.
f ’TESTC tkn | cn = \_ → en
f ’DFLT tkn+1 | cn+1 = en+1
f xk = let cs = f ’TESTC S xk
in if isEmpty cs then f ’DFLT xk
else (chooseValue cs) ()
Obviously, this modified scheme avoids the potentially duplicated condition checking in standard rules, but it is more sophisticated since it requires the handling of
sets of continuations. Depending on the implementation of set functions, this might
be impossible if the values are operations. If the results computed by set functions
are actually sets (and not multi-sets), this scheme cannot be applied since sets require an equality operation on elements in order to eliminate duplicated elements.
Fortunately, this scheme is applicable with PAKCS (Hanus et al. 2016), which
computes multi-sets as results of set functions so that it does not require equality
on elements. Thus, we compare the run times of both schemes for some of the
operations shown above which contain complex applicability conditions (functional
patterns). All benchmarks were executed on a Linux machine (Debian Jessie) with
an Intel Core i7-4790 (3.60Ghz) processor and 8GB of memory. Figure 1 shows
the run times (in seconds) to evaluate some operations with both schemes. These
benchmarks indicate that the new scheme might yield a reasonable performance
gain, although this clearly depends on the particular example. A further alternative
transformation scheme is discussed in the following section.
3
The operation chooseValue non-deterministically chooses some value of the given set.
16
Sergio Antoy and Michael Hanus
System:
PAKCS 1.14.0 (Hanus et al. 2016)
Operation:
Arguments:
isSet
[1..1000]
isSet
[1000,1..1000]
lookup
5001
[(1,.)..(5000,.)]
lookup
5000
[(1,.)..(5000,.)]
queens
6
Sect. 5:
7.04
4.78
3.56
3.54
0.23
Sect. 6.1:
2.27
2.28
1.81
3.57
0.23
Fig. 1. Performance comparison of different transformation schemes.
6.2 Transforming Default Rules into Standard Rules
In some situations, the behavior of a default rule can be provided by a set of standard rules. Almost universally, standard rules are more efficient. An example of this
situation is provided with the operation zip. In Example 2 this operation is defined
with a default rule. A definition using standard rules is shown at the beginning
of Sect. 3. The input/output relations of the two definitions are identical. In this
section, we introduce a few concepts to describe how to obtain, under sufficient
conditions, a set of standard rules that behave as a default rule.
The programs considered in this section are constructor-based (O’Donnell 1977)
(the extension to functional patterns is discussed later). Thus, there are disjoint
sets of operation symbols, denoted by f , g, . . ., and constructor symbols, denoted
by c, d , . . . An f -rooted pattern is an expression of the form f tn where f is an
operation symbol of arity n, each ti is an expression consisting of variables and/or
constructor symbols only, and f tn is linear, i.e., there are no repeated occurrences
of some variable. A pattern is an f -rooted pattern for some operation f . A pattern
is ground if it does not contain any variable. A program rule has the form l = r
where the left-hand side l is a pattern (the extension to conditional rules is discussed
later). Given a redex t and a step t → u, u is called a contractum (of t ). Although
Curry allows non-linear patterns for the convenience of the programmer, they are
transformed into linear ones through a simple syntactic transformation.
In the following, we first consider a specific class of programs, called inductively
sequential, where the rules of each operation can be organized in a definitional tree
(Antoy 1992).
Definition 1 (Definitional tree)
The symbols rule, exempt, and branch, appearing below, are uninterpreted functions for classifying the nodes of a tree. A partial definitional tree with an f -rooted
pattern p is either a rule node rule(p = r ), an exempt node exempt (p), or a branch
node branch(p, x , Tk ), where x is a variable in p (also called the inductive variable),
{c1 , . . . , ck } is the set of all the constructors of the type of x , the substitution σi
maps x to ci xai (where xai are all fresh variables and ai is the arity of ci ), and
Ti is a partial definitional tree with pattern σi (p) (for i = 1, . . . , k ). A definitional
tree T of an operation f is a finite partial definitional tree with pattern f xn , where
Default Rules for Curry
17
n is the arity of f and xn are pairwise different variables, such that T contains
all and only the rules defining f (up to variable renaming). In this case, we call f
inductively sequential.
Definitional trees have a comprehensible graphical representation. For instance, the
definitional tree of the operation “++” defined in Example 1 is shown in Fig. 2. In
this graphical representation, the pattern of each node is shown. The root node is a
branch and its children are rule nodes. The inductive variable of the branch is the
left operand of “++”. Referring to Def. 1, σ1 maps this variable to “[]” and σ2 to
(x : xs). For rule nodes, the right-hand side of the rule is shown below the arrow.
Exempt nodes are marked by the keyword exempt, as shown in Figures 3 and 4.
- ++ ys ◗
◗◗◗◗
(x:xs) ++ ys
[] ++ ys
s
sss
ys
x : (xs ++ ys)
Fig. 2. A definitional tree of the operation “++”
.
For the sake of completeness, we sketch how definitional trees are used by the
evaluation strategy. The details can be found in (Antoy 1997). We discuss how to
compute a rewrite of an expression rooted by an operation. More general cases are
reduced to that. It can be shown that any step so computed is needed. Thus, let t
be an expression rooted by an operation f and T a definitional tree of f . A traversal
of T finds the deepest node N in T whose pattern p matches t . Such a node, and
pattern, exist for every t . If N is a rule node, then t is a redex and is reduced. If
N is an exempt node, then the computation is aborted because t has no value as
in, e.g., head [], the head of an empty list. If N is a branch node, then the match
of the inductive variable of p is an expression t ′ rooted by some operation and the
strategy recursively seeks to compute a step of t ′ .
Before presenting our transformation, we state an important property of definitional trees.
Definition 2 (Mutually exclusive and exhaustive patterns)
Let f be an operation symbol and S a set of f -rooted patterns. We say that the
patterns of S are mutually exclusive iff for any ground f -rooted pattern p, no two
distinct patterns of S match p, and we say that the patterns of S are exhaustive
iff for any ground f -rooted pattern p, there exists a pattern in S that matches p.
Lemma 1 (Uniqueness)
Let f be an operation defined by a set of standard rules. If T is a definitional tree
of the rules of f , then the patterns in the leaves of T are exhaustive and mutually
exclusive.
18
Sergio Antoy and Michael Hanus
Proof
Let p be any ground f -rooted pattern, π the pattern in a node N of T , and suppose
that π matches p. Initially, we show that if N is not a leaf of T , there is exactly
one child N ′ of N such the pattern π ′ of N ′ matches p. Let x be the inductive
variable of π and q the subexpression of p matched by x . Since p is ground and
q is a proper subexpression of p, q is rooted by some constructor symbol c. Let
{c1 , . . . ck } be the set of all the constructors of the type defining c and let ai be the
arity of ci , for all appropriate i. By Def. 1, N has k children with patterns σi (π),
where σi = {x 7→ ci x ai } and x ai is a fresh variable, for all appropriate i. Hence,
exactly one of these patterns matches p since ci x ai matches q iff ci = c. Going
back to the proposition’s claim, since the pattern in the root of T matches p, by
induction on the depth of T , there is exactly one leaf whose pattern matches p.
Inductive sequentiality is sufficient, but not necessary for a set of exhaustive and
mutually exclusive patterns. We will later show a non-inductively sequential operation with exhaustive and mutually exclusive patterns. Nevertheless, inductive
sequentiality supports a constructive method to transform default rules. Since not
every definitional tree is useful to define our transformation, we first restrict the
set of definitional trees.
Definition 3 (Minimal definitional tree)
A definitional tree is minimal iff there is some rule node below any branch node of
the tree.
For example, consider the operation isEmpty defined by the single rule
isEmpty [] = True
Fig. 3 shows a non-minimal tree of the rules defining isEmpty. The right child of
the root is a branch node that has no rule node below it. In a minimal tree of the
rules defining isEmpty, the right child would be an exempt node.
isEmpty ❙
❧❧
❧❧❧❧
❙❙❙❙
❙
isEmpty (-:-)
❘
isEmpty []
True
♠♠
♠♠♠
isEmpty (-:[])
(exempt)
❘❘❘❘
❘
isEmpty (-:(-:-))
(exempt)
Fig. 3. A non-minimal definitional tree of the operation isEmpty
.
We now investigate sufficient conditions for the equivalence between an operation
defined with a default rule and an operation defined by standard rules only.
Definition 4 (Replacement of a default rule)
Let f be an operation defined by a set of standard rules and a default rule f xk = t ,
where xk are pairwise different variables and t some expression, and let T be a
minimal definitional tree of the standard rules of f . Let N1 , N2 , . . . Nn be the exempt
19
Default Rules for Curry
nodes of T , tki the pattern of node Ni and σi the substitution {xk 7→ tki }, for
1 6 i 6 n. The following set of standard rules of f is called a replacement of the
default rule of f :
σi (f xk = t ),
for 1 6 i 6 n
(1)
Fig. 4 shows a minimal definitional tree of the single standard rule of operation
zip defined at the beginning of Sect. 3. The right-most leaf of this tree holds this
rule. Since this leaf is below both branch nodes, the definitional tree is minimal
according to Def. 3. The remaining two leaves hold the patterns that match all and
only the combinations of arguments to which the default rule would be applicable.
These patterns are more instantiated than that of the default rule, but we will see
that any expression reduced by these rules does not need any additional evaluation
with respect to the default rule.
zip [] (exempt)
zip - PPP
PPP
P
zip (x:xs) P
q
qqq
♣♣
♣♣♣
zip (x:xs) []
(exempt)
PPP
PPP
PP
zip (x:xs) (y:ys)
(x,y) : zip xs ys
Fig. 4. A definitional tree of the standard rule of operation zip defined in Sect. 3
.
Lemma 2 (Correctness)
Let f be an operation defined by a set S of standard rules and a default rule r
of the form f xk = t , where each xi is a variable, for all appropriate i, and some
expression t , and let R be the replacement of r . For any ground f -rooted pattern
p, p is reduced at the root to some q by the default rule r iff p is reduced at the
root to q by some rule of R.
Proof
The proof is done in two steps. First, we prove that p is reduced by r iff p is
reduced by some rule of R. Then we prove that the contracta by the two rules
are the same. By Lemma 1, the patterns in the rules of S ∪ R are exhaustive and
mutually exclusive. Therefore, p is reduced by r if and only if p is not reduced by
any rule of S if and only if p is reduced by some rule of R. We now prove the equality
of the contracta. In the remainder of this proof, all the substitutions are restricted
to xk , the argument variables of r . If p is reduced by r with some match σ, then
p = σ(f xk ) and q = σ(t ). Pattern p is also reduced by some rule of R which, by
Def. 4, is of the form σi (r ), for some substitution σi . Consequently, p = σ ′ (σi (f xk ))
for some match σ ′ . Since p is ground, σ = σ ′ ◦ σi . Thus, the contractum of p by
the rule of R is σ ′ (σi (t )) = σ(t ) = q.
20
Sergio Antoy and Michael Hanus
In Def. 4, the replacement of a default rule is constructed for a minimal definitional
tree. The hypothesis of minimality is not used in the proof of Lemma 2. The reason is
that the lemma claims a property of f -rooted ground patterns. During the execution
of a program, the default rule may be applied to some f -rooted expression e that
may neither be a pattern nor ground. The hypothesis of minimality ensures that,
in this case, no additional evaluation of e is required when a replacement rule is
applied instead of the default rule. This fact is counter intuitive and non-trivial
since the pattern of the default rule matches any f -rooted expression, whereas the
patterns in the replacement rules do not, except in the degenerate case in which
the set of standard rules is empty. However, a default rule is applicable only if
no standard rule is applicable. Therefore, expression e must have been evaluated
“enough” to determine that no standard rule is applicable. The following lemma
shows that this evaluation is just right for the application of a replacement rule.
Lemma 3 (Evaluation)
Let e be an f -rooted expression reduced by the default rule of f according to the
transformational semantics of Sect. 5. Let T be a minimal definitional tree of (the
standard rules of) f . There exists an exempt node of T whose pattern matches e.
Proof
First note that the standard rules of f and the rules of f ’TEST, as defined in Sect. 5,
have identical left-hand sides. Hence T is also a minimal definitional tree of the
rules of f ’TEST, which are used to check the applicability of the default rule.
To prove the claim, we construct a path N0 , N1 , . . . Np in the definitional tree T of
f with the following invariant properties: (a) the pattern πi of each Ni unifies with
e, and, (b) if the last node Np is a leaf of T , Np is an exempt node. Establishing
the invariant: N0 is the root node of T . By definition, its pattern π0 is f xk , where
xk are fresh distinct variables. Hence π0 unifies with e so that invariant (a) holds.
Furthermore, if N0 is a leaf of T , then it cannot be a rule node, otherwise e would
never be reduced by a default rule. Hence N0 is an exempt node, i.e., invariant (b)
holds. Maintaining the invariant: We assume that the invariant (a) holds for node
Nk , for some k > 0. If Nk is a leaf of T , then, as in the base case, Nk must be
an exempt node. Hence we assume Nk is a branch of T and show that invariant
(a) can be extended to some child Nk +1 of Nk . Since Nk is a branch node, e and
πk unify. For each child N ′ of Nk , let π ′ be the pattern of N ′ , and let v be the
inductive variable of the branch node Nk . By the definition of T , π ′ = σ ′ (πk ),
where σ ′ = {v 7→ c xac }, c is a constructor symbol of arity ac , and xi is a fresh
variable for any appropriate i. Let σ be the match of πk to e and t = σ(v ). If t is a
variable, then any child of Nk satisfies invariant (a). Otherwise, t must be rooted by
some constructor symbol, say d , for the following reasons. Because T is minimal,
there are one or more rule nodes below Nk . The pattern in any of these rules is
an instance of πk that has some constructor symbol in the position matched by v .
Hence, unless t is constructor-rooted, it would be impossible to tell which, if any,
of these rules reduces e, hence it would be impossible to say whether e must be
reduced by a standard rule or the default rule. Hence, Nk +1 is the child in which v
is mapped to d xad so that invariant (a) also holds for Nk +1 .
Default Rules for Curry
21
We define the replacement of a default rule by a set of standard rules under four
assumptions. We assess the significance of these assumptions below.
Inductive sequentiality. The standard rules are inductively sequential. This is a
very mild requirement in practice. For instance, every operation of the Curry Prelude, except for the non-deterministic choice operator “?” shown in Sect. 2, is inductively sequential. Non-inductively sequential operations are problematic to evaluate
efficiently. E.g., the following operation, adapted from (Berry 1976, Prop. II.2.2),
is defined by rules that do not admit a definitional tree:
f False True x
= ...
f x
False True = ...
f True x
False = ...
To apply f , the evaluation to constructor normal form of two out of the three
arguments is both necessary and sufficient. No practical way is known to determine
which these two arguments are without evaluating all three. Furthermore, since
the evaluation of an argument may not terminate, the three arguments must be
evaluated concurrently (but see (Antoy and Middeldorp 1996)).
Most general pattern. We assumed in our transformation that the pattern of the
default rule is most general, i.e., the arguments of the operation are all variables.
Choosing the most general pattern keeps the statement of Lemma 3 simple and
direct. With this assumption, no extra evaluation of the arguments is needed for
the application of a replacement rule. To relax this assumption, we can modify
Def. 4 as follows. If the left-hand side of the default rule is f uk , we look for a most
general unifier, say σi , of uk and tki . Then rule σi (f uk → t ) is in the replacement
of the default rule iff such a σi exists.
Unconditional rules. Both standard rules and the default rule are unconditional.
Adding a condition to the default rule is straightforward, similar to the transformation shown in Sect. 5. The condition of a default rule is directly transferred to each
replacement rule by extending display (1) in Def. 4 with the condition. By contrast,
conditions in standard rules require some care. With a modest loss of generality,
assume that the standard rules have a definitional tree where each leaf node has a
conditional rule of the form:
f tk | c = t
(2)
where c is a Boolean expression and t is any expression. Lemma 1 proves that if p
is any f -rooted ground pattern matched by f tk no other standard rule matches p.
Hence, p is reduced at the root by the default rule of f iff c is not satisfied by p.
Therefore, we need the following rule in the replacement of the default rule
f tk | ¬c = t
(3)
where ¬c denotes the “negation” of c, i.e., the condition satisfied by all the patterns
matched by f tk that do not satisfy c. In the spirit of functional logic programming,
c is evaluated non-deterministically. For example, consider an operation that takes
a list of colors, say Red, Green and Blue, and removes all Red occurrences from the
list:
22
Sergio Antoy and Michael Hanus
data Color = Red | Green | Blue
remred cs | cs == x++[Red]++y
= remred (x++y)
where x,y free
remred ’default cs = cs
The first rule is applied if there exist x and y that satisfy the condition. E.g., for
the list [Red,Green,Red,Blue] there are two such combinations of x and y. Thus, the
“negation” of this condition must negate the existence of any such x and y. This
can be automatically done according to the transformational semantics presented
in Sec. 5, but applied to a single rule. This example’s replacement of the default
rule is shown below:
remred cs | isEmpty (remred ’TEST S cs) = cs
remred ’TEST cs | _++[Red]++_ == cs = ()
Constructor patterns. The standard rules defining an operation have constructor
patterns. Curry also provides functional patterns, presented in Sec. 2. Rules defined
by functional patterns can be transformed into ordinary rules (Antoy and Hanus 2005,
Def. 4) by moving the functional pattern matching into the condition of a rule.
Hence, the absence of functional patterns from our discussion is not an intrinsic
limitation. Since functional patterns are quite expressive, operations defined with
functional patterns often consist of a single program rule and a default rule (as in
all examples shown in in Sect. 4). For instance, the previous operation remred can
be defined with a functional pattern as follows:
remred (x++[Red]++y) = remred (x++y)
remred ’default
cs = cs
Hence, the improved transformation scheme presented in Sect. 6.1 is still useful and
should be applied in combination with the transformation shown in this section.
7 Benchmarking
To show the practical advantage of the transformation described in the previous
section, we evaluated a few simple operations defined in a typical functional programming style with default rules. For instance, the Boolean conjunction can be
defined with a default rule:
and True True = True
and ’default _ _ = False
The replacement of the default rule consists of two rules so that the transformation
yields the following standard rules:
and True True = True
and True False = False
and False _
= False
Similarly, the computation of the last element of a list can be defined with a default
rule:
Default Rules for Curry
23
last [x] = x
last’default (_:xs) = last xs
Our final example extracts all values in a list of optional (“Maybe”) values:
catMaybes []
= []
catMaybes (Just x : xs) = x : catMaybes xs
catMaybes ’default (_:xs) = catMaybes xs
With the introduction of default rules, the order of evaluation may become more
arbitrary, even though only needed steps are executed. For example, in the first
definition of operation and both arguments must be evaluated, in any order, for the
application of the standard rule. If the evaluation of one argument does not terminate and the other one evaluates to False, the order in which the two arguments are
evaluated becomes observable. This situation is not directly related to the presence
of a default rule. There are two “natural” inductive definitions of operation and, one
evaluates the first argument first, as in the second definition of and, and another
evaluates the second argument first. From the single standard rule of and, we cannot say which of the two definitions was intended. If the default rule of operation
and is replaced by a set of standard rules, as per Sec. 6.2, the resulting definition,
which is inductively sequential, will explicitly and arbitrarily encode which of the
two arguments is to be evaluated first.
As discussed earlier, functional logic computations execute narrowing steps, i.e.,
steps in which some variable of an expression is instantiated and the rule reducing
the expression depends on the instantiation of the variable. For example, consider
again the and operation for its simplicity. The evaluation of and x True, where x
is a free variable, narrows x to True to apply the standard rule and narrows x to
False to apply the default rule. In a narrowing step, a variable is instantiated by
the unification of the expression being evaluated and the left-hand side of a rule.
This does not work with a default rule, since the arguments in the left-hand side
are themselves variables. In particular, the transformational semantics of and has
no rule to unify x with False. To obtain the intended behavior in narrowing steps
variables are instantiated by generators (Antoy and Hanus 2006). In the example
being discussed, the Boolean generator is True ? False.
Figure 5 shows the run times (in seconds) to evaluate the operations discussed
in this section with the different transformation schemes (i.e., the scheme of Sect. 5
and the replacement of default rules presented in this section) and different Curry
implementations (where “call size” denotes the number of calls to and and the
lengths of the input lists for the other examples). The benchmarks were executed
on the same machine as the benchmarks in Sect. 6.1. The results clearly indicate
the advantage of replacing default rules by standard rules, in particular for PAKCS,
which has a less sophisticated implementation of set functions than KiCS2.
8 Related Work
In this section, we compare our proposal of default rules for Curry with existing
proposals for other rule-based languages.
24
Sergio Antoy and Michael Hanus
System:
Operation / call size:
PAKCS 1.14.0 (Hanus et al. 2016)
zip / 1000
and / 100000
last / 2000
catMaybes / 2000
Sect. 5:
3.66
8.46
2.53
2.45
Sect. 6.2:
0.01
0.25
0.01
0.01
System:
KiCS2 0.5.0 (Braßel et al. 2011)
zip / 106
and / 106
last / 105
catMaybes / 106
Sect. 5:
2.72
1.35
0.38
0.40
Sect. 6.2:
0.04
0.08
0.01
0.01
Operation / call size:
Fig. 5. Performance comparison of different schemes for different compilers for some
operations discussed in this section.
The functional programming language Haskell (Peyton Jones 2003) has no explicit concept of default rules. Since Haskell applies the rules defining a function
sequentially from top to bottom, it is a common practice in Haskell to write a “catch
all” rule as a final rule to avoid writing several nearly identical rules (see example
zip at the beginning of Sect. 3). Thus, our proposal for default rules increases the
similarities between Curry and Haskell. However, our approach is more general,
since it also supports logic-oriented computations, and it is more powerful, since it
ensures optimal evaluation for inductively sequential standard rules, in contrast to
Haskell (as shown in Sect. 5).
Since Haskell applies rules in a sequential manner, it is also possible to define
more than one default rule for a function, e.g., where each rule has a different specificity. This cannot be directly expressed with our default rules where at most one
default rule is allowed. However, one can obtain the same behavior by introducing
a sequence of auxiliary operations where each operation has one default rule.
The logic programming language Prolog (Deransart et al. 1996) is based on backtracking where the rules defining a predicate are sequentially applied. Similarly to
Haskell, one can also define “catch all” rules as the final rules of predicate definitions. In order to avoid the unintended application of these rules, one has to put
“cut” operators in the preceding standard rules. As already discussed in Sect. 3,
these cuts are only meaningful for instantiated arguments, otherwise the completeness of logic programming might be destroyed. Hence, this kind of default rules
can be used only if the predicate is called in a particular mode, in contrast to
our approach. The completeness for arbitrary modes might require the addition
of concepts from Curry into Prolog, like the demand-driven instantiation of free
variables.
Various encapsulation operators have been proposed for functional logic programs
Default Rules for Curry
25
(Braßel et al. 2004) to encapsulate non-deterministic computations in some data
structure. Set functions (Antoy and Hanus 2009) have been proposed as a strategyindependent notion of encapsulating non-determinism to deal with the interactions
of laziness and encapsulation (see (Braßel et al. 2004) for details). One can also use
set functions to distinguish successful and non-successful computations, similarly to
negation-as-failure in logic programming, exploiting the possibility to check result
sets for emptiness. When encapsulated computations are nested and performed
lazily, it turns out that one has to track the encapsulation level in order to obtain
intended results, as discussed in (Christiansen et al. 2013). Thus, it is not surprising
that set functions and related operators fit quite well to our proposal. Actually,
many explicit uses of set functions in functional logic programming to implement
negation-as-failure can be implicitly and more tersely encoded with our concept of
default rules, as shown in Examples 7 and 8.
Default rules and negation-as-failure have been also explored in (López-Fraguas and Sánchez-Hernández 2004;
Sánchez-Hernández 2006) for functional logic programs. In these works, an operator, fails, is introduced to check whether every reduction of an expression to a
head-normal form is not successful. (López-Fraguas and Sánchez-Hernández 2004)
proposes the use of this operator to define default rules for functional logic programming. However, the authors propose a scheme where the default rule is applied if
no standard rule was able to compute a head normal form. This is quite unusual
and in contrast to functional programming (and our proposal) where default rules
are applied if pattern matching and/or conditions of standard rules fail, but the
computations of the rules’ right-hand sides are not taken into account to decide
whether a default rule should be applied. The same applies to an early proposal for
default rules in an eager functional logic language (Moreno-Navarro 1994). Since
the treatment of different sources of non-determinism and their interaction were
not explored at that time, nested computations with failures are not considered by
these works. As a consequence, the operator fails might yield unintended results
if it is used in nested expressions. For instance, if we use fails instead of set functions to implement the operation isUnit defined in Example 4, the evaluation of
isUnit failed yields the value False in contrast to our intended semantics.
Finally, we proposed in (Antoy and Hanus 2014) to change Curry’s rule selection
strategy to a sequential one. However, it turned out that this change has drawbacks
w.r.t. the evaluation strategy, since formerly optimal reductions are no longer possible in particular cases. For instance, consider the operation f defined in Sect. 5
and the call f loop 2. In a sequential rule selection strategy, one starts by testing
whether the first rule is applicable. Since both arguments are demanded by this
rule, one might evaluate them from left to right (as done in the implementation
(Antoy and Hanus 2014)) so that this evaluation does not terminate. This problem is avoided with our proposal which returns 2 even in the presence of a default
rule for f. Moreover, the examples presented in (Antoy and Hanus 2014) can be
expressed with default rules in a similar way.
26
Sergio Antoy and Michael Hanus
9 Conclusions
We proposed a new concept of default rules for Curry. Default rules are available
in many rule-based languages, but a sensible inclusion into a functional logic language is demanding. Therefore, we used advanced features for encapsulating search
to define and implement default rules. Thanks to this approach, typical logic programming features, like non-determinism and evaluating operations with unknown
arguments, are still applicable with our new semantics. This distinguishes our approach from similar concepts in logic programming which simply cut alternatives.
Our approach can lead to more elegant and comprehensible declarative programs,
as shown by several examples in this paper. Moreover, many uses of negationas-failure, which are often implemented in functional logic programs by complex
applications of encapsulation operators, can easily be expressed with default rules.
Since encapsulated search is more costly than simple pattern matching, we have
also shown some opportunities to implement default rules more efficiently. In particular, if the standard rules are inductively sequential and unconditional, one can
replace the default rules by a set of standard rules so that the usage of encapsulated
search can be completely avoided.
Acknowledgments. The authors are grateful to Sandra Dylus and the anonymous
reviewers for their suggestions to improve a previous version of this paper. This
material is based in part upon work supported by the National Science Foundation
under Grant No. 1317249.
References
Antoy, S. 1992. Definitional trees. In Proc. of the 3rd International Conference on
Algebraic and Logic Programming. Springer LNCS 632, 143–157.
Antoy, S. 1997. Optimal non-deterministic functional logic computations. In 6th Int’l
Conf. on Algebraic and Logic Programming (ALP’97). Vol. 1298. Springer LNCS,
Southampton, UK, 16–30.
Antoy, S., Echahed, R., and Hanus, M. 2000. A needed narrowing strategy. Journal
of the ACM 47, 4, 776–822.
Antoy, S. and Hanus, M. 2005. Declarative programming with function patterns. In
Proceedings of the International Symposium on Logic-based Program Synthesis and
Transformation (LOPSTR’05). Springer LNCS 3901, 6–22.
Antoy, S. and Hanus, M. 2006. Overlapping rules and logic variables in functional logic
programs. In Proceedings of the 22nd International Conference on Logic Programming
(ICLP 2006). Springer LNCS 4079, 87–101.
Antoy, S. and Hanus, M. 2009. Set functions for functional logic programming. In
Proceedings of the 11th ACM SIGPLAN International Conference on Principles and
Practice of Declarative Programming (PPDP’09). ACM Press, 73–82.
Antoy, S. and Hanus, M. 2010. Functional logic programming. Communications of the
ACM 53, 4, 74–85.
Antoy, S. and Hanus, M. 2014. Curry without Success. In Proc. of the 23rd International
Workshop on Functional and (Constraint) Logic Programming (WFLP 2014). CEUR
Workshop Proceedings, vol. 1335. CEUR-WS.org, 140–154.
Default Rules for Curry
27
Antoy, S. and Middeldorp, A. 1996. A sequential strategy. Theoretical Computer
Science 165, 75–95.
Berry, G. 1976. Bottom-up computation of recursive programs.
Braßel, B., Hanus, M., and Huch, F. 2004. Encapsulating non-determinism in functional logic computations. Journal of Functional and Logic Programming 2004, 6.
Braßel, B., Hanus, M., Peemöller, B., and Reck, F. 2011. KiCS2: A new compiler
from Curry to Haskell. In Proc. of the 20th International Workshop on Functional and
(Constraint) Logic Programming (WFLP 2011). Springer LNCS 6816, 1–18.
Christiansen, J., Hanus, M., Reck, F., and Seidel, D. 2013. A semantics for weakly
encapsulated search in functional logic programs. In Proc. of the 15th International
Symposium on Principle and Practice of Declarative Programming (PPDP’13). ACM
Press, 49–60.
Deransart, P., Ed-Dbali, A., and Cervoni, L. 1996. Prolog - the standard: reference
manual. Springer.
González-Moreno, J., Hortalá-González, M., López-Fraguas, F., and
Rodrı́guez-Artalejo, M. 1999. An approach to declarative programming based on a
rewriting logic. Journal of Logic Programming 40, 47–87.
Hanus, M. 2011. Declarative processing of semistructured web data. In Technical Communications of the 27th International Conference on Logic Programming. Vol. 11. Leibniz
International Proceedings in Informatics (LIPIcs), 198–208.
Hanus, M. 2013. Functional logic programming: From theory to Curry. In Programming
Logics - Essays in Memory of Harald Ganzinger. Springer LNCS 7797, 123–168.
Hanus, M., Antoy, S., Braßel, B., Engelke, M., Höppner, K., Koj, J., Niederau,
P., Sadre, R., and Steiner, F. 2016. PAKCS: The Portland Aachen Kiel Curry
System. Available at http://www.informatik.uni-kiel.de/∼pakcs/.
Hanus (ed.), M. 2016. Curry: An integrated functional logic language (vers. 0.9.0).
Available at http://www.curry-language.org.
López-Fraguas, F. and Sánchez-Hernández, J. 2004. A proof theoretic approach to
failure in functional logic programming. Theory and Practice of Logic Programming 4, 1,
41–74.
Moreno-Navarro, J. 1994. Default rules: An extension of constructive negation for
narrowing-based languages. In Proc. Eleventh International Conference on Logic Programming. MIT Press, 535–549.
O’Donnell, M. J. 1977. Computing in Systems Described by Equations. Springer LNCS
58.
Peyton Jones, S., Ed. 2003. Haskell 98 Language and Libraries—The Revised Report.
Cambridge University Press.
Reddy, U. 1985. Narrowing as the operational semantics of functional languages. In
Proc. IEEE Internat. Symposium on Logic Programming. Boston, 138–151.
Sánchez-Hernández, J. 2006. Constructive failure in functional-logic programming:
From theory to implementation. Journal of Universal Computer Science 12, 11, 1574–
1593.
Slagle, J. 1974. Automated theorem-proving for theories with simplifiers, commutativity,
and associativity. Journal of the ACM 21, 4, 622–642.
| 6cs.PL
|
arXiv:1609.00325v1 [math.GR] 1 Sep 2016
CONJUGACY SEARCH PROBLEM AND THE
ANDREWS–CURTIS CONJECTURE
DMITRY PANTELEEV AND ALEXANDER USHAKOV
Abstract. We develop new computational methods for studying potential counterexamples to the Andrews–Curtis conjecture, in particular,
Akbulut–Kurby examples AK(n). We devise a number of algorithms in
an attempt to disprove the most interesting counterexample AK(3). To
improve metric properties of the search space (which is a set of balanced
presentations of 1) we introduce a new transformation (called an ACMmove here) that generalizes the original Andrews-Curtis transformations
and discuss details of a practical implementation. To reduce growth of
the search space we introduce a strong equivalence relation on balanced
presentations and study the space modulo automorphisms of the underlying free group. Finally, we prove that automorphism-moves can be
applied to AK(n)-presentations. Unfortunately, despite a lot of effort
we were unable to trivialize any of AK(n)-presentations, for n > 2.
Keywords. Andrews-Curtis conjecture, Akbulut-Kurby presentations,
trivial group, conjugacy search problem, computations.
2010 Mathematics Subject Classification. 20-04, 20F05, 20E05.
1. Introduction
The Andrews–Curtis conjecture (AC-conjecture, or ACC) is a long-standing
open problem in low-dimensional topology and combinatorial group theory.
It was proposed by Andrews and Curtis in [1] while categorizing possible
counterexamples to the Poincaré conjecture. Later, Wright in [29] formulated an equivalent conjecture about 3-deformations of 2-CW-complexes associated with all finitely presented groups, thus showing that Zeeman conjecture [30] implies AC-conjecture. It is known that Zemman conjecture
also implies the Poincaré conjecture and is implied by Poincaré in some
cases, e.g. [12]. Despite recent progress and solution to Poincaré conjecture,
validity of the AC-conjecture remains open.
Although most of motivating examples come from topology, the conjecture is usually formulated in the language of combinatorial group theory, as
a question of equivalence of presentations of the trivial group. In this paper
we use the language of combinatorial group theory, omitting any topological
aspects of the problem.
Date: September 2, 2016.
The second author has been partially supported by NSF grant DMS-1318716.
1
2
D. PANTELEEV AND A. USHAKOV
1.1. Balanced presentations of the trivial group. Let X = {x1 , . . . , xn },
F = F (X) be the free group on X, and R a finite subset of F . The normal
closure of R, denoted by ncl(R), is the smallest normal subgroup in F (X)
containing R. A pair (X; R) defines a quotient group F/ ncl(R),
P denoted
by hX | Ri, and is called a presentation of hX | Ri. The sum r∈R |r| is
called the total length of the presentation (X; R) and is denoted by L(R).
We say that R ⊆ F (X) is symmetrized if R contains only cyclically reduced
words and is closed under taking inverses and cyclic permutations. Denote
by R? the minimal symmetrized set containing R (with all the words cyclically reduced). A presentation (X; R) is symmetrized if R = R? . A finite
presentation can be efficiently symmetrized and symmetrization does not
change the computational properties of the fundamental problems.
We say that a group presentation (X; R) is balanced if |X| = |R|. Some
group presentations define the trivial group 1. The “most trivial” presentation of 1 on generators {x1 , . . . , xn } is, of course, (x1 , . . . , xn ; x1 , . . . , xn )
called the canonical presentation of 1 on {x1 , . . . , xn }. Define a set Bn ⊆ Fnn
of balanced relator-tuples of the trivial group:
Bn = { (r1 , . . . , rn ) | ncl(r1 , . . . , rn ) = Fn } .
We use vector notation for tuples in Fnn . The problem of deciding if (X; R)
defines the trivial group is undecidable (see [24, 4]). It is an open problem
if the same is true for balanced presentations (see Magnus’ problem [18,
Problem 1.12]).
1.2. Transformations of group presentations. There are several types
of transformations that for a general group presentation (x1 , . . . , xn ; r1 , . . . , rn )
produce a new presentation (x1 , . . . , xn ; r10 , . . . , rn0 ) on the same set of generators of the same group. The Andrews-Curtis transformations AC1 , AC2 , AC3
(or simply AC-moves) are of that type:
(AC1 ) ri → ri rj for i 6= j,
(AC2 ) ri → ri−1 ,
(AC3 ) ri → w−1 ri w for some w ∈ Fn .
The transformations AC1 , AC2 can be recognized as Nielsen transformations
of the tuple (r1 , . . . , rn ) and the transformation AC3 is a conjugation of any
element in a tuple. Since the AC-moves are invertible, we can say that u
and v are AC-equivalent (and write u ∼AC v) if there exists a sequence of
AC-moves transforming u into v.
More generally, a transformation (named here an ACM-move) that replaces a single element ui in u with an element u0i satisfying:
0
u±1
i ∼ ui in hx1 , . . . , xn | u1 , . . . , ui−1 , ui+1 , . . . , un i
produces an isomorphic presentation. It is easy to see that AC-moves are
particular types of the ACM-move. Also, the ACM-move can be recognized
as a slightly generalized M -transformation of [9]. It is easy to see that u
ON ANDREWS–CURTIS CONJECTURE
3
can be transformed to v by AC-moves if and only if the same can done by
ACM-moves. Therefore, to check AC-equivalence one can use ACM-moves.
Yet another transformation of a group presentation that does not change
the group is an automorphism move, which is an application of ϕ ∈ Aut(Fn )
to every component of u. It is not known if the system of AC-moves with
automorphism-moves is equivalent to the system of AC-moves. More on
automorphism-moves in Section 3.
1.3. The conjecture. Denote by C the set of all tuples that can be obtained
from the canonical tuple by a sequence of AC-moves. More generally, for
u ∈ Bn denote by Cu the set of tuples in Bn AC-equivalent to u.
The Andrews-Curtis conjecture [2] states that C = Bn , i.e., every balanced presentation of the trivial group can be converted to the canonical
presentation by a sequence of AC-moves.
Despite nearly 50 years of research the conjecture is still open. It is
widely believed that the AndrewsCurtis conjecture is false with most theoretic works attempting to disprove it. A common approach is to fix some
group G, a homomorphism ϕ : Fn → G, and investigate if for any u ∈ Bn
there exists a sequence of AC-moves taking ϕ(u) into (ϕ(x1 ), . . . , ϕ(xn )).
Clearly, if the answer is negative for some choice of G and ϕ, then the original conjecture does not hold. Several classes of groups were investigated
that way, e.g., solvable groups [21], finite groups [5], the Grigorchuk group
[23], but the (negative) answer is not found.
1.4. Potential counterexamples. A big obstacle towards the solution of
the problem is that there is no algorithm to test if a particular balanced
presentation of the trivial group satisfies the conjecture, or not. There is a
number of particular balanced presentations that are not known to satisfy
the conjecture.
• Akbulut–Kurby examples: AK(n) = x, y | xyx = yxy, xn+1 = y n
for n ≥ 3.
• Miller–Schupp examples: x, y | x−1 y 2 x = y 3 , x = w , where w has
exponent sum 0 on x.
• B. H. Neumann example x, y, z | z −1 yz = y 2 , x−1 zx = z 2 , y −1 xy = x2 .
These examples are referred to as potential counterexamples to ACC. More
examples of balanced presentations of 1 (known to be AC-equivalent to the
canonical presentation) can be found in [15, 16, 8]. It was shown in [19] by
means of a computer experiment that there are no counterexamples of total
length 12 or less. Later it was shown in [13] that every balanced presentation
of total length 13 is either AC-equivalent to the canonical presentation or to
AK(3) = hx, y | x3 = y 4 , xyx = yxyi,
which makes AK(3) the shortest potential counterexample.
4
D. PANTELEEV AND A. USHAKOV
1.5. Computational approach to disproving a counterexample. To
check if a given tuple u is AC-equivalent to the canonical presentation, one
can enumerate equivalent presentations (by applying AC-moves) until the
canonical presentation is found (see [19, 13, 6]). There are several general
computational problems associated with that approach that we would like
to mention here:
• Cu is infinite and there is no terminating condition which allows an
enumeration procedure to stop with a negative answer. The enumeration procedure can only terminate with a positive answer when
it finds the canonical presentation.
• Lengths of tuples are unbounded.
• Cu has exponential growth.
To alleviate some of the problems one can bound the lengths of the words
in tuples by some constant L and do not process a tuple v which is ACequivalent to the given u if v contains a (cyclic) word of length greater
than L. This approach (used in [13, 6]) allows to use fixed memory slots
for words and makes the search space finite. Also, it is a good heuristic to
process shorter tuples first.
In this paper we consider the case n = 2 only. We use compact memory
representation for balanced pairs (u, v) ∈ B2 . We represent each letter by
a 2-bit number, thus packing 32 letters into a 64-bit machine-word. This
approach saves memory and allows to implement operations such as a cyclic
shift in just a few processor instructions, compared to a usual approach
which includes several memory writes.
1.6. Our work. In this paper we develop new efficient techniques to enhance algorithmic search in C (or Cu ). Our work is similar to previous computational investigations of ACC, but goes much further. The presentation
AK(3) is the main object of study and most of the algorithms are tested on
AK(3). Our big goal was to prove that AK(3) is not a counterexample, i.e.,
it satisfies ACC. Unfortunately, we were not able to achieve our goal. Below
we list the key features of our work.
• In Section 2 we show that ACM-moves can be used in practice. Notice that for n = 2 that requires enumerating (short!) conjugates
in a one relator group for a given element. The later problem does
not have an efficient solution as of now. It is not even known if the
conjugacy problem is decidable or not in one relator groups. Based
on techniques described in [22] we design a heuristic procedure enumerating short conjugates and discuss the details of implementation.
• We prove in Section 3 that automorphism-moves can be used with
the AC-moves for Akbulut-Kurby presentations AK(n), regardless
of whether the conjecture holds for AK(n).
• In Section 4 we introduce an equivalence relation ∼ on pairs in B2
and define normal forms for the equivalence classes. We show that
in practice equivalence of two pairs can be checked and normal forms
ON ANDREWS–CURTIS CONJECTURE
5
computed. That allows us to work with the quotient space B2 / ∼
which elements are (infinite) equivalence classes. Working in B2 / ∼
we work with large blocks of elements from B2 . Thus, we can say
that the space B2 / ∼ is much smaller than B2 , even though both
sets are infinite countable.
• In Section 5 we use heuristics to investigate if certain properties of
one-relator groups described in [8] and [16] could be the reason of
our unsuccessful search of trivialization for AK(3).
• In Section 6 we present results of our experiments.
2. ACM-move
In this section we describe our implementation of the ACM-move, i.e., an
algorithm which for a given pair u, v ∈ F = F (x, y) constructs a subset of
the set:
U = U (u, v) = u0 ∈ F (x, y) u0 ∼ u in hx, y | vi and |u0 | ≤ L ,
where L ∈ N is a fixed parameter value. Ideally, the algorithm should
construct the whole set U (u, v). The algorithm is based on weighted Xdigraphs.
2.1. Weighted X-digraphs. Formally, a weighted X-digraph is a tuple
(V, E, µ, γ) where (V, E) defines a directed graph, µ : E → X ± is the labeling
function, and γ : E → Z is the weight function.
x,k
We often use the following notation a → b for the edge with origin a,
terminus b, label x, and weight k. We say that an edge b
x,k
e−1 .
x−1 ,−k
→
a is the
inverse to e = a → b and denote it by
We say that a weighted Xdigraph Γ is:
• folded, if for every a ∈ V and x ∈ X ± there exists at most one edge
with the origin a labeled with x;
• inverse, if with every edge e the graph Γ contains e−1 ;
• rooted, if Γ comes with a designed vertex called the root.
A path p in Γ is a sequence of adjacent edges e1 . . . en , its label is µ(p) =
µ(e1 ) . . . µ(ek ) and the weight γ(p) = γ(e1 ) + . . . + γ(ek ). A circuit is a path
with the same origin and terminus.
An inverse weighted labeled digraph Γ with a root v0 and a number N ∈
N ∪ {∞} is called a pseudo conjugacy graph for u in G = hx, y | vi if the
following conditions are satisfied:
(CG1) µ(l) ∼v uγ(l) for any circuit l.
(CG2) N = ∞ or uN = 1 in G.
The simplest nontrivial example of a pseudo conjugacy graph Loop(u) for u
in hx, y | vi is shown in Fig. 1.
To implement ACM-move we generate a sufficiently large pseudo-conjugacy
graph Γ for u and then “harvest” circuits of weight 1. A large pseudoconjugacy graph Γ can be generated starting with Loop(u) and applying
6
D. PANTELEEV AND A. USHAKOV
(w1 , 0)
(w0 , 0)
v0
(wk , 1)
(wk−1 , 0)
Figure 1. A pseudo conjugacy graph Loop(u).
R-completion procedure D times. R-completion is a variation of coset enumeration first described in [28, 22] and reviewed for more precise complexity
bounds below. Harvest is shortly discussed in Section 2.4.
2.2. Operations on weighted X-digraphs. Here we shortly describe several operations with graphs used later in the sequel.
2.2.1. Vertex-identification. Given a folded X-digraph Γ and distinct v1 , v2 ∈
V (Γ) define a graph Id(Γ, v1 , v2 ) obtained from Γ as follows:
• add a new vertex v to V (Γ);
x,a
x,a
• for each edge vi → u add an edge v → u;
x,a
x,a
• for each edge u → vi add an edge u → v;
• remove v1 and v2 .
In general, the graph Id(Γ, v1 , v2 ) is not folded.
2.2.2. Weight-shift. For v ∈ V (Γ) and δ ∈ Z define a graph Shift(v, δ) obtained from Γ by changing weights of edges incident to v as follows:
x,a
• the weight a of e = v → u (where v 6= u) is increased by δ;
x,a
• the weight a of e = u → v (where v 6= u) is decreased by δ.
It is easy to see that a weight-shift preserves the weights of circuits in Γ and,
hence, preserves the property to be a pseudo-conjugacy graph. The number
of arithmetic operations required for this operation is clearly bounded by
the number of edges Ev incident on v.
x,a
2.2.3. Folding. If Γ is not folded, then it contains two edges e1 = u → v1
x,b
and e2 = u → v2 . Consider several cases.
• If v1 = v2 and a ≡ b mod N , then we remove one of the edges.
• If v1 = v2 and a 6≡ b mod N , then we replace the modulus N with
the number gcd(N, b − a) and remove one of the edges.
• If v1 6= v2 , then:
– apply Shift(v1 , a−b) (or Shift(v2 , b−a)) to achieve γ(e1 ) = γ(e2 );
– identify v1 and v2 ;
– remove e2 .
ON ANDREWS–CURTIS CONJECTURE
7
It is straightforward to check that the described operation produces a pseudoconjugacy graph from a pseudo-conjugacy graph Γ (see [28] for more detail).
Since folding e1 with e2 decreases the number of edges, a sequence of folds
eventually stops with a folded graph. The final result of folding is unique
up to shifts of weights. Denote it by Fold(Γ).
2.2.4. R-completion. Recall that a (finite) set R ⊂ F (X) is called symmetrized if R with every r ∈ R contains all cyclic permutations of r. To complete a given weighted X-digraph Γ with (symmetrized) relators R ⊂ F (X)
means to add a circuit at v labeled with r to Γ of weight 0 for every v ∈ V
and r ∈ R. It is easy to check that if Γ is a pseudo-conjugacy graph then
the result is a pseudo-conjugacy graph as well. It clearly requires linear time
(in |Γ|) to R-complete a given graph Γ. In general, the result is not folded.
2.3. Complexity of weighted X-digraph folding. Let Γ = (V, E, µ, γ)
be a weighted X-digraph. It follows from the description of folding that
Fold(Γ) as an X-digraph (i.e., if we forget about the weight function) is the
same as the result of folding of the X-digraph (V, E, µ). The only difference between weighted X-digraph folding and X-digraph folding is weightprocessing (applications of Shift(v1 , a−b) or Shift(v2 , b−a) in Section 2.2.3).
The idea is to modify the procedure and take weights into account.
2.3.1. X-digraph folding. Recall that X-digraph folding can be done in nearly
linear time O(|V | log∗ (|V |)), where log∗ is inverse-Ackermann function (see
[27]). In some sense, folding of an X-digraph Γ describes the following
equivalence relation ∼ on V (Γ):
• v1 ∼ v2 ⇔ there exists a path from v1 to v2 in Γ with µ(p) = ε;
and results into the graph Γ/ ∼. To effectively represent the sets of identified
vertices (equivalence classes) one can use compressed tree representation (as
in [26]). Each vertex v ∈ V contains a pointer p(v) to its parent, the root
points to itself. Vertex and its parent always belong to the same equivalence class, thus each tree represents an equivalence class. This presentation
allows to compare and merge two classes very efficiently, which results in
O(|V | log∗ (|V |)) complexity bound.
2.3.2. Weighted X-digraph folding. To achieve a similar complexity bound
for weighted X-digraph folding we need to take into account shifts of weight.
Since in the middle of the folding process we work with equivalence classes
(as in Section 2.3.1 above), weight-shift requires shifting weights for a whole
class. To avoid shifting weights many times with each vertex v we keep a
number δ(v) called shift value. That defines the total shift ∆(v) of a vertex
v as:
(
0,
if p(v) = v;
∆(v) =
δ(v) + ∆(v 0 ), if p(v) = v 0 6= v.
Hence, to perform weight-shift of v2 while identifying v1 and v2 (case v1 6= v2
in the Folding procedure, section 2.2.3) we set p(v2 ) = v1 and instead of
8
D. PANTELEEV AND A. USHAKOV
doing Shif t(v, c) immediately, we just set δ(v) = c. Comparison and merge
of two vertex equivalence classes can be easily extended to tree presentations
with shifts δ. Therefore, the following proposition holds.
Proposition 2.1. The number of additions performed by the folding procedure is O(|V | log∗ (|V |)).
We would like to point out that the values of the weight function can grow
exponentially fast (linearly in binary). Nevertheless, in all our experiments
we never encountered values greater than 264 .
Corollary 2.2. Let (X; R) be a symmetrized presentation. The total number of additions required to apply R-completion to a weighted graph Γ is
O (|Γ| L (R) log∗ (|Γ| L (R))).
2.4. Harvest. Here we describe a procedure that for a weighted folded Xdigraph Γ and L ∈ N finds all circuits in Γ of weight 1 and length up to
L. Since the number of such circuits is expected to grow exponentially with
L one can not expect a very efficient implementation. Nevertheless, certain
heuristics allow to speed up enumeration significantly.
For every vertex v ∈ Γ we find all reduced paths P in Γ from v of length
up to dL/2e and distribute them into bins Pu,x,ν,l :
Pu,x,γ,l = {p ∈ P | t(p) = u, γ(p) = γ, |p| = l, µ(p) ends with x}.
Then we consider pairs of “compatible” bins:
T = {(Pu1 ,x1 ,γ1 ,l1 , Pu2 ,x2 ,γ2 ,l2 ) | u1 = u2 , x1 6= x2 , γ1 −γ2 ≡N 1, 0 ≤ l1 −l2 ≤ 1}.
The set of circuits at v is constructed as:
{p1 p−1
2 | p1 ∈ P1 , p2 ∈ P2 , (P1 , P2 ) ∈ T }.
Finally, the vertex v is removed from Γ and the same process is applied to
another vertex in Γ. Note that we may skip vertices which has no adjacent
edges of a non-trivial weight. The number of operations is bounded by
(2 · |X|)dL/2e , but it was much smaller in practice.
2.5. ACM-move implementation efficiency. The implementation of the
ACM-move described above constructs a subset of the set of conjugates for
a given u in hx, y | vi of bounded length. In general, it can be a proper
subset of U (u, v). The result depends on the value of D – the number of
completion steps used to construct pseudo-conjugacy graphs. We denote it
by UD (u, v). In this section we shortly prove that:
∞
[
UD (u, v) = U (u, v),
D=1
and define the parameter δ (called depth) of a conjugate u0 ∈ U (u, v) responsible for “complexity” of u.
For a set S ⊂ R2 let ∂S be its boundary and S the closure of S in R2 . Let
D be a finite connected planar X-digraph with set of vertices V (D) and set of
ON ANDREWS–CURTIS CONJECTURE
9
edges E(D). Let C(D) be a set of cells of D which are connected and simply
connected bounded components of R2 \ D. The unbounded component of
R2 \ D is called the outer cell of D denoted by cout . An edge e ∈ E(D)
is free if it does not belong to ∂c for any c ∈ C(D). For any e ∈ E(D)
we denote its label by µ(e) ∈ X ± . The boundary of a cell c ∈ C(D)
traversed in a counterclockwise direction starting from some vertex of c
makes a closed path e1 . . . en giving the word µ(c) = µ(e1 ) . . . µ(en ) ∈ (X ± )∗
called a boundary label of c. Depending on a starting vertex we get a cyclic
permutation of the same word.
For the rest of this subsection let D be a finite connected planar X-digraph
with a base vertex v0 ∈ V (D)∩∂cout . The graph D is a van Kampen diagram
over hX | Ri if µ(c) ∈ R? for every c ∈ C(D). The boundary label µ(D) of
D is the boundary label of ∂cout read starting from v0 in a counterclockwise
direction. Note that we need also to specify the first edge to read from
v0 , that is, the starting boundary position, but it is not important for our
considerations so we omit this issue.
Now let us exclude one of the cells from C(D) and call it the inner cell
cin of D. Denote v0 by vout and pick any vertex vin ∈ V (D) ∩ ∂cin . We call
D an annular (Schupp) diagram (see [25]) over hX | Ri if µ(c) ∈ R? for any
c ∈ C(D). Its two boundary labels µin (D) = µ(cin ) and µout (D) = µ(cout )
read in a counterclockwise direction from vin and vout correspondingly, are
called the inner and outer labels of D. For any w1 , w2 ∈ (X ± )∗ we have
that w1 ∼G w2 if and only if there exists an annular diagram D over hX | Ri
with µin (D) = w1 and µout (D) = w2 .
We measure diagram complexity using a notion of depth (introduced
in [28]). For a van Kampen or annular diagram D define the dual graph
D∗ = (V ∗ , E ∗ ) as an undirected graph with V ∗ = C(D) ∪ cout (for annular
diagrams we add cin ) and E ∗ = {(c1 , c2 ) | ∂c1 ∩ ∂c2 6= ∅}. We denote the
graph distance in D∗ by d∗ .
The depth of a (generalized) van Kampen diagram D is defined by:
δ(D) = max d∗ (c, cout ).
c∈C(D)
The depth of an annular diagram D is:
δ∼ (D) = max [min (d∗ (c, cout ), d∗ (c, cin ))] .
c∈C(D)
(There is a similar notion of a diagram radii (see [10, 7]).) Define the
conjugate depth of two words w1 , w2 ∈ F (X) as:
δ∼ (w1 , w2 ) =
min
D is
an annular
diagram
{ δ(D) | µin (D) = w1 , µout (D) = w2 }
if w1 ∼G w2 and ∞ otherwise. The next theorem shows a relation between
complexity of the conjugacy search problem and the conjugacy depth.
Theorem (Theorem 3.5 in [20]). There exists an algorithm which for a
given finite symmetrized presentation hX | Ri and words w1 , w2 ∈ F (X)
10
D. PANTELEEV AND A. USHAKOV
terminates with the affirmative answer if and only if w1 ∼G w2 . Furthermore, its complexity can be bounded above by:
2δ∼ (w1 ,w2 )
Õ |w1 ||w2 |L(R)
.
For our purposes it will be useful to define another characteristic of annular diagrams, the inner conjugacy depth:
in
δ∼
(D) = max d∗ (c, cin ),
c∈C(D)
and the conjugacy depth from w1 to w2 as:
w1
δ∼
(w2 ) =
min
D is
an annular
diagram
w1
{ δ∼
(D) | µin (D) = w1 , µout (D) = w2 } ,
if w1 ∼G w2 and ∞ otherwise.
Theorem 2.3. Assume that u and u0 are conjugate in hx, y | vi and δ =
u (u0 ). If δ ≤ D (where D is the number of R-completions applied to
δ∼
Loop(u)), then our implementation of the ACM-move applied for (u, v) produces the pair (u0 , v).
Proof. Same proof as that of [22, Theorem 17.6.12].
It easily follows from Corollary 2.2 that, in general, D iterations of Rcompletion procedure require exponential time. Fortunately, in our experiments with AK(3), we observed that the value D = 2 is sufficient. Application of more than two R-completions did not produce any additional
conjugates and did not change highlighted figures in Table 1, section 6.1.
3. Nielsen automorphisms and AC-equivalence
In this section we discuss automorphism-moves, namely applications of
an automorphism ϕ ∈ Aut(F2 ):
(u, v) → (ϕ(u), ϕ(v)).
It is not known if adding these transformations to AC-moves results in an
equivalent system of transformations or not (even for presentations of 1).
Nevertheless, the following is true.
Lemma 3.1 ([9, Proposition 1(iii)] or [19, Lemma 1]). If (u, v) can be
transformed into (x, y) using AC-moves and automorphisms, then (u, v) can
be transformed into (x, y) using AC-moves only.
With any (u, v) we can associate ϕ(u,v) ∈ End(F2 ) defined by ϕ(u,v) (x) =
u and ϕ(u,v) (y) = v. That way we can treat B2 as a monoid. By Lemma
3.2, B2 naturally acts on AC-components.
Lemma 3.2. Assume that (u, v) ∼AC (u0 , v 0 ) and ϕ ∈ B2 . Then:
(ϕ(u), ϕ(v)) ∼AC (ϕ(u0 ), ϕ(v 0 )).
ON ANDREWS–CURTIS CONJECTURE
11
Proof. Clearly, it is sufficient to prove the result for the case when (u0 , v 0 )
is obtained from (u, v) by a single ACM-move, i.e., we may assume that
u ∼ u0 in hx, y | vi and v = v 0 . Hence, ϕ(u) ∼ ϕ(u0 ) in hx, y | ϕ(v)i and
(ϕ(u), ϕ(v)) ∼AC (ϕ(u0 ), ϕ(v 0 )).
Lemma 3.3 immediately implies Lemma 3.1 since (x, y) ∼AC (ϕ(x), ϕ(y))
for every ϕ ∈ Aut(F (X)).
Lemma 3.3. Let u, v ∈ F (X), ϕ ∈ End(F (X)), and (u, v) ∼AC (ϕ(u), ϕ(v)).
Then for any u0 , v 0 ∈ F (X) the following holds:
(u, v) ∼AC (u0 , v 0 ) ⇒ (u0 , v 0 ) ∼AC (ϕ(u0 ), ϕ(v 0 ))
Proof. As above, we may assume that u ∼ u0 in hx, y | vi and v = v 0 . Hence
(ϕ(u0 ), ϕ(v 0 )) ∼AC (ϕ(u), ϕ(v)) ∼AC (u, v) ∼AC (u0 , v 0 ).
With (u, v) ∈ B2 we can associate a monoid:
EndAC (u, v) = { ϕ ∈ End(F2 ) | (u, v) ∼AC (ϕ(u), ϕ(v)) } ,
under the usual composition. Lemma 3.3 implies that:
• EndAC (u, v) = EndAC (u0 , v 0 ) whenever (u, v) ∼AC (u0 , v 0 );
• Aut(F2 ) ≤ EndAC (x, y).
• EndAC (x, y) = { ϕ | ϕ(x) = u, ϕ(y) = v, (u, v) ∈ B2 } if and only if
ACC holds.
Below we prove that Aut(F2 ) ≤ EndAC (AK(n)) for every n ∈ N.
3.1. Akbulut-Kurby examples. Lemmas 3.4, 3.5, and 3.6 show that:
ϕ(AK(n)) ∼AC AK(n),
for some ϕ ∈ Aut(F2 ). Proofs were obtained using a computer program.
Below for brevity we use X and Y as x−1 and y −1 .
Lemma 3.4. ϕ(AK(n)) ∼AC AK(n) for ϕ ∈ Aut(F2 ) defined by ϕ(x) = y,
ϕ(y) = x.
Proof. The pair (ϕ(xyxY XY ), ϕ(xk Y k+1 )) = (yxyXY X, y k X k+1 ) can be
modified as follows:
∼AC2
∼ACM
∼AC3
(xyxY XY, y k X k+1 )
(xyxY XY, c−1 xk Y k+1 c) where c = xyx
(xyxY XY, xk Y k+1 ).
Appendix A provides more detail for each ACM-move used.
Lemma 3.5. ϕ(AK(n)) ∼AC AK(n) for ϕ ∈ Aut(F2 ) defined by ϕ(x) = x,
ϕ(y) = Y .
12
D. PANTELEEV AND A. USHAKOV
Proof. The pair (ϕ(xyxY XY ), ϕ(xk Y k+1 )) = (xY xyXy, xk y k+1 ) can be
modified as follows:
∼AC2
∼AC3
∼ACM
∼AC3
∼ACM
∼AC3
∼AC2
∼ACM
∼AC3
∼AC2
∼ACM
∼AC3
∼ACM
∼AC3
∼ACM
∼AC3
(xY xyXy, Y k+1 X k )
(xyXyxY, Y k+1 X k )
(xyXyxY, c−1 xyXy k+1 xY k+2 c), c = y k+1
(xyXyxY, xyXy k+1 xY k+2 )
(c−1 xyXyxY Y Xyc, xyXy k+1 xY k+2 ), c = y k+2 XyxY
(xyXyxY Y Xy, xyXy k+1 xY k+2 )
(xyXyxY Y Xy, y k+2 XY k+1 xY X)
(xyXyxY Y Xy, c−1 xyXy k XY k c), c = xyXy
(xyXyxY Y Xy, xyXy k XY k )
(Y xyyXY xY X, xyXy k XY k )
(c−1 Y xyyXY xY Xc, xyXy k XY k ), c = Y k−1 xY k−1 xY Y Xy
(Y xyyXY xY X, xyXy k XY k )
(c−1 xk Y k+1 c, xyXy k XY k ), c = Y k−1 xY k−1 xY Y Xy
(xk Y k+1 , xyXy k XY k )
(xk Y k+1 , c−1 xyxY XY c), c = Y k
(xk Y k+1 , xyxY XY )
Lemma 3.6. ϕ(AK(n)) ∼AC AK(n) for ϕ ∈ Aut(F2 ) defined by ϕ(x) = x,
ϕ(y) = yx.
Proof. The pair (ϕ(xyxY XY ), ϕ(xk Y k+1 )) = (xyxY XXY, xk−1 (Y X)k Y )
can be modified as follows:
= (xyxY XXY, xk−1 (Y X)k Y )
∼AC3 (Y xyxY XX, xk−1 (Y X)k Y )
∼AC3 (xxyXY Xy, xk−1 (Y X)k Y )
∼ACM (xxyXY Xy, c−1 xk−1 yX k−1 yXY c), c = xk−3 (Y X)k Y
∼AC3 (xxyXY Xy, xk−1 yX k−1 yXY )
∼AC2 (Y xyxY XX, xk−1 yX k−1 yXY )
∼ACM (c−1 xk Y k+1 c, xk−1 yX k−1 yXY ), c = xyX k−1 yXY Xy
∼AC3 (xk Y k+1 , xk−1 yX k−1 yXY )
∼AC2 (xk Y k+1 , yxY xk−1 Y X k−1 )
∼ACM (xk Y k+1 , c−1 xyxY XY c), c = yX k−1 yXY
∼AC3 (xk Y k+1 , xyxY XY )
Proposition 3.7. ϕ(AK(n)) ∼AC AK(n) for any n > 2 and ϕ ∈ Aut(F2 ).
Proof. Automorphisms considered in Lemmas 3.4, 3.5 and 3.6 generate Aut(F2 ).
Hence the proposition holds for any ϕ ∈ Aut(F2 ).
The next corollary implies that adding automorphism-moves to AC-moves
does not increase orbits for AK(n) presentations.
ON ANDREWS–CURTIS CONJECTURE
13
Corollary 3.8. If (u, v) ∼AC AK(3), then (ϕ(u), ϕ(v)) ∼AC AK(3) for
every ϕ ∈ Aut(F2 ).
It is natural to raise a question if a similar result holds for all balanced
presentations of 1. We performed experiments with several randomly generated Miller–Schupp presentations. The results were not always positive,
i.e., for some presentations we were unable to prove AC-equivalence to their
automorphic images.
Conjecture. It is not true that for every (u, v) ∈ B2 and ϕ ∈ Aut(F2 )
(u, v) ∼AC (ϕ(u), ϕ(v)).
Note that the conjecture above immediately implies negative answer to
ACC. In Table 2 the reader can find particular balanced presentations suspected to satisfy the conjecture above.
4. Canonical forms of presentations
For a given relation tuple u ∈ Bn the search space Cu is infinite and no
computer procedure can exhaust all of its elements. To reduce the search
space one can introduce an equivalence relation ∼ on Cu (or on Fnn ), define
efficiently computable representatives N F (u) for equivalence classes, and
study the quotient space:
Bn / ∼ = {NF(u) | u ∈ Bn }.
That way one can achieve compression of the search space as a single element
NF(u) representing its (infinite) equivalence class. Clearly, coarser relations
on Fnn give better compression.
Below we consider two equivalence relations on B2 . Same results hold for
Bn with n > 2. The first one (referred to as a cyclic relation here) was used
by Casson in a series of unpublished work (according to Bowman–McCaul)
and by Bowman–McCaul (who “followed” Casson). The second relation is
new and is significantly stronger.
4.1. Cyclic relation. Let ∼ be the transitive closure of the following pairs
in F22 :
• (u, v) ∼ (v, u),
• (u, v) ∼ (u−1 , v),
• (u, v) ∼ (u, c−1 vc),
where u, v, c are arbitrary words in F2 . We call ∼ a cyclic relation on F22 .
To define canonical representatives for the cyclic relation we do the fol−1
lowing. Fix any order on generators, say x1 < x−1
1 < x2 < x2 and denote
by < the corresponding shortlex order on F2 and, further, the corresponding
lexicographic order on F22 . Let (u, v) ∈ F22 . It is easy to see that taking
the least cyclic permutation of u±1 , the least cyclic permutation of v ±1 ,
and “sorting” the obtained words, produces the least representative of the
equivalence class of (u, v), denoted by NF(u, v). Clearly, so-defined normal
form is an efficiently computable.
14
D. PANTELEEV AND A. USHAKOV
It easily follows from the definition that NF(u, v) ∈ B2 for any (u, v) ∈ B2 .
Hence, AC-moves can be naturally defined on B2 / ∼:
ν
NF(u, v) = (u, v) 7→ NF(ν(u, v)),
where ν is an AC-move. The problem with this approach is that computing the cyclic normal form negates applications of the AC3 -move. That
can result in the component Cu being broken into subcomponents (i.e., Cu
can become disconnected). In particular, Bowman–McCaul implementation
(found here http://www.math.utexas.edu/users/sbowman/ac-bfs.tar.
gz) does not take the normal form of a pair obtained by an AC3 -move,
which completely negates the advantage of using normal forms. As we explain below, ACM-moves can solve this problem.
4.2. Cyclic relation with automorphisms. Define an equivalence relation ∼ on F22 by taking a closure of the following pairs:
•
•
•
•
(u, v) ∼ (v, u),
(u, v) ∼ (u−1 , v),
(u, v) ∼ (u, c−1 vc),
(u, v) ∼ (ϕ(u), ϕ(v)),
where u, v, c are arbitrary words in F2 and ϕ is an arbitrary automorphism in Aut(F2 ). Note that the defined relation makes (u, v) equivalent to
(ϕ(u), ϕ(v)) which, in general, is not known to be AC-equivalent to (u, v).
Hence, it is possible that an equivalence class of (u, v) contains an element
which is not AC-equivalent to (u, v). Nevertheless, the following is true.
Proposition 4.1. For every u, v, u0 , v 0 ∈ F2 if
• AK(n) ∼AC (u, v), and
• (u, v) ∼ (u0 , v 0 ),
then AK(n) ∼AC (u0 , v 0 ).
Proof. Follows from Corollary 3.8.
Proposition 4.1 allows us to replace the original component C = CAK(3)
with its factor C/ ∼ which is much smaller. The problem that taking a
normal form of a pair negates AC3 -moves is still relevant if we use the
original AC-moves. That is where ACM-moves really help. It follows from
Theorem 2.3 that choosing sufficiently large value of the parameter D we
can produce any conjugate of u in hx, y | vi.
As in Section 4.1, the normal form of the pair (u, v) is defined as the
minimal pair in its equivalence class. Below we show that normal forms can
be computed efficiently. Our main tool is the following classic result.
Theorem (Whitehead theorem, see [17, Proposition 4.20]). Let w1 , . . . , wt ,
w10 , . . . , wt0 be cyclic words in a free group F such that:
w10 = α(w1 ), . . . , wt0 = α(wt )
ON ANDREWS–CURTIS CONJECTURE
15
P 0
P 0
for some α ∈ Aut(F ). If
|wk | is minimal among all
|α (wk )| for
α0 ∈ Aut(F ), then α = τ1 . . . τn , n ≥ 0,P
where τ1 , . . . , τn arePWhitehead
automorphisms and forP
each i the
k |τ1 · · · τi (wk )| ≤
k |wk | with
Plength
0
strict inequality unless
|wk | = |wk |.
Recall [17, Section 1.4] that Whitehead automorphisms are automorphisms of two types:
• (length-preserving) automorphisms that permute the letters X ± ;
• automorphisms that for some fixed “multiplier” a ∈ X ± carry each
of the elements x ∈ X into one of x, xa, a−1 x, or a−1 xa.
There are exactly 20 Whitehead automorphisms for a free group of rank 2.
According to the Whitehead theorem if the total length of a given tuple of
cyclic words can be decreased by an application of an automorphism, then
it can be decreased by an application of a single Whitehead automorphism.
Hence, to compute the normal form of a pair (u, v) we do the following.
• First, minimize the total length |u|+|v| of (u, v) by applying 12 nonlength-preserving Whitehead automorphisms while the total length
decreases.
• Then, construct a set of all equivalent pairs of the least total length
by applying all automorphisms.
• Finally, choose the least cyclic normal form among the pairs of the
least length.
The procedure described above is efficient except, maybe, the second step,
where we construct the set of all pairs of the least total length. Currently
there are no theoretical polynomial bounds on the size of that set. Nevertheless, in our computations the maximal size observed was 112 for AK(3)
equivalent presentations with |r| ≤ 20 bound. The average size of the set of
all pairs of the least total length was 9.
5. Groups with high Dehn function
One potential challenge for computer enumeration techniques is described
by Bridson in [8] and Lishak in [16]. Both papers use a similar idea based
on properties of the following one-relator group:
(1)
x, y | y −1 x−1 yxy −1 xy = x2 ,
introduced by Baumslag in [3], satisfying the inequality:
(2)
Dehn(n) ≥ T ower2 (log2 (n)),
first observed in [11]. Lishak constructs a particular sequence of balanced
presentations parametrized by n ∈ N:
un = (r, wn y −1 ) ∈ B2 ,
where wn ∈ F (X), satisfying the following conditions.
• un is AC-equivalent to the canonical presentation.
16
D. PANTELEEV AND A. USHAKOV
• The number of steps required to obtain the canonical presentation
is super-exponential in n.
The later property comes as a consequence of the inequality (2).
Being very curious about possibility that that is the reason why our program fails to find AC-trivialization of AK(3) we tested all words obtained
in our experiments. For each word r we attempted to bound the Dehn function of the group hx, y | ri. For that purpose we used D. Holt’s package
[14] to identify automatic groups (automatic groups have at most quadratic
Dehn functions) that left us with 5356 “perhaps non-automatic” one-relator
groups. Among those 1205 we classified as Baumslag–Solitar type presentations, i.e., the presentations with the relation (un )v = um for some u, v ∈ F2 .
None of Baumslag–Solitar type presentations satisfied the condition u ∼F v,
i.e., no presentations were identified as Baumslag-type presentations (1).
Clearly, this is a heuristic approach and we can not guarantee that our list
of presentations does not contain Baumslag groups, as isomorphism problem for one-relator groups is not known to be decidable/undecidable. Also,
we were unable to classify 4151 remaining presentations. In case someone
would like to further investigate this, we have published the obtained lists
here: https://github.com/stevens-crag/ak3_types.
In light of these heuristic results it seems to be a very interesting computational problem to classify short one-relator groups hx, y | ri with |r| ≤ 20.
Find precise upper bounds on their Dehn functions.
Conjecture. Baumslag’s group x, y | y −1 x−1 yxy −1 xy = x2 has the highest Dehn function among all one-relator groups.
6. Results
The described algorithms were tested on several known potential counterexamples. Our attention was mainly focused on AK(3) and Miller–
Schupp presentations. To test performance and compare with other experimental results we also ran our programs on AK(2) and other presentations
that are known to be AC-equivalent to the canonical presentation.
As we already mentioned in Section 2.4, we set a bound L on the length
on the conjugates obtained during harvest phase. We also set a limit on the
total length of pairs to be 2L + 2. Notice that we need to do that as taking a
normal form described in Section 4.2 can increase length of one of the words
beyond L (which is allowed in our implementation). Experiments were run
on a machine with two 8-core 3.1 Ghz Intel Xeon CPU E5-2687W and 64
GB RAM.
6.1. Enumeration of AK(3)-equivalent presentations. As shown in
Section 3, automorphism-moves can be used together with ACM-moves
when applied to AK(3)-equivalent presentations. In particular, one can use
normal forms from Section 4.2 to compress the array of stored presentations.
Table 1 shows dynamics of growth of a component of CAK(3) constructed by
ON ANDREWS–CURTIS CONJECTURE
17
our program for different values of L. Each cell in Table 1 corresponds to a
value L and a value T and presents the number of pairs of the total length
equal to T constructed by the program with the single-word-bound L.
T\L
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
10
4
10
70
64
220
98
240
10
20
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
11
4
10
70
86
416
392
764
442
746
438
112
6
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
12
4
10
70
86
454
398
1382
522
1624
570
1462
42
110
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
13
4
10
70
86
458
590
2854
2004
3870
2812
4474
3400
4350
4306
710
52
0
0
0
0
0
0
0
0
0
0
0
0
0
0
14
4
10
70
86
458
590
3226
2082
8334
3714
9194
3858
11246
5384
13548
494
1652
2
0
0
0
0
0
0
0
0
0
0
0
0
15
4
10
70
86
458
590
3226
3352
16948
12288
21678
12978
22422
17930
28176
26008
30934
20430
5854
326
0
0
0
0
0
0
0
0
0
0
16
4
10
70
86
458
590
3226
3352
19666
12584
41492
15458
42550
19668
51590
27874
77162
24146
62178
3338
6314
62
0
0
0
0
0
0
0
0
17
4
10
70
86
458
590
3226
3356
19690
23174
101544
61100
102262
62874
96714
76930
123178
128556
159086
122164
151550
128556
22772
1848
0
0
0
0
0
0
18
4
10
70
86
458
590
3226
3356
19690
23174
128356
64686
236860
83902
196098
83864
230774
138478
368336
130302
353810
150518
374246
19030
51496
522
0
0
0
0
19
4
10
70
86
458
590
3226
3356
19692
23188
128380
150264
631000
375818
538380
289920
445036
355754
546680
597064
730650
538278
872784
762768
1016332
848998
209668
19332
0
0
20
4
10
70
86
458
590
3226
3356
19692
23192
128388
150276
843778
394172
1269016
364040
953378
405746
1041462
639362
1758270
585132
1519374
813708
2112918
946260
2414958
120852
270942
12062
Table 1. Each cell shows the number of pairs AC-equivalent
to AK(3) of total length T obtained by the program when run
with the length bound L. Highlighted cells do not increase
when L is increased.
It took our program 10 days to finish enumeration with the bound L = 20,
consuming 207 days of CPU time. The running time with the bound L = 21
is expected to be 60 days. We decided not to proceed beyond the value
L = 20. Memory usage during the experiments was moderate and never
exceeded 8Gb. CPU time is the main obstacle. However, we can notice that
the numbers in rows of Table 1 stabilize, at least for values T = 13, . . . , 20.
For instance, we can conjecture that the number of normal forms of AK(3)equivalent presentations of total length 20 or less is 3356 and there is no
canonical one among them.
6.2. Old non-counterexamples. We also tested our program on some
balanced presentations that were eliminated from the list of potential counterexamples before us. Our program trivializes any of them almost immediately (in less than 10 seconds on a single computational core) in less than 5
ACM-moves.
18
D. PANTELEEV AND A. USHAKOV
• AK(2): (x2 y −3 , xyx(yxy)−1 ) ∼AC (x, xyx−1 yxy −3 ) ∼AC (x, y)
• Gordon presentation (x−1 yx2 y −1 : xy 3 x−1 y −4 ) ∼AC (xyx−1 y −2 , x) ∼AC
(x, y), also considered in [6].
6.3. Miller–Schupp type presentations. We analyzed several randomly
generated Miller–Schupp presentations:
x, y | x−1 y 2 x = y 3 , x = w ,
where w has exponent sum 0 on x. We attempted to trivialize them or
show AC-equivalence with their automorphic images. Both tasks were dealt
with different success. Table 2 contains pairs (u, v) for which the program
failed to prove equivalence (u, v) ∼AC (ϕ(u), ϕ(v)) for ϕ ∈ Aut(F2 ) defined
by y → y −1 , y → yx and x → y, y → x. (In particular, we could not
trivialize the corresponding presentations.) Table 3 contains Miller-Schupp
type presentations for which the program proved automorphic equivalence
(u, v) ∼AC (ϕ(u), ϕ(v)) for any ϕ ∈ Aut(F2 ), but was not able to trivialize them. Table 4 contains trivializable Miller-Schupp presentations. The
purpose of Tables 2, 3, and 4 is to provide reference for future experiments.
(xyyyXYY,xxxyXYXY)
(xyyyXYY,xxxYXXyXXY)
(xyyyXYY,xxxyxyXXXy)
(xyyyXYY,xxxxyXyXXY)
(xyyyXYY,xxxyXyXXXy)
(xyyyXYY,xxxYxyXXXY)
(xyyyXYY,xxxxYXyXXY)
(xyyyXYY,xxxxyXXyXY)
(xyyyXYY,xxxyxYXXXY)
(xyyyXYY,xxxyXXYXXy)
(xyyyXYY,xxxYXyXy)
(xyyyXYY,xxxxYXyXXy)
(xyyyXYY,xxxxYXYXXY)
(xyyyXYY,xxxxyXXyXy)
(xyyyXYY,xxxxYXXYXY)
(xyyyXYY,xxxxYXXyXy)
(xyyyXYY,xxxyxYXXXy)
(xyyyXYY,xxxxYXYXXy)
(xyyyXYY,xxxxyXXyXy)
(xyyyXYY,xxxxyXyXXy)
(xyyyXYY,xxxYxyXXXy)
(xyyyXYY,xxxxyXXYXy)
(xyyyXYY,xxxxyXYXXY)
(xyyyXYY,xxxyXyXXXY)
(xyyyXYY,xxxxyXXYXY)
(xyyyXYY,xxxxYXXYXy)
(xyyyXYY,xxxxYXXyXY)
(xyyyXYY,xxxyxyXXXY)
(xyyyXYY,xxxxyXYXXy)
Table 2. Miller-Schupp pairs (u, v) with unknown equivalence (u, v) ∼AC (ϕ(u), ϕ(v)) for ϕ ∈ Aut(F2 ).
(xxxyXXY,xyyyyXYYY)
(xxxyXXY,xxyyyXYY)
(xxxYXXy,xyyyxYYYY)
(xxxYXXy,xxYYxyyy)
(xyyyXYY,xxxyXyXy)
(xyyyXYY,xxxyXYXy)
(xxxyXXY,xyyyXYYYY)
(xxxYXXy,xxYYYXyy)
(xyyyXYY,xxxxYXXXY)
(xyyyXYY,xxxyyXXy)
(xxxyXXY,xxyyxYYY)
(xyyyXYY,xxxyyXXY)
(xyyyXYY,xxxyXyXY)
(xxxyXXY,xyyyyxYYY)
(xyyyXYY,xxxxyXXXy)
(xyyyXYY,xxxYXyXY)
Table 3. Miller-Schupp pairs (u, v) with equivalence
(u, v) ∼AC (ϕ(u), ϕ(v)) for ϕ ∈ Aut(F2 ), but not known
if trivializable.
(XyyxYYY,xxYYYXYxYXYY)
(XyyxYYY,xxYXyXyyxyy)
(XyyxYYY,xyxYXyy)
(XyyxYYY,xyxYYXyyyyyy)
(XyyxYYY,xxyyyXYYXyxY)
(XyyxYYY,xxyyyXYxYXYY)
(XyyxYYY,xxYYXYXYXyy)
(XyyxYYY,xyXyyxYXyyxY)
(XyyxYYY,xxYXyxyyyXY)
(XyyxYYY,xyxYYYYYYXy)
(XyyxYYY,xyyxYYYXYYY)
(XyyxYYY,xyyyXyyXyy)
Table 4. Trivializable Miller-Schupp presentations.
ON ANDREWS–CURTIS CONJECTURE
19
7. Conclusion
Despite a lot of effort, we were unable to disprove any new Akbulut–
Kurby type presentations. In fact, the numbers in Table 1 rows stabilize as
the value of the parameter L increases, suggesting that the AC-equivalence
class of AK(3) does not contain the canonical presentation, thus supporting
a common opinion that ACC does not hold.
Appendix A. Used ACM-moves justification
In this section we prove the one-relator groups identities used in lemmas 3.4, 3.5 and 3.6 for the ACM moves. Every proof demonstrates that
c−1 uc(u0 )−1 = 1 in hx, y | vi.
A.1. Used in lemma 3.4.
c = xyx, u = xk Y k+1 , v = xyxY XY , u0 = y k X K+1 :
XY X · xk (Y k+1 · xyx · xk+1 )Y k
=XY X(xk · xyx · Y K )
(Y xyx = xy)
=XY Xxyx (xyxY = yx)
=1
A.2. Used in lemma 3.5.
u = xyXy k+1 xY k+2 , v = xyXyxY , c = y k+1 , u0 = Y k+1 X k :
(Y k+1 ) · xyXy k+1 x(Y k+2 · y k+1 ) · xk (y k+1 ) = (xk+1 yXy k+1 )x · Y
=(xk yXy k ) · xY = . . . = yXxY = 1
u = xyXyxY Y Xy, v = xyXy k+1 xY k+2 , c = y k+2 XyxY , u0 = xyXyxY :
yXY xY k+2 · xyXyxY Y X(y · y k+2 XyxY · yXY xY )X
=yXY (xY k+2 xyXy)xY Y X · y k+2 X = yXY · Y k · xY (Y Xy k+2 X)
=(yX)Y k+1 (xY · Xy k+1 ) = Y k+1 · y k+1 = 1
u = xyXy k XY k , v = xyXyxY Y Xy, c = xyXy, u0 = y k+2 XY k+1 xY X:
(Y xY X) · (xyXy k )XY k · xyXy · xyXy k+1 x(Y k+2 )
=XY k xy(XyxyXy k+1 )x · Y 3 = X(Y k xy · y)yXy k · xY 3
=X · Y k−1 xy(Xyx · yXy k )xY 3 = X(Y k−1 xy · y)yXy k−1 · xY 3
=Xxy · yyX · xY 3 = 1
20
D. PANTELEEV AND A. USHAKOV
u = xk Y k+1 , v = xyXy k XY k , c = Y k−1 xY k−1 xY Y Xy,
u0 = Y xyyXY xY X:
(Y xyyX)y k−1 Xy k−1 · xk Y k+1 · Y k−1 xY k−1 xY Y Xy
·xyXy(xY Y Xy) = (y k−1 X)y k−1 xk Y 2k xY k−1 xY Y Xy(xyXy)
=(y k · y k−1 )xk Y 2k xY k−1 xY Y X(y)
=y k x(Xy k · xk )Y 2k xY k−1 xY Y X
(xyXy k X = y k )
(shift)
=y k x · y(Xy k · xk−1 )Y 2k xY k−1 xY Y X
...
(Xy k x = yXy k )
=y k x · y k X(y k · Y 2k )xY k−1 xY Y X
=y k x(y k X · Y k · x)Y k−1 xY Y X = (y k x) · xY · Y k−1 xY (Y X)
=(Xy k · x)Y k xY = (Xy k · x)Y k (xY ) = y k · Y k = 1
u = xyxY XY , v = xk Y k+1 , c = Y k , u0 = y k xY k xY X:
(y k · xy)xY XY · Y k · xyX(y k XY k ) = xY X(Y Y k )xyX · y k+1 =
xY (X · X k x)yXy k+1 = x(Y · Y k+1 · y)Xy k+1 = . . . = Y k+1 y k+1
A.3. Used in lemma 3.6.
u = xk−1 yX k−1 yXY , v = xxyXY Xy, c = xk−3 (Y X)k Y ,
u0 = xk−1 (Y X)k Y :
y(xy)k (X k−3 · xk−1 )yX k−1 yXY · (xk−3 (Y X)k Y · y(xy)k X k−1 )
=y(xy)k · x2 yX k−1 (yXY · X)X = (yx)k yx2 yX k−1 · XX(Y · X)
= . . . = yx2 yX k−1 XX · xk Y X = yx2 yXY X = 1
u = xk Y k+1 , v = xk−1 yX k−1 yXY , c = xyX k−1 yXY Xy,
u0 = Y xyxY XX:
(Y xyxY )xk−1 Y (X · xk )Y k+1 · xyX k−1 yXY Xy · xx(yXY Xy)
=(xk−1 Y xk−1 Y k+1 ) · xyX k−1 yXY Xy(xx) = . . .
=Y (xk−1 · xyX k−1 yXY )Xy = Y · x · Xy = 1
u = xyxY XY , v = xk Y k+1 , c = yX k−1 yXY ,
u0 = yxY xk−1 Y X k−1 :
(yxY xk−1 Y ) · xyxY X(Y · y)X k−1 yXY · xk−1 (yX k−1 yXY )
=(x)yxY X k yX(Y xk−1 ) = (y)xY X k yX · (y k )
=(xY )X k y(X · xk ) = X k y · y k = 1
ON ANDREWS–CURTIS CONJECTURE
21
References
[1] J. J. Andrews and M. L. Curtis. Free groups and handlebodies. Proceedings of the
American Mathematical Society, 16(2):192–195, 1965.
[2] J. J. Andrews and M. L. Curtis. Extended Nielsen operations in free groups. Amer.
Math. Month., 73:21–28, 1966.
[3] G. Baumslag. A non-cyclic one-relator group all of whose finite factor groups are
cyclic. J. Australian Math. Soc., 10:497–498, 1969.
[4] W. Boone. The word problem. Proc. Natl. Acad. Sci., 44:1061–1065, 1958.
[5] A. Borovik, A. Lubotzky, and A. Myasnikov. The finitary andrews-curtis conjecture.
In Infinite Groups: Geometric, Combinatorial and Dynamical Aspects, volume 248 of
Progress in Mathematics, pages 15–30. Birkhuser Basel, 2005.
[6] R. Bowman and S. McCaul. Fast searching for andrews-curtis trivializations. Experiment. Math., 15(2):193–198, 2006.
[7] N. Brady, T. Riley, and H. Short. The geometry of the word problem for finitely
generated groups. Advanced Courses in Mathematics CRM Barcelona. Birkhauser,
2007.
[8] M. Bridson. The complexity of balanced presentations and the andrews-curtis conjecture. arXiv preprint arXiv:1504.04187, 2015.
[9] R. Burns and O. Macedonska. Balanced presentations of the trivial group. Bull. London Math. Soc., 25:514–526, 1992.
[10] S. Gersten and T. Riley. Filling length in finitely presentable groups. Geometriae
Dedicata, 92:41–58, 2002.
[11] S. M. Gersten. Dehn functions and l1-norms of finite presentations. In Algorithms and
Classification in Combinatorial Group Theory, pages 195–225, Berlin, 1992. Springer.
[12] D. Gillman and D. Rolfsen. The Zeeman conjecture for standard spines is equivalent
to the Poincaré conjecture. Topology, 22(3):315–323, 1983.
[13] G. Havas and C. Ramsay. Breadth-first search and the Andrews–Curtis conjecture.
Int. J. Algebra Comput., 13(01):61–68, 2003.
[14] D. Holt. kbmag2 (version 2.4). Downloaded from http://homepages.warwick.ac.
uk/~mareg/download/kbmag2/, December 2000.
[15] S. Ivanov. On balanced presentations of the trivial group. Invent. Math., 165:525–549,
2006.
[16] B. Lishak. Balanced finite presentations of the trivial group. Preprint. Available at
http://arxiv.org/abs/1504.00418, 2015.
[17] R. Lyndon and P. Schupp. Combinatorial group theory. Springer, 2015.
[18] V. Mazurov and E. Khukhro. Unsolved Problems in Group Theory. The Kourovka
Notebook. No. 18. Available at http://arxiv.org/abs/1401.0300, 2014.
[19] A. D. Miasnikov. Genetic algorithms and the andrews-curtis conjecture. International
Journal of Algebra and Computation, 9:671–686, 1999.
[20] P. Morar and A. Ushakov. Search problems in groups and branching processes. Int.
J. Algebra Comput., 25:445–480, 2015.
[21] A. G. Myasnikov. Extended nielsen transformations and the trivial group. Mat. Zametki, 35:491–495, 1984.
[22] A. G. Myasnikov, V. Shpilrain, and A. Ushakov. Non-commutative Cryptography
and Complexity of Group-theoretic Problems, chapter 17. Mathematical surveys and
monographs. American Mathematical Society, 2011.
[23] A. Myropolska. Andrews-Curtis and Nielsen equivalence relations on some infinite
groups. Preprint. Available at http://arxiv.org/abs/1304.2668, 2013.
[24] P. Novikov. On the algorithmic unsolvability of the word problem in group theory.
Proc. Steklov Inst., 44:1–143, 1955.
[25] P. E. Schupp. On Dehn’s algorithm and the conjugacy problem. Mathematische Annalen, 178:119–130, 1968.
22
D. PANTELEEV AND A. USHAKOV
[26] R. Tarjan and J. van Leeuwen. Worst-case analysis of set union algorithms. J. ACM,
31(2):245–281, March 1984.
[27] N. Touikan. A fast algorithm for stallings folding process. International Journal of
Algebra and Computation, 16(06):1031–1045, 2006.
[28] A. Ushakov. Fundamental search problems in groups. PhD thesis, CUNY/Graduate
Center, 2005.
[29] P. Wright. Group presentations and formal deformations. Transactions of the American Mathematical Society, 208:161–169, 1975.
[30] E. C. Zeeman. On the dunce hat. Topology, 2(4):341–358, 1963.
Department of Mathematics, Stevens Institute of Technology, Hoboken,
NJ, USA
E-mail address: dpantel1,[email protected]
| 4math.GR
|
THE NORMALIZED VOLUME OF A SINGULARITY IS LOWER
SEMICONTINUOUS
arXiv:1802.09658v1 [math.AG] 27 Feb 2018
HAROLD BLUM AND YUCHEN LIU
Abstract. We show that in any Q-Gorenstein flat family of klt singularities, normalized volumes are lower semicontinuous with respect to the Zariski topology. A quick
consequence is that smooth points have the largest normalized volume among all klt
singularities. Using an alternative characterization of K-semistability developed by Li,
Liu and Xu, we show that K-semistability is a very generic or empty condition in any
Q-Gorenstein flat family of log Fano pairs.
1. Introduction
Given an n-dimensional complex klt singularity (x ∈ (X, D)), Chi Li [Li15a] introduced the normalized volume function on the space Valx,X of real valuations of C(X)
centered at x. More precisely, for any such valuation v, its normalized volume is defined
c x,(X,D) (v) := AX,D (v)n vol(v), where AX,D (v) is the log discrepancy of v with respect
as vol
to (X, D) according to [JM12, BdFFU15], and vol(v) is the volume of v according to
[ELS03]. Then we can define the normalized volume of a klt singularity (x ∈ (X, D)) by
c x,(X,D) (v)
c X, D) := min vol
vol(x,
v∈Valx,X
c was shown recently in [Blu16]. We also denote
where the existence of minimizer of vol
c X) := vol(x,
c X, 0).
vol(x,
The normalized volume of a klt singularity x ∈ (X, D) carries interesting information of
c X, D) ≤
its geometry and topology. It was shown by the second author and Xu that vol(x,
n
n and equality holds if and only if (x ∈ X \ Supp(D)) is smooth (see [LiuX17, Theorem
A.4] or Theorem 33). By [Xu14] the local algebraic fundamental group π̂1loc (X, x) of a
klt singularity x ∈ X is always finite. Moreover, assuming the conjectural finite degree
formula of normalized volumes [LiuX17, Conjecture 4.1], the size of π̂1loc (X, x) is bounded
c X) (see Remark 37). If X is a Gromov-Hausdorff limit of Kählerfrom above by nn /vol(x,
c X) = nn · Θ(x, X)
Einstein Fano manifolds, then Li and Xu [LX17] showed that vol(x,
where Θ(x, X) is the volume density of a closed point x ∈ X (see [HS16, SS17] for
background materials).
In this article, it is shown that the normalized volume of a singularity is lower semicontinuous in families.
Theorem 1. Let π : (X , D) → T together with a section σ : T → X be a Q-Gorenstein
flat family of complex klt singularities over a normal variety T . Then the function t 7→
c
vol(σ(t),
Xt , Dt ) on T (C) is lower semicontinuous with respect to the Zariski topology.
Date: February 28, 2018.
1
2
HAROLD BLUM AND YUCHEN LIU
One quick consequence of Theorem 1 is that smooth points have the largest normalized
volumes among all klt singularities (see Theorem 33 or [LiuX17, Theorem A.4]). Another
natural consequence is that if X is a Gromov-Hausdorff limit of Kähler-Einstein Fano
manifolds, then the volume density function x 7→ Θ(x, X) on X(C) is lower semicontinuous in the Zariski topology, which is stronger than being lower semicontinuous in the
Euclidean topology mentioned in [SS17] (see Corollary 35).
We also state the following natural conjecture on constructibility of normalized volumes
of klt singularities (see also [Xu17, Conjecture 4.11]).
Conjecture 2. Let π : (X , D) → T together with a section σ : T → X be a Q-Gorenstein
flat family of complex klt singularities over a normal variety T . Then the function t 7→
c
vol(σ(t),
Xt , Dt ) on T (C) is constructible.
Verifying the Zariski openness of K-semistability is an important step in the construction of an algebraic moduli space of K-polystable Q-Fano varieties. In a smooth family
of Fano manifolds, Odaka [Oda13] and Donaldson [Don15] showed that the locus of fibers
admitting Kähler-Einstein metrics (or equivalently, being K-polystable) with discrete automorphism groups is Zariski open. This was generalized by Li, Wang and Xu [LWX14]
where they proved the Zariski openness of K-semistability in a Q-Gorenstein flat families of smoothable Q-Fano varieties in their construction of the proper moduli space of
smoothable K-polystable Q-Fano varieties (see [SSY16, Oda15] for related results). A
common feature is that analytic methods were used essentially in proving these results.
Using the alternative characterization of K-semistability by the affine cone construction
developed by Li, the second author, and Xu in [Li15b, LL16, LX16], we apply Theorem
1 to prove the following result on weak openness of K-semistability. Unlike the results
described in the previous paragraph, our result is proved using purely algebraic methods and hence can be applied to Q-Fano families with non-smoothable fibers (or more
generally, families of log Fano pairs).
Theorem 3. Let ϕ : (Y, E) → T be a Q-Gorenstein flat family of complex log Fano pairs
over a normal base T . If (Yo, Eo ) is log K-semistable for some closed point o ∈ T , then
the following statements hold:
(1) There exists an intersection U of countably many Zariski open neighborhoods of
o, such that (Yt , Et ) is log K-semistable for any closed point t ∈ U. In particular,
(Yt , Et ) is log K-semistable for a very general closed point t ∈ T .
(2) Denote by η the generic point of T , then the geometric generic fiber (Yη̄ , Eη̄ ) is log
K-semistable.
(3) Assume Conjecture 2 is true, then such U from (1) can be chosen as a genuine
Zariski open neighborhood of o.
The following corollary generalizes [Li17, Theorem 4] and follows easily from Theorem
3.
Corollary 4. Suppose a complex log Fano pair (Y, E) specially degenerates to a log Ksemistable log Fano pair (Y0 , E0 ), then (Y, E) is also log K-semistable.
Our strategy to prove Theorem 1 is to study invariants of ideals instead of invariants of
valuations. From Liu’s characterization of normalized volume by normalized multiplicities
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
3
of ideals (see [Liu16, Theorem 27] or Theorem 5), we know
c
vol(σ(t),
Xt , Dt ) = inf lct(Xt , Dt ; a)n · e(a)
a
where the infimum is taken over all ideals a ⊂ OXt cosupported at σ(t). These ideals
are parametrized by a relative Hilbert scheme of X /T with countably many components.
Clearly a 7→ lct(Xt , Dt ; a) is lower semicontinuous on the Hilbert scheme, but a 7→ e(a)
may only be upper semicontinuous. Thus, it is unclear what semicontinuity properties
a 7→ lct(a)n · e(a) may have.
To fix this issue, we introduce the normalized colength of singularities ℓc
c,k (σ(t), Xt , Dt )
n
by taking the infimum of lct(Xt , Dt ; a) · ℓ(Oσ(t),Xt /a) for ideals a satisfying a ⊃ mkσ(t)
and ℓ(Oσ(t),Xt /a) ≥ ck n . The normalized colength function behaves better in families since the colength function a 7→ ℓ(Oσ(t),Xt /a) is always locally constant in the
Hilbert scheme, so a 7→ lct(Xt , Dt ; a)n · ℓ(Oσ(t),Xt /a) is constructibly lower semicontinuous on the Hilbert scheme. Thus, the properness of Hilbert schemes implies that
t 7→ ℓc
c,k (σ(t), Xt , Dt ) is constructibly lower semicontinuous on T . Then we prove a key
equality between the asymptotic normalized colength ℓd
c,∞ (σ(t), Xt , Dt ) and the normalized
c
volume vol(σ(t), Xt , Dt ) when c is small (see Theorem 12) using local Newton-Okounkov
bodies following [Cut13, KK14] (see Lemma 13) and convex geometry (see Appendix A).
Then by establishing a uniform approximation of volumes by colengths (see Theorem
16) and generalizing Li’s Izumi and properness estimates [Li15a] to families (see Theorems 21 and 22), we show that the normalized colengh functions uniformly approximate
the normalized volume function from above (see Theorem 27). Putting these ingredients
together, we get the proof of Theorem 1.
This paper is organized as follows. In Section 2, we give the preliminaries including
notations, normalized volumes of singularities, and Q-Gorenstein flat families of klt pairs.
In Section 3.1, we introduce the concept of normalized colengths of singularities. We
show in Theorem 12 that the normalized volume of a klt singularity is the same as its
asymptotic normalized colength. The proof of Theorem 12 uses a comparison of colengths
and multiplicities established in Lemma 13. In Section 3.2, we study the normalized
volumes and normalized colength after algebraically closed field extensions. In Section 4,
we establish a uniform approximation of volume of a valuation by colengths of its valuation
ideals. In Section 5, we generalize Li’s Izumi and properness estimates to families. The
results from Sections 4 and 5 enable us to prove the uniform approximation of normalized
volumes by normalized colengths from above in families (see Section 6.1). The proofs of
main theorems are presented in Section 6.2. We give applications of our main theorems
in Section 6.3. Theorem 33 generalizes the inequality part of [LiuX17, Theorem A.4]. We
show that the volume density function on a Gromov-Hausdorff limit of Kähler-Einstein
manifolds is lower semicontinuous in the Zariski topology (see Corollary 35). We give
an effective upper bound on the degree of finite quasi-étale maps over klt singularities
on Gromov-Hausdorff limits of Kähler-Einstein Fano manifolds (see Theorem 36). In
Appendix A we provide certain convex geometric results on lattice points counting that
are needed in proving Lemma 13. In Appendix B, we provide results on constructbility
of Hilbert–Samuel funtions that are needed in proving uniform approximation results in
Section 4.
4
HAROLD BLUM AND YUCHEN LIU
Acknowledgements. The first author would like to thank his advisor Mircea Mustaţă
for numerous useful discussions and his constant support. In addition, he would like to
thank Mattias Jonsson, Ilya Smirnov, and Tommaso de Fernex for many useful conversations. The second author would like to thank Chi Li and Chenyang Xu for fruitful
discussions.
We wish to thank János Kollár, Linquan Ma, Sam Payne, Xiaowei Wang, and Ziquan
Zhuang for helpful comments. We are also grateful to Ruixiang Zhang for his help on
Proposition 39.
2. Preliminaries
2.1. Notations. In this paper, all varieties are assumed to be irreducible, reduced, and
defined over a (not necessarily algebraically closed) field k of characteristic 0. For a variety
T over k, we denote the residue field of any scheme-theoretic point t ∈ T by κ(t). Given
a morphism π : X → T between varieties over k, we write Xt := X ×T Spec(κ(t)) for the
scheme theoretic fiber over t ∈ T . We also denote the geometric fiber of π over t ∈ T by
Xt := X ×T Spec(κ(t)). Suppose X is a variety over k and x ∈ X is a k-rational point.
Then for any field extension K/k, we denote (xK , XK ) := (x, X) ×Spec(k) Spec(K).
Let X be a normal variety over k and D be an effective Q-divisor on X. We say that
(X, D) is a Kawamata log terminal (klt) pair if (KX +D) is Q-Cartier and KY −f ∗ (KX +D)
has coefficients > −1 on some log resolution f : Y → (X, D). A klt pair (X, D) is called
a log Fano pair if in addition X is proper and −(KX + D) is ample. A klt pair (X, D)
together with a closed point x ∈ X is called a klt singularity (x ∈ (X, D)).
Let (X, D) be a klt pair. For an ideal sheaf a on X, we define the log canonical threshold
of a with respect to (X, D) by
lct(X, D; a) := inf
E
AX,D (ordE )
,
ordE (a)
where the infimum is taken over all prime divisors E on a log resolution f : Y → (X, D).
We will often use the notation lct(a) to abbreviate lct(X, D; a) once the klt pair (X, D) is
specified. If a is co-supported at a single closed point x ∈ X, we define the Hilbert–Samuel
multiplicity of a as
ℓ(Ox,X /am )
e(a) := lim
m→∞
mn /n!
where n := dim(X) and ℓ(Ox,X /am ) denotes the length of Ox,X /am as an Ox,X -module.
2.2. Valuations. Let X be a variety defined over a field k and x ∈ X closed point. By a
valuation of the function field K(X), we mean a valuation v : K(X)× → R that is trivial
on k. By convention, we set v(0) := +∞. Such a valuation v has center x if v is ≥ 0 on
Ox,X and > 0 on the maximal ideal of Ox,X . We write Valx,X for the set of valuations of
K(X) with center x.
To any valuation v ∈ Valx,X and m ∈ Z>0 there is an associated valuation ideal defined
locally by am (v) := {f ∈ OX | v(f ) ≥ m}. Note that am (v) is mx -primary for each
m ∈ Z>0 . For an ideal a ⊂ OX and v ∈ Valx,X , we set
v(a) := min{v(f ) | f ∈ a · Ox,X } ∈ [0, +∞].
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
5
2.3. Normalized volumes of singularities. Let k be an algebraically closed field of
characteristic 0. For an n-dimensional klt singularity x ∈ (X, D) over k, C. Li [Li15a]
c x,(X,D) : Valx,X → R>0 ∪ {+∞}. Recall
introduced the normalized volume function vol
that for v ∈ Valx,X ,
(
n
c x,(X,D) (v) := AX,D (v) · vol(v) if AX,D (v) < +∞ ,
vol
+∞
if AX,D (v) = +∞
where AX,D (v) and vol(v) denote the log discrepancy and volume of v. As defined in
[ELS03], the volume of v is given by
vol(v) := lim sup
m→∞
ℓ(Ox,X /am (v))
.
mn /n!
By [ELS03, Mus02, LM09, Cut13],
e(am (v))
.
m→∞
mn
The log discrepancy of v, denoted AX,D (v), is defined in [JM12, BdFFU15] (and [LL16]
for the case of klt pairs).
The normalized volume of the singularity x ∈ (X, D) is given by
vol(v) = lim
c X, D) :=
vol(x,
inf
v∈Valx,X
c x,(X,D) (v).
vol
When k is uncountable, the above infimum is a minimum [Blu16].
The following characterization of normalized volumes using log canonical thresholds
and multiplicities of ideals is crucial in our study. Note that the right hand side of (2.1)
was studied by de Fernex, Ein and Mustaţă [dFEM04] when x ∈ X is smooth and D = 0.
Theorem 5 ([Liu16, Theorem 27]). With the above notation, we have
(2.1)
c X, D) =
vol(x,
inf
a : mx -primary
lct(X, D; a)n · e(a).
The following theorem provides an alternative characterization of K-semistability using
the affine cone construction. Here we state the most general form, and special cases can
be found in [Li15b, LL16].
Theorem 6 ([LX16, Proposition 4.6]). Let (Y, E) be a log Fano pair of dimension (n − 1)
over an algebraically closed field k of characteristic 0. For r ∈ N satisfying L := −r(KY +
E) is Cartier, the affine cone X = C(Y, L) is defined by X := Spec ⊕m≥0 H 0 (Y, L⊗m ). Let
D be the Q-divisor on X corresponding to E. Denote by x the cone vertex of X. Then
c X, D) ≤ r −1 (−KY − E)n−1 ,
vol(x,
and the equality holds if and only if (Y, E) is log K-semistable.
2.4. Q-Gorenstein flat families of klt pairs. In this section, the field k is not assumed
to be algebraically closed.
Definition 7.
6
HAROLD BLUM AND YUCHEN LIU
(a) Given a normal variety T , a Q-Gorenstein flat family of klt pairs over T consists of
a surjective flat morphism π : X → T from a variety X , and an effective Q-divisor
D on X avoiding codimension 1 singular points of X , such that the following
conditions hold:
• All fibers Xt are connected, normal and not contained in Supp(D);
• KX /T + D is Q-Cartier;
• (Xt , Dt ) is a klt pair for any t ∈ T .
(b) A Q-Gorenstein flat family of klt pairs π : (X , D) → T together with a section
σ : T → X is called a Q-Gorenstein flat family of klt singularities. We denote by
σ(t) the unique closed point of Xt lying over σ(t) ∈ Xt .
Proposition 8. Let π : (X , D) → T be a Q-Gorenstein flat family of klt pairs over a
normal variety T . The following hold.
(1) There exists a closed subset Z of X of codimension at least 2, such that Zt has
codimension at least 2 in Xt for every t ∈ T , and π : X \ Z → T is a smooth
morphism.
(2) X is normal.
(3) For any morphism f : T ′ → T from a normal variety T ′ to T , the base change
πT ′ : (XT ′ , DT ′ ) = (X , D) ×T T ′ → T ′ is a Q-Gorenstein flat family of klt pairs
over T ′ , and KXT ′ /T ′ + DT ′ = g ∗ (KX /T + D) where g : XT ′ → X is the base change
of f .
Proof. (1) Assume π is of relative dimension n. Let Z := {x ∈ X | dimκ(x) ΩX /T ⊗ κ(x) >
n}. It is clear that Z is Zariski closed. Since k is of characteristic 0, Zt = Z ∩ Xt is the
singular locus of Xt . Hence codimXt Zt ≥ 2 because Xt is normal.
(2) From (1) we know that Z is of codimension at least 2 in X , and X \ Z is smooth
over T . Thus X \ (Z ∪ π −1 (Tsing )) is regular, and Z ∪ π −1 (Tsing ) has codimension at least
2 in X . So X satisfies property (R1 ). Since π is flat, for any point x ∈ Xt we have
depth(Ox,X ) = depth(Ox,Xt ) + depth(Ot,T ) by [Mat80, (21.C) Corollary 1]. Hence it is
easy to see that X satisfies property (S2 ) since both Xt and T are normal. Hence X is
normal.
(3) Let ZT ′ := Z ×T T ′ , and note that XT ′ \ ZT ′ is smooth over T ′ . Since the fibers of
πT ′ and T ′ are irreducible, XT ′ is also irreducible. Thus the same argument of (2) implies
that XT ′ satisfies both (R1 ) and (S2 ), which means XT ′ is normal. Since π|X \Z is smooth,
we know that KXT ′ /T ′ + DT ′ and g ∗ (KX /T + D) are Q-linearly equivalent after restriting
to XT ′ \ ZT ′ . Since ZT ′ is of codimension at least 2 in XT ′ , the Q-linear equivalence over
XT ′ \ ZT ′ extends to XT ′ . Thus we finish the proof.
Definition 9.
(a) Let Y be a normal projective variety. Let E be an effective Qdivisor on Y . We say that (Y, E) is a log Fano pair if (Y, E) is a klt pair and
−(KY + E) is Q-Cartier and ample. We say Y is a Q-Fano variety if (Y, 0) is a
log Fano pair.
(b) Let T be a normal variety. A Q-Gorenstein family of klt pairs ϕ : (Y, E) → T is
called a Q-Gorenstein flat family of log Fano pairs if ϕ is proper and −(KY/T + E)
is ϕ-ample.
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
7
The following proposition states a well known result on the behaviour of the log canonical threshold in families. See [Amb16, Corollary 1.10] for a similar statement. The proof
is omitted because it follows from arguments similar to those in [Amb16].
Proposition 10. Let π : (X , D) → T be a Q-Gorenstein flat family of klt pairs over a
normal variety T . Let a be an ideal sheaf of X . Then
(1) The function t 7→ lct(Xt , Dt ; at ) on T is constructible;
(2) If in addition V (a) is proper over T , then the function t 7→ lct(Xt , Dt ; at ) is lower
semicontinuous with respect to the Zariski topology on T .
3. Comparison of normalized volumes and normalized colengths
3.1. Normalized colengths of klt singularities.
Definition 11. Let x ∈ (X, D) be a klt singularity over an algebraically closed field k of
characteristic 0. Denote its local ring by (R, m) := (Ox,X , mx ).
(a) Given constants c ∈ R>0 and k ∈ N, we define the normalized colength of x ∈
(X, D) with respect to c, k as
lct(a)n · ℓ(R/a).
ℓc
inf
c,k (x, X, D) := n! ·
mk ⊂a⊂m
ℓ(R/a)≥ck n
Note that the assumption mk ⊂ a ⊂ m implies a is an m-primary ideal.
(b) Given a constant c ∈ R>0 , we define the asymptotic normalized colength function
of x ∈ (X, D) with respect to c as
c
ℓd
c,∞ (x, X, D) := lim inf ℓc,k (x, X, D).
k→∞
It is clear that ℓc
c,k is an increasing function in c. The main result in this section is the
following theorem.
Theorem 12. For any klt singularity x ∈ (X, D) over an algebraically closed field
characteristic 0, there exists c0 = c0 (x, X, D) > 0 such that
c
(3.1)
ℓd
whenever 0 < c ≤ c0 .
c,∞ (x, X, D) = vol(x, X, D)
k of
Proof. We first show the “≤” direction. Let us take a sequence of valuations {vi }i∈N such
c i ) = vol(x,
c X, D). We may rescale vi so that vi (m) = 1 for any i. Since
that limi→∞ vol(v
c i )}i∈N are bounded from above, by [Li15a, Theorem 1.1] we know that there exists
{vol(v
C1 > 0 such that AX,D (vi ) ≤ C1 for any i ∈ N. Then by Li’s Izumi type inequality [Li15a,
Theorem 3.1], there exists C2 > 0 such that ordm (f ) ≤ vi (f ) ≤ C2 ordm (f ) for any i ∈ N
and any f ∈ R. As a result, we have mk ⊂ ak (vi ) ⊂ m⌈k/C2 ⌉ for any i, k ∈ N. Thus
e(m) n
e(m)
ℓ(R/ak (vi )) ≥ ℓ(R/m⌈k/C2 ⌉ ) ∼ n!C
n k . Let us take c0 = 2n!C n , then for k ≫ 1 we have
2
2
ℓ(R/ak (vi )) ≥ c0 k n for any i ∈ N. Therefore, for any i ∈ N we have
n
n
c
ℓ[
c0 ,∞ (x, X, D) ≤ n! lim inf lct(ak (vi )) ℓ(R/ak (vi )) = lct(a• (vi )) vol(vi ) ≤ vol(vi ).
k→∞
In the last inequality we use lct(a• (vi )) ≤ AX,D (vi ) as in the proof of [Liu16, Theorem
c
c
27]. Thus ℓ[
c0 ,∞ (x, X, D) ≤ limi→∞ vol(vi ) = vol(x, X, D). This finishes the proof of the
“≤” direction.
8
HAROLD BLUM AND YUCHEN LIU
c
For the “≥” direction, we will show that ℓd
c,∞ (x, X, D) ≥ vol(x, X, D) for any c > 0.
By a logarithmic version of the Izumi type estimate [Li15a, Theorem 3.1], there exists
a constant c1 = c1 (x, X, D) > 0 such that v(f ) ≤ c1 AX,D (v)ordm (f ) for any valuation
v ∈ Valx,X and any function f ∈ R. For any m-primary ideal a, there exists a divisorial
valuation v0 ∈ Valx,X computing lct(a) by [Liu16, Lemma 26]. Hence we have the following
Skoda type estimate:
lct(a) =
AX,D (v0 )
AX,D (v0 )
1
≥
=
.
v0 (a)
c1 AX,D (v0 )ordm (a)
c1 ordm (a)
Let 0 < δ < 1 be a positive number. If a 6⊂ m⌈δk⌉ and ℓ(R/a) ≥ ck n , then
lct(a)n · ℓ(R/a) ≥
ck n
c
≥ n n.
n
n
c1 (⌈δk⌉ − 1)
c1 δ
c
If we choose δ sufficiently small such that δ n ·cn1 vol(x,
X, D) ≤ n!c, then for any m-primary
k
⌈δk⌉
n
ideal a satisfying m ⊂ a 6⊂ m
and ℓ(R/a) ≥ ck we have
Thus it suffices to show
c
n! · lct(a) · ℓ(R/a) ≥ vol(x,
X, D).
c
vol(x,
X, D) ≤ n! · lim inf
inf
k→∞ mk ⊂a⊂m⌈δk⌉
ℓ(R/a)≥ck n
lctn (a)ℓ(R/a).
By Lemma 13, we know that for any ǫ > 0 there exists k0 = k0 (δ, ǫ, (R, m)) such that for
any k ≥ k0 we have
n! ·
inf
mk ⊂a⊂m⌈δk⌉
lctn (a)ℓ(R/a) ≥ (1 − ǫ)
Hence the proof is finished.
inf
mk ⊂a⊂m⌈δk⌉
c X, D).
lct(a)n e(a) ≥ (1 − ǫ)vol(x,
The following result on comparison between colengths and multiplicities is crucial in the
proof of Theorem 12. Note that Lemma 13 is a special case of Lech’s inequality [Lec60,
Theorem 3] when R is a regular local ring.
Lemma 13. Let (R, m) be an n-dimensional analytically irreducible Noetherian local domain. Assume that the residue field R/m is algebraically closed. Then for any positive
numbers δ, ǫ ∈ (0, 1), there exists k0 = k0 (δ, ǫ, (R, m)) such that for any k ≥ k0 and any
ideal mk ⊂ a ⊂ m⌈δk⌉ , we have
n! · ℓ(R/a) ≥ (1 − ǫ)e(a).
Proof. By [KK14, 7.8] and [Cut13, Section 4], R admits a good valuation ν : R → Zn for
some total order on Zn . Let S := ν(R \ {0}) ⊂ Nn and C(S) be the closed convex hull of
S. Then we know that
• C(S) is a strongly convex cone;
• There exists a linear functional ξ : Rn → R such that C(S) \ {0} ⊂ ξ>0 ;
• There exists r0 ≥ 1 such that for any f ∈ R \ {0}, we have
(3.2)
ordm (f ) ≤ ξ(ν(f )) ≤ r0 ordm (f ).
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
9
Suppose a is an ideal satisfying mk ⊂ a ⊂ m⌈δk⌉ . Then we have ν(mk ) ⊂ ν(a) ⊂ ν(m⌈δk⌉ ).
By (3.2), we know that
S ∩ ξ≥r0 k ⊂ ν(a) ⊂ S ∩ ξ≥δk .
Similarly, we have S ∩ ξ≥r0 ik ⊂ ν(ai ) ⊂ S ∩ ξ≥δik for any positive integer i.
Let us define a semigroup Γ ⊂ Nn+1 as follows:
Γ := {(α, m) ∈ Nn × N : x ∈ S ∩ ξ≤2r0 m }.
For any m ∈ N, denote by Γm := {α ∈ Nn : (α, m) ∈ Γ}. It is easy to see Γ satisfies
[LM09, (2.3-5)], thus [LM09, Proposition 2.1] implies
#Γm
lim
= vol(∆),
m→∞ mn
where ∆ := ∆(Γ) is a convex body in Rn defined in [LM09, Section 2.1]. It is easy to see
that ∆ = C(S) ∩ ξ≤2r0 .
Let us define Γ(k) := {(α, i) ∈ Nn × N : (α, ik) ∈ Γ}. Then we know that ∆(k) :=
∆(Γ(k) ) = k∆. For an ideal a and k ∈ N satisfying mk ⊂ a ⊂ m⌈δk⌉ , we define
(k)
Γ(k)
: α ∈ ν(ai )}.
a := {(α, i) ∈ Γ
(k)
(k)
Then it is clear that Γa also satisfies [LM09, (2.3-5)]. Since ν(ai ) = (S ∩ ξ>2r0 ik ) ∪ Γa,i
(k)
(k)
and R/m is algebraically closed, we have ℓ(R/ai ) = #(Γi \ Γa,i ) because ν has onedimensional leaves. Again by [LM09, Proposition 2.11], we have
(k)
(k)
#(Γi \ Γa,i )
ℓ(R/ai )
=
lim
= vol(∆(k) ) − vol(∆(k)
n!e(a) = lim
a ),
n
n
i→∞
i→∞
i
i
(k)
(k)
(k)
(k)
where ∆a := ∆(Γa ). Since Γa,i ⊂ ν(ai ) ⊂ ξ≥δik , we know that ∆a ⊂ ξ≥δk . Denote by
(k)
∆′ := C(S) ∩ ξ<δ , then it is clear that ∆a ⊂ k(∆ \ ∆′ ).
On the other hand,
(k)
(k)
n
ℓ(R/a) = #(Γ1 \ Γa,1 ) ≥ #Γk − #(∆(k)
a ∩ Z ).
(k)
Denote by ∆a,k := k1 ∆a , then ∆a,k ⊂ ∆ \ ∆′ . Since vol(∆a,k ) ≤ vol(∆) − vol(∆′ ), there
exists positive numbers ǫ1 , ǫ2 depending only on ∆ and ∆′ such that
ǫ2
ǫ1
vol(∆) − .
(3.3)
vol(∆a,k ) ≤ vol(∆) − vol(∆′ ) ≤ 1 −
ǫ
ǫ
k
⌈δk⌉
Let us pick k0 such that for any k ≥ k0 and any m ⊂ a ⊂ m , we have
(k)
#(∆a ∩ Zn )
#Γk
≥
(1
−
ǫ
)vol(∆),
≤ vol(∆a,k ) + ǫ2 .
1
kn
kn
Here the second inequality is guaranteed by applying Proposition 39 to ∆a,k as a sub
convex body of a fixed convex body ∆. Thus
(k)
#Γk #(∆a ∩ Zn )
ℓ(R/a) − (1 − ǫ)n!e(a)
≥ n −
− (1 − ǫ)(vol(∆) − vol(∆a,k ))
kn
k
kn
≥ (1 − ǫ1 )vol(∆) − vol(∆a,k ) − ǫ2 − (1 − ǫ)(vol(∆) − vol(∆a,k ))
= (ǫ − ǫ1 )vol(∆) − ǫ(∆a,k ) − ǫ2
≥ 0.
10
HAROLD BLUM AND YUCHEN LIU
Here the last inequality follows from (3.3). Hence we finish the proof.
3.2. Normalized volumes under field extensions. In the rest of this section, we
use Hilbert schemes to describe normalized volumes of singularities after a field extension K/k. Let (X, D) be a klt pair over k and x ∈ X be a k-rational point. Let
Zk := Spec(Ox,X /mkx,X ) denote the k-th thickening of x. Consider the Hilbert scheme
Hk,d := Hilbd (Zk /k). For any field extension K/k we know that Hk,d(K) parametrizes
ideal sheaves c of XK satisfying c ⊃ mkxK ,XK and ℓ(OxK ,XK /c) = d. In particular, any
scheme-theoretic point h ∈ Hk,d corresponds to an ideal b of Oxκ(h) ,Xκ(h) satisfying those
two conditions, and we denote by h = [b].
Proposition 14. Let k be a field of characteristic 0. Let (X, D) be a klt pair over k. Let
x ∈ X be a k-rational point. Then
(1) For any field extension K/k with K algebraically closed, we have
ℓc
c,k (xK , XK , DK ) = n! ·
inf
d≥ck n , [b]∈Hk,d
d · lct(Xκ([b]) , Dκ([b]) ; b)n .
(2) With the assumption of (1), we have
c K , XK , DK ) = vol(x
c k̄ , Xk̄ , Dk̄ ).
vol(x
Proof. (1) We first prove the “≥” direction. By definition, ℓbc,k (xK , XK , DK ) is the infimum of n! · lct(XK , DK ; c)n ℓ(OXK /c) where c is an ideal on XK satisfying mkxK ⊂ c ⊂
mxK and ℓ(OXK /c) =: d ≥ ck n . Hence [c] represents a point in Hk,d(K). Suppose [c]
is lying over a scheme-theoretic point [b] ∈ Hk,d , then it is clear that (XK , DK , c) ∼
=
(Xκ([b]), Dκ([b]) , b) ×Spec(κ([b])) Spec(K). Hence lct(XK , DK ; c) = lct(Xκ([b]) , Dκ([b]); b), and
the “≥” direction is proved.
Next we prove the “≤” direction. By Proposition 10, we know that the function [b] 7→
cl
lct(Xκ([b]) , Dκ([b]) ; b) on Hk,d is constructible and lower semicontinuous. Denote by Hk,d
the set of closed points in Hk,d. Since the set of closed points are dense in any stratum of
Hk,d with respect to the lct function, we have the following equality:
n! ·
inf
d≥ck n , [b]∈Hk,d
d · lct(Xκ([b]) , Dκ([b]) ; b)n = n! ·
inf
cl
d≥ck n , [b]∈Hk,d
d · lct(Xκ([b]) , Dκ([b]) ; b)n
cl
Any [b] ∈ Hk,d
satisfies that κ([b]) is an algebraic extension of k. Since K is algebraically closed, κ([b]) can be embedded into K as a subfield. Hence there exists a
point [c] ∈ Hk,d (K) lying over [b]. Thus similar arguments implies that lct(XK , DK ; c) =
lct(Xκ([b]) , Dκ([b]) ; b), and the “≤” direction is proved.
c
(2) From (1) we know that ℓc
c,k (xK , XK , DK ) = ℓc,k (xk̄ , Xk̄ , Dk̄ ) for any c, k. Hence it
follows from Theorem 12.
The following corollary is well-known to experts. We present a proof here using normalized volumes.
Corollary 15. Let (Y, E) be a log Fano pair over a field k of characteristic 0. The
following are equivalent:
(i) (Yk̄ , Ek̄ ) is log K-semistable;
(ii) (YK , EK ) is log K-semistable for some field extension K/k with K = K;
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
11
(iii) (YK , EK ) is log K-semistable for any field extension K/k with K = K.
We say that (Y, E) is geometrically log K-semistable if one (or all) of these conditions
holds.
Proof. Let us take the affine cone X = C(Y, L) with L = −r(KY + E) Cartier. Let D
be the Q-divisor on X corresponding to E. Denote byx ∈ X the cone vertex of X. Let
K/k be a field extension with K = K. Then Theorem 6 implies that (YK , EK ) is log
c K , XK , DK ) = r −1 (−KY − E)n−1 . Hence the corollary is
K-semistable if and only if vol(x
a consequence of Proposition 14 (2).
We finish this section with a natural speculation. Suppose x ∈ (X, D) is a klt singularity
over a field k of characteristic zero that is not necessarily algebraically closed. The
definition of normalized volume of singularities extend verbatimly to x ∈ (X, D) which
c X, D). Then we expect vol(x,
c X, D) = vol(x
c k̄ , Xk̄ , Dk̄ ), i.e.
we also denote by vol(x,
normalized volumes are stable under base change to algebraic closures. Such a speculation
should be a consequence of the Stable Degeneration Conjecture (SDC) stated in [Li15a,
c
Conjecture 7.1] and [LX17, Conjecture 1.2] which roughly says that a vol-minimizing
valuation vmin over xk̄ ∈ (Xk̄ , Dk̄ ) is unique and quasi-monomial, so vmin is invariant
under the action of Gal(k̄/k) and hence has the same normalized volume as its restriction
to x ∈ (X, D).
4. Uniform approximation of volumes by colengths
In this section, we prove the following result that gives an approximation of the volume
of valuation by the colengths of its valuation ideals. The result is a consequence of
arguments in [Blu16, Section 3.4] (which in turn relies on ideas in [ELS03]) and properties
of the Hilbert–Samuel function.
Theorem 16. Let π : (X , D) → T together with a section σ : T → X be a Q-Gorenstein
flat family of klt singularities. Set n = dim(X ) − dim(T ). For every A ∈ R>0 and ǫ > 0,
there exists a positive integer N so that the following holds: If t ∈ T and v ∈ Valσ(t),Xt
satisfies v(mσ(t) ) = 1 and AXt ,Dt (v) ≤ A, then
ℓ(Oσ(t),Xt /am (v))
≤ vol(v) + ǫ
mn /n!
for all positive integers m divisible by N.
We begin by approximating the volume of a valuation by the multiplicity of its valuation
ideals.
Proposition 17. Let x ∈ (X, D) be a klt singularity defined over an algebraically closed
field k and r a positive integer such that r(KX + D) is Cartier. Fix v ∈ Valx,X satisfying
v(mx ) = 1 and AX,D (v) < +∞.
(a) If x ∈ Xsing ∪ Supp(D), then for all m ∈ Z>0 we have
1/n
e(am (v))1/n
⌈AX,D (v)⌉ e(mx )1/n e (OX (−rD) · JacX +mm
x )
≤ vol(v)1/n +
+
.
m
m
m
12
HAROLD BLUM AND YUCHEN LIU
(b) If x ∈
/ Xsing ∪ Supp(D), then for all m ∈ Z>0 we have
⌈AX,D (v)⌉ e(mx )1/n
e(am (v))1/n
≤ vol(v)1/n +
.
m
m
Proof. Fix v ∈ Valx,X satisfying v(mx ) = 1 and AX,D (v) < +∞. To simplify notation, we
set a• := a• (v) and A := ⌈AX,D (v)⌉. By [Blu16, Theorem 7.2],
(4.1)
(JacX ·OX (−rD))ℓ a(m+A)ℓ ⊂ (am )ℓ
for all m, ℓ ∈ Z>0 . Since v(mx ) = 1, we see mm
x ⊂ am for all m ∈ Z>0 . As in the proof of
[Blu16, Proposition 3.7], it follows from the previous inclusion combined with (4.1) that
(4.2)
ℓ
ℓ
(JacX ·OX (−rD) + mm
x ) a(m+A)ℓ ⊂ (am ) .
for all m ∈ Z>0 . We now apply Teissier’s Minkwoski inequality [Laz04, Example 1.6.9] to
the previous inclusion and find that
(4.3)
1/n
ℓ e(am )1/n ≤ ℓ · e(JacX ·OX (−rD) + mm
+ e(a(m+A)ℓ )1/n .
x )
Dividing both sides of (4.3) by m · ℓ and taking the limit as ℓ → ∞ gives
1/n
e(am )1/n
e(JacX ·OX (−rD) + mm
m+A
x )
vol(v)1/n .
≤
+
m
m
m
Since mm
x ⊂ am for all m ∈ Z>0 , vol(v) ≤ e(mx ) and the desired inequality follows. In the
case when x ∈
/ Xsing ∪ Supp(D), the stronger inequality follows from a similar argument
and the observation that (JacX ·OX (−rD)) is trivial in a neighborhood of x.
Before proceeding, we recall basic properties of the Jacobian ideal. Let f : A → B be
a flat finite type morphism of Noetherian rings. Assume the induced map Spec(B) →
Spec(A) has constant relative dimension d. The relative Jacobian ideal JacB/A is defined
as the d-th fitting ideal of ΩB/A . This means, we take a free resolution
ψ
B ⊕p → B ⊕m → ΩB/A → 0
and view ψ as a matrix with elements in B. Now, JacB/A is given by the (m − d)-minors
of ψ. The previous definition gives a local description for JacX/Y when X → Y is a flat
finite type morphism of Noetherian schemes.
Lemma 18. With the same assumptions as in Theorem 16, there exists a nonempty open
set U ⊂ T such that JacXt = JacX /T ·OXt for all t ∈ U.
Proof. It is sufficient to consider the case when X and T are affine. Now, choose a free
resolution E1 → E0 → ΩX /T . By generic flatness, we may choose a nonempty open set
U ⊂ T such that the restriction of E1 → E0 → ΩX /T to Xt is a free resolution of ΩXt /k(t) .
Thus, JacXt = JacX /T ·OXt for all t ∈ U.
Proposition 19. With the same assumptions as in Theorem 16, fix a positive integer
r such that r(KX /T + D) is Cartier. Then, for every ǫ > 0, there exists M so that the
following holds: If t ∈ T satisfies σ(t) ∈ V (JacXt ) ∪ Supp(Dt ), then
e JacXt ·OXt (−rDt ) + mm
σ(t)
≤ ǫ.
mn
for all m ≥ M.
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
13
Proof. To simplify notation, set Z = {t ∈ T |σ(t) ∈ V (JacXt ) ∪ Supp(Dt )}. By the
Noetherianity of T , it suffices to prove the following claim: for each ǫ > 0, there exists a
e(JacXt +mm
)
σ(t)
nonempty open set U ⊂ T and a positive integer M such that the
≤ ǫ for all
mn
m ≥ M and t ∈ Z ∩ U.
We proceed to prove the claim. Fix ǫ > 0 and apply Lemma 18 to choose a nonempty
open affine set U ⊂ T such that JacXt = JacX /T ·OX for all t ∈ U. Now, by applying
Chevalley’s Theorem, we see Z ∩ U is constructible in U. Thus, we may shrink U further
so that either (1) U ∩ Z = ∅ or (2) U ∩ Z = U holds. In case (1), the claim is complete.
Therefore, we assume (2) holds and shrink U accordingly.
Next, choose a nonzero element g ∈ JacX /T ·OX (−rD)(U) such that the restriction of
e t = mσ(t) · Rt
g to Oσ(t),Xt , denoted gt , is nonzero for all t ∈ U. Set Rt := Oσ(t),Xt /gt and m
for each t ∈ U. Note that Rt is n − 1 dimensional for all t ∈ U . Now,
m
m
e(JacXt ·OXt (−rDt ) + mm
σ(t) ) ≤ e(gt + mσ(t) ) ≤ n! · ℓ(Oσ(t),Xt /(gt + mσ(t) )) · e(mσ(t) ),
m
where the first inequality follows from the inclusion JacXt ·OXt (−rDt ) + mm
σ(t) ⊂ gt + mσ(t)
and the second is precisely Lech’s inequality [Lec60, Theorem 3]. Thus,
e(JacXt ·OXt (−rDt ) + mm
em
ℓ(Rt /m
σ(t) )
t )
(4.4)
· e(mσ(t) ).
≤ n!
mn
mn
em
for all t ∈ U. By Proposition 42, we may shrink U again so that e(mσ(t) ) and ℓ(Rt /m
t ))
m
n−1
e t ) = O(m ).
are constant for all t ∈ U and m ≥ 1. Since dim Rt = n−1, we have ℓ(Rt /m
Therefore, there exists an integer M so that
em
ℓ(Rt /m
t )
n!
· e(mσ(t) ) ≤ ǫ
mn
for all m ≥ M and t ∈ U. This completes the claim.
The following proposition is a consequence of results in Appendix B.
Proposition 20. Keep the assumptions and notation in Theorem 16, and fix an integer
k ∈ Z>0 . Then, for any ǫ > 0, there exists M ∈ Z>0 so that the following holds: For any
point t ∈ T and ideal a ⊂ Oσ(t),Xt satisfying mkσ(t) ⊂ a ⊂ mσ(t) ,
ℓ(Oσ(t),Xt /am )
≤ e(a) + ǫ
mn /n!
for all m ≥ M.
Proof. Set d := max{ℓ(Oσ(t),Xt /mkσ(t) ) | t ∈ T }, and consider the union of Hilbert schemes
S
k
H := dm=1 Hilbm (Zk /T ), where Zk = Spec(OX /Iσ(T
) ). Let τ denote the morphism
H → T . A point h ∈ H corresponds to the ideal bh = b·OX ×T κ(h) , where b is the universal
ideal sheaf on X ×T H. By applying Proposition 42 to the irreducible components of H
endowed with reduced scheme structure, we see that the set of functions {Hbh | h ∈ H}
is finite.
Next, fix ǫ > 0. By the previous paragraph, there exists M ∈ Z>0 so that
(4.5)
Hbh (m))
≤ e(bh ) + ǫ
mn /n!
14
HAROLD BLUM AND YUCHEN LIU
for all m ≥ M. Now, consider a point t ∈ T and an ideal a ⊂ OXt satisfying mkσ(t) ⊂ a ⊂
mσ(t) . Since
ℓ(Oσ(t),Xt /a) ≤ ℓ(Oσ(t),Xt /mkσ(t) ) = ℓ(Oσ(t),Xt /mkσ(t) ) ≤ d,
there is a map ρ : Spec(κ(t)) → H such that a = bρ(0) · OX ×T κ(t) . Therefore, Ha = Hbρ(0)
and (4.5) implies
ℓ(Oσ(t),Xt /am )
≤ e(a) + ǫ
mn /n!
for all m ≥ M.
We will now deduce Theorem 16 from Propositions 17, 19, and 20.
Proof of Theorem 16. To simplify notation, we set
Wt = {v ∈ Valσ(t),Xt | v(mσ(t) ) = 1 and AXt ,Dt (v) ≤ A}
for each t ∈ T . In order to prove the theorem, it suffices to prove the following claim: for
every ǫ > 0, there exists an integer N so that if t ∈ T , then
ℓ(Oσ(t),Xt /am (v)) 1/n
≤ vol(v)1/n + ǫ
(m)n /n!
for all v ∈ Wt and m ∈ Z>0 divisible by N. Indeed, if v ∈ Wt , then vol(v) ≤ e(mσ(t) ).
Since the set {e(mσ(t) ) | t ∈ T } is bounded from above by Proposition 42, the claim implies
the conclusion of the theorem.
We now fix ǫ > 0 and proceed to bound the latter two terms in Proposition 17.1. First,
we apply Proposition 20 to find a positive integer M1 so that
A · e(mσ(t) )1/n
≤ ǫ/4
M1
for all t ∈ T . Next, we apply Proposition 19 to find a positive integer M2 so that the
following holds: if t ∈ T and σ(t) ∈ V (JacXt ) ∪ Supp(Dt ), then
′
′
e(JacXt ·OXt (−rDt ) + mσ(t) m )1/n
e(JacXt ·OXt (−rDt ) + mσ(t) m )1/n
=
< ǫ/4.
m′
m′
for all m′ ≥ M2 . Now, set m′ := max{M1 , M2 }. Proposition 17 implies that if t ∈ T , then
(4.6)
e(am′ (v))1/n
≤ vol(v)1/n + ǫ/2
m′
for all v ∈ Wt .
′
Next, note that if t ∈ T and v ∈ Wt , then mm
⊂ am′ (v). Therefore, we may apply
σ(t)
Proposition 20 to find an integer M such that if t ∈ T and v ∈ Wt , then
!1/n
ℓ(Oσ(t),Xt /(am′ (v))ℓ )
e(am′ (v))1/n
+ ǫ/2
≤
(4.7)
(m′ · ℓ)n /n!
m′
for all ℓ ≥ M. Thus, if t ∈ T and v ∈ Wt , then
ℓ(Oσ(t),Xt /(am′ ·ℓ (v))) 1/n e(am′ (v))1/n
≤
+ ǫ/2 ≤ vol(v) + ǫ
(m′ · ℓ)n /n!
m′
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
15
for all ℓ ≥ M, where the first inequality follows from (4.7) combined with the inclusion
am′ (v)ℓ ⊂ am′ ·ℓ (v) and the second inequality from (4.6). Therefore, setting N := m′ · M
completes the claim.
5. Li’s Izumi and properness estimates in families
In this section, we generalize results of [Li15a] to families of klt singularities. These
results will be used to prove Theorem 27.
Theorem 21 (Izumi-type Estimate). Let π : (X , D) → T together with a section σ :
T → X be a Q-Gorenstein flat family of klt singularities over a variety T . There exists
a constant K0 > 0 so that the following holds: If t ∈ T and v ∈ Valσ(t),Xt satisfies
AXt ,Dt (v) < +∞, then
v(g) ≤ K0 · AXt ,Dt (v) · ordσ(t) (g)
for all g ∈ Oσ(t),Xt .
Theorem 22 (Properness Estimate). Let π : (X , D) → T together with a section σ :
T → X be a Q-Gorenstein flat family of klt singularities over a variety T . There exists
a constant K1 > 0 so that the following holds: If t ∈ T and v ∈ Valσ(t),Xt satisfies
AXt ,Dt (v) < +∞, then
K1 · AXt ,Dt (v)
≤ AXt ,Dt (v)n · vol(v).
v(mσ(t) )
The proofs of these theorems rely primarily on the result and techniques found in
[Li15a]. The main new ingredient can be found in Proposition 25, which is proved using
arguments of [BFJ14] and [Li15a, Appendix II].
5.1. Order functions. Let X be a normal variety defined over an algebraically closed
field k and x ∈ X a closed point. For g ∈ Ox,X the order of vanishing of g at x is defined
as
ordx (g) := max{j ≥ 0 | g ∈ mjx }.
If X is smooth at x, then ordx is a valuation of the function field of X. In the singular
case, ordx may fail to be a valuation. For example, the inequality
′
′
ordx (g n+n ) ≥ ordx (g n ) + ordx (g n )
c x , which is
may be strict. Following [BFJ14], we consider an alternative function ord
defined by
c x (g) := lim 1 ordx (g n ) = sup 1 ordx (g n ).
ord
n→∞ n
n n
+
Let ν : X → X denote the normalized blowup of mx and write
!
r
X
ai Ei ,
mx · OX + = OX −
i=1
where the Ei are prime divisors on X and each ai ∈ Z>0 . The following statement,
c x in terms of the
which was proved in [BFJ14, Theorem 4.3], gives an interpretation of ord
exceptional divisors of ν.
16
HAROLD BLUM AND YUCHEN LIU
Proposition 23. For any function g ∈ Ox,X and m ∈ Z>0 ,
c x (g) = mini=1,...,r ordEi (g) and
(1) ord
ai
c x (g) ≥ m if and only if g ∈ mm .
(2) ord
x
Building upon results in [BFJ14, Section 4.1], we show a comparison between ordx and
c
ordx .
Proposition 24. If there exists a Q-divisor D such that (X, D) is klt pair, then
c x (g) ≤ (n + 1)ordx (g)
ordx (g) ≤ ord
for all g ∈ Ox,X and n = dim(X).
c x (g) as a supremum. For the
Proof. The first inequality follows from the definition of ord
second inequality, assume m := ordx (g) > 0 and note that g ∈
/ mm+1
. Since
x
mm+n
⊂ J ((X, D), mm+n
) = mm+1
· J ((X, D), mxn−1 ) ⊂ mx m+1 ,
x
x
x
where the first inclusion follows from the fact that (X, D) is klt and the second from
c x (g) < m + n, and the
. Therefore, ord
Skoda’s Theorem [Laz04, 9.6.39], we see g ∈
/ mm+n
x
claim is complete.
5.2. Izumi type estimates. The propositions in the section concern the following setup,
which will arise in the proof of Theorem 21. Let x ∈ (X, D) be an affine klt singularity
over an algebraically closed field k. Fix a projective compactification X ⊂ X and a
resolution of singularities π : Y → X. Assume there exists a very ample line bundle L on
Y and the restriction of π to X, denoted π : Y → X, is a log resolution of (X, D, mx ).
Proposition 25. There exists a constant C0 so that the following holds: For any closed
point y ∈ π −1 (x) and g ∈ Ox,X , we have
ordy (π ∗ g) ≤ C0 · ordx (g).
P
Furthermore, if we write mx · OY = OY (− ri=1 ai Ei ) where each Ei is a prime divisor
on Y and ai ∈P
Z>0 , then there is a formula for such a constant C0 given in terms of the
coefficients of
ai Ei , the intersection numbers (Ei · Ej · Ln−2 ) for 1 ≤ i, j ≤ r, and the
dimension of X.
The proposition is a refined version of [Li15a, Theorem 3.2]. Its proof relies on ideas in
[BFJ14] and [Li15a, Appendix II].
Proof. Fix a closed point y ∈ π −1 (x) and an element g ∈ Ox,X . Let ρ : By Y → Y denote
the blowup of Y at y with exceptional divisor F0 . We write µ := π ◦ ρ and Fi for the strict
transform of Ei . Consider the divisor G given by the closure of {µ∗ g = 0} and write
G=
r
X
i=0
e
bi Fi + G,
where no Fi lies in the support of G̃. Note that b0 = ordy (π ∗ g) and bi := ordEi (g)
for 1 ≤ i ≤ r. Since Y → X factors through the normalized blowup of X along mx ,
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
17
Proposition 23 implies
ordEi (g)
bi
c x (g).
= min
≤ ord
1≤i≤r
1≤i≤r ai
ai
To simplify notation, we set a := max{ai }.
Our goal will be to find a constant C such that b0 ≤ C · bi for all i ∈ {1, . . . , r}. After
finding such a C, we will have that
c x (g) ≤ (C/a)(n + 1)ordx (g),
ordy (π ∗ g) = b0 ≤ (C/a)ord
min
where the last inequality follows from Proposition 24. Thus, the desired inequality will
hold with C0 = (C/a)(n + 1).
We now proceed to find such a constant C. Set M = ρ∗ L − (1/2)F0 and note that M
is ample [Laz04, Example 5.1.6]. For each i ∈ {1, . . . , r}, we consider
r
X
e · Fi · M n−2 ≤ G · Fi · M n−2 = 0,
bj (Fi · Fj · M n−2 ) = G · Fi · M n−2 − G
j=0
where the last equality follows from the fact that G is a principal divisor in a neighborhood
of π −1 (0). Now, we set
cij := (Fi · Fj · M n−2 ).
and see
X
bj cij ≤ −bi cii ≤ bi |cii |.
j6=i
Note that if i 6= j, then cij 6= 0 if and only if Fi ∩ Fj 6= ∅. When that is the case,
bj ≤
(5.1)
|cii |
bi .
cij
Computing the cij in terms of intersection numbers on Y , we find that for i, j ∈ {1, . . . , r}
(
(Ei · Ej · Ln−2 ) − (1/2)n−2 if 0′ ∈ Ei ∩ Ej
cij =
.
(Ei · Ej · Ln−2 )
otherwise
Additionally,
(
(1/2)n−2
c0i =
0
if 0′ ∈ Ei
.
otherwise
Now, for each i, j ∈ {1, . . . , r} such that i 6= j and Ei ∩ Ej 6= ∅, we set
Cij =
|(Ei · Ei · Ln−2 )| + (1/2)n−2
.
(Ei · Ej · Ln−2 ) − (1/2)n−2
Note that |cii |/cij ≤ Cij . For each i, we set
Ci0 =
|(Ei · Ei · Ln−2 )| + (1/2)n−2
.
(1/2)n−2
Similarly, note that |cc0iii | ≤ Ci0 if 0′ ∈ Ei .
Now, set C ′ = max{1, Cij , Ci0 }. By our choice of C ′ , if i, j ∈ {0, 1, . . . , r} are distinct
and Fi ∩ Fj 6= ∅, then bj ≤ C ′ · bi . Now, Zariski’s Main Theorem implies ∪Fi is connected.
Therefore, we set C = 1+C ′ +C ′2 +· · · C ′r and conclude b0 ≤ C·bi for all i ∈ {1, . . . , r}.
18
HAROLD BLUM AND YUCHEN LIU
Proposition 26. There exists a constant C1 such that the following holds: If v ∈ Valx,X
satisfies AX,D (v) < +∞ and y ∈ cY (v), then
v(g) ≤ C1 · AX,D (v)ordy (π ∗ g)
for all g ∈ Ox,X . Furthermore, if KY − π ∗ (KX + D) has coefficients > −1 + ǫ with
0 < ǫ < 1, then the condition holds when C1 := 1/ǫ.
Proof. A proof of the statement can be found in the proof [Li15a, Theorem 3.1] in the
case when D = 0. The more general statement follows from a similar argument.
5.3. Proofs of Theorems 21 and 22.
Proof of Theorem 21. It is sufficient to prove the theorem in the case when both X and
T are affine. We will show that there exists a nonempty open set U ⊂ T and a constant
K0 > 0 such that the conclusion of the theorem holds for all t ∈ U. By induction on the
dimension of T , the proof will be complete.
Fix a (relative) projective compactification π : X → T . Denote the ideal sheaf of
σ(T ) in X by Iσ(T ) . Fix a projective resolution of singularities ρ : Y → X such that its
restriction to X , denoted ρ : Y → X , is a log resolution of (X , D, Iσ(T ) ). Set µ = π ◦ ρ.
We write
!
k
k
X
X
∗
ai Ei
bi Ei and KY − ρ (KX + D) =
Iσ(T ) · OY = OY −
i=1
i=1
where each Ei is a prime divisor on Y. We order these prime divisors so that each Ei
dominates T if and only if 1 ≤ i ≤ r for some positive integer r ≤ k.
By generic smoothness, there exists a nonempty open set U1 ⊂ T such that Yt → Xt
is a log resolution of (Xt , Dt , mσ(t) ) for all t ∈ U1 and µ−1 (U1 ) → U1 is smooth. Further
shrinking U1 , we may assume Ei,t 6= ∅ if and only if 1 ≤ i ≤ r. Let us assume i ≤ r for
the rest of the proof. Now, we have
!
r
r
X
X
mσ(t) · OYσ(t) = OY −
bi Ei,t and KYt − ρt∗ (KXt + Dt ) =
ai Ei,t
i=1
i=1
for each t ∈ U1 . Note that the divisors Ei |t may have multiple irreducible components.
Next, we apply [SPA, Tag 0551] to find and an étale morphism T ′ → U1 , with T ′
irreducible and so that all irreducible components of the generic fiber of E ′ = Ei ×T T ′ → T ′
are geometrically irreducible. Denote by (X ′ , D ′ , Y ′ , Ei′) := (X , D, Y, Ei) ×T T ′ , and η ′ the
generic points of T ′ . Write
E ′ i,η′ = E ′ i,1,η′ ∪ · · · ∪ E ′ i,mi ,η′
′
for the decomposition of Ei,η
′ into irreducible components, and set Ei,j equal to the closure
′
′
of Ei,j,η′ in Y . Applying [SPA, Tag 0559], we may find an open subset U ′ ⊂ T ′ so that
each divisor Ei,j,t is geometrically irreducible for all t ∈ U ′ . Further shrinking U ′ , we may
assume that the divisors Ei,j,t for 1 ≤ i ≤ r and 1 ≤ j ≤ mi are distinct. We choose
U ⊂ T to be a nonempty open subset contained in the image of U ′ in T .
We seek to find a constant C0 such that if t ∈ U, then
(5.2)
ordσ(t) (g) ≤ C0 · ordy (ρt∗ (g))
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
19
for all g ∈ Oσ(t),Xt and y ∈ ρt−1 (σ(t)). Since (X ′ , D ′, Y ′ )|t′ ∼
= (X , D, Y)|t if t is the
′
image of t , it suffices to establish an inequality of the form (5.2) on the singularities
σ(t′ ) ∈ (Xt′′ , Dt′′ ) for t′ ∈ U ′ . Let L be a line bundle on Y such that Lt is very ample for
all t ∈ T , and write L′ for the pullback of L to Y ′ . Now, fix 1 ≤ i1 , i2 ≤ r such that
bi1 , bi2 > 0. For fixed 1 ≤ j1 ≤ mi1 and 1 ≤ j2 ≤ mi2 , the function that sends U ′ ∋ t′
to (Ei1 ,j1 ,t′ · Ei2 ,j2 ,t′ · L′t′ n−2 ) is constant [Kol96, Lemma VI.2.9]. From our choice of U ′ , we
know that Ei,j,t′ is irreducible for any t′ ∈ U ′ . Therefore, we may apply Proposition 25 to
find such a constant C0 such that the desired inequality holds for all t′ ∈ U ′ .
Next, choose 0 < ǫ < 1 so that ai < 1 − ǫ for all 1 ≤ i ≤ r. Set C1 := 1/ǫ. By
Proposition 26, if t ∈ U and v ∈ Valσ(t),Xt , then
(5.3)
v(g) ≤ C1 · AXt ,Dt ordy (ρt∗ g)
for all g ∈ Oσ(t),Xt and y ∈ cYt (v). Combining (5.2) and (5.3), we see that the desired
inequality holds when K0 = C0 · C1 .
Proof of Theorem 22. The theorem follows immediately Theorem 21 and [Li15a, Theorem
4.1], as in the proof of [Li15a, Theorem 4.3].
6. Proofs and applications
6.1. A convergence result for normalized colengths.
Theorem 27. Let π : (X , D) → T together with a section σ : T → X be a Q-Gorenstein
flat family of klt singularities. For every ǫ > 0, there exists a constant c1 > 0 and integer
N so that the following holds: if t ∈ T , then
c
ℓd
c,m (σ(t), Xt , Dt ) ≤ vol(σ(t), Xt , Dt ) + ǫ
for all m divisible by N and 0 < c ≤ c1 .
Before beginning the proof of the previous theorem, we record the following statement.
Proposition 28. Let π : (X , D) → T together with a section σ : T → X be a QGorenstein flat family of klt singularities. There exists a constant A so that
o
n
c
c
Xt , Dt ) = inf vol(v)
vol(σ(t),
| v ∈ Valσ(t),Xt with v(mσ(t) ) = 1 and AXt ,Dt (v) ≤ A
for all t ∈ T .
c
Xt , Dt ) ≤ B
Proof. We first note that there exists a real number B so that vol(σ(t),
n
n
c
for all t ∈ T . Indeed, vol(σ(t), Xt , Dt ) ≤ lct(mσ(t) ) e(mσ(t) ) = lct(mσ(t) ) e(mσ(t) ) and the
function that sends t ∈ T to lct(mσ(t) )n e(mσ(t) ) takes finitely many values by Propositions
10 and 42. Thus,
o
n
c
c
c
vol(σ(t), Xt , Dt ) = inf vol(v) | v ∈ Valσ(t),Xt with v(mσ(t) ) = 1 and vol(v) ≤ B
for all t ∈ T .
Next, fix a constant K1 ∈ R>0 satisfying the conclusion of Theorem 22. If v ∈ Valσ(t),Xt
c
≤ B, then AXt ,Dt (v) ≤ B/K2 . Therefore, the proposition
satisfies v(mσ(t) ) = 1 and vol(v)
holds with A := B/K2 .
20
HAROLD BLUM AND YUCHEN LIU
Proof of Theorem 27. Fix ε > 0 and a constant A ∈ R>0 satisfying the conclusion of the
previous proposition. To simplify notation, set
Wt = {v ∈ Valσ(t),Xt | v(mσ(t) ) = 1 and A(Xt ,Dt ) (v) ≤ A}
for each t ∈ T . We proceed by proving the following two claims.
Claim 1: There exist constants c1 ∈ R>0 and M1 ∈ Z>0 such that the following holds: if
t ∈ T , then
n
ℓd
c,m (σ(t), Xt , Dt ) ≤ inf n! · lct(am (v)) ℓ(Oσ(t),X /am (v))
t
v∈Wt
for all 0 < c < c1 , and m ≥ M1 .
Proposition 42 implies there exist constants c1 > 0 and M1 ∈ Z>0 such that
n
ℓ(Oσ(t),Xt /mm
σ(t) ) ≥ m · c1
for all t ∈ T and m ≥ M1 . Now, consider v ∈ Wt for some t ∈ T . Since v(mσ(t) ) = 1,
mm
⊂ am (v) for all m ∈ Z>0 . Therefore, ℓ(Oσ(t),Xt /am (v)) ≥ ℓ(Oσ(t),Xt /mm
)). The
σ(t)
σ(t)
claim now follows from definition of ℓd
c,m (σ(t), Xt , Dt ).
Claim 2: There exists M2 ∈ Z>0 such that the following holds: if t ∈ T and v ∈ Wt ,
then
c
+ε
n! · lct(am (v))n · ℓ(Oσ(t),X /am (v)) ≤ vol(v)
t
for all integers m divisible by M2 .
By Theorem 16, there exists M2 ∈ Z>0 such the following holds: If t ∈ T and v ∈ Wt ,
then
(6.1)
ℓ(Oσ(t),Xt /am (v))
≤ vol(v) + ε/An
n
m /n!
for all integers m divisible by M2 . Note that
m · lct(am (v)) ≤ lct(a• (v)) ≤ AXt ,Dt (v).
Therefore, multiplying (6.1) by (m · lct(am (v)))n yields the desired result.
We return to the proof of the corollary. Fix constants M1 , M2 , and c1 satisfying the
conclusions of the Claims 1 and 2. Set M = M1 · M2 . Now, if t ∈ T , m is a postive integer
divisible by M, and c satisfies 0 < c < c1 , then
n
ℓd
c,m (σ(t), Xt , Dt ) ≤ inf n! · lct(am (v)) ℓ(Oσ(t),Xt /am (v))
v∈Wt
c
≤ inf vol(v)
+ε
v∈Wt
c
Xt , Dt ) + ε,
= vol(σ(t),
where the first (in)equality follows from Claim 1, the second from Claim 2, and the third
from our choice of A.
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
21
6.2. Proofs. The following theorem is a stronger result that implies Theorem 1.
Theorem 29. Let π : (X , D) → T together with a section σ : T → X be a Q-Gorenstein
flat family of klt singularities over a field k of characteristic 0. Then the function t 7→
c
vol(σ(t),
Xt , Dt ) on T is lower semicontinuous with respect to the Zariski topology.
k
Proof. Let Zk → T be the k-th thickening of the section σ(T ), i.e. Zk = SpecT (OX /Iσ(T
) ).
k
Let dk := maxt∈T ℓ(Oσ(t),Xt /mσ(t),Xt ). For any d ∈ N, denote Hk,d := Hilbd (Zk /T ). Since
n
Zk is proper over T , we know that Hk,d is also proper over T . Let Hk,d
be the normalization
of Hk,d. Denote by τk,d : Hk,d → T . After pulling back the universal ideal sheaf on
n
n
X ×T Hk,d over Hk,d to Hk,d
, we obtain an ideal sheaf bk,d on X ×T Hk,d
. Denote by
n
n
n
πk,d : (X ×T Hk,d , D ×T Hk,d ) → Hk,d the projection, then πk,d provides a Q-Gorenstein
flat family of klt pairs.
n
Following the notation of Proposition 14, assume h is scheme-theoretic point of Hk,d
lying over [b] ∈ Hk,d . Denote by t = τk,d ([b]) ∈ T . By construction, the ideal sheaf bk,d,h on
X ×T Spec(κ(h)) is the pull back of b under the flat base change Spec(κ(h)) → Spec(κ([b])).
Hence
lct((X , D) ×T Spec(κ(h)); bk,d,h ) = lct((X , D) ×T Spec(κ([b])); b).
For simplicity, we abbreviate the above equation to lct(bk,d,h ) = lct(b). Applying Propon
sition 10 to the family πk,d and the ideal bk,d implies that the function Φn : Hk,d
→ R>0
n
defined as Φ (h) := lct(bk,d,h ) is constructible and lower semicontinuous with respect to
n
the Zariski topology on Hk,d
. Since lct(bk,d,h ) = lct(b), Φn descends to a function Φ
on Hk,d as Φ([b]) := lct(b). Since Hk,d is proper over T , we know that the function
φ : T → R>0 defined as
φ(t) := n! · nmin Φ([b])n
ck ≤d≤dk
−1
(t)
[b]∈τk,d
is constructible and lower semicontinuous with respect to the Zariski topology on T .
Then Proposition 14 implies φ(t) = ℓc
Thus we conclude that t 7→
c,k (σ(t), Xt , Dt ).
c
ℓc,k (σ(t), Xt , Dt ) is constructible and lower semicontinuous with respect to the Zariski
topology on T .
Let us fix ǫ > 0 and a scheme-theoretic point o ∈ T . By Theorem 27, there exist c1 > 0
and N ∈ N such that
ǫ
c
(6.2)
vol(σ(t),
Xt , Dt ) ≥ ℓc
c,k (σ(t), Xt , Dt ) −
2
for any t ∈ T , k divisible by N and 0 < c ≤ c1 . Since t 7→ ℓc
c,k (σ(t), Xt , Dt ) is constructibly
lower semicontinuous on T , there exists a Zariski open neighborhood U of o such that
c
(6.3)
ℓc
c,k (σ(t), Xt , Dt ) ≥ ℓc,k (σ(o), Xo , Do ) for any t ∈ U.
By Theorem 12, there exist c0 > 0 and N0 ∈ N such that
ǫ
c
ℓc
c,k (σ(o), Xo , Do ) ≥ vol(σ(o), Xo , Do ) −
2
for any 0 < c ≤ c0 and any k ≥ N0 . Let us choose c = min{c0 , c1 } and k = N · N0 . Then
combining (6.2), (6.3) and (6.4) yields that
(6.4)
c
c
Xt , Dt ) ≥ vol(σ(o),
Xo , Do ) − ǫ for any t ∈ U.
vol(σ(t),
22
HAROLD BLUM AND YUCHEN LIU
The proof is finished.
The following theorem is a stronger result that implies Theorem 3.
Theorem 30. Let ϕ : (Y, E) → T be a Q-Gorenstein flat family of log Fano pairs over a
field k of characteristic 0. Assume that some geometric fiber (Yo , Eo) is log K-semistable
for a point o ∈ T . Then
(1) There exists an intersection U of countably many Zariski open neighborhoods of o,
such that (Yt , Et) is log K-semistable for any point t ∈ T . If, in addition, k = k̄ is
uncountable, then (Yt , Et ) is log K-semistable for a very general closed point t ∈ T .
(2) The geometrically log K-semistable locus
T K-ss := {t ∈ T : (Yt , Et ) is log K-semistable}
is stable under generalization.
Proof. (1) For r ∈ N satisfying L = −r(KY/T + E) is Cartier, we define the relative affine
cone X of (Y, L) by
X := SpecT ⊕m≥0 ϕ∗ (L⊗m ).
Assume r is sufficiently large, then it is easy to see that ϕ∗ (L⊗m ) is locally free on T for
all m ∈ N. Thus we have Xt ∼
= Spec ⊕m≥0 H 0 (Yt , L⊗m
t ) := C(Yt , Lt ). Let D be the Qdivisor on X corresponding to E. By [Kol13, Section 3.1], the projection π : (X , D) → T
together with the section of cone vertices σ : T → X is a Q-Gorenstein flat family of klt
singularities.
Since (Yo , Eo) is K-semistable, Theorem 6 implies
c
vol(σ(o),
Xo , Do ) = r −1 (−KYo − Eo )n−1 .
Then by Theorem 29, there exists an intersection U of countably many Zariski open
c
c
Xt , Dt ) ≥ vol(σ(o),
Xo , Do) for any t ∈ U. Since
neighborhoods of o, such that vol(σ(t),
the global volumes of log Fano pairs are constant in Q-Gorenstein flat families, we have
c
c
vol(σ(t),
Xt , Dt ) ≥ vol(σ(o),
Xo , Do ) = r −1 (−KYo − Eo)n−1 = r −1 (−KYt − Et )n−1 .
Then Theorem 6 implies that (Yt , Et ) is K-semistable for any t ∈ U.
(2) Let o ∈ T K-ss be a scheme-theoretic point. Then by (1) there exists countably many
Zariski open neighborhoods Ui of o such that ∩i Ui ⊂ T K-ss . If t is a generalization of o,
then t belongs to all Zariski open neighborhoods of o, so t ∈ T K-ss .
Proof of Theorem 3. It is clear that (1) and (2) follows from Theorem 30. For (3), the
constructibility of normalized volumes implies that the set U in the proof of Theorem 30
(1) can be chosen as a Zariski open neighborhood of o. Then the same argument in the
proof of Theorem 30 (1) works.
The following corollary is a stronger result that implies Corollary 4.
Corollary 31. Let π : (Y, E) → T be a Q-Gorenstein family of complex log Fano pairs.
Assume that π is isotrivial over a Zariski open subset U ⊂ T , and (Yo , Eo ) is log Ksemistable for a closed point o ∈ T \ U. Then (Yt , Et ) is log K-semistable for any t ∈ U.
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
23
Proof. Since (Yo , Eo) is log K-semistable, Theorem 30 implies that (Yt , Et ) is log Ksemistable for very general closed point t ∈ T . Hence there exists (hence any) t ∈ U
such that (Yt , Et ) is log K-semistable.
Remark 32. If the ACC of normalized volumes (in bounded families) were true, then
Conjecture 2 follows by applying Theorem 1. Moreover, we suspect that a much stronger
result on discreteness of normalized volumes away from 0 (see also [LiuX17, Question
4.3]) might be true, but we don’t have much evidence yet.
6.3. Applications. In this section we present applications of Theorem 1. The following
theorem generalizes the inequality part of [LiuX17, Theorem A.4].
Theorem 33. Let x ∈ (X, D) be a complex klt singularity of dimension n. Let a be the
c X, D) ≤ (1 − a)nn .
largest coefficient of components of D containing x. Then vol(x,
Proof. Suppose Di is the component of D containing x with coefficient D. Let Din be the
normalization of Di . By applying Theorem 1 to pr2 : (X×Din , D×Din ) → Din together with
c
c X, D)
the natural diagonal section σ : Din → X × Din , we have that vol(x,
X, D) ≤ vol(y,
for a very general closed point y ∈ Di . We may pick y to be a smooth point in both X and
c X, D) ≤ vol(0,
c An , aAn−1 ) where An−1 is a coordinate hyperplane of An . Let
D, then vol(x,
us take local coordinates (z1 , · · · , zn ) of An such that An−1 = V (z1 ). Then the monomial
1
valuation va on An with weights ((1 − a)−1 , 1, · · · , 1) satisfies AAn (v) = 1−a
+ (n − 1),
1
n−1
ordva (A ) = 1−a and vol(va ) = (1 − a). Hence
c
c 0,(An ,aAn−1 ) (va ) = (AAn (v) − aordv (An−1 ))n · vol(va ) = (1 − a)nn .
vol(x,
X, D) ≤ vol
a
The proof is finished.
Theorem 34. Let (X, D) be a klt pair over C. Then
c
(1) The function x 7→ vol(x,
X, D) on X(C) is lower semicontinuous with respect to
the Zariski topology.
(2) Let Z be an irreducible subvariety of X. Then for a very general closed point z ∈ Z
we have
c X, D) = sup vol(x,
c X, D).
vol(z,
x∈Z
In particular, there exists a countable intersection U of non-empty Zariski open
c X, D)|U is constant.
subsets of Z such that vol(·,
Proof. Part (1) follows quickly by applying Theorem 1 to pr2 : (X × X, D × X) → X
together with the diagonal section σ : X → X × X. For part (2), denote by Z n the
normalization of Z. Then the proof follows quickly by applying Theorem 1 to pr2 :
(X × Z n , D × Z n ) → Z n together with the natural diagonal section σ : Z n → X × Z n .
Next we study the case when X is a Gromov-Hausdorff limit of Kähler-Einstein Fano
c X) = nn · Θ(x, X) is lower semicontinuous
manifolds. Note that the function x 7→ vol(x,
with respect to the Euclidean topology on X by [SS17, LX17]. The following corollary
improves this result and follows easily from part (1) of Theorem 34.
Corollary 35. Let X be a Gromov-Hausdorff limit of Kähler-Einstein Fano manifolds.
c X) = nn · Θ(x, X) on X(C) is lower semicontinuous with
Then the function x 7→ vol(x,
respect to the Zariski topology.
24
HAROLD BLUM AND YUCHEN LIU
The following theorem partially generalizes [SS17, Lemma 3.3 and Proposition 3.10].
Theorem 36. Let X be a Gromov-Hausdorff limit of Kähler-Einstein Fano manifolds.
Let x ∈ X be any closed point. Then for any finite quasi-étale morphism of singularities
π : (y ∈ Y ) → (x ∈ X), we have deg(π) ≤ Θ(x, X)−1 . In particular, we have
(1) |π̂1loc (X, x)| ≤ Θ(x, X)−1 .
(2) For any Q-Cartier Weil divisor L on X, we have ind(x, L) ≤ Θ(x, X)−1 where
ind(x, L) denotes the Cartier index of L at x.
c Y) =
Proof. By [LX17, Theorem 1.7], the finite degree formula holds for π, i.e. vol(y,
c X). Since vol(y,
c Y ) ≤ nn by [LiuX17, Theorem A.4] or Theorem 33 and
deg(π) · vol(x,
c X) = nn · Θ(x, X) by [LX17, Corollary 5.7], we have deg(π) ≤ nn /vol(x,
c X) =
vol(x,
−1
Θ(x, X) .
Remark 37. If the finite degree formula [LiuX17, Conjecture 4.1] were true for any klt
c X) holds for any finite quasi-étale morphism
singularity, then clearly deg(π) ≤ nn /vol(x,
π : (y, Y ) → (x, X) between n-dimensional klt singularities. In particular, we would get
c X) where π̂ loc (X, x) is known to be finite
an effective upper bound |π̂1loc (X, x)| ≤ nn /vol(x,
1
by [Xu14, BGO17] (see [LiuX17, Theorem 1.5] for a partial result in dimension 3).
Theorem 38. Let V be a K-semistable complex Q-Fano variety of dimension (n − 1).
Let q be the largest integer such that there exists a Weil divisor L satisfying −KV ∼Q qL.
Then
q · (−KV )n−1 ≤ nn .
Proof. Consider the orbifold cone X := C(V, L) = Spec(⊕m≥0 H 0 (V, OV (⌊mL⌋)) with
the cone vertex x ∈ X. Let X̃ := SpecV ⊕m≥0 OV (⌊mL⌋) be the partial resolution
of X with exceptional divisor V0 . Then by [Kol04, 40-42], x ∈ X is a klt singularity,
and (V0 , 0) ∼
= (V, 0) is a K-semistable Kollár component over x ∈ X. Hence [LX16,
c x,X . By [Kol04, 40-42] we have AX (ordV ) =
Theorem A] implies that ordV0 minimizes vol
0
n−1
q, vol(ordV0 ) = (L ). Hence
c X) = AX (ordV )n vol(ordV ) = q n (Ln−1 ) = q(−KV )n−1 ,
vol(x,
0
0
c X) ≤ nn by [LiuX17, Theorem A.4] or Theorem
and the proof is finished since vol(x,
33.
Appendix A. Asymptotic lattice points counting in convex bodies
In this appendix, we will prove the following proposition.
Proposition 39. For any positive number ǫ, there exists k0 = k0 (ǫ, n) such that for any
closed convex body ∆ ⊂ [0, 1]n and any integer k ≥ k0 , we have
(A.1)
#(k∆ ∩ Zn )
− vol(∆) ≤ ǫ.
kn
Proof. We do induction on dimensions. If n = 1, then k∆ is a closed interval of length
kvol(∆), hence we know
kvol(∆) − 1 ≤ #(k∆ ∩ Z) ≤ kvol(∆) + 1.
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
25
So (A.1) holds for k0 = ⌈1/ǫ⌉.
Next, assume that the proposition is true for dimension n − 1. Denote by (x1 , · · · , xn )
the coordinates of Rn . Let ∆t := ∆ ∩ {xn = t} be the sectional convex body in [0, 1]n−1 .
Let [t− , t+ ] be the image
R t+ of ∆ under the projection onto the last coordinate. Then we
know that vol(∆) = t− vol(∆t )dt. By induction hypothesis, there exists k1 ∈ N such
that
ǫ
#(k∆t ∩ Zn−1 )
ǫ
vol(∆t ) − ≤
≤ vol(∆t ) +
for any k ≥ k1 .
n−1
3
k
3
It is clear that
X
#(k∆ ∩ Zn ) =
#(k∆t ∩ Zn−1 ),
t∈[t− ,t+ ]∩ k1 Z
so for any k ≥ k1 we have
(A.2)
#(k∆ ∩ Zn ) − k n−1 ·
X
t∈[t− ,t+ ]∩ k1 Z
ǫ
1
2ǫ
vol(∆t ) ≤ k n−1 · #([t− , t+ ] ∩ Z) ≤ k n .
3
k
3
Next, we know that the function t 7→ vol(∆t )1/(n−1) is concave on [t− , t+ ] by the BrunnMinkowski theorem. In particular, we can find t0 ∈ [t− , t+ ] such that g(t) := vol(∆t )
reaches its maximum at t = t0 . Hence g is increasing on [t− , t0 ] and decreasing on [t0 , t+ ].
Then applying Proposition 40 to g|[t− ,t0 ] and g|[t0,t+ ] respectively yields
Z
Z
t0
vol(∆t )dt −
t−
t+
t0
vol(∆t )dt −
1
k
1
k
Since 0 ≤ vol(∆t0 ) ≤ 1, we have
Z t+
1
vol(∆t )dt −
(A.3)
k
t−
X
vol(∆t ) ≤
2
,
k
X
vol(∆t ) ≤
2
.
k
X
vol(∆t ) ≤
5
.
k
t∈[t− ,t0 ]∩ k1 Z
t∈[t0 ,t+ ]∩ k1 Z
t∈[t+ ,t− ]∩ k1 Z
Therefore, by setting k0 = max(k1 , ⌈15/ǫ⌉), the inequality (A.1) follows easily by combining (A.2) and (A.3).
Proposition 40. For any monotonic function g : [a, b] → [0, 1] and any k ∈ N, we have
Z b
1 X
2
g(s)ds −
g(t) ≤ .
k
k
a
1
t∈[a,b]∩ k Z
and bk := ⌊kb⌋
,
Proof. We may assume that g is an increasing function. Denote ak := ⌈ka⌉
k
k
R
t
1
1
so [a, b] ∩ k Z = [ak , bk ] ∩ k Z. Since t−1/k g(s)ds ≤ g(t)/k whenever t ∈ [ak + 1/k, bk ], we
have
Z bk
X
1 X
1
g(t) ≤
g(t),
g(s)ds ≤
k
k
ak
1
1
t∈[ak +1/k,bk ]∩ k Z
t∈[a,b]∩ k Z
26
HAROLD BLUM AND YUCHEN LIU
Similarly,
R t+1/k
t
g(s)ds ≥ g(t)/k for any t ∈ [ak , bk − 1/k], we have
Z bk
X
1
1 X
1
g(s)ds ≥
g(t) ≥
g(t) − .
k
k
k
ak
1
1
t∈[ak ,bk −1/k]∩ k Z
t∈[a,b]∩ k Z
It is clear that ak ∈ [a, a + 1/k] and bk ∈ [b − 1/k, b], so we have
Z b
Z bk
Z b
Z bk
2
g(s)ds ≥
g(s)ds − ,
g(s)ds ≤
g(s)ds.
k
ak
a
ak
a
As a result, we have
1
k
X
t∈[a,b]∩ k1 Z
1
g(t) − ≤
k
Z
b
g(s)ds ≤
a
1
k
X
t∈[a,b]∩ k1 Z
g(t) +
2
k
Appendix B. Families of Ideals and the Hilbert–Samuel Function
The following proposition concerns the behavior of the Hilbert–Samuel function along
a family of ideals. The statement is not new. The proof we give follows arguments found
found in [FM00].
Definition 41. If (R, m) is a local ring and I is an m-primary ideal, then the Hilbert–
Samuel function of I, denoted HI : N → N, is given by HI (m) := ℓR (R/I m ). Note that
e(I) = limn→∞ HI (m)/mn , where n = dim(R).
Proposition 42. Let π : X → T be a morphism of finite type k-schemes. Assume T is
integral and π has a section σ : T → X . If a ⊂ OX is an ideal and at = a · OXσ(t) is
mσ(t) -primary for all t ∈ T , then T has a filtration
∅ = T0 ⊂ T1 ⊂ · · · T1 ⊂ Tm = T
such that for every 1 ≤ i ≤ m, Ti is closed in T and the function Ti \ Ti−1 ∋ t 7→ Hat is
constant.
Proof. To prove the result, it is sufficient to show that there exists a nonempty open set
proceed to find such a set U.
U ⊂ T such that Hat is constant for all
P t ∈ U. iWei+1
ℓ(a
/a
For each t ∈ T , we have Hat (m) = m−1
t ). Therefore, we consider the finitely
t
i=0
i
i+1
generated OX -algebra gra := ⊕i≥0 a /a . By generic flatness, we may choose a nonempty
open set U ⊂ T such that both OX |π−1 (U ) and gra |π−1 (U ) are flat over U.
For each i ∈ N, the function U ∋ t 7→ dimκ(t) (ai /ai+1 |t ) is constant, since each ai /ai+1
is flat over U and ai /ai+1 |t has zero dimensional support for each t ∈ U. Since κ(t) ≃
Oσ(t),X /mσ(t) , dimκ(t) (ai /ai+1 |t ) = ℓ(ai /ai+1 |t ) for all t ∈ T . Furthermore, Lemma 43,
proved below, implies ai /ai+1 |t = ait /ai+1
for all t ∈ U. Therefore, U ∋ t 7→ ℓ(ait /ai+1
t
t ) is
constant, and the proof is complete.
Before stating the following lemma, we introduce some notation. Let A be a ring, I ⊂ A
an ideal, and M an A-module. We set
M I mM
grI (M) :=
.
I m+1 M
m≥0
NORMALIZED VOLUME OF A SINGULARITY IS LOWER SEMICONTINUOUS
27
Lemma 43. Let B → A be a morphism of rings, I ⊂ A an ideal, and M ∈ Mod(A). If
grI M and M are both flat over B, then for any N ∈ Mod(B)
(grI M) ⊗B N ≃ grI (M ⊗B N).
Proof. We follow the argument given in [FM00]. Consider the surjective map αm :
(I m M) ⊗B N → I m (M ⊗B N). We claim that, for each m ∈ Z>0 , αm is injective and
I m M is flat over B.
In order to prove the claim, we induct on m. The claim holds when m = 0, since α0
is clearly an isomorphism and M is flat over B by assumption. Next, consider the exact
sequence
0 → I m+1 M → I m M → I m M/I m+1 M → 0
and assume the claim holds for a positive integer m. Since I m M and I m M/I m+1 M are
flat over B, so is I m+1 M. By the flatness of I m M/I m+1 M, we may tensor by N to get an
exact sequence
0 → I m+1 M ⊗B N → I m M ⊗B N → I m M/I m+1 M ⊗B N → 0.
By the above exact sequence, the injectivity of αm implies the injectivity of αm+1 . Now
that the claim has been proven, the lemma follows from applying the claim to the previous
short exact sequence.
References
[Amb16] F. Ambro: Variation of log canonical thresholds in linear systems. Int. Math. Res. Not. (2016),
no. 14, 4418–4448.
[BGO17] B. Bhatt, O. Gabber and M. Olsson: Finiteness of étale fundamental groups by reduction
modulo p. Preprint available at arXiv:1705.07303.
[Blu16] H. Blum: Existence of valuations with smallest normalized volume. To appear in Compos. Math.,
available at arXiv:1606.08894.
[BdFFU15] S. Boucksom, T. de Fernex, C. Favre and S. Urbinati: Valuation spaces and multiplier ideals
on singular varieties. Recent advances in algebraic geometry, 29-51, London Math. Soc. Lecture Note
Ser., 417, Cambridge Univ. Press, Cambridge, 2015.
[BFJ14] S. Boucksom, C. Favre, and M. Jonsson: A refinement of Izumi’s theorem. In Valuation theory
in interaction, 55-81, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich (2014).
[Cut13] S. Cutkosky: Multiplicities associated to graded families of ideals. Algebra Number Theory 7
(2013), no. 9, 2059-2083.
[dFEM04] T. de Fernex, L. Ein and M. Mustaţă: Multiplicities and log canonical threshold. J. Algebraic
Geom. 13 (2004), no. 3, 603-615.
[Don15] S. Donaldson: Algebraic families of constant scalar curvature Kähler metrics. Surveys in differential geometry 2014. Regularity and evolution of nonlinear equations, 111-137, Surv. Differ. Geom.,
19, Int. Press, Somerville, MA, 2015.
[ELS03] L. Ein, R. Lazarsfeld and K. Smith: Uniform approximation of Abhyankar valuation ideals in
smooth function fields. Amer. J. Math. 125 (2003), no. 2, 409-440.
[FM00] H. Flenner and M. Manaresi: Equimultiplicity and equidimensionality of normal cones. In: Recent
progress in intersection theory (Bologna, 1997). Trends Math., 199–215.
[Fuj15] K. Fujita: Optimal bounds for the volumes of Kähler-Einstein Fano manifolds. To appear in
Amer. J. Math., available at arXiv:1508.04578.
[HS16] H. Hein and S. Sun: Calabi-Yau manifolds with isolated conical singularities. To appear in Publ.
Math. Inst. Hautes Études Sci., available at arXiv:1607.02940.
[JM12] M. Jonsson and M. Mustaţă: Valuations and asymptotic invariants for sequences of ideals. Ann.
Inst. Fourier (Grenoble) 62 (2012), no. 6, 2145-2209 (2013).
28
HAROLD BLUM AND YUCHEN LIU
[KK14] K. Kaveh and A. Khovanskii: Convex bodies and multiplicities of ideals. Proc. Steklov Inst. Math.
286 (2014), no. 1, 268-284.
[Kol96] J. Kollár: Rational curves on algebraic varieties. Ergebnisse der Mathematik und ihrer Grenzgebiete, 32. Springer-Verlag, Berlin, 1996.
[Kol04] J. Kollár: Seifert Gm -bundles. Preprint available at arXiv:math/0404386.
[Kol13] J. Kollár: Singularities of the minimal model program, volume 200 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2013. With a collaboration of S. Kovács.
[Laz04] R. Lazarsfeld: Positivity in algebraic geometry. I-II. Ergebnisse der Mathematik und ihrer Grenzgebiete, 48–49. Springer-Verlag, Berlin, 2004.
[LM09] R. Lazarsfeld and M. Mustaţă: Convex bodies associated to linear series. Ann. Sci. Éc. Norm.
Supér. (4) 42 (2009), no. 5, 783-835.
[Lec60] C. Lech: Note on multiplicities of ideals. Ark. Mat. 4 (1960), 63-86.
[Li15a] C. Li: Minimizing normalized volumes of valuations. To appear in Math. Z., available at
arXiv:1511.08164.
[Li15b] C. Li: K-semistability is equivariant volume minimization. To appear in Duke Math. J., available
at arXiv:1512.07205.
[Li17] C. Li: Yau-Tian-Donaldson correspondence for K-semistable Fano manifolds. J. Reine Angew.
Math. 733 (2017), 55-85.
[LL16] C. Li and Y. Liu: Kähler-Einstein metrics and volume minimization. To appear in Adv. Math.,
available at arXiv:1602.05094.
[LWX14] C. Li, X. Wang and C. Xu: On proper moduli space of smoothable Kähler-Einstein Fano varieties. Preprint available at arXiv:1411.0761v3.
[LX16] C. Li and C. Xu: Stability of valuations and Kollár components. Preprint available at
arXiv:1604.05398.
[LX17] C. Li and C. Xu: Stability of valuations: higher rational rank. Preprint available at
arXiv:1707.05561.
[Liu16] Y. Liu: The volume of singular Kähler-Einstein Fano varieties. To appear in Compos. Math.,
available at arXiv:1605.01034.
[LiuX17] Y. Liu and C. Xu: K-stability of cubic threefolds. Preprint available at arXiv:1706.01933.
[Mat80] H. Matsumura: Commutative algebra. Second edition. Mathematics Lecture Note Series, 56.
Benjamin/Cummings Publishing Co., Inc., Reading, Mass., 1980. xv+313 pp.
[Mus02] M. Mustaţǎ. On multiplicities of graded sequences of ideals. J. Algebra 256 (2002), 229–249.
[Oda13] Y. Odaka: On the moduli of Kähler-Einstein Fano manifolds. Proceeding of Kinosaki algebraic
geometry symposium, 2013.
[Oda15] Y. Odaka: Compact moduli spaces of Kähler-Einstein Fano varieties. Publ. Res. Inst. Math. Sci.
51 (2015), no. 3, 549-565.
[SS17] C. Spotti and S. Sun: Explicit Gromov-Hausdorff compactifications of moduli spaces of KählerEinstein Fano manifolds. Preprint available at arXiv:1705.00377.
[SSY16] C. Spotti, S. Sun and C. Yao: Existence and deformations of Kähler-Einstein metrics on smoothable Q-Fano varieties. Duke Math. J. 165 (2016), no. 16, 3043-3083.
[SPA] The Stacks Project Authors: Stracks Project. Available at http://stacks.math.columbia.edu.
[Xu14] C. Xu: Finiteness of algebraic fundamental groups. Compos. Math. 150 (2014), no. 3, 409-414.
[Xu17] C. Xu: Interaction between singularity theory and the minimal model program. Submitted to 2018
ICM Algebraic and Complex Geometry Session, available at arXiv:1712.01041.
Department of Mathematics, University of Michigan, Ann Arbor, MI 48109–1043, USA
E-mail address: [email protected]
Department of Mathematics, Yale University, New Haven, CT 06511, USA.
E-mail address: [email protected]
| 0math.AC
|
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL
IDEALS
arXiv:1311.5927v1 [math.CO] 22 Nov 2013
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Abstract. The critical ideals of a graph are the determinantal ideals of the generalized Laplacian
matrix associated to a graph. In this article we provide a set of minimal forbidden graphs for the set
of graphs with at most three trivial critical ideals. Then we use these forbidden graphs to characterize
the graphs with at most three trivial critical ideals and clique number equal to 2 and 3.
1. Introduction
Given a connected graph G = (V, E) and a set of indeterminates XG = {xu : u ∈ V (G)}, the
generalized Laplacian matrix L(G, XG ) of G is the matrix with rows and columns indexed by the
vertices of G given by
(
xu
if u = v,
L(G, XG )uv =
−muv otherwise,
where muv is the number of the edges between vertices u and v of G.
Definition 1.1. For all 1 ≤ i ≤ |V |, the i-th critical ideal of G is the determinantal ideal given by
Ii (G, XG ) = h{det(m) : m is an i × i submatrix of L(G, XG )}i ⊆ Z[XG ].
We say that a critical ideal is trivial when it is equal to h1i. Critical ideals were firstly defined
in [8] and have been studied in a more general framework in [2]. The algebraic co-rank γ(G) of G
is the number of trivial critical ideals of G. The algebraic co-rank allows to separate the set of all
simple connected graphs in the following graph families:
Γ≤i = {G : G is a simple connected graph with γ(G) ≤ i}.
In [8] it was proven that if H is an induced subgraph of G, then Ii (H, XH ) ⊆ Ii (G, XG ) for all
i ≤ |V (H)|. Thus γ(H) ≤ γ(G). This implies that the set Γ≤i is closed under induced subgraphs.
Therefore we can say that a graph G is forbidden for Γ≤k when γ(G) ≥ k + 1. Moreover, we define
the set of minimal forbidden as follows:
Definition 1.2. Let Forb(Γ≤k ) be the set of minimal (under induced subgraphs property) forbidden
graphs for Γ≤k .
A graph G is called γ-critical if γ(G − v) < γ(G) for all v ∈ V (G). Then Forb(Γ≤k ) is equal to
the set of γ-critical graphs with γ(G) ≥ k + 1 and γ(G − v) ≤ k for all v ∈ V (G). Given a family F
of graphs, a graph G is called F-free if no induced subgraph of G is isomorphic to a member of F.
Thus G ∈ Γ≤k if and only if G is Forb(Γ≤k )-free. Therefore, characterizing Forb(Γ≤k ) leads to a
characterization of Γ≤k .
2000 Mathematics Subject Classification. Primary 05C25; Secondary 05C50, 05E99.
Key words and phrases. Critical ideal, Critical group, Laplacian matrix, Forbidden induced subgraph.
Both authors were partially supported by CONACyT grant 166059, the first author was partially supported by
CONACyT and the second author was partially supported by SNI.
1
2
CARLOS A. ALFARO AND CARLOS E. VALENCIA
In [1], these ideas were used to obtain a characterization of Γ≤1 and Γ≤2 . More precisely, it was
found that Forb(Γ≤1 ) = {P3 } and thus Γ≤1 consists of the complete graphs. On the other hand,
Forb(Γ≤2 ) consists of 5 graphs: P4 , K2 \ P2 , K6 \ M2 , cricket and dart. And Γ≤2 consists of the
graphs isomorphic to an induced subgraph of one of the following graphs: tripartite complete graph
Kn1 ,n2 ,n3 or Tn1 ∨ (Kn2 + Kn3 ), where G + H denote the disjoint union of the graphs of G and H,
and G ∨ H denote the join of G and H.
The main goal of this paper is to provide a set of minimal forbidden graphs for Γ≤3 and a partial
description of Γ≤3 . Specifically, we will characterize the graphs in Γ≤3 with clique number equal
to 2 and 3. Therefore, we prove that if a graph G ∈ Γ≤3 has clique number at most 3, then G is
an induced subgraph of one graph in the family F1 (see Figure 1). The converse is stronger, each
induced subgraph of a graph in F1 belongs to Γ≤3 , but not all graphs in F11 have clique number less
than 4.
n5
n2
n1
n7
n1
n6
n4
n3
(i) graph G1
(ii) family of graphs F11
(iii) family of graphs F12
Figure 1. The family of graphs F1 . A black vertex represents a clique of cardinality
nv , a white vertex represents a stable set of cardinality nv and a gray vertex represents
a single vertex.
This article is divided as follows. In Section 1.1, we will show how the characterization of Γ≤i
leads to a characterization of the graphs having critical group with i invariant factors equal to 1.
In Section 2, we will construct a graph Gd by replacing its vertices by cliques and stable sets. And
we will show a novel method to verify whether the k-th critical ideal of Gd is trivial or not. This
method will be applied to prove that each induced subgraph of a graph in the family F1 belongs to
Γ≤3 . In Section 3, we will give a family of minimal forbidden graphs for Γ≤3 , and it will be used to
prove that a graph G ∈ Γ≤3 has clique number 2 if and only if G is an induced subgraph of a graph
in F2 (see Figure 3). Section 4 is devoted to prove that if a graph G ∈ Γ≤3 and has clique number
3, then G is an induced subgraph of one graph in the family F1 .
1.1. Applications to the critical group. The Laplacian matrix L(G) of G is the evaluation of
L(G, XG ) at XG = DG , where DG is the degree vector of G. By considering L(G) as a linear map
L(G) : ZV → ZV , the cokernel of L(G) is the quotient module ZV /Im L(G). The torsion part of
this module is the critical group K(G) of G. The critical group has been studied intensively on
several contexts over the last 30 years: the group of components [14, 15], the Picard group [3, 4],
the Jacobian group [3, 4], the sandpile group [7], chip-firing game [4, 17], or Laplacian unimodular
equivalence [10, 18].
It is known [13, Theorem 1.4] that the critical group of a connected graph G with n vertices can
be described as follows:
K(G) ∼
= Zd1 ⊕ Zd2 ⊕ · · · ⊕ Zdn−1 ,
where d1 , d2 , ..., dn−1 are positive integers with di | dj for all i ≤ j. These integers are called invariant
factors of the Laplacian matrix of G. Besides, if ∆i (G) is the greatest common divisor of the i-minors
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
3
of the Laplacian matrix L(G) of G, then the i-th invariant factor di is equal to ∆i (G)/∆i−1 (G), where
∆0 (G) = 1 (for details see [12, Theorem 3.9]).
Definition 1.3. Given an integer number k, let fk (G) be the number of invariant factors of the
critical group of G equal to k. Also, let Gi = {G : G is a simple connected graph with f1 (G) = i}.
The study and characterization of Gi is of great interest. In particular, some results and conjectures
on the graphs with cyclic critical group can be found in [15, Section 4] and [20, Conjectures 4.3 and
4.4]. On the other hand, it is easy to see [7, 14, 18] that G1 consists only of the complete graphs.
Besides, several people (see [17]) posed interest on the characterization of G2 and G3 . In this sense,
few attempts have been done. In [19] it was characterized the graphs in G2 whose third invariant
factor is equal to n, n − 1, n − 2, or n − 3. Later in [6] the characterizations of the graphs in G2
with a cut vertex, and the graphs in G2 with number of independent cycles equal to n − 2 are given.
Recently, a complete characterization of G2 was given in [1]. However, nothing is known about G3 .
A crucial result linking critical groups and critical ideals is [8, Theorem 3.6], which states if DG
is the degree vector of G, and d1 | · · · | dn−1 are the invariant factors of K(G), then
+
* i
Y
dj = h∆i (G)i for all 1 ≤ i ≤ n − 1.
Ii (G, DG ) =
j=1
Thus if the critical ideal Ii (G, XG ) is trivial, then ∆i (G) and di are equal to 1. Equivalently, if
∆i (G) and di are not equal to 1, then the critical ideal Ii (G, XG ) is not trivial.
The critical ideals behave better than critical group under induced subgraph property. It is not
difficult to see that unlike of Γ≤k , Gk is not closed under induced subgraphs. For instance, the cone
of the claw graph c(K1,3 ) belongs to G2 , but the claw graph K1,3 belongs to G3 . Also, K6 \ {2P2 }
belongs to G3 , meanwhile K5 \ {2P2 } belongs to G2 . Moreover, if H is an induced subgraph of G,
then it is not always true that K(H) E K(G). For example, K(K4 ) ∼
= Z35 .
= Z24 5 K(K5 ) ∼
On the other hand, a consequence of [8, Theorem 3.6] is that Gi ⊆ Γ≤i for all i ≥ 0. Therefore,
after an analysis of the i-th invariant factor of the Laplacian matrix of the graphs in Γ≤i , the
characterization of Gi can be obtained. See for instance [1] for the characterizations of G1 and G2 .
2. Cliques, stable sets and critical ideals
Let G = (V, E) a simple graph. Suppose that V ′ is a subset of V , the induced subgraph G[V ′ ]
of G is the subgraph of G whose vertex set is V ′ and whose edge set is the set of those edges of G
that have both ends in V ′ . If E ′ is a subset of E, the edge-induced subgraph G[E ′ ] is the subgraph
of G whose edge set is E ′ and whose vertex set consists of all ends of edges of E ′ . Let P, Q be two
subsets of V , we denote by E(P, Q) the set of edges of G with one end in P and the other end in
Q. A clique of G is a subset S of V of mutually adjacent vertices, and the maximum size of a clique
of G is the clique number ω(G) of G. A subset S of V is called an independent set, or stable set, of
G if no two vertices of S are adjacent in G. The graph with n vertices whose vertex set induces a
stable set is called the trivial graph, and denoted by Tn . The cardinality of a maximum stable set
in G is called the stability number of G and is denoted by α(G).
Given a simple graph G = (V, E) and a vector d ∈ ZV , the graph Gd is constructed as follows.
For each vertex u ∈ V is associated a new vertex set Vu , where Vu is a clique of cardinality −du if
du is negative, and Vu is a stable set of cardinality du if du is positive. And each vertex in Vu is
adjacent with each vertex in Vv if and only if u and v are adjacent in G. Then the graph G is called
the underlying graph of Gd .
4
CARLOS A. ALFARO AND CARLOS E. VALENCIA
A convenient way to visualize Gd is by means of a drawing of G, where the vertex u is colored in
black if du is negative, and colored in white if du is positive. We indicate the cardinality of Vu by
writing it inside the drawing of vertex u. When |du | = 1, we may color u in gray and avoid writing
the cardinality (see Figure 1). On the other hand, it will be useful to avoid writing the cardinality
when |du | = 2 (see Figure 2).
In general, the computation of the Gröbner bases of the critical ideals is more than complicated.
However, in the rest of this section we will show a novel method, developed in [2], to decide for
i ≤ |V (G)| whether the i-th critical ideal of Gd is trivial or not.
′
For V ′ ⊆ V (G) and d ∈ ZV , we define φ(d) as follows:
(
0
if dv > 0,
φ(d)v =
−1 if dv < 0.
Theorem 2.1. [2, Theorem 3.7] Let n ≥ 2 and G be a graph with n vertices. For V ′ ⊆ V (G),
′
1 ≤ j ≤ n and d ∈ ZV , the critical ideal Ij (Gd , XGd ) is trivial if and only if the evaluation of
Ij (G, XG ) at XG = φ(d) is trivial.
Thus the procedure of verify whether a family of graphs belongs to Γ≤i becomes in an evaluation
of the i-th critical ideal of the underlying graph of the family.
Let G2 be the underlying graph of the family of graphs F11 (see Figure 1.ii) with vertex set
V = {v1 , v2 , v3 , v4 , v5 , v6 , v7 }. Let d ∈ ZV such that d1 , d2 , d3 are positive integers and d4 , d5 , d6 , d7
are negative integers. Thus φ(d) = (0, 0, 0, −1, −1, −1, −1). By using a computer algebra system we
can check that
I4 (G2 , XG2 ) = h2, x1 , x2 , x3 , x4 + 1, x5 + 1, x6 + 1, x7 + 1i.
Since the evaluation I4 (G2 , XG2 ) at XG2 = φ(d) is equal to h2i, then by Theorem 2.1 the critical
ideal I4 (Gd2 , XGd ) is non-trivial. Therefore, each graph in this family of graphs has algebraic co-rank
2
at most 3. Now let G3 be the underlying graph of the family of graphs F12 (see Figure 1.iii) with
vertex set V = {v1 , v2 , v3 , v4 , v5 , v6 }. Let d1 be positive integer. By using a computer algebra system
we can check that
I4 (G3 , XG3 ) = hx21 + 5x1 + 5, x1 + x2 + 3, x1 + x3 + 3, x1 + x4 + 3, x1 + x5 + 3, x1 + x6 + 3i.
Since the evaluation of I4 (G3 , XG3 ) at x1 = 0 is non-trivial, then by Theorem 2.1 the critical ideal
I4 (Gd3 , XGd ) is non-trivial. Therefore, each graph in F12 has algebraic co-rank at most 3.
3
On the other hand, it can be verified that γ(G1 ), γ(G2 ) and γ(G3 ) are equal to 3. Then the
graphs in F1 have algebraic co-rank 3, and so each induced subgraph of a graph in F1 has algebraic
co-rank at most 3.
Proposition 2.2. Each graph in F1 belongs to Γ≤3 .
3. A description of Γ≤3
It is possible to compute the algebraic co-rank of all connected graphs with at most 9 vertices
using the software Macaulay2 [9] and Nauty [16]. The computation of the algebraic co-rank of the
connected graphs with at most 8 vertices required at most 3 hours on a MacBookPro with a 2.8 GHz
Intel i7 quad core processor and 16 GB RAM. Besides, the computation of the algebraic co-rank of
the connected graphs with 9 vertices required 4 weeks of computation on the same computer.
Let F3 be the family of graphs shown in Figure 2. This family represents the graphs in Forb(Γ≤3 )
with at most 8 vertices. Since there exists no minimal forbidden graph with 9 vertices for Γ≤3 , then
it is likely that F3 = Forb(Γ≤3 ).
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
P5
G6,1
G6,2
G6,3
G6,4
G6,5
G6,6
G6,7
G6,8
5
G6,9
G6,10
G6,11
G6,12
G6,13
G6,14
G6,15
G6,16
G6,17
G6,18
G6,19
G6,20
G6,21
G6,22
G6,23
G6,24
G6,25
G6,26
G6,27
G7,1
G7,2
G7,3
G7,4
G7,13
G7,14
G7,5
G7,6
G7,15
G7,7
G7,8
G7,16 G7,17
G7,9
G8,1
G7,10
G8,2
G7,11
G8,3
G7,12
G8,4
Figure 2. The family of graphs F3 . A black vertex represents a clique of cardinality
2, a white vertex represents a stable set of cardinality 2 and a gray vertex represent
a single vertex.
Proposition 3.1. Each graph in F3 belongs to Forb(Γ≤3 ).
Proof. It can easily checked, using a computer algebra system, that each graph in F3 is γ-critical
and has algebraic co-rank equal to 4.
One of the main results of this article is the following:
Theorem 3.2. If a graph G ∈ Γ≤3 has clique number at most 3, then G is an induced subgraph of
a graph in F1 .
We divide the proof in two characterizations: the graphs in Γ≤3 with clique number equal to 2 and
3. The converse of Theorem 3.2 is stronger, by Proposition 2.2 we have that each induced subgraph
of a graph in F1 belongs to Γ≤3 . However, it is not difficult to recognize the graphs in F11 with
clique number greater or equal than 4.
Theorem 3.3. Let G be a simple connected graph with ω(G) = 2. Then, G is F3 -free if and only if
G is isomorphic to an induced subgraph of a graph in F2 (see Figure 3).
n1
n2
(i) Kn1 ,n2
n1
n2
(ii) F21
n1
n2
(iii) F22
n1
n2
(iv) F23
Figure 3. The family of graphs F2 . A white vertex represents a stable set of
cardinality nv and a gray vertex represents a single vertex.
6
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Since each graph in F2 is isomorphic to an induced subgraph of a graph in F11 , then Proposition
2.2 implies that each graph in F2 belongs to Γ≤3 . Note that the graphs in Γ≤1 and Γ≤2 are induced
subgraph of a graph in F11 (see Figure 1.ii).
Theorem 3.4. Let G be a simple connected graph with ω(G) = 3. Then, G is F3 -free if and only if
G is isomorphic to an induced subgraph of a graph in F1 with clique number 3.
Section 4 is devoted to the proof of the Theorem 3.4. Now we give the proof of Theorem 3.3.
Proof of Theorem 3.3. Since each graph in F2 belongs to Γ≤3 , then each graph in F2 is F3 -free.
Thus, we get one implication.
Suppose G is F3 -free. Let a, b ∈ V (G) such that ab ∈ E(G). Since ω(G) = 2, then there is no
vertex adjacent with a and b at the same time. For a vertex v ∈ V (G), the neighbor set NG (v) of v
in G is the set of all vertices adjacent with v. Let A = NG (a) − b and B = NG (b) − a. Clearly, each
of the sets A and B induces a trivial graph. Let us define
A = {u ∈ A : ∃v ∈ B such that uv ∈ E},
A′ = {u ∈ A : ∄v ∈ B such that uv ∈ E}
B = {u ∈ B : ∃v ∈ A such that uv ∈ E}, and B ′ = {u ∈ B : ∄v ∈ A such that uv ∈ E}.
Thus we have two possible cases: when A and B are not empty and when A and B are empty.
First we consider when A and B are not empty. In this case we have the following statements:
Claim 3.5. One of the sets A′ or B ′ is empty, and the other has cardinality at most one.
Proof. Suppose A′ and B ′ are not empty. Let u ∈ A, v ∈ B, s ∈ A′ and t ∈ B ′ . Then the vertex
set {a, b, u, v, s, t} induces a graph isomorphic to G6,9 , which is impossible. Now suppose A′ has
cardinality more than 1. Take w1 , w2 ∈ A′ . Then {x, y, u, v, w1 , w2 } induces a graph isomorphic to
G6,3 ; a contradiction. Thus A′ has cardinality at most one.
Claim 3.6. The edge set E(A, B) induces either a complete bipartite graph or a complete bipartite
graph minus an edge.
Proof. First note that each vertex in A is incident with each vertex in B, except for at most one
vertex. It is because if u ∈ A and v1 , v2 , v3 ∈ B such that uv1 ∈ E(G) and uv2 , uv3 ∈
/ E(G), then
the induced subgraph G[{x, y, u, v1 , v2 , v3 }] is isomorphic to G6,3 ; which is impossible. In a similar
way, each vertex in B is incident with each vertex in A, except for at most one vertex. Thus the
edge set E(A, B) must be equal to the edge set {uv : u ∈ A, v ∈ B} minus a matching. In fact,
the cardinality of this matching must be at most one. Otherwise, if u, v ∈ A and s, t ∈ B such that
us, vt ∈
/ E(G) and ut, vs ∈ E(G), then the induced subgraph G[{t, u, a, v, s}] is isomorphic to P5 ;
which is a contradiction.
Claim 3.7. It is not possible that, at the same time, the edge set E(A, B) induces a complete bipartite
graph minus an edge and A′ ∪ B ′ 6= ∅.
Proof. Suppose both situations occur at the same time. Let u1 , u2 ∈ A and v1 , v2 ∈ B such that
u1 v1 , u1 v2 , u2 v1 ∈ E(G) and u2 v2 ∈
/ E(G). And let w ∈ A′ . Then the vertex set {u1 , u2 , v1 , v2 , y, w}
induces a graph isomorphic to G6,9 ; which is a contradiction.
Thus there are three possible cases:
(a) E(A, B) = {uv : u ∈ A, v ∈ B} and A′ ∪ B ′ = ∅,
(b) E(A, B) = {uv : u ∈ A, v ∈ B} and A′ = T1 , B ′ = ∅ or
(c) E(A, B) induces a bipartite complete graph minus an edge and A′ ∪ B ′ = ∅.
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
7
Let V∅ denote the set of vertices that are not adjacent with a nor b. In what follows we will
describe the vertex set V∅ .
Case (a). Let w1 , w2 ∈ V∅ , u1 , u2 ∈ A, and v1 ∈ B. Note that it is not possible that a vertex in V∅
is adjacent with a vertex in A and a vertex in B at the same time, because then we get ω(G) ≥ 3.
Moreover
Claim 3.8. There exist no two vertices in V∅ such that one is adjacent with a vertex in A and the
other one is adjacent with a vertex in B.
Proof. Suppose w1 u1 , w2 v1 ∈ E(G). There are two cases: either w1 w2 ∈ E(G) or w1 w2 ∈
/ E(G).
If w1 w2 ∈ E(G), then G[{w1 , w2 , v1 , b, a}] ≃ P5 ; which is impossible. And if w1 w2 ∈
/ E(G), then
G[{w1 , w2 , v1 , u1 , b, a}] ≃ G6,9 ; which is a contradiction.
Without loss of generality, suppose w1 ∈ V∅ is adjacent with u1 ∈ A. Thus
Claim 3.9. The vertex set V∅ has cardinality at most 1.
Proof. There are two possible cases: either each vertex in V∅ is adjacent with a common vertex
in A or not. Suppose there exists w2 ∈ V∅ such that u1 is adjacent with both w1 and w2 . Then
w1 w2 ∈
/ E(G), because otherwise w1 , w2 and u1 induce a K3 . Thus the vertex set {w1 , w2 , u1 , v1 , a, b}
induces a graph isomorphic to G6,3 ; which is forbidden. Then this case is not possible. Now suppose
w1 and w2 are not adjacent with a common vertex in A. We have the following possible cases:
(1) w1 u1 , w2 u2 ∈ E(G),
(2) w1 u1 , w2 u2 , w1 w2 ∈ E(G), or
(3) w1 w2 ∈ E(G).
This yields a contradiction since in case (1) the vertex set {w1 , u1 , v1 , u2 , w2 } induces a graph isomorphic to P5 , in case (2) the vertex set {w2 , w1 , u1 , v1 , b} induces a graph isomorphic to P5 , and in
case (3) the vertex set {w2 , w1 , u1 , v1 , b} induces a graph isomorphic to P5 .
If |A| = 2, there are two possibilities: either w1 is adjacent with each vertex in A or w1 is adjacent
with only one vertex of A. If |A| ≥ 3, then w1 is adjacent with either all vertex in A or only one
vertex in A. It is because if w1 is adjacent with both u1 , u2 ∈ A and w is not adjacent with u3 ∈ A,
then the vertex set {w1 , u1 , u2 , u3 , a, b} induces a graph isomorphic to G6,3 . Note that when w1 is
adjacent with each vertex in A, then the graph is isomorphic to an induced subgraph of a graph in
F22 . Meanwhile, when w1 is adjacent only with one vertex of A, then the graph is isomorphic to an
induced subgraph of a graph in F21 . Finally, when V∅ = ∅, the graph is a complete bipartite graph.
Case (b). Suppose without loss of generality that A′ 6= ∅. By Claims 3.8 and 3.9, the vertex set
V∅ has cardinality at most one. Let w ∈ V∅ , u1 ∈ A, u2 ∈ A′ and v1 ∈ B. We have two cases: w
is adjacent with either a vertex in A or a vertex in B. Let us consider when w is adjacent with a
vertex in A. Here we have two possibilities: either wu2 ∈ E(G) or wu2 ∈
/ E(G). However, none of
the two cases is allowed, since in the first case we get that the vertex set {w, u2 , a, b, v1 } induces a
graph isomorphic to P5 , and in the second case the vertex set {w, u2 , a, b, u1 , v1 } induces a graph
isomorphic to G6,9 . Thus the remaining case is that w is adjacent with a vertex in B. In this case w
must be adjacent with u2 and each vertex in B, because otherwise the graph P5 appears as induced
subgraph. Note that this graph is isomorphic to an induced subgraph of a graph in F23 .
Case(c). Let w ∈ V∅ , u1 , u2 ∈ A, and v1 , v2 ∈ B such that u1 v2 ∈
/ E(G) and u1 v1 , u2 v1 , u2 v2 ∈ E(G).
Note that if w is adjacent with u1 or v2 , then w is adjacent with both u1 and v2 , because since
u1 v2 ∈
/ E(G), we would obtain P5 as induced subgraph; which is not possible. In this case, when w
is adjacent with u1 and v2 , the graph is isomorphic to an induced subgraph of F23 . Now we consider
8
CARLOS A. ALFARO AND CARLOS E. VALENCIA
when w is adjacent with a vertex in (A − u1 ) ∪ (B − v2 ). Without loss of generality, we can suppose
w is adjacent with u2 . The vertex w is not adjacent with v1 or v2 , because otherwise a clique of
cardinality 3 is obtained. On the other hand, w is not adjacent with u1 , because otherwise w would
be adjacent with both vertices u1 and v1 . Thus w is adjacent only with u2 , but the vertex set
{w, u1 , u2 , v2 , a, b, } induces a graph isomorphic to G6,9 ; a contradiction. Thus V∅ is empty, and the
graph is isomorphic to an induced subgraph of F22 .
Now we consider the case when A and B are empty. One of the vertex sets A′ or B ′ has
cardinality at most one, because otherwise A′ ∪ B ′ ∪ {a, b} would contain G6,1 as induced subgraph.
Thus, let us assume that A′ = {u} and |B ′ | > 1. Let V∅ denote the vertex set whose vertices are not
adjacent with both a and b.
Claim 3.10. If A′ and B ′ are not empty, and w ∈ V∅ is adjacent with a vertex in A′ ∪ B ′ , then w
is adjacent with each vertex in A′ ∪ B ′ .
Proof. Let v ∈ B ′ . Suppose one of the edges uw or vw does not exist. Then {w, u, a, b, v} induces a
graph isomorphic to P5 ; which is a contradiction.
B′
Note that the vertex set V∅ induces a stable set, because otherwise ω(G) > 2. Thus when A′ and
are not empty, the graph G is isomorphic to an induced subgraph of a graph in F23 .
Now consider the case when B = ∅ and |A| > 1.
Claim 3.11. Each vertex w ∈ V∅ is adjacent with either a unique vertex in A or each vertex in A.
Proof. The result is easy to check when |A| ≤ 2. So suppose A has cardinality greater or equal than
3. Let u1 , u2 , u3 ∈ A such that w is adjacent with both u1 , u2 , and w is not adjacent with u3 . Since
the vertex set {w, u1 , u2 , u3 , a, b} induces a graph isomorphic to G6,3 , then we get a contradiction
and the result follows.
Claim 3.12. Let w ∈ V∅ such that it is adjacent with u ∈ A. If V∅ has cardinality greater than 1,
then each vertex in V∅ is adjacent either with u or with each vertex in A.
Proof. Suppose there exists w′ ∈ V∅ such that w′ is not adjacent with u. Let u′ ∈ A such that w′ is
adjacent with w′ . There are four possible cases:
•
•
•
•
wu′ , ww′ ∈ E(G),
ww′ ∈ E(G) and wu′ ∈
/ E(G),
wu′ ∈ E(G) and ww′ ∈
/ E(G),
wu′ , ww′ ∈
/ E(G).
In the cases when ww′ ∈ E(G), the vertex set {w, u, a, u′ , w′ } induces a graph isomorphic to P5 ;
those cases do not occur. On the other hand if wu′ ∈ E(G) and ww′ ∈
/ E(G), then {w, w′ , u, u′ , a, b}
induces a graph isomorphic to G6,9 , and this case can not occur. Finally, if wu′ , ww′ ∈
/ E(G), then
′
′
the vertex set {w, u, a, u , w } induces a graph isomorphic to P5 . Which is a contradiction. Thus
each pair of vertices in V∅ are adjacent with the same vertices in A. And the result follows.
Thus there are two cases: when each vertex in V∅ is adjacent with a unique vertex u in A, and
when each vertex in V∅ is adjacent with each vertex in A. Note that V∅ induces a stable set, because
otherwise ω(G) > 2. In the first case we have that either |V∅ | ≥ 2 or the vertex set A − u is empty. It
is because otherwise G would have G6,1 as induced subgraph. Therefore, in each case G is isomorphic
to a graph in F11 .
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
9
4. Proof of Theorem 3.4
One implication is easy because Proposition 2.2 implies that each graph in F1 is F3 -free. The
other implication is much more complex.
Suppose G is F3 -free. Let W = {a, b, c} be a clique of cardinality 3. For each X ⊆ {a, b, c}, let
VX = {u ∈ V (G) : NG (u) ∩ {a, b, c} = X}. Note that V∅ denote the vertex set whose vertices are
not adjacent with a vertex in {a, b, c}.
Since ω(G) = 3, then the vertex set Va,b,c is empty, and for each pair {x, y} ⊂ {a, b, c}, the vertex
set Vx,y induces a stable set. Furthermore
Claim 4.1. For x ∈ {a, b, c}, the induced subgraph G[Vx ] is isomorphic to either Km,n , 2K2 , K2 +K1 ,
or Tn .
Proof. First consider K3 , P4 , K2 + T2 and P3 + T1 as induced subgraphs of G[Vx ]. Since G[K3 ∪ x] ≃
K4 , G[P4 ∪ {x, y}] ≃ G6,12 , G[K2 + T2 ∪ {a, b, c}] ≃ G7,1 , G[P3 + T1 ∪ {x, y}] ≃ G6,4 and all of them
are forbidden for G, then the graphs K3 , P4 , K2 + T2 and P3 + T1 are forbidden in G[Vx ]. Thus
ω(G[Vx ]) ≤ 2.
If ω(G[Vx ]) = 2, then there exist u, v ∈ Vx such that uv ∈ E(G[Vx ]). Clearly, NG[Vx ] (u) ∩
NG[Vx ] (v) = ∅, and each vertex set NG[Vx ] (u) and NG[Vx ] (u) induces a trivial subgraph. Since G[Vx ]
is P4 -free, then st ∈ E(Vx ) for all s ∈ NVx (u) \ {v} and t ∈ NVx (v) \ {u}. Therefore, each component
in G[Vx ] is a complete bipartite subgraph.
If a component in G[Vx ] has cardinality at least three, then G[Vx ] does not have another component, because the existence of another component makes that P3 + T1 appears as an induced
subgraph in G[Vx ]; which is impossible. If there is a component in G[Vx ] of cardinality at least two,
then there is at most another component in G[Vx ] since K2 + T2 is forbidden in G[Vx ]. And thus the
result turns out.
Claim 4.2. If there is x ∈ {a, b, c} such that the induced subgraph G[Vx ] has at least 2 components
or is isomorphic to a complete bipartite with at least three vertices, then the vertex set Vy is empty
for each y ∈ {a, b, c} − x.
Proof. Let G[Vx ] be as above. Then there are two vertices u, v ∈ Vx that are not adjacent. Suppose
Vy is not empty, that is, there is w ∈ Vy . There are three possibilities:
• uw, vw ∈ E(G),
• uw ∈ E(G) and vw ∈
/ E(G), and
• uw, vw ∈
/ E(G).
But in each case the vertex set {a, b, c, u, v, w} induces a graph isomorphic to G6,15 , G6,11 , and
G6,2 , respectively. Since these graphs are forbidden, then we obtain a contradiction and then Vy is
empty.
Claim 4.3. If there is x ∈ {a, b, c} such that G[Vx ] has at least 2 components or is isomorphic
to a complete bipartite with at least three vertices, then the vertex set Vx,y is empty for each y ∈
{a, b, c} − x.
Proof. Let G[Vx ] be as above. Then there are two vertices u1 , u2 ∈ Vx that are not adjacent. Suppose
Vx,y 6= ∅. Let v ∈ Vx,y . There are three possibilities:
• u1 v ∈ E(G) and u2 v ∈ E(G),
• u1 v ∈ E(G) and u2 v ∈
/ E(G), and
• u1 v ∈
/ E(G) and u2 v ∈
/ E(G).
10
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Then in each case, G[{a, b, c, u1 , u2 , v}] is isomorphic to G6,16 , G6,12 and G6,4 , respectively. Since
these graphs are forbidden, we have that Vx,y is empty.
Remark 4.4. Claims 4.2 and 4.3 imply that when Vx has connected 2 components or is a complete
bipartite graph of at least 3 vertices, then the only non-empty vertex sets are V∅ , Vx and Vy,z , where
x, y and z are different elements of {a, b, c} Moreover, by Claim 4.1, the vertex set Vx is one of the
following vertex sets:
• Tn , where n ≥ 2,
• complete bipartite graph with cardinality at least 3,
• K1 + K2 or 2K2 .
Next result describes the induced subgraph G[Va ∪ Vb ∪ Vc ] when each set Vx is connected of
cardinality at most 2.
Claim 4.5. Suppose for each x ∈ {a, b, c} the vertex set Vx is connected of cardinality at most 2. If
for all x ∈ {a, b, c} the set Vx is not empty, then G[Va ∪ Vb ∪ Vc ] is isomorphic (where x, y and z are
different elements of {a, b, c}) to one of the following sets:
• Vx ∨ (Vy + Vz ) where Vx = K1 , Vy = Km , Vz = Kn and m, n ∈ {1, 2},
• Vx ∨ (Vy + Vz ) where Vx = K2 , Vy = K1 , Vz = K1 , or
• Vx ∨ (Vy ∨ Vz ) where Vx = K1 , Vy = K1 , Vz = K1 .
If Vz = ∅, then G[Vx ∪ Vy ] is isomorphic to one of the following sets:
• Vx + Vy , where Vx = Km , Vy = Kn and m, n ∈ {1, 2}, or
• Vx ∨ Vy , where Vx = K1 , Vy = Km and m ∈ {1, 2}.
If Vy = Vz = ∅, then Vx is isomorphic to K1 or K2 .
Proof. It is not difficult to prove that either E(Vx , Vy ) is empty or E(Vx , Vy ) induces a complete
bipartite graph. The result follows by checking the possibilities with a computer algebra system.
In the rest of the proof, for each case obtained in Remark 4.4 and Claim 4.5, we will analyze the
remaining edges sets and the vertex set V∅ . As before, we may refer to x, y or z as different elements
of {a, b, c}. Each case will be consider in a subsection.
4.1. When Va ∪ Vb ∪ Vc = ∅. Now we describe the induced subgraph G[Va,b ∪ Va,c ∪ Vb,c ].
Claim 4.6. If the vertex sets Vx,y and Vy,z are not empty, and the edge set E(Vx,y , Vy,z ) is empty,
then |Vx,y | = |Vy,z | = 1.
Proof. Suppose |Vx,y | ≥ 2 and Vy,z 6= ∅. Take u, u′ ∈ Vx,y and v ∈ Vy,z . The result follows since the
vertex set {a, b, c, u, u′ , v} induces a graph isomorphic to the forbidden graph G6,16 .
Claim 4.7. If E(Vx,y , Vy,z ) 6= ∅, E(Vx,y , Vx,z ) = ∅ and E(Vy,z , Vx,z ) = ∅, then Vx,z is empty.
Proof. Let u ∈ Vx,y and v ∈ Vy,z such that uv ∈ E(G). Suppose there exists w ∈ Vx,z . Then
uw, vw ∈
/ E(G). Since the induced subgraph G[{a, b, c, u, w, v}] is isomorphic to G6,24 , then we get
a contradiction. Then Vx,z is empty.
Claim 4.8. If E(Vx,y , Vy,z ) 6= ∅, then E(Vx,y , Vy,z ) induces either a complete bipartite graph or a
complete bipartite graph minus an edge.
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
11
Proof. Let u ∈ Vx,y . Suppose there exist v, v ′ ∈ Vy,z such that uv, uv ′ ∈
/ E(G). Since the induced
subgraph G[{a, b, c, u, v, v ′ }] is isomorphic to G6,16 , which is forbidden, then the vertex u is adjacent
with at least all but one vertices in Vy,z . In a similar way, we have that each vertex in Vy,z is adjacent
with at at least all but one vertices in Vx,y . Therefore, the edge set E(Vx,y , Vy,z ) induces a complete
bipartite graph minus a matching. Now suppose this matching has cardinality greater or equal to 2.
Then there exist u, u′ ∈ Vx,y and v, v ′ ∈ Vy,z such that uv ′ , u′ v ∈
/ E(G) and uv, u′ v ′ ∈ E(G). Since
′
′
G[{v, u, x, u , v }] is isomorphic to P5 , then we get a contradiction and the matching has cardinality
at most 1. Therefore, E(Vxy , Vy,z ) induces a complete bipartite graph or a complete bipartite graph
minus an edge.
Claim 4.9. If E(Vx,y , Vy,z ) 6= ∅, E(Vy,z , Vx,z ) 6= ∅ and E(Vx,y , Vx,z ) = ∅, then Vx,y = {v1 }, Vx,z =
{v2 } and one of the following two statements holds:
• E(v1 , Vy,z ) = {v1 v : v ∈ Vy,z } and E(v2 , Vy,z ) = {v2 v : v ∈ Vy,z }, or
• there exists v3 ∈ Vy,z such that E(v1 , Vy,z ) = {v1 v : v ∈ Vy,z − v3 } and E(v2 , Vy,z ) = {v2 v :
v ∈ Vy,z − v3 }.
Proof. Since there is no edge joining a vertex in Vx,y and a vertex in Vx,z , then Claim 4.6 implies
that |Vx,y | = |Vx,z | = 1. Let Vx,y = {v1 } and Vx,z = {v2 }. By Claim 4.8, each edge set E(Vx,y , Vy,z )
and E(Vy,z , Vx,z ) induces either a complete bipartite graph or a complete bipartite graph minus an
edge. If both sets E(Vx,y , Vy,z ) and E(Vy,z , Vx,z ) induce either a complete bipartite graph, then we
are done. So it remains to check two cases:
• when one edge set induces a complete bipartite graph and the other one induces a complete
bipartite graph minus an edge, and
• when both edge sets induce a complete bipartite graph minus an edge.
First consider the former case. Suppose there exists v3 ∈ Vy,z such that E(v1 , Vy,z ) = {v1 v : v ∈ Vy,z −
v3 } and E(v2 , Vy,z ) = {v2 v : v ∈ Vy,z }. Since G[{a, b, c, v1 , v2 , v3 }] is isomorphic to G6,24 , then this
case is not possible. Now consider second case. Let v3 , v4 ∈ Vy,z . There are two possible cases: either
v1 v3 , v2 v3 ∈
/ E(G), or v1 v3 , v2 v4 ∈
/ E(G). Suppose v1 v3 , v2 v4 ∈
/ E(G). Since G[{v4 , v1 , z, v2 , v3 }] is
isomorphic to P5 , then this case is impossible. And therefore v1 v3 , v2 v3 ∈
/ E(G).
Claim 4.10. If each edge set E(Vx,y , Vy,z ) is not empty, then the induced subgraph G[Va,b ∪Va,c ∪Vb,c ]
is isomorphic to one of the following graphs:
• a complete tripartite graph,
• a complete tripartite graph minus an edge, or
• a complete tripartite graph minus the edges v1 v2 , v2 v3 , v3 v1 , where v1 ∈ Vx,y , v2 ∈ Vy,z ,
v3 ∈ Vx,z .
Proof. By Claim 4.8, the induced subgraph G[Va,b ∪ Va,c ∪ Vb,c ] is a complete tripartite graph minus
at most three edges. We will analyze the cases where 2 or 3 edges have been removed.
Suppose G[Va,b ∪ Va,c ∪ Vb,c ] induces a complete tripartite graph minus 2 edges. Since E(Vx,y , Vy,z )
induces complete bipartite graph or complete bipartite graph minus an edge, then the 2 edges cannot
be removed from a unique edge set E(Vx,y , Vy,z ). Then there are two possibilities:
• there exist u, u′ ∈ Vx,y , v ∈ Vy,z and w ∈ Vx,z such that uv, u′ w ∈
/ E(G), or
• there exist u ∈ Vxy , v ∈ Vy,z and w ∈ Vx,z such that uv, uw ∈
/ E(G).
In the first case the graph induced by the set {x, v, u, u′ , w, z} is isomorphic to the forbidden graph
G6,17 ; which is impossible. And in the second case, the the induced subgraph G[{a, b, c, u, v, w}] is
12
CARLOS A. ALFARO AND CARLOS E. VALENCIA
isomorphic to G6,13 , which is impossible. Thus the case where 2 edges are removed from G[Va,b ∪
Va,c ∪ Vb,c ] is not possible.
Suppose G[Va,b ∪ Va,c ∪ Vb,c ] is a complete tripartite graph minus 3 edges. Since E(Vx,y , Vy,z )
induces a complete bipartite graph or a complete bipartite graph minus an edge, then the 3 edges
cannot be removed from a unique edge set E(Vx,y , Vy,z ). Thus there are four possible cases:
(a)
(b)
(c)
(d)
v1 v4 , v3 v6 , v5 v2
v1 v3 , v2 v5 , v5 v4
v2 v4 , v4 v5 , v5 v2
v1 v3 , v3 v5 , v5 v2
∈
/ E(G),
∈
/ E(G),
∈
/ E(G),
∈
/ E(G),
where
where
where
where
v1 , v2
v1 , v2
v1 , v2
v1 , v2
∈ Vx,y ,
∈ Vx,y ,
∈ Vx,y ,
∈ Vx,y ,
v3 , v4
v3 , v4
v3 , v4
v3 , v4
∈ Vy,z
∈ Vy,z
∈ Vy,z
∈ Vy,z
and
and
and
and
v5 , v6 ∈ Vx,z
v5 ∈ Vx,z
v5 ∈ Vx,z and
v5 ∈ Vx,z .
Cases (a), (b) and (d) are impossible, the argument is the following. Case (a) is impossible because
the induced subgraph G[{v2 , v3 , v5 , v6 , x, y}] is isomorphic to the forbidden graph G6,17 . In case (b),
the induced subgraph G[{v2 , v5 , v4 , a, b, c}] is isomorphic to the forbidden graph G6,24 . And in case
(d), the induced subgraph G[{v2 , v3 , v5 , a, b, c}] is isomorphic to the forbidden graph G6,24 . Thus
when 3 edges are removed the only possible case is (c).
By previous Claims we have the following cases:
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
when the set Vx,y is the only not empty set,
when Vx,z = ∅, Vx,y = {vxy }, Vy,z = {vyz } and vxy vyz ∈
/ E(G),
when Vx,z = ∅ and E(Vx,y , Vy,z ) induces a bipartite complete graph,
when Vx,z = ∅ and E(Vx,y , Vy,z ) induces a bipartite complete graph minus an edge,
when Va,b = {vab }, Va,c = {vac }, Vb,c = {vbc } and vab vac , vab vbc , vac vbc ∈
/ E(G),
when Vx,z = {vxz }, Vy,z = {vyz }, vxz vyz ∈
/ E(G), and the edge sets E(vxz , Vx,y ) and
E(vyz , Vx,y ) induce a complete bipartite graph,
when Vx,z = {vxz }, Vy,z = {vyz }, vxz vyz ∈
/ E(G), and there exists vxy ∈ Vx,y such that
E(vxz , Vx,y ) = {vxz v : v ∈ Vx,y − vxy } and E(vyz , Vx,y ) = {vyz v : v ∈ Vx,y − vxy }.
when G[Va,b ∪ Va,c ∪ Vb,c ] is isomorphic to a complete tripartite graph,
when G[Va,b ∪ Va,c ∪ Vb,c ] is isomorphic to a complete tripartite graph minus an edge,
when G[Va,b ∪ Va,c ∪ Vb,c ] is isomorphic to a complete tripartite graph, where v1 ∈ Vx,y ,
v2 ∈ Vy,z , v3 ∈ Vx,z and v1 v2 , v2 v3 , v3 v1 ∈
/ E(G).
Now we describe the vertex set V∅ .
Remark 4.11. Let w, w′ ∈ V∅ . Suppose w is adjacent with a vertex in Vx,y and with w′ . Then the
vertex w′ is adjacent with a vertex in Vx,y , because otherwise the shortest path from w′ to z would
contains a graph isomorphic to P5 . Thus each vertex in V∅ is adjacent with a vertex in Vx,y for some
{x, y} ⊂ {a, b, c}.
Claim 4.12. If w ∈ V∅ is adjacent with v ∈ Vx,y , then either w is adjacent only with v and with
no other vertex in Vx,y , or w is adjacent with each vertex in Vx,y . Moreover, if each vertex in V∅
is adjacent with a vertex in Vx,y , then either exists a vertex v ∈ Vv,x such that each vertex in V∅ is
adjacent with v, or each vertex in V∅ is adjacent with each vertex in Vx,y .
Proof. Since the first statement is easy when Vx,y has cardinality at most 2, then we assume that Vx,y
has cardinality at least 3. Let v ′ , v ′′ ∈ Vx,y . Suppose w is adjacent with v and v ′ but not adjacent
with v ′′ . Since G[{x, z, v, v ′ , v ′′ , w}] is isomorphic to G6,3 , then we get a contradiction. And then w
is adjacent only with v or with v, v ′ and v ′′ .
Let w, w′ ∈ V∅ . Suppose there is v ∈ Vx,y such that wv ∈ E(G) and w′ v ∈
/ E(G). The vertex w is
′
not adjacent with w , because otherwise by Remark 4.11 we get a contradiction. Let v ′ ∈ Vx,y such
that w′ is adjacent with w′ . Thus there are two possible cases:
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
13
• ww′ , wv ′ ∈
/ E(G), and
• wv ′ ∈ E(G) and ww′ ∈
/ E(G).
Since in the first case G[{w, v, x, v ′ , w′ }] is isomorphic to P5 and in the second case G[{x, z, v, v ′ , w, w′ }]
is isomorphic to G6,9 , then we get a contradiction and thus there is no vertex in Vx,y adjacent with
a vertex in V∅ and not adjacent with other vertex in V∅ .
Claim 4.13. Let v ∈ Vx,y . If each vertex in V∅ is adjacent with v, then V∅ induces either K2 or it
is a trivial graph. If furthermore there exists v ′ ∈ Vx,y such that no vertex in V∅ is adjacent with v ′ ,
then V∅ is a clique of cardinality at most 2.
Proof. First note that P3 is forbidden as induced subgraph in G[V∅ ]. It is because if the vertices
w1 , w2 , w3 ∈ V∅ induce a graph isomorphic to P3 , then G[{x, y, v, w1 , w2 , w3 }] ≃ G6,7 . Now we will
see that each component in G[V∅ ] is a clique. Let C be a component in G[V∅ ]. Suppose C is not
a clique, then there are two vertices not adjacent in C, say w and w′ . Let P be the smallest path
contained in C between w and w′ . The length of P is greater or equal to 3. So P3 is an induced
subgraph of P , and hence of C. Which is a contradiction, and therefore, C is a clique. On the other
hand, the graph K2 + K1 is forbidden as induced subgraph in G[V∅ ]. It is because if w1 , w2 , w3 ∈ V∅
such that G[{w1 , w2 , w3 }] ≃ K2 + K1 , then G[{x, y, v, w1 , w2 , w3 }] ≃ G6,8 . Therefore, if G[V∅ ] has
more than one component, then each component has cardinality one.
Let v, v ′ ∈ Vx,y such that each vertex in V∅ is adjacent with v, and no vertex in V∅ is adjacent
with v ′ . Suppose V∅ induces a stable set of cardinality at least 2. Take w, w′ ∈ V∅ . Then we get a
contradiction since the induced graph G[{w, w′ , v, v ′ , x, z}] is isomorphic to G6,1 . Thus V∅ is a clique
of cardinality at most 2.
Thus by Claims 4.12 and 4.13, in case (1) we have the following possible cases:
• V∅ is a clique of cardinality at most 2, and each vertex in V∅ is adjacent with only one vertex
in Vx,y , and
• V∅ is a trivial graph and E(V∅ , Vx,y ) induces a complete bipartite graph.
Note that these graphs are isomorphic to an induced subgraph of a graph in F11 .
Claim 4.14. If E(Vx,y , Vy,z ) = ∅ and E(V∅ , Vx,y ∪ Vy,z ) 6= ∅, then V∅ is a clique of cardinality at
most 2, and each vertex in V∅ is adjacent each vertex in Vx,y ∪ Vy,z
Proof. Let vxy ∈ Vx,y , vyz ∈ Vy,z such that vxy vyz ∈
/ E(G). It is easy to see that if w is adjacent
with vxy or vyz , then w is adjacent with both vertices, because otherwise G has an induced subgraph
isomorphic to P5 . Thus each vertex in V∅ is adjacent with each vertex in Vx,y ∪ Vy,z . Now suppose
w, w′ ∈ V∅ such that w and w′ are not adjacent. Since the induced subgraph G[{w, w′ , x, y, vxy , vyz }]
is isomorphic to G6,15 , then we get a contradiction and V∅ induces a clique of cardinality at most
2.
By previous Claim we get that in case (2), V∅ is a clique of cardinality at most 2 and each vertex
in V∅ is adjacent each vertex in Vx,y ∪ Vy,z . This graph is isomorphic to an induced subgraph of a
graph in F11 .
Claim 4.15. If the edge set E(Vx,y , Vy,z ) induces a complete bipartite graph and E(V∅ , Vx,y ∪Vy,z ) 6= ∅,
then V∅ is a clique of cardinality at most 2, and each vertex in V∅ is adjacent only to one vertex in
Vx,y ∪ Vy,z .
14
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Proof. First we prove that there each vertex in V∅ is adjacent with a vertex in only one of the sets
Vx,y or Vy,z . Suppose there exists a vertex w ∈ V∅ adjacent with v ∈ Vx,y and u ∈ Vy,z . Since the
induced subgraph G[{w, a, b, c, v, u}] is isomorphic to G6,24 , then we get a contradiction and each
vertex in V∅ is adjacent only with vertices of one of the vertex sets Vx,y or Vy,z . Suppose there are
two vertices w, w′ ∈ V∅ such that w is adjacent with v ∈ Vx,y , and w′ is adjacent with u ∈ Vy,z .
Since G[{w, w′ , u, v, x, z}] is isomorphic to G6,9 , then this is impossible and the vertices of V∅ are
adjacent only to vertices in one of the vertex sets either Vx,y or Vy,z . Now suppose that w ∈ V∅ is
adjacent with two vertices in Vx,y , say v and v ′ . Take u ∈ Vy,z . So u is adjacent with both v and
v ′ . Since the induced subgraph G[{w, v, v ′ , u, a, b, c}] is isomorphic to G7,10 , then this cannot occur.
Thus each vertex in V∅ is adjacent only with one vertex in Vx,y ∪ Vy,z . Finally suppose V∅ induces a
trivial graph of cardinality at least 2. Let w, w′ ∈ V∅ adjacent with v ∈ Vx,y . Take u ∈ Vy,z , so u is
adjacent with both v. Since the induced subgraph G[{w, w, v, u, x, z}] is isomorphic to G6,6 , we get
a contradiction and the result follows.
By previous Claim we have that in case (3) the vertex set V∅ induces a clique of cardinality at
most 2 and each vertex in V∅ is adjacent only to one vertex in Vx,y ∪ Vy,z . In this case, the graph is
isomorphic to an induced subgraph of a graph in F11 .
Claim 4.16. Let u ∈ Vx,y and v, v ′ ∈ Vy,z such that u is adjacent with v but not with v ′ . If w ∈ V∅ ,
then w is not adjacent with v.
Proof. Suppose w is adjacent with v. Note that if w is adjacent with u or v ′ , then w is adjacent with
both u and v ′ . Thus there are two cases: either w is adjacent only with v, or w is adjacent with v, v ′
and u. Both cases are impossible because in the former case G[{x, z, u, v, v ′ , w}] is isomorphic to G6,9 ,
meanwhile in the second case G[{y, z, u, v, v ′ , w}] is isomorphic to G6,17 ; which is a contradiction.
Consider case (4). Let u ∈ Vx,y and v ∈ Vy,z such that u is not adjacent with v. By Claim 4.16,
each vertex in V∅ is adjacent with u or with v. It is not difficult to see that in fact each vertex in
V∅ is adjacent with both u and v, because otherwise G would contain P5 as induced subgraph. By
applying Claim 4.14 to the induced subgraph G[{u, v} ∪ V∅ ], we get that V∅ is a clique of cardinality
at most 2. And this graph is isomorphic to an induced subgraph of a graph in F11 .
Now consider case (5), by Claim 4.14, we get that V∅ is a clique of cardinality at most 2, and each
vertex in V∅ is adjacent each vertex in Va,b ∪ Vb,c ∪ Va,c . And this graph is isomorphic to an induced
subgraph of a graph in F11 .
Claim 4.17. Let u1 ∈ Vx,y , u2 ∈ Vy,z and u3 ∈ Vx,z such that u1 is adjacent with both u2 and u3 ,
and u2 u3 ∈
/ E(G). If w ∈ V∅ , then w is not adjacent with u1 .
Proof. Suppose w is adjacent with u1 . Note that if w is adjacent with u2 or u3 , then w is adjacent
with both u2 and u3 . Thus there are two cases: either w is adjacent only with u1 , or w is adjacent
with u1 , u2 and u3 . Both cases are impossible because in the former case G[{x, y, u1 , u2 , u3 , w}]
is isomorphic to G6,12 , meanwhile in the second case G[{x, z, u1 , u2 , u3 , w}] is isomorphic to G6,25 ;
which is a contradiction.
Consider the Cases (6) and (9). Let vxz ∈ Vx,z and vyz ∈ Vy,z such that vxz is not adjacent with
vyz . By Claims 4.14 and 4.17, each vertex in V∅ is adjacent only with both vxz and vyz . And by
Claim 4.14, we get that V∅ is a clique of cardinality at most 2. And these graphs are isomorphic to
an induced subgraph of a graph in F11 .
Now consider the Case (7) and (10). Let vxy ∈ Vx,y , vxz ∈ Vx,z and vyz ∈ Vy,z such that vxy is not
adjacent with vxz and vyz , and vxz is not adjacent with vyz . By Claims 4.17 and 4.14, each vertex in
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
15
V∅ is adjacent only with vxy , vxz and vyz . And by Claim 4.14, we get that V∅ is a clique of cardinality
at most 2. And these graphs are isomorphic to an induced subgraph of a graph in F11 .
Finally in the case (8), by Claim 4.15 we have that the vertex set V∅ is a clique of cardinality
at most 2 and each vertex in V∅ is adjacent with only one vertex in Va,b ∪ Vb,c ∪ Va,c . This case
corresponds to a graph that is isomorphic to an induced subgraph of a graph in F11 .
4.2. Case Vx = Tn , where n ≥ 2. First we will obtain that E(Vx , Vy,z ) satisfies one of the following
statements:
• it induces a complete bipartite graph,
• there exists a vertex s ∈ Vx , we called the apex, such that E(Vx , Vy,z ) = {uv : u ∈ Vx −
s and v ∈ Vy,z }, or
• there exists a vertex s ∈ Vy,z , we called the apex, such that E(Vx , Vy,z ) = {uv : u ∈
Vx and v ∈ Vy,z − s}.
To do it, we will prove the following claims.
Claim 4.18. Let u1 , u2 ∈ Vx and v1 , v2 ∈ Vy,z such that u1 v1 , u2 v2 ∈
/ E(G). Then either u1 = u2 or
v1 = v2 .
Proof. Suppose u1 6= u2 and v1 6= v2 . There are three possible cases:
• u1 v2 , u2 v1 ∈
/ E(G),
• u1 v2 ∈
/ E(G) and u2 v1 ∈ E(G), or
• u1 v2 , u2 v1 ∈ E(G).
The first two cases are not possible since the induced subgraph G[{x, y, u1 , u2 , v1 , v2 }] would be
isomorphic to G6,1 and G6,9 , respectively. In the last case, the induced subgraph G[{x, u1 , u2 , v1 , v2 }]
is isomorphic to P5 ; which is impossible. Thus the result follows.
The last claim implies that all non-edges in E(Vx , Vy,z ) are incident to a one vertex: the apex s.
Suppose the vertex s is in Vx , and there are vertices v1 , v2 ∈ Vy,z such that sv1 ∈ E(G), and
sv2 ∈
/ E(G). By Claim 4.18, each vertex in Vx − s is adjacent with v1 and v2 . Then the induced
subgraph G[{a, b, c, u1 , u2 , v1 , v2 }] ≃ G7,9 , that is impossible. This implies that if the vertex s ∈ Vx is
not adjacent with a vertex in Vy,z , then s is not adjacent with all vertices in Vy,z . A similar argument
yields that if the apex vertex s is in Vy,z and s is not adjacent with a vertex in Vx , then s is not
adjacent with all vertices in Vx .
Thus, we have three cases:
(a) E(Vx , Vy,z ) is complete bipartite minus the edges between a vertex s (the apex) in Vx and all
vertices of Vy,z ,
(b) E(Vx , Vy,z ) is complete bipartite minus the edges between a vertex s (the apex) in Vy,z and all
vertices in Vx , and
(c) E(Vx , Vy,z ) is complete bipartite.
Claim 4.19. If |Vx | ≥ 3 and v ∈ Vy,z , then E(v, Vx ) either it induces a complete bipartite graph or
it is empty.
Proof. Let u1 , u2 , u3 ∈ Vx . Suppose one of the two following possibilities happen: vu1 ∈ E(G) and
vu2 , vu3 ∈
/ E(G) , or vu1 , vu2 ∈ E(G) and vu3 ∈
/ E(G). In the first case the induced subgraph
G[{u1 , u2 , u3 , v, x, y}] is isomorphic to G6,3 , meanwhile in the second case the induced subgraph
G[{u1 , u2 , u3 , v, x, y, z}] is isomorphic to G7,2 . Then both cases cannot occur, and we get the result.
16
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Thus case (a) occur only when |Vx | = 2.
In what follows we describe the vertex set V∅ , that is, the vertex set whose vertices have no edge
in common with the vertex set {a, b, c}.
Remark 4.20. Let w ∈ V∅ . The vertex w is adjacent with a vertex in Vx ∪Vy,z , because otherwise the
shortest path from w to {a, b, c} would contains the graph P5 as induced subgraph. Let u1 , u2 ∈ Vx .
If w is adjacent with u1 or u2 , then w is adjacent with both vertices, because otherwise the induced
subgraph G[{a, b, c, u1 , u2 , w}] would be isomorphic to G6,6 , which is forbidden.
In case (a), we will see that each vertex in V∅ is adjacent with each vertex in Vx ∪ Vy,z . Let w in
V∅ . Supppose s ∈ Vx is the vertex that is not adjacent with any vertex in Vy,z . If w ∈ V∅ is adjacent
with one of the vertices in {s} ∪ Vy,z , then w must to be adjacent with s and each vertex in Vy,z ,
because otherwise let v ∈ Vy,z , then the induced subgraph G[{x, y, s, w, v}] would be isomorphic to
P5 . Then by Remark 4.20, w is adjacent with each vertex in Vx ∪ Vy,z .
Claim 4.21. The vertex set V∅ induces a stable set.
Proof. Suppose w1 , w2 ∈ V∅ are adjacent. Since both w1 and w2 are adjacent with u ∈ Vx − s and
v ∈ Vy,z , then the induced subgraph G[{u, v, w1 , w2 }] is isomorphic to K4 ; which is forbidden. Thus
w1 and w2 are not adjacent, and therefore V∅ is a stable set.
Thus this case corresponds to a graph that is isomorphic to an induced subgraph of a graph in
F11 .
Now consider case (b). Let u1 , u2 ∈ Vx and s ∈ Vy,z such that s is not adjacent with u1 and u2 .
Claim 4.22. If w ∈ V∅ is adjacent with a vertex in Vy,z \ {s}, then each vertex in V∅ is adjacent
with each vertex in Vx ∪ Vy,z .
Proof. Let w ∈ V∅ . Suppose w is adjacent with v ∈ Vy,z \ {s}. If w is not adjacent with both vertices
u1 , u2 ∈ Vx , then we get that the induced subgraph G[{a, b, c, w, u1 , u2 , v}] is isomorphic to G7,2 ,
which cannot be. Thus w is adjacent with at least one vertex in Vx . Then by Remark 4.20, w is
adjacent with both vertices u1 and u2 . If there exist v ′ ∈ Vy,z \ {s} such that w is not adjacent
with v ′ , then the vertex set {w, u1 , x, y, v, v ′ } would induce a graph isomorphic to G6,15 . On the
other hand, if w is not adjacent with s, then the vertex set {w, u1 , x, y, s} induces the subgraph P5 .
Therefore, w is adjacent with each vertex in Vx ∪ Vy,z .
Suppose there is another vertex w′ ∈ V∅ . By the above argument if w′ is adjacent with a vertex
in Vy,z \ {s}, then it must be adjacent with each vertex in Vx ∪ Vy,z . Also if w′ is adjacent with
a vertex in {s} ∪ Vx , then w′ must be adjacent with each vertex in {s} ∪ Vx . So suppose w′ is
adjacent with each vertex in s ∪ Vx and w′ is not adjacent with each vertex in Vy,z \ {s}. Then
there are two possibilities: either ww′ ∈
/ E(G) or ww′ ∈ E(G). Let v ∈ Vy,z \ {s}. In the first case
′
G[{x, y, s, v, w, w }] ≃ G6,15 and in the second case G[{x, y, v, w, w′ }] ≃ P5 . Since both graphs are
forbidden, then we get a contradiction and thus w′ is adjacent with v. And therefore w′ is adjacent
with each vertex in Vx ∪ Vy,z .
Claim 4.23. Either each vertex in V∅ is adjacent with each vertex in Vx ∪ {s}, or each vertex in V∅
is adjacent with each vertex in Vx ∪ Vy,z .
Proof. Let w ∈ V∅ . Clearly, if w is adjacent with a vertex in Vx ∪ {v}, then w is adjacent with each
vertex in Vx ∪ {v}, because otherwise P5 would be an induced subgraph. By Claim 4.22, we have
that if there is a vertex w ∈ V∅ adjacent with a vertex in Vy,z , different to the apex s, then each
vertex in V∅ is adjacent with each vertex in Vx and Vy,z \ {s}. Thus we get the result.
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
17
Claim 4.24. The vertex set V∅ induces either a clique of cardinality at most 2 or a trivial graph.
Proof. First note that P3 is forbidden as induced subgraph in V∅ . It is because if w1 , w2 , w3 ∈ V∅
induce P3 , then G[{x, y, u1 , w1 , w2 , w3 }] ≃ G6,7 . Now we will get that each component in G[V∅ ] is
a clique. Let C be a component in G[V∅ ]. Suppose C is not a clique, then it has two vertices not
adjacent, say u and v. Let P be the smallest path in C between u and v. Thus the length of P is
greater or equal to 3. So P3 is an induced subgraph of P , and hence of C. Therefore, C is a complete
graph.
On the other hand, the graph K2 + K1 is forbidden for G[V∅ ]. It is because if w1 , w2 , w3 ∈ V∅
such that G[{w1 , w2 , w3 }] ≃ K2 + K1 , then G[{x, y, u1 , w1 , w2 , w3 }] ≃ G6,8 ; which cannot happen.
Therefore, if G[V∅ ] has more than one component, then each component has cardinality one.
In the first case of Claim 4.23, if |Vy,z | ≥ 2, then V∅ is either K1 or K2 , because if u1 ∈ Vx ,
v ∈ Vy,z \ {s}, and w1 , w2 ∈ V∅ are adjacent, then G[{w1 , w2 , u1 , v, x, y}] ≃ G6,3 . Otherwise if
Vy,z = {s}, then both possibilities in Claim 4.24 are allowed. In the second case of Claim 4.23, if
|Vy,z | ≥ 3, we have that V∅ = ∅, because if w ∈ V∅ is adjacent with u1 , u2 ∈ Vx and with two vertices
v1 , v2 ∈ Vy,z \ {s}, we get G[{x, y, u1 , u2 , v1 , v2 , w}] ≃ G7,13 as forbidden subgraph. If |Vy,z | = 2, then
V∅ is trivial since ω(G) = 3. And if Vy,z = {s}, then both possibilities in Claim 4.24 are allowed.
With no much effort the reader can see that each of these cases corresponds to a graph isomorphic
to an induced subgraph of a graph in F11 .
Case (c). By Claim 4.23, there are two possible cases:
• either each vertex in V∅ is adjacent with each vertex in Vx ∪ Vy,z , or
• each vertex in V∅ is adjacent with each vertex in Vx , and no vertex in V∅ is adjacent with
any vertex in Vy,z .
In first case when |Vy,z | ≥ 2, the vertex set V∅ is empty. Because if w ∈ V∅ is adjacent with the
vertices v1 , v2 ∈ Vy,z , and u1 , u2 ∈ Vx , then the induced subgraph G[{x, y, u1 , u2 , v1 , v2 , w}] ≃ G7,13
which is forbidden. Then the vertex set V∅ is empty. Otherwise when |Vy,z | = 1, then the vertex set
V∅ must be a stable set, because if there exist two adjacent vertices in V∅ , then by taking a vertex
in Vx and a vertex in Vy,z we get K4 that is forbidden. Finally, in the case when each vertex in V∅
is adjacent with each vertex in Vx , and no vertex in V∅ is adjacent with a vertex in Vy,z , we get that
V∅ is a clique of cardinality at most 2. It is because if w1 , w2 ∈ V∅ are such that w1 w2 ∈
/ E(G), then
we get G[{w1 , w2 , u1 , v, x, y}] ≃ G6,3 that is forbidden. And each of these graphs are isomorphic to
an induced subgraph of a graph in F11 .
4.3. Case Vx is a complete bipartite graph of cardinality at least 3. Assume Vx is a complete
bipartite graph of cardinality at lest three with (A, B) the bipartition of Vx .
Claim 4.25. If v ∈ Vy,z , then E(v, Vx ) 6= ∅.
Proof. Suppose E(v, Vx ) = ∅. There exist u1 , u2 , u3 ∈ Vx such that G[{u1 , u2 , u3 }] ≃ P3 . And then
we get contradiction since the induced subgraph G[{u1 , u2 , u3 , x, y, v}] ≃ G6,7 ; which is forbidden.
Claim 4.26. Let u ∈ Vx and v ∈ Vy,z . If u and v are adjacent, then v is adjacent with each vertex
in the part (A or B) containing u.
Proof. Suppose u ∈ A and v is not adjacent with any vertex in A \ {u}. If |A| = 1, then the
result follows, so we may assume |A| ≥ 2. Since G[Vx ] is connected and has cardinality at least
3, then there exists u′ ∈ A and u′′ ∈ B such that G[{u, u′′ , u′ }] ≃ P3 . We have u′ and v are
not adjacent. There are two possibilities: either u′′ v ∈ E(G) or u′′ v ∈
/ E(G). In the first case
18
CARLOS A. ALFARO AND CARLOS E. VALENCIA
the induced subgraph G[{u′ , u′′ .u, v, y}] is isomorphic to P5 , and in the second case the induced
subgraph G[{x, y, u, u′ , u′′ , v}] is isomorphic to G6,17 . Since both cases are forbidden, we get a
contradiction.
Previous Claims suggest to divide the vertex set Vy,z in three subsets:
A , the vertices in V
• Vy,z
y,z that are adjacent with each vertex in A,
B
• Vy,z , the vertices in Vy,z that are adjacent with each vertex in B, and
AB , the vertices in V
• Vy,z
y,z that are adjacent with each vertex in A ∪ B.
In what follows, we assume |A| ≥ |B|.
A and V B is no more than 1.
Claim 4.27. The cardinality of the sets Vy,z
y,z
A . Let u ∈ A, and u′ ∈ B. Since G[{u, u′ , v, v ′ , x, y}] is
Proof. Suppose there exist v, v ′ ∈ Vy,z
B is similar.
isomorphic to G6,15 , which is forbidden, then we get a contradiction. The case Vy,z
B or V AB is empty.
Claim 4.28. If |B| ≥ 2 and Vy,z 6= ∅, then one of the sets Vy,z
y,z
B and v ′ ∈ V AB . Let u, u′ ∈ B, and u′′ ∈ A. Thus G[{v, v ′ , u, u′ , u′′ , x, y}] ≃
Proof. Suppose v ∈ Vy,z
y,z
G7,13 . Which is impossible.
A or V AB is empty.
Since |A| ≥ 2, then by applying previous Claim to A, one of the sets Vy,z
y,z
Thus the possible cases we have are the following:
(a)
(b)
(c)
(d)
(e)
(f)
Vy,z = ∅,
B ∪ V AB = ∅ and |V A | = 1,
Vy,z
y,z
y,z
A
AB = ∅ and |V B | = 1,
Vy,z ∪ Vy,z
y,z
AB = ∅ and |V A | = |V B | = 1,
Vy,z
y,z
y,z
A = ∅, |B| = 1, |V B | = 1 and |V AB | ≥ 1, and
Vy,z
y,z
y,z
A ∪ V B = ∅ and |V AB | ≥ 1.
Vy,z
y,z
y,z
Now we describe V∅ , that is, the set of vertices not adjacent with any vertex in {a, b, c}. Let
w ∈ V∅ . The vertex w is adjacent with a vertex in Vx ∪ Vy,z , because otherwise the shortest path
from w to {x, y} would contains the graph P5 as induced subgraph.
Claim 4.29. Let w ∈ V∅ . If w is adjacent with a vertex in Vx , then w is adjacent with each vertex
in the parts of the partition (A, B) with cardinality greater or equal to 2.
Proof. Let v ∈ Vx be a vertex adjacent with w. Suppose v ∈ B. we will prove two things: (1) if
|B| ≥ 2, then w is adjacent with each vertex in B, and (2) w is adjacent with each vertex in A.
Let us consider case when |B| ≥ 2. We will see that w is adjacent with each vertex in B. Suppose
there is a vertex v ′ ∈ B not adjacent with w. Take u ∈ A. Thus there are two possibilities: either u
and w are adjacent or not. The case uw ∈ E(G) is impossible because G[{u, v, v ′ , w, x, y}] ≃ G6,10 ,
which is forbidden. Meanwhile, the case uw ∈
/ E(G) is impossible because G[{u, v, v ′ , w, x, y}] ≃
G6,14 , which is forbidden. Thus w is adjacent with v ′ , and therefore w is adjacent with each vertex
in B.
Now we see that w is adjacent with each vertex in A. Note that in this case |B| may be equal
to 1. Suppose w is not adjacent with any vertex in A. Let u, u′ ∈ A. Since the induced subgraph
G[{w, v, u, u′ , x, y}] is isomorphic to the forbidden graph G6,5 , then we get a contradiction. Thus w
is adjacent with a vertex in A. Now applying the previous case (1) to A, we get that w is adjacent
with each vertex in A.
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
19
Next Claim show us what happens in the case when |B| = 1.
Claim 4.30. If |B| = 1 and E(V∅ , Vx ) 6= ∅, then only one of the edges sets E(V∅ , Vx ) or E(V∅ , A)
induces a complete bipartite graph.
Proof. Let w, w′ ∈ V∅ , u ∈ B, and u′ ∈ A. By Claim 4.29, vertices w and w′ are adjacent with each
vertex in A. Suppose w is adjacent with u, and w′ is not adjacent with u. There are two possibilities:
either w and w′ are adjacent or not. If ww′ ∈ E(G), then the induced subgraph G[w, w′ , u, x, y] is
isomorphic to P5 ; which is impossible. On the other hand, if ww′ ∈
/ E(G), then G[{w, w′ , u, u′ , x, y}]
is isomorphic to G6,10 , which is forbidden. So we get a contradiction, and the result follows.
Claim 4.31. Let w ∈ V∅ . If w is adjacent with v ∈ Vy,z , then w is adjacent with each vertex in
A ∪ V AB and
the parts of partition (A, B) with cardinality greater or equal to 2. Moreover, if v ∈ Vy,z
y,z
|B| = 1, then w is adjacent with the unique vertex in B.
Proof. Let w ∈ V∅ and v ∈ Vy,z such that w and v are adjacent. By Claim 4.25, E(v, Vx ) 6= ∅, and
A , v ∈ V AB , or v ∈ V B .
therefore there are three cases: v ∈ Vy,z
y,z
y,z
A . If E(w, V ) = ∅, then by taking u, u ∈ A, the forbidden induced subgraph
Suppose v ∈ Vy,z
x
′
G[W ∪ {w, v, u, u }] ≃ G7,2 would appear and we get a contradiction. Thus E(w, Vx ) 6= ∅. By Claim
4.29, w is adjacent with each vertex in the parts of Vx with cardinality greater or equal to 2. When
|B| = 1, take u′′ ∈ B. If wu′′ ∈
/ E(G), then G[{w, v, y, x, u′′ }] would be isomorphic to the forbidden
graph P5 . Therefore, wu′′ ∈ E(G).
AB . In a similar way than in previous case we get that E(w, V ) 6= ∅, and v is
Suppose v ∈ Vy,z
x
adjacent with each vertex in the parts of the partition (A, B) of cardinality greater or equal to 2. In
the case when |B| = 1, suppose u ∈ B with wu ∈
/ E(G). Take u′ ∈ A, we know that u′ w ∈ E(G).
Since G[{w, u, u′ , x, y, v}] ≃ G6,24 , then we get a contradiction. Thus wu ∈ E(G).
B . It is easy to see that v is adjacent with each vertex in A, because otherwise
Now suppose v ∈ Vy,z
P5 would appear. Therefore E(w, Vx ) 6= ∅, and by Claim 4.29, we have that w is adjacent with each
vertex in the parts of the partition (A, B) of cardinality greater or equal to 2.
AB ) = ∅. Moreover, if E(V , B) 6= ∅ and E(V , V B ) 6= ∅,
Claim 4.32. If E(V∅ , B) 6= ∅, then E(V∅ , Vy,z
∅
∅ y,z
AB = ∅.
then Vy,z
AB . Suppose w and v are adjacent. Take u ∈ A and u ∈ B. Since v
Proof. Let w ∈ V∅ and v ∈ Vy,z
1
2
AB
is in Vy,z and u1 is adjacent with u2 , then G[{v, u1 , u2 }] is isomorphic to K3 . On the other hand,
by Claims 4.29 and 4.30, w is adjacent with each vertex in A ∪ B. Thus the induced subgraph
G[{w, v, u1 , u2 }] is isomorphic to K4 that is forbidden, and therefore w cannot be adjacent with v.
B , and suppose there exists a vertex v ∈ V AB . Since w
Let w is adjacent with u2 ∈ B and v ′ ∈ Vy,z
y,z
is not adjacent with v, then G[{w, v, v ′ , u2 , x, y}] is isomorphic to G6,15 ; which it is not possible.
Claim 4.33. Let w ∈ V∅ and v ∈ Vy,z . If w and v are adjacent, then each vertex in V∅ is adjacent
with v.
Proof. Suppose w′ ∈ V∅ such that w′ is not adjacent with v. By Claims 4.29 and 4.30, both vertices
w and w′ are adjacent with each vertex in A. Let u ∈ A. There are four cases obtained by the
A
combinations of the following possible cases: either ww′ ∈ E(G) or ww′ ∈
/ E(G), and either v ∈ Vy,z
B . When ww ′ ∈ E(G) and v ∈ V B , the induced subgraph G[{w ′ , w, v, y, x}] is isomorphic to
or v ∈ Vy,z
y,z
B , the induced subgraph G[{w ′ , u, w, v, y}]
P5 ; which is not possible. When ww′ ∈
/ E(G) and v ∈ Vy,z
A , the induced subgraph
is isomorphic to P5 ; which is not possible. When ww′ ∈ E(G) and v ∈ Vy,z
20
CARLOS A. ALFARO AND CARLOS E. VALENCIA
A,
G[{w′ , w, v, y, x}] is isomorphic to P5 ; which is not possible. Finally, when ww′ ∈
/ E(G) and v ∈ Vy,z
′
′
the induced subgraph G[{w , w, x, y, v, u}] is isomorphic to G6,11 ; which is not possible. Thus w is
adjacent with v.
A and v ′ ∈ V B . If w ∈ V is adjacent with v or v ′ , then w is adjacent with
Claim 4.34. Let v ∈ Vy,z
∅
y,z
both v and v ′ .
Proof. First suppose w is adjacent with v and not with v ′ . Let u ∈ A. By Claim 4.31, w is adjacent
with a vertex u ∈ A. Thus G[{w, u, x, y, v ′ }] is isomorphic to P5 that is a contradiction. Therefore w
and v ′ are adjacent. Now suppose w is adjacent with v ′ and not with v. Let u ∈ B. If u is adjacent
with w, then a P5 is obtained in a similar way than previous case. Thus assume u is not adjacent
with w. Then G[{w, x, y, u, v, v ′ }] is isomorphic to G6,9 ; which is impossible. Therefore, w and v are
adjacent.
AB 6= ∅, then E(V , B) = ∅.
Claim 4.35. If Vy,z
∅
AB . By Claim 4.29, w is adjacent with u and
Proof. Let u1 , u2 ∈ A, u3 ∈ B, w ∈ V∅ and v ∈ Vy,z
1
u2 . Suppose w is adjacent with u3 . By Claim 4.32, the vertices w and v are not adjacent. Thus
G[{x, y, u1 , u2 , u3 , v, w}] is isomorphic to G7,10 , which is a contradiction. Therefore, E(V∅ , B) =
∅.
Thus applying previous Claims to cases (a) - (f), we obtain the following possibilities:
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Vy,z = ∅, |B| = 1, and E(V∅ , A) induces a complete bipartite graph,
Vy,z = ∅, |B| ≥ 1, and E(V∅ , Vx ) induces a complete bipartite graph,
B ∪ V AB = ∅, |V A | = 1, |B| ≥ 1 and E(V , V ∪ V ) induces a complete bipartite graph,
Vy,z
y,z
∅ x
y,z
y,z
A
AB
B
Vy,z ∪ Vy,z = ∅, |Vy,z | = 1, |B| = 1 and E(V∅ , A ∪ Vy,z ) induces a complete bipartite graph,
A ∪ V AB = ∅, |V B | = 1, |B| ≥ 1 and E(V , V ∪ V ) induces a complete bipartite graph,
Vy,z
y,z
∅ x
y,z
y,z
AB = ∅, |V A | = |V B | = 1, |B| ≥ 1 and E(V , V ∪ V ) induces a complete bipartite graph,
Vy,z
x
y,z
∅
y,z
y,z
A
B
AB
Vy,z ∪ Vy,z = ∅, |Vy,z | ≥ 1, |B| = 1 and E(V∅ , A) induces a complete bipartite graph,
A ∪V B = ∅, |V AB | ≥ 1, |V B | = 1, |B| = 1 and E(V , A∪V B ) induces a complete bipartite
Vy,z
∅
y,z
y,z
y,z
y,z
graph.
With a similar argument as in Claim 4.24, we obtain that V∅ is either trivial or K2 . In cases
(1), (4), (7) and (8), V∅ cannot be Tn with n ≥ 2, since taking w, w′ ∈ V∅ , u ∈ A, u′ ∈ B, then
G[{w, w′ , u, u′ , x, y}] ≃ G6,2 . On the other hand, since ω(G) = 3, then V∅ is not isomorphic to K2
in the cases (2), (3), (5) and (6). It is not difficult to see that in each case G is isomorphic to an
induced subgraph of a graph in F11 .
4.4. Case when Vx induces K1 +K2 or 2K2 . Through this case we assume that Vx = {u1 , u2 , u3 , u4 }
such that u1 u2 , u3 u4 ∈ E(G). That is, G[{u1 , u2 , u3 , u4 }] ≃ 2K2 . Let A = {u1 , u2 } and B = {u3 , u4 }.
The following discussion also applies when one of the vertex set A or B has cardinality 1.
Claim 4.36. If v ∈ Vy,z , then E(v, Vx ) 6= ∅. Moreover, if v ∈ Vy,z , then v is adjacent with each
vertex in one of the following sets A, B or Vx .
Proof. Suppose there is no edge joining v and a vertex in Vx . Since G[{u1 , u2 , u3 , x, y, v}] ≃ G6,6 is a
forbidden induced subgraph, then a contradiction is obtained. Then v is adjacent with some vertex
in Vx .
Suppose v is adjacent with one of the vertices in A. We will prove that v cannot be adjacent only
with one vertex of A, say u1 . Thus suppose v is adjacent with u1 , and v is not adjacent with u2 , u3
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
21
and u4 . Since G[{u1 , u2 , u3 , v, x, y}] ≃ G6,11 , then v is adjacent with u2 or a vertex in B, say u3 . If
v is adjacent with u2 , then we are done. So we assume v is adjacent with u1 , but v is not adjacent
with u2 . This is not possible because G[{u1 , u2 , u3 , v, y, z}] ≃ G6,10 . Therefore, either v is adjacent
with both u1 and u2 , or v is not adjacent with neither of u1 and u2
Claim 4.37. If |Vy,z | ≥ 2, then each vertex in Vy,z is adjacent with each vertex of only one of the
following sets A, B or Vx .
Proof. Let v, v ′ ∈ Vy,z . Consider the following cases:
(a) v is adjacent with each vertex in A and v ′ is adjacent with each vertex in B,
(b) v is adjacent with each vertex in Vx and v ′ is adjacent with each vertex in B, and
(c) v is adjacent with each vertex in Vx and v ′ is adjacent with each vertex in A.
Case (a) is impossible because G[{u1 , v, y, v ′ , u3 }] ≃ P5 , which is forbidden. On the other hand,
cases (b) and (c) are not allowed because G[{a, b, c, v, v ′ , u1 , u3 }] ≃ G7,9 ; which is forbidden. Thus,
the result follows.
Thus E(Vx , Vy,z ) satisfies only one of the following three cases:
(1) E(Vy,z , A) induces a complete bipartite graph,
(2) E(Vy,z , B) induces a complete bipartite graph, or
(3) E(Vy,z , Vx ) induces a complete bipartite graph.
Now we describe V∅ , the set of vertices not adjacent with any vertex in {a, b, c}. Let w ∈ V∅ . The
vertex w is adjacent with a vertex in Vx ∪ Vy,z , because otherwise the shortest path from w to {x, y}
would contains the graph P5 as induced subgraph.
Claim 4.38. If w ∈ V∅ is adjacent with a vertex in Vx , then w is adjacent with each vertex in Vx .
Proof. Suppose w is adjacent with u1 and not with u3 . Since G[W ∪ {w, u1 , u3 }] ≃ G7,4 is forbidden,
w also must be adjacent with u3 . In a similar way we get the opposite case and it turns out the
result.
Consider Cases (1) and (2). The following arguments works on both cases. First note that if there
exists w ∈ V∅ adjacent with v ∈ Vy,z , then w is adjacent with each vertex in Vx . The reason is the
following. Suppose w is not adjacent with any vertex in Vx . Since v is not adjacent with a vertex
in Vx , say u, and the vertices w and u are not adjacent, then we have that G[{w, v, y, x, u}] ≃ P5 ;
which is a contradiction. Then w and u are adjacent. And by Claim 4.38, the edge set E(w, Vx )
induces a complete bipartite graph. On the other hand, we can prove in a similar way that if there
exists w ∈ V∅ adjacent with each vertex in Vx , then w is adjacent with each vertex in Vy,z . Therefore,
each vertex in V∅ is adjacent with each vertex in Vx ∪ Vy,z . Furthermore, the set V∅ is a stable set.
It is because if w, w′ ∈ V∅ were adjacent, then by taking u ∈ Vx and v ∈ Vy,z such that u and
v are adjacent, the induced subgraph G[{w, w′ , u, v, x, y}] would be isomorphic to G6,20 ; which is
impossible. These cases correspond to a graph isomorphic to an induced subgraph of a graph in F11 .
Now let us consider case (3).
Claim 4.39. If w ∈ V∅ is adjacent with a vertex in Vy,z , then each vertex in V∅ is adjacent with
each vertex in Vx ∪ Vy,z .
Proof. Let w ∈ V∅ and v ∈ Vy,z such that wv ∈ E(G). Suppose w is not adjacent with any vertex
in Vx . Take u2 , u3 ∈ Vx such that u2 u3 ∈ E(G). Since the induced subgraph G[{a, b, c, w, u2 , u3 , v}]
22
CARLOS A. ALFARO AND CARLOS E. VALENCIA
is isomorphic to G7,2 , then we get a contradiction; and w is adjacent with a vertex in Vx . Thus by
Claim 4.38, w is adjacent with each vertex in Vx .
Suppose that there exists v ′ ∈ Vy,z such that w is not adjacent with v ′ . Since the vertex set
{w, u1 , x, y, v, v ′ } would induce G6,15 , this does not occur. Therefore, w is adjacent with each vertex
in Vx ∪ Vy,z .
Suppose there is another vertex w′ ∈ V∅ . By the above argument, if w′ is adjacent with a vertex
in Vy,z , then it must be adjacent with each vertex in Vx ∪ Vy,z . And we are done. On the other
hand, by Claim 4.38, if w′ is adjacent with a vertex in Vx , then w′ must be adjacent with each
vertex in Vx . So suppose w′ is adjacent with each vertex in Vx , but not adjacent with each vertex
in Vy,z . Then there are two possibilities: either ww′ ∈
/ E(G) or ww′ ∈ E(G). In the first case
′
G[{x, y, u1 , v, w, w }] ≃ G6,11 and in the second case G[{x, y, v, w, w′ }] ≃ P5 . Since both graphs are
forbidden, then w′ is adjacent with v. And therefore w′ is adjacent with each vertex in Vx ∪ Vy,z .
By Claims 4.38 and 4.39, we obtain that there are two possible cases: either each vertex in V∅
is adjacent with each vertex in Vx ∪ Vy,z , or each vertex in V∅ is adjacent only with each vertex in
Vx . Consider first case. If there exists a vertex w ∈ V∅ , then w adjacent with the vertices v ∈ Vy,z ,
and u1 , u2 ∈ Vx . Thus G[{w, v, u1 , u2 }] is isomorphic to K4 that is not allowed, then V∅ is empty.
Now consider second case. Let w ∈ V∅ and v ∈ Vy,z . Thus w is adjacent with u1 , u2 and u3 . Since
G[{w, u1 , u2 , u3 , v, x, y}] is isomorphic to G7,9 , we get a contradiction. Then V∅ = ∅. The graph in
this case is isomorphic to an induced subgraph of a graph in F11 .
4.5. Cases when G[Va ∪ Vb ∪ Vc ] = Vx ∨ (Vy + Vz ) is one of the following graphs: K1 ∨ 2K1 ,
K1 ∨ (K1 + K2 ), K1 ∨ 2K2 , or K2 ∨ 2K1 . For the sake of clarity, we suppose E(Va , Vb ) and E(Va , Vc )
induce a complete bipartite graph, and E(Vb , Vc ) is empty. Now we are going to obtain some claims
that describe the edge sets joining Vx and Vy,z .
Claim 4.40. Let x, y ∈ {a, b, c}. If E(Vx , Vy ) is not empty, then E(Vx , Vxy ) and E(Vy , Vxy ) are
empty.
Proof. Let vx ∈ Vx , vy ∈ Vy and vxy ∈ Vx,y . Suppose vxy is adjacent with both vx and vy . Then
G[{a, b, c, vxy , vx , vy }] ≃ G6,24 ; which is a contradiction. Now suppose vxy is adjacent with vx and
not with vy . In this case G[{a, b, c, vxy , vx , vy }] ≃ G6,17 ; which is impossible. And therefore result
turns out.
Claim 4.40 implies that E(Va,b , Va ∪ Vb ) = ∅ and E(Va,c , Va ∪ Vc ) = ∅.
Claim 4.41. Let x, y ∈ {a, b, c}. If Vx,y 6= ∅ and E(Vx , Vy ) = ∅, then each edge set E(Vx , Vx,y ) and
E(Vy , Vx,y ) induces a complete bipartite graph.
Proof. Let vx ∈ Vx , vy ∈ Vy and vxy ∈ Vx,y . Suppose vxy is not adjacent with both vx and vy . Then
G[{a, b, c, vxy , vx , vy }] ≃ G6,5 ; which is a contradiction. Finally, suppose vxy is adjacent with vx and
not with vy . Since G[{a, b, c, vx,y , vx , vy }] ≃ G6,14 is forbidden, then we get a contradiction. And the
result turns out.
Claim 4.41 implies that each vertex in Vb,c is adjacent with each vertex in Vb ∪ Vc .
Claim 4.42. If Vb,c 6= ∅, then E(Va , Vb,c ) is empty.
Proof. Let va ∈ Va , vb ∈ Vb and vbc ∈ Vb,c . Suppose vbc is adjacent with va . Then G[{a, b, c, va , vb , vbc }]
is isomorphic to G6,25 that is forbidden. Therefore, there is no edge joining a vertex in Va with a
vertex in Vb,c .
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
23
Claim 4.43. Let x ∈ {b, c}. If Va,x 6= ∅, then E(Va,x , Vx ) induces a complete bipartite graph.
Proof. Let vb ∈ Vb , vc ∈ Vc and vax ∈ Va,x . Suppose vax is not adjacent with vx . Since G[{a, b, c, vax , vb , vc }]
is isomorphic to G6,10 that is forbidden, then each vertex in Vx is adjacent with each vertex in Va,x .
Claim 4.44. The edge set E(Va,b , Va,c ) induces a complete bipartite graph.
Proof. Let va ∈ Va , vb ∈ Vb , vab ∈ Va,b , and vac ∈ Va,c . We know that va is not adjacent with both
vab and vac , and vb is adjacent with vac , but vb is not adjacent with vab . Suppose vab and vac are
not adjacent. Then G[{b, c, va , vb , vab , vac , }] ≃ G6,9 ; which is impossible. And therefore, each vertex
in Va,b is adjacent with each vertex in Va,c .
Claim 4.45. Let x ∈ {b, c} The edge set E(Va,x , Vb,c ) induces a complete bipartite graph.
Proof. Let vb ∈ Vb , vc ∈ Vc , vax ∈ Va,x , and vbc ∈ Vb,c . We know that vb vbc , vbc vc , vc vab , vb vac ∈ E(G)
and vx vax , vb vc ∈
/ E(G). Suppose vax and vbc are not adjacent. Then G[{b, c, vb , vc , vax , vbc , }] ≃ G6,14 ;
which is impossible. Therefore each vertex in Va,x is adjacent with each vertex in Vb,c .
Now we analyze V∅ .
Claim 4.46. If w ∈ V∅ , then w cannot be adjacent with any vertex in Va ∪ Vb ∪ Vc .
Proof. Let x ∈ {b, c}, y ∈ {b, c} − x, va ∈ Va and vx ∈ Vx . Suppose w ∈ V∅ such that E(w, Va ∪ Vb ∪
Vc ) 6= ∅. Consider the following cases:
(a) w is adjacent only with va ,
(b) w is adjacent only with vx , or
(c) w is adjacent with both va and vx .
Cases (a) and (b) are impossible, because G would have P5 as induced subgraph obtained by
G[{w, va , vx , x, y}] and G[{w, vx , va , a, y}], respectively. Finally in case (c), the induced subgraph
G[{a, b, c, w, va , vx }] is isomorphic to G6,21 ; which is impossible.
Claim 4.47. There exists no vertex in V∅ adjacent with a vertex in Vb,c .
Proof. Let w ∈ V∅ , va ∈ Va , vb ∈ Vb , vc ∈ Vc , and vbc ∈ Vb,c . Suppose w is adjacent with vbc . A
contradiction is obtained since G[{w, va , vb , vc , vbc , w}] ≃ G6,11 . Thus w is not adjacent with any
vertex in Vb,c
Claim 4.48. There exists no vertex in V∅ adjacent with a vertex in Va,b ∪ Va,c .
Proof. Let x ∈ {b, c}, y ∈ {b, c} − x, va ∈ Va , vb ∈ Vb and vax ∈ Va,x . Suppose w is adjacent with vax .
Since the induced subgraph G[{w, vax , va , vb , vc }] is isomorphic to G6,5 , then we get a contradiction.
And then the result follows.
Claims 4.46, 4.47 and 4.48 imply that no vertex in V∅ is adjacent with a vertex in G \ V∅ , which
implies that V∅ = ∅. Thus the graph is isomorphic to an induced subgraph of a graph in F11 .
4.6. Case G[Va ∪ Vb ∪ Vc ] = K1,1,1 , where each vertex set Vx = K1 . Let Va = {va }, Vb = {vb },
and Vc = {vc }. By Claim 4.40, the edge sets E(Vx,y , Vx ) and E(Vx,y , Vy ) are empty for x, y ∈ {a, b, c}.
By Claim 4.42, the edge set E(Vx,y , Vz ) is empty for x, y, z ∈ {a, b, c}. Now let vxy ∈ Vx,y . Since
G[{vy , vz , z, x, vx,y }] is isomorphic to P5 which is a forbidden, then Vxy is empty for each pair x, y ∈
{a, b, c}. On the other hand, by Claim 4.46 the edge set V∅ is empty. This graph is isomorphic to
G1 , see Figure 1.i.
24
CARLOS A. ALFARO AND CARLOS E. VALENCIA
4.7. Case Vz = ∅ and G[Vx ∪ Vy ] = Vx + Vy , where Vx = Km , Vy = Kn and m, n ∈ {1, 2}.
Without loss of generality, suppose Vc = ∅, and E(Va , Vb ) is empty. By Claim 4.41, each vertex in
Va,b is adjacent with each vertex in Va ∪ Vb .
Claim 4.49. Let x ∈ {a, b}. If Vx,c 6= ∅, then E(Vx , Vx,c ) = ∅.
Proof. Let y ∈ {a, b} − x, vxc ∈ Vx,c , va ∈ Va and vb ∈ Vb . Suppose vx and vxc are adjacent. There are two possible cases: either vy and vxc are adjacent or not. Since in the first case
G[{a, b, c, va , vb , vxc }] ≃ G6,17 and in the second case G[{va , vb , y, vxc , c}] is isomorphic to P5 , then we
get a contradiction. Thus vx and vxc are not adjacent.
By Claim 4.43, each vertex in Va,c is adjacent with each vertex in Vb , and each vertex in Vb,c is
adjacent with each vertex in Va .
Claim 4.50. Each vertex in Va,b is adjacent with each vertex in Va,c ∪ Vb,c .
Proof. Let vab ∈ Va,b , va ∈ Va and vb ∈ Vb . Suppose there exists vxc ∈ Vxc with x ∈ {a, b} such that
vab vxc ∈
/ E(G). Since G[{va , vb , vab , vxc , c}] is isomorphic to P5 , then we get a contradiction; and the
vertices vab and vxc are adjacent. And the result turns out.
Claim 4.51. Each vertex in Va,c is adjacent with each vertex in Vb,c .
Proof. Suppose there are vac ∈ Va,c and vbc ∈ Vb,c such that vbc vac ∈
/ E(G). Let va ∈ Va and vb ∈ Vb
Since G[{vb , vac , c, vbc , va }] ≃ P5 , then we get a contradiction.
Now we describe the vertex set V∅ , that is, the set of vertices that are not adjacent with any
vertex in {a, b, c}.
Claim 4.52. If w ∈ V∅ , then w cannot be adjacent with any vertex in Va ∪ Vb .
Proof. Let va ∈ Va and vb ∈ Vb . Suppose w ∈ V∅ such that E(w, Va ∪ Vb ) 6= ∅. Consider the following
cases:
(a) w is adjacent only with va , or
(b) w is adjacent with both va and vb .
Cases (a) and (b) are impossible, because G would have P5 as induced subgraph obtained by
G[{w, vb , b, a, va }] and G[{va , w, vb , b, c}], respectively. Thus w is not adjacent with any vertex in
Va ∪ Vb .
Claim 4.53. There is no vertex w ∈ V∅ adjacent with a vertex in Va,b .
Proof. Suppose w ∈ V∅ is adjacent with vab ∈ Va,b . Let va ∈ Va and vb ∈ Vb . Since G[{va , vb , va,b , w, a, c}]
is isomorphic to G6,2 , then we get a contradiction and w and vab are not adjacent.
Claim 4.54. There is no vertex w ∈ V∅ adjacent with a vertex in Va,c ∪ Vb,c .
Proof. Let x ∈ {a, b}, vxc ∈ Vx,c , y ∈ {a, b} − x and vy ∈ Vy . Suppose w ∈ V∅ is adjacent with vxc .
Since G[{vy , y, c, vxc , w}] is isomorphic to P5 , then we get a contradiction and w is not adjacent with
vxc .
Thus there is no edge between W and G \ W , and therefore W is empty. Therefore, the graph is
isomorphic to an induced subgraph of a graph in F11 .
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
25
4.8. Case Vz = ∅ and G[Vx ∪ Vy ] = Vx ∨ Vy , where Vx = K1 , Vy = Km and m ∈ {1, 2}. Without
loss of generality, suppose Vc = ∅, and E(Va , Vb ) is complete. By Claim 4.40, E(Va,b , Vx ) = ∅ for
x ∈ {a, b}.
Claim 4.55. Let x ∈ {a, b}. Either E(Vx,c , Va ) induces a complete bipartite graph and E(Vx,c , Vb ) =
∅, or E(Vx,c , Va ) = ∅ and E(Vx,c , Vb ) induces a complete bipartite graph.
Proof. Let va ∈ Va , vb ∈ Vb , vxc ∈ Vx,c and y ∈ {a, b} − x. First note that vxc cannot be not adjacent
with va and vb at the same time, because otherwise G[{va , vb , y, c, vxc }] would be isomorphic to P5 .
Also vxc cannot be adjacent with va and vb at the same time, because otherwise G[{a, b, c, va , vb , vxc }]
would be isomorphic to G6,25 . Thus vxc is adjacent only with one vertex either va or vb .
′ ∈ V . If v
′
Suppose vxc is adjacent with vy . Let vy′ ∈ Vy and vxc
x,c
xc is not adjacent with vy , then
G[{vy , vy′ , x, c, vx , vxc }] is isomorphic to G6,21 . Thus E(Vy , vxc ) induces a complete bipartite graphs.
′ are not adjacent, then G[{v , v , v , c, v ′ }] is isomorphic to P .
On the other hand, if vy and vxc
a b xc
5
xc
Then E(Vx,c , Vy ) induces a complete bipartite graph.
′ ∈ V . If v ′ and v
Suppose vxc is adjacent with vx . Let vx′ ∈ Vx and vxc
x,c
xc are not adjacent, then
x
′
G[{va , vb , vx , x, c, vxc }] is isomorphic to G6,22 . Thus E(Vx , vxc ) induces a complete bipartite graph.
′ and v are not adjacent, then G[{v , v , v , c, v ′ }] is isomorphic to P .
On the other hand, if vxc
5
x
y x xc
xc
Thus E(Vx,c , Vx ) induces a complete bipartite graph.
Claim 4.56. If Va,c and Vb,c are not empty, then each vertex in Va,c ∪ Vb,c is adjacent with each
vertex in either Va or Vb .
Proof. Let vac ∈ Va,c , vbc ∈ Vb,c , va ∈ Va and vb ∈ Vb .
Suppose vac is adjacent with va , and vbc is adjacent with vb . Then there are two cases: either vac
and vbc are adjacent or not. In the first case G[{a, b, c, vb , vac , vbc }] is isomorphic to G6,24 ; then this
case is impossible. And in the second case G[{a, b, va , vb , vac , vbc }] is isomorphic to G6,21 ; then this
case is not possible.
Suppose vac is adjacent with vb and vbc is adjacent with va . Then there are two cases: either vac
and vbc are adjacent or not. In the first case G[{a, b, c, va , vb , vac , vbc }] is isomorphic to G7,13 ; then
this case is impossible. And in the second case G[{a, b, c, vb , vac , vbc }] is isomorphic to G6,17 ; then
this case is not possible.
Thus vac and vbc are adjacent with the same vertex: either va or vb . And the result follows.
Claim 4.57. If Va,c and Vb,c are not empty, then the set E(Vac , Vbc ) induces a complete bipartite
graph.
Proof. Let x ∈ {a, b} and y ∈ {a, b} − x, vac ∈ Va,c , vbc ∈ Vb,c , va ∈ Va and vb ∈ Vb . Suppose vxc and
Vyc are not adjacent, and vxc and vyc are adjacent with vx . Since G[{x, y, vx , vy , vx,c , vy,c }] ≃ G6,15 ,
we get a contradiction. And then E(Vac , Vbc ) induces a complete bipartite graph.
Claim 4.58. Let x ∈ {a, b}. If Vx,c 6= ∅, then E(Vx,c , Va,b ) induces a complete bipartite graph.
Proof. Let y ∈ {a, b} − x, vxc ∈ Vx,c , vab ∈ Va,b , va ∈ Va and vb ∈ Vb . Suppose vxc and vab are not
adjacent. There are two cases: either vxc is adjacent with vx or vxc is adjacent with vy . If vxc is
adjacent with vx , then the induced subgraph G[{vab , y, vy , vx , vxc }] is isomorphic to P5 ; which is a
contradiction. Then this case is impossible. On the other hand, if vxc is adjacent with vy , then the
induced subgraph G[{vac , va , vb , a, b, vxc }] is isomorphic to G6,9 ; which is a contradiction. Thus this
case is also impossible, and therefore E(Vx,c , Va,b induces a complete bipartite graph.
26
CARLOS A. ALFARO AND CARLOS E. VALENCIA
Now let us describe V∅ . By Claim 4.46, there exists no vertex in V∅ adjacent with a vertex in
Va ∪ Vb .
Claim 4.59. There is no vertex in V∅ adjacent with a vertex in Va,b .
Proof. Let vab ∈ Va,b and w ∈ V∅ . Suppose vab and w are adjacent. Since G[{w, va,b , a, va , vb }] is
isomorphic to P5 , then we get a contradiction. And therefore vab and w are not adjacent.
Claim 4.60. Let x ∈ {a, b}. There is no vertex in V∅ adjacent with a vertex in Vx,c .
Proof. Let y ∈ {a, b} − x, vxc ∈ Vx,c , vab ∈ Va,b , va ∈ Va and vb ∈ Vb . Suppose the vertex w ∈ V∅ is
adjacent with vxc . There are two cases: either vxc is adjacent with x or vxc is adjacent with y. First
case is impossible since G[{w, vxc , vx , vy , y}] is isomorphic to P5 ; which is forbidden. And second
case cannot occur because G[{w, y, vxc , c, va , vb }] ≃ G6,9 . Then it follows that w is adjacent with no
vertex in Vx,c .
Thus by previous Claims, the vertex set V∅ is empty, because there is no vertex in V∅ adjacent
with a vertex in G \ V∅ . Then G is isomorphic to an induced subgraph of a graph in F11 .
4.9. Case Vy ∪ Vz = ∅ and Vx is K1 or K2 . Without loss of generality, suppose Vb = Vc = ∅ and
Va = {u1 , u2 }.
Claim 4.61. Let x ∈ {b, c}. If Va,x 6= ∅, then either E(Va , Va,x ) induces a complete bipartite graph
or E(Va , Va,x ) is empty.
Proof. Let u ∈ Va . Suppose there exist v1 , v2 ∈ Va,x such that uv1 ∈ E(G) and uv2 ∈
/ E(G). Since
the induced subgraph G[{a, b, c, u, v1 , v2 }] is isomorphic to G6,16 , then we get a contradiction. Thus
this case is not possible and therefore u is adjacent with either each vertex in Va,x or no vertex in
Va,x . Now we are going to discard the possibility that E(u1 , Va,x ) induces a complete bipartite graph
and E(u2 , Va,x ) = ∅. Suppose there is a vertex v ∈ Va,x such that u1 v ∈ E(G) and u2 v ∈
/ E(G).
Since the induced subgraph G[{c, b, v, u1 , u2 }] is isomorphic to P5 , then we get a contradiction. Thus
either E(Va , Va,x ) induces a complete bipartite graph or E(Va , Va,x ) is empty.
Claim 4.62. Let u ∈ Va . If Vb,c 6= ∅, then E(u, Vb,c ) satisfies only one of the following:
• it induces a complete bipartite graph,
• it is an empty edge set, or
• it induces a complete bipartite graph minus an edge.
Proof. Suppose there exist v1 , v2 , v3 ∈ Vb,c such that uv1 ∈ E(G) and uv2 , uv3 ∈
/ E(G). Since the
induced subgraph G[{a, b, u, v1 , v2 , v3 }] is isomorphic to G6,3 , then we get a contradiction. Thus this
case is not possible and the result follows.
Claim 4.63. If Va = {u1 , u2 } and Vb,c 6= ∅, then E(Va , Vb,c ) satisfies one of the following:
•
•
•
•
•
it
it
it
it
it
induces a complete bipartite graph,
is an empty edge set,
induces a complete bipartite graph minus an edge,
induces a perfect matching and |Va | = |Vb,c | = 2, or
induces a complete bipartite graph minus two edges u1 v and u2 v, where v ∈ Vb,c .
Proof. Since cases where |Vb,c | ≤ 2 can be checked easily with a computer algebra system or with
similar arguments to the rest of the proof, then we asume |Vb,c | ≥ 3. By Claim 4.62, we only have
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
27
to check the possibilities of the edge sets E(u1 , Vb,c ) and E(u2 , Vb,c ). The possible cases we have to
discard are the following:
• E(u1 , Vb,c ) = ∅ and E(u2 , Vb,c ) induces a complete bipartite graph,
• E(u1 , Vb,c ) = ∅ and E(u2 , Vb,c ) induces a complete bipartite graph minus an edge, and
• each edge set E(u1 , Vb,c ) and E(u2 , Vb,c ) induces a complete bipartite graph minus an edge
and the two removed edges don’t share a common vertex.
Let v1 , v2 , v3 ∈ Vb,c . Suppose we are in the first case. Thus u1 is adjacent with no vertex in Vb,c ,
and u2 is adjacent with each vertex in Vb,c . Since the induced subgraph G[{a, b, u1 , u2 , v1 , v2 }] is
isomorphic to G6,15 , then we get a contradiction and this case is not possible. Suppose we are in the
second case. Thus u1 is adjacent with no vertex in Vb,c , and u2 is adjacent only with each vertex
in Vb,c − v1 . Since the induced subgraph G[{v1 , b, v2 , u2 , u1 }] is isomorphic to P5 , then we get a
contradiction and this case is not possible. Finally, suppose we are in the third case. Thus u1 is
adjacent with each vertex in Vb,c − v1 , and u2 is adjacent with each vertex in Vb,c − v2 . Since the
induced subgraph G[{a, u1 , u2 , v1 , v2 , v3 }] is isomorphic to G6,5 , then we get a contradiction and this
case is not possible. And the result turns out.
Claim 4.64. If E(Va , Va,b ) and E(Va , Va,c ) are empty, then E(Va,b , Va,c ) induces a complete bipartite
graph.
Proof. Let vab ∈ Va,b and vac ∈ Va,c . Suppose vab and vac are not adjacent. Since the induced
subgraph G[{a, b, c, u1 , vab , vac }] is isomorphic to G6,12 , then we get a contradiction. And thus each
vertex in Va,b is adjacent with each vertex in Va,c .
Claim 4.65. If each of the edge sets E(Va , Va,b ) and E(Va , Va,c ) induces a complete bipartite graph,
then |Va | = |Va,b | = |Va,c | = 1 and E(Va,b , Va,c ) = ∅.
Proof. First we will prove that E(Va,b , Va,c ) = ∅. Let vab ∈ Va,b and vac ∈ Va,c . Suppose vab and vac
are adjacent. Since the induced subgraph G[{a, b, c, vab , vac , va }] is isomorphic to G6,26 , then we get
a contradiction. And thus E(Va,b , Va,c ) = ∅.
′ ∈V
Now let x ∈ {b, c}. Suppose Va,x has cardinality at least 2. Let y ∈ {b, c} − x, vax , vax
a,x and
′
v ∈ Va,y . Then vax v, vax v ∈
/ E(G). Since the induced subgraph G[{a, y, vax , vax , va , v}] is isomorphic
to G6,16 , then we get a contradiction. And then |Va,b | = |Va,c | = 1
Finally suppose that |Va | = 2. Let vab ∈ Va,b and vac ∈ Va,c . Since G[{vab , vac , u1 , u2 , a, x}] is
isomorphic to G6,23 , then we get a contradiction. Thus Va has cardinality at most 1.
Claim 4.66. Let x ∈ {b, c} and y ∈ {b, c} − x. If E(Va , Va,x ) = ∅ and E(Va , Va,y ) induces a complete
bipartite graph, then E(Va,b , Va,c ) induces a complete bipartite graph.
Proof. Let vab ∈ Va,b and vac ∈ Va,c . Suppose vab and vac are not adjacent. Since the induced subgraph G[{va , vay , y, x, vax }] is isomorphic to P5 , then we get a contradiction. And thus E(Va,b , Va,c )
induces a complete bipartite graph.
Claim 4.67. Let x ∈ {b, c}. If Va,x 6= ∅, Vb,c 6= ∅ and E(Va , Va,x ) = ∅, then only one of the following
statements is true:
• E(Va , Vb,c ) = ∅ and E(Va,x , Vb,c ) induces a complete graph, or
• each edge set E(Va , Vb,c ) and E(Va,x , Vb,c ) induce a complete bipartite graph.
Proof. We will analyze the following four cases:
• when E(Va , Vb,c ) = ∅,
28
CARLOS A. ALFARO AND CARLOS E. VALENCIA
• when E(Va , Vb,c ) induces a complete bipartite graph,
• when there exist u ∈ Va and v1 , v2 ∈ Vb,c such that uv1 ∈ E(G) and uv2 ∈
/ E(G), and
• when there exists v ∈ Vb,c such that u1 v ∈ E(G) and u2 v ∈
/ E(G), where u1 , u2 ∈ Va .
Consider the first case. Let vax ∈ Va,x and vbc ∈ Vb,c . Suppose vax and vbc are not adjacent. Since
the induced subgraph G[{a, b, c, u1 , vax , vbc }] is isomorphic to G6,14 , then we get a contradiction. And
thus E(Va,x , Vb,c ) induces a complete bipartite graph.
Now consider the second case. Let vax ∈ Va,x and vbc ∈ Vb,c . Suppose vax and vbc are not
adjacent. Since the induced subgraph G[{a, b, c, u1 , vax , vbc }] is isomorphic to G6,17 , then we get a
contradiction. And thus E(Va,x , Vb,c ) induces a complete bipartite graph.
In what follows we will prove that the last two cases are impossible which implies that the rest
possibilities of Claim 4.63 are impossible.
Now consider the third case. Let vax ∈ Va,x . Suppose there exist u ∈ Va and v1 , v2 ∈ Vb,c such
that uv1 ∈ E(G) and uv2 ∈
/ E(G). There are four possible cases:
•
•
•
•
vax v1 , vax v2 ∈ E(G),
vax v1 , vax v2 ∈
/ E(G),
vax v1 ∈ E(G) and vax v2 ∈
/ E(G), or
vax v1 ∈
/ E(G) and vax v2 ∈ E(G).
Since in first case G[{a, b, c, u, vax , v1 , v2 }] ≃ G7,10 , in second case G[{b, c, u, vax , v1 , v2 }] ≃ G6,10 , in
third case G[{b, c, u, vax , v1 , v2 }] ≃ G6,14 and in fourth case G[{u, v1 , vy , v2 , vax }] ≃ P5 , then we get a
contradiction.
In the last case, we get a contradiction because otherwise by first case: va,x v ∈
/ E(G), but by
second case: va,x v ∈ E(G). Thus this case is impossible.
Claim 4.68. Let x ∈ {b, c}. If Va,x 6= ∅, Vb,c 6= ∅, and E(Va , Va,x ) induces a complete bipartite
graph, then the edge sets E(Va , Vb,c ) and E(Va,x , Vb,c ) induce complete bipartite graphs.
Proof. Let vax ∈ Va,x , v ∈ Vb,c and u ∈ Va . Suppose uv ∈
/ E(G). There are two cases: either
vvax ∈ E(G) or vvax ∈
/ E(G). Since the induced subgraph G[{a, b, c, u, v, vax }] is isomorphic to G6,24
and G6,22 , in the first case and in the second case, respectively, then we get a contradiction. Thus
uv ∈ E(G) and therefore the edge set E(Va , Vb,c ) induces a complete bipartite graph.
Now suppose vvax ∈
/ E(G). By previous result, uv ∈ E(G). Since the induced subgraph
G[{a, b, c, u, v, vax }] is isomorphic to G6,25 , then we get a contradiction and thus vvax ∈ E(G).
Therefore E(Va,x , Vb,c ) induces a complete bipartite graph.
By previous Claims we have the following cases:
(1) E(Va , Va,b ∪ Vac ∪ Vb,c ) = ∅, and each edge set E(Va,b , Va,c ) and E(Va,b ∪ Va,c , Vb,c ) induces a
complete bipartite graph,
(2) E(Va , Va,b ∪ Vac ) = ∅, and each edge set E(Va,b , Va,c ) and E(Va ∪ Va,b ∪ Va,c , Vb,c ) induces a
complete bipartite graph,
(3) E(Va , Va,b ) = ∅, and each edge set E(Va ∪ Va,b , Va,c ) and E(Va ∪ Va,b ∪ Va,c , Vb,c ) induces a
complete bipartite graph,
(4) E(Va,b , Va,c ) = ∅, and each edge set E(Va , Va,b ∪ Va,c ∪ Vb,c ) and E(Va,b ∪ Va,c , Vb,c ) induces a
complete bipartite graph and |Va | = |Va,b | = |Va,c | = 1,
(5) Vb,c = ∅, E(Va , Va,b ∪ Vac ) = ∅, and E(Va,b , Va,c ) induces a complete bipartite graph,
(6) Vb,c = ∅, E(Va , Va,x ) = ∅, and E(Va ∪ Va,x , Va,y ) induces a complete bipartite graph, where
x ∈ {a, b} and y ∈ {a, b} − x,
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
29
(7) Vb,c = ∅, E(Va , Va,b ∪ Va,c ) induce a complete bipartite graph, E(Va,b , Va,c ) = ∅ and |Va | =
|Va,b | = |Va,c | = 1,
(8) Va,y = ∅, E(Va , Va,x ∪ Vb,c ) = ∅, and E(Va,x , Vb,c ) induces a complete bipartite graph, where
x ∈ {a, b} and y ∈ {a, b} − x,
(9) Va,y = ∅, E(Va , Va,x ) = ∅, and E(Va ∪ Va,x , Vb,c ) induces a complete bipartite graph, where
x ∈ {a, b} and y ∈ {a, b} − x,
(10) Va,y = ∅, E(Va , Va,x ∪ Vb,c ) and E(Va,x , Vb,c ) induce a complete bipartite graph, where x ∈
{a, b} and y ∈ {a, b} − x,
(11) Va,b ∪ Va,c = ∅, and E(Va , Vb,c ) induces a complete bipartite graph,
(12) Va,b ∪ Va,c = ∅ and E(Va , Vb,c ) = ∅,
(13) Va,b ∪ Va,c = ∅, and E(Va , Vb,c ) induces a complete bipartite graph minus an edge,
(14) Va,b ∪ Va,c = ∅, |Vb,c | ≥ 2 and E(Va , Vb,c ) induces a complete bipartite graph minus two edges
u1 v and u2 v, where v ∈ Vb,c ,
(15) Va,b ∪ Va,c = ∅, |Va | = |Vb,c | = 2 and E(Va , Vb,c ) induces a perfect matching,
(16) Va,y ∪ Vb,c = ∅ and E(Va , Va,x ) induces a complete bipartite graph, where x ∈ {a, b} and
y ∈ {a, b} − x, and
(17) Va,y ∪ Vb,c = ∅ and E(Va , Va,x ) = ∅, where x ∈ {a, b} and y ∈ {a, b} − x.
Now we describe V∅ , the set of vertices that are not adjacent with any vertex in {a, b, c}. Let
w ∈ V∅ . The vertex w is adjacent with a vertex in Va ∪ Va,b ∪ Va,c ∪ Vb,c , because otherwise the
shortest path from w to {a, b, c} would contains the graph P5 as induced subgraph.
Claim 4.69. If Va,x 6= ∅ for some x ∈ {b, c}, then E(V∅ , Va ∪ Va,x ) = ∅.
Proof. Let w ∈ V∅ , vax ∈ Va,x and u ∈ Va . Suppose w is adjacent with u or vax . Then there are
three possible cases:
(a) w is adjacent only with u,
(b) w is adjacent only with vax , or
(c) w is adjacent with both vertices u and vax .
First consider case (a). This case has two subcases: either u is adjacent with vax or not. If they are
adjacent, then G[{u, w, vax , b, c}] is isomorphic to P5 . Since it is forbidden, then va is not adjacent
with vax . On the other hand, if they are not adjacent, then G[{a, b, c, w, u, vax }] is isomorphic to
G6,7 . Therefore, case (a) is impossible. Now consider case (b). There are two possible cases: either
u is adjacent with vax or not. If they are adjacent, then G[{a, b, c, w, u, vax }] is isomorphic to G6,14 ;
which is forbidden. Thus u is not adjacent with vax . But if they are not adjacent, then we have that
G[{a, b, c, w, u, vax }] is isomorphic to G6,10 . Thus case (b) cannot occur. Finally, consider case (c).
The subcases are: either u is adjacent to vax or not. If they are adjacent, then G[{a, b, c, w, u, vax }] is
isomorphic to G6,22 ; which is forbidden. Then u is not adjacent with vax But if they are not adjacent,
then the induced subgraph G[{u, w, vax , b, c}] is isomorphic to P5 ; which is forbidden. Thus, we get
that case (c) is impossible. And therefore, w is not adjacent with u or vax . From which the result
follows.
Claim 4.70. Let x ∈ {b, c}. If Va,x 6= ∅, then E(V∅ , Vb,c ) = ∅. Moreover, if Va,x 6= ∅, then V∅ = ∅.
Proof. Let w ∈ V∅ , vax ∈ Va,x , va ∈ Va and vbc ∈ Vb,c . Suppose w is adjacent with vbc . By Claim
4.69, we have that w is not adjacent with a vertex in Va ∪ Va,x . By Claims 4.67 and 4.68, the vertex
vax is adjacent with vbc , and one of the following three cases occur:
(a) va is adjacent with vax and vbc ,
(b) va is adjacent only with vbc , and
30
CARLOS A. ALFARO AND CARLOS E. VALENCIA
(c) va is not adjacent with both vax and vbc .
Case (a) is not possible because the induced subgraph G[{va , vax , b, c, vbc , w}] would be isomorphic
to G6,12 ; which is impossible. In case (b), we have that G[{va , vax , b, c, vbc , w}] is isomorphic to
G6,4 . Since it is forbidden, then this case is not possible. Finally, case (c) is not possible since
G[{va , a, x, vbc , w}] is isomorphic to P5 that is forbidden. Thus w is not adjacent with vbc . And there
is no vertex in V∅ adjacent with a vertex in G \ V∅ . Therefore V∅ is empty.
Thus by previous Claim, in cases (1) to (10), (16) and (17) the vertex set V∅ is empty. The cases
when Va,b ∪ Va,c = ∅ (cases (4) and (7)) correspond to a graph isomorphic to an induced subgraph
of F12 . The rest of these cases correspond to a graph isomorphic to an induced subgraph of F11 .
Claim 4.71. Let wa ∈ V∅ such that wa is adjacent with u ∈ Va , and wa is not adjacent with a vertex
in Vb,c . Let wbc ∈ V∅ such that wbc is adjacent with v ∈ Vb,c and wbc is not adjacent with a vertex in
Va . Let w ∈ V∅ such that w is adjacent with u′ ∈ Va and v ′ ∈ Vb,c . If E(Va , Vb,c ) induces a complete
bipartite graph, then no two vertices of {wa , wbc , w} exist at the same time.
Proof. Suppose wa and wbc exist at the same time. There there are two possible cases: either wa
and wbc are adjacent or not. If wa and wbc are adjacent, then G[{b, a, u, wa , wbc }] is isomorphic to
P5 ; which is impossible. If wa and wbc are not adjacent, then G[{b, c, u, v, wa , wbc }] is isomorphic to
G6,6 ; which is impossible. Then wa and wbc do not exist at the same time.
Suppose wa and w exist at the same time. There there are two possible cases: either u = u′ or
u 6= u′ . Suppose u = u′ , then we have two possible cases: either wa and w are adjacent or not. If
wa and w are adjacent, then G[{wa , w, v ′ , b, a}] is isomorphic to P5 ; which is impossible. But if wa
and w are not adjacent, then G[{b, c, u, v ′ , wa , w}] is isomorphic to G6,11 ; which is impossible. Then
u 6= u′ . Thus suppose u 6= u′ . Note that in the induced subgraph G[{a, b, c, u, v ′ , wa , w}], the vertex
w is only adjacent with v ′ , and wa is only adjacent with u. Then applying previous (wa , wbc ) case
in this induced subgraph, we get that w and wa do not exist at the same time.
Suppose wbc and w exist at the same time. There are two possible cases: either v = v ′ or v 6= v ′ .
Suppose v = v ′ , then we have two possible cases: either wbc and w are adjacent or not. If wa and
wbc are adjacent, then G[{wbc , w, u, a, b}] is isomorphic to P5 ; which is impossible. But if wa and
wbc are not adjacent, then G[{b, c, u, v, wbc , w}] is isomorphic to G6,11 ; which is impossible. Then
v 6= v ′ . Suppose v 6= v ′ . Note that in the induced subgraph G[{a, b, c, u, v, wa , w}], the vertex w is
only adjacent with u, and wbc is only adjacent with v. Thus by applying first case in this induced
subgraph, we get that w and wbc do not exist at the same time.
Claim 4.72. Let w ∈ V∅ , u ∈ Va and v ∈ Vb,c . If E(Va , Vb,c ) induces a complete bipartite graph and
w is adjacent with u and v, then E(w, Va ∪ Vb,c ) induces a complete graph.
Proof. First we see that if v ′ ∈ Vb,c − v, then w is adjacent with v ′ . Suppose w and v ′ are not
adjacent. Since the induced subgraph G[{a, b, u, v, v ′ , w}] is isomorphic to the forbidden graph G6,15 ,
then we get a contradiction. Thus w is adjacent with each vertex in Vb,c . Now we see that if
u′ ∈ Va − u, then w is adjacent with u′ . Suppose w and u′ are not adjacent. Since the induced
subgraph G[{a, b, u, u, u′ , w}] is isomorphic to G6,15 , then we get a contradiction. Therefore, w is
adjacent with each vertex in Va ∪ Vb,c .
Claim 4.73. If |Va | = 2 and E(V∅ , Va ) 6= ∅, then either each vertex in V∅ is adjacent only with one
vertex in Va and V∅ is a clique, or each vertex in V∅ is adjacent with u1 and u2 , and V∅ is trivial.
Proof. Let w, w′ ∈ V∅ and i, j ∈ {1, 2} with i 6= j. By proving that the following cases are not possible, it follows that the only possible cases are that either E({u1 , u2 }, {w, w′ }) is equal to {wui , w′ uj }
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
31
and ww′ ∈ E(G), or E({u1 , u2 }, {w, w′ }) is equal to {wui , wuj , w′ ui , w′ uj } and ww′ ∈ E(G). Which
implies the result.
(a)
(b)
(c)
(d)
(e)
(f)
ww′ , wuj , w′ uj ∈
/ E(G) and wui , w′ ui
ww′ , wuj , w′ ui ∈
/ E(G) and wui , w′ uj
′
wuj , w ui ∈
/ E(G) and ww′ , wui , w′ uj
′
′
ww , w uj ∈
/ E(G) and wui , wuj , w′ ui
′
w uj ∈
/ E(G) and ww′ , wui , wuj , w′ ui
w′ uj , ww′ , wui , wuj , w′ ui ∈ E(G).
∈ E(G),
∈ E(G),
∈ E(G),
∈ E(G),
∈ E(G), and
Since in cases (a), (b) and (d) the induced subgraph G[{a, b, u1 , u2 , w, w′ }] is isomorphic to G6,2 , G6,8 ,
G6,10 , respectively, then these cases are impossible. On the other hand, in case (c) the induced graph
G[{w, w′ , uj , a, b}] is isomorphic to P5 : which is impossible. Also in case (e) the induced subgraph
G[{w′ , w, uj , a, b}] is isomorphic to P5 ; which is not possible. Finally in case (f) the induced subgraph
G[{a, b, c, u1 , u2 , w, w′ }] is isomorphic to G7,14 ; which is not possible.
Claim 4.74. If E(Va , Vb,c ) induces a complete bipartite graph, and each vertex in V∅ is adjacent only
with one vertex in Va , then there is no vertex in V∅ adjacent with a vertex in Vb,c .
Proof. Let w ∈ V∅ , v ∈ Vb,c . Suppose w is adjacent with u1 and v, but w is not adjacent with u2 .
Since the induced subgraph G[{a, b, w, u1 , u2 , v}] is isomorphic to G6,17 , then we get a contradiction
and the result follows.
Claim 4.75. Let w ∈ V∅ and u ∈ Va . If E(Va , Vb,c ) induces a complete bipartite graph, E(w, Vb,c ) 6= ∅
and E(w, Va ) = ∅, then each vertex in V∅ is adjacent with only one vertex in Vb,c , and V∅ is a clique
of cardinality at most 2.
Proof. Let v, v ′ ∈ Vb,c . Suppose w is adjacent with v and v ′ . Since G[{a, b, c, u, v, v ′ , w}] is isomorphic
to G7,9 , then we get a contradiction. And we have that w is adjacent only with one vertex in Vb,c
Now let w, w′ ∈ V∅ and v, v ′ ∈ Vb,c such that wv ∈ E(G) and w′ v ′ ∈
/ E(G). Suppose v 6= v ′ . There
′
′
are two cases: either w and w are adjacent or not. If w and w are adjacent, then the induced
subgraph G[{b, c, v, u, w, w′ }] is isomorphic to G6,6 , then we get a contradiction and ww′ ∈
/ E(G).
Now if w and w′ are not adjacent, then the induced subgraph G[{w, v, u, v ′ , w′ }] is isomorphic to P5
and we get a contradiction. Thus v = v ′ .
Finally, let w, w′ ∈ V∅ adjacent with v ∈ Vb,c . Suppose w and w′ are not adjacent. Since
the induced subgraph G[{a, b, v, u, w, w′ }] is isomorphic to G6,3 , then we get a contradiction and
ww′ ∈ E(G).
In case (11), by Claims 4.71, 4.72, 4.73, 4.74 and 4.75, we have the following possible cases:
• each vertex in V∅ is adjacent with each vertex in Va ∪ Vb,c , and V∅ is a trivial graph,
• each vertex in V∅ is adjacent only with a vertex in Va , and V∅ is a clique of cardinality at
most 2, or
• each vertex in V∅ is adjacent only with a vertex in Vb,c , and V∅ is a clique of cardinality at
most 2.
Each of these cases corresponds to a graph isomorphic to an induced subgraph of F11 .
Remark 4.76. Let w ∈ V∅ , u ∈ Va and v ∈ Vb,c . If uv ∈
/ E(G), and w adjacent with u or v, then w
is adjacent with both vertices u and v.
In case (12), by Claim 4.73 and Remark 4.76, we have that each vertex in V∅ is adjacent with
each vertex in Va ∪ Vb,c , and
32
CARLOS A. ALFARO AND CARLOS E. VALENCIA
• V∅ is a clique or a trivial graph when |Va | = 1, or
• V∅ is a clique when |Va | = 2.
It is not difficult to see that each of these graphs are isomorphic to an induced subgraph of F11
Claim 4.77. Let w ∈ V∅ , u ∈ Va and v ∈ Vb,c . Suppose u is adjacent with each vertex in Vb,c − v.
If E(w, Vb,c ∪ {u}) 6= ∅, then one of the following statements holds:
• each vertex in V∅ is adjacent with u and v,or
• E(w, Vb,c ∪ {u}) induces a complete bipartite graph and |Vb,c − v| = 1.
Proof. First case follows by Remark 4.76. Now we check when w is adjacent with a vertex in Vb,c − v.
Let v ′ ∈ Vb,c . Then u is adjacent with v ′ . Suppose w is adjacent only with v ′ . Since the induced
subgraph G[{a, b, u, v, v ′ , w}] is isomorphic to G6,9 , then we have a contradiction and w is adjacent
also with u or v. Thus by Remark 4.76, w is adjacent with u, v and v ′ . Finally by applying Claim
4.72 to the induced subgraph G[{w, u} ∪ (Vb,c − v)], we get that w also is adjacent with each vertex
in Vb,c − v. Thus w is adjacent with each vertex in {u} ∪ Vb,c .
Suppose the cardinality of Vb,c − v is at least 2. Take v, v ′ ∈ Vb,c − v. Thus w is adjacent with
v, v ′ and v ′′ . Since the induced subgraph G[{u, v, v ′ , v ′′ , a, w}] is isomorphic to G6,5 , then we get a
contradiction. Thus the cardinality of Vb,c − v is at most 1.
Claim 4.78. Let w ∈ V∅ and v ∈ Vb,c . Suppose u1 is adjacent with each vertex in Vb,c , and u2 is
adjacent with each vertex in Vb,c − v. If E(w, Vb,c ) 6= ∅, then either w is adjacent only with u2 and
v, or |Vb,c | = 1 and E(w, Va ∪ Vb,c ) induces a complete bipartite graph.
Proof. First case follows by Remark 4.76. Now we prove that if w is adjacent with u1 , then w
is adjacent with u2 and v. Suppose w is adjacent only with u1 . Since the induced subgraph
G[{a, b, u1 , u2 , v, w}] is isomorphic to G6,11 , then we have a contradiction and w is adjacent also with
u2 or v. Thus by Remark 4.76, w is adjacent with u1 , u2 and v. Finally, suppose |Vb,c | ≥ 2. Let
v ′ ∈ Vb,c . By applying Claim 4.72 to the induced subgraph G[{w, u1 } ∪ Vb,c ], we get that w also
is adjacent with each vertex in Vb,c . Thus w is adjacent with each vertex in Va ∪ Vb,c . But since
G[{a, u1 , u2 , v, v ′ , w}] is isomorphic to G6,18 , then we get a contradiction, and |Vb,c | = 1.
Let u ∈ Va and v ∈ Vb,c such that uv ∈
/ E(G). Therefore, in case (13), by Claims 4.73, 4.77 and
4.78, we have that one of the following cases holds:
• V∅ is a clique of cardinality at most 2, and each vertex in V∅ is adjacent with u and v,
• Va = {u}, |Vb,c \ {v}| ≤ 1, V∅ is trivial and each vertex in V∅ is adjacent with each vertex in
{u} ∪ Vb,c , or
• |Va | = 2, V∅ is trivial and Vb,c = {v} and each vertex in V∅ is adjacent with u1 , u2 and v.
It can be checked that each of these graphs is isomorphic to an induced subgraph or a graph in F11 .
Claim 4.79. Let w ∈ V∅ and v ∈ Vb,c such that v is not adjacent with u1 and u2 . Suppose |Vb,c | ≥ 2
and E(Va , Vb,c − v) induces a complete bipartite graph. If E(w, Vb,c ) 6= ∅, then V∅ = {w}, and w is
adjacent with u1 , u2 and v.
Proof. By Remark 4.76, we have that if w is adjacent with one vertex of {v, u1 , u2 }, then w is
adjacent with v, u1 and u2 . Now we see that w is not adjacent to any other vertex. Suppose w is
adjacent with v ′ ∈ Vb,c − v. First consider the case when w is not adjacent with v, u1 and u2 . Since
G[{b, v, v ′ , u1 , u2 , w}] is isomorphic to G6,6 , then we get a contradiction and thus w is adjacent with
v, v ′ , u1 , and u2 . But if w is adjacent with v, v ′ , u1 , and u2 , then G[{b, v, v ′ , u1 , u2 , w}] is isomorphic
to G6,20 ; which is impossible. Then w is not adjacent with any vertex in Vb,c − v.
SMALL CLIQUE NUMBER GRAPHS WITH THREE TRIVIAL CRITICAL IDEALS
33
Suppose there exist another vertex w′ ∈ V∅ . By previous discussion w′ is adjacent with v, u1
and u2 . Since G[{a, b, w, w′ , v, v ′ }] is isomorphic to G6,1 , then we get a contradiction. Thus V∅ has
cardinality at most 1.
Let v ∈ Vb,c such that u1 v, u2 v ∈
/ E(G). Therefore, in case (14), by Claims 4.73 and 4.79, we have
that V∅ = {w} and w is adjacent with u1 , u2 and v. Then G is isomorphic to an induced subgraph
of F11 .
Claim 4.80. Let Vb,c = {v, v ′ } such that vu1 , v ′ u2 ∈ E(G) and vu2 , v ′ u1 ∈
/ E(G). If w ∈ V∅ and
E(w, Va ∪ Vb,c ) 6= ∅, then w is adjacent with u1 , u2 , v and v ′ .
Proof. Since w is adjacent with a vertex in {u1 , u2 , v, v ′ }, then w is adjacent with u1 and v ′ , or w is
adjacent with u2 and v. Suppose w is adjacent with u1 and v ′ , but w is not adjacent with u2 and
v. Since G[{a, u1 , u2 , v, v ′ , w}] is isomorphic to G6,11 , then we get a contradiction, and therefore w
is also adjacent with u2 and v.
Let Vb,c = {v, v ′ } such that vu1 , v ′ u2 ∈ E(G) and vu2 , v ′ u1 ∈
/ E(G). Thus in case (15), by Claims
4.73 and 4.80, we have that V∅ is trivial and each vertex in V∅ is adjacent only with u1 , u2 , v and v ′ .
Therefore, G is isomorphic to an induced subgraph of F11 .
References
[1] C.A. Alfaro and C.E. Valencia, Graphs with two trivial critical ideals, to appear in Discrete Appl. Math.
[2] C.A. Alfaro, H.H. Corrales and C.E. Valencia, Critical ideals of signed graphs with twin vertices, preprint.
[3] R. Bacher, P. de la Harpe and T. Nagnibeda, The lattice of integral flows and the lattice of integral cuts on a
finite graph, Bull. Soc. Math. France 125 (1997) 167-198.
[4] N. Biggs, Chip-firing and the critical group of a graph, J. Alg. Combin. 9 (1999) 25-46.
[5] J.A. Bondy and U.S.R. Murty, Graph Theory, Grad. Texts in Math., vol. 244, Springer, 2008.
[6] W.H. Chan, Y. Hou, W.C. Shiu, Graphs whose critical groups have larger rank, Acta Math. Sinica 27 (2011)
1663–1670.
[7] R. Cori and D. Rossin, On the sandpile group of dual graphs, European J. Combin. 21 (4) (2000) 447–459.
[8] H.H. Corrales and C.E. Valencia, On the critical ideals, to appear.
[9] D.R. Grayson and M.E. Stillman, Macaulay2, a software system for research in algebraic geometry, available at
www.math.uiuc.edu/Macaulay2/.
[10] R. Grone, R. Merris and W. Watkins, Laplacian unimodular equivalence of graphs. In: R. Brualdi, S. Friedland
and V. Klee (Eds.) Combinatorial and Graph-Theoretical Problems in Linear Algebra, Springer-Verlag (1993)
175-180.
[11] C. Godsil and G. Royle, Algebraic Graph Theory, Grad. Texts in Math., vol. 207, Springer-Verlag, New York,
2001.
[12] N. Jacobson, Basic Algebra I, Second Edition, W. H. Freeman and Company, New York, 1985.
[13] D.J. Lorenzini, Arithmetical Graphs, Math. Ann. 285 (1989) 481–501.
[14] D.J. Lorenzini, A finite group attached to the laplacian of a graph, Discrete Math. 91 (1991) 277–282.
[15] D.J. Lorenzini, Smith normal form and Laplacians, J. Combin. Theory B 98 (2008) 1271-1300.
[16] B.D. McKay, nauty Users Guide (Version 2.4), available at http://cs.anu.edu.au/∼bdm/nauty/.
[17] C. Merino, The chip-firing game, Discrete Math. 302 (2005) 188–210.
[18] R. Merris, Unimodular Equivalence of Graphs, Linear Algebra Appl. 173 (1992),181-189
[19] Y. Pan and J. Wang, A note on the third invariant factor of the Laplacian matrix of a graph, preprint,
math.CO/0912.3608v1
[20] D.G. Wagner, The critical group of a directed graph, preprint, arXiv:math/0010241v1 [math.CO]
Departamento de Matemáticas, Centro de Investigación y de Estudios Avanzados del IPN, Apartado
Postal 14–740, 07000 Mexico City, D.F.
E-mail address: [email protected],[email protected]
| 0math.AC
|
Connectivity of Natura 2000 forest sites
in Europe
Christine Estreguil 1,* , Giovanni Caudullo 1 , and Daniele de Rigo 1,2
1
European Commission, Joint Research Centre, Institute for Environment and Sustainability ,
Via E. Fermi 2749, I-21027 Ispra (VA), Italy
arXiv:1406.1501v2 [cs.CE] 20 Jun 2014
2
Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria,
Via Ponzio 34/5, I-20133 Milano, Italy
Background/Purpose: In the context of the European Biodiversity policy, the Green Infrastructure Strategy is one supporting tool to mitigate fragmentation, inter-alia to increase the spatial
and functional connectivity between protected and unprotected areas. The Joint Research Centre has developed an integrated model to provide a macro-scale set of indices to evaluate the
connectivity of the Natura 2000 network, which forms the backbone of a Green Infrastructure
for Europe. The model allows a wide assessment and comparison to be performed across countries in terms of structural (spatially connected or isolated sites) and functional connectivity
(least-cost distances between sites influenced by distribution, distance and land cover).
Main conclusion: The Natura 2000 network in Europe shows differences among countries in
terms of the sizes and numbers of sites, their distribution as well as distances between sites.
Connectivity has been assessed on the basis of a 500 m average inter-site distance, roads and
intensive land use as barrier effects as well as the presence of "green" corridors. In all countries
the Natura 2000 network is mostly made of sites which are not physically connected. Highest
functional connectivity values are found for Spain, Slovakia, Romania and Bulgaria. The more
natural landscape in Sweden and Finland does not result in high inter-site network connectivity
due to large inter-site distances. The distribution of subnets with respect to roads explains the
higher share of isolated subnets in Portugal than in Belgium.
Cite as:
Estreguil, C., Caudullo, G., de Rigo, D., 2014. Connectivity of Natura 2000 forest sites in Europe.
F1000Posters 2014, 5: 485. DOI: 10.6084/m9.figshare.1063300. ArXiv: 1406.1501
* Corresponding author: [email protected]
Europe’s Green Infrastructure Strategy is a holistic policy initiative integrating nature, biodiversity and sustainable
development [2]; it is one supporting tool to
implement the European Biodiversity Strategy and achieve 2020 targets [8]. One of its
aims is to mitigate fragmentation and render protection more effective. One condition to achieve this is to increase the spatial
T
and functional connectivity between natural and semi-natural protected and unprotected areas.
Protected areas such as Natura 2000
(N2K) [6] sites form the backbone of a Green
Infrastructure for Europe [4, 9]. Besides
their size and quality, their connectivity contributes to the movement and dispersal of
animals and plants.
HE
1
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
Figure 1: Illustration of the site rasterization method: five N2K sites (left) resulting three subnets (right,
from [12]).
from the perspective of influencing factors
related to site distribution, grey infrastructure and inter-site distance: connectivity
indices, isolated sites.
The unprotected landscape plays a role
in enhancing or reducing the conservation
and resilience of protected habitats; for example, grey infrastructure (artificial lands
and roads) and intensive land use often
pose the biggest threats or disturbances for
biodiversity conservation. There is a need
for conservation tools accounting for connectivity in tandem with landscape planning [16].
Connectivity measures should not necessarily be to link individual habitat patches
with physical structures (such as corridors
of similar habitat), but to ensure the existence of required functional connections
between protected sites (e.g. inter-site distances or/and landscape permeability) and
identify potentially isolated sites.
The Joint Research Centre has developed
an integrated spatially explicit model to provide a macro-scale European vision of the
connectivity of N2K sites on [12, 13]:
Pattern of Natura 2000 site network
The network of Natura 2000 sites designated under the "Habitat" [6] and "Birds"
Directives [7] covers circa 18% of the European Union (728,744 km2 ); 80% of the
sites include forest. All N2K polygons [5]
representing the extracted site areas have
been converted to a raster layer in order to
generate sub-network (subnet) components
which were formed by one or more Natura
2000 sites in cases of overlap (sites physically
connected) (Figure 1).
There are differences among countries in
terms of the sizes and numbers of sites, their
distribution as well as distances between
sites (Figure 2). In the small frames of Figure 2, Ireland (frame 1) has a well-connected
network; Portugal (frame 2) has large and
isolated subnets; Spain (frame 3) has large
and connected sites; Germany (frame 4) has
small and densely distributed sites; Sweden
(frame 5) has mainly small and distant subnets; Bulgaria (frame 6) has large subnets.
• Site network pattern in terms of size and
number.
• Intra-site network structural connectivity with morphological criteria: complex,
simple subnets.
• Inter-site network functional connectivity
2
Figure 2: Natura 2000 forest site network with six zoom frames showing network differences.
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
3
Extract
Sitestwithtfocalthabitat
4
Rasterize
Merge
Landtuse)covertshares1
Artificial
Agriculture
Natural
Landscapehmatrix
Sharehofhroadhnetwork
LANDtCOVERtSHARES
Subnettstatistics1
N2Khcoverage
NumberhofhN2Khsites
Numberhofhsubnets
SubnethsizehCmax,hmedianL
N2KhSubnethmap
Reclassify
Structuralhsubnethmap
Sharetof1
Simplehsubnets l,tnodeb
Complexhsubnets l>,tnodeb
(tSimpletlnodesb
(tComplextllinksmtnodesb
Reclassify
frictiontvaluestllogtscaleb
forteachtlandtuse
Frictionhmap
Inter-subnet
least-costhpath
TXTtfiles
Connectivitytindices1
RPC sitetareatweighetdtconnectivity
RAPC unweightedtconnectivity
Sharetin isolatedhsubnets
Functionalhconnectivity
CONEFOR
INTER(SITEtNETWORKtFUNCTIONALtCONNECTIVITY
MSPAhmap
coremtedgemtbridge
loopmtbranchmtislet
GUIDOStMSPA
INTRA(SITEtNETWORKtSTRUCTURALtCONNECTIVITY
Figure 3: The data input and information flow to compute the analysis on Natura 2000 network connectivity. The
intermediate steps generate ::categorical:: raster ::matrix:: layers from the input data. The categorical variables are
then transformed into ::nonnegative:: friction maps and functional distance (cost) estimates, from which least-cost
paths are computed. Most derived indices are dimensionless ::proportion:: values in [0, 1] with the exception of the
::nonnegative:: indices summarising the number of N2K sites and of subnets, and their size. The notation in the
workflow follows the Semantic Array Programming paradigm as applied in [11, 13] where the array-based semantics
is based on [3]. Details on the less obvious data-transformation modules may be found in [14].
11nonnegative11
11proportion11
Simplifiedhlandhusehmap
Xtclasses1tartificialm
pintensiveptagriculturem
natural)semi(natural
Mainhroadhnetwork
highwaysmttrunksmtprimary
roadstwithouttbridges)tunnels
11binary11
Reclassify
Extract
Rasterize
11categorical11
CLC
OSM
Pythontprocessing
withtESRItspatialtlibraries
Pythontprocessing
Indextfamily
Derivedtindices
Specializedtsoftware
Derivedtmap
Intermediatetstep
Inputtdata
NvK
SITE
NETWORKtPATTERN
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
Figure 4: Structural connectivity morphological analysis of subnets based on GUIDOS software (left), reclassification into simple/complex subnets (right, from [12]).
of landscape resistance to species dispersal
and to identify "green" corridors, a new European land use based friction map was created from the Corine Land Cover map of year
2006 at 100 m spatial resolution [10] and the
OpenStreetMap layer [1, 15].
Spatially-explicit model of
connectivity of protected areas
The model is based on two available software packages (free-download GUIDOS [18]
and free software Conefor [17]) integrated
with GIS Python programming tools for automated processing [19] (Figure 3). The
approach is harmonized and applies both
structural and functional criteria. It represents a compromise between biological
species models and commonly used connectedness measures.
Figure 5: Illustration of functional connectivity and
parameters: a i and a j refer to subnet areas, p i , j is
the probability of connectivity, A L the area of the landscape unit.
Intra-site network structural
connectivity
The spatial configuration of the sites is characterized in terms of simple subnets, made
of one node, and of complex subnets, made
of several interconnected nodes and links
(Figure 4). For each country two structural
connectivity indices are proposed:
• Share of Complex subnets.
• Share of Simple subnets.
The probability of connectivity (pi j ) is measured as a function of the landscape resistance (costi j ) between subnets (ai , a j ). It applies 50% probability of connectivity at 500
m average inter-subnets distance (dist 50% ).
It also accounts for the presence of "green"
corridors in-between subnets (Figure 5).
Inter-site network functional
connectivity
The model is based on a probabilistic power
weighted dispersal function [13]. As a proxy
5
ln(0.5)
d i st 50%
k
=
pi , j
= e k · costi , j
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
Figure 6: Functional connectivity illustrated for two subnets; the tunnel along the main road (left) is accounted
for identifying the least-cost (from [12]).
Intra-site
structural
connectivity
Site network pattern
Country
name
2
Natura
Number Number Subnet size (km )
2000
sites
subnets Median
Max
coverage
Simple Complex
subnet subnet
Inter-site functional
connectivity
Land cover shares
Road
RPC
RAPC
Isolated
AUSTRIA
15.8%
203
297
1.4
2813
80.4%
19.6%
6.7%
22.3%
71.0%
2.2%
5.8%
17.5%
7.4%
BELGIUM
13.6%
397
1036
0.9
385
76.8%
23.2%
24.5%
43.3%
32.2%
3.7%
5.3%
23.3%
4.2%
BULGARIA
35.1%
273
131
12.6
10656
60.4%
39.6%
7.3%
50.9%
41.8%
0.7%
16.9%
21.3%
7.6%
CYPRUS
12.4%
35
44
1.1
635
75.0%
25.0%
11.1%
47.3%
41.6%
1.9%
9.6%
42.6%
6.8%
CZECH REP.
15.0%
782
665
0.8
2196
84.5%
15.5%
8.1%
44.6%
47.3%
1.4%
4.5%
10.6%
3.2%
GERMANY
16.0%
4287
6931
0.8
4398
82.9%
17.1%
11.3%
47.7%
41.0%
2.4%
3.8%
12.1%
1.7%
DENMARK
7.2%
185
178
5.1
223
76.9%
23.1%
9.2%
67.7%
23.1%
1.3%
1.2%
9.9%
36.5%
ESTONIA
18.9%
506
699
1.3
729
90.9%
9.1%
3.1%
23.0%
73.9%
0.6%
4.9%
17.2%
5.0%
SPAIN
26.3%
1100
849
5
25191
77.8%
22.2%
3.9%
47.3%
48.7%
1.8%
10.9%
13.9%
7.5%
FINLAND
12.9%
1524
2952
0.7
7557
93.1%
6.9%
2.4%
6.2%
91.4%
0.7%
5.0%
4.5%
FRANCE
13.2%
1401
3965
0.5
5412
87.2%
12.8%
6.9%
43.9%
49.2%
1.8%
3.5%
9.3%
2.8%
GREECE
25.0%
258
227
30.9
2930
90.3%
9.7%
3.7%
33.5%
62.7%
1.3%
5.3%
9.1%
41.9%
HUNGARY
20.1%
377
611
2.5
2632
80.4%
19.6%
8.3%
67.2%
24.5%
1.5%
6.0%
12.9%
2.8%
IRELAND
13.0%
217
284
0.8
2624
80.6%
19.4%
3.4%
10.7%
85.9%
1.4%
8.7%
38.5%
2.8%
ITALY
17.2%
1873
1230
5.1
2238
87.1%
12.9%
5.5%
36.1%
58.4%
2.0%
3.0%
8.3%
12.6%
LITHUANIA
13.0%
417
312
2.1
948
83.4%
16.6%
5.0%
51.8%
43.3%
1.7%
3.4%
11.5%
3.5%
LUXEMBOURG
22.1%
57
79
0.5
88
76.0%
24.0%
14.2%
33.7%
52.1%
6.6%
11.5%
43.7%
0.0%
LATVIA
12.0%
302
296
2.7
936
94.9%
5.1%
2.8%
25.5%
71.7%
1.7%
2.9%
11.6%
5.4%
MALTA
12.2%
9
9
0.9
30
88.9%
11.1%
35.8%
4.0%
60.1%
6.7%
9.7%
57.9%
11.1%
13.0%
NETHERLANDS
Urban
Agricult. Natural
9.5%
8.5%
144
184
3.1
1012
80.4%
19.6%
17.1%
36.8%
46.0%
3.6%
3.5%
14.3%
POLAND
20.5%
846
854
2.6
9729
84.7%
15.3%
5.5%
56.9%
37.6%
1.0%
7.1%
11.8%
4.4%
PORTUGAL
17.1%
61
69
4.9
1842
85.5%
14.5%
6.2%
33.1%
60.7%
2.9%
4.9%
29.9%
18.8%
ROMANIA
22.6%
404
371
6.5
18649
77.9%
22.1%
8.4%
49.6%
42.0%
1.2%
11.3%
14.7%
7.8%
SWEDEN
10.9%
2919
3088
0.6
15908
93.8%
6.2%
2.3%
8.3%
89.5%
0.6%
5.7%
4.0%
6.3%
SLOVAKIA
30.1%
361
207
1.2
6837
90.4%
9.6%
8.4%
44.9%
46.7%
1.5%
15.5%
20.1%
6.3%
5.7%
464
1090
0.9
2704
82.7%
17.3%
10.2%
28.0%
61.8%
2.7%
1.4%
9.4%
6.4%
UN. KINDOM
Table 1: Country based table of indices highlighting their highest value (orange) and their lowest (purple).
6
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
For each country three functional connectivity indices are applied to the subnets1 :
• Site area weighted Root Probability of Connectivity (RPC) which is sensitive to the
size of subnets.
v
u n n
uX X
u
ai · a j · p i j
u
u i =1 j =1
RPC = t
A 2L
Country based results on the
Natura 2000 network connectivity
In all countries, the N2K network is mostly
made of simple subnets. Share of complex
subnets ("physically connected" sites) range
from 40% in Bulgaria to 5% in Latvia.
Large numbers of subnets and similar
small median sizes are found in Germany
and France, their inter-site connectivity with
emphasis on the sites’ sizes is rather low
(RPC approx. 3%) but higher when the intersite distances and landscape are at focus
(RAPC approx. 10%).
Highest RPC are found for Spain, Slovakia, Romania and Bulgaria. The more natural landscape in Sweden and Finland does
not result in high inter-site network connectivity due to large inter-site distances. The
distribution of subnets with respect to roads
explains the higher share of isolated subnets
in Portugal than in Belgium.
• Root un-weighted Average Probability of
Connectivity (RAPC) which is sensitive to
the unprotected landscape resistance and
functional distances between subnets.
v
u n n
uX X
u
pi j
u
t i =1 j =1
RAPC =
n2
• Share of functionally Isolated subnets in
the Natura 2000 network.
Figure 7: European map of the root un-weighted average probability of connectivity (RAPC) and the share of
N2K subnets functionally isolated.
1 RPC and RAPC are special cases of the general family of indices Power Weighted Probability of Dispersal
(PWPD) [13]. They also belong to its simplified formulation s-PWPD [11] as instances in which the s-PWPD
parameters {α, β, γ1 , γ2 } are respectively {0, 1/2, 1, 1} for RPC and {0, 1/2, 0, 0} for RAPC.
7
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
Figure 8: National profile of site area weighted root probability of connectivity (RPC). The chart includes
the N2K cover percentage per country and a proxy of the gap in connectivity computed as their difference
(from [12]).
For further datails please refer to:
Estreguil, C., Caudullo, G., San-Miguel-Ayanz, J., 2013. Connectivity of Natura 2000 Forest
Sites. EUR 26087EN. Luxemburg: Publications Office of the European Union. JRC 83104.
DOI:10.2788/95065
Estreguil, C., de Rigo, D., Caudullo, G., 2014. A proposal for an integrated modelling framework to characterise habitat pattern. Environmental Modelling & Software 52, 176-191.
DOI:10.1016/j.envsoft.2013.10.011
References
[1] Bennett, J., 2010. OpenStreetMap. Packt Publishing. ISBN: 978-1-84719-750-4 (page 5).
[2] Bennett, G., Bento Pais, R., Berry, P., Didicescu, P. S., Fichter, M., Hlaváč, V., Hoellen, K., Jones-Walters, L.,
Miko, L., Onida, M., Plesník, J., Smith, D., Wakenhut, F., 2010. Green Infrastructure Implementation:
Proceedings of the European Commission Conference 19 November 2010. (Ed: Karhu, J.). European
Commission, 28 pp.
http://ec.europa.eu/environment/nature/ecosystems/green_infrastructure.htm (page 1).
[3] de Rigo, D., 2012. Applying semantic constraints to array programming: the module ”check_is” of the
Mastrave modelling library. In: Semantic Array Programming with Mastrave - Introduction to Semantic
Computational Modelling. http://mastrave.org/doc/mtv_m/check_is (page 4).
[4] Directorate-General for Environment (DG ENV), 2012. The Multifunctionality of Green Infrastructure.
Science for Environment Policy.
http://ec.europa.eu/environment/nature/ecosystems/docs/Green_Infrastructure.pdf
(page 1).
8
Estreguil C., Caudullo G., de Rigo D., 2014. Connectivity of Natura 2000 forest sites in Europe.
[5] Directorate-General for Environment (DG ENV), 2012. Natura 2000 data - the European network of
protected sites. Temporal coverage: 2011. European Environment Agency web portal.
http://www.eea.europa.eu/data-and-maps/data/ds_resolveuid/60860bd4-28d6-44aa-93c7d9354a8205e3 (page 2).
[6] European Commission, 1992. Council directive 92/43/EEC of 21 may 1992 on the conservation of natural
habitats and of wild fauna and flora. Official Journal of the European Union 35 (L 206), 7-50. (pages 1 and 2).
[7] European Commission, 2010. Directive 2009/147/EC of the European Parliament and of the Council of 30
November 2009 on the conservation of wild birds. Official Journal of the European Union 53 (L 20), 7-25.
(page 2).
[8] European Commission, 2011. Our life insurance, our natural capital: an EU biodiversity strategy to 2020.
Brussel, COM (2011) 244 final. http://ec.europa.eu/environment/nature/biodiversity/
comm2006/pdf/2020/1_EN_ACT_part1_v7%5B1%5D.pdf (page 1).
[9] European Commission, 2013. Green Infrastructure (GI) - Enhancing Europe’s Natural Capital. Brussel,
COM (2013) 249 final.
http://eur-lex.europa.eu/LexUriServ/LexUriServ.do?uri=COM:2013:0249:FIN:EN:PDF
(page 1).
[10] European Environment Agency, 2012a. Corine Land Cover 2006 raster data - version 16. European
Environment Agency web portal. http://www.eea.europa.eu/data-andmaps/data/ds_resolveuid/ef13cef8-2ef5-49ae-9545-9042457ce4c6 (page 5).
[11] Estreguil, C., Caudullo, G., de Rigo, D., Whitmore, C., San-Miguel-Ayanz, J., 2012. Reporting on European
forest fragmentation: standardized indices and web map services. IEEE Earthzine 5 (2), 384031+.
http://www.earthzine.org/?p=384031 (2nd quarter theme: Forest Resource Information). (pages 4
and 7).
[12] Estreguil, C., Caudullo, G., San-Miguel-Ayanz, J., 2013. Connectivity of Natura 2000 Forest Sites. EUR
26087EN. Luxemburg: Publications Office of the European Union. JRC 83104. DOI:10.2788/95065 (pages 2,
5, 6, and 8).
[13] Estreguil, C., de Rigo, D., Caudullo, G., 2014. A proposal for an integrated modelling framework to
characterise habitat pattern. Environmental Modelling & Software 52, 176-191.
DOI:10.1016/j.envsoft.2013.10.011 (pages 2, 4, 5, and 7).
[14] Estreguil, C., de Rigo, D., Caudullo, G., 2014. Supplementary materials for: A proposal for an integrated
modelling framework to characterise habitat pattern.
http://mastrave.org/bib/Estreguil_etal_EMSsuppl_2014.pdf
(Extended version of the supplementary materials as published in Environmental Modelling & Software 52,
176-191, DOI:10.1016/j.envsoft.2013.10.011). (page 4).
[15] Haklay, M., Weber, P., 2008. OpenStreetMap: User-Generated Street Maps. Pervasive Computing 7(4). doi:
10.1109/MPRV.2008.80. (page 5).
[16] McHugh, N., Thompson, S., 2011. A rapid ecological network assessment tool and its use in locating
habitat extension areas in a changing landscape. Journal for Nature Conservation 19 (2011) 236-244.
DOI:10.1016/j.jnc.2011.02.002. (page 2).
[17] Saura, S., Torné, J., 2009. Conefor Sensinode 2.2: a software package for quantifying the importance of
habitat patches for landscape connectivity. Environmental Modelling & Software 24 (1), 135-139.
DOI:10.1016/j.envsoft.2008.05.005. (page 5).
[18] Soille, P., Vogt, P., 2009. Morphological segmentation of binary patterns. Pattern Recogn. Lett. 30 (4),
456-459. DOI:10.1016/j.patrec.2008.10.015. (page 5).
[19] Van Rossum, G., Drake Jr., F., 2011. The Python Language Reference Manual (version 3.2). Network Theory
Limited, ISBN 978-1-906966-14-0. (page 5).
9
| 5cs.CE
|
arXiv:1709.01427v1 [stat.ML] 5 Sep 2017
Stochastic Gradient Descent: Going As Fast As
Possible But Not Faster
Alice Schoenauer Sebag
Altschuler&Wu lab, Dpt of Pharm. Chem.
UCSF, San Francisco, CA 94158
[email protected]
Marc Schoenauer
INRIA-CNRS-UPSud-UPSay
TAU, U. Paris-Sud, 91405 Orsay
[email protected]
Michèle Sebag
CNRS-UPSud-INRIA-UPSay
TAU, U. Paris-Sud, 91405 Orsay
[email protected]
Abstract
When applied to training deep neural networks, stochastic gradient descent (SGD)
often incurs steady progression phases, interrupted by catastrophic episodes in
which loss and gradient norm explode. A possible mitigation of such events is to
slow down the learning process.
This paper presents a novel approach to control the SGD learning rate, that uses
two statistical tests. The first one, aimed at fast learning, compares the momentum
of the normalized gradient vectors to that of random unit vectors and accordingly
gracefully increases or decreases the learning rate. The second one is a change
point detection test, aimed at the detection of catastrophic learning episodes; upon
its triggering the learning rate is instantly halved.
Both abilities of speeding up and slowing down the learning rate allows the proposed approach, called SAL E RA, to learn as fast as possible but not faster. Experiments on standard benchmarks show that SAL E RA performs well in practice, and
compares favorably to the state of the art.
Machine Learning (ML) algorithms require efficient optimization techniques, whether to solve convex
problems (e.g., for SVMs), or non-convex ones (e.g., for Neural Networks). In the convex setting,
the main focus is on the order of the convergence rate [Nesterov, 1983, Defazio et al., 2014]. In
the non-convex case, ML is still more of an experimental science. Significant efforts are devoted
to devising optimization algorithms (and robust default values for the associated hyper-parameters)
tailored to the typical regime of ML models and problem instances (e.g. deep convolutional neural
networks for MNIST [Le Cun et al., 1998] or ImageNet [Deng et al., 2009]) [Duchi et al., 2010,
Zeiler, 2012, Schaul et al., 2013, Kingma and Ba, 2014, Tieleman and Hinton, 2012].
As the data size and the model dimensionality increase, mainstream convex optimization methods
are adversely affected. Hessian-based approaches, which optimally handle convex optimization
problems however ill-conditioned they are, do not scale up and approximations are required [Martens
et al., 2012]. Overall, Stochastic Gradient Descent (SGD) is increasingly adopted both in convex and
non-convex settings, with good performances and linear tractability [Bottou and Bousquet, 2008,
Hardt et al., 2015].
Within the SGD framework, one of the main issues is to know how to control the learning rate:
the objective is to reach a satisfactory learning speed without triggering any catastrophic event,
manifested by the sudden rocketing of the training loss and the gradient norm. Finding "how much is
not too much" in terms of learning rate is a slippery game. It depends both on the current state of the
system (the weight vector) and the current mini-batch. Often, the eventual convergence of SGD is
ensured
√ by decaying the learning rate as O(t) [Robbins and Monro, 1951, Defazio et al., 2014] or
O( t) [Zinkevich, 2003] with the number t of mini-batches. While learning rate decay effectively
prevents catastrophic events, it is a main cause for the days or weeks of computation behind the many
breakthroughs of deep learning. Many and diverse approaches have thus been designed to achieve the
learning rate adaptation [Amari, 1998, Duchi et al., 2010, Schaul et al., 2013, Kingma and Ba, 2014,
Tieleman and Hinton, 2012, Andrychowicz et al., 2016] (more in Section 1).
This paper proposes a novel approach to adaptive SGD, called SAL E RA (Safe Agnostic LEraning
Rate Adaptation). SAL E RA is based on the conjecture that, if learning catastrophes are well taken
care of, the learning process can speed up whenever successive gradient directions show general
agreement about the direction to go.
The frequent advent of catastrophic episodes, long observed by neural net practitioners [Goodfellow
et al., 2016, Chapter 8] raises the question of how to best mitigate their impact. The answer depends
on whether these events could be anticipated with some precision. Framing catastrophic episodes
as random events1 , we adopt a purely curative strategy (as opposed to a preventive one): detecting
and instantly curing catastrophic episodes. Formally, a sequential cumulative sum change detection
test, the Page-Hinkley (PH) test [Page, 1954, Hinkley, 1970] is adapted and used to monitor the
learning curve reporting the minibatch losses. If a change in the learning curve is detected, the system
undergoes an instant cure by halving the learning rate and backtracking to its former state. Such
instant cure can be thought of in terms of a dichotomic approximation of line search (see e.g. Defazio
et al. [2014], Eq. (3)).
Once the risk of catastrophic episodes is well addressed, the learning rate can be adapted in a more
agile manner: the AL E RA (Agnostic LEarning Rate Adaptation) process increases (resp. decreases)
the learning rate whenever the correlation among successive gradient directions is higher (resp. lower)
than random, by comparing the actual gradient momentum and the agnostic momentum built from
random unit vectors.
The contribution of the paper is twofold. First, it proposes an original and efficient way to control
learning dynamics (section 2.1). Secondly, it opens a new approach for handling catastrophic events
and salvaging a significant part of doomed-to-fail runs (section 2.2). The experimental validation
thereof compares favorably with the state of the art on the MNIST and CIFAR-10 benchmarks
(section 3).
1
Related work
SGD was revived in the last decade as an effective method for training deep neural networks with
linear computational complexity in the size of the dataset [Bottou and Bousquet, 2008, Hardt et al.,
2015]. SGD faces two limitations, depending on the learning rate: too large, and the learning
trajectory leads to catastrophic episodes; too small, and its convergence takes ages. The dynamic
adjustment of the learning rate was therefore acknowledged a key issue since the early SGD days
[Robbins and Monro, 1951].
Dealing with catastrophic events in deep learning The exploding gradient problem, described
in [Goodfellow et al., 2016, Chapter 8] as the encounter of steep cliff structures in the derivative
landscape during learning, is frequently met while training neural networks (and even more so when
training recurrent neural networks [Bengio et al., 1994]).
When it comes to dealing with such events, most published work focuses on creating the conditions
so that they do not arise. Among the possibilities are the use of regularizations, e.g., L1 or L2
regularization [Pascanu et al., 2013] or Max-norm regularization [Srivastava et al., 2014]. Gradient
clipping, constraining the gradient norm to remain smaller than a constant [Pascanu et al., 2013],
is another possibility. The introduction of batch normalization [Ioffe and Szegedy, 2015] also
helps diminishing the frequency of such events. Finally, proper initialization [Glorot and Bengio,
2010, Sutskever et al., 2013] or unsupervised pre-training [Erhan et al., 2010], i.e. initializing the
optimization trajectory in a good region of the parameter space, also diminish the frequency of such
events.
Learning rate adaptation Addressing the slow speed of SGD, learning rate adaptation has been
acknowledged a key issue since the late 80s (see George and Powell [2006] for a review). Using the
1
As far as tractable computational resources are involved in their prediction.
2
information contained in the correlation of successive gradient directions was already at the heart of
the delta-delta and delta-bar-delta update rules proposed by Jacobs [1988]. Briefly, the delta-bar-delta
rule states that for each parameter, if the current gradient and the relaxed sum of past gradients have
the same sign, the learning rate is incremented additively; and if they are of opposite sign, then the
learning rate is decremented multiplicatively. Decrementing the learning rates faster than increasing
them was already advocated by the author to adapt faster in case of catastrophic events.
The natural gradient descent (NGD) approach [Amari, 1998] considers the Riemaniann geometry
of the parameter space, using the Fisher information matrix (estimated by the gradient covariance
matrix), to precondition the gradient. Due to its quadratic complexity in the dimension of the
parameter space, NGD approximations have been designed for deep networks [Pascanu and Bengio,
2014]. Notably, approaches such as the Hessian-Free from Martens [2010] can be interpreted as
NGD [Pascanu and Bengio, 2014].
A DAGRAD [Duchi et al., 2010] also uses the information of past gradients to precondition each
update in a parameter-wise manner, dividing the learning rate by the sum of squared past gradients.
Several approaches have been proposed to refine A DAGRAD and mitigate its learning rate decay to 0,
including A DADELTA [Zeiler, 2012], RMSProp [Tieleman and Hinton, 2012], and A DAM [Kingma
and Ba, 2014]. A DAM is based on estimating the first and second moments of the gradient w.r.t.
each parameter, and using their ratio to update the parameters. Moment estimates are maintained
by exponential moving averages of different weight factors, such that by default the inertia of the
first moment is higher by two orders of magnitude than the second. As will be seen, SAL E RA also
builds upon the use of the gradient second moment, with the difference that it is compared with a
fixed agnostic counterpart.
In [Schaul et al., 2013], the learning rate is computed at each time-step to approximately maximally decrease the expected loss, where the loss function is locally approximated by a parabola.
Finally, Andrychowicz et al. [2016] address learning rate adaptation as a reinforcement learning
problem, exploiting the evidence gathered in the current time steps to infer what would have been the
good decisions earlier on, and accordingly optimizing a hyper-parameter adjustment policy.
More remotely related are the momentum approaches, both its classic [Polyak, 1964] and Nesterov
versions [Nesterov, 1983] as derived by Sutskever et al. [2013], that rely on a relaxed sum of past
gradients for indicating a more robust descent direction than the current gradient.
2
SAL E RA
SAL E RA involves two components: a learning rate adaptation scheme, which ensures that the
learning system goes as fast as it can; and a catastrophic event manager, which is in charge of
detecting undesirable behaviors and getting the system back on track.
2.1
Agnostic learning rate adaptation
Rationale The basic idea of the proposed learning rate update is to compare the current gradient
descent to a random walk with uniformly chosen gradient directions. Indeed, the sum of successive
normalized gradient vectors, referred to as cumulative path in the following, has a larger norm than
the sum of uniformly drawn unit vectors if and only if gradient directions are positively correlated.
In such cases, the learning process has a global direction and the process can afford to speed up.
On the opposite, if the norm of the cumulative path is smaller than its random equivalent, gradient
directions are anti-correlated: the process is alternating between opposite directions (e.g., bouncing
on the sides of a narrow valley, or hovering around some local optimum) and the learning rate should
be decreased.
The AL E RA scheme takes inspiration from the famed CMA-ES [Hansen and Ostermeier, 2001] and
NES [Wierstra et al., 2014] algorithms, today considered among the best-performing derivative-free
continuous optimization algorithms. These approaches de facto implement natural gradient optimization [Amari, 1998] and instantiate the Information-Geometric Optimization paradigm [Ollivier et al.,
2017] in the space of normal distributions on IRd . Formally, CMA-ES maintains a normal distribution.
The variance of the normal distribution, aka step-size, is updated on the basis of a comparison of the
cumulative path of the algorithm (moving exponential average of successive steps) with that of a
random walk with Gaussian moves of fixed step-sizes. This mechanism is said to be agnostic as it
makes no assumption whatsoever on the properties of the optimization objective.
3
The AL E RA algorithm The partial adaptation of the CMA-ES scheme to the minimization of a
loss function L on a d-dimensional parameter space is defined as follows. Let θt be the solution at
time t, gt = ∇θ L(θt ) the gradient of the current loss, and ηt the current learning rate. SGD computes
the solution at time t + 1 by θt+1 = θt − ηt gt . Let k.k denote the L2 norm and < ., . > the associated
dot product.
Definition. For t > 0, the exponential moving average of the normalized gradients with weight α,
pt , and its random equivalent rt are defined as:
p0 = 0; r0 = 0
pt = α gt /kgt k + (1 − α) pt−1
rt = α ut + (1 − α) rt−1
(1)
(2)
(3)
where (ut )t>0 are independent random unit vectors in IRd .
Proposition. For t > 0, the expectation µ(t) and variance σ 2 (d, t) of krt k2 as defined above are:
α
µ(t) = E(krt k2 ) =
[1 − (1 − α)2t ]
(4)
2−α
σ 2 (d, t) = V ar(krt k2 ) =
1
2α2 (1 − α)2
[1 − (1 − α)t ][1 − (1 − α)t−1 ]
d (2 − α)2 ((1 − α)2 + 1)
(5)
Proof: Appendix A.
Let µ and σ 2 (d) denote the limits of respectively µ(t) and σ 2 (d, t) as t → +∞. At time step t,
the AL E RA scheme updates the cumulative path pt by comparing its norm with the distribution
of the agnostic momentum defined from µ and σ 2 (d). The learning rate is increased or decreased
depending on the normalized gap between the squared norm of pt and rt :
kpt k2 − µ
ηt+1 = ηt exp C
(6)
σ(d)
with C a hyper-parameter of the algorithm.
The approach is implemented in the AL E RA algorithm (non-greyed lines in Algorithm 1). Given
hyper-parameters α and C, as well as η0 , the initial learning rate, and ρ the mini-batch size, each
iteration over a mini-batch computes the new exponential moving average of the normalized gradient
(line 13), performs the agnostic update of the learning rate (line 14) before updating parameter θ
the usual way (line 15). The learning rate can be controlled in a layer-wise fashion, independently
maintaining an exponentiated moving average and updating the learning rate for each layer of the
neural network (lines 13-14). This algorithm is used in all experiments of Section 3.
Parameter-wise learning rate adaptation As noted by Kingma and Ba [2014], the parameterwise control of the learning rate is desirable in some contexts. The above scheme is extended to
achieve the parameter-wise update of the learning rate as follows. For t > 0 let r2t,i denote the
squared ith coordinate of rt . It is straightforward to show that the expectation (respectively the
standard deviation) of r2t,i is the expectation of rt divided by d (respectively the standard deviation of
√
rt divided by d).
Given t > 0 and i ∈ [|1; d|], the squared ith coordinate of pt noted r2t,i , can thus likewise be adjusted
by comparison with its random counterpart r2t,i . The update therefore becomes:
(
)
r2 − µ
0 1 t,i
ηi,t+1 = ηi,t exp C √
(7)
d σ(d)
See Appendix B for a full derivation of the parameter-wise algorithm.
2.2
Catastrophic event management
As said, the ability to learn fast requires an emergency procedure, able both to detect an emergency
and to recover from it.
4
Algorithm 1: S-AL E RA: Agnostic LEarning Rate Adaptation and Page-Hinkley change detection
1
2
3
4
Input: Model with loss function L
Parameters : Memory rate α, factor C
//algorithm parameters
Initial learning rate η0 , mini-batch ratio ρ
// run parameters
Initialize : t ← 0; p ← 0; η ← η0 ; init(θ)
// initialization
¯ ← 0; ∆ ← L(θ, first mini-batch)/10
L, Lmin , `, `,
// initialize Page-Hinkley
while stopping criterion not met do
M B ← new mini-batch ; t ← t + 1
// perform forward pass
` = ρL(θ, MB) + (1 − ρ)`
¯
`¯ ← (` + t`)/(t
+ 1)
// empirical mean of batch losses
5
¯
L ← L + (` − `)
6
Lmin ← min(Lmin , L)
7
if L − Lmin > ∆ then
8
θ ← θ(b) ; η ← η/2
9
¯ ` ← 0; t ← 0
L, Lmin , `,
10
11
12
13
14
15
16
17
// cumulated deviations from mean
// lower bound of deviations
// Page-Hinkley triggered: backtrack
// and re-initialize Page-Hinkley2
else
θ(b) ← θ
g ← ∇θ L(θ, MB)
p ← αg/kgk + (1 − α)p
η ← η exp C(kpk22 − µ)/σ(d)
θ ← θ − ηg
end
end
// save for possible backtracks
// compute gradient with backward pass
// exponential moving average of normalized gradients
// agnostic learning rate update
// standard parameter update
Recovery In a healthy learning regime, the training error should decrease along time − up to the
noise due to the inter-batch variance − unless the learning system abruptly meets a cliff structure
[Goodfellow et al., 2016], usually blamed on too large a learning rate in an uneven gradient landscape.
In a convex noiseless optimization setting, if computationally tractable, the best strategy is to compute
(an approximation of) the optimal learning rate η ∗ through line search3 . In such context, as a thought
experiment, let η − > 0 be such that, if used to update θt , the resulting θt+1 would yield the same
performance as θt . For η > η − , θt+1 yields a worse performance than θt , and if continued the
optimization process is likely to diverge. For η < η ∗ , θt+1 yields a performance improvement. For
η ∗ < η < η − , θt+1 yields a performance improvement too; the further θ trajectory is likely to bounce
back and forth on the walls of the optimum valley.
Overall, the safety zone for the learning rate is ]0, η − [ (with the caveat that the safety zone is narrower
for ill-conditioned optimization problems). The proposed safeguard strategy primarily aims to detect
when η steps outside of its safety zone (see the change point detection test below), and to apply a
correction as to get back in it. Upon change detection in the mini-batch loss, SAL E RA implements
a straightforward correction: halving the learning rate and recovering the last solution before test
−
triggering. The halving process is iterated if needed, sending ηt back in ] η2 , η − ] exponentially fast
(except for the perturbations in the gradient due to the mini-batch variance).
The rationale for the halving trick is based on a trade-off between the number of successive dividing
iterations, that could indeed be made even smaller by using a larger dividing factor, and the required
standard AL E RA iterations that will be needed to reach the optimal learning rate η ∗ after having
reached the safety zone again. The choice of the 2 dividing factor is further discussed in Appendix C.
2
There are in fact two time counters in the full SAL E RA Algorithm, the global usual one and the one related
to the Page-Hinkley test. Only one is used here for simplicity reasons.
3
See Defazio et al. [2014] on how to handle the noise in the case of a composite loss function.
5
Detection SAL E RA applies a change detection test on the signal given as the minibatch loss `t .
The PH detection test [Page, 1954, Hinkley, 1970] is chosen as it provides optimal guarantees about
the trade-off between the detection delay upon a change (affecting the average or standard deviation
of the signal) and the mean time between false alarms. For t > 0, it maintains the empirical mean `¯t
of the signal, and the cumulative deviation4 Lt from the empirical mean (Lt = Lt−1 + (`t − `¯t )).
Finally, it records the empirical bounds of Lt (Lmin = mint Lt ; Lmax = maxt Lt ). In case of a
stationary signal, the expectation of Lt is 0 by construction; the PH change test is thus triggered when
the gap between Lt and its empirical bounds is higher than a problem-dependent threshold ∆, which
controls the alarm rate.
The PH test is implemented in the SAL E RA algorithm as follows (greyish lines in Algorithm 1).
∆ is set to L(θ, first mini-batch)/λ, with λ = 10, that is, one tenth of the empirical loss on the
first minibatch in all experiments (this issue is further discussed in Section 3.2). Variables `¯t , L
and Lmax are maintained (lines 5- 6). In the learning context, a decrease of the loss signal is
welcomed and expected. Only the case of an increasing signal is thus monitored. Upon test triggering
(Lmax − L > ∆), the learning rate is halved and the weight vector θ is reset to the last solution
before then (line 8), and the PH test is reinitialized (line 9).
3
Experiments
The goal of the following experiments is to validate the algorithmic ideas introduced in section 2 by
comparing their application with that of widely used optimization techniques (see Section 1) on some
straightforward NN architectures.
3.1
Experimental Setting
Datasets All experiments are performed on the MNIST [Le Cun et al., 1998] and CIFAR10 [Krizhevsky, 2009] datasets, which respectively contain 60k and 50k training examples. Both
contain 10k test examples, which are to be classified in 10 classes. The data is normalized according
to the mean and standard deviation along each coordinate on the training set.
Algorithms Adagrad, NAG, and Adam are used as baselines. The agnostic adaptation rule and the
change detection can be applied independently. In order to separate their effect, 3 original algorithms
are studied here: ALeRA (the white lines in Algorithm 1) implements the agnostic learning rate
adaptation without the change detection; Ag-Adam uses the same agnostic adaptation for the learning
rate on top of the A DAM algorithm. Finally, the change detection mechanism is implemented with the
agnostic adaptation, yielding SALeRA as described in Algorithm 1, as well as its parameter-wise
version SPALeRA (Algorithm 2 in Appendix B).
Hyperparameters The exploration of the hyper-parameter space for all algorithms has been done
on a grid of possible values (with the exception of Adagrad which has no hyper-parameter):
- NAG: the momentum γ ∈ {.8, .9, .99, .999}.
- Adam: β1 ∈ {.8, .9, .99} and β2 ∈ {.99, .999, .9999};
- ALeRA: α ∈ {.001, .01, .1, .25} and C ∈ {3.10−8 , 3.10−7 , 3.10−6 , 3.10−5 } (see Algorithm 1). The parameters for SALeRA and SPALeRA are the same, as there is no additional
parameter for the Page-Hinkley part.
- Ag-Adam: the recommended values for A DAM (β1 = .9 and β2 = .999) are used for the
A DAM part, the same values than for AL E RA are used for the AL E RA part.
The initial learning rate η0 ranges from 10−5 to 1 depending on the algorithm. Finally, the mini-batch
size was set to either 1% or 1h of the training set size. All reported results are based on 5 independent
runs performed for each hyperparameter set unless otherwise specified.
Network Models All experiments consider the following 4 network architectures:
- M0: a softmax regression model with cross-entropy loss (i.e. no hidden layers),
4
The PH test takes into account the extreme value phenomenon by considering an upper cumulative deviation
and a lower cumulative deviation, defined from Lt by adding (resp. subtracting) a margin δ to Lt . In SAL E RA
δ is set to 0.
6
- M2: 2 fully connected hidden layers with ReLU activation, on top of M0. The hidden layers
are of respective sizes (500, 300) for MNIST, (1 500, 900) for CIFAR-10.
- M2b is identical to Model M2 above, except that Batch Normalization layers [Ioffe and
Szegedy, 2015] are added in each hidden layer.
- M4: LeNet5-inspired convolutional models [Le Cun et al., 1998]. These models contain 2
convolutional layers with max-pooling followed by 2 fully connected layers, all with ReLU
activation. They are of respective sizes (32, 64, 128, 128) for MNIST, and (32, 64, 384, 384)
for CIFAR-10. Batch normalization is used in each layer.
These architectures are not specifically optimized for the task at hand, but rather chosen to compare
the performances of past and novel algorithms in a wide variety of situations.
Experimental conditions All computations are performed on 46 GPUs (5 TITAN X(Pascal), 9
GTX 1080 and 32 Tesla K80) using the Torch library [Collobert et al., 2011] in double precision. A
typical run on a TITAN X (Pascal) GPU for 20k mini-batches of size 50 for CIFAR-10 on M4 takes
between 8 and 10 minutes for all algorithms.
Metrics MNIST and CIFAR-10 are classification problems. We therefore report classification
accuracy on the test set at 5 epochs (i.e. 5 full passes on the training set) and 20 epochs (end of all
runs), as well as their standard deviations on the 5 independent runs.
3.2
Experimental results
MNIST
M0
M2
M2b
M4b
CIFAR
M0
M0
M2
M2
M2b
M2b
M4b
M4b
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
NAG
7.75 (.14)
7.59 (.07)
1.95 (.17)
1.58 (.08)
1.82 (.13)
1.47 (.10)
.85 (.08)
.72 (.09)
60.37 (.55)
59.73 (.19)
45.82 (.93)
45.08 (.32)
45.01 (.84)
42.50 (.48)
27.74 (.48)
27.45 (.39)
Adagrad
7.73 (.06)
7.51 (.09)
2.00 (.06)
1.71 (.08)
1.72 (.07)
1.48 (.06)
1.02 (.09)
.82 (.08)
60.49 (.71)
59.76 (.36)
44.81 (.62)
43.59 (.51)
44.18 (.62)
43.79 (.25)
34.93 (.96)
29.15 (.67)
A DAM
7.72 (.12)
7.43 (.10)
2.07 (.11)
1.56 (.06)
1.81 (.07)
1.57 (.94)
.89 (.31)
.80 (.08)
60.60 (.45)
59.81 (.24)
45.68 (.39)
44.43 (.50)
44.30 (.96)
43.60 (.64)
28.50 (.68)
27.84 (.59)
AG -A DAM
AL E RA
7.31 (.09)
7.51 (.09)
7.29 (.09)
7.43 (.03)
1.93 (.17)
1.86 (.11)
1.57 (.04)
1.55 (.10)
1.66 (.08)
1.59 (.08)
1.53 (.05)
1.43 (.04)
.91 (.06)
.82 (.30)
.79 (.05)
.63 (.05)
59.62 (.27) 59.89 (.19)
59.34 (.24) 59.31 (.11)
44.91 (.42) 44.69 (17.58)
43.25 (.40) 43.19 (.21)
43.33 (.33) 43.08 (.17)
42.72 (.33) 42.12 (.15)
25.60 (.29) 28.61 (.50)
25.30 (.18) 26.35 (.64)
SAL E RA
7.51 (.09)
7.44 (.04)
1.87 (.05)
1.59 (.09)
1.59 (.08)
1.48 (.09)
.82 (.30)
.64 (.07)
59.69 (.33)
59.31 (.25)
44.42 (.24)
42.72 (.41)
43.08 (.17)
42.50 (.29)
28.61 (.50)
25.93 (.64)
SPAL E RA
8.03 (.09)
7.60 (.08)
1.93 (.14)
1.55 (.09)
1.78 (.08)
1.50 (.09)
.82 (.14)
.63 (.11)
61.32 (.65)
59.71 (.48)
45.74 (.85)
44.48 (.65)
44.92 (.89)
43.23 (.27)
29.61 (.99)
27.94 (.22)
Table 1: Best performances (test error over all tested parameter settings) at 5 and 20 epochs of NAG,
A DAGRAD, A DAM, AG -A DAM, AL E RA and SAL E RA (average and standard deviation on 5 runs).
For each line, the best results are in bold, and results less than 1 std.dev. away are in italics.
Learning performances The experimental evidence (Table 1) shows that AG -A DAM quite often
slightly but statistically significantly improves on A DAM. A possible explanation is that AG -A DAM
has a more flexible adjustment of the learning rate than A DAM (possibly increasing η by a few
orders of magnitude). In many cases, AL E RA and SAL E RA yield similar results; indeed, whenever
AL E RA does not meet catastrophic episodes, AL E RA and SAL E RA have the same behaviors. A
representative run, where AL E RA and SAL E RA undergo catastrophic episodes is depicted on Fig.
1 (both runs with same random seed). AL E RA faces a series of catastrophic episodes, where the
training error reaches up to 80%. It eventually stabilizes itself with a medium training loss, but with
test error above 80%. In the meanwhile, SAL E RA reacts upon the first catastrophic episode around
epoch 8, by halving the learning rate on each layer. It faces a further catastrophic episode around
epoch 13, and halves the learning rates again. Overall, it faces less frequent and less severe (in terms
of train loss and test error deteriorations) accidents. Eventually, SAL E RA recovers acceptable train
and test errors.
7
MNIST M0
M2
M2b
M4b
CIFAR
M0
M2
M2b
M4b
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
5ep.
20ep.
NAG
7.91 (.14)
7.59 (.07)
1.95 (.17)
1.64 (.06)
1.85 (.05)
1.47 (.10)
.85 (.08)
.76 (.12)
60.73 (.44)
59.73 (.19)
46.62 (.12)
45.08 (.32)
45.48 (.67)
43.86 (.52)
27.74 (.48)
27.70 (.76)
A DAM
AG -A DAM
AL E RA
7.87 (.25)
7.86 (.20)
7.96 (.13)
7.45 (.08)
7.34 (.08)
7.53 (.10)
2.30 (.17)
2.04 (.15)
1.89 (.07)
1.59 (.06)
1.80 (.45)
1.68 (.04)
1.93 (.09)
1.99 (.11)
1.71 (.11)
1.62 (.09)
1.82 (.26)
1.59 (.05)
1.03 (.08)
1.08 (.13)
.99 (.11)
.80 (.08)
.91 (.08)
.83 (.01)
60.74 (.39) 60.55 (.49)
60.46 (.81)
60.08 (.22) 59.34 (.24)
59.31 (.11)
45.71 (.54) 45.15 (.15) 44.84 (17.52)
44.44 (.61) 44.30 (.42) 43.92 (1.43)
44.90 (.50) 43.33 (.33) 44.15 (.29)
43.60 (.64) 43.11 (.35)
42.70 (.31)
29.34 (.83) 28.15 (.21)
28.61 (.50)
28.00 (.72) 27.46 (.54)
27.60 (.26)
SAL E RA
7.96 (.13)
7.53 (.06)
1.87 (.05)
1.68 (.05)
1.72 (.04)
1.59 (.05)
.99 (.14)
.83 (.01)
60.43 (.42)
59.37 (.10)
45.01 (.14)
43.68 (.25)
44.49 (.18)
42.70 (.31)
28.61 (.50)
27.60 (.26)
SPAL E RA
8.06 (.10)
7.60 (.08)
1.94 (.13)
1.63 (.05)
1.80 (.08)
1.52 (.02)
.94 (.09)
.63 (.11)
61.71 (.76)
59.79 (.57)
46.08 (.90)
44.62 (.56)
45.45 (.57)
43.51 (.80)
30.09 (.73)
28.05 (.43)
Table 2: Best performance (test error) for the robust algorithm-parameter settings (over all runparameters settings for η0 and batch size): momentum .9 for NAG, β1 = .8 and β2 = .9999 for
A DAM, α = .001 and C = 3.10−6 for AG -A DAM α = .01 and C = 3.10−6 for AL E RA and
SAL E RA, and α = .1 and C = 3.10−8 for SPAL E RA.
It is interesting to note that the learning rates on Fig. 1 are constantly increasing, in contradiction
with common knowledge. In practice, the learning rate behavior depends on the dataset, the neural
architecture and the seed, and can be very diverse (constant decrease, constant increase, or most of
the time, an increase followed by a decrease). The diverse learning rate behavior is viewed as an
original feature of the proposed approach, made possible by the ability to detect, and recover from,
catastrophic explosions of the training loss.
The actual behavior of all algorithms is depicted for CIFAR-10, model M4 on Fig. 2, with AG -A DAM,
AL E RA and SAL E RA respectively getting first, second and third rank in terms of test error at 20
epochs. In terms of optimization per se, SAL E RA (respectively AL E RA) reaches a training error
close to 0 at epoch 10 (resp. epoch 20) whereas AG -A DAM reaches a plateau after epoch 15. In the
meanwhile, the test error decreases and the test loss increases for all three algorithms. A tentative
interpretation for this fact is that the neural net yields more crisp output, close to 0 or 1; this does not
change the error while increasing the loss. This result suggests several perspectives for further work
(Section 4).
In order to determine to which extent the best results in Table 1 depend on the hyper-parameter
settings, and define best configurations for the considered benchmarks, a sensitivity analysis was
performed, comparing the results on all models and all epochs for each setting, and choosing the one
with lowest sum of ranks. The results for these robust settings are displayed table 2 (as A DAGRAD
has no hyper-parameter, it is not mentioned there). The proposed hyper-parameters for AL E RA and
SAL E RA are found to be α = .01 and C = 3.10−6 . Interestingly enough, we find that the optimal
hyper-parameters for A DAM are β1 = .8 and β2 = .9999 instead of β1 = .9 and β2 = .999 as is
suggested in the original paper [Kingma and Ba, 2014]. The proposed approaches still show some
advantage over A DAM and NAG, though they seem more sensitive to their parameter tuning. It is left
for further work to derive precise recommendations depending on model characteristics.
Catastrophe management performances Let us define a failed run as a run which attains more
than 80% test error after 20 epochs. AL E RA is observed to have 18.3% failed runs over the parameter
range as defined in 3.1. Our catastrophe management scheme makes it possible for SAL E RA to
avoid approximately 40% of these failures, reaching a rate of failure of 11.7% on the same parameter
range.
It would of course be possible to further diminish the failure rate of SAL E RAby setting the PH
alarm threshold ∆ to L(θ, first mini-batch)/λ with λ = 100 (or 1, 000), rather than 10 as used
in all experiments reported above. However, this would potentially interfere with learning rate
adaptation by triggering learning rate halvings even when there is no serious alert to be made, thus
preventing it to be as bold as possible. Indeed, setting λ = 100 causes a decline of more than half
8
Figure 1: Comparison, on a representative run on MNIST M2 (with same random seed), of AL E RA
and SAL E RA: test error (top row), minibatch error (middle row) and learning rates (bottom row). A
first catastrophic episode is met around epoch 8: SAL E RA reacts by dividing the learning rate for all
three layers. Note that the further catastrophic events met by SAL E RA are more rare and less severe
than for AL E RA: SAL E RA eventually yields a better training loss and a considerably better test
error than AL E RA. Better seen in color.
Figure 2: Comparative behaviors on CIFAR-10, model M4 (best configuration for each algorithm,
see text). Left: test error. Middle: test loss. Right: mini-batch error. Better seen in color.
9
of SAL E RA performances (data not shown), even though it manages to approximately halve the
number of failed runs. Furthermore, setting λ = 1 000 does not further diminish the failure rate,
therefore only harming SAL E RA learning performances. On the other hand, setting λ = 1 does not
significantly improve the best performances, but has a higher failure rate.
One tenth of the initial loss is therefore a very good balance between the aggressive learning rate
adaptation scheme and its braking counterpart, at least on these datasets and architectures.
4
Discussion
The first proposed contribution relies on the comparison of the gradient momentum with a fixed
reference. It is meant to estimate the overall correlation among the sequence of gradients, which
can be thought of as a signal-to-noise ratio in the process generated from the current solution, the
objective and the successive mini-batches. Depending on this ratio, the process can be accelerated or
slowed down. The AL E RA procedure, which implements this idea, proves to be significantly able to
increase and decrease the learning rate. Furthermore, this process can be plugged on A DAM, with a
performance improvement on average.
The price to pay for this flexibility is that it increases the risk of catastrophic episodes, with instant
rocketing of the training loss and gradient norm. The proposed approach relies on the conjecture
that catastrophic episodes can be rigorously observed and detected. A second conjecture is that a
neural net optimizer is almost doomed to face such episodes along the optimization process. These
events are mostly detrimental to optimization: before the run, one often chooses small learning rates
(and thus slow convergence) to prevent them, and during the run, they are mostly not recovered from.
Based on these two conjectures, the second contribution of the paper is an agnostic and principled
way to detect and address such episodes. The detection relies on the Page-Hinkley change point
detection test. As soon as an event is detected, learning rates are halved and the previous solution
recovered.
A short-term perspective for further research is to apply the proposed approach to recurrent neural
networks, and to consider more complex datasets. Another perspective is to replace the halving trick
by approximating the line search, e.g. by exploiting the gaps between the actual momentum and the
reference one, for several values of the momentum weight factor. A third perspective regards the
adaptation of the PH detection threshold ∆ during learning. In some runs, the PH test is triggered
over and over, resulting in a very small learning rate preventing any further improvement. The goal is
to adapt ∆ when the PH mechanism is re-initialized (line 9 of Algorithm 1) from the current loss
values. Another perspective is to apply SAL E RA to
Acknowledgments
We heartfully thank Steve Altschuler and Lani Wu for making this work possible, supporting it, as
well as for very insightful discussions. We also thank Yann Ollivier and Sigurd Angenent for other
insightful discussions, and the anonymous reviewers of a preliminary version of this paper for their
accurate and constructive comments.
References
S.-I. Amari. Natural gradient works efficiently in learning. Neural Comput., 10(2):251–276, Feb.
1998. ISSN 0899-7667. doi: 10.1162/089976698300017746. URL http://dx.doi.org/10.
1162/089976698300017746.
M. Andrychowicz, M. Denil, S. G. Colmenarejo, M. W. Hoffman, D. Pfau, T. Schaul, and N. de Freitas.
Learning to learn by gradient descent by gradient descent. In D. D. Lee, M. Sugiyama, U. von
Luxburg, I. Guyon, and R. Garnett, editors, NIPS 16, pages 3981–3989, 2016.
Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is
difficult. Trans. Neur. Netw., 5(2):157–166, Mar. 1994. ISSN 1045-9227. doi: 10.1109/72.279181.
URL http://dx.doi.org/10.1109/72.279181.
L. Bottou and O. Bousquet. The tradeoffs of large scale learning. In J. Platt, D. Koller, Y. Singer, and
S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20, pages 161–
10
168. NIPS Foundation (http://books.nips.cc), 2008. URL http://leon.bottou.org/papers/
bottou-bousquet-2008.
R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A matlab-like environment for machine
learning. In BigLearn, NIPS Workshop, 2011.
A. Defazio, F. R. Bach, and S. Lacoste-Julien. SAGA: A fast incremental gradient method with
support for non-strongly convex composite objectives. In Z. Ghahramani, M. Welling, C. Cortes,
N. D. Lawrence, and K. Q. Weinberger, editors, NIPS14, pages 1646–1654, 2014.
J. Deng, K. Li, M. Do, H. Su, and L. Fei-Fei. Construction and Analysis of a Large Scale Image
Ontology. Vision Sciences Society, 2009.
J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic
optimization. Technical Report UCB/EECS-2010-24, EECS Department, University of California, Berkeley, Mar 2010. URL http://www2.eecs.berkeley.edu/Pubs/TechRpts/2010/
EECS-2010-24.html.
D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, and S. Bengio. Why does unsupervised
pre-training help deep learning? J. Mach. Learn. Res., 11:625–660, Mar. 2010. ISSN 1532-4435.
URL http://dl.acm.org/citation.cfm?id=1756006.1756025.
A. P. George and W. B. Powell. Adaptive stepsizes for recursive estimation with applications in
approximate dynamic programming. Mach. Learn., 65(1):167–198, Oct. 2006. ISSN 0885-6125.
doi: 10.1007/s10994-006-8365-9. URL http://dx.doi.org/10.1007/s10994-006-8365-9.
X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks.
In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics,
AISTATS 2010, Chia Laguna Resort, Sardinia, Italy, May 13-15, 2010, pages 249–256, 2010. URL
http://www.jmlr.org/proceedings/papers/v9/glorot10a.html.
I. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. MIT Press, 2016. http://www.
deeplearningbook.org.
N. Hansen and A. Ostermeier. Completely derandomized self-adaptation in evolution strategies.
Evolutionary Computation, 9(2):159–195, 2001.
M. Hardt, B. Recht, and Y. Singer. Train faster, generalize better: Stability of stochastic gradient
descent. CoRR, abs/1509.01240, 2015. URL http://arxiv.org/abs/1509.01240.
D. Hinkley. Inference about the change point from cumulative sum-tests. Biometrika, 58:509–523,
1970.
S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing
internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning,
ICML 2015, Lille, France, 6-11 July 2015, pages 448–456, 2015. URL http://jmlr.org/
proceedings/papers/v37/ioffe15.html.
R. A. Jacobs. Increased rates of convergence through learning rate adaptation. Neural Networks, 1
(4):295–307, 1988.
D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014.
URL http://dblp.uni-trier.de/db/journals/corr/corr1412.html#KingmaB14.
A. Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009.
Y. Le Cun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document
recognition. In Proceedings of the IEEE, pages 2278–2324, 1998.
J. Martens. Deep learning via hessian-free optimization. In Proceedings of the 27th International
Conference on Machine Learning (ICML-10), June 21-24, 2010, Haifa, Israel, pages 735–742,
2010. URL http://www.icml2010.org/papers/458.pdf.
J. Martens, I. Sutskever, and K. Swersky. Estimating the hessian by back-propagating curvature. In
ICML. icml.cc / Omnipress, 2012.
11
Y. Nesterov. A method of solving a convex programming problem with convergence rate O(1/sqr(k)).
Soviet Mathematics Doklady, 27:372–376, 1983.
Y. Ollivier, L. Arnold, A. Auger, and N. Hansen. Information-geometric optimization algorithms: A
unifying picture via invariance principles. Journal of Machine Learning Research, 18(18):1–65,
2017. URL http://jmlr.org/papers/v18/14-467.html.
E. Page. Continuous inspection schemes. Biometrika, 41:100–115, 1954.
R. Pascanu and Y. Bengio. Revisiting natural gradient for deep networks. In International Conference
on Learning Representations 2014 (Conference Track), Apr. 2014. URL http://arxiv.org/
abs/1301.3584.
R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training recurrent neural networks.
In Proceedings of the 30th International Conference on International Conference on Machine
Learning - Volume 28, ICML’13, pages III–1310–III–1318. JMLR.org, 2013. URL http://dl.
acm.org/citation.cfm?id=3042817.3043083.
B. Polyak. Some methods of speeding up the convergence of iteration methods. USSR Computational
Mathematics and Mathematical Physics, 4(5):1–17, 1964.
H. Robbins and S. Monro. A stochastic approximation method. Annals of Mathematical Statistics,
22:400–407, 1951.
T. Schaul, S. Zhang, and Y. Le Cun. No More Pesky Learning Rates. In Proceedings of the 30th
International Conference on International Conference on Machine Learning, volume 28, pages
343–351, 2013.
N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A simple way
to prevent neural networks from overfitting. J. Mach. Learn. Res., 15(1):1929–1958, Jan. 2014.
ISSN 1532-4435. URL http://dl.acm.org/citation.cfm?id=2627435.2670313.
I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of initialization and momentum
in deep learning. In Proceedings of the 30th International Conference on International Conference
on Machine Learning, volume 28 of ICML’13, pages III–1139–III–1147. JMLR.org, 2013. URL
http://dl.acm.org/citation.cfm?id=3042817.3043064.
T. Tieleman and G. Hinton. Lecture 6.5 - RMSProp, COURSERA: Neural Networks for Machine
Learning. Technical report, Technical report, 2012.
D. Wierstra, T. Schaul, T. Glasmachers, Y. Sun, J. Peters, and J. Schmidhuber. Natural evolution
strategies. J. Mach. Learn. Res., 15(1):949–980, 2014. ISSN 1532-4435.
M. D. Zeiler. ADADELTA: an adaptive learning rate method. CoRR, abs/1212.5701, 2012. URL
http://arxiv.org/abs/1212.5701.
M. Zinkevich. Online convex programming and generalized infinitesimal gradient ascent. In
Proceedings of the Twentieth International Conference on International Conference on Machine
Learning, ICML’03, pages 928–935. AAAI Press, 2003. ISBN 1-57735-189-4. URL http:
//dl.acm.org/citation.cfm?id=3041838.3041955.
12
Appendix A: Derivation of formula 4
Let us define r0 = 0, and for t > 0
rt = αut + (1 − α)rt−1
with ut a unit vector uniformly drawn in IRd . We have by recurrence for t > 0
rt = α
t
X
(1 − α)(t−l) ul
l=1
Thus
t
X
X
(1 − α)2t−l−k < ul , uk >
krt k2 = α2 (1 − α)2(t−l) kul k2 +
l=1
(8)
l6=k
We are dealing with unit vectors which are uniformly drawn. Thus with δ denoting the Kronecker
delta
∀ l > 0 kul k2 = 1
∀ l, k > 0 E(< ul , uk >) = δlk
Taking the expectation in 8 we have
t
X
E(krt k ) = α
(1 − α)2(t−l)
2
2
(9)
l=1
that is,
1 − (1 − α)2t
1 − (1 − α)2
α
=
1 − (1 − α)2t
2−α
E(krt k2 ) = α2
The derivation of formula 5 is similar.
13
(10)
(11)
Appendix B: Safe Parameter-wise Agnostic LEarning Rate Adaptation
Algorithm 2 gives the detailed implementation of the parameter-wise version of both AL E RA (white
lines) and SAL E RA (plus greyed lines), as briefly described in√
Section 2.1. In line 17, denotes the
0
0
coordinate-wise multiplication, and µ = µ/d, σ(d) = σ(d)/ d.
Algorithm 2: SPAL E RA: Parameter-wise AL E RA and Page-Hickley change detection
1
2
3
4
Input: Model with parameter θ in Rd and loss function L
Parameters : Memory rate α, factor C
//algorithm parameters
Initial learning rate η0 , mini-batch ratio ρ
// run parameters
Initialize : t = 0; p ← 0; ∀ i ∈ [|1; d|] ηi ← 1; init(θ)
// initialization
¯
L, Lmin , `, `, ← 0; ∆ ← L(θ, first mini-batch)/10
// initialize Page-Hinkley
while stopping criterion not met do
M B ← new mini-batch ; t ← t + 1
// perform forward pass
` = ρL(θ, MB) + (1 − ρ)`
¯
`¯ ← (` + t`)/(t
+ 1)
// empirical mean of batch losses
5
¯
L ← L + (` − `)
6
Lmin ← min(Lmin , L)
7
if L − Lmin > ∆ then
8
θ ← θ(b) ; η ← η/2
9
¯ ` ← 0; t ← 0
L, Lmin , `,
10
11
12
13
14
15
16
17
18
19
// cumulated deviations from mean
// lower bound of deviations
// Page-Hinkley triggered: backtrack
// and re-initialize Page-Hinkley
else
θ(b) ← θ
// save for possible backtracks
g ← ∇θ L(θ, MB)
// compute gradient with backward pass
p ← αg/kgk + (1 − α)p
// exponential moving average of normalized gradients
for i=1,d do
ηi ← η0 ηi exp C(p2i − µ0 )/σ(d)0
// parameter-wise agnostic learning rate update
end
θ ←θ−η g
// standard parameter update
end
end
14
Appendix C. Analysis of the dividing factor in 1D
We use here the context and notations of Section 2.2, and make it even simpler by assuming one
dimension (the gradient direction does not change), and minimizing the 1D parabola F (θ) = 12 aθ2 .
It is straightforward to show that, independently of the current solution θt , the optimal value for the
learning rate η is η ∗ = 1/a, and that the value above which the loss will deteriorate is η − = 2η ∗ .
Let us assume that the current learning rate is ηt > η − , and that the recovery phase of SAL E RA will
be used to prevent further catastrophic event, with a dividing factor ζ > 1, and let us show that ζ = 2
is the best trade-off between bringing η back into ]0, η − ] and then reaching the optimal value η ∗ , as
discussed in Section 2.2.
After the PH test has been triggered, a first phase brings ηt below η − by successive divisions by ζ.
The number of such divisions is S = logζ ηη−t + 1 = log( ηη−t )/ log(ζ).
−
The second phase uses the standard AL E RA procedure to reach η ∗ from this value ηt+1 ∈] ηζ , η − ].
In the 1D context, let us assume a simplified procedure, that updates η by multiplying it by some
1 − ε if η > η ∗ and by 1 + ε if ηt < η ∗ , for some small ε.
Let us consider two cases, depending on whether ζ is smaller or greater than 2. We can compute the
expectation of the number of iteration of SAL E RA that are needed to reach η ∗ from ηt+1 .
−
If ζ ≤ 2, ηζ ≥ η ∗ , hence the further standard update phase decreases η by multiplying it by 1 − ε
until becoming less than (and very close to) η ∗ . The length of this update phase increases with ηt+1 ,
thus by construction it is minimal for ζ = 2. In the meanwhile, the length of the division phase
decreases as ζ increases; thus the optimal value for ζ ∈]1, 2] is ζ = 2.
If ζ ≥ 2, then ηηt+1
∈] ζ1 , 1[. Let us consider the two intervals ] ζ1 , 12 [ and ] 12 , 1[, and U and T the
−
respective expectations of the number of AL E RA iterations to reach η ∗ . The expectation of the total
number of iteration is the sum of U and T, weighted by the probability of arriving in the respective
ζ−2
ζ
intervals, i.e., the lengths of these intervals. These weights are hence 12 ζ−1
for U and 12 ζ−1
for T.
The expected value for T is independent of ζ. For u ∈] 12 , 1[, and only counting the number of
log(2u)
multiplications needed to get close to 12 , the number of iterations is − log(1−ε)
≡ log(2u)
. Integrating
ε
1
over ] − 2 , 1[ gives
Z
1
1 1
1
1
T =
log(2u) = [t log(t) − t + t log(2)]11 = (log(2) − )
2
ε 12
ε
ε
2
Similarly, the expected value for U can be computed over the interval ζ1 , 21 [ using the same approach.
Only counting again the number of multiplications needed to get close to 12 , it comes
U=
1
ε
1
2
Z
− log(2u) =
1
ζ
1
1
1
[−t log(t) + t − t log(2)] 21 =
ζ
ε
ε
1 log(ζ) 1 − log(2)
−
−
2
ζ
ζ
Let us summarize now the different computational costs involved after a catastrophe has been detected
and before the η ∗ grail is reached. The cost of the dividing iterations only involves a forward pass
on the current minibatch. Let us denote this cost by cF . On the other hand, the standard AL E RA
iterations have a larger cost, involving a forward pass plus a backward pass and the weights update.
Let us denote this cost by cB .
We are looking for the value of ζ that will minimize the total cost of reaching η ∗ after a catastrophic
event has been detected, i.e., that minimizes
1ζ −2
1 ζ
cF S + cB (
U+
T)
2ζ −1
2ζ −1
or, equivalently, that minimizes
C
1 ζ − 2 1 log(ζ) 1 − log(2)
1 ζ
1
J(ζ) =
+
−
−
+
(log(2) − )
(12)
log(ζ) 2 ζ − 1 2
ζ
ζ
2ζ −1
2
15
Average cost of SALeRA
C=0.1
C=0.01
C=0.001
J
0.5
0.4
0.3
3
5
7
9
Zeta
Figure 3: Cost function J (Equation 12) for different values of constant C. The plots for smaller
values of C are indistinguishable from that of C = 0.001.
with
C=
cF
ηt
ε log( − )
cB
η
It is easy to empirically check (Figure 3) that J(ζ) has its global minimum between 3 and 5, that
depends on the value of the constant C, assumed small here (C ∈]0, 0.1]). Then J increases to some
asymptotic value. However, the value ζ = 2 was initially chosen for historical reasons, by reference
to the famed doubling trick frequently used in different areas of Machine Learning. In the light of
these results in the simple 1D case, further work will investigate slightly larger values.
16
| 9cs.NE
|
Discrete Mathematics and Theoretical Computer Science
DMTCS vol. 19:1, 2017, #16
arXiv:1601.06456v5 [math.CO] 1 Jun 2017
On universal partial words
Herman Z. Q. Chen1 Sergey Kitaev2 Torsten Mütze3 Brian Y. Sun4∗
1
2
3
4
School of Science, Tianjin Chengjian University, P.R. China
Department of Computer and Information Sciences, University of Strathclyde, Glasgow, UK
Institut für Mathematik, TU Berlin, Germany
College of Mathematics and System Science, Xinjiang University, Urumqi, Xinjiang 830046, P.R. China
received 10th Nov. 2016, revised 8th May 2017, accepted 16th May 2017.
A universal word for a finite alphabet A and some integer n ≥ 1 is a word over A such that every word in An appears
exactly once as a subword (cyclically or linearly). It is well-known and easy to prove that universal words exist for
any A and n. In this work we initiate the systematic study of universal partial words. These are words that in addition
to the letters from A may contain an arbitrary number of occurrences of a special ‘joker’ symbol ✸ ∈
/ A, which can
be substituted by any symbol from A. For example, u = 0✸011100 is a linear partial word for the binary alphabet
A = {0, 1} and for n = 3 (e.g., the first three letters of u yield the subwords 000 and 010). We present results on
the existence and non-existence of linear and cyclic universal partial words in different situations (depending on the
number of ✸s and their positions), including various explicit constructions. We also provide numerous examples of
universal partial words that we found with the help of a computer.
Keywords: universal word, partial word, De Bruijn graph, Eulerian cycle, Hamiltonian cycle
1 Introduction
De Bruijn sequences are a centuries-old and well-studied topic in combinatorics, and over the years
they found widespread use in real-world applications, e.g., in the areas of molecular biology [CPT11],
computer security [Ral82], computer vision [PSCF05], robotics [Sch01] and psychology experiments
[SBSH97]. More recently, they have also been studied in a more general context by constructing universal cycles for other fundamental combinatorial structures such as permutations or subsets of a fixed
ground set (see e.g. [Hur90, CDG92, HRW12, SW14]).
In the context of words over a finite alphabet A, we say that a word u is universal for An if u contains
every word of length n ≥ 1 over A exactly once as a subword. We distinguish cyclic universal words
and linear universal words. In the cyclic setting, we view u as a cyclic word and consider all subwords
of length n cyclically across the boundaries of u. In the linear setting, on the other hand, we view u as
a linear word and only consider subwords that start and end within the index range of letters of u. From
∗ The last author was supported by the Scientific Research Program of the Higher Education Institution of Xinjiang Uygur
Autonomous Region (No. XJEDU2016S032) and the Natural Science Foundation of Xinjiang University.
ISSN 1365–8050
c 2017 by the author(s)
Distributed under a Creative Commons Attribution 4.0 International License
2
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
this definition it follows that the length of a cyclic or linear universal word must be |A|n or |A|n + n − 1,
respectively. For example, for the binary alphabet A = {0, 1} and for n = 3, u = 0001011100 is a
linear universal word for A3 . Observe that a cyclic universal word for An can be easily transformed into a
linear universal word for An , so existence results in the cyclic setting imply existence results for the linear
setting. Note also that reversing a universal word, or permuting the letters of the alphabet yields a universal
word again. The following classical result is the starting point for our work (see [dB46, Yoe62, Lem71]).
Theorem 1 For any finite alphabet A and any n ≥ 1, there exists a cyclic universal word for An .
The standard proof of Theorem 1 is really beautiful and concise, using the De Bruijn graph, its line
graph and Eulerian cycles (see [CDG92] and Section 2 below).
1.1 Universal partial words
In this paper we consider the universality of partial words, which are words that in addition to letters from
A may contain any number of occurrences of an additional special symbol ✸ ∈
/ A. The idea is that every
occurrence of ✸ can be substituted by any symbol from A, so we can think of ✸ as a ‘joker’ or ‘wildcard’
symbol. Formally, we define A✸ := A ∪ {✸} and we say that a word v = v1 v2 · · · vn ∈ An appears as
a factor in a word u = u1 u2 · · · um ∈ Am
✸ if there is an integer i such that ui+j = ✸ or ui+j = vj for
all j = 1, 2, . . . , n. In the cyclic setting we consider the indices of u in this definition modulo m. For
example, in the linear setting and for the ternary alphabet A = {0, 1, 2}, the word v = 021 occurs twice
as a factor in u = 120✸120021 because of the subwords 0✸1 and 021 of u, whereas v does not appear as
a factor in u′ = 12✸11✸.
Partial words were introduced in [BB99], and they too have real-world applications (see [BS08] and
references therein). In combinatorics, partial words appear in the context of primitive words [BS05], of
(un)avoidability of sets of partial words [BSBK+ 09, BBSGR10], and also in the study of the number
of squares [HHK08] and overlap-freeness [HHKS09] in (infinite) partial words. The concept of partial
words has been extended to pattern-avoiding permutations in [CJJK11].
The notion of universality given above extends straightforwardly to partial words, and we refer to a
universal partial word as an upword for short. Again we distinguish cyclic upwords and linear upwords.
The simplest example for a linear upword for An is ✸n := ✸✸ · · · ✸, the word consisting of n many ✸s,
which we call trivial. Let us consider a few more interesting examples of linear upwords over the binary
alphabet A = {0, 1}. We have that ✸✸0111 is a linear upword for A3 , whereas ✸✸01110 is not a linear
upword for A3 , because replacing the first two letters ✸✸ by 11 yields the same factor 110 as the last
three letters. Similarly, 0✸1 is not a linear upword for A2 because the word 10 ∈ A2 does not appear as a
factor (and the word 01 ∈ A2 appears twice as a factor).
1.2 Our results
In this work we initiate the systematic study of universal partial words. It turns out that these words are
rather shy animals, unlike their ordinary counterparts (universal words without ‘joker’ symbols). That
is, in stark contrast to Theorem 1, there are no general existence results on upwords, but also many
non-existence results. The borderline between these two cases seems rather complicated, which makes
the subject even more interesting (this is true also for non-binary alphabets, as the constructions of the
follow-up paper [GGH+ 16] indicate). In addition to the size of the alphabet A and the length n of the
factors, we also consider the number of ✸s and their positions in an upword as problem parameters.
3
On universal partial words
n
1
2
3
4
5
k
1
1
2
1
2
3
4
1
2
3
4
5
6
7
8
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
✸
✸011 (Thm. 9, Thm. 17)
— (Thm. 6)
✸00111010 (Thm. 9)
0✸011100 (Thm. 10)
— (Thm. 6)
— (Thm. 7)
✸00011110100101100 (Thm. 9)
0✸010011011110000 (Thm. 10)
01✸0111100001010 (Thm. 10)
— (Thm. 6)
— (Thm. 7)
01100✸011110100
— (Thm. 7)
0011110✸0010110
✸0000111110111001100010110101001000 (Thm. 9)
0✸01011000001101001110111110010001 (Thm. 10)
01✸011000001000111001010111110100 (Thm. 10)
011✸0111110000010100100011010110 (Thm. 10)
— (Thm. 6)
00101✸0010011101111100000110101
010011✸010000010101101111100011
0100110✸01000001110010111110110
01110010✸0111110110100110000010
010011011✸010001111100000101011
0101000001✸01011111001110110001
01010000011✸0101101111100010011
001001101011✸001010000011111011
0011101111100✸00110100010101100
01010000010011✸0101101111100011
001000001101011✸001010011111011
Tab. 1: Examples of linear upwords for An , A = {0, 1}, with a single ✸ at position k from the beginning or end
for n = 1, 2, 3, 4, 5 and all possible values of k (upwords where the ✸ is closer to the end of the word than to the
beginning can be obtained by reversal). A dash indicates that no such upword exists.
We first focus on linear upwords. For linear upwords containing a single ✸, we have the following
results: For non-binary alphabets A (i.e., |A| ≥ 3) and n ≥ 2, there is no linear upword for An with a
single ✸ at all (Theorem 5 below). For the binary alphabet A = {0, 1}, the situation is more interesting
(see Table 1): Denoting by k the position of the ✸, we have that for n ≥ 2, there is no linear upword for
An with k = n (Theorem 6), and there are no linear upwords in the following three cases: n = 3 and
k = 4, and n = 4 and k ∈ {5, 7} (Theorem 7). We conjecture that these are the only non-existence cases
4
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
n
2
3
4
5
✸✸ (Cor. 12)
✸✸0111 (Cor. 12, Thm. 17)
✸001011✸
✸00011✸1001011 (Thm. 13)
✸0001011✸10011
001✸110✸001
✸0100✸101011000001110111110010
✸0000111✸100010010101100110111 (Thm. 13)
✸00001001✸10001101011111011001
✸0000100111✸100011001010110111
✸00001010111✸10001101100100111
0✸0011✸0100010101101111100000
0✸010110✸00011101111100100110
0✸0101110✸0001101100100111110
0✸010111110✸00011011001001110
0✸0101101110✸0001100100111110
00✸0011✸00101011011111010000
01✸01100101110✸0100000111110
01✸0110010111110✸01000001110
01✸0100000101011000111110✸110
001✸0101✸001110111110000010
011✸011010010✸0111110000010
011✸0110101001000✸011111000
011✸0111110001101010000010✸10
011✸011010000011111000100101✸1
01001✸1110✸010000011011001
Tab. 2: Examples of linear upwords for An , A = {0, 1}, with two ✸s for n = 2, 3, 4, 5.
for a binary alphabet (Conjecture 8). To support this conjecture, we performed a computer-assisted search
and indeed found linear upwords for all values of 2 ≤ n ≤ 13 and all possible values of k other than
the ones excluded by the beforementioned results. Some of these examples are listed in Table 1, and the
remaining ones are available on the third author’s website [www]. We also prove the special cases k = 1
and k ∈ {2, 3, . . . , n − 1} of our conjecture (Theorems 9 and 10, respectively).
For linear upwords containing two ✸s we have the following results: First of all, Table 2 shows examples of linear upwords with two ✸s for the binary alphabet A = {0, 1} for n = 2, 3, 4, 5. We establish
a sufficient condition for non-existence of binary linear upwords with two ✸s (Theorem 11), which in
particular shows that a (1 − o(1))-fraction of all ways of placing two ✸s among the N = Θ(2n ) positions
does not yield a valid upword. Moreover, we conclude that there are only two binary linear upwords where
the two ✸s are adjacent (Corollary 12), namely ✸✸ for n = 2 and ✸✸0111 for n = 3 (see Table 2). We
also construct an infinite family of binary linear upwords with two ✸s (Theorem 13).
Let us now discuss cyclic upwords. Note that the trivial solution ✸n is a cyclic upword only for n = 1.
On universal partial words
5
For the cyclic setting we have the following rather general non-existence result: If gcd(|A|, n) = 1, then
there is no cyclic upword for An (Corollary 16). In particular, for a binary alphabet |A| = 2 and odd
n, there is no cyclic upword for An . In fact, we know only of a single cyclic upword for the binary
alphabet A = {0, 1} and any n ≥ 2, namely ✸001✸110 for n = 4 (up to cyclic shifts, reversal and letter
permutations).
1.3 Outline of this paper
This paper is organized as follows. In Section 2 we introduce some notation and collect basic observations
that are used throughout the rest of the paper. In Sections 3 and 4 we prove our results on linear upwords
containing a single or two ✸s, respectively. Section 5 contains the proofs on cyclic upwords. Finally,
Section 6 discusses possible directions for further research, including some extensions of our results to
non-binary alphabets.
2 Preliminaries
For the rest of this paper, we assume w.l.o.g. that the alphabet is A = {0, 1, . . . , α − 1}, so α ≥ 2 denotes
the size of the alphabet. We often consider the special case α = 2 of the binary alphabet, and then for
x ∈ {0, 1} we write x for the complement of x. Moreover, for any word u, we let |u| denote its length. As
we mentioned before, reversing a universal word and/or permuting the letters of the alphabet again yields
a universal word. We can thus assume w.l.o.g. that in an upword u the letters of A appear in increasing
order from left to right, i.e., the first occurence of symbol i is before the first occurence of symbol j
whenever i < j. Moreover, if u can be factored as u = xyz, where x and z do not contain any ✸s, then
we can assume that |x| ≤ |z|.
One standard approach to prove the existence of universal words is to define a suitable graph and
to search for a Hamiltonian path/cycle in this graph (another more algebraic approach uses irreducible
polynomials). Specifically, the De Bruijn graph GnA has as vertices all elements from An (all words of
length n over A), and a directed edge from a vertex u to a vertex v whenever the last n − 1 letters of u are
the same as the first n − 1 letters of v. We call such an edge (u, v) an x-edge, if the last letter of v equals
x. Figure 1 (a) and (b) shows the graph GnA , A = {0, 1}, for n = 2 and n = 3, respectively. Clearly,
a linear universal word for An corresponds to a Hamiltonian path in GnA , and a cyclic universal word to
n−1
a Hamiltonian cycle in this graph. Observe furthermore that GnA is the line graph of GA
. Recall that
the line graph L(G) of a directed graph G is the directed graph that has a vertex for every edge of G,
and a directed edge from e to e′ if in G the end vertex of e equals the starting vertex of e′ . Therefore,
the problem of finding a Hamiltonian path/cycle in GnA is equivalent to finding an Eulerian path/cycle in
n−1
GA
. The existence of an Eulerian path/cycle follows from the fact that the De Bruijn graph is connected
and that each vertex has in- and out-degree α (this is one of Euler’s famous theorems [Eul36], see also
[BJG08, Theorem 1.6.3]). This proves Theorem 1. In fact, this existence proof can be easily turned into
an algorithm to actually find (many) universal words (using Hierholzer’s algorithm [HW73] or Fleury’s
algorithm [Fle83]).
We now discuss how this standard approach of proving the existence of universal words can be extended
to universal partial words. Specifically, we collect a few simple but powerful observations that will be used
in our proofs later on.
For any vertex v of GnA , we let Γ+ (v) and Γ− (v) denote the sets of out-neighbours and in-neighbours
6
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
G3A = L(G2A )
G2A
000
00
S(u, 1, n) S(u, 2, n)
01 100
10
001
010
S(u, 3, n) S(u, 5, n)
11
S(u, 4, n)
101
110
011
111
(a)
(b)
S(u, 6, n)
H(u, n), u = 0✸011100
(c)
Fig. 1: The De Bruijn graphs G2A (a) and G3A = L(G2A ) (b) for A = {0, 1} with a spanning subgraph H(u, n) of
G3A for the linear upword u = 0✸011100 for A3 (H(u, n) is shown by solid edges). Part (c) if the figure shows a
schematic drawing of the graph H(u, n). H(u, n) is the line graph of the highlighted sequences of edges in G2A .
of v, respectively (both are sets of vertices of GnA ). As we mentioned before, we clearly have |Γ+ (v)| =
|Γ− (v)| = α.
Observation 2 For any vertex v = v1 v2 · · · vn of GnA and its set of out-neighbours Γ+ (v), there are
α − 1 vertices different from v with the same set of out-neighbours Γ+ (v), given by xv2 v3 · · · vn , where
x ∈ A \ {v1 }. For any vertex v = v1 v2 · · · vn of GnA and its set of in-neighbours Γ− (v), there are
α − 1 vertices different from v with the same set of in-neighbours Γ− (v), given by v1 v2 · · · vn−1 x, where
x ∈ A \ {vn }.
For any linear upword u for An , we define a spanning subgraph H(u, n) of the De Bruijn graph GnA as
follows, see Figure 1 (c): For any i = 1, 2, . . . , N − n + 1, we let S(u, i, n) denote the set of all words that
are obtained from the subword of u of length n starting at position i by replacing any occurences of ✸ by
a letter from the alphabet A. Clearly, if there are d many ✸s in this subword, then there are αd different
possibilities for substitution, so we have |S(u, i, n)| = αd . Note that the sets S(u, i, n) form a partition
of the vertex set of GnA (and H(u, n)). The directed edges of H(u, n) are given by all the edges of GnA
induced between every pair of consecutive sets S(u, i, n) and S(u, i + 1, n) for i = 1, 2, . . . , N − n.
For example, for the linear upword u = 0✸011100 over the binary alphabet A = {0, 1} for n = 3
we have S(u, 1, n) = {000, 010}, S(u, 2, n) = {001, 101}, S(u, 3, n) = {011}, S(u, 4, n) = {111},
S(u, 5, n) = {110} and S(u, 6, n) = {100}, and the spanning subgraph H(u, n) of G3A is shown in
Figure 1 (c). To give another example with the same A and n, for the linear upword u = ✸✸0111 we have
S(u, 1, n) = {000, 010, 100, 110}, S(u, 2, n) = {001, 101}, S(u, 3, n) = {011}, S(u, 4, n) = {111},
and then H(u, n) is a binary tree of depth 2 with an additional edge emanating from the root.
7
On universal partial words
H(u, n) (solid edges)
vx
vx = v1 · · · vn−1 x, x ∈ A \ {vn }
v
v = v1 v2 · · · vn
...
S(u, k + 1, n)
S(u, 1, n)
S(u, k, n)
Fig. 2: Illustration of the proof of Lemma 4.
The following observation follows straightforwardly from these definitions.
Observation 3 Let u = u1 u2 · · · uN be a linear upword for An . A vertex in S(u, i, n), i = 1, 2, . . . , N −
n, has out-degree 1 in H(u, n) if ui+n ∈ A, and out-degree α if ui+n = ✸. A vertex in S(u, i, n),
i = 2, 3, . . . , N − n + 1, has in-degree 1 in H(u, n) if ui−1 ∈ A, and in-degree α if ui−1 = ✸. The
vertices in S(u, 1, n) have in-degree 0, and the vertices in S(u, N − n + 1, n) have out-degree 0.
By this last observation, the graph H(u, n) is determined only by the positions of the ✸s in u. Intuitively, the ✸s lead to branching in the graph H(u, n) due to the different possibilities of substituting
symbols from A. In particular, if u has no ✸s, then H(u, n) is just a spanning path of GnA (i.e., a Hamiltonian path, so we are back in the setting of Theorem 1). So when searching for a linear universal partial
word u with a particular number of ✸s at certain positions, we essentially search for a copy of the spanning subgraph H(u, n) in GnA . We will exploit this idea both in our existence and non-existence proofs.
For the constructions it is particularly useful (and for our computer-searches it is computationally much
more efficient) to not search for a copy of H(u, n) in GnA directly, but to rather search for the correspondn−1
ing sequences of edges in GA
, which can be seen as generalizations of Eulerian paths that were used
before in the proof of Theorem 1 (see Figure 1 (a)). For example, to search for a linear upword u with a
single ✸ at position k ∈ {1, 2, . . . , n − 1}, we can prescribe the first k − 1 letters and the n letters after
the ✸ (with a particular choice of symbols from A, or by iterating over all possible choices), and search
n−1
that remains when deleting from it all edges that correspond
for an Eulerian path in the subgraph of GA
to the prescribed prefix of u (see the proofs of Theorems 9 and 10 below). This idea can be generalized
straightforwardly to search for upwords with other ✸ patterns (see for example the proof of Theorem 13
below).
The next lemma will be used repeatedly in our proofs (both for existence and non-existence of upwords). The proof uses the previous two graph-theoretical observations to derive dependencies between
letters of an upword.
Lemma 4 Let u = u1 u2 · · · uN be a linear upword for An , A = {0, 1, . . . , α − 1}, n ≥ 2, such that
uk = ✸ and uk+n 6= ✸ (we require k + n ≤ N ). Then for all i = 1, 2, . . . , n − 1 we have that if ui 6= ✸,
then uk+i = ui . Moreover, we have that if un 6= ✸, then α = 2 and uk+n = un .
Proof: By Observation 3, each vertex in the set S(u, k + 1, n) has in-degree α in H(u, n), and each
vertex in S(u, k, n) has out-degree 1. By Observation 2, for each v = v1 v2 · · · vn ∈ S(u, k + 1, n) there
are α − 1 other vertices (different from the ones in S(u, k + 1, n)) in GnA with the same set Γ− (v) of α
8
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
many in-neighbors, namely vx := v1 · · · vn−1 x, where x ∈ A \ {vn } (see Figure 2). As the in-degree
of every vertex of GnA is exactly α, and in H(u, n) all vertices except the ones in S(u, 1, n) already have
in-degree at least 1, it follows that each of the vertices vx must be equal to one of the vertices in S(u, 1, n).
It follows that if ui 6= ✸ then uk+i 6= ✸ and ui = vi = uk+i for all i = 1, 2, . . . , n − 1. Moreover,
if un 6= ✸ and α ≥ 3, then there are at least two vertices vx , x ∈ A \ {vn }, ending with different
symbols x, each of which must be equal to one of the vertices in S(u, 1, n), which is impossible because
all words in this set end with the same symbol un . It follows that if un 6= ✸ then we must have α = 2
✷
and un = x 6= vn = uk+n , so uk+n = un .
3 Linear upwords with a single ✸
3.1 Non-existence results
Our first result completely excludes the existence of linear upwords with a single ✸ for non-binary alphabets.
Theorem 5 For A = {0, 1, . . . , α − 1}, α ≥ 3, and any n ≥ 2, there is no linear upword for An with a
single ✸.
Proof: Suppose that such an upword u = u1 u2 · · · uk−1 ✸uk+1 · · · uN exists. We claim that the ✸ in u is
preceded or followed by at least n symbols from A. If not, then u would have at most αn different factors,
which is strictly less than αn for α ≥ 3 and n ≥ 2. So we assume w.l.o.g. that the ✸ in u is followed
by at least n symbols from A, i.e., k + n ≤ N . By Lemma 4 we have ui = ✸ or uk+i = ui for all
i = 1, 2, . . . , n − 1 and un = ✸, which implies k = n and therefore un+i = ui for all i = 1, . . . , n − 1.
But this means that the word v := un+1 · · · u2n ∈ An appears twice as a factor in u starting at positions
1 and n + 1 (in other words, the vertex v ∈ S(u, n + 1, n) is identical to a vertex from S(u, 1, n) in
H(u, n)), a contradiction.
✷
Our next result excludes several cases with a single ✸ for a binary alphabet.
Theorem 6 For A = {0, 1} and any n ≥ 2, there is no linear upword for An with a single ✸ at position
n from the beginning or end.
Proof: We first consider the case n = 2. Suppose that there is an upword u = u1 ✸u3 for An . Assuming
w.l.o.g. that u1 = 0, we must have u3 = 1, otherwise the word 00 would appear twice as a factor. But
then the word 10 does not appear as a factor in u = 0✸1, while 01 appears twice, a contradiction.
For the rest of the proof we assume that n ≥ 3. Suppose there was an upword u =
u1 u2 · · · un−1 ✸un+1 · · · uN with N = 2n −1. Note that N −n ≥ n, or equivalently 2n ≥ 2n+1, holds by
our assumption n ≥ 3, so the ✸ in u is followed by at least n more symbols from A. Applying Lemma 4
yields that un+i = ui for all i = 1, . . . , n − 1, which means that the word v := un+1 · · · u2n ∈ An
appears twice as a factor in u starting at positions 1 and n + 1, a contradiction.
✷
In contrast to Theorem 5, for a binary alphabet we can only exclude the following three more (small)
cases in addition to the cases excluded by Theorem 6 (all the exceptions are marked in Table 1).
Theorem 7 For A = {0, 1}, there is no linear upword for An with a single ✸ at position k from the
beginning or end in the following three cases: n = 3 and k = 4, and n = 4 and k ∈ {5, 7}.
On universal partial words
9
Proof: Suppose that there is an upword u = u1 u2 u3 ✸u5 u6 u7 for the case n = 3. Applying Lemma 4
twice to u and its reverse we obtain that u5 u6 u7 = u1 u2 u3 and u1 u2 u3 = u5 u6 u7 , a contradiction.
To prove the second case suppose that there is an upword of the form u = u1 u2 u3 u4 ✸u6 · · · u15
for n = 4. Applying Lemma 4 twice to u and its reverse we obtain that u has the form u =
u1 u2 u3 u4 ✸u1 u2 u3 u4 u10 u11 u1 u2 u3 u4 . We assume w.l.o.g. that u1 = 0. The word z := 0000 must
appear somewhere as a factor in u, and since u12 = u1 = 1, the only possible starting positions for z in
u are 1, 2, . . . , 8. However, the starting positions 1, 2, 5, 6, 7 can be excluded immediately, as they would
cause z to appear twice as a factor in u. On the other hand, if z starts at positions 3, 4 or 8, then the
neighboring letters must both be 1, causing 0101, 1010 or 1101, respectively, to appear twice as a factor
in u, a contradiction.
The proof of the third case proceeds very similarly to the second case, and allows us to conclude that
such an upword u must have the form u = u1 u2 u3 u4 u5 u6 ✸u1 u2 u3 u4 u3 u4 u5 u6 . We assume w.l.o.g. that
u3 = 0. The word z := 0000 must appear somewhere as a factor in u, and since u12 = u3 = 1 the only
possible starting positions for z in u are 1, 2, . . . , 8. The starting positions 1, 3, 4, 6, 8 can be excluded
immediately, as they would cause z to appear twice as a factor in u. On the other hand, if z starts at
positions 2, 5 or 7, then the neighboring letters must both be 1, causing 0011, 0101 or 0000, respectively,
to appear twice as a factor in u, a contradiction.
✷
3.2 Existence results
We conjecture that for a binary alphabet and a single ✸, the non-existence cases discussed in the previous
section are the only ones.
Conjecture 8 For A = {0, 1} and any n ≥ 1, there is a linear upword for An containing a single ✸ at
position k in every case not covered by Theorem 6 or Theorem 7.
Recall the numerical evidence for the conjecture discussed in the introduction. In the remainder of this
section we prove some cases of this general conjecture.
Theorem 9 For A = {0, 1} and any n ≥ 2, there is a linear upword for An with a single ✸ at the first
position that begins with ✸0n−1 1.
Note that by Lemma 4, every linear upword for An with a single ✸ of the form u = ✸u2 u3 · · · uN
satisfies the conditions u2 = u3 = · · · = un = un+1 , i.e., w.l.o.g. it begins with ✸0n−1 1 (up to letter
permutations).
Proof: Consider the word v = v1 v2 · · · vn+1 := ✸0n−1 1 and the corresponding three edges
n−1
(0n−1 , 0n−1 ), (10n−2 , 0n−1 ) and (0n−1 , 0n−2 1) in the De Bruijn graph GA
. Denote the graph obn−1
tained from GA by removing these three edges and the isolated vertex 0n−1 by G′ . Clearly, the edges
of G′ form a connected graph, and every vertex in G′ has in- and out-degree exactly two, except the vertex
y := 0n−2 1 which has one more out-edge than in-edges and the vertex z := 10n−2 which has one more
in-edge than out-edges. Therefore, G′ has an Eulerian path starting at y and ending at z, and this Eulerian
path yields the desired upword that begins with v.
✷
For any binary word w ∈ Ak , A = {0, 1}, and any n ≥ 1, we write c(w, n) = c1 c2 · · · cn for
the word given by ci = wi for i = 1, 2, . . . , k, ci = ci−k for all i = k + 1, k + 2, . . . , n − 1 and
10
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
V1
...
V0
1
vk−1
v31
vk1
v21
...
v10
v20
v30
0
vk−1
vk0
v11 = vk+1
V2
vk+2
Fig. 3: Subgraph of Gn−1
constructed in the proof of Theorem 10.
A
cn = cn−k . Informally speaking, c(w, n) is obtained by concatenating infinitely many copies of w,
truncating the resulting word at length n and complementing the last symbol. For example, we have
c(011, 7) = 0110111 and c(011, 8) = 01101100. Using this terminology, the starting segment of the
linear upword from Theorem 9 can be written as ✸c(0, n). The next result is a considerable extension of
the previous theorem.
Theorem 10 For A = {0, 1}, any n ≥ 3 and any k ∈ {2, 3, . . . , n − 1}, there is a linear upword for An
with a single ✸ at the k-th position that begins with 01k−2 ✸c(01k−1 , n).
The idea of the proof of Theorem 10 is a straightforward generalization of the approach we used to
n−1
prove Theorem 9 before, and boils down to showing that the De Bruijn graph GA
without the edges that
are given by the prescribed upword prefix still has an Eulerian path.
Proof: The words 0✸c(01, 3)100 = 0✸011100, 0✸c(01, 4)11011110000 = 0✸010011011110000 and
01✸c(011, 4)100001010 = 01✸0111100001010 from Table 1 show that the statement is true for n = 3
and n = 4. For the rest of the proof we assume that n ≥ 5. Consider the word w = w1 w2 · · · wk+n :=
01k−2 ✸c(01k−1 , n). For i = 1, 2, . . . , k we let vi0 and vi1 denote the two words from S(w, i, n − 1)
obtained by substituting ✸ in w by 0 or 1, respectively. Moreover, let vk+1 = wk+1 · · · wk+n−1 be the
unique word from S(w, k +1, n−1) and vk+2 = wk+2 · · · wk+n the unique word from S(w, k +2, n−1),
and define V 0 := {vi0 | i = 1, 2, . . . , k}, V 1 := {vi1 | i = 1, 2, . . . , k}, V 2 := {vk+1 , vk+2 } and
V ′ := V 0 ∪ V 1 ∪ V 2 . We proceed to show that |V ′ | = 2k + 1, i.e., only two of the words just defined
coincide, namely v11 = vk+1 (v11 is given by the first n − 1 letters of w = 01k−1 c(01k−1 , n), and vk+1
is given by the first n − 1 letters of c(01k−1 , n), which are equal). In other words, the corresponding
n−1
set of vertices in GA
has size 2k + 1 (see Figure 3). If k = 2, then this can be verified directly by
considering the number of leading and trailing 0s and 1s of the vertices in V 0 , V 1 and V 2 . We now
assume that k ≥ 3. Every word from V 0 , except possibly v10 , contains the factor 00 exactly once and is
uniquely identified by the position of this factor, proving that |V 0 | = k. The words in V 1 are all uniquely
identified by the number of leading 1s, which equals 0 for v11 and k − i + 1 for i = 2, 3, . . . , k, implying
that |V 1 | = k. We now show that V 0 and V 1 are disjoint. To prove this we use again that all the words
in V 0 , except possibly v10 , contain the factor 00, and that moreover no word from V 1 contains this factor.
However v10 does not contain the factor 00 only in the case k = n − 1, and then v10 starts and ends with 0,
unlike any of the words from V 1 in this case, proving that V 0 and V 1 are disjoint. It remains to show that
vk+2 ∈
/ V 0 ∪V 1 . If k = n−1, then vk+2 = 1n−1 and all other words from V 0 and V 1 contain at least one
On universal partial words
11
0, so vk+2 ∈
/ V 0 ∪ V 1 . If k ≤ n − 2, then the word vk+2 = wk+2 · · · wk+n satisfies wk+n = wn , i.e., its
last letter and the one k positions to the left of it are complementary (recall the definition of c(01k−1 , n)),
a property that does not hold for any of the words in V 1 , implying that vk+2 ∈
/ V 1 . Moreover, in this case
0
all words from V contain the factor 00 exactly once and are uniquely identified by the position of this
factor, and vk+2 might contain the factor 00 only at the last two positions, so the only potential conflict
could arise in the case k = n−2 when v10 = 01n−4 00 ends with 00. However, in this case vk+2 = 1n−3 00
is still different from v10 . We conclude that vk+2 ∈
/ V 0 ∪ V 1 in all cases. Combining these observations
′
0
1
2
shows that |V | = |V | + |V | + |V | − 1 = 2k + 1, as claimed.
0
1
Consider the set of 2k + 1 edges E ′ := {(vi0 , vi+1
) | i = 1, 2, . . . , k − 1} ∪ {(vi1 , vi+1
) | i =
n−1
0
1
1, 2, . . . , k − 1} ∪ {(vk , vk+1 ), (vk , vk+1 ), (vk+1 , vk+2 )} in the De Bruijn graph GA (see Figure 3).
They span a subgraph on V ′ that has the following pairs of out-degrees and in-degrees: (1, 0) for the
vertex v10 , (0, 1) for the vertex vk+2 , (1, 1) for the vertices vi0 and vi1 , i = 2, 3, . . . , k, (2, 2) for the vertex
v11 = vk+1 .
n−1
We denote the graph obtained from GA
by removing the edges in E ′ and the isolated vertex v11 =
′
′
vk+1 by G . Clearly, every vertex in G has the same in- and out-degree (1 or 2), except the vertex vk+2
which has one more out-edge than in-edges, and the vertex v10 which has one more in-edge than out-edges.
To complete the proof of the theorem we show that G′ contains an Eulerian path (which must start at vk+2
and end at v10 ), and to do this, it suffices (by the before-mentioned degree conditions) to show that G′ is
connected.
We first consider the case k ≤ n − 2: From any vertex v ∈ G′ , we follow 0-edges until we either reach
the vertex 0n−1 or a vertex from V ′ for which the next 0-edge is from E ′ (this could happen right at the
beginning if v ∈ V ′ ). In this case we follow 1-edges until we reach the vertex 1n−1 , and from there we
follow 0-edges until we reach 0n−1 . (We only ever follow edges in forward direction.) We claim that in
this process we never use an edge from E ′ , which shows that G′ is connected. To see this suppose we
encounter a vertex v ′ from V ′ for which the next 0-edge is from E ′ . This means that v ′ has k − 1 trailing
1s (here we use that k ≤ n − 2), so following a 1-edge leads to a vertex that has k trailing 1s, and in
the next step to a vertex that has k + 1 trailing 1s. Note that the vertices in V ′ \ {vk+2 } have at most
k − 1 trailing 1s, and vk+2 has at most k trailing 1s, so we avoid any edges from E ′ on our way to 1n−1 .
Moreover, on the way from 1n−1 to 0n−1 via 1n−1−i 0i , i = 1, 2, . . . , n − 1, we do not use any edges
from E ′ either, because any vertex from V ′ \ {vk+2 } that starts with a 1 has at least two transitions from
1s to 0s, or vice versa, when reading it from left to right (using again k ≤ n − 2), and 0n−1 ∈
/ V ′.
Now consider the case k = n − 1: From any vertex v ∈ G′ , we follow 0-edges until we either reach
the vertex 0n−1 or the only vertex v10 = 01n−3 0 from V ′ \ {vk+1 } for which the next 0-edge is from
E ′ . In this case we follow a single 1-edge to 1n−3 01 = v31 , and from there we follow 0-edges until we
reach 0n−1 . Similarly to before, we need to argue that we never use an edge from E ′ in this process. On
the way from 1n−3 01 = v31 to 0n−1 we never use any edges from E ′ , because any vertices on this path
except the first one 1n−3 01 and the last two 10n−2 and 0n−1 contain the factor 010, so all these vertices
are different from V ′ (for n ≥ 5 and k = n − 1 no word from V ′ contains 010 as a factor), implying that
all edges except possibly the last one are safe. However, since 0n−1 ∈
/ V ′ , the last edge is safe, too.
These arguments show that G′ is connected, so it has an Eulerian path, and this Eulerian path yields the
desired upword that begins with w. This completes the proof.
✷
12
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
4 Linear upwords with two ✸s
In this section we focus on binary alphabets. Many of the non-existence conditions provided in this section
can be generalized straightforwardly to non-binary alphabets, as we briefly discuss in Section 6 below.
4.1 Non-existence results
Theorem 11 For A = {0, 1} and any n ≥ 5, there is no linear upword for An with two ✸s of the form
u = x✸y✸z if |x|, |y|, |z| ≥ n or |x| = n − 1 or |z| = n − 1 or |y| ≤ n − 2.
As Table 2 shows, there are examples of linear upwords with two ✸s whenever the conditions in Theorem 11 are violated. Put differently, for every upword u = x✸y✸z in the table for n ≥ 5 we have that
one of the numbers |x|, |y|, |z| is at most n − 1, |x| 6= n − 1, |z| 6= n − 1 and |y| ≥ n − 1. Note that
already by the first condition |x|, |y|, |z| ≥ n, a (1 − o(1))-fraction of all choices of placing two ✸s among
N = Θ(2n ) positions are excluded as possible candidates for upwords.
Proof: We first assume that |x|, |y|, |z| ≥ n, i.e., yn , zn ∈ A. Applying Lemma 4 yields zi = yi = xi ∈ A
for i = 1, 2, . . . , n − 1 and zn = yn = xn , so the word y1 y2 · · · yn = z1 z2 · · · zn appears twice as a factor
in u, a contradiction.
We now assume that |x| = n − 1 (the case |z| = n − 1 follows by symmetry). Note that the number
of factors of u is at most 2(|y| + 1) + 4(|z| + 1): This is because every subword ending at the first ✸
or at a letter from y contains at most one ✸, giving rise to two factors, and every subword ending at the
second ✸ or at a letter from z contains at most two ✸s, giving rise to four factors. This number is at most
2n + 4n = 6n for |y|, |z| ≤ n − 1, which is strictly less than 2n for n ≥ 5. Therefore, we must have
|y| ≥ n or |z| ≥ n in this case. We assume w.l.o.g. that |y| ≥ n, i.e., yn ∈ A. Applying Lemma 4 yields
yi = xi ∈ A for i = 1, 2, . . . , n − 1, implying that the word y1 y2 · · · yn appears twice as a factor in u, a
contradiction.
We now assume that |y| ≤ n−2. In this case we must have |x| ≥ n or |z| ≥ n, because if |x|, |z| ≤ n−1
then the number of factors of u is at most 2(|y| + 1) + 4(|z| + 1) ≤ 2(n − 1) + 4n ≤ 6n, which is strictly
less than 2n for n ≥ 5. We assume w.l.o.g. that |z| ≥ n. Let k := |y| + 1 ≤ n − 1 and consider
the subword y ′ := y✸z1 z2 · · · zn−k of u, which is well-defined since |z| ≥ n (k is the position of the
✸ in y ′ ). Since k ≤ n − 1 we have yn′ = zn−k ∈ A. Applying Lemma 4 yields that |x| = |y|.
′
Moreover, if k = 1 (|x| = |y| = 0) then the same lemma yields y2′ = y3′ = · · · = yn−1
= yn′ , i.e.,
z1 = z2 = · · · = zn−2 = zn−1 and zn−1 = zn−3 , a contradiction. On the other hand, if k ≥ 2, then
zi+kℓ = yi = xi for all i = 1, 2, . . . , k − 1 and ℓ = 0, 1, . . . with i + kℓ ≤ n − 1, i.e., the factors obtained
from the subword y ′ in u appear twice, starting at position 1 and position k + 1, a contradiction.
✷
Corollary 12 For A = {0, 1} and any n ≥ 2, ✸✸ for n = 2 and ✸✸0111 for n = 3 are the only linear
upwords for An containing two ✸s that are adjacent (up to reversal and letter permutations).
Proof: The non-existence of linear upwords with two adjacent ✸s for n ≥ 5 follows from Theorem 11,
because for such an upword u = x✸✸z the subword y between the two ✸s is empty, so |y| = 0 ≤ n − 2.
For n = 4 and |y| = 0 the estimate in the third part in the proof of Theorem 11 can be strengthened
to show that if |x|, |z| ≤ n − 1, then the number of factors of u is strictly less than 4n ≤ 2n unless
u = u1 u2 u3 ✸✸u6 u7 u8 , which means we can continue the argument as before, leading to a contradiction.
The exceptional case u = u1 u2 u3 ✸✸u6 u7 u8 can be excluded as follows: Applying Lemma 4 shows that
On universal partial words
13
u2 = u6 and u3 = u7 , and then it becomes clear that the factor 0000, at whatever position within u
it is placed, would appear twice. For n = 3 the only possible linear upwords with two adjacent ✸s by
Lemma 4 are u = ✸✸u3 u3 u3 u6 , which leads to ✸✸0111 (w.l.o.g. u3 = 0, and for 111 to be covered we
must have u6 = 1), and u = u1 ✸✸u4 is impossible because u1 0u4 appears twice as a factor (starting at
positions 1 and 2). For n = 2 the only possible linear upword with two ✸s is ✸✸.
✷
4.2 Existence results
Our next result provides an infinite number of binary linear upwords with two ✸s (see Table 2).
Theorem 13 For A = {0, 1} and any n ≥ 4, there is a linear upword for An with two ✸s that begins
with ✸0n−1 1n−2 ✸10n−2 1.
Proof: Consider the word w = w1 w2 · · · w3n−1 := ✸0n−1 1n−2 ✸10n−21. It is easy to check that
w yields 3n + 1 different factors x1 x2 · · · xn ∈ An , and each of these factors gives rise to an edge
n−1
(x1 x2 · · · xn−1 , x2 x3 · · · xn ) in the De Bruijn graph GA
. The set E ′ of these edges and their end
′
vertices V form a connected subgraph that has in- and out-degree 1 for all vertices in V ′ except for
v0′ := 0n−1 , v1′ := 1n−1 , v2′ := 10n−2 and v3′ := 1n−2 0 which have in- and out-degree 2, and y := 0n−2 1
and z := 01n−2 which have in-degree 2 and out-degree 1, or in-degree 1 and out-degree 2, respectively.
n−1
We denote the graph obtained from GA
by removing the edges in E ′ and the vertices v0′ , v1′ , v2′ and v3′
′
′
by G . Clearly, every vertex in G has the same in- and out-degree (1 or 2), except the vertex y which has
only one outgoing edge, and the vertex z which has only one incoming edge. To complete the proof of the
theorem we show that G′ contains an Eulerian path (which must start at y and end at z), and to do this, it
suffices (by the before-mentioned degree conditions) to show that G′ is connected.
If n = 4, then G′ consists only of the edges (y, 010), (010, 101), (101, z) (a connected graph), so for
the rest of the proof we assume that n ≥ 5. Consider a vertex v in G′ other than z.
If v ends with 0, consider the (maximum) number k of trailing 0s. Note that k ≤ n − 3, as the vertices
v2′ and v0′ that correspond to the cases k ∈ {n − 2, n − 1} are not in G′ . From v we follow 1-edges and
0-edges alternatingly, starting with a 1-edge, until we either reach the vertex s := 1n−3 01 or the vertex
t := 010101 · · · ∈ An−1 (this could happen right at the beginning if v = t). From s or t we follow
1-edges until the vertex z.
If v ends with 1, then we do the following: If v 6= s we follow a single 0-edge, and then proceed as
before until the vertex z. If v = s we directly follow 1-edges until z. (Note that we only ever follow edges
in forward direction.)
We claim that in this process we never use an edge from E ′ , which shows that G′ is connected. To see
this we first consider the case that we start at a vertex v with k ≤ n − 3 trailing 0s. If k ≥ 2, then the
vertex reached from v via a 1-edge is not in V ′ , because no vertex in V ′ has a segment of k ≤ n − 3
consecutive 0s surrounded by 1s. Also, none of the next vertices before reaching t is from V ′ , because all
contain the factor 0010, unlike any word in V ′ . If k = 1, then the vertex reached from v by following a
1-edge is either s ∈ V ′ (then we stop) or not in V ′ , as no other vertex from V ′ ends with 101. If it is not
in V ′ , then the next vertex reached via a 0-edge could be in V ′ , but all the subsequent vertices until (and
including) t are not, since they all contain the factor 0101, unlike any word in V ′ . This shows that none
of the edges traversed from v to s or t is from E ′ . Moreover, none of the vertices traversed between s and
z or between t and z is from V ′ , because they all contain the factor 0101 or 1011, unlike any word in V ′ ,
so we indeed reach z without using any edges from E ′ .
14
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
Now consider the case that we start at a vertex v that ends with 1. The only interesting case is v 6= s.
There are only two 0-edges in E ′ starting at a vertex that ends with 1, namely the edges starting at v1′ and
s. However, v is different from v1′ because v1′ is not part of G′ , and v is different from s by assumption.
We conclude that the 1-edge we follow is not from E ′ .
These arguments show that G′ is connected, so it has an Eulerian path, and this Eulerian path yields the
desired upword that begins with w. This completes the proof.
✷
5 Cyclic upwords
Throughout this section, all indices are considered modulo the size of the corresponding word. All the notions introduced in Section 2 can be extended straightforwardly to cyclic upwords, where factors are taken
cyclically across the word boundaries. In particular, when defining the graph H(u, n) for some cyclic upword = u1 u2 · · · uN we consider the subsets of words S(u, i, n) cyclically for all i = 1, 2, . . . , N . Then
the first two statements of Observation 3 hold for all vertices S(u, i, n), i = 1, 2, . . . , N . The next lemma
is the analogue of Lemma 4 for cyclic upwords.
Lemma 14 Let u = u1 u2 · · · uN be a cyclic upword for An , where A = {0, 1, . . . , α − 1} and n ≥ 2. If
uk = ✸ then uk+n = ✸.
Proof: Suppose that uk = ✸ and uk+n 6= ✸. By Observation 3, each vertex in the set S(u, k + 1, n)
has in-degree α in H(u, n), and each vertex in S(u, k, n) has out-degree 1. By Observation 2, for each
v = v1 v2 · · · vn ∈ S(u, k + 1, n) there are α − 1 other vertices (different from the ones in S(u, k + 1, n))
in GnA with the same set Γ− (v) of α many in-neighbors, namely vx := v1 · · · vn−1 x, where x ∈ A \ {vn }.
As the in-degree of every vertex of GnA is exactly α, and in H(u, n) all vertices already have in-degree at
least 1, it follows that the vertices vx can not be part of H(u, n), a contradiction to the fact that H(u, n)
is a spanning subgraph of GnA .
✷
Lemma 15 immediately yields the following corollary, which captures various rather severe conditions
that a cyclic upword must satisfy, relating its length N , the size α of the alphabet, and the value of the
parameter n.
Corollary 15 Let u = u1 u2 · · · uN be a cyclic upword for An , where A = {0, 1, . . . , α − 1} and n ≥ 2,
with at least one ✸. Then we have N = αn−d for some d, 1 ≤ d ≤ n − 1, such that n divides dN .
Proof: By Lemma 14, for any ✸ in u, the other two symbols in distance n from it must be ✸s as well.
Thus, the indices 1, 2, . . . , N are partitioned into gcd(n, N ) many residue classes modulo n, and all
symbols at positions from the same residue class are either all ✸s or all letters from A. Let d denote the
number of ✸s among any n consecutive symbols of u, then we have 1 ≤ d ≤ n − 1 (there is at least one
✸, but not all letters can be ✸s), and any starting position in u gives rise to αd different factors, implying
that N = αn−d . Furthermore, the d many ✸s within any n consecutive letters of u are partitioned into
n/ gcd(n, N ) many blocks with the same ✸ pattern, so n/ gcd(n, N ) must divide d, and this condition is
equivalent to n dividing d gcd(n, N ) and to n dividing dN .
✷
As an immediate corollary of our last result, we can exclude the existence of cyclic upwords for many
combinations of α and n.
On universal partial words
15
Corollary 16 Let A = {0, 1, . . . , α − 1} and n ≥ 2. If gcd(α, n) = 1, then there is no cyclic upword for
An . In particular, for α = 2 and odd n, there is no cyclic upword for An .
Proof: Suppose that such an upword u = u1 u2 · · · uN exists. Then by Corollary 15 we have N = αn−d
for some d, 1 ≤ d ≤ n − 1, such that n divides dN . However, as gcd(α, n) = 1, n does not divide
N = αn−d , so n must divide d, which is impossible, yielding a contradiction.
✷
By Corollaries 15 and 16, for a binary alphabet (α = 2), the only remaining potential parameter values
for cyclic upwords are n = 2 and d = 1, n = 4 and d ∈ {1, 2}, n = 6 and d = 3, n = 8 and
d ∈ {1, 2, . . . , 6}, n = 10 and d = 5, n = 12 and d ∈ {3, 6, 9}, etc. The case n = 2 and d = 1 can
be easily exluded: w.l.o.g. such a word has the form ✸0, leading to the factor 00 appearing twice (and 11
does not appear as a factor at all). However, for n = 4 and d = 1 we have the cyclic upword ✸001✸110,
which we already mentioned in the introduction. This is the only cyclic upword for a binary alphabet that
we know of. Cyclic upwords for any even alphabet size α ≥ 4 and n = 4 have been constructed in the
follow-up paper [GGH+ 16].
6 Outlook
In this paper we initiated the systematic study of universal partial words, and we hope that our results and
the numerous examples of upwords provided in the tables (see also the extensive data available on the
website [www]) generate substantial interest for other researchers to continue this exploration, possibly
in one of the directions suggested below.
Concerning the binary alphabet A = {0, 1}, it would be interesting to achieve complete classification
of linear upwords containing a single ✸, as suggested by Conjecture 8. For two ✸s such a task seems
somewhat more challenging (recall Table 2, Theorem 11 and see the data from [www]). Some examples
of binary linear upwords with three ✸s are listed in Table 3, and deriving some general existence and
non-existence results for this setting would certainly be of interest.
The next step would be to consider the situation of more than three ✸s present in a linear upword. The
following easy-to-verify example in this direction was communicated to us by Rachel Kirsch [GGH+ 16].
Theorem 17 For A = {0, 1} and any n ≥ 2, ✸n−1 01n is a linear upword for An with n − 1 many ✸s.
Complementing Theorem 17, we can prove the following non-existence result in this direction, but it
should be possible to obtain more general results.
Theorem 18 For A = {0, 1}, any n ≥ 4 and any 2 ≤ d ≤ n − 2, there is no linear upword for An that
begins with ✸d xd+1 xd+2 . . . xn+2 with xi ∈ A for all i = d + 1, . . . , n + 2.
The proof of Theorem 18 is easy by applying Lemma 4 to the first and second ✸. We leave the details
to the reader.
It would also be interesting to find examples of binary cyclic upwords other than ✸001✸110 for n = 4
mentioned before.
Finally, a natural direction would be to search for (linear or cyclic) upwords for non-binary alphabets,
but we anticipate that no non-trivial upwords exist in most cases (recall Theorem 5). As evidence for this
we have the following general non-existence result in this setting.
Theorem 19 For A = {0, 1, . . . , α − 1}, α ≥ 3, and any d ≥ 2, for large enough n there is no linear or
cyclic upword for An with exactly d many ✸s.
16
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
n
3
4
5
✸✸✸
✸✸✸01111 (Thm. 17)
✸✸001✸11010
✸001✸110✸00
0✸001✸110✸0
✸0010✸0111✸10011011000001
✸0000111✸10001001101100101✸1
✸00001110✸100010100110101111✸
✸0000100111✸10001101100101✸1
✸0000101110✸1000110101001111✸
✸00001111101✸10001011001✸01
✸000010101110✸10001101001111✸
✸0000101001110✸1000110101111✸
✸00001101100111✸1000100101✸1
✸0000110101001110✸1000101111✸
✸00001101100100111✸1000101✸1
✸000010010101111100✸1101✸00
0✸1100✸001111101101000101✸1
Tab. 3: Examples of linear upwords for An , A = {0, 1}, with three ✸s for n = 3, 4, 5.
Theorem 19 shows in particular that for a fixed alphabet size α and a fixed number d ≥ 2 of diamonds,
there are only finitely many possible candidates for upwords with d diamonds (which in principle could all
be checked by exhaustive search). The proof idea is that for fixed d and large enough n, such an upword
must contain a ✸ and a symbol from A in distance n, and then applying Lemma 4 or Lemma 14 yields a
contradiction (recall the proof of Theorem 5). We omit the details here. On the positive side, upwords for
even alphabet sizes α ≥ 4 and n = 4 have been constructed in [GGH+ 16] (and these upwords are even
cyclic).
A question that we have not touched in this paper is the algorithmic problem of efficiently generating
upwords. As a preliminary observation in this direction we remark here that some of the linear upwords
constructed in Theorem 9 and 10 can also be obtained by straightforward modifications of the FKM
de Bruijn sequences constructed in [FM78, FK86], for which efficient generation algorithms are known
[RSW92].
Acknowledgements
The authors thank Martin Gerlach for his assistance in our computer searches, Rachel Kirsch and her
collaborators [GGH+ 16], as well as Artem Pyatkin for providing particular examples of small upwords.
The second author is grateful to Sergey Avgustinovich for helpful discussions on universal partial words,
and to Bill Chen and Arthur Yang for their hospitality during the author’s visit of the Center for Combinatorics at Nankai University in November 2015. This work was supported by the 973 Project, the PCSIRT
Project of the Ministry of Education and the National Science Foundation of China. We also thank the
On universal partial words
17
anonymous referees of this paper for several valuable suggestions and references that helped improving
the presentation.
References
[BB99]
J. Berstel and L. Boasson. Partial words and a theorem of Fine and Wilf. Theoret. Comput.
Sci., 218(1):135–141, 1999. WORDS (Rouen, 1997).
[BBSGR10] B. Blakeley, F. Blanchet-Sadri, J. Gunter, and N. Rampersad. On the complexity of deciding
avoidability of sets of partial words. Theoret. Comput. Sci., 411(49):4263–4271, 2010.
[BJG08]
J. Bang-Jensen and G. Z. Gutin. Digraphs: Theory, Algorithms and Applications. Springer,
2nd edition, 2008.
[BS05]
F. Blanchet-Sadri. Primitive partial words. Discrete Applied Mathematics, 148(3):195–213,
2005.
[BS08]
F. Blanchet-Sadri. Open problems on partial words. In G. Bel-Enguix, M. D. JiménezLópez, and C. Martín-Vide, editors, New Developments in Formal Languages and Applications, pages 11–58. Springer Berlin Heidelberg, Berlin, Heidelberg, 2008.
[BSBK+ 09] F. Blanchet-Sadri, N. C. Brownstein, A. Kalcic, J. Palumbo, and T. Weyand. Unavoidable
sets of partial words. Theory Comput. Syst., 45(2):381–406, 2009.
[CDG92]
F. Chung, P. Diaconis, and R. Graham. Universal cycles for combinatorial structures. Discrete Math., 110(1-3):43–59, 1992.
[CJJK11]
A. Claesson, V. Jelínek, E. Jelínková, and S. Kitaev. Pattern avoidance in partial permutations. Electron. J. Combin., 18(1):Paper 25, 41, 2011.
[CPT11]
P. E. C. Compeau, P. A. Pevzner, and G. Tesler. How to apply De Bruijn graphs to genome
assembly. Nature biotechnology, 29(11):987–991, 2011.
[dB46]
N. G. de Bruijn. A Combinatorial Problem. Koninklijke Nederlandse Akademie v. Wetenschappen, 49:758–764, 1946.
[Eul36]
L. Euler. Solutio problematis ad geometriam situs pertinentis. Comment. Academiae Sci. I.
Petropolitanae, 8:128–140, 1736.
[FK86]
H. Fredricksen and I. J. Kessler. An algorithm for generating necklaces of beads in two
colors. Discrete Math., 61(2-3):181–188, 1986.
[Fle83]
M. Fleury. Deux problemes de geometrie de situation. Journal de mathematiques elementaires, pages 257–261, 1883.
[FM78]
H. Fredricksen and J. Maiorana. Necklaces of beads in k colors and k-ary de Bruijn sequences. Discrete Math., 23(3):207–210, 1978.
18
Herman Z. Q. Chen, Sergey Kitaev, Torsten Mütze, Brian Y. Sun
[GGH+ 16] B. Goeckner, C. Groothuis, C. Hettle, B. Kell, P. Kirkpatrick, R. Kirsch, and R. Solava.
Universal partial words over non-binary alphabets. arXiv:1611.03928, Nov 2016.
[HHK08]
V. Halava, T. Harju, and T. Kärki. Square-free partial words. Inform. Process. Lett.,
108(5):290–292, 2008.
[HHKS09]
V. Halava, T. Harju, T. Kärki, and P. Séébold. Overlap-freeness in infinite partial words.
Theoret. Comput. Sci., 410(8-10):943–948, 2009.
[HRW12]
A. E. Holroyd, F. Ruskey, and A. Williams. Shorthand universal cycles for permutations.
Algorithmica, 64(2):215–245, 2012.
[Hur90]
G. H. Hurlbert. Universal cycles: On beyond de Bruijn. ProQuest LLC, Ann Arbor, MI,
1990. Thesis (Ph.D.)–Rutgers The State University of New Jersey - New Brunswick.
[HW73]
C. Hierholzer and C. Wiener. Ueber die Möglichkeit, einen Linienzug ohne Wiederholung
und ohne Unterbrechung zu umfahren. Math. Ann., 6(1):30–32, 1873.
[Lem71]
A. Lempel. m-ary closed sequences. J. Combinatorial Theory Ser. A, 10:253–258, 1971.
[PSCF05]
J. Pagès, J. Salvi, C. Collewet, and J. Forest. Optimised De Bruijn patterns for one-shot
shape acquisition. Image and Vision Computing, 23(8):707 – 720, 2005.
[Ral82]
A. Ralston. De Bruijn sequences—a model example of the interaction of discrete mathematics and computer science. Math. Mag., 55(3):131–143, 1982.
[RSW92]
F. Ruskey, C. Savage, and T. M. Y. Wang. Generating necklaces. J. Algorithms, 13(3):414–
430, 1992.
[SBSH97]
H. Sohn, D. L. Bricker, J. R. Simon, and Y. Hsieh. Optimal sequences of trials for balancing
practice and repetition effects. Behavior Research Methods, Instruments, & Computers,
29(4):574–581, 1997.
[Sch01]
E. R. Scheinerman. Determining planar location via complement-free De Brujin sequences
using discrete optical sensors. IEEE Transactions on Robotics and Automation, 17(6):883–
889, Dec 2001.
[SW14]
B. Stevens and A. Williams. The coolest way to generate binary strings. Theory Comput.
Syst., 54(4):551–577, 2014.
[www]
currently http://www.math.tu-berlin.de/~muetze.
[Yoe62]
M. Yoeli. Binary ring sequences. Amer. Math. Monthly, 69:852–855, 1962.
| 7cs.IT
|
Online Stochastic Matching: New Algorithms and Bounds∗
Brian Brubach†, Karthik A. Sankararaman‡, Aravind Srinivasan§, and Pan Xu¶
arXiv:1606.06395v3 [cs.DS] 15 Nov 2017
Department of Computer Science, University of Maryland, College Park, MD 20742, USA
Original Version: May 2016
This Version: November 2017
Abstract
Online matching has received significant attention over the last 15 years due to its close
connection to Internet advertising. As the seminal work of Karp, Vazirani, and Vazirani has an
optimal (1−1/e) competitive ratio in the standard adversarial online model, much effort has gone
into developing useful online models that incorporate some stochasticity in the arrival process.
One such popular model is the “known I.I.D. model” where different customer-types arrive
online from a known distribution. We develop algorithms with improved competitive ratios for
some basic variants of this model with integral arrival rates, including: (a) the case of general
weighted edges, where we improve the best-known ratio of 0.667 due to Haeupler, Mirrokni and
Zadimoghaddam [12] to 0.705; and (b) the vertex-weighted case, where we improve the 0.7250
ratio of Jaillet and Lu [13] to 0.7299.
We also consider an extension of stochastic rewards, a variant where each edge has an
independent probability of being present. For the setting of stochastic rewards with non-integral
arrival rates, we present a simple optimal non-adaptive algorithm with a ratio of 1 − 1/e. For
the special case where each edge is unweighted and has a uniform constant probability of being
present, we improve upon 1 − 1/e by proposing a strengthened LP benchmark.
One of the key ingredients of our improvement is the following (offline) approach to bipartitematching polytopes with additional constraints. We first add several valid constraints in order
to get a good fractional solution f; however, these give us less control over the structure of
f. We next remove all these additional constraints and randomly move from f to a feasible
point on the matching polytope with all coordinates being from the set {0, 1/k, 2/k, . . . , 1} for
a chosen integer k. The structure of this solution is inspired by Jaillet and Lu (Mathematics of
Operations Research, 2013) and is a tractable structure for algorithm design and analysis. The
appropriate random move preserves many of the removed constraints (approximately with high
probability and exactly in expectation). This underlies some of our improvements and could be
of independent interest.
∗
A preliminary version of this appeared in the European Symposium on Algorithms (ESA), 2016
Email: [email protected]
Supported in part by NSF Awards CNS 1010789 and CCF 1422569.
‡
Email: [email protected]
Supported in part by NSF Awards CNS 1010789 and CCF 1422569.
§
Email: [email protected]
Supported in part by NSF Awards CNS 1010789 and CCF 1422569, and by research awards from Adobe, Inc.
¶
Email: [email protected]
Supported in part by NSF Awards CNS 1010789 and CCF 1422569.
†
1
1
Introduction.
Applications to Internet advertising have driven the study of online matching problems in recent
years [20]. In these problems, we consider a bipartite graph G = (U, V, E) in which the set of
vertices U is available offline while the set of vertices in V arrive online. Whenever some vertex
v arrives, it must be matched immediately (and irrevocably) to (at most) one vertex in U . Each
offline vertex u can be matched to at most one v. In the context of Internet advertising, U is the
set of advertisers and V is the set of impressions. The edges E define the impressions that interest
a particular advertiser. When an impression v arrives, we must choose an available advertiser (if
any) to match with it. WLOG we consider the case where v ∈ V can be matched at most once.
Since advertising forms the key source of revenue for many large Internet companies, finding good
matching algorithms and obtaining even small performance gains can have high impact.
In the stochastic known I.I.D. model of arrival, we are given a bipartite graph G = (U, V, E)
and a finite online time horizon T (here we assume T = n). In each round, a vertex v is sampled
with replacement from a known distribution over V . The sampling distributions are independent
and identical over all of the T online rounds. This captures the fact that we often have historical
data about the impressions and can predict the frequency with which each type of impression
will arrive. Edge-weighted matching [9] is a general model in the context of advertising: every
advertiser gains a given revenue for being matched to a particular type of impression. Here, a
type of impression refers to a class of users (e.g., a demographic group) who are interested in the
same subset of advertisements. A special case of this model is vertex-weighted matching [1], where
weights are associated only with the advertisers. In other words, a given advertiser has the same
revenue generated for matching any of the user types interested in it. In some modern business
models, revenue is not generated upon matching advertisements, but only when a user clicks on the
advertisement: this is the pay-per-click model. From historical data, one can assign the probability
of a particular advertisement being clicked by a type of user. Works including [21, 22] capture this
notion of stochastic rewards by assigning a probability to each edge.
One unifying theme in most of our approaches is the use of an LP benchmark with additional
valid constraints that hold for the respective stochastic-arrival models. We use the optimal solution
of this LP to guide our online actions. To do so, we use various modifications of dependent
randomized rounding.
2
Preliminaries and technical challenges.
In the Unweighted Online Known I.I.D. Stochastic Bipartite Matching problem, we are given a
bipartite graph G = (U, V, E). The set U is available offline while the vertices v arrive online
and are drawn with replacement from an I.I.D. distribution on V . For each v ∈ V , we are given
an arrival rate rv , which is the expected number of times v will arrive. With the exception of
Sections 5, this paper will focus on the integral arrival rates setting where all rv ∈ Z+ . For reasons
described in [12], we can further assume WLOG that each v has rv = 1 under the assumption of
integral arrival rates. In this case, we have that |V | = n where n is the total number of online
rounds.
In the vertex-weighted variant, every vertex u ∈ U has a weight wu and we seek a maximum
weight matching. In the edge-weighted variant, every edge e ∈ E has a weight we and we again
seek a maximum weight matching. In the stochastic rewards variant, each edge has both a
2
weight we and a probability pe of being present once we probe edge e1 and we seek to maximize
the expected weight of the matching.
Asymptotic assumption and notation. We will always assume n is large and analyze
algorithms as n goes to infinity: e.g., if x ≤ 1 − (1 − 2/n)n , we will just write this as “x ≤ 1 − 1/e2 ”
instead of the more-accurate “x ≤ 1 − 1/e2 + o(1)”. These suppressed o(1) terms will subtract
at most o(1) from our competitive ratios. Another fact to note is that the competitive ratio is
defined slightly differently than usual for this set of problems (Similar to notation used in [20]). In
E[ALG]
particular, it is defined as E[OPT]
. Algorithms can be adaptive or non-adaptive. When v arrives,
an adaptive algorithm can modify its online actions based on the realization of the online vertices
thus far, but a non-adaptive algorithm has to specify all of its actions before the start of the online
phase. Throughout, we use “WS” to refer to the worst case instance for various algorithms.
2.1
LP benchmark for deterministic rewards.
As in prior work (e.g, see [20]), we use the following LP to upper bound the optimal offline expected
performance and also use it to guide our algorithm in the case where rewards are deterministic.
For the case of stochastic rewards, we use slightly modified LPs, whose definitions we defer until
Sections 5 and 6. We first show an LP for the unweighted variant, then describe the changes for
the vertex-weighted and edge-weighted settings. As usual, we have a variable fe for each edge. Let
P
∂(w) be the set of edges adjacent to a vertex w ∈ U ∪ V and let fw = e∈∂(w) fe .
maximize
X
fe
(1)
e∈E
subject to
X
fe ≤ 1
∀u ∈ U
(2)
X
fe ≤ 1
∀v ∈ V
(3)
∀e ∈ E
(4)
e∈∂(u)
e∈∂(v)
0 ≤ fe ≤ 1 − 1/e
2
′
fe + fe′ ≤ 1 − 1/e
∀e, e ∈ ∂(u), ∀u ∈ U
(5)
Variants: The objective function is to maximize u∈U e∈∂(u) fe wu in the vertex-weighted
P
variant and maximize e∈E fe we in the edge-weighted variant (here we refers to w(u,v) ).
Constraint 2 is the matching constraint for vertices in U . Constraint 3 is valid because each
vertex in V has an arrival rate of 1. Constraint 4 is used in [19] and [12]. It captures the fact that
the expected number of matches for any edge is at most 1 − 1/e. This is valid for large n because
the probability that a given vertex does not arrive during n rounds is 1/e. Constraint 5 is similar
to the previous one, but for pairs of edges. For any two neighbors of a given u ∈ U , the probability
that neither of them arrive is 1/e2 . Therefore, the sum of variables for any two distinct edges in
∂(u) cannot exceed 1 − 1/e2 . Notice that constraints 4 and 5 reduces the gap between the optimal
LP solution and the performance of the optimal online algorithm. In fact, without constraint 4, we
cannot in general achieve a competitive ratio better than 1 − 1/e.
P
1
P
The edge realization process is independent for different edges. At each step, the algorithm “probes” an edge.
With probability pe the edge e exists and with the remaining probability it does not. Once realization of an edge is
determined, it does not affect the random realizations for the rest of the edges.
3
Note that the work of [19] does not use an LP to upper-bound the optimal value of the offline
instance. Instead they use Monte-Carlo simulations wherein they simulate the arrival sequence and
compute the vector f by approximating (via Monte-Carlo simulation) the probability of matching
an edge e in the offline optimal solution. We do not use a similar approach for our problems for
a few reasons. (1) For the weighted variants, namely the edge and vertex-weighted versions, the
number of samples depends on the maximum value of the weight, making it expensive. (2) In the
unweighted version, the running time of the sampling based algorithm is O(|E|2 n4 ); on the other
hand, we show in Section 2.6 that the LP based algorithm can be solved much faster, Õ(|E|2 ) time
in the worst case and even faster than that in practice. (3) For the stochastic rewards setting,
the offline problem is not known to be polynomial-time solvable. The paper [4] shows under the
assumption of constant p and OPT = ω(1/p), they can obtain a (1 − ǫ) approximation to the
optimal solution. However, these assumptions are too strong to be used in our setting.
Finally, for the stochastic rewards setting, one might be tempted to use an LP to achieve the
same property obtained from Monte-Carlo simulation via adding extra constraints. In the context
of uniform stochastic rewards where each edge e is associated with a uniform constant probability
p, what we really need is:
X
1 − exp(−|S|p)
fe ≤
∀S ⊆ ∂(u),
(6)
p
e∈S
To guarantee this via the LP, a straightforward approach is to add this family of constraints into
the LP. However, the number of such constraints is exponential and there seems to be no obvious
separation oracle. We overcome this challenge by showing it suffices to ensure that Inequality (6)
above holds for all S with |S| ≤ 2/p, which is a constant and thus the resultant LP is polynomial
solvable.
2.2
Overview of vertex-weighted algorithm and contributions.
A key challenge encountered by [13] was that their special LP could lead to length four cycles of
type C1 shown in Figure 1. In fact, they used this cycle to show that no algorithm could perform
better than 1 − 2/e2 ≈ 0.7293 using their LP. They mentioned that tighter LP constraints such as 4
and 5 in the LP from Section 2 could avoid this bottleneck, but they did not propose a technique
to use them. Note that the {0, 1/3, 2/3} solution produced by their LP was an essential component
of their Random List algorithm.
We show a randomized rounding algorithm to construct a similar, simplified {0, 1/3, 2/3} vector
from the solution of a stricter benchmark LP. This allows for the inclusion of additional constraints,
most importantly constraint 5. Using this rounding algorithm combined with tighter constraints, we
will upper bound the probability of a vertex appearing in the cycle C1 from Figure 1 at 2−3/e ≈ 0.89.
(See Lemma 4) Additionally, we show how to deterministically break all other length four cycles
which are not of type C1 without creating any new cycles of type C1 . Finally, we describe an
algorithm which utilizes these techniques to improve previous results in both the vertex-weighted
and unweighted settings.
For this algorithm, we first solve the LP in Section 2 on the input graph. In Section 4, we
show how to use the technique in sub-section 2.7 to obtain a sparse fractional vector. We then
present a randomized online algorithm (similar to the one in [13]) which uses the sparse fractional
vector as a guide to achieve a competitive ratio of 0.7299. Previously, there was gap between the
best unweighted algorithm with a ratio of 1 − 2e−2 due to [13] and the negative result of 1 − e−2
4
Figure 1: This cycle is the source of the negative result described by Jaillet and Lu [13]. Thick
edges have fe = 2/3 while thin edges have fe = 1/3.
(C1 )
u1
v1
u2
v2
due to [19]. We take a step towards closing that gap by showing that an algorithm can achieve
0.7299 > 1 − 2e−2 for both the unweighted and vertex-weighted variants with integral arrival rates.
In doing so, we make progess on Open Questions 3 and 4 from the book [20]. 2
2.3
Overview of edge-weighted algorithm and contributions.
A challenge that arises in applying the power of two choices to this setting is when the same edge
(u, v) is included in both matchings M1 and M2 . In this case, the copy of (u, v) in M2 can offer no
benefit and a second arrival of v is wasted. To use an example from related work, Haeupler et al. [12]
choose two matchings in the following way. M1 is attained by solving an LP with constraints 2, 3
and 4 and rounding to an integral solution. M2 is constructed by finding a maximum weight
matching and removing any edges which have already been included in M1 . A key element of their
proof is showing that the probability of an edge being removed from M2 is at most 1 − 1/e ≈ 0.63.
The approach in this paper is to construct two or three matchings together in a correlated
manner to reduce the probability that some edge is included in all matchings. We show a general
technique to construct an ordered set of k matchings where k is an easily adjustable parameter.
For k = 2, we show that the probability of an edge appearing in both M1 and M2 is at most
1 − 2/e ≈ 0.26.
For the algorithms presented, we first solve an LP on the input graph. We then round the LP
solution vector to a sparse integral vector and use this vector to construct a randomly ordered set of
matchings which will guide our algorithm during the online phase. We begin Section 3 with a simple
warm-up algorithm which uses a set of two matchings as a guide to achieve a 0.688 competitive
ratio, improving the best known result for this problem. We follow it up with a slight variation
that improves the ratio to 0.7 and a more complex 0.705-competitive algorithm which relies on a
convex combination of a 3-matching algorithm and a separate pseudo-matching algorithm.
2.4
Overview of stochastic rewards and contributions.
This algorithm (Algorithm 9) is presented in Section 5 and 6. We believe the known I.I.D. model
with stochastic rewards is an interesting new direction motivated by the work of [21] and [22] in the
adversarial model. We introduce a new, more general LP (see LP (9)) specifically for this setting
and show that a simple algorithm using the LP solution directly can achieve a competitive ratio
of 1 − 1/e, which is proved to be optimal among all non-adaptive algorithms. In [22], it is shown
that no randomized algorithm can achieve a ratio better than 0.62 < 1 − 1/e in the adversarial
2
Open Questions 3 and 4 state the following: “In general, close the gap between the upper and lower bounds. In
some sense, the ratio of 1 − 2e−2 achieved in [13] for the integral case, is a nice ‘round’ number, and one may suspect
that it is the correct answer.”
5
Table 1: Summary of Contributions
Problem
Previous Work
This Paper
Edge-Weighted (Section 3)
0.667 [12]
0.705
Vertex-Weighted (Section 4)
0.725 [13]
0.7299
Unweighted
1 − 2/e2 [13]
0.7299 (> 1 − 2/e2 )
Stochastic Rewards (Section 5 and 6)
N/A
1 − e−1 for general version
0.702 for the restricted version
model. Hence, achieving a 1 − 1/e for the i.i.d. model shows that this lower bound does not extend
to this model. Further, the paper [5] shows that using LP (9) one cannot achieve a ratio better
than 1 − 1/e. We discuss some challenges relating to why the techniques used in prior work do not
directly extend to this model. Finally, we consider a restricted version of the problem where each
edge is unweighted and has a uniform constant probability p ∈ (0, 1] under integral arrival rates.
By proposing a family of valid constraints, we are able to show that in this restricted setting, one
can indeed beat 1 − 1/e.
2.5
Summary of our contributions
Theorem 1. For vertex-weighted online stochastic matching with integral arrival rates, online
algorithm VW achieves a competitive ratio of at least 0.7299.
Theorem 2. For edge-weighted online stochastic matching with integral arrival rates, there exists
an algorithm which achieves a competitive ratio of at least 0.7 and algorithm EW[q] with q = 0.149251
achieves a competitive ratio of at least 0.70546.
Theorem 3. For edge-weighted online stochastic matching with arbitrary arrival rates and stochastic rewards, online algorithm SM (9) achieves a competitive ratio of 1 − 1/e, which is optimal all
among all non-adaptive algorithms.
Theorem 4. For unweighted online stochastic matching with integral arrival rates and uniform
stochastic rewards, there exists an adaptive algorithm which achieves a competitive ratio of at least
0.702.
2.6
Runtime of algorithm.
In this section, we discuss the implementation details of our algorithm. All of our algorithms solve
an LP in the pre-processing step. The dimension of the LP is determined by the constraint matrix
which consists of O(|E| + |U | + |V |) rows and O(|E|) columns. However, note that the number of
non-zero entries in this matrix is of the order O((|U | + |V |)|E|). Some recent work (e.g., [17]) shows
that such sparse programs can be solved in time Õ(|E|2 ) using interior point methods (which are
6
known to perform very well in practice). This sparsity in the LP is the reason we can solve very
large instances of the problem. The second critical step in pre-processing is to perform randomized
rounding. Note that we have O(|E|) variables and in each step of the randomized rounding due to
[11], they incur a running time of O(|E|). Hence the total running time to obtain a rounded solution
is of the order O(|E|2 ). Finally, recall that both these operations are part of the pre-processing step.
Hence in the online phase the algorithm incurs a per-time-step running time of at most O(|U |) for
the stochastic rewards case (in fact, a smarter implementation using binary search runs as fast as
O(log |U |)) and O(1) for the edge-weighted and the vertex-weighted algorithms in Section 3 and 4.
2.7
LP rounding technique DR[f, k].
For the algorithms presented, we first solve the benchmark LP in sub-section 2.1 for the input
instance to get a fractional solution vector f. We then round f to an integral solution F using a
two step process we call DR[f, k]. The first step is to multiply f by k. The second step is to apply
the dependent rounding techniques of Gandhi, Khuller, Parthasarathy, and Srinivasan [11] to this
new vector. In this paper, we will always choose k to be 2 or 3. This is because a vertex in V may
appear more than once, but probably not more than two or three times.
While dependent rounding is typically applied to values between 0 and 1, the useful properties
extend naturally to our case in which kfe may be greater than 1 for some edge e. To understand this
process, it is easiest to imagine splitting each kfe into two edges with the integer value fe′ = ⌊kfe ⌋
and fractional value fe′′ = kfe − ⌊kfe ⌋. The former will remain unchanged by the dependent
rounding since it is already an integer while the latter will be rounded to 1 with probability fe′′ and
0 otherwise. Our final value Fe would be the sum of those two rounded values. The two properties
of dependent rounding we use are:
1. Marginal distribution: For every edge e, let pe = kfe − ⌊kfe ⌋. Then, Pr[Fe = ⌈kfe ⌉] = pe
and Pr[Fe = ⌊kfe ⌋] = 1 − pe .
2. Degree-preservation:
For any vertex w ∈ U ∪ V , let its fractional degree kfw be
P
P
e∈∂(w) kfe and integral degree be the random variable Fw =
e∈∂(w) Fe . Then Fw ∈
{⌊kfw ⌋, ⌈kfw ⌉}.
2.8
Related work.
The study of online matching began with the seminal work of Karp, Vazirani, Vazirani [14], where
they gave an optimal online algorithm for a version of the unweighted bipartite matching problem
in which vertices arrive in adversarial order. Following that, a series of works have studied various
related models. The book by Mehta [20] gives a detailed overview. The vertex-weighted version
of
this problem was introduced by Aggarwal, Goel and Karande [1], where they give an optimal
1 − 1e ratio for the adversarial arrival model. The edge-weighted setting has been studied in the
adversarial model by Feldman, Korula, Mirrokni and Muthukrishnan [9], where they consider an
additional relaxation of “free-disposal”.
In addition to the adversarial and known I.I.D. models, online matching is also studied under
several other variants such as random arrival order, unknown distributions, and known adversarial
distributions. In the setting of random arrival order, the arrival sequence is assumed to be a random
permutation over all online vertices, see e.g., [6, 15, 16, 18]. In the case of unknown distributions, in
each round an item is sampled from a fixed but unknown distribution. If the sampling distributions
7
are required to be the same during each round, it is called unknown I.I.D. ([7, 8]); otherwise, it is
called adversarial stochastic input ([7]). As for known adversarial distributions, in each round an
item is sampled from a known distribution, which is allowed to change over time ([2, 3]). Another
variant of this problem is when the edges have stochastic rewards. Models with stochastic rewards
have been previously studied by [21, 22] among others, but not in the known I.I.D. model.
Related Work in the Vertex-Weighted and Unweighted Settings. The vertex-weighted
and unweighted settings have many results starting with Feldman, Mehta, Mirrokni and Muthukrishnan [10] who were the first to beat 1 − 1/e with a competitive ratio of 0.67 for the unweighted
problem. This was improved by Manshadi, Gharan, and Saberi [19] to 0.705 with an adaptive
algorithm. In addition, they showed that even in the unweighted variant with integral arrival
rates, no algorithm can achieve a ratio better than 1 − e−2 ≈ 0.86. Finally, Jaillet and Lu [13]
presented an adaptive algorithm which used a clever LP to achieve 0.725 and 1 − 2e−2 ≈ 0.729 for
the vertex-weighted and unweighted problems, respectively.
Related Work in the Edge-Weighted Setting. For this model, Haeupler, Mirrokni, Zadimoghaddam [12] were the first to beat 1 − 1/e by achieving a competitive ratio of 0.667. They use
a discounted LP with tighter constraints than the basic matching LP (a similar LP can be seen in
2.1) and they employ the power of two choices by constructing two matchings offline to guide their
online algorithm.
√Other Related Work. Devanur et al [8] gave an algorithm which achieves a ratio of 1 −
1/ 2πk for the Adwords problem in the Unknown I.I.D. arrival model with knowledge of the
optimal budget utilization and when the bid-to-budget ratios are at most 1/k. Alaei et al. [2]
considered the Prophet-Inequality Matching problem,
√ in which v arrives from a distinct (known)
distribution Dt , in each round t. They gave a 1 − 1/ k + 3 competitive algorithm, where k is the
minimum capacity of u.
3
Edge-weighted matching with integral arrival rates
3.1
A simple 0.688-competitive algorithm.
As a warm-up, we describe a simple algorithm which achieves a competitive ratio of 0.688 and
introduces key ideas in our approach. We begin by solving the LP in sub-section 2.1 to get a
fractional solution vector f and applying DR[f, 2] as described in Subsection 2.7 to get an integral
vector F. We construct a bipartite graph GF with Fe copies of each edge e. Note that GF will
have max degree 2 since for all w ∈ U ∪ V , Fw ≤ ⌈2fw ⌉ ≤ 2 and thus we can decompose it into
two matchings using Hall’s Theorem. Finally, we randomly permute the two matchings into an
ordered pair of matchings, [M1 , M2 ]. These matchings serve as a guide for the online phase of the
algorithm, similar to [12]. The entire warm-up algorithm for the edge-weighted model, denoted by
EW0 , is summarized in Algorithm 1.
3.1.1
Analysis of algorithm EW0 .
We will show that EW0 (Algorithm 1) achieves a competitive ratio of 0.688. Let [M1 , M2 ] be our
randomly ordered pair of matchings. Note that there might exist some edge e which appears in
8
Algorithm 1: [EW0 ]
1
2
3
4
5
6
Construct and solve the benchmark LP in sub-section 2.1 for the input instance.
Let f be an optimal fractional solution vector. Call DR[f, 2] to get an integral vector F.
Create the graph GF with Fe copies of each edge e ∈ E and decompose it into two
matchings.
Randomly permute the matchings to get a random ordered pair of matchings, say
[M1 , M2 ].
When a vertex v arrives for the first time, try to assign v to some u1 if (u1 , v) ∈ M1 ;
when v arrives for the second time, try to assign v to some u2 if (u2 , v) ∈ M2 .
When a vertex v arrives for the third time or more, do nothing in that step.
both matchings if and only if fe > 1/2. Therefore, we consider three types of edges. We say an
edge e is of type ψ1 , denoted by e ∈ ψ1 , if and only if e appears only in M1 . Similarly e ∈ ψ2 , if
and only if e appears only in M2 . Finally, e ∈ ψb , if and only if e appears in both M1 and M2 . Let
P1 , P2 , and Pb be the probabilities of getting matched for e ∈ ψ1 , e ∈ ψ2 , and e ∈ ψb , respectively.
According to the result in Haeupler et al. [12], Lemma 1 bounds these probabilities.
Lemma 1. (Proof details in Section 3 of [12]) Given M1 and M2 , in the worst case (1) P1 = 0.5808;
(2) P2 = 0.14849 and (3) Pb = 0.632.
We can use Lemma 1 to prove that the warm-up algorithm EW0 achieves a ratio of 0.688 by
examining the probability that a given edge becomes type ψ1 , ψ2 , or ψb .
Proof. (Analysis for EW0 )
Consider the following two cases.
• Case 1: 0 ≤ fe ≤ 1/2: By the marginal distribution property of dependent rounding,
there can be at most one copy of e in GF and the probability of including e in GF is 2fe .
Since an edge in GF can appear in either M1 or M2 with equal probability 1/2, we have
Pr[e ∈ ψ1 ] = Pr[e ∈ ψ2 ] = fe . Thus, the ratio is (fe P1 + fe P2 )/fe = P1 + P2 = 0.729.
• Case 2: 1/2 ≤ fe ≤ 1 − 1/e: Similarly, by marginal distribution, Pr[e ∈ ψb ] = Pr[Fe =
⌈2fe ⌉] = 2fe −⌊2fe ⌋ = 2fe −1. It follows that Pr[e ∈ ψ1 ] = Pr[e ∈ ψ2 ] = (1/2)(1−(2fe −1)) =
1 − fe . Thus, the ratio is (noting that the first term is from case 1 while the second term is
from case 2) ((1 − fe )(P1 + P2 ) + (2fe − 1)Pb )/fe ≥ 0.688, where the WS is for an edge e with
fe = 1 − 1/e.
3.2
A 0.7-competitive algorithm.
In this section, we describe an improvement upon the previous warm-up algorithm to get a competitive ratio of 0.7. We start by making an observation about the performance of the warm-up
algorithm. After solving the LP, let edges with fe > 1/2 be called large and edges with fe ≤ 1/2
be called small. Let L and S, be the sets of large and small edges, respectively. Notice that in the
previous analysis, small edges achieved a much higher competitive ratio of 0.729 versus 0.688 for
9
large edges. This is primarily due to the fact that we may get two copies of a large edge in GF . In
this case, the copy in M1 has a better chance of being matched, since there is no edge which can
“block” it (i.e. an edge with the same offline neighbor that gets matched first), but the copy that
is in M2 has no chance of being matched.
To correct this imbalance, we make an additional modification to the fe values before applying
DR[f, k]. The rest of the algorithm is exactly the same. Let η be a parameter to be optimized later.
For all large edges ℓ ∈ L such that fℓ > 1/2, we set fℓ = fℓ + η. For all small edges s ∈ S which are
adjacent to some large edge, let ℓ ∈ L be the largest edge adjacent to s such that fℓ > 1/2. Note
that it is possiblefor s to have two large neighbors, but we only care about the largest one. We set
ℓ +η)
.
fs = fs 1−(f
1−fℓ
In other words, we increase the values of large edges while ensuring that for all w ∈ U ∪ V ,
fw ≤ 1 by reducing the values of neighboring small edges proportional to their original values. Note
that it is not possible for two large edges to be adjacent since they must both have fe > 1/2. For
all other small edges which are not adjacent to any large edges, we leave their values unchanged.
We then apply DR[f, 2] to this new vector, multiplying by 2 and applying dependent rounding as
before.
3.2.1
Analysis.
We can now prove Theorem 2.
Proof. As in the warm-up analysis, we’ll consider large and small edges separately
• 0 ≤ fs ≤ 21 : Here we have two cases
– Case 1: s is not adjacent to any large edges.
In this case, the analysis is the same as the warm-up algorithm and we still get a 0.729
competitive ratio for these edges.
– Case 2: s is adjacent to some large edge ℓ.
For this case, let fℓ be the value of the largest neighboring edge in the original LP
solution. Then s achieves a ratio of
fs
1 − (fℓ + η)
(0.1484 + 0.5803)/fs =
1 − fℓ
1 − (fℓ + η)
(0.1484 + 0.5803)
1 − fℓ
This follows from Lemma 1; in particular, the first two terms are the result of how we
set fs in the algorithm, while the two numbers, 0.1484 and 0.5803, are the probabilities
that s is matched when it is in M2 and M1 , respectively. Note that for fℓ ∈ [0, 1) this
is a decreasing function with respect to fℓ . So the worst case is fℓ = 1 − 1/e (due to
constraint 4 in LP 2.1) and we have a ratio of
•
1
2
1 − (1 − 1/e + η)
(0.1484 + 0.5803) =
1 − (1 − 1/e)
< fℓ ≤ 1 − 1e :
1/e − η
(0.1484 + 0.5803)
1/e
Here, the ratio is ((1 − (fℓ + η))(P1 + P2 ) + (2(fℓ + η) − 1)Pb )/fℓ , where the WS is for an
edge e with fℓ = 1 − 1/e. This follows because of the fact that it is a decreasing function
10
with respect to fℓ . To see that it is a decreasing function, note that it can be rearranged as
(P1 + P2 − Pb + η(2Pb − P1 − P2 ) + fℓ (2Pb − P1 − P2 ))/fℓ . Substituting the appropriate values,
we have that the value of (2Pb − P1 − P2 ) = 0.535. Hence, the expression can be written as
(c1 + 0.535fℓ )/fℓ . If we show that c1 ≥ 0, we are done. The optimal value of η we choose
turns out to be 0.0142. Hence we have c1 = P1 + P2 − Pb + 0.0142 · 0.535 = 0.1048 > 0.
Choosing the optimal value of η = 0.0142, yields an overall competitive ratio of 0.7 for this
new algorithm. We now need to show that this value of η ensures that both fℓ and fs are less
than 1 after modification. Since fℓ ≤ 1 − 1/e we have that fℓ + η ≤ 1 − 1/e + 0.0142 ≤ 1. Note
that fℓ ≥ 1/2. Hence, the modified value of fs is always less than or equal to the original value,
ℓ +η)
is decreasing in the range fℓ ∈ [1/2, 1 − 1/e] and has a value less than 0.98 at
since 1−(f
1−fℓ
fℓ = 1/2.
3.3
A 0.705-competitive algorithm.
In the next few subsections, we describe our final edge-weighted algorithm with all of the attenuation
factors. To keep it modular, we give the following guide to the reader. We note that the definition of
large and small edges given below in Subsection 3.3.1 is different from the definition in the previous
subsection.
• Section 3.3.1 describes the main algorithm which internally invokes two algorithms, EW1 and
EW2 , which are described in sections 3.3.2 and 3.3.3, respectively.
• Theorem 2 proves the final competitive ratio. This proof depends on the performance guarantees of EW1 and EW2 , which are given by Lemmas 2 and 3, respectively.
• The proof of Lemma 2 depends on claims 4, 5, and 6 (Found in the Appendix). Each of those
claims is a careful case-by-case analysis. Intuitively, 4 refers to the case where the offline
vertex u is incident to one large edge and one small edge (here the analysis is for the large
edge), 5 refers to the case where u is incident to three small edges and 6 refers to the case
where u is incident to a small edge and large edge (here the analysis is for the small edge).
• The proof of Lemma 3 depends on claims 7 and 8 (Found in the Appendix). Again, both of
those claims are proven by a careful case-by-case analysis. Since there are many cases, we
have given a diagram of the cases when we prove them.
3.3.1
A 0.705-competitive algorithm.
In this section, we describe an algorithm EW (Algorithm 2), that achieves a competitive ratio
of 0.705. The algorithm first solves the benchmark LP in subsection 2.1 and obtains a fractional
optimal solution f. By invoking DR[f, 3], it obtains a random integral solution F. Notice that from
LP constraint 4 we see fe ≤ 1 − 1/e ≤ 2/3. Therefore after DR[f, 3], each Fe ∈ {0, 1, 2}. Consider
the graph GF where each edge e is associated with the value of Fe . We say an edge e is large
if Fe = 2 and small if Fe = 1 (note that this differs from the definition of large and small in
Subsection 3.2).
We design two non-adaptive algorithms, denoted by EW1 and EW2 , which take the sparse graph
GF as input. The difference between the two algorithms EW1 and EW2 is that EW1 favors the small
11
edges while EW2 favors the large edges. The final algorithm is to take a convex combination of EW1
and EW2 i.e. run EW1 with probability q and EW2 with probability 1 − q.
Algorithm 2: EW[q]
1
2
3
Solve the benchmark LP in sub-section 2.1 for the input. Let f be the optimal solution
vector.
Invoke DR[f, 3] to obtain the vector F.
Independently run EW1 and EW2 with probabilities q and 1 − q respectively on GF .
The details of algorithm EW1 and EW2 and the proof of Theorem 2 are presented in the following
sections.
3.3.2
Algorithm EW1 .
In this section, we describe the randomized algorithm EW1 (Algorithm 3). Suppose we view the
graph of GF in another way where each edge has Fe copies. Let PM[F, 3] refer to the process of
constructing the graph GF with Fe copies of each edge, decomposing it into three matchings, and
randomly permuting the matchings. EW1 first invokes PM[F, 3] to obtain a random ordered triple
of matchings, say [M1 , M2 , M3 ]. Notice that from the LP constraint 4 and the properties of DR[f, 3]
and PM[F, 3], an edge will appear in at most two of the three matchings. For a small edge e = (u, v)
in GF , we say e is of type Γ1 if u has two other neighbors v1 and v2 in GF with F(u,v1 ) = F(u,v2 ) =1.
We say e is of type Γ2 if u has exactly one other neighbor v1 with F(u,v1 ) = 2. WLOG we can
P
assume that for every u, Fu = e∈∂(u) Fe = 3; otherwise, we can add a dummy node v ′ to the
neighborhood of u.
Note, we use the terminology, assign v to u to denote that edge (u, v) is matched by the
algorithm if u is not matched until that step.
Algorithm 3: EW1 [h]
1
2
3
4
5
Invoke PM[F, 3] to obtain a random ordered triple matchings, say [M1 , M2 , M3 ].
When a vertex v comes for the first time, assign v to some u1 with (u1 , v) ∈ M1 .
When v comes for the second time, assign v to some u2 with (u2 , v) ∈ M2 .
When v comes for the third time, if e is either a large edge or a small edge of type Γ1 then
assign v to some u3 with e = (u3 , v) ∈ M3 . However, if e is a small edge of type Γ2 then
with probability h, assign v to some u3 with e = (u3 , v) ∈ M3 ; otherwise, do nothing.
When v comes for the fourth or more time, do nothing in that step.
Here, h is a parameter we will fix at the end of analysis. Let R[EW1 , 1/3] and R[EW1 , 2/3] be
the competitive ratio for a small edge and large edge respectively.
Lemma 2. For h = 0.537815, EW1 achieves a competitive ratio R[EW1 , 2/3] = 0.679417, R[EW1 , 2/3] =
0.751066 for a large and small edge respectively.
Proof. In case of the large edge e, we divide the analysis into three cases where each case corresponds
to e being in one of the three matchings. And we combine these conditional probabilities using
Bayes’ theorem to get the final competitive ratio for e. For each of the two types of small edges,
12
we similarly condition them based on the matching they can appear in, and combine them using
Bayes’ theorem. A complete version of proof can be found in Section A.1.1 of Appendix.
3.3.3
Algorithm EW2 .
Algorithm EW2 (Algorithm 5) is a non-adaptive algorithm which takes GF as input and performs
well on the large edges. Recall that the previous algorithm, EW1 , first invokes PM[F, 3] to obtain a
random ordered triple of matchings. In contrast, EW2 will invoke a routine, denoted by PM∗ [F, 2]
(Algorithm 4), to generate a (random ordered) pair of pseudo-matchings from F. Recall that F is
an integral solution vector where ∀e we have Fe ∈ {0, 1, 2}. WLOG, we can assume that Fv = 1
for every v in GF ; otherwise we can dummy vertex to ensure this is the case.
Algorithm 4: PM∗ [F, 2][y1 , y2 ]
1
2
Suppose v has two neighbors in GF , say u1 , u2 , with e1 = (u1 , v) being a large edge while
e2 = (u2 , v) being a small edge. Add e1 to the primary matching M1 and e2 to the
secondary matching M2 .
Suppose v has three neighbors in GF and the incident edges are ∂(v) = (e1 , e2 , e3 ). Take
a random permutation of ∂(v), say (π1 , π2 , π3 ) ∈ Π(∂(v)). Add π1 to M1 with
probability y1 and π2 to M2 with probability y2 .
Here 0 ≤ y1 , y2 ≤ 1 are parameters which will be fixed after the analysis. Algorithm 5 describes
EW2 .
Algorithm 5: [EW2 ][y1 , y2 ]
1
2
3
Invoke PM∗ [F, 2][y1 , y2 ] to generate a random ordered pair of pseudo-matchings, say
[M1 , M2 ].
When a vertex v comes for the first time, assign v to some u1 if (u1 , v) ∈ M1 ; When v
comes for the second time, try to assign v to some u2 if (u2 , v) ∈ M2 .
When a vertex v comes for the third or more time, do nothing in that step.
Let R[EW2 , 1/3] and R[EW2 , 2/3] be the competitive ratios for small edges and large edges,
respectively.
Lemma 3. For y1 = 0.687 and y2 = 1, EW2 [y1 , y2 ] achieves a competitive ratio of R[EW2 , 2/3] =
0.8539 and R[EW2 , 1/3] = 0.4455 for a large and small edge respectively.
Proof. We analyze this on a case-by-case basis by considering the local neighborhood of the edge. A
large edge can have two possible cases in its neighborhood, while a small edge can have eight possible
cases. This is because of the fact that a large edge can have only small edges in its neighborhood
while a small edge can have both large and small edges in its neighborhood. Choosing the worst
case among the two for large edge and the worst case among the eight for the small edge, we prove
the claim. Complete details of the proof can be found in section A.1.2 of Appendix.
3.3.4
Convex combination of EW1 and EW2 .
In this section, we prove theorem 2.
13
Proof. Let (a1 , b1 ) be the competitive ratios achieved by EW1 for large and small edges, respectively.
Similarly, let (a2 , b2 ) denote the same for EW2 . We have the following two cases.
• 0 ≤ fe ≤ 31 : By marginal distribution property of DR[f, 3], we know that Pr[Fe = 1] = 3fe .
Thus, the final ratio is
3fe (qb1 /3 + (1 − q)b2 /3)/fe = qb1 + (1 − q)b2
• 1/3 ≤ fe ≤ 1 − 1/e: By the same properties of DR[f, 3], we know that Pr[Fe = 2] = 3fe − 1
and Pr[Fe = 1] = 2 − 3fe . Thus, the final ratio is
(3fe − 1)(2qa1 /3 + 2(1 − q)a2 /3) + (2 − 3fe )(qb1 /3 + (1 − q)b2 /3) /fe
The competitive ratio of the convex combination is maximized at q = 0.149251 with a value of
0.70546.
4
Vertex-weighted stochastic I.I.D. matching with integral arrival
rates.
In this section, we consider vertex-weighted online stochastic matching on a bipartite graph G
under the known I.I.D. model with integral arrival rates. We present an algorithm in which each
offline vertex u has a competitive ratio of at least 0.72998 > 1 − 2e−2 . Recall that after invoking
DR[f, 3], we can obtain a (random) integral vector F with Fe ∈ {0, 1, 2}. Define H = F/3 and let
GH be the graph induced by H. Each edge e in GH thus takes a value He ∈ {0, 1/3, 2/3}. Notice
. P
that for each u, Hu = e∈∂(u) He ≤ 1, which implies that u has at most 3 neighbors in GH (we
ignore all edges e with He = 0). In this section, we focus on the sparse graph GH . The main idea
of our algorithm is as follows.
1. Solve the vertex-weighted benchmark LP in Section 2.1. Let f be an optimal solution vector.
2. Invoke DR[f, 3] to obtain an integral vector F and a fractional vector H with H = F/3.
3. Apply a series of modifications to H and transform it to another solution H′ (See Section
4.2).
4. Run the Randomized List Algorithm [13] based on H′ , denoted by RLA[H′ ], on the graph GH .
We first briefly describe how we overcome the bottleneck case for the algorithm in [13] and
then explain the algorithm in full detail. The WS for the vertex-weighted case in [13] is shown in
Figure 2, which happens at a node u with a competitive ratio of 0.725 (recall that [13] analyze their
algorithm by considering cases for various neighborhood structure at a given offline vertex). From
the analysis of [13], we have that the node u1 (in Figure 2) has a competitive ratio of at least 0.736.
Hence, we can boost the performance of u at the cost of u1 . Specifically, we increase the value of
H(u,v1) and decrease the value H(u1 ,v1 ) . Cases (10) and (11) in Figure 4 illustrate this. After this
modification, the new WS for vertex-weighted is now the C1 cycle shown in both Figures 1 and 2.
14
In fact, this is the WS for the unweighted case in [13] as well. However, Lemma 4 implies that C1
cycles can be avoided with probability at least 3/e − 1. This helps us improve the ratio even for
the unweighted case in [13]. Lemma 4 describes this formally.
Lemma 4. For any given u ∈ U , u appears in a C1 cycle after DR[f, 3] with probability at most
2 − 3/e.
Proof. Consider the graph GH obtained after DR[f, 3]. Notice that for some vertex u to appear in
a C1 cycle, it must have a neighboring edge with He = 2/3. Now we try to bound the probability
of this event. It is easy to see that for some e ∈ ∂(u) with fe ≤ 1/3, Fe ≤ 1 after DR[f, 3], and
hence He = Fe /3 ≤ 1/3. Thus only those edges e ∈ ∂(u) with fe > 1/3 will possibly be rounded
P
to He = 2/3. Note that, there can be at most two such edges in ∂(u), since e∈∂(u) fe ≤ 1. Hence,
we have the following two cases.
• Case 1:
∂(u) contains only one edge e with fe > 1/3. Let q1 = Pr[He = 1/3] and
q2 = Pr[He = 2/3] after DR[f, 3]. By DR[f, 3], we know that E[He ] = E[Fe ]/3 = q2 (2/3) +
q1 (1/3) = fe .
Notice that q1 + q2 = 1 and hence q2 = 3fe − 1. Since this is an increasing function of fe and
fe ≤ 1 − 1/e from LP constraint 4, we have q2 ≤ 3(1 − 1/e) − 1 = 2 − 3/e.
• Case 2: ∂(u) contains two edges e1 and e2 with fe1 > 1/3 and fe2 > 1/3. Let q2 be
the probability that after DR[f, 3], either He1 = 2/3 or He2 = 2/3. Note that, these two
events are mutually exclusive since Hu ≤ 1. Using the analysis from case 1, it follows that
q2 = (3fe1 − 1) + (3fe2 − 1) = 3(fe1 + fe2 ) − 2.
From LP constraint 5, we know that fe1 + fe2 ≤ 1 − 1/e2 , and hence q2 ≤ 3(1 − 1/e2 ) − 2 <
2 − 3/e.
Now we present the details of RLA based on a given H′ in Section 4.1 and then discuss the two
modifications transforming H to H′ in Section 4.2. We give a formal statement of our algorithm
in Section 4.3 and the related analysis.
4.1
RLA algorithm.
Now we discuss how to apply RLA based on H′ to the sparse graph GH . Let δH (v) be the set
. P
of neighboring nodes of v in GH . Here we assume WLOG that Hv = e∈∂(v) He = 1 and thus
each v has at least 2 neighbors in GH since each non-zero He satisfies He ∈ {1/3, 2/3} (we are
in a better situation when Hv < 1). Additionally, we will see in Section 4.2 that during the two
modifications, we have the sum of all edge values incident to v (i.e., Hv ) unchanged and hence we
have H′v = Hv = 1 for each v.
Each time when a vertex v comes, RLA first generates a random list Rv , which is a permutation
over δH (v), based on H′ as follows.
• If |δH (v)| = 2, say δH (v) = (u1 , u2 ), then sample a random list Rv such that
′
′
Pr[Rv = (u1 , u2 )] = H(u
, Pr[Rv = (u2 , u1 )] = H(u
1 ,v)
2 ,v)
15
(7)
• If |δH (v)| = 3, say δH (v) = (u1 , u2 , u3 ). Then we sample a permutation of (i, j, k) over {1, 2, 3}
such that
′
H(u
j ,v)
′
(8)
Pr[Rv = (ui , uj , uk )] = H(ui ,v) ′
′
H(uj ,v) + H(u
k ,v)
We can verify that the sampling distributions described in Equations (7) and (8) are valid since
P
H′v = e∈∂(v) He′ = 1.
The full details of the Random List Algorithm, RLA[H′ ], are shown in Algorithm 6.
Algorithm 6: RLA[H′ ] (Random List Algorithm based on H′ )
1
2
4.2
When a vertex v comes, generate a random list Rv satisfying Equation (7) or (8)
If all u in the list are matched, then drop the vertex v; otherwise, assign v to the first
unmatched u in the list.
Two kinds of modifications to H.
As stated earlier, we first modify H before running the RLA algorithm. In this section, we describe
the modifications.
4.2.1
The first modification to H
(C1 )
(WS)
1
u1
1
u
(C2 )
(C3 )
u1
v1
u1
v1
u1
v1
u2
v2
u2
v2
u2
v2
u1
v1
u1
v1
u2
v2
u2
v2
v1
v2
1
u2
1
u3
Figure 2: Left: The WS for Jaillet and Lu [13] for their vertex-weighted case. Right: The three
possible types of cycles of length 4 after applying DR[f, 3]. Thin edges have He = 1/3 and thick
edges have He = 2/3.
The first modification is to break the cycles deterministically. There are three possible cycles
of length 4 in the graph GH , denoted C1 , C2 , and C3 . In [13], they give an efficient way to break
C2 and C3 , as shown in Figure 2. Cycle C1 cannot be modified further and hence, is the bottleneck
for their unweighted case. Notice that, while breaking the cycles of C2 and C3 , new cycles of C1
can be created in the graph. Since our randomized construction of solution H gives us control on
the probability of cycles C1 occurring, we would like to break C2 and C3 in a controlled way, so as
not to create any new C1 cycles. This procedure is summarized in Algorithm 7 and its correctness
is proved in Lemma 6.
16
Algorithm 7: [Cycle breaking algorithm] Offline Phase
1
2
3
While there is some cycle of type C2 or C3 , Do:
Break all cycles of type C2 .
Break one cycle of type C3 and return to the first step.
Lemma 5. After applying Algorithm 7 to GH , we have (1) the value Hw is preserved for each
w ∈ U ∪ V ; (2) no cycle of type C2 or C3 exists; (3) no new cycle of type C1 is added.
Proof of Lemma 5. Lemma 5 follows from the following three Claims:
Claim 1. Breaking cycles will not change the value Hw for any w ∈ U ∪ V .
Claim 2. After breaking a cycle of type C2 , the vertices u1 , u2 , v1 , and v2 can never be part of
any length four cycle.
Claim 3. When all length four cycles are of type C1 or C3 , breaking exactly one cycle of type C3
cannot create a new cycle of type C1 .
Proof of Claim 1. As shown in Figure 2, we increase and decrease edge values fe in such a
way that their sums Hw at any vertex w will be preserved.
Notice that C2 cycles can be freely broken without creating new C1 cycles. After removing all
cycles of type C2 , removing a single cycle of type C3 cannot create any cycles of type C1 . Hence,
Algorithm 7 removes all C2 and C3 cycles without creating any new C1 cycles.
Proof of Claim 2. Consider the structure after breaking a cycle of type C2 . Note that the
edge (u2 , v2 ) has been permanently removed and hence, these four vertices together can never be
part of a cycle of length four. The vertices u1 and v1 have Hu1 = 1 and Hv1 = 1 respectively.
So they cannot have any other edges and therefore cannot appear in any length four cycle. The
vertices u2 and v2 can each have one additional edge, but since the edge (u2 , v2 ) has been removed,
they can never be part of any cycle with length less than six.
Proof of Claim 3. First, we note that since no edges will be added during this process, we
cannot create a new cycle of length four or join with a cycle of type C1 . Therefore, the only cycles
which could be affected are of type C3 . However, every cycle c of type C3 falls into one of two cases:
Case 1:
c is the cycle we are breaking. In this case, c cannot become a cycle of type C1 since
we remove two of its edges and break the cycle.
Case 2: c is not the cycle we are breaking. In this case, c can have at most one of its edges
converted to a 2/3 edge. Let c′ be the length four cycle we are breaking. Note that c and c′
will differ by at least one vertex. When we break c′ , the two edges which are converted to
2/3 will cover all four vertices of c′ . Therefore, at most one of these edges can be in c.
Note that breaking one cycle of type C3 could create cycles of type C2 , but these cycles are
always broken in the next iteration, before breaking another cycle of type C3 .
4.2.2
The Second modification to H.
Informally, this second modification decreases the rates of lists associated with those nodes u with
Hu = 1/3 or Hu = 2/3 and increases the rates of lists associated with nodes u with Hu = 1. We
will illustrate our intuition on the following example.
17
1/3 u1
v1
1
1/3 u1
u
0.1
v1
v2
2/3 u2
1
u
0.9
Figure 3: An example of the need for the second modification. For the left: competitive analysis
shows that in this case, u1 and u2 can achieve a high competitive ratio at the expense of u. For the
right: an example of balancing strategy by making v1 slightly more likely to pick u when it comes.
Consider the graph G in Figure 3. Let thin and thick edges represent He = 1/3 and He = 2/3
respectively. We will now calculate the competitive ratio after applying RLA on G. Let Pu denote
the probability that u gets matched after the algorithm. Let Bu denote the event that among the
n random lists, there exists a list starting with u and Gvu denote the event that among the n lists,
there exists successive lists such that (1) Each of those lists starts with a u′ 6= u and u′ ∈ δ(v) and
(2) The lists arrive in an order which ensures u will be matched by the algorithm. From lemma 4
and Corollary 1 in [13], the following lemma follows:
Lemma 6. Suppose u is not a part of any cycle of length 4. We have
Pu = 1 − (1 − Pr[Bu ])
Y
v∼u
(1 − Pr[Gvu ]) + o(1/n)
For the node u, we have Pr[Bu ] = 1 − e−1 . From definition, Gvu1 is the event that among the n
lists, the random list Rv1 = (u1 , u) comes at least twice. Notice that the list Rv1 = (u1 , u) comes
1
with probability 3n
. Thus we have Pr[Gvu1 ] = Pr[X ≥ 2] = 1 − e−1/3 (1 + 1/3), where X ∼ Pois(1/3).
20
Similarly, we can get Pr[Gvu2 ] = 1 − e−2/3 (1 + 2/3) and the resultant Pu = 1 − 9e
2 ∼ 0.699. Observe
−1/3
−2/3
that Pu1 ≥ Pr[Bu1 ] = 1 − e
and Pu2 ≥ Pr[Bu2 ] = 1 − e
. Let R[RLA, 1], R[RLA, 1/3] and
R[RLA, 2/3] be the competitive ratio achieved by RLA for u, u1 and u2 respectively. Hence, we have
R[RLA, 1] ∼ 0.699 while R[RLA, 1/3] ≥ 3(1 − e−1/3 ) ∼ 0.8504 and R[RLA, 2/3] ≥ 0.729.
Intuitively, one can improve the worst case ratio by increasing the arrival rate for Rv1 = (u, u1 )
′
while reducing that for Rv1 = (u1 , u). Suppose one modifies H(u1 ,v1 ) and H(u,v1 ) to H(u
= 0.1
1 ,v1 )
′
and H(u,v1 ) = 0.9, the arrival rate for Rv1 = (u, u1 ) and Rv1 = (u1 , u) gets modified to 0.1/n and
0.9/n respectively. The resulting changes are Pr[Bu ] = 1 − e−0.9−1/3 , Pr[Gvu1 ] = 1 − e−0.1 (1 + 0.1),
R[RLA, 1] = 0.751, Pr[Bu1 ] = 1 − e−1/3 , Pr[Gvu11 ] ∼ 0.227 and R[RLA, 1/3] ≥ 0.8. Hence, the
′
performance on WS instance improves. Notice that after the modifications, Hu′ = H(u,v
+H(u,v2 ) =
1)
0.9 + 1/3.
Figure 4 describes the various modifications applied to H vector. The values on top of the edge,
denote the new values. Cases (11) and (12) help improve upon the WS described in Figure 2.
4.3
4.3.1
Vertex-Weighted Algorithm VW
Analysis of algorithm VW.
The full details of our vertex-weighted algorithm are stated as follows.
18
1/3 u1
2/3 u1
0.1
v
1
u2
0.9
1
u2
1
u3
0.2
0.4
1
u2
1/3 u1
0.4
2/3 u2
1
u3
(5)
2/3 u1
2/3 u1
0.3
2/3 u2
0.7
1
(9)
0.85
1
u2
0.15
0.2
1/3 u1
v
1/3 u2
0.65
u3
0.25
0.25
1
u2
1
u3
0.4
(3)
1
u3
(6)
v
2/3 u2
v
(2)
v
1/3 u1
0.6
v
(1)
2/3 u1
2/3 u1
0.15
0.1
0.1
u1
v
1/3 u1
1
0.5
0.75
(8)
1
u1
v
1 − x2
x2
1
(11)
(10)
0.25
v
2/3 u2
1 − x1
u2
0.45
v
0.8
x1
v
(4)
(7)
1
0.1
0.45
v
u2
(12)
Figure 4: Illustration for second modification to H. The value assigned to each edge represents the
value after the second modification. Here, x1 = 0.2744 and x2 = 0.15877.
Algorithm 8: VW
1 Construct and solve the LP in sub-section 2.1 for the input instance.
2 Invoke DR[f, 3] to output F and H. Apply the two kinds of modifications to morph H to
H′ .
′
3 Run RLA[H ] on the graph GH .
The algorithm VW consists of two different random processes: sub-routine DR[f, 3] in the offline
phase and RLA in the online phase. Consequently, the analysis consists of two parts. First, for a
given graph GH , we analyze the ratio of RLA[H′ ] for each node u with Hu = 1/3, Hu = 2/3 and
Hu = 1. The analysis is similar to [13]. Second, we analyze the probability that DR[f, 3] transforms
each u, with fractional fu values, into the three discrete cases seen in the first part. By combining
the results from these two parts we get the final ratio.
Let us first analyze the competitive ratio for RLA[H′ ]. For a given H and GH , let Pu be the
probability that u gets matched in RLA[H′ ]. Notice that the value Pu is determined not just by
the algorithm RLA itself, but also the modifications applied to H. We define the competitive ratio
of a vertex u achieved by RLA as Pu /Hu , after modifications. Lemma 7 gives the respective ratio
values. The proof can be found in section A.2.1 in the Appendix.
Lemma 7. Consider a given H and a vertex u in GH . The respective ratios achieved by RLA after
the modifications are as described below.
19
• If Hu = 1, then the competitive ratio R[RLA, 1] = 1 − 2e−2 ∼ 0.72933 if u is in the first cycle
C1 and R[RLA, 1] ≥ 0.735622 otherwise.
• If Hu = 2/3, then the competitive ratio R[RLA, 2/3] ≥ 0.7847.
• If Hu = 1/3, then competitive ratio R[RLA, 1/3] ≥ 0.7622.
Now we have all essentials to prove Theorem 1.
Proof. From Lemmas 4 and 5, we know that any u is present in cycle C1 with probability at most
(2 − 3/e).
Consider a node u with 2/3 ≤ fu ≤ 1 and let q1 , q2 , q3 be the probability that after DR[f, 3] and
the first modification, Hu = 1 and u is in the first cycle C1 , Hu = 1 and u is not in C1 , Hu = 2/3
respectively. From Lemma 7, we get that the final ratio for u should be at least
(0.72933q1 + 0.735622q2 + (2/3) ∗ 0.7847q3 )/(q1 + q2 + (2/3)q3 )
Minimizing the above expression subject to (1) q1 + q2 + q3 = 1; (2) 0 ≤ qi , 1 ≤ i ≤ 3; (3)
q1 ≤ 2 − 3/e, we get a minimum value of 0.729982 for q1 = 2 − 3/e and q2 = 3/e − 1.
For any node u with 0 ≤ u ≤ 2/3, we know that the ratio is at least the min value of R[RLA, 2/3]
and R[RLA, 1/3], which is 0.7622. This completes the proof of Theorem 1.
5
Non-integral arrival rates with stochastic rewards.
The setting here is strictly generalized over the previous sections in the following ways. Firstly,
it allows an arbitrary arrival rate (say rv ) which can be fractional for each stochastic vertex v.
P
Notice that, v rv = n where n is the total number of rounds. Secondly, each e = (v, u) ∈ E is
associated with a value pe , which captures the probability that the edge e = (u, v) is present when
we probe it, i.e., we try to assign v to u. We assume this process is independent of the stochastic
arrival of each v. We will show that the simple non-adaptive algorithm introduced in [12] can be
extended to this general case. This achieves a competitive ratio of (1 − 1e ). Note that Manshadi
et al. [19] show that no non-adaptive algorithm can possibly achieve a ratio better than (1 − 1/e)
for the non-integral arrival rates, even for the case of all pe = 1. Thus, our algorithm is an optimal
non-adaptive algorithm for this model.
We use an LP similar to [13] for the case of non-integral arrival rates. For each e = (u, v) ∈ E,
let fe be the expected number of probes on edge e. When there are multiple copies of v, we count
the sum of probes among all copies of e in the offline optimal and thus some realizations of fe can
be greater than 1.
max
X
we fe pe :
(9)
e∈E
s.t.
X
fe pe ≤ 1,
X
f e ≤ rv ,
e∈∂(u)
e∈∂(v)
20
∀u ∈ U
∀v ∈ V
(10)
(11)
Our algorithm is summarized in Algorithm 9. Notice that the last constraint ensures that step
2 in the algorithm is valid. Let us now prove theorem 3.
Algorithm 9: SM
1
2
Construct and solve LP (9). WLOG assume {fe |e ∈ E} is an optimal solution.
When a vertex v arrives, assign v to each of its neighbors u with a probability
f(u,v)
rv .
Proof. Let B(u, t) be the event that u is safe at beginning of round t and A(u, t) to be the event
that vertex u is matched during the round t conditioned on B(u, t). From the algorithm, we know
Pr[A(u, t)] ≤
P
v∼u
rv fu,v
n rv p e
≤ n1 , which follows by P r[B(u, t)] = Pr
i
hV
t−1
i=1 (¬A(u, i)) ≥ 1 −
1
n
t−1
.
Consider an edge e = (u, v) in the graph. Notice that the probability that e gets matched in
SM should be
Pr[e is matched] =
≥
6
n
X
Pr[v arrives at t and B(u, t) ] ·
t=1
n
X
t=1
1
1−
n
t−1
f e pe
rv
1
rv fe pe
≥ 1−
f e pe
n rv
e
Integral arrival rates with uniform stochastic rewards.
In this section, we consider a special case of the model studied in Section 5 and show that we can
indeed surpass the 1 − 1/e barrier. We specialize the model in the following two ways. (1) We
consider the unweighted case with uniform constant edge probabilities (i.e., we = 1 and pe = p for
some constant p ∈ (0, 1] for all e ∈ E). The constant p is arbitrary, but independent of the problem
parameters. (2) Each vertex v that comes online has an integral arrival rate rv (as usual WLOG
rv = 1 and |V | = n). We refer to this special model as unweighted online stochastic matching
with integral arrival rates and uniform stochastic rewards. Note that even for this special case,
given an offline instance (i.e., the sequence of realizations for the online arrival), it is unclear if
we can efficiently solve or approximate the exact offline optimal within (1 − ǫ) without any extra
assumptions. Hence we cannot directly apply the Monte-Carlo simulation technique in [19] to
approximate the exact expected offline optimal within an arbitrary desired accuracy. Here we
present a strengthened LP as the benchmark to upper bound the offline optimal.
max p ·
s.t.
X
fe :
(12)
e∈E
X
fe · p ≤ 1,
∀u ∈ U
(13)
X
fe ≤ 1,
∀v ∈ V
(14)
fe p ≤ 1 − exp(−|S|p),
∀S ⊆ ∂(u), |S| ≤ 2/p
(15)
e∈∂(u)
e∈∂(v)
X
S⊆∂(u)
21
Lemma 8. LP (12) is a valid upper bound for the expected offline optimal.
Proof. It suffices to show that the constraint (15) is valid (the correctness of the other constraints
follows from the previous section). Consider a given S ⊆ ∂(u). Let X(I) be the number of copies
of all edges in S in an offline instance I and Y (I) be the probability that any edge in S is included
in an offline optimal on I. From a straightforward calculation, it follows that Y (I) ≤ 1 − (1 − p)X(I)
and using linearity of expectation we get,
E[Y (I)] ≤ 1 − E[(1 − p)X(I) ]
By definition we have E[Y (I)] = e∈S fe · p. Since n is assumed to be sufficiently large, X(I)
follows a Poisson distribution with mean |S|. Substituting the fact that E[(1 − p)X(I) ] = exp(−p|S|)
into the equation above, we prove the lemma.
P
Note that it is impossible to beat 1 − 1/e using LP (12) as the benchmark without the extra
constraint (15) (see the hardness instance shown in [5]). Our main idea in the online phase is
primarily based on [19]. In the offline phase, we first solve LP (12) and get an optimal solution
{fe∗ }. When a vertex v arrives, we generate a random list of two choices based on {fe∗ |e ∈ ∂(v)},
denoted by Lv = (Lv (1), Lv (2)), where Lv (1), L(2) ∈ ∂(v). Our online decision based on Lv is as
follows: if Lv (1) = (u, v) is safe, i.e., u is available, then match v to u; else if the second choice
Lv (2) is safe match v to Lv (2). The random list Lv generated based on {fe∗ |e ∈ ∂(v)} satisfies the
following two properties:
(P1) : Pr[Lv (1) = e] = fe∗ and Pr[Lv (2) = e] = fe∗ for each e ∈ ∂(v).
(P2) : Pr[Lv (1) = e ∧ Lv (2) = e] = max 2fe − 1, 0 for each e ∈ ∂(v).
Algorithm 10:
1
2
3
Solve LP (12) and let {fe∗ |e ∈ E} be an optimal solution.
When a vertex v arrives, generate a random list Lv of two choices based on {fe∗ |e ∈ ∂(v)}
such that Lv satisfies Property (P1) and (P2).
If Lv (1) = (u, v) is safe, i.e., u is available, then assign v to u; else if the second choice
Lv (2) is safe, match v to it.
There are several ways to generate Lv satisfying (P1) and (P2). One simple way is shown in
Section 4 of [19]. We can verify that all of the calculations shown in [19] can be extended here if we
incorporate the independent process that each e will be present with probability p after we assign
v to u. Hence, the final ratio is as follows (this can be viewed as a counterpart to Equation (15) on
page 11 of [19]).
E[ALG]
≥ min
u∈U
E[OPT]
where fu′ =
that
′
(1 − e−fu ) + qu′ e−2 − (qu′ )2 e−1
fu′
′
∗
e∈∂(u) fe · p ≤ 1 and qu = p ·
P
qu′ ≤ p ·
X
e=(u,v)∈∂(u)
P
1
2
.
′
′
= F (fu , qu )
(16)
e=(u,v)∈∂(u) Pr[Lv (2) = e ∧ Lv (1) 6= e] . Observe
Pr[Lv (2) = e] = p ·
22
− e−1 − e−2 fu′ (1 − fu′ )
X
e∈∂(u)
fe∗ = fu′ ≤ 1
We can verify that for each given fu′ ≤ 1, the RHS expression in inequality (16) is an increasing
function of qu′ during the interval [0, 1]. Thus an important step is to lower bound qu′ for a given
fu′ . The following key lemma can be viewed as a counterpart to Lemma 4.7 of [19]:
Lemma 9. For each given fu′ ≥ ln 2/2, we have that qu′ ≥ fu′ − (1 − ln 2).
Proof. Consider a given u with fu′ ≥ ln 2/2. Define ∆ = fu′ − qu′ . Thus we have:
∆=p·
=p·
=p·
=p·
X
e=(u,v)∈∂(u)
fe∗ − Pr[Lv (2) = e ∧ Lv (1) 6= e]
X
e=(u,v)∈∂(u)
Pr[Lv (2) = e] − Pr[Lv (2) = e ∧ Lv (1) 6= e]
X
Pr[Lv (2) = e ∧ Lv (1) = e]
e=(u,v)∈∂(u)
X
e∈∂(u)
From Property P1
max 2fe∗ − 1, 0
From Property P2
Thus to lower bound qu′ , we essentially need to maximize ∆. Let S ∗ ⊆ ∂(u) be the set of edges
in ∂(u) with fe∗ ≥ 1/2, which is called a contributing edge. Thus we have
∆=p·
Observe that
X
e∈∂(u)
max 2fe∗ − 1, 0 = p ·
X
e∈S ∗
(2fe∗ − 1) =
X
e∈S ∗
2pfe∗ − p|S ∗ |
X
2f ′
2
p ∗
|S | ≤
fe∗ · p ≤ fu′ ⇒ |S ∗ | ≤ u ≤
2
p
p
e∈S ∗
From Constraint (15), we have
into Equation (17), we get
P
∗
e∈S ∗ (pfe )
(17)
(18)
≤ 1 − exp(−|S ∗ |p). Substituting this inequality back
∆ ≤ 2 − 2 exp − |S ∗ | · p − |S ∗ | · p
It is easy to verify that when fu′ ≥ ln 2/2, the above expression has a maximum value of 1 − ln 2
when |S ∗ | · p = ln 2. Thus we have that ∆ ≤ 1 − ln 2 and qu′ ≥ fu′ − (1 − ln 2).
Let us now complete the proof of Theorem 4.
Proof. We need to prove that F (fu′ , qu′ ) defined in 16 has a lower bound of 0.702 for all fu′ ∈ [0, 1].
Consider the first case when fu′ ≤ ln 2/2. It is easy to verify that F (fu′ , qu′ ) ≥ F (fu′ , 0) ≥
F (ln 2/2, 0) ∼ 0.8. Consider the second case when fu′ ≥ ln 2/2. From Lemma 9, we have qu′ ≥
fu′ − (1 − ln 2). Once again, simple calculations show that
F (fu′ , qu′ ) ≥ F fu′ , fu′ − (1 − ln 2) ≥ F (1, 1 − (1 − ln 2)) ∼ 0.702
23
7
Conclusion and future directions.
In this paper, we gave improved algorithms for the Edge-Weighted and Vertex-Weighted models.
Previously, there was a gap between the best unweighted algorithm with a ratio of 1 − 2e−2 due
to [13] and the negative result of 1 − e−2 due to [19]. We took a step towards closing that gap
by showing that an algorithm can achieve 0.7299 > 1 − 2e−2 for both the unweighted and vertexweighted variants with integral arrival rates. In doing so, we made progess on Open Questions 3
and 4 in the online matching and ad allocation survey [20]. This was possible because our approach
of rounding to a simpler fractional solution allowed us to employ a stricter LP. For the edgeweighted variant, we showed that one can significantly improve the power of two choices approach
by generating two matchings from the same LP solution. For the variant with edge weights, nonintegral arrival rates, and stochastic rewards, we presented a (1 − 1/e)-competitive algorithm. This
showed that the 0.62 < 1−1/e bound given in [22] for the adversarial model with stochastic rewards
does not extend to the known I.I.D. model.
A natural next step in the edge-weighted setting is to use an adaptive strategy. For the vertexweighted problem, one can easily see that the stricter LP we use still has a gap. In addition,
we only utilize fractional solutions {0, 1/3, 2/3}. However, dependent rounding gives solutions in
{0, 1/k, 2/k, . . . , ⌈k(1 − 1/e)⌉/k}; allowing for random lists of length greater than three. Stricter
LPs and longer lists could both yield improved results. In the stochastic
model with non
rewards
integral arrival rates, an open question is to either improve upon the 1 − 1e ratio in the general
case. In this work, we showed how for certain restrictions it is possible to beat 1 − 1/e. However,
the serious limitation comes from the fact that a polynomial sized LP is insufficient to capture the
complexity of the problem.
Acknowledgments.
The authors would like to thank Aranyak Mehta and the anonymous
reviewers for their valuable comments, which have significantly helped improve the presentation of
this paper.
References
[1] Aggarwal, G., Goel, G., Karande, C., and Mehta, A. Online vertex-weighted bipartite
matching and single-bid budgeted allocations. In Proceedings of the twenty-second annual
ACM-SIAM symposium on Discrete Algorithms (2011), SIAM, pp. 1253–1264.
[2] Alaei, S., Hajiaghayi, M., and Liaghat, V. Online prophet-inequality matching with
applications to ad allocation. In Proceedings of the 13th ACM Conference on Electronic Commerce (2012), ACM, pp. 18–35.
[3] Alaei, S., Hajiaghayi, M., and Liaghat, V. The online stochastic generalized assignment
problem. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and
Techniques. Springer, 2013, pp. 11–25.
[4] Assadi, S., Khanna, S., and Li, Y. The stochastic matching problem with (very) few
queries. In Proceedings of the 2016 ACM Conference on Economics and Computation (2016),
ACM, pp. 43–60.
24
[5] Brubach, B., Sankararaman, K. A., Srinivasan, A., and Xu, P. Attenuate locally, win
globally: An attenuation-based framework for online stochastic matching with timeouts. In
Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems (Richland,
SC, 2017), AAMAS ’17, International Foundation for Autonomous Agents and Multiagent
Systems, pp. 1223–1231.
[6] Devanur, N. R., and Hayes, T. P. The adwords problem: online keyword matching with
budgeted bidders under random permutations. In Proceedings of the 10th ACM conference on
Electronic commerce (2009), ACM, pp. 71–78.
[7] Devanur, N. R., Jain, K., Sivan, B., and Wilkens, C. A. Near optimal online algorithms
and fast approximation algorithms for resource allocation problems. In Proceedings of the 12th
ACM conference on Electronic commerce (2011), ACM, pp. 29–38.
[8] Devanur, N. R., Sivan, B., and Azar, Y. Asymptotically optimal algorithm for stochastic
adwords. In Proceedings of the 13th ACM Conference on Electronic Commerce (2012), ACM,
pp. 388–404.
[9] Feldman, J., Korula, N., Mirrokni, V., Muthukrishnan, S., and Pál, M. Online ad
assignment with free disposal. In Internet and network economics. Springer, 2009, pp. 374–385.
[10] Feldman, J., Mehta, A., Mirrokni, V., and Muthukrishnan, S. Online stochastic
matching: Beating 1-1/e. In Foundations of Computer Science (FOCS) (2009), IEEE, pp. 117–
126.
[11] Gandhi, R., Khuller, S., Parthasarathy, S., and Srinivasan, A. Dependent rounding
and its applications to approximation algorithms. Journal of the ACM (JACM) 53, 3 (2006),
324–360.
[12] Haeupler, B., Mirrokni, V. S., and Zadimoghaddam, M. Online stochastic weighted
matching: Improved approximation algorithms. In Internet and Network Economics, vol. 7090
of Lecture Notes in Computer Science. Springer Berlin Heidelberg, 2011, pp. 170–181.
[13] Jaillet, P., and Lu, X. Online stochastic matching: New algorithms with better bounds.
Mathematics of Operations Research 39, 3 (2013), 624–646.
[14] Karp, R. M., Vazirani, U. V., and Vazirani, V. V. An optimal algorithm for on-line
bipartite matching. In Proceedings of the twenty-second annual ACM symposium on Theory
of computing (1990), ACM, pp. 352–358.
[15] Kesselheim, T., Radke, K., Tönnis, A., and Vöcking, B. An optimal online algorithm for weighted bipartite matching and extensions to combinatorial auctions. In European
Symposium on Algorithms (ESA). Springer, 2013, pp. 589–600.
[16] Korula, N., and Pál, M. Algorithms for secretary problems on graphs and hypergraphs.
In Automata, Languages and Programming. Springer, 2009, pp. 508–520.
[17] Lee, Y. T., and Sidford, A. Efficient inverse maintenance and faster algorithms for linear programming. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual
Symposium on (2015), IEEE, pp. 230–249.
25
[18] Mahdian, M., and Yan, Q. Online bipartite matching with random arrivals: an approach
based on strongly factor-revealing LPs. In Proceedings of the forty-third annual ACM symposium on Theory of computing (2011), ACM, pp. 597–606.
[19] Manshadi, V. H., Gharan, S. O., and Saberi, A. Online stochastic matching: Online
actions based on offline statistics. Mathematics of Operations Research 37, 4 (2012), 559–573.
[20] Mehta, A. Online matching and ad allocation. Foundations and Trends in Theoretical
Computer Science 8, 4 (2012), 265–368.
[21] Mehta, A., and Panigrahi, D. Online matching with stochastic rewards. In Foundations
of Computer Science (FOCS) (2012), IEEE, pp. 728–737.
[22] Mehta, A., Waggoner, B., and Zadimoghaddam, M. Online stochastic matching with
unequal probabilities. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on
Discrete Algorithms (2015), SIAM.
A
Appendix
A.1
A.1.1
Complementary materials in section 3 (Edge-weighted model).
Proof of Lemma 2
We will prove Lemma 2 using the following three Claims. Recall that we had one kind of large
edge, while two kinds of small edges. Hence, the following claim characterizes the performance of
each of them.
Claim 4. For a large edge e, EW1 [h] (3) with parameter h achieves a competitive ratio of R[EW1 , 2/3] =
0.67529 + (1 − h) ∗ 0.00446.
Claim 5. For a small edge e of type Γ1 , EW1 [h] (3) achieves a competitive ratio of R[EW1 , 1/3] =
0.751066, regardless of the value h.
Claim 6. For a small edge e of type Γ2 , EW1 [h] (3) achieves a competitive ratio of R[EW1 , 1/3] =
0.72933 + h ∗ 0.040415.
By setting h = 0.537815, the two types of small edges have the same ratio and we get that
EW1 [h] achieves (R[EW1 , 2/3], R[EW 1 , 1/3]) = (0.679417, 0.751066). Thus, this proves Lemma 2.
Proof of Claim 4
Proof. Consider a large edge e = (u, v1 ) in the graph GF . Let e′ = (u, v2 ) be the other small edge
incident to u. Edges e and e′ can appear in [M1 , M2 , M3 ] in the following three ways.
• α1 : e ∈ M1 , e′ ∈ M2 , e ∈ M3 .
26
• α2 : e′ ∈ M1 , e ∈ M2 , e ∈ M3 .
• α3 : e ∈ M1 , e ∈ M2 , e′ ∈ M3 .
Notice that the random triple of matchings [M1 , M2 , M3 ] is generated by invoking PM[F, 3]. From
the property of PM[F, 3], we know that αi will occur with probability 1/3 for 1 ≤ i ≤ 3. For α1
and α2 , we can ignore the second copy of e in M3 and from Lemma 1 we have
Pr[e is matched | α1 ] ≥ 0.580831 and Pr[e is matched | α2 ] ≥ 0.148499
For α3 , we have
Pr[e is matched | α3 ] =
n
X
1
n
1−
n
X
1
t=1
+
t=1
n
+(1 − h)
2
n
t−1
+
n
X
1 t−1
t=1
(t − 1)(t − 2)
2n2
n
1−
n
2
1−
n
!
n
X
t−1
1 1
t=1
n
n3
3
2
n
t−2
t−3
2
1−
n
t−4
≥ 0.621246 + (1 − h) ∗ 0.00892978
Hence, we have
Pr[e is matched] =
3
2
1X
Pr[e is matched | αi ] ≥ R[EW1 , 2/3]
3 i=1
3
where R[EW1 , 2/3] = 0.67529 + (1 − h) ∗ 0.00446489.
Proof of Claims 5 and 6
Proof. Consider a small edge e = (u, v) of type Γ1 . Let e1 and e2 be the two other small edges
incident to u. For a given triple of matchings [M1 , M2 , M3 ], we say e is of type ψ1 if e appears in
M1 while the other two in the remaining two matchings. Similarly, we define the type ψ2 and ψ3
for the case where e appears in M2 and M3 respectively. Notice that the probability that e is of
type ψi , 1 ≤ i ≤ 3 is 1/3.
Similar to the calculations in the proof of Claim 4, we have Pr[e is matched| ψ1 ] ≥ 0.571861,
Pr[e is matched| ψ2 ] ≥ 0.144776 and Pr[e is matched| ψ3 ] ≥ 0.0344288. Therefore we have
Pr[e is matched] =
3
1
1X
Pr[e is matched | ψi ] ≥ R[EW1 , 1/3]
3 i=1
3
where R[EW1 , 1/3] = 0.751066.
Consider a small edge e = (u, v) of type Γ2 , we define type βi , 1 ≤ i ≤ 3, if e appears in
Mi while the large edge e′ incident to u appears in the remaining two matchings. Similarly, we
have Pr[e is matched| ψ1 ] ≥ 0.580831, Pr[e is matched| ψ2 ] ≥ 0.148499 and Pr[e is matched| ψ3 ] ≥
h ∗ 0.0404154.
Hence, the ratio for a small edge of type Γ2 is R[EW1 , 1/3] = 0.72933 + h ∗ 0.0404154.
27
A.1.2
Proof of Lemma 3
We will prove Lemma 3 using the following two Claims.
Claim 7. For a large edge e, EW2 [y1 , y2 ] (5) achieves a competitive ratio of
R[EW2 , 2/3] = min 0.948183 − 0.099895y1 − 0.025646y2 , 0.871245
Claim 8. For a small edge e, EW2 [y1 , y2 ] (5) achieves a competitive ratio of R[EW2 , 1/3] = 0.4455,
when y1 = 0.687, y2 = 1.
Therefore, by setting y1 = 0.687, y2 = 1 we get that R[EW2 , 2/3] = 0.8539 and R[EW2 , 1/3] =
0.4455, which proves Lemma 3.
Proof of Claim 7
Proof. Figure 5 shows the two possible configurations for a large edge.
(A)
(B)
v1
v1
u
u
v2
v2
Figure 5: Diagram of configurations for a large edge e = (u, v1 ). Thin and Thick lines represent
small and large edges respectively.
Consider a large edge e = (u, v1 ) with the configuration (A). From PM∗ [F, 2][y1 , y2 ], we know
that e will always be in M1 while e′ = (u, v2 ) will be in M1 and M2 with probability y1 /3 and y2 /3
respectively.
We now have the following cases
• α1 : e ∈ M1 and e′ ∈ M1 . This happens with probability y1 /3 and Pr[e is matched | α1 ] ≥
0.432332.
• α2 : e ∈ M1 and e′ ∈ M2 . This happens with probability y2 /3 and Pr[e is matched | α2 ] ≥
0.580831.
• α3 : e ∈ M1 and e′ 6∈ M1 , e′ 6∈ M2 . This happens with probability (1 − y1 /3 − y2 /3) and
Pr[e is matched| α1 ] ≥ 0.632121.
28
Therefore we have
y2
Pr[e is matched | α2 ]
3
y1 y2
+(1 −
− ) Pr[e is matched | α3 ]
3
3
2
(0.948183 − 0.099895y1 − 0.025646y2 )
≥
3
Pr[e is matched] =
y
1
3
Pr[e is matched | α1 ] +
Consider the configuration (B). From PM∗ [F, 2][y1 , y2 ], we know that e will always be in M1
and e′ = (u, v2 ) will always be in M2 . Thus we have
Pr[e is matched] = Pr[e is matched | α2 ] =
2
∗ 0.871245
3
Hence, this completes the proof of Claim 7.
Proof of Claim 8
Proof. Figure 6 shows all possible configurations for a small edge.
(1a)
(1b)
v1
u
(2a)
v1
(2b)
v1
u
v2
u
v2
v1
u
v2
v2
v3
v3
(3a)
v1
(3b)
v1
(4a)
v1
(4b)
v1
u
v2
u
v2
u
v2
u
v2
v3
v3
v3
v3
Figure 6: Diagram of configurations for a small edge e = (u, v1 ). Thin and Thick lines represent
small and large edges respectively.
Similar to the proof of Claim 7, we will do a case-by-case analysis on the various configurations.
Let ei = (u, vi ) for 1 ≤ i ≤ 3 and E be the event that e1 gets matched. For a given ei , denote
ei ∈ M0 if ei ∈
/ M1 , ei 6∈ M2 .
• (1a). Observe that e1 ∈ M2 and e2 ∈ M1 . Thus we have Pr[E] =
1
3
∗ 0.44550.
• (1b). Observe that we have two cases: {α1 : e2 ∈ M1 , e1 ∈ M1 } and {α2 : e2 ∈ M1 , e1 ∈ M2 }.
Case α1 happens with probability y1 /3 and the conditional probability is Pr[E |α1 ] = 0.432332.
Case α2 happens with probability y2 /3 and the conditional is Pr[E |α2 ] = 0.148499. Thus we
have
1
Pr[E] = y1 /3 ∗ Pr[E |α1 ] + y2 /3 ∗ Pr[E |α2 ] ≥ (0.432332y1 + 0.148499y2 )
3
29
• (2a). Observe that e1 ∈ M2 , e2 ∈ M2 , e3 ∈ M2 . Pr[E] =
1
3
∗ 0.601704
• (2b). Observe that we have two cases: {α1 : e1 ∈ M1 , e2 ∈ M2 , e3 ∈ M2 } and {α2 : e1 ∈
M2 , e2 ∈ M2 , e3 ∈ M2 }. Case α1 happens with probability y1 /3 and the conditional is
Pr[E |α1 ] = 0.537432. Case α2 happens with probability y2 /3 and conditional is Pr[E |α2 ] =
0.200568. Thus we have
1
Pr[E] = y1 /3 ∗ Pr[E |α1 ] + y2 /3 ∗ Pr[E |α2 ] ≥ (0.537432y1 + 0.200568y2 )
3
• (3a). Observe that we have three cases: {α1 : e1 ∈ M2 , e2 ∈ M1 , e3 ∈ M2 }, {α2 : e1 ∈
M2 , e2 ∈ M2 , e3 ∈ M2 } and {α2 : e1 ∈ M2 , e2 ∈ M0 , e3 ∈ M2 }. Case α1 happens with probability y1 /3 and conditional is Pr[E |α1 ] = 0.13171. Case α2 happens with probability y2 /3
and conditional is Pr[E |α2 ] = 0.200568. Case α3 happens with probability (1 − y1 /3 − y2 /3)
and conditional is Pr[E |α3 ] = 0.22933.
Similarly, we have
Pr[E] = y1 /3 ∗ Pr[E |α1 ] + y2 /3 ∗ Pr[E |α2 ] + (1 − y1 /3 − y2 /3) ∗ Pr[E |α3 ]
1
≥ (0.13171y1 + 0.200568y2 + (3 − y1 − y2 )0.22933)
3
• (3b). Observe that we have six cases.
– α1 : e1 ∈ M1 , e2 ∈ M1 , e3 ∈ M2 . Pr[α1 ] = y12 /9 and Pr[E |α1 ] = 0.4057.
– α2 : e1 ∈ M1 , e2 ∈ M2 , e3 ∈ M2 . Pr[α2 ] = y1 y2 /9 and Pr[E |α2 ] = 0.5374.
– α3 : e1 ∈ M1 , e2 ∈ M0 , e3 ∈ M2 . Pr[α3 ] = y1 /3(1 − y1 /3 − y2 /3) and Pr[E |α3 ] = 0.58083.
– α4 : e1 ∈ M2 , e2 ∈ M1 , e3 ∈ M2 . Pr[α4 ] = y1 y2 /9, Pr[E |α4 ] = 0.1317.
– α5 : e1 ∈ M2 , e2 ∈ M2 , e3 ∈ M2 . Pr[α5 ] = y22 /9, Pr[E |α5 ] = 0.2006.
– α6 : e1 ∈ M2 , e2 ∈ M0 , e3 ∈ M2 . Pr[α6 ] = y2 /3(1 − y1 /3 − y2 /3)/3 and Pr[E |α6 ] =
0.22933.
Therefore we have
Pr[E] ≥
1
0.135241y12 + 0.223033y1 y2 + 0.066856y22
3
+y1 (3 − y1 − y2 )0.193610 + y2 (3 − y1 − y2 )0.076443
30
• (4a). Observe that we have following six cases.
– α1 : e1 ∈ M2 , e2 ∈ M1 , e3 ∈ M1 . Pr[α1 ] = y12 /9 and Pr[E |α1 ] = 0.08898.
– α2 : e1 ∈ M2 , e2 ∈ M2 , e3 ∈ M2 . Pr[α2 ] = y22 /9 and Pr[E |α2 ] = 0.2006.
– α3 : e1 ∈ M2 , e2 ∈ M0 , e3 ∈ M0 . Pr[α3 ] = (1 − y1 /3 − y1 /3)2 , and Pr[E |α3 ] = 0.2642.
– α4 : e1 ∈ M2 while either e2 ∈ M1 , e3 ∈ M2 or e2 ∈ M2 , e3 ∈ M1 . Pr[α2 ] = 2y1 y2 /9 and
Pr[E |α4 ] = 0.1317.
– α5 : e1 ∈ M2 while either e2 ∈ M1 , e3 ∈ M0 or e2 ∈ M0 , e3 ∈ M1 . Pr[α5 ] = 2y1 /3(1 −
y1 /3 − y2 /3) and Pr[E |α5 ] = 0.14849.
– α6 : e1 ∈ M2 while either e2 ∈ M2 , e3 ∈ M0 or e2 ∈ M0 , e3 ∈ M2 . Pr[α5 ] = 2y2 /3(1 −
y1 /3 − y2 /3) and Pr[E |α6 ] = 0.22933.
Therefore we have
Pr[E] ≥
1
0.029661y12 + 2 ∗ 0.043903y1 y2 + 0.066856y22 + 2y1 (3 − y1 − y2 )0.0494997
3
+2y2 (3 − y1 − y2 )(0.076443) + (3 − y1 − y2 )2 0.0880803
• (4b). Observe that in this configuration, we have additional six cases to the ones discussed
in (4a). Let αi be the cases defined in (4a) for each 1 ≤ i ≤ 6. Notice that each Pr[αi ] has a
multiplicative factor of y2 /3. Now, consider the six new cases.
– β1 : e1 ∈ M1 , e2 ∈ M1 , e3 ∈ M1 . Pr[α1 ] = y13 /27 and Pr[E |α1 ] = 0.3167.
– β2 : e1 ∈ M1 , e2 ∈ M2 , e3 ∈ M2 . Pr[α2 ] = y1 y22 /27 and Pr[E |α2 ] = 0.5374.
– β3 : e1 ∈ M1 , e2 ∈ M0 , e3 ∈ M0 . Pr[α3 ] = y1 /3 ∗ (1 − y1 /3 − y2 /3)2 and Pr[E |α3 ] = 0.632.
– β4 : e1 ∈ M1 and either e2 ∈ M1 , e3 ∈ M2 or e2 ∈ M2 , e3 ∈ M1 . Pr[α2 ] = 2y12 y2 /27 and
Pr[E |α4 ] = 0.4057.
– β5 : e1 ∈ M1 and either e2 ∈ M1 , e3 ∈ M0 or e2 ∈ M0 , e3 ∈ M1 . Pr[α5 ] = 2y12 /9 ∗ (1 −
y1 /3 − y2 /3) and Pr[E |α5 ] = 0.4323.
– β6 : e1 ∈ M1 and either e2 ∈ M2 , e3 ∈ M0 or e2 ∈ M0 , e3 ∈ M2 . Pr[α5 ] = 2y1 y2 /9 ∗ (1 −
y1 /3 − y2 /3) and Pr[E |α6 ] = 0.58083.
31
Hence, we have
Pr[E] ≥
1
0.632y1 − 0.133133y12 + 0.0093y13 + 0.264241y2
3
−0.11127y1 y2 + 0.01170y12 y2 − 0.0232746y22 + 0.00488y1 y22 + 0.00068y23
Setting y1 = 0.687, y2 = 1, we get that the competitive ratio for a small edge is 0.44550. The
bottleneck cases are configurations (1a) and (1b).
A.2
A.2.1
Supplemental materials in section 4
Proof of Lemma 7 (Vertex-weighted and Unweighted)
When Hu = 1 and u is in the cycle C1 , [13] show that the competitive ratio of u is 1 − 2e−2 . Hence,
for the remaining cases, we use the following Claims.
Claim 9. If Hu = 1 and u is not in C1 , then we have R[RLA, 1] ≥ 0.735622.
Claim 10. R[RLA, 2/3] ≥ 0.7870.
Claim 11. R[RLA, 1/3] ≥ 0.8107.
Recall that Bu is the event that among the n random lists, there exists a list starting with u
and Gvu is the event that among the n lists, there exist successive lists such that (1) all start with
some u′ which are different from u but are neighbors of v; and (2) they ensure u will be matched.
Notice that Pu is the probability that u gets matched in RLA[H′ ]. For each u, we compute
Pr[Bu ] and Pr[Gvu ] for all possibilities of v ∼ u and using Lemma 6 we get Pu . First, we discuss
two different ways to calculate Pr[Gvu ]. For some cases, we use a direct calculation, while for the
rest we use the Markov-chain approximation method.
u
1−y
y
x
u
v
′
Figure 7: Case 1 in calculation of Pr[Gvu ]
Two ways to compute the value Pr[Gvu ]
1. Case 1: Consider the case when v has two neighbors as shown in Figure 7. Assume v has two
′
′
′
neighbors u and u′ and after modifications, H(u
′ ,v) = y, H(u,v) = 1 − y and Hu′ = x. Thus,
the second certificate event Gvu corresponds to the event (1) a list starting with u′ comes
at some time 1 ≤ i < n; (2) the list Rv = (u′ , u) comes for a second time at some j with
32
1
u1
1/3 u2
2/3 u3
0.75
0.1
v
0.15
Figure 8: Case 2 in calculation of Pr[Gvu ]
i < j ≤ n. Note that the arrival rate of a list starting with u′ is Hu′ ′ = x/n and the rate of
list Rv = (u′ , u) is y/n. Therefore we have
Pr[Gvu ] =
n−1
X
i=1
x/n(1 − x/n)(i−1) (1 − (1 − y/n)(n−i)
x − e−y x + (−1 + e−x )y
(if x 6= y)
x−y
∼ 1 − e−x (1 + x) (if x = y)
∼
(19)
(20)
(21)
2. Case 2: Consider the case when v has three neighbors. The value Pr[Gvu ] is approximated
using the Markov Chain method, similar to [13]. Let us use the following example to illustrate
the method.
Consider the following case as shown in Figure 8 (v has three neighbors u, u1 and u2 with
′
Hu = 1, Hu1 = 1/3 and Hu2 = 2/3). Recall that after modifications, we have H(u
=b=
1 ,v)
′
′
0.1, H(u2 ,v) = c = 0.15 and H(u,v) = d = 0.75. We simulate the process of u getting matched
resulting from several successive random lists starting from either u1 or u2 by an n-step
Markov Chain as follows. We have 5 states: s1 = (0, 0, 0), s2 = (0, 1, 0), s3 = (0, 0, 1), s4 =
(0, 1, 1) and s5 = (1, ∗, ∗) and the three numbers in each triple correspond to u, u1 and u2
being matched(or not) respectively. State s5 corresponds to u being matched; the matched
status of u1 and u2 is irrelevant. The chain initially starts in s1 and the probability of being in
state s5 after n steps gives an approximation to Pr[Gvu ]. The one-step transition probability
matrix M is shown as follows.
c + 1/3
b
, M1,3 =
, M1,1 = 1 − M1,2 − M1,3
n
n
c + 1/3
bc
bd
=
+
, M2,5 =
,
n
(c + d)n
(c + d)n
= 1 − M2,4 − M2,5
cb
cd
b
, M3,5 =
= +
n (b + d)n
(b + d)n
= 1 − M3,4 − M3,5
b+c
, M4,4 = 1 − M4,5
=
n
=1
M1,2 =
M2,4
M2,3
M3,4
M3,3
M4,5
M5,5
Mi,j = 0 for all other i, j
33
2
Notice that M1,3 = c+1/3
and not 3n
since after modifications, the arrival rate of a list starting
n
with u2 decreases correspondingly.
Let us now prove the three Claims 9, 10 and 11. Here we give the explicit analysis for the
case when Hu = 1. For the remaining cases, similar methods can be applied. Hence, we omit the
analysis and only present the related computational results which leads to the conclusion.
Proof of Claim 9
Notice that u is not in the cycle C1 and thus Lemma 6 can be used. Figure 9 describes all
possible cases when a node u ∈ U has Hu = 1. (We ignore all those cases when Hu < 1, since they
will not appear in the WS.)
1
u
1
0.9
u
v
1/3 u1
u
v
2/3 u1
0.1
1
0.85
1 − x1
1
v
x1
0.15
1
u
u1
1 − x2
x2
1
v
u1
v1
(α1 )
1
(α2 )
u
1
1
1
u1
u
u
1
u1
1/3 u2
1
u
1/3 u1
2/3 u2
1
x2
u1
u
0.4
v
1 − x1
1
u1
1 − x2
v
2/3 u1
0.6
v1
0.45
0.45
(β2 )
v
0.1
0.65
0.15
1
u
1
u1
2/3 u2
0.4
0.2
0.2
(β3 )
1
v
u
1/3 u1
1/3 u2
(β6 )
(β5 )
1
u
(α4 )
v
(β1 )
1
1
x1
v
u2
(α3 )
0.8
0.1
0.1
(β7 )
(β4 )
1
v
u
2/3 u1
2/3 u2
0.5
0.25
v
0.25
(β8 )
v
0.2
(β9 )
Figure 9: Vertex-weighted Hu = 1 cases. The value assigned to each edge represents the value after
the second modification. No value indicates no modification. Here, x1 = 0.2744 and x2 = 0.15877.
Let v1 and v2 be the two neighbors of u with H(u,v1 ) = 2/3 and H(u,v2 ) = 1/3. In total, there
are 4 × 10 combinations, where v1 is chosen from some αi , 1 ≤ i ≤ 4 and v2 is chosen from some
βi , 1 ≤ i ≤ 9. For Hu = 1, we need to find the worst combination among these such that the value
34
Pu is minimized. We can find this WS using the Lemma 6.
Q
For each type of αi , βj , we compute the values it will contribute to the term (1 − Bu ) v∼u (1 −
Pr[Gvu ]). For example, assume v1 is of type α1 , denoted by v1 (α1 ). It contributes e−0.9 to the term
Q
(1 − Bu ) and (1 − Pr[Gvu1 ]) to v∼u (1 − Pr[Gvu ]), thus the total value it contributes is γ(v1 , α1 ) =
e−0.9 (1 − Pr[Gvu1 ]). Similarly, we can compute all γ(v1 , αi ) and γ(v2 , βj ). Let i∗ = arg maxi γ(v1 , αi )
and j ∗ = arg maxj γ(v2 , βj ). The WS is for the combination {v1 (αi∗ ), v2 (βj ∗ )} and the resulting
value of Pu and R[RLA, 1] is as follows:
Pu = 1 − γ(v1 , αi∗ )γ(v2 , βj ∗ )
R[RLA, 1] = Pu /Hu = Pu
Here is a list of γ(v1 , αi ) and γ(v2 , βj ), for each 1 ≤ i ≤ 4 and 1 ≤ j ≤ 9.
• α1 : We have Pr[Gvu ] = 1 − e−0.1 ∗ 1.1 and γ(v, α1 ) = e−0.1 ∗ 1.1 ∗ e−0.9 = 0.404667.
• α2 : Pr[Gvu ] ≥ 1 − e−0.15 ∗ 1.15 and γ(v, α2 ) ≤ 0.423.
Notice that after modifications, Hu′ 1 ≥ 0.15. Hence, we use this and Equation 19 to compute
the lower bound of Pr[Gvu ].
• α3 : Pr[Gvu ] ≥ 0.0916792 and γ(v, α3 ) ≤ 0.439667.
Notice that for any large edge e incident to a node u with Hu = 1 (before modification), we
′
≥ 0.7256 and
have after modification, He′ ≥ 1 − 0.2744 = 0.7256. Thus we have H(u
1 ,v1 )
v
′
Hu1 ≥ 1. From Equation 19, we get Pr[Gu ] ≥ 0.0916792.
• α4 : Pr[Gvu ] ≥ 0.0307466 and γ(v, α4 ) ≤ 0.417923.
Notice that for any small edge e incident to a node u with Hu = 1 (before modification), we
have after modification, He′ ≥ 0.15877. Thus, we have Hu′ 1 ≥ 3 ∗ 0.15877.
• β1 : Pr[Gvu ] = 0.1608 and γ(v, β1 ) = 0.601313.
• β2 : Pr[Gvu ] ≥ 0.208812 and γ(v, β2 ) ≤ 0.601313.
′
After modifications, we have H(u
≥ 0.2744 and thus we get Hu′ 1 ≥ 1.
1 ,v1 )
• β3 : Pr[Gvu ] ≥ 0.251611 and γ(v, β2 ) ≤ 0.63852.
′
After modifications, we have H(u
≥ 0.2744 and thus we get Hu′ 1 ≥ 1 − 0.15877 + 0.2744.
1 ,v1 )
• β4 : Pr[Gvu ] = 0.121901 and γ(v, β4 ) = 0.588607.
35
• β5 : Pr[Gvu ] = 0.1346 and γ(v, β5 ) = 0.551803.
• β6 : Pr[Gvu ] ≥ 0.1140 and γ(v, β6 ) ≤ 0.593904.
• β7 : Pr[Gvu ] = 0.0084 and γ(v, β7 ) = 0.4455.
• β8 : Pr[Gvu ] ≥ 0.0397 and γ(v, β8 ) ≤ 0.582451.
• β9 : Pr[Gvu ] ≥ 0.0230 and γ(v, β9 ) ≤ 0.510039.
Using the computed values above, let us compute the ratio of a node u with Hu = 1.
• If u has three neighbors, then the WS configuration is when each of the three neighbors of u
is of type β3 . This is because, the value of γ(v, β3 ) is the largest. The resultant ratio is 0.73967.
• If u has two neighbors, then the WS configuration is when one of the neighbor is of type β1
(or β2 ) and the other is of type α3 . The resultant ratio is 0.735622.
Proof of Claim 10
The proof is similar to that of Claim 9. The Figure 10 shows all possible configurations of a
node u with Hu = 2/3. Note that the WS cannot have F (v) < 1 and hence we omit them here.
For a neighbor v of u, if H(u,v) = 2/3, then v is in one of αi , 1 ≤ i ≤ 3; if H(u,v) = 1/3, then v is in
one of βi , 1 ≤ i ≤ 8. We now list the values γ(v, αi ) and γ(v, βj ), for each 1 ≤ i ≤ 3 and 1 ≤ j ≤ 8.
• α1 : We have Pr[Gvu ] = 1 − e−0.25 ∗ 1.25 and γ(v, α1 ) = e−0.25 ∗ 1.25 ∗ e−0.75 = 0.459849.
• α2 : We have Pr[Gvu ] ≥ 0.0528016 and γ(v, α1 ) ≤ 0.470365.
• α3 . We have Pr[Gvu ] ≥ 0.13398 and γ(v, α3 ) ≤ 0.475282.
• β1 : We have Pr[Gvu ] = 1 − e−0.7 ∗ 1.7 and γ(v, β1 ) = 0.625395.
• β2 : We have Pr[Gvu ] ≥ 0.226356 and γ(v, β2 ) ≤ 0.665882.
• β3 : We have Pr[Gvu ] ≥ 0.1819 and γ(v, β3 ) ≤ 0.669804.
• β4 : We have Pr[Gvu ] ≥ 0.1130 and γ(v, β4 ) ≤ 0.635563.
36
2/3 u
2/3 u
0.75
v
1/3 u1
(α1 )
2/3 u
0.3
2/3 u
0.7
1
2/3 u
2/3 u
v
1
u1
1/3 u2
(β5 )
1
u1
1
u2
0.85
(β2 )
2/3 u
1/3 u2
u1
0.2
0.65
0.15
0.4
(α3 )
v
u1
(β1 )
1/3 u1
1
0.15
v
2/3 u1
v
(α2 )
0.3
0.6
v
2/3 u1
0.25
2/3 u
0.7
0.2
0.4
2/3 u
v
0.4
2/3 u1
2/3 u2
(β3 )
2/3 u
v
1
u1
2/3 u2
(β6 )
0.25
0.5
0.25
v
(β4 )
2/3 u
v
1/3 u1
v
2/3 u2
(β7 )
(β8 )
Figure 10: Vertex-weighted Hu = 2/3 cases. The value assigned to each edge represents the value
after the second modification. No value indicates no modification.
• β5 : We have Pr[Gvu ] ≥ 0.0587 and γ(v, β5 ) ≤ 0.674471.
• β6 : We have Pr[Gvu ] ≥ 0.1688 and γ(v, β6 ) ≤ 0.680529.
• β7 : We have Pr[Gvu ] ≥ 0.1318 and γ(v, β7 ) ≤ 0.676155.
• β8 : We have Pr[Gvu ] ≥ 0.0587 and γ(v, β8 ) ≤ 0.674471.
Hence, the WS structure is when u is such that Hu = 2/3 and has one neighbor of type α3 . The
resultant ratio is 0.7870.
Proof of Claim 11
The Figure 11 shows the possible configurations of a node u with Hu = 1/3. Again, we omit
those cases where Hv < 1.
We now list the values γ(v, αi ), for each 1 ≤ i ≤ 8.
• α1 : We have Pr[Gvu ] = 1 − e−0.75 ∗ 1.75 and γ(v, α1 ) = 0.643789.
• α2 : We have Pr[Gvu ] ≥ 0.282256 and γ(v, α2 ) ≤ 0.649443.
37
1/3 u
1/3 u
1/3 u
0.25
0.1
v
2/3 u1
v
0.75
1
u1
(α1 )
1/3 u
v
2/3 u2
u1
1
u2
0.9
(α2 )
1/3 u
2/3 u1
1
1
u1
2/3 u2
(α5 )
0.15
0.65
0.2
0.1
0.45
1/3 u
v
0.45
1/3 u1
1/3 u2
(α3 )
1/3 u
v
1
u1
0.1
0.8
0.1
1/3 u2
(α6 )
(α7 )
v
(α4 )
1/3 u
v
1/3 u1
v
2/3 u2
(α8 )
Figure 11: Vertex-weighted Hu = 1/3 cases. The value assigned to each edge represents the value
after the second modification. No value indicates no modification.
• α3 : We have Pr[Gvu ] ≥ 0.1935 and γ(v, α3 ) ≤ 0.729751.
• α4 : We have Pr[Gvu ] ≥ 0.0587 and γ(v, α4 ) ≤ 0.674471.
• α5 : γ(v, α5 ) ≤ 0.674471.
• α6 : We have Pr[Gvu ] ≥ 0.1546 and γ(v, α6 ) ≤ 0.727643.
• α7 : We have Pr[Gvu ] ≥ 0.1938 and γ(v, α7 ) ≤ 0.72948.
• α8 : γ(v, α8 ) ≤ 0.674471.
Hence, the WS for node u with Hu = 1/3 is when u has one neighbor of type α3 . The resultant
ratio is 0.8107.
38
This figure "JailletFig7P26.jpg" is available in "jpg" format from:
http://arxiv.org/ps/1606.06395v3
| 8cs.DS
|
On the tail behavior of a class of multivariate
conditionally heteroskedastic processes
Rasmus Søndergaard Pedersen∗ and Olivier Wintenberger†
arXiv:1701.05091v3 [math.ST] 5 Dec 2017
October 20, 2017
Abstract
Conditions for geometric ergodicity of multivariate autoregressive conditional heteroskedasticity (ARCH) processes, with the so-called BEKK (Baba,
Engle, Kraft, and Kroner) parametrization, are considered. We show for a
class of BEKK-ARCH processes that the invariant distribution is regularly
varying. In order to account for the possibility of different tail indices of the
marginals, we consider the notion of vector scaling regular variation (VSRV),
closely related to non-standard regular variation. The characterization of the
tail behavior of the processes is used for deriving the asymptotic properties
of the sample covariance matrices.
AMS 2010 subject classifications: 60G70, 60G10, 60H25, 39A50.
Keywords and phrases: Stochastic recurrence equations, Markov processes, regular variation, multivariate ARCH, asymptotic properties, geometric ergodicity.
1
Introduction
The aim of this paper is to investigate the tail behavior of a class of multivariate
conditionally heteroskedastic processes. Specifically, we consider the BEKK-ARCH
∗
Department of Economics, University of Copenhagen. Øster Farimagsgade 5, DK-1353 Copenhagen K, Denmark. † Department of Mathematical Sciences, University of Copenhagen, Universitetsparken 5, DK-2100 Copenhagen Ø, Denmark & Sorbonne Universités, UPMC Univ. Paris
06, LSTA, Case 158 4 Place Jussieu, 75005 Paris, France.
We are grateful for comments and suggestions from the editor-in-chief (Thomas Mikosch), an associate editor, and two referees, which have led to a much improved manuscript. Moreover, we
thank Sebastian Mentemeier for valuable comments. Pedersen greatly acknowledges funding from
the Carlsberg Foundation. Financial support by the ANR network AMERISKA ANR 14 CE20
0006 01 is gratefully acknowledged by Wintenberger. Correspondence to: Rasmus S. Pedersen,
[email protected].
1
(or BEKK(1,0,l)) process, introduced by Engle and Kroner (1995), satisfying
1/2
t∈N
Xt = Ht Zt ,
Ht = C +
l
X
|
Ai Xt−1 Xt−1
A|i ,
(1.1)
(1.2)
i=1
with (Zt : t ∈ N) i.i.d., Zt ∼ N (0, Id ), C a d × d positive definite matrix, A1 , ..., Al ∈
M (d, R) (the set of d × d real matrices), and some initial value X0 . Due to the
assumption that Zt is Gaussian, it holds that Xt can be written as the stochastic
recurrence equation (SRE)
Xt = M̃t Xt−1 + Qt ,
(1.3)
with
M̃t =
l
X
mit Ai
(1.4)
i=1
and (mit : t ∈ N) is an i.i.d. process mutually independent of (mjt : t ∈ N) for i 6= j,
with mit ∼ N (0, 1). Moreover (Qt : t ∈ N) is an i.i.d. process with Qt ∼ N (0, C)
mutually independent of (mit : t ∈ N) for all i = 1, ..., l.
To our knowledge, the representation in (1.3)-(1.4) of the BEKK-ARCH process
is new. Moreover, the representation will be crucial for studying the stochastic
properties of the process. Firstly, we find a new sufficient condition in terms of the
matrices A1 , ..., Al in order for (Xt : t ≥ 0) to be geometrically ergodic. In particular,
for the case l = 1, we derive a condition directly related to the eigenvalues of A1 , in
line with the strict stationarity condition found by Nelson (1990) for the univariate
ARCH(1) process. This condition is milder compared to the conditions found in the
existing body of literature on BEKK-type processes. Secondly, the representation
is used to characterize the tails of the stationary solution to (Xt : t ∈ N).
Whereas the tail behavior of univariate GARCH processes is well-established, see
e.g. Basrak et al. (2002b), few results on the tail behavior of multivariate GARCH
processes exist. Some exceptions are the multivariate constant conditional correlation (CCC) GARCH processes, see e.g. Stărică (1999), Pedersen (2016), and Matsui
and Mikosch (2016), and a class of factor GARCH processes, see Basrak and Segers
(2009). This existing body of literature relies on rewriting the (transformed) process
on companion form that obeys a non-negative multivariate SRE. The characterization of the tails of the processes then follows by an application of Kesten’s Theorem
(Kesten (1973)) for non-negative SREs. Such approach is not feasible when analyz-
2
ing BEKK-ARCH processes, as these are stated in terms of an Rd -valued SRE in
(1.3). For some special cases of the BEKK-ARCH process, we apply existing results
for Rd -valued SREs in order to show that the stationary distribution for the BEKKARCH process is multivariate regularly varying. Specifically, when the matrix M̃t
in (1.4) is invertible (almost surely) and has a law that is absolutely continuous
with respect to the Lebesgue measure on M (d, R) (denoted ID BEKK-ARCH) we
argue that the classical results of Kesten (1973, Theorem 6), see also Alsmeyer and
Mentemeier (2012), apply. Moreover, when M̃t is the product of a positive scalar
and a random orthogonal matrix (denoted Similarity BEKK-ARCH) we show that
the results of Buraczewski et al. (2009) apply. Importantly, we do also argue that the
results of Alsmeyer and Mentemeier (2012) rely on rather restrictive conditions that
can be shown not to hold for certain types of BEKK-ARCH processes, in particular
the much applied process where l = 1 and A1 is diagonal, denoted Diagonal BEKKARCH. Specifically, and as ruled out in Alsmeyer and Mentemeier (2012), we show
that the Diagonal BEKK-ARCH process exhibits different marginal tail indices, i.e.
P(±Xt,i > x)/ci x−αi → 1 as x → ∞ for some constant ci > 0, i = 1, ..., d (denoted
Condition M). In order to analyze this class of BEKK-ARCH processes, where the
tail indices are allowed to differ among the elements of Xt , we introduce a new notion of vector scaling regular variation (VSRV) distributions, based on element-wise
scaling of Xt instead of scaling by an arbitrary norm of Xt . We emphasize that
the notion of VSRV is similar to the notion of non-standard regular variation (see
Resnick (2007, Chapter 6)) under the additional Condition M. In addition, in the
spirit of Basrak and Segers (2009), we introduce the notion of VSRV processes with
particular attention to Markov chains and characterize their extremal behavior. We
argue that the stationary distribution of the Diagonal BEKK-ARCH process is expected to be VSRV, which is supported in a simulation study. Proving that the
VSRV property applies requires that new multivariate renewal theory is developed,
and we leave such task for future research.
The rest of the paper is organized as follows. In Section 2, we state sufficient
conditions for geometric ergodicity of the BEKK-ARCH process and introduce the
notion of vector-scaling regular varying (VSRV) distributions. We show that the
distribution of Xt satisfies this type of tail-behavior, under suitable conditions. In
Section 3 we introduce the notion of VSRV processes and state that certain BEKKARCH processes satisfy this property. Moreover, we consider the extremal behavior
of the process, in terms of the asymptotic behavior of maxima and extremal indices.
Lastly, we consider the convergence of point processes based on VSRV processes. In
Section 4, we consider the limiting distribution of the sample covariance matrix of
3
Xt , which relies on point process convergence. Section 5 contains some concluding
remarks on future research directions.
Notation: Let GL(d, R) denote the set of d × d invertible real matrices. With
M (d, R) the set of d × d real matrices and A ∈ M (d, R), let ρ(A) denote the spectral
radius of A. With ⊗ denoting the Kronecker product, for any real matrix A let
A⊗p = A⊗A⊗· · ·⊗A (p factors). For two matrices, A and B, of the same dimension,
A B denotes the elementwise product of A and B. Unless stated otherwise, k · k
denotes an arbitrary matrix norm. Moreover, Sd−1 = {x ∈ Rd : kxk = 1}. For
two matrices A and B of the same dimensions, A
B means that Aij > Bij for
some i, j. For two positive functions f and g, f (x) ∼ g(x), if limx→∞ f (x)/g(x) = 1.
Let L(X) denote the distribution of X. By default, the mode of convergence for
distributions is weak convergence.
2
2.1
Stationary solution of the BEKK-ARCH model
Existence and geometric ergodicity
We start out by stating the following theorem that provides a sufficient condition
for geometric ergodicity of the BEKK-ARCH process. To our knowledge, this result
together with Proposition 2.3 below are new.
Theorem 2.1. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4), suppose that
(
inf
n∈N
"
1
E log
n
n
Y
!#)
M̃t
< 0.
(2.1)
t=1
Then (Xt : t = 0, 1, ...) is geometrically ergodic, and for the associated stationary
solution, E[kXt ks ] < ∞ for some s > 0.
The proof of the theorem follows by (Alsmeyer, 2003, Theorems 2.1-2.2, Example
2.6.d, and Theorem 3.2) and is hence omitted.
Remark 2.2. A sufficient condition for the existence of finite higher-order moments
of Xt can be obtained from Theorem 5 of Feigin and Tweedie (1985). In particular, if ρ(E[M̃t⊗2n ]) < 1 for some n ∈ N, then, for the strictly stationary solution,
P
2
E[kXt k2n ] < ∞. For example, ρ( li=1 A⊗2
i ) < 1 implies that E[kXt k ] < ∞. This
result complements Theorem C.1 of Pedersen and Rahbek (2014) that contains conditions for finite higher-order moments for the case l = 1.
For the case where M̃t contains only one term, i.e. l = 1, the condition in (2.1)
simplifies and a condition for geometric ergodicity can be stated explicitly in terms
of the eigenvalues of the matrix A1 :
4
Proposition 2.3. Let Xt satisfy (1.1)-(1.2) with l = 1 and let A := A1 . Then a
necessary and sufficient condition for (2.1) is that
1
ρ(A) < exp
[−ψ(1) + log(2)] = 1.88736...,
2
(2.2)
where ψ(·) is the digamma function.
Proof. The condition (2.1) holds if and only if there exists n ∈ N such that
n
Y
"
E log
!#
< 0.
M̃t
(2.3)
t=1
Let mt := m1t . It holds that
"
E log
n
Y
!#
"
= E log
M̃t
A
n
n
Y
!#
mt
t=1
t=1
= log (kAn k) − nE [− log(|mt |)]
1
n
= log (kA k) − n
[−ψ(1) + log(2)] ,
2
and hence (2.3) is satisfied if
log kAn k1/n <
1
[−ψ(1) + log(2)] .
2
The result now follows by observing that kAn k1/n → ρ(A) as n → ∞.
Remark 2.4. It holds that ρ(A⊗2 ) = (ρ(A))2 . Hence the condition in (2.2) is equivalent to
1
1
ρ(A ) < exp {−ψ(1) + log(2)} = exp −ψ
2
2
⊗2
= 3.56...,
which is similar to the strict stationary condition found for the ARCH coefficient of
the univariate ARCH(1) process with Gaussian innovations; see Nelson (1990).
Boussama et al. (2011) derive sufficient conditions for geometric ergodicity of the
P
P
|
GARCH-type BEKK process, where Ht = C + pi=1 Ai Xt−i Xt−i
A|i + qj=1 Bj Ht−j Bj| ,
Ai , Bj ∈ M (d, R), i = 1, ..., p, j = 1, ..., q. Specifically, they show that a sufficient
P
Pq
⊗2
condition is ρ( pi=1 A⊗2
i +
j=1 Bj ) < 1. Setting p = 1 and q = 0, this condition
simplifies to ρ(A⊗2
1 ) < 1, which is stronger than the condition derived in (2.2).
Below, we provide some examples of BEKK-ARCH processes that are geometrically ergodic and that will be studied in detail throughout this paper.
5
Example 2.5 (ID BEKK-ARCH). Following Alsmeyer and Mentemeier (2012),
we consider BEKK processes with corresponding SRE’s satisfying certain irreducibility and density conditions (ID), that is conditions (A4)-(A5) in Section A.1 in the
appendix. Specifically, we consider the bivariate BEKK-ARCH process in (1.1)-(1.2)
with
Ht = C +
4
X
|
A|i ,
Ai Xt−1 Xt−1
i=1
where
a1 0
A1 =
0 0
0 0
A2 =
,
a2 0
0 a3
A3 =
,
0 0
0 0
A4 =
(2.4)
0 a4
and
a1 , a2 , a3 , a4 6= 0.
(2.5)
Writing Xt as an SRE, we obtain
Xt = M̃t Xt−1 + Qt ,
(2.6)
with
M̃t =
4
X
Ai mit
(2.7)
i=1
where (m1t ), (m2t ), (m3t ), (m4t ) are mutually independent i.i.d. processes with mit ∼
N (0, 1). Assuming that a1 , a2 , a3 , a4 are such that the top Lyapunov exponent of
(M̃t ) is strictly negative, we have that the process is geometrically ergodic.
Notice that one could consider a more general d-dimensional process with the
same structure as in (2.4)-(2.7), but with M̃t containing d2 terms such that M̃t has
a Lebesgue density on M (d, R), as clarified in Example 2.10 below. Moreover, one
could include additional terms to M̃t , say a term containing a full matrix A or an
autoregressive term, as presented in Remark 2.8 below. We will focus on the simple
bivariate process, but emphasize that our results apply to more general processes.
Example 2.6 (Similarity BEKK-ARCH). Consider the BEKK process in (1.1)(1.2) with l = 1 and A := A1 = aO, where a is a positive scalar and O is an
orthogonal matrix. This implies that the SRE (1.3) has M̃t = amt O. By definition,
6
M̃t is a similarity with probability one, where we recall that a matrix is a similarity
if it can be written as a product of a positive scalar and an orthogonal matrix. From
Proposition 2.3, we have that if a < exp {(1/2) [−ψ(1) + log(2)]} = 1.88736..., then
the process is geometrically ergodic. An important process satisfying the similarity
|
property is the well-known scalar BEKK-ARCH process, where Ht = C +aXt−1 Xt−1
,
√
a > 0. Here A = aId , with Id the identity matrix.
Example 2.7 (Diagonal BEKK-ARCH). Consider the BEKK-ARCH process in
(1.1)-(1.2) with l = 1 such that A := A1 is diagonal. We refer to this process as the
Diagonal BEKK-ARCH process. Relying on Proposition 2.3, the process is geometrically ergodic, if each diagonal element of A is less than exp {(1/2) [−ψ(1) + log(2)]} =
1.88736... in modulus.
As discussed in Bauwens et al. (2006), diagonal BEKK models are typically used
in practice, e.g. within empirical finance, due to their relatively simple parametrization. As will be shown below, even though the parametrization is simple, the tail
behavior is rather rich in the sense that each marginal of Xt has different tail indices,
in general.
Remark 2.8. As an extension to (1.1)-(1.2), one may consider the autoregressive
BEKK-ARCH (AR BEKK-ARCH) process
1/2
Xt = A0 Xt−1 + Ht Zt ,
Ht = C +
l
X
t∈N
|
A|i ,
Ai Xt−1 Xt−1
i=1
with A0 ∈ M (R, d). This process has recently been studied and applied by Nielsen
and Rahbek (2014) for modelling the term structure of interest rates. Notice that
the process has the SRE representation
Xt = M̃t Xt−1 + Qt ,
M̃t = A0 +
l
X
mit Ai .
i=1
Following the arguments used for proving Theorem 2.1, it holds that the AR BEKKARCH process is geometrically ergodic if condition (2.1) is satisfied. Interestingly,
as verified by simulations in Nielsen and Rahbek (2014) the Lyapunov condition
may hold even if the autoregressive polynomial has unit roots, i.e. if A0 = Id + Π,
where Π ∈ M (R, d) has reduced rank.
7
2.2
Multivariate regularly varying distributions
The stationary solution of the BEKK-ARCH process (see Theorem 2.1) can be
written as
Xt =
∞ Y
i
X
M̃t−j+1 Qt−i ,
t ∈ Z.
(2.8)
i=0 j=1
Even if the random matrices M̃t are light-tailed under the Gaussian assumption, the
Q
maximum of the products ( Tt=1 M̃t )T ≥0 may exhibit heavy tails when T → ∞. More
precisely, the tails of the stationary distribution are suspected to have an extremal
behavior as a power law function: For any u ∈ Sd−1 ,
P(u| X0 > x) ∼ C(u)x−α(u) ,
x → ∞,
(2.9)
with α(u) > 0 and C(u0 ) > 0 for some u0 ∈ Sd−1 . The cases where α(u) = α and
C(u) > 0 for all u ∈ Sd−1 are referred as Kesten’s cases, because of the seminal
paper Kesten (1973), and are the subject of the monograph by Buraczewski et al.
(2016). A class of multivariate distributions satisfying this property is the class of
multivariate regularly varying distributions (de Haan and Resnick (1977)):
Definition 2.9. Let R̄d0 := R̄d \{0}, R̄ := R∪{−∞, ∞}, and B̄0d be the Borel σ-field
of R̄d0 . For an Rd -valued random variable X and some constant scalar x > 0, define
µx (·) := P(x−1 X ∈ ·)/P(kXk > x). Then X and its distribution are multivariate
regularly varying if there exists a non-null Radon measure µ on B̄0d which satisfies
µx (·) → µ(·)
vaguely, as x → ∞.
(2.10)
For any µ-continuity set C and t > 0, µ(tC) = t−α µ(C), and we refer to α as the
index of regular variation.
We refer to de Haan and Resnick (1977) for the notion of vague convergence and
additional details. Below, we provide two examples of multivariate regularly varying
BEKK processes.
Example 2.10 (ID BEKK-ARCH, continued). Consider the ID BEKK-ARCH
process (2.4)-(2.7) from Example 2.5. By verifying conditions (A1)-(A7) of Theorem 1.1 of Alsmeyer and Mentemeier (2012), stated in Section A.1 in the appendix,
we establish that the process is multivariate regularly varying.
Since (m1t , m2t , m3t , m4t ) and Qt are Gaussian, we have that (A1)-(A2) hold.
8
Moreover,
a1 m1t a3 m3t
M̃t =
a2 m2t a4 m4t
(2.11)
is invertible with probability one, which ensures that (A3) is satisfied. From (2.11)
we also notice that the distribution of M̃t has a Lebesgue density on M (d, R) which
is strictly positive in a neighborhood of I2 . This ensures that the irreducibility and
density conditions (A4)-(A5) are satisfied. The fact that Qt ∼ N (0, C) and independent of M̃t implies that condition (A6) holds. Lastly, condition (A7) holds by
the fact that (m1t , m2t , m3t , m4t ) and Qt are Gaussian. By Theorem 1.1 of Alsmeyer
and Mentemeier (2012) we have established the following proposition:
Proposition 2.11. Let Xt satisfy (2.4)-(2.7) such that the top Lyapunov exponent
of (M̃t ) is strictly negative. Then for the stationary solution (Xt ), there exists α > 0
such that
lim tα P(x| X0 > t) = K(x),
t→∞
x ∈ S1 ,
(2.12)
for some finite, positive, and continuous function K on S1 .
The proposition implies that each marginal of the distribution of X0 is regularly
varying of order α. By Theorem 1.1.(ii) of Basrak et al. (2002a), we conclude that
X0 is multivariate regularly varying whenever α is a non-integer. Moreover, since
X0 is symmetric, the multivariate regular variation does also hold if α is an odd
integer, see Remark 4.4.17 in Buraczewski et al. (2016).
The proposition does also apply if a1 = 0 or a4 = 0. This can be seen by
Q
observing that nk=1 M̃k has a strictly positive density on M (d, R) for n sufficiently
large, which is sufficient for establishing conditions (A4)-(A5).
Example 2.12 (Similarity BEKK-ARCH, continued). The Similarity BEKKARCH, introduced in Example 2.6, fits into the setting of Buraczewski et al. (2009),
see also Section 4.4.10 of Buraczewski et al. (2016). Specifically, using the representation M̃t = a|mt |sign(mt )O, we have that
(i) E[log(|mt a|)] < 0 if a < exp {(1/2) [−ψ(1) + log(2)]},
(ii) P(M̃t x + Qt = x) < 1 for any x ∈ Rd , and
(iii) log(|amt |) has a non-arithmetic distribution.
Then, due to Theorem 1.6 of Buraczewski et al. (2009), we have the following proposition:
9
Proposition 2.13. Let Xt satisfy (1.1)-(1.2) with l = 1 such that A := A1 = aO,
where a > 0 and O is an orthogonal matrix. If a < exp {(1/2) [−ψ(1) + log(2)]} =
1.88736..., then the process has a unique strictly stationary solution (Xt ) with Xt
multivariate regularly varying with index α > 0 satisfying E[(|mt |a)α ] = 1.
In the following example, we clarify that the Diagonal BEKK-ARCH process, introduced in Example 2.7, does not satisfy the conditions of Theorem 1.1 of Alsmeyer
and Mentemeier (2012). Moreover, we argue that the marginals may have different
tail indices, which motivates the notion of vector scaling regular variation, introduced in the next section.
Example 2.14 (Diagonal BEKK-ARCH, continued). Consider the diagonal
BEKK-ARCH process in Example 2.7, i.e. (1.1)-(1.2) with l = 1 such that A := A1
is diagonal, mt := m1t , and Mt := M̃t = mt A. For this process, the distribution of
Mt is too restricted to apply the results by Alsmeyer and Mentemeier (2012), as in
Example 2.10. Specifically, the irreducibility condition (A4) in Appendix A.1 can
be shown not to hold, as clarified next. It holds that
(
P kx
|
n
Y
Mk k
−1
x
|
k=1
n
Y
!
)
Mk ∈ U
(
= P |
k=1
n
Y
−1
|
n −1
mk | kx A k
k=1
(
= P sign
n
Y
!
)
|
n
mk x A ∈ U
k=1
n
Y
!
)
mk kx| An k−1 x| An ∈ U .
k=1
Hence for any x ∈ Sd−1 we can always find a non-empty open U ⊂ Sd−1 such that
(
max P sign
n∈N
n
Y
!
)
|
n −1 |
n
mk kx A k x A ∈ U
= 0.
(2.13)
k=1
As an example, for d = 2, choose x = (1, 0)| . Then kx| An k−1 x| An ∈ {(−1, 0)} ∪
{(1, 0)} for any n ∈ N. We conclude that condition (A4) does not hold for the
diagonal BEKK-ARCH process.
Note that, each element of Xt = (Xt,1 , ..., Xt,d )| of the diagonal BEKK-ARCH
process can be written as an SRE,
t ∈ Z,
Xt,i = Aii mt Xt−1,i + Qt,i ,
i = 1, . . . , d.
By Theorem 4.1 of Goldie (1991), the stationary solution of the marginal equation
exists if and only if E[log(|Aii m0 |)] < 0. In that case there exists a unique αi > 0
such that E[|m0 |αi ] = |Aii |−αi and
P(±X0,i > x) ∼ ci x−αi
where
ci =
10
E[|X1,i |αi − |Aii m1 X0,i |αi ]
.
2αi E[|Aii m1 |αi log(|Aii m1 |)]
Hence each marginal of X0 may in general have different tail indices. More precisely,
the tail indices are different if the diagonal elements of A, i.e. the Aii s, are, and the
heaviest marginal tail index αi0 corresponds to the largest diagonal coefficient Ai0 i0 .
When i0 is unique, i.e. αi0 < αi for all i = 1, ..., d except i 6= i0 , the distribution X0
can be considered as multivariate regularly varying with index αi0 and with a limit
measure µ with degenerate marginals i 6= i0 .
2.3
Vector scaling regularly varying distributions
The previous Example 2.14 shows that the Diagonal BEKK-ARCH process fits into
the case where α(u) in (2.9) is non-constant. Such cases have not attracted much
attention in the existing body of literature. However, recent empirical studies, such
as Matsui and Mikosch (2016), see also Damek et al. (2017), may suggest that it is
more realistic to consider different marginal tail behaviors when modelling multidimensional financial observations. The idea is to use a vector scaling instead of the
scaling P(kXk > x) in Definition 2.9 that reduced the regular variation properties
of the vector X to the regular variation properties of the norm kXk only. More
precisely, let (Xt ) be a stationary process in Rd and let x = (x1 , . . . , xd )| ∈ Rd .
−1 |
Denote also x−1 = (x−1
1 , . . . , xd ) .
In our framework, we consider distributions satisfying the following condition:
Condition M Each marginal of X0 is regularly varying of order αi > 0, i = 1, ..., d.
The slowly varying functions `i (t) → ci > 0 as t → ∞, i = 1, ..., d.
Indeed, the Diagonal BEKK-ARCH process introduced in Example 2.14 satisfies
Condition M. Moreover, any regularly varying distribution satisfying the Kesten
property (2.9) satisfies Condition M. In particular, the ID and Similarity BEKKARCH processes, introduced in Examples 2.5 and 2.6 respectively, satisfy Condition
M.
We introduce the notion of vector scaling regular variation as the nonstandard
regular variation of the book of Resnick (2007) under Condition M, extended to
negative components (Resnick, 2007, Sections 6.5.5-6.5.6):
Definition 2.15. The distribution of the vector X0 is vector scaling regularly varying (VSRV) if and only if it satisfies Condition M and it is non-standard regularly
varying, i.e. there exists a normalizing sequence x(t) and a Radon measure µ with
non-null marginals such that
tP(x(t)−1
X0 ∈ ·) → µ(·),
11
vaguely.
(2.14)
The usual way of analyzing non-standard regularly varying vectors is to consider a componentwise normalization that is standard regularly varying in the sense
of Definition 2.9. Specifically, when X0 = (X0,1 , ..., X0,d )| satisfies Definition 2.15,
−1
α1
αd |
(c−1
1 (X0,1 /|X0,1 |)|X0,1 | , ..., cd (X0,d /|X0,d |)|X0,d | ) satisfies Definition 2.9 with index one. Throughout we find it helpful to focus on the non-normalized vector X0 in
order to preserve the multiplicative structure of the tail chain introduced in Section
3.2 below, which is used for analyzing the extremal properties of VSRV processes.
In the following proposition we state the VSRV vector X0 has a polar decomposition. In the case where Condition M is not satisfied, note that the polar decomposition holds on a transformation of the original process. Under Condition M, the
natural radius notion is k · kα , where
αi
kxkα := max c−1
i |xi | .
1≤i≤d
(2.15)
Notice that the homogeneity of k · kα , due to Condition M, will be essential for the
proof.
Proposition 2.16. Suppose that the vector X0 satisfies Condition M. Then X0 is
VSRV if and only if there exists a tail vector Y0 ∈ Rd with non-degenerate marginals
such that
L(((ci t)−1/αi )1≤i≤d X0 | kX0 kα > t) →t→∞ L(Y0 ),
(2.16)
where k · kα is defined in (2.15). Moreover, kY0 kα is standard Pareto distributed.
Notice that a similar vector scaling argument has been introduced in Lindskog
et al. (2014).
Proof. Adapting Theorem 4 of de Haan and Resnick (1977), the definition of vector
scaling regularly varying distribution of X0 in (2.14) implies (2.16). Conversely,
under Condition M, we have that |X0,k |αk is regularly varying of order 1 for all
1 ≤ k ≤ d with slowly varying functions `i (t) ∼ ci . Moreover kX0 kα is regularly
varying from the weak convergence in (2.16) applied on the Borel sets {kX0 kα > ty},
y ≥ 1. Thus, kX0 kα is regularly varying of order 1 with slowly varying function `(t).
One can rewrite (2.16) as
`(t)−1 tP(x(t)−1
X0 ∈ ·, kX0 kα > t) → P(Y0 ∈ ·).
Using the slowly varying property of `, we obtain, for any > 0,
`(t)−1 tP(x(t)−1
X0 ∈ ·, kX0 kα > t) → −1 P(Y0 ∈ ·).
12
Then by marginal homogeneity of k · kα ,
`(t)−1 tP(x(t)−1
X0 ∈ ·, kx(t)−1
X0 kα > ) → −1 P(Y0 ∈ ·).
Notice that `(t)t−1 > 0 is non-increasing as it is the tail of kX0 kα . So there exists a
change of variable t = h(t0 ) so that `(t)−1 t = t0 and
t0 P(x(h(t0 ))−1
X0 ∈ ·, kx(h(t0 ))−1
X0 kα > ) → −1 P(Y0 ∈ ·).
We obtain the existence of µ for x0 = x ◦ h in (2.14) such that µ(·, kxkα > ) = P(·),
which is enough to characterize µ entirely, choosing > 0 arbitrarily small.
The spectral properties of VSRV X0 can be expressed in terms of the tail vector
Y0 . Notice that for any u ∈ {+1, 0, −1}d , there exists c+ (u) ≥ 0 satisfying
lim P max
t→∞
1≤i≤d
αi
c−1
i (ui X0,i )+
> t | kX0 kα > t = c+ (u).
−1 |
d
Consider c−1 (u X0 )α+ , where c−1 = (c−1
1 , . . . , cd ) and for x ∈ R and α =
(α1 , ..., αd )| , (x)α+ = ((x1 )α+1 , ..., (xd )α+d )| . If c+ (u) is non-null, by a continuous mapping argument, c−1 (u X0 )α+ satisfies
L(t−1 c−1
(u
X0 )α+ | k(u
X0 )+ kα > t) →t→∞ L(c+ (u)−1 (u
Y0 )α+ ),
(2.17)
and c−1 (u X0 )α+ is regularly varying of index 1. By homogeneity of the limiting
measure in the multivariate regular variation (2.10), we may decompose the limit as
a product
P((k(u
X0 )+ kα > ty, c−1 (u X0 )α+ /k(u
P(k(u X0 )+ kα > t)
X0 )+ kα ∈ ·)
→ y −α PΘu (·),
for any y ≥ 1. Such limiting distribution is called a simple max-stable distribution,
and PΘu , supported by the positive orthant, is called the spectral measure of c−1
(u X0 )α+ , see de Haan and Resnick (1977) for more details. By identification of
the two expressions of the same limit, we obtain the following proposition.
Proposition 2.17. With Y0 defined in Proposition 2.16, the distribution of (u
Y0 )α+ /k(u Y0 )+ kα , if non-degenerate, is the spectral measure of c−1 (u X0 )α+ ∈
[0, ∞)d . Moreover, it is independent of k(u Y0 )+ kα , and c+ (u)−1 k(u Y0 )+ kα is
standard Pareto distributed.
Proof. That c+ (u)−1 k(u Y0 )+ kα is standard Pareto distributed follows from the
convergence in (2.17) associated with the regularly varying property, ensuring the
13
homogeneity of the limiting measure. Then, using again the homogeneity in (2.17),
it follows that (u Y0 )α+ /k(u Y0 )+ kα and c+ (u)−1 k(u Y0 )+ kα are independent.
Example 2.18 (Diagonal BEKK-ARCH, continued). We have not been able to
establish the existence of Y0 satisfying (2.16), except the case of the scalar BEKKARCH where the diagonal elements of A are identical. In this case the process is
a special case of the Similarity BEKK-ARCH, see Example 2.6. Even in this case,
the characterization of the spectral distribution is not an easy task because of the
diagonality of A, ruling out Theorem 1.4 of Buraczewski et al. (2009). In Section
A.2 in the appendix we have included some estimates of the spectral measure of
X0 for the bivariate case. The plots suggest that the tails of the process are indeed
dependent. We emphasize that new multivariate renewal theory should be developed
in order to prove that the Diagonal-ARCH model is VSRV. We leave such task for
future research.
3
Vector-scaling regularly varying time series and
their extremal behavior
The existence of the tail vector in Proposition 2.16 allows us to extend the asymptotic
results of Perfekt (1997) to VSRV vectors taking possibly negative values. In order
to do so, we use the notion of tail chain from Basrak and Segers (2009) adapted to
VSRV stationary sequences with eventually different tail indices.
3.1
Vector scaling regularly varying time series
We introduce a new notion of multivariate regularly varying time series based on
VSRV of Xt .
Definition 3.1. The stationary process (Xt ) is VSRV if and only if there exists a
process (Yt )t≥0 , with non-degenerate marginals for Y0 , such that
L(((ci t)−1/αi )1≤i≤d
(X0 , X1 , . . . , Xk ) | kX0 kα > t) →t→∞ L(Y0 , . . . , Yk ),
for all k ≥ 0. The sequence (Yt )t≥0 is called the tail process.
Following Basrak and Segers (2009), we extend the notion of spectral measure
to the one of spectral processes for any VSRV stationary process:
14
Definition 3.2. The VSRV stationary process (Xt ) admits the spectral process (Θt )
if and only if
L(kX0 k−1
α (X0 , X1 , . . . , Xk ) | kX0 kα > t) →t→∞ L(Θ0 , . . . , Θk ),
for all k ≥ 0.
By arguments similar to the ones in the proof of Proposition 2.17, it follows that
the VSRV properties also characterize the spectral process of (c−1 (u Xt )α+ )t≥0 ,
with X0 following the stationary distribution, which has the distribution of ((u
Yt )α+ /k(u Y0 )+ kα )t≥0 . We have the following proposition.
Proposition 3.3. For a VSRV stationary process (Xt ), where Y0 has non-degenerate
marginals and kY0 kα is standard Pareto distributed, the spectral process of any nondegenerate (c−1 (u Xt )α+ )t≥0 is distributed as ((u Yt )α+ /k(u Y0 )+ kα )t≥0 and
independent of k(u Y0 )+ kα . Moreover c+ (u)−1 k(u Y0 )+ kα is standard Pareto
distributed.
3.2
The tail chain
In the following, we will focus on the dynamics of the tail process (Yt )t≥1 in Definition
3.1, given the existence of Y0 . We will restrict ourselves to the case where (Xt ) is a
Markov chain, which implies that (Yt ) is also a Markov chain called the tail chain;
see Perfekt (1997). We have the following proposition.
Proposition 3.4. Let (Xt ) satisfy (1.1)-(1.2) be a VSRV stationary process. With
M̃t defined in (1.4), the tail process (Yt ) admits the multiplicative form
Yt+1 = M̃t+1 Yt ,
t ≥ 0.
(3.1)
Proof. Following the approach of Janssen and Segers (2014), one first notices that
the existence of the kernel of the tail chain does not depend on the marginal distribution. Thus the characterization of the kernel extends automatically from the
usual multivariate regular variation setting to the vector scaling regular variation
one. It is straightforward to check Condition 2.2 of Janssen and Segers (2014). We
conclude that the tail chain has the multiplicative structure in (3.1).
The tail chain for VSRV process satisfying (1.1)-(1.2) is the same no matter the
values of the marginal tail indices; for the multivariate regularly varying case with
common tail indices it coincides with the tail chain of Janssen and Segers (2014)
15
under Condition M. Notice that we can extend the tail chain Yt backward in time
(t < 0) using Corollary 5.1 of Janssen and Segers (2014).
3.3
Asymptotic behavior of the maxima
From the previous section, we have that the tail chain (Yt ) quantifies the extremal
behavior of (Xt ) in (1.1)-(1.2). Let us consider the asymptotic behavior of the
component-wise maxima
max(X1 , . . . , Xn ) = (max(X1,k , . . . , Xn,k ))1≤k≤d .
Let u = (1, . . . , 1) = 1 ∈ Rd and assume that c+ (1) = limt→∞ P(X0
x(t) |
|X0 |
x(t)) is positive. Recall that for (Xt ) i.i.d., the suitably scaled maxima
converge to the Fréchet distribution; see de Haan and Resnick (1977), i.e. for any
x = (x1 , . . . , xd )| ∈ Rd+ , defining un (x) such that nP(X0,i > un,i (x)) ∼ x−1
i , 1 ≤ i ≤
d, we have
P(max(X1 , . . . , Xn ) ≤ un (x)) → exp(−A∗ (x)),
if and only if (X0 )+ is vector scaling regularly varying. In such case, due to Condition
M, we have the expression
1
(Y0,i )α+k
.
A (x) = c+ (1)E
max
k(Y0 )+ kα 1≤i≤d ci xi
#
"
∗
(3.2)
Let us assume the following Condition, slightly stronger than (2.1):
There exists p > 0 such that lim E[kM̃1 · · · M̃n kp ]1/n < 1.
(3.3)
n→∞
Theorem 3.5. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4), suppose that
condition (3.3) holds. Suppose that the stationary distribution is VSRV. Assuming
the existence of Y0 in Definition 3.1, we have that
P(max(Xm , . . . , Xn ) ≤ un (x)) → exp(−A(x)),
where A(x) admits the expression
c+ (1)E max
1≤i≤d
maxk≥0
Q
1≤j≤k
M̃k−j Y0
αk
i +
k(Y0 )+ kα ci xi
− max
1≤i≤d
maxk≥1
Q
1≤j≤k
M̃k−j Y0
k(Y0 )+ kα ci xi
αk
i +
(. 3.4)
Proof. We verify the conditions of Theorem 4.5 of Perfekt (1997). Condition B2 of
16
Perfekt (1997) is satisfied under the more tractable Condition 2.2 of Janssen and
Segers (2014). Indeed, the tail chain depends only on the Markov kernel and one
can apply Lemma 2.1 of Janssen and Segers (2014), because it extends immediately
to the vector scaling regularly varying setting. Condition D(un ) of Perfekt (1997)
holds by geometric ergodicity of the Markov chain for a sequence un = C log n, with
C > 0 sufficiently large. Lastly, the finite clustering condition,
lim lim supn→∞ P[max(|Xm |, . . . , |XC log n |)
m→∞
un (x) | |X0 |
un (x)] = 0,
(3.5)
holds for any C > 0 using the same reasoning as in the proof of Theorem 4.6 of
Mikosch and Wintenberger (2013) under the drift condition (DCp ) for some p < α =
min{αi : 1 ≤ i ≤ d}. As (Xt ) is also standard α regularly varying, actually the drift
condition holds thanks to Condition (3.3) on some sufficiently large iterations of the
Markov kernel. Finally, as (3.5) is a special case of Condition D∞ (c log n) of Perfekt
(1997), we obtain the desired result with the characterization given in Theorem 4.5
of Perfekt (1997)
A(x) =
Z
(0,∞)d \(0,x)
P (Tj ≤ x, k ≥ 1 | T0 = y) ν(dy),
αi
where (Tk )k≥0 is the tail chain of the standardized Markov chain (c−1
i (Xk,i )+ )1≤i≤d ,
k ≥ 0. As µ restricted to (0, ∞)d \ (0, 1)d is the distribution of Y0 , we assume that
xi ≥ 1 for all 1 ≤ i ≤ d so that we identify ν as the distribution of
αi
(c−1
i (Y0,i )+ )1≤i≤d
αi
max c−1
i (Y0,i )+ /xi > 1.
under the constraint
1≤i≤d
Thus we have
αi
−1
αi
A(x) = P c−1
i (Yk,i )+ /xi ≤ 1, k ≥ 1, 1 ≤ i ≤ d, max ci (Y0,i )+ /xi > 1 .
1≤i≤d
To obtain an expression that is valid for any xi > 0, we exploit the homogeneity
property, and we obtain
αi
αi
A(x) = P max max (ci xi )−1 Yk,i
> 1 − P max max (ci xi )−1 Yk,i
>1
k≥0 1≤i≤d
"
−1
maxk≥0 max1≤i≤d (ci xi )
= c+ (1)E
k(Y0 )+ kα
k≥1 1≤i≤d
maxk≥1
(Yk,i )α+i
−
max1≤i≤d (ci xi )−1 (Yk,i )α+i
k(Y0 )+ kα
because c+ (1)−1 k(Y0 )+ kα is standard Pareto distributed and independent of the
spectral process (Yk )α+ /k(Y0 )+ kα . This expression is homogeneous and extends to
any possible x by homogeneity.
17
#
3.4
Extremal indices
As the random coefficients M̃t in (1.4) may be large, consecutive values of Xt can be
large. In the univariate case, one says that the extremal values appear in clusters.
An indicator of the average length of the cluster is the inverse of the extremal index,
an indicator of extremal dependence; see Leadbetter et al. (1983).
Thus, the natural extension of the extremal index is the function θ(x) = A(x)/A∗ (x),
with A∗ (x) and A(x) defined in (3.2) and (3.4), respectively. Notice that there is no
reason why θ should not depend on x. When xi ≥ c+ (1), for 1 ≤ i ≤ d, we have the
more explicit expression in terms of the spectral process,
αi
αi
θ(x) = P Yk,i
≤ ci xi , k ≥ 1, 1 ≤ i ≤ d | Y0,i
> ci xi , 1 ≤ i ≤ d .
(3.6)
However, the extremal index θi of the marginal index (Xt,i ) is still well-defined. It
depends on the complete dependence structure of the multivariate Markov chain
thanks to the following proposition:
Proposition 3.6. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4) satisfying
(3.3) and assuming the existence of Y0 in Definition 3.1, the extremal index, θ,
defined in (3.6), is a positive continuous function bounded from above by 1 that can
be extended to (0, ∞]d \ {∞, . . . , ∞}. The extremal indices of the marginals are
θi = θ(∞, . . . , ∞, xi , ∞, . . . , ∞)
E
k(Y0 )+ k−1
α
maxk≥0
Q
M̃k−j Y0
1≤j≤k
=
h
E
αi
i +
− maxk≥1
αi
k(Y0 )+ k−1
α (Y0,i )+
Q
1≤j≤k
M̃k−j Y0
αi
i +
i
.
Proof. Except for the positivity of the extremal index, the result follows by Proposition 2.5 in Perfekt (1997). The positivity is ensured by applying Corollary 2 in
Segers (2005).
Example 3.7 (Diagonal BEKK-ARCH, continued). Suppose that X0 is VSRV
as conjectured in Example 2.18. It follows from the tail chain approach of Janssen
and Segers (2014) that the stationary Markov chain (Xt ) is regularly varying. Thanks
αi
to the diagonal structure of the matrices M̃k = Amk , one can factorize k(Y0 )+ k−1
α (Y0,i )
αi
in the expression of θi provided in Proposition 3.6. Since k(Y0 )+ k−1
and mk
α (Y0,i )
are independent for k ≥ 1, we recover a similar expression as in the remarks after
18
Theorem 2.1 in de Haan et al. (1989):
αi
θi = E max Akii
k≥0
Y
αi
mj − max Akii
k≥1
1≤j≤k
+
Y
mj .
1≤j≤k
+
We did not manage to provide a link between the θi and the extremal index θ(x) of
the (multivariate) stationary solution (Xt ) of the Diagonal BEKK-ARCH. Due to the
different normalising sequences in the asymptotic extremal result given in Theorem
3.5, the extremal index θ(x) depends on the constants ci , i = 1, ..., d. For x∗i =
c+ (1), 1 ≤ i ≤ d, the expression (3.6) gets more simple because c+ (1)−1 k(Y0 )+ kα is
standard Pareto distributed and supported on [1, ∞):
θ(x∗ ) = P Akii
mj Y0,i ≤ (ci c+ (1))1/αi , k ≥ 1, 1 ≤ i ≤ d .
Y
1≤j≤k
One can check that θ(x∗ ) ≥ θi0 where 1 ≤ i0 ≤ d satisfies Ai0 i0 ≥ Aii , 1 ≤ i ≤ d
so that i0 is the marginal with smallest tail and extremal indices. Thus the inverse
of the extremal index of the multidimensional Diagonal BEKK-ARCH is not larger
than the largest average length of the marginals clusters. It can be interpreted as
the fact that the largest clusters are concentrated along the i0 axis, following the
interpretation of the multivariate extremal index given on p. 423 of Beirlant et al.
(2006).
3.5
Convergence of point processes
Let us consider the vector scaling point process on Rd
Nn (·) =
n
X
δ((ci n)−1/αi )1≤i≤d
Xt (·),
n ≥ 0.
(3.7)
t=1
We want to characterize the asymptotic distribution of the point process Nn when
n → ∞. We refer to Resnick (2007) for details on the convergence in distribution
for random measures. In order to characterize the limit, we adapt the approach
of Davis and Hsing (1995) to the multivariate VSRV case similar to Davis and
Mikosch (1998). The limit distribution will be a cluster point process admitting the
expression
N (·) =
∞ X
∞
X
j=1 t=1
δ
(ci Γj )−1/αi
1≤i≤d
Qj,t
(·),
(3.8)
where Γj , j = 1, 2, ..., are arrival times of a standard Poisson process, and (Qj,t )t∈Z ,
19
j = 1, 2, ..., are mutually independent cluster processes. Following Basrak and Tafro
(2016), we use the back and forth tail chain (Yt ) to describe the cluster process:
Consider the process (Zt ), satisfying
L (Zt )t∈Z = L (Yt )t∈Z | sup kYt kα ≤ 1 ,
t≤−1
which is well defined when the anti-clustering condition (3.5) is satisfied. Then we
have
j ≥ 1,
(Z
)
L (Qj,t )t∈Z = L L−1
t t∈Z ,
Z
with LZ = supt∈Z kZt kα . Notice that the use of the pseudo-norm k · kα and the fact
that kY0 kα is standard Pareto are crucial to mimic the arguments of Basrak and
Tafro (2016). The limiting distribution of the point process Nn coincides with the
one of N :
Theorem 3.8. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4), suppose that
(3.3) holds, and assume that Y0 in Definition 3.1 exists. With Nn defined in (3.7)
and N defined in (3.8),
d
Nn → N,
n → ∞.
Proof. Let us denote sign the operator sign(x) = x/|x|, x ∈ R, applied coordinatewise to vectors in Rd . We apply Theorem 2.8 of Davis and Mikosch (1998)
to the transformed process (c−1 sign(Xt ) |Xt |α )t∈Z which is standard regularly
varying of order 1. In order to do so, one has to check that the anti-clustering
condition (3.5) is satisfied and that the cluster index of its max-norm is positive.
This follows from arguments developed in the proof of Theorem 3.5. The mixing
condition of Davis and Mikosch (1998) is implied by the geometric ergodicity of
P
(Xt ). Thus, the limiting distribution of the point process nt=1 δn−1 c−1 sign(Xt ) |Xt |α
P∞
P
for some cluscoincides with the one of the cluster point process ∞
t=1 δΓ−1
j=1
j Q̃j,t
ter process (Q̃j,t )t∈Z . A continuous mapping argument yields the convergence of Nn
P
P∞
to ∞
j=1
t=1 δ((ci Γj )−1/αi )1≤i≤d sign(Q̃j,t ) |Q̃j,t |α . The limiting cluster process coincide
with Qj,t in distribution thanks to the definition of VSRV processes.
4
Sample covariances
In this section, we derive the limiting distribution of the sample covariances for
certain BEKK-ARCH processes. Consider the sample covariance matrix,
Γn,X =
n
1X
Xt Xt| .
n t=1
20
Let vech(·) denote the half-vectorization operator, i.e. for a d × d matrix A = [aij ],
vech(A) = (a11 , a21 , ..., ad1 , a22 , ..., ad2 , a33 , ..., add )| (d(d + 1)/2 × 1). The derivation
of the limiting distribution of the sample covariance matrix relies on using the multidimensional regularly varying properties of the stationary process (vech(Xt Xt| ) :
t ∈ Z). Let an−1 denote the normalization matrix,
−1/αi −1/αj
cj
.
1≤i,j≤d
an−1 = n−1/αi −1/αj ci
Using Theorem 3.8 and adapting the continuous mapping argument of Proposition
3.1 of Davis and Mikosch (1998) yield the following result.
Proposition 4.1. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4) satisfying
(3.3) and assuming the existence of Y0 in Definition 3.1, we have
n
X
δvech(an−1 )
t=1
d
vech(Xt Xt| )
→
∞ X
∞
X
δvech(P` )
`=1 t=1
n → ∞,
vech(Q`,t Q|`,t ) ,
where
−1/αi −1/αj −1/αi −1/αj
.
cj
ci
1≤i,j≤d
P` = Γ`
Let us define αi,j = αi αj /(αi + αj ) and assume that αi,j 6= 1 and αi,j 6= 2 for all
1 ≤ i ≤ j ≤ d. Note that αi,j is a candidate for the tail index of the cross product
Xt,i Xt,j and that αi,i = αi /2, 1 ≤ i ≤ d. Actually it is the case under some extra
assumptions ensuring that the product Y0,i Y0,j is non null, see Proposition 7.6 of
Resnick (2007). In line with Theorem 3.5 of Davis and Mikosch (1998), we then get
our main result on the asymptotic behavior of the empirical covariance matrix
Theorem 4.2. Let Xt satisfy (1.1)-(1.2). With M̃t defined in (1.4), suppose that
(3.3) holds, and assume that Y0 in Definition 3.1 exists. Moreover, for any (i, j)
such that 1 < αi,j < 2, suppose that
lim lim sup Var n−1/αi,j
ε→0
n→∞
n
X
t=1
Xt,i Xt,j 1|Xt,i Xt,j |≤n1/αi,j ε = 0.
(4.1)
Then
√
1−1/αi,j
n∧n
(Γn,X − E[Γn,X ]1αi,j >1 )i,j
d
→ S,
n → ∞,
1≤j≤i≤d
where Si,j is an αi,j ∧ 2-stable random variable for 1 ≤ i ≤ j ≤ d and non-degenerate
for i = j.
When Theorem 4.2 applies, as αi,j ≥ (αi ∧ αj )/2, the widest confidence interval
21
on the covariance estimates is supported by the i0 th marginal satisfying αi0 ≤ αi
for all 1 ≤ i ≤ d.
In order to apply Theorem 4.2, the main difficulty is to show that the condition
(4.1) holds. However, notice that Theorem 4.2 applies simultaneously on the crossproducts with αi,j ∈
/ [1, 2] with no extra assumption. Next, we apply Theorem 4.2
to the ongoing examples.
Example 4.3 (Diagonal BEKK-ARCH, continued). Consider the diagonal BEKKARCH process and the cross products Xt,i Xt,j for some i ≤ j and any t ∈ Z. From
Hölder’s inequality (which turns out to be an equality in our case), we have
E[|Aii Ajj m20 |αi,j ] = E[|Aii m0 |αi |]αi,j /αi E[|Ajj m0 |αj ]αi,j /αj = 1.
Thus, (Xt,i Xt,j ), which is a function of the Markov chain (Xt ), satisfies the drift
condition (DCp ) of Mikosch and Wintenberger (2013) for all p < αi,j . Then, one
can show that (4.1) is satisfied using the same reasoning as in the proof of Theorem
4.6 of Mikosch and Wintenberger (2013).
Example 4.4 (Similarity BEKK-ARCH, continued). If αi,j ∈
/ [1, 2], the limiting distribution of the sample covariance matrix for the Similarity BEKK-ARCH
follows directly from Theorem 4.2. If αi,j ∈ (1, 2) the additional condition (4.1) has
to be checked. Relying on the same arguments as in Example 4.3, one would have
to verify that the condition (DCp ) of Mikosch and Wintenberger (2013) holds for
the Similarity BEKK-ARCH process, which appears a difficult task as it requires to
find a suitable multivariate Lyapunov function. We leave such task for future investigation. Consider the special case of the scalar BEKK-ARCH process introduced in
√
Example 2.6. Here A = aId , with Id the identity matrix, such that M̃t is diagonal.
In the case αi,j ∈ (1, 2) for a least some pair (i, j), the limiting distribution of the
sample covariance is derived along the lines of Example 4.3. Specifically, this relies
on assuming that a < exp {(1/2) [−ψ(1) + log(2)]} such that a stationary solution
exists, and noting that the index of regular variation for each marginal of Xt is given
√
by α satisfying E[| amt |α ] = 1.
Example 4.5 (ID BEKK-ARCH, continued). Whenever αi,j ∈
/ [1, 2], the limiting distribution of the sample covariance matrix for the ID BEKK-ARCH follows
directly from Theorem 4.2. Similar to Example 4.4 we leave for future investigation
to show whether condition (4.1) holds.
The previous examples are important in relation to variance targeting estimation
of the BEKK-ARCH model, as considered in Pedersen and Rahbek (2014). For the
22
univariate GARCH process, Vaynman and Beare (2014) have shown that the limiting
distribution of the (suitably scaled) variance targeting estimator follows a singular
stable distribution when the tail index of the process lies in (2, 4). We expect a
similar result to hold for the BEKK-ARCH process.
5
Concluding remarks
We have found a mild sufficient condition for geometric ergodicity of a class of
BEKK-ARCH processes. By exploiting the the processes can be written as a multivaraite stochastic recurrence equation (SRE), we have investigated the tail behavior
of the invariant distribution for different BEKK-ARCH processes. Specifically, we
have demonstrated that existing Kesten-type results apply in certain cases, implying
that each marginal of the invariant distribution has the same tail index. Moreover,
we have shown for certain empirically relevant processes, existing renewal theory is
not applicable. In particular, we show that the Diagonal BEKK-ARCH processes
may have component-wise different tail indices. In light of this property, we introduce the notion of vector scaling regular varying (VSRV) distributions and processes.
We study the extremal behavior of such processes and provide results for convergence of point processes based on VSRV processes. It is conjectured, and supported
by simulations, that the Diagonal BEKK-ARCH process is VSRV. However, it remains an open task to verify formally that the property holds. Such task will require
the development of new multivariate renewal theory.
Our results are expected to be important for future research related to the statistical analysis of the Diagonal BEKK-ARCH model. As recently shown by Avarucci
et al. (2013), the (suitably scaled) maximum likelihood estimator for the general
BEKK-ARCH model (with l = 1) does only have a Gaussian limiting distribution,
if the second-order moments of Xt is finite. In order to obtain the limiting distribution in the presence of very heavy tails, i.e. when E[kXt k2 ] = ∞, we believe
that non-standard arguments are needed, and in particular the knowledge of the
tail-behavior is expected to be crucial for the analysis. We leave additional considerations in this direction to future research.
References
Alsmeyer, G. (2003): “On the Harris recurrence of iterated random Lipschitz
functions and related convergence rate results,” Journal of Theoretical Probability,
16, 217–247.
23
Alsmeyer, G. and S. Mentemeier (2012): “Tail behaviour of stationary solutions of random difference equations: The case of regular matrices,” Journal of
Difference Equations and Applications, 18, 1305–1332.
Avarucci, M., E. Beutner, and P. Zaffaroni (2013): “On moment conditions for quasi-maximum likelihood estimation of multivariate ARCH models,”
Econometric Theory, 29, 545–566.
Basrak, B., R. A. Davis, and T. Mikosch (2002a): “A characterization of
multivariate regular variation,” The Annals of Applied Probability, 12, 908–920.
——— (2002b): “Regular variation of GARCH processes,” Stochastic Processes and
their Applications, 99, 95–115.
Basrak, B. and J. Segers (2009): “Regularly varying multivariate time series,”
Stochastic Processes and their Applications, 119, 1055–1080.
Basrak, B. and A. Tafro (2016): “A complete convergence theorem for stationary regularly varying multivariate time series,” Extremes, 19, 549–560.
Bauwens, L., S. Laurent, and J. V. K. Rombouts (2006): “Multivariate
GARCH models: A survey,” Journal of Applied Econometrics, 21, 79–109.
Beirlant, J., Y. Goegebeur, J. Segers, and J. L. Teugels (2006): Statistics
of extremes: theory and applications, John Wiley & Sons.
Boussama, F., F. Fuchs, and R. Stelzer (2011): “Stationarity and geometric
ergodicity of BEKK multivariate GARCH models,” Stochastic Processes and their
Applications, 121, 2331–2360.
Buraczewski, D., E. Damek, Y. Guivarc’h, A. Hulanicki, and R. Urban (2009): “Tail-homogeneity of stationary measures for some multidimensional
stochastic recursions,” Probability Theory and Related Fields, 145, 385–420.
Buraczewski, D., E. Damek, and T. Mikosch (2016): Stochastic Models with
Power-Law Tails: The Equation X = AX + B, Springer Series in Operations
Research and Financial Engineering, Springer International Publishing.
Damek, E., M. Matsui, and W. Świątkowski (2017): “Componentwise different tail solutions for bivariate stochastic recurrence equations,”
https://arxiv.org/abs/1706.05800.
24
Davis, R. A. and T. Hsing (1995): “Point process and partial sum convergence
for weakly dependent random variables with infinite variance,” The Annals of
Probability, 879–917.
Davis, R. A. and T. Mikosch (1998): “The sample autocorrelations of heavytailed processes with applications to ARCH,” The Annals of Statistics, 26, 2049–
2080.
de Haan, L. and S. I. Resnick (1977): “Limit theory for multivariate sample
extremes,” Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete, 40,
317–337.
de Haan, L., S. I. Resnick, H. Rootzén, and C. G. de Vries (1989): “Extremal behaviour of solutions to a stochastic difference equation with applications
to ARCH processes,” Stochastic Processes and their Applications, 32, 213–224.
Einmahl, J. H., L. de Haan, and V. I. Piterbarg (2001): “Nonparametric
estimation of the spectral measure of an extreme value distribution,” The Annals
of Statistics, 29, 1401–1423.
Engle, R. F. and K. F. Kroner (1995): “Multivariate simultaneous generalized
ARCH,” Econometric Theory, 11, 122–150.
Feigin, P. and R. Tweedie (1985): “Random coefficient autoregressive processes:
a Markov chain analysis of stationarity and finiteness of moments,” Journal of
Time Series Analysis, 6, 1–14.
Goldie, C. M. (1991): “Implicit renewal theory and tails of solutions of random
equations,” The Annals of Applied Probability, 1, 126–166.
Janssen, A. and J. Segers (2014): “Markov tail chains,” Journal of Applied
Probability, 51, 1133–1153.
Kesten, H. (1973): “Random difference equations and renewal theory for products
of random matrices,” Acta Mathematica, 131, 207–248.
Kulik, R., P. Soulier, and O. Wintenberger (2015): “The tail empirical
process of regularly varying functions of geometrically ergodic Markov chains,”
https://arxiv.org/abs/1511.04903.
Leadbetter, M., G. Lindgren, and H. Rootzén (1983): Extremes and related properties of random sequences and processes, Springer series in statistics,
Springer-Verlag.
25
Lindskog, F., S. I. Resnick, and J. Roy (2014): “Regularly varying measures on metric spaces: Hidden regular variation and hidden jumps,” Probability
Surveys, 11, 270–314.
Matsui, M. and T. Mikosch (2016): “The extremogram and the crossextremogram for a bivariate GARCH(1,1) process,” Advances in Applied Probability, 48, 217–233.
Mikosch, T. and O. Wintenberger (2013): “Precise large deviations for dependent regularly varying sequences,” Probability Theory and Related Fields, 156,
851–887.
Nelson, D. B. (1990): “Stationarity and persistence in the GARCH(1,1) model,”
Econometric Theory, 6, 318–334.
Nielsen, H. B. and A. Rahbek (2014): “Unit root vector autoregression with
volatility induced stationarity,” Journal of Empirical Finance, 29, 144–167.
Pedersen, R. S. (2016): “Targeting estimation of CCC-GARCH models with
infinite fourth moments,” Econometric Theory, 32, 498–531.
Pedersen, R. S. and A. Rahbek (2014): “Multivariate variance targeting in the
BEKK-GARCH model,” The Econometrics Journal, 17, 24–55.
Perfekt, R. (1997): “Extreme value theory for a class of Markov chains with
values in Rd ,” Advances in Applied Probability, 29, 138–164.
Resnick, S. I. (2007): Heavy-tail phenomena: probabilistic and statistical modeling,
Springer Science & Business Media.
Segers, J. (2005): “Generalized Pickands estimators for the extreme value index,”
Journal of Statistical Planning and Inference, 128, 381–396.
Stărică, C. (1999): “Multivariate extremes for models with constant conditional
correlations,” Journal of Empirical Finance, 6, 515–553.
Vaynman, I. and B. K. Beare (2014): “Stable limit theory for the variance
targeting estimator,” in Essays in Honor of Peter C. B. Phillips, ed. by Y. Chang,
T. B. Fomby, and J. Y. Park, Emerald Group Publishing Limited, vol. 33 of
Advances in Econometrics, chap. 24, 639–672.
26
A
Appendix
A.1
Theorem 1.1 of Alsmeyer and Mentemeier (2012)
Consider the general SRE
Yt = At Yt−1 + Bt
(A.1)
with (At , Bt ) a sequence of i.i.d. random variables with generic copy (A, B) such
that A is a d × d real matrix and B takes values in Rd . Consider the following
conditions of Alsmeyer and Mentemeier (2012):
• (A1) E[log+ (kAk)] < ∞, where k · k denotes the operator norm.
• (A2) E[log+ (kBk)] < ∞.
• (A3) P[A ∈ GL(d, R)] = 1.
• (A4) maxn∈N P {kx| ni=1 Ai k−1 (x| ni=1 Ai ) ∈ U } > 0, for any x ∈ Sd−1 and
any non-empty open subset U of Sd−1 .
Q
Q
• (A5) Let Vδ denote the open δ-ball in GL(d, R) and let LEB denote the
Lebesgue measure on M (d, R). It holds that for any Borel set A ∈ M (d, R),
Q 0
P( ni=1
Ai ∈ A) ≥ γ0 1Vc (Γ0 ) (A)LEB(A) for some Γ0 ∈ GL(d, R), n0 ∈ N, and
c, γ0 > 0.
• (A6) P(A0 v + B0 = v) < 1 for any v ∈ Rd .
• (A7) There exists κ0 > 0 such that
E[ infd−1 kx| A0 kκ0 ] ≥ 1,
x∈S
E[kA0 kκ0 log+ kA0 k] < ∞,
and 0 < E[kB0 kκ0 ] < ∞.
Theorem A.1 (Alsmeyer and Mentemeier (2012, Theorem 1.1)). Consider the SRE
Q
in (A.1)) suppose that β := limn→∞ n−1 log(k ni=1 Ai k) < 0 and that (A1)-(A7)
hold, then there exists a unique κ ∈ (0, κ0 ] such that
lim n−1 log(k
n→∞
n
Y
Ai kκ ) = 0.
i=1
Moreover, the SRE has a strictly stationary solution satisfying,
lim tκ P(x| Y0 > t) = K(x)
t→∞
for all x ∈ Sd−1 ,
where K is a finite positive and continuous function on Sd−1 .
27
A.2
Estimation of the spectral measure for the bivariate
diagonal BEKK-ARCH process
In this section we consider the estimation of the spectral measure of the diagonal
BEKK-ARCH process presented in Example 2.14. Specifically, we consider a special
case of the BEKK-ARCH process in (1.1)-(1.2), where d = 2:
Xt = mt AXt−1 + Qt ,
with {Qt : t ∈ N} an i.i.d. process with Qt ∼ N (0, C) independent of {mt : t ∈ N},
and
A
0
11
.
A=
0 A22
Following the approach for i.i.d. sequences of vectors given in Einmahl et al.
(2001), we consider the following estimator of the spectral measure of Xt = (Xt,1 , Xt,2 )| :
Φ̂(θ) =
T
1X
,
1
(2)
k t=1 {Rt(1) ∨Rt(2) ≥T +1−k,arctan T +1−Rt(1) ≤θ}
T +1−R
θ ∈ [0, π/2],
t
(j)
where Rt denotes the rank of Xt,j among X1,j , ..., XT,j , j = 1, 2, i.e.
(j)
Rt
:=
T
X
1{Xi,j ≥Xt,j } .
i=1
Here k is a sequence satisfying k(T ) → ∞ and k(T ) = o(T ). Einmahl et al. (2001)
showed that this estimator is consistent for i.i.d. series. We expect a similar result to
hold for geometrically ergodic processes. The reason is that the asymptotic behavior
of the empirical tail process used in Einmahl et al. (2001) has been extended to such
cases in Kulik et al. (2015).
We consider the estimation of the spectral measure for different values of C, A11 ,
and A22 . In particular, the matrix C is
1 c
C = 10−5
,
c 1
c ∈ {0, 0.5},
and the values A11 and A22 are determined according to choices of the tail indices
of Xt,1 and Xt,2 , respectively. I.e. A11 and A22 satisfy E[|mt |αi ] = |Aii |−αi and are
determined by analytical integration. Specifically, with φ(·) the pdf of the standard
28
normal distribution,
Z ∞
αi = 0.5 ⇒ Aii = (
|m|0.5 φ(m)dm)−1/0.5 ≈ 1.479
−∞
αi = 2.0 ⇒ Aii = 1
αi = 3.0 ⇒ Aii = (8/π)−1/6 ≈ 0.8557
αi = 4.0 ⇒ Aii = 3−1/4 ≈ 0.7598
Figure A.1 contains plots of the estimates of the spectral measure. The estimates
Φ̂(θ) are based on one realization of the process with T = 2,000 and a burn-in period
of 10,000 observations.
1.5
k=100
k=300
k=500
k=200
k=400
1.5
α1 = 2, α2 = 2, c = 0.5
1.0
^
Φ(θ)
^
Φ(θ)
α1 = 2, α2 = 2, c = 0
0.5
0.0
1.0
0.5
0.1
0.2
0.3
0.4
0.5
0.0
0.1
0.2
(θ/π)
2.0
α1 = 3, α2 = 4, c = 0
1.5
^
Φ(θ)
^
Φ(θ)
1.5
1.0
0.1
0.2
0.3
0.4
0.5
0.0
0.1
0.2
0.4
0.5
0.3
0.4
0.5
(θ/π)
α1 = 4, α2 = 0.5, c = 0
1.5
^
Φ(θ)
^
Φ(θ)
0.3
1.0
2.0
1.0
0.5
0.0
0.5
α1 = 3, α2 = 4, c = 0.5
(θ/π)
2.0
1.5
0.4
0.5
0.5
0.0
0.3
(θ/π)
α1 = 4, α2 = 0.5, c = 0.5
1.0
0.5
0.1
0.2
0.3
0.4
0.5
(θ/π)
0.0
0.1
0.2
(θ/π)
Figure A.1: Nonparametric estimates for k = 100, 200, 300, 400, 500 and for various
choices of α1 , α2 , and c.
29
| 10math.ST
|
arXiv:1702.00686v1 [cs.SY] 1 Feb 2017
Co-simulation: State of the art
Cláudio Gomes ([email protected])
Casper Thule ([email protected])
David Broman ([email protected])
Peter Gorm Larsen ([email protected])
Hans Vangheluwe ([email protected])
February 3, 2017
Contents
1 Introduction
1.1 Motivation . . . . .
1.2 Co-simulation . . . .
1.3 Need for the Survey
1.4 Outline . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
4
6
6
2 Modeling, Simulation, and Co-simulation
2.1 Dynamical Systems – Models of Real Systems
2.2 Simulators – Computing the Behavior Trace .
2.3 Simulation Units - Mock-ups of Reality . . .
2.4 Compositional Co-simulation . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7
7
8
10
10
3 Discrete Event Based Co-simulation
3.1 DE Simulation units . . . . . . . . .
3.2 DE Co-simulation Orchestration . .
3.3 Challenges . . . . . . . . . . . . . . .
3.3.1 Causality . . . . . . . . . . .
3.3.2 Determinism and Confluence
3.3.3 Dynamic Structure . . . . . .
3.3.4 Distribution . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
13
19
19
19
20
20
4 Continuous Time Based Co-simulation
4.1 CT Simulation Units . . . . . . . . . . . . . . . . . . .
4.2 CT Co-simulation Orchestration . . . . . . . . . . . .
4.3 Challenges . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Modular Composition – Algebraic Constraints
4.3.2 Algebraic loops . . . . . . . . . . . . . . . . . .
4.3.3 Consistent Initialization of Simulators . . . . .
4.3.4 Compositional Convergence – Error Control . .
4.3.5 Compositional Stability . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
20
20
22
27
27
31
33
33
36
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4.3.6
4.3.7
Compositional Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Real-time Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
40
5 Hybrid Co-simulation Approach
5.1 Hybrid Co-simulation Scenarios . . . . . . . . . . . . . .
5.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1 Semantic Adaptation . . . . . . . . . . . . . . . .
5.2.2 Predictive Step Sizes . . . . . . . . . . . . . . . .
5.2.3 Event Location . . . . . . . . . . . . . . . . . . .
5.2.4 Discontinuity Identification . . . . . . . . . . . .
5.2.5 Discontinuity Handling . . . . . . . . . . . . . .
5.2.6 Algebraic Loops, Legitimacy, and Zeno Behavior
5.2.7 Stability . . . . . . . . . . . . . . . . . . . . . . .
5.2.8 Theory of DE Approximated States . . . . . . .
5.2.9 Standards for Hybrid Co-simulation . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
41
41
43
43
43
43
44
44
44
45
45
45
6 Classification
6.1 Methodology .
6.2 Taxonomy . . .
6.3 State of the Art
6.4 Discussion . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
45
46
46
47
47
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
7 Concluding Remarks
48
A Historical Perspective of Co-simulation
A.1 One Formalism and Dynamic Iteration . . . . . . . .
A.2 Two Formalisms: Digital and Analog Co-simulation
A.3 Multi-abstraction/Multi-Formalism Co-simulation .
A.4 Black-box Co-simulation . . . . . . . . . . . . . . . .
A.5 Real-time Co-simulation . . . . . . . . . . . . . . . .
A.6 Many simulation units: Large Scale Co-simulation .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
B State of the Art in Co-simulation Frameworks
C Co-Simulation Scenario Categorization
C.1 Non-Functional Requirements . . . . . .
C.1.1 Fault Tolerance . . . . . . . . . .
C.1.2 Configuration Reusability . . . .
C.1.3 Performance . . . . . . . . . . .
C.1.4 IP Protection . . . . . . . . . . .
C.1.5 Parallelism . . . . . . . . . . . .
C.1.6 Distribution . . . . . . . . . . . .
C.1.7 Hierarchy . . . . . . . . . . . . .
C.1.8 Scalability . . . . . . . . . . . . .
C.1.9 Platform Independence . . . . .
C.1.10 Extensibility . . . . . . . . . . .
C.1.11 Accuracy . . . . . . . . . . . . .
2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
77
77
78
78
79
79
79
79
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
123
123
123
123
123
124
125
126
126
127
127
127
127
C.1.12 Open source . . . . . . . . . .
C.2 Simulator Requirements . . . . . . .
C.2.1 Information Exposed . . . . .
C.2.2 Causality . . . . . . . . . . .
C.2.3 Time Constraints . . . . . . .
C.2.4 Rollback Support . . . . . . .
C.2.5 Availability . . . . . . . . . .
C.3 Framework Requirements . . . . . .
C.3.1 Standard . . . . . . . . . . .
C.3.2 Coupling . . . . . . . . . . .
C.3.3 Number of Simulation Units .
C.3.4 Domain . . . . . . . . . . . .
C.3.5 Dynamic structure . . . . . .
C.3.6 Co-simulation Rate . . . . . .
C.3.7 Communication Step Size . .
C.3.8 Strong Coupling Support . .
C.3.9 Results Visualization . . . . .
C.3.10 Communication Approach . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
D List of Acronyms
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
128
128
129
132
134
136
138
140
140
140
142
144
146
146
148
150
152
154
157
Abstract
It is essential to find new ways of enabling experts in different disciplines to collaborate
more efficient in the development of ever more complex systems, under increasing market
pressures. One possible solution for this challenge is to use a heterogeneous model-based
approach where different teams can produce their conventional models and carry out their
usual mono-disciplinary analysis, but in addition, the different models can be coupled for
simulation (co-simulation), allowing the study of the global behavior of the system. Due to its
potential, co-simulation is being studied in many different disciplines but with limited sharing
of findings. Our aim with this work is to summarize, bridge, and enhance future research in
this multidisciplinary area.
We provide an overview of co-simulation approaches, research challenges, and research opportunities, together with a detailed taxonomy with different aspects of the state of the art of
co-simulation and classification for the past five years.
The main research needs identified are: finding generic approaches for modular, stable and
accurate coupling of simulation units; and expressing the adaptations required to ensure that
the coupling is correct.
1
1.1
Introduction
Motivation
Truly complex engineered systems that integrate physical, software and network aspects, are emerging [153, 185]. Due to external pressure, the development of these systems has to be concurrent and
distributed, that is, divided between different teams and/or external suppliers, each in their own
domain and each with their own tools. Each participant develops a partial solution to a constituent
3
system, that needs to be integrated with all the other partial solutions. The later in the process
the integration is done, the less optimal it is [228].
Innovative and truly optimal multi-disciplinary solutions can only be achieved through a holistic
development process [237] where the partial solutions developed independently are integrated sooner
and more frequently. Furthermore, the traditional activities carried out at the partial solution level
—such as requirements compliance check, or design space exploration— can be repeated at the
global level, and salient properties spanning multiple constituent systems can be studied.
Modeling and simulation can improve the development of the partial solutions (e.g., see Friedman
and Ghidella [92]), but falls short in fostering this holistic development process [34]. To understand
why, one has to observe that: a) models of each partial solution cannot be exchanged or integrated
easily, because these are likely developed by different specialized tools; and b) externally supplied
models have Intellectual Property (IP), hence cannot be cheaply disclosed to system integrators.
1.2
Co-simulation
Co-simulation is proposed as a solution to overcome these important challenges. It consists of
the theory and techniques to enable global simulation of a coupled system via the composition of
simulators. Each simulator is a black box mock-up of a constituent system, developed and provided
by the team that is responsible for that system. This allows each team/supplier to work on its
part of the problem with its own tools without having the coupled system in mind, and eases the
relationship between system integrators and suppliers: the latter provide virtual mock-ups of the
solutions they are selling without revealing their IP; the former can perform early conformance
checks and evaluate different designs from multiple competing suppliers.
An alternative to co-simulation is co-modelling, were models are described in a unified language,
and then simulated. There are advantages to this approach but each domain has its own particularities when it comes to simulation (e.g., see [63, 167, 246]) making it impractical to find a language
and simulation algorithm that fits all.
As part of the systematic review that led to the current document (see section 6.1 for details), we took note of the approaches to co-simulation and the publications in applications of
co-simulation. The approaches to co-simulation shaped the taxonomy in section 6.2 and the applications of co-simulation shows that in the last five years, co-simulation has been applied in many
different engineering domains, as fig. 1 shows. In concrete, the publications are:
Automotive - [4, 26, 28, 29, 39, 44, 69, 74, 82, 132, 156, 168, 213, 237, 258, 264, 266]
Electricity Production and Distribution - [3, 5, 33, 78, 95, 96, 111, 120, 139, 157–159, 163,
205, 224, 243, 248, 255, 267]
HVAC - [73, 88, 114, 188, 192, 251]
IC and SoC Design - [209]
Maritime - [194, 195]
Robotics - [143, 198, 199, 265]
A closer look at the publications shows, however, that the average reported co-simulation scenario
includes only two simulators, each a mock-up of a constituent system from a different domain.
While this gives evidence that co-simulation enhances the development multi-domain systems, it
is not yet up-to-par with the scale of Cyber-Physical Systems (CPSs). The unexplored potential
is recognized in a number of completed and ongoing projects that address co-simulation (MOD-
4
ELISAR1 , DESTECS2 , INTO-CPS3 , ACOSAR4 , ACoRTA5 ), and is one of the reasons why the
Functional Mock-up Interface (FMI) Standard was created.
12
11
10
9
Publications
8
Application Domain
Automotive
Electricity Production and Distribution
HVAC
IC and SoC Design
Maritime
Robotics
7
6
5
4
3
2
1
0
2011
2012
2013
2014
2015
2016
Year
Figure 1: Research publications of co-simulation applications over the past five years.
The FMI standard6 defines the interfaces that enable modelling and simulation tools to cooperate
in a simulation, while keeping the IP in the models protected. As of December 2016, 15 tools
implement the standard7 and many others are planning to support it.
The standard was originally created for the coupling of continuous systems, but has been used
(or abused) to simulate hybrid systems. Extensions have been proposed in the literature to facilitate
and standardize this simulation (e.g., zero step size transitions [36, 232], or step size prediction [46]),
and we complement those by adding to the understanding of the the challenges arising in hybrid
system co-simulation. Co-simulation is not a new concept and there are many approaches in the
state of the art, in particular, discrete event based approaches, that have been studied and may
shed light on how hybrid system co-simulation can be performed.
Contribution. We present a survey and a taxonomy, as an attempt to bridge, relate and
classify the many co-simulation approaches in the state of the art.
1 https://itea3.org/project/modelisar.html
2 http://www.destecs.org/
3 http://into-cps.au.dk/
4 https://itea3.org/project/acosar.html
5 http://www.v2c2.at/research/ee-software/projects/acorta/
6 https://www.fmi-standard.org
7 This
number includes only the tools that have passed the crosscheck in https://www.fmi-standard.org/tools
5
1.3
Need for the Survey
Despite the growing interest in the benefits and scientific challenges of co-simulation, to the best
of our knowledge, no existing survey attempts to cover the heterogeneous communities in which it
is being studied. The lack of such a survey means that the same techniques are being proposed
independently with limited sharing of findings. To give an example, the use of dead-reckoning
models is a well known technique in discrete event co-simulation [152], but only very recently it was
used in a continuous time co-simulation approach [221]. Our hypothesis is that bridging these
different co-simulation approaches means that solutions and techniques can be exchanged, and a
deeper understanding of hybrid system co-simulation can be attained.
Scientifically, co-simulation —multi-disciplinary by its very nature— mixes the following fields
of research:
1. Numerical analysis – Accuracy and stability of the coupled system have to be studied [10, 51,
66, 105, 107, 109, 128, 197].
2. Differential Algebraic System Simulation – The composition of co-simulation units is, in the
most general sense, made through algebraic constraints [214, 219, 229].
3. Hybrid Systems – co-simulation scenarios, in the most general sense, are hybrid systems
[47, 154, 172–174, 184, 251, 263]
4. Optimization – the heterogeneous capabilities of co-simulation units pose interesting tradeoffs
[46, 236].
5. Hierarchy – Systems of systems are hierarchical and the corresponding co-simulation scenarios
should be hierarchical as well [96]. Compositionality properties of co-simulations becomes an
interesting research challenge.
6. Formal Verification – The co-simulation orchestration algorithm, also known as the master,
can be certified to be correct under certain assumptions about the co-simulation scenarios
[46, 89, 99, 100, 179].
7. System Testing – Co-simulation can be used for exhaustively testing a set of black-box constituent systems, with a non-deterministic environment [150, 164].
8. Dynamic Structure Systems – Subsystems can have different dependencies depending on
whom, and which level of abstraction, they interact with [21–23, 234].
9. Multi-paradigm Modeling – Subsystems can have have different models at different levels of
abstraction [246]. The relationships between the multiple levels have to be known so that
correct dynamic switching between levels abstraction can be made.
1.4
Outline
To help structure the characteristics of the simulators and how they interact, we distinguish two
main approaches for co-simulation: Discrete Event (DE), described in section 3, and Continuous
Time (CT), described in section 4. Both of these can be, and are, used for the co-simulation
of continuous, discrete, or hybrid coupled systems. We call Hybrid co-simulation, described in
6
section 5, a co-simulation approach that mixes the DE and CT approaches 8 . section 6 categorizes
the features provided by co-simulation frameworks, and classifies the state of the art with that
taxonomy. Finally, section 7 concludes this publication. The section below provides the terminology
used in the rest of the survey.
2
2.1
Modeling, Simulation, and Co-simulation
Dynamical Systems – Models of Real Systems
In this section we present, in an informal manner, the concepts that will be used throughout the
document.
A dynamical system is a model of a real system (for instance a physical system or a computer
system) characterized by a state and a notion of evolution rules. The state is a set of point values
in a state space. The evolution rules describe how the state evolves over an independent variable,
usually time. For example, a traffic light, the real system, can be modeled by a dynamical system
that, at any time, can be in one of four possible states (red, yellow, green, or off). The evolution
rules may dictate that it changes from red to green after some time (e.g., 60 seconds). Another
example is a mass-spring-damper, modeled by a set of first order Ordinary Differential Equations
(ODEs). The equations describe how the state —position and velocity of the mass— changes
continuously over the simulated time. In contrast with the traffic light system, where the state
cannot take an infinite number of different values over a finite duration of simulated time, the state
of the mass-spring-damper can.
The behavior trace is the set of trajectories followed by the state (and outputs) of a dynamical
system. For example, a state trajectory x can be defined as a mapping between a time base T and
the set of reals R, that is, x : T → R. fig. 2 shows a possible behavior trace for each of the example
systems described before. In this example, the time base is R.
Figure 2: Examples of behavior traces.
We refer to the time variable t ∈ T by simulated time —or simply time, when no ambiguity
exists— defined over a time base T (typical the real numbers R), as opposed to the wall-clock time
τ ∈ WcT , which is the time that passes in the real world [94]. When computing the behavior
trace of a dynamical system over an interval [0, t] of simulated time, a computer takes τ units of
8 Note that in this survey, we are focusing on timed formalisms (also called models of computation) and how they
interact in a hybrid co-simulation environment. Other formalisms, with no or only logical notion of time (such as
dataflow and synchronous reactive), are not discussed in this survey. For an overview of formalisms and models of
computation, please see the book on Ptolemy II [200] and the following survey [45].
7
wall-clock time that depend on t. τ can therefore be used to measure the run-time performance of
simulators. fig. 3a highlights different kinds of simulation, based on the relationship between τ and
t. In real-time simulation, the relationship between t and τ is t = ατ , for a given α > 0. In most
cases α = 1 is required, but making sure this is obeyed by the simulation algorithm is one of the
main challenges in real-time simulation, and by extension, of co-simulation. In as-fast-as-possible
—or analytical — simulation, the relationship between τ and t is not restricted. Simulation tools
that offer interactive visualization allow the user to pause the simulation and/or set the relationship
between τ and t.
Knowing when a dynamical system can be used to predict the behavior of a real system is
crucial. The experimental frame describes, in an abstract way, a set of assumptions in which
the behavior trace of the dynamical system can be compared with the one of the real system
[24, 230, 245, 246, 262]. By real system we mean either an existing physical system, or a fictitious
—to be engineered— one. Validity is then the difference between the behavior trace of the dynamical
system and the behavior trace of the real system, measured under the assumptions specified by the
experimental frame. This is what conveys predictive power to dynamical systems. For example,
Hooke’s law, in the mass-spring-damper system, can only be used to predict the reaction force of
the spring for small deformations. For the traffic light dynamical system, the experimental frame
includes the assumption that the transition from the red light to the green light is instantaneous. It
is a valid assumption, provided that the executing platform in which the controller software runs,
has enough computing power [76, 181, 247]. A model is invalid when, within the experimental frame
assumptions, its behavior trace is so different than the one of the real system, that it cannot be
used to predict properties of the real system.
In order to be practical, the behavior trace of a dynamical system has to highlight just the
features of interest of the real system that are relevant for the tasks at hand [145]. In the traffic
light model, the precise amount of wall-clock time a transition from red to green takes is unknown,
but deemed small enough to be neglected. In the mass-spring-damper, Hooke’s law was chosen
because the maximum displacement of the mass will not be large when the context in which the
system will be used is taken into account.
Finally, we consider only those dynamical systems for which it is possible to obtain its meaning,
i.e. the behavior trace, even if only an approximation.
2.2
Simulators – Computing the Behavior Trace
There are two generally accepted ways of obtaining the behavior trace of a dynamical system:
Translational Translate the dynamical system into another model, which can be readily used
to obtain the behavior trace. Obtaining the analytical solution of the mass-spring-damper
equations is an example of this approach. For instance, if the traffic light model is expressed
in the Statechart formalism, it can be translated into a DEVS model, as done in Borland [40],
which can be used to obtain the behavior trace.
Operational Use of a solver – an algorithm that takes the dynamical system as input, and outputs
a behavior trace. For the mass-spring-damper example, a numerical solver can be used to
obtain an approximation of the behavior trace.
We focus on the latter.
A simulator is an algorithm that takes a dynamical system and computes its behavior trace.
Running in a digital computer, it is often the case that a simulator will only be able to approximate
that trace. Two aspects contribute to the error in these approximations: inability to calculate a
8
1.0
● ● ●
●
Solution
x1 approx.
x1 analytical
v1 approx.
v1 analytical
●
●
●
●
●
●
●
●
●
●
●
●
●
0.5
●
●
●
●
●
●
●
●
●
0.0
●
●
●
●
●
●
●
● ● ● ● ● ● ● ● ● ● ●
● ●
● ●
● ●
● ●
● ●
● ●
●
● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ●
● ● ●
● ● ● ● ● ●
● ● ●
● ● ● ● ● ● ●
● ● ●
● ● ●
● ● ● ● ● ● ● ●
● ●
● ● ● ● ●
● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●
● ●
● ●
●
● ●
●
● ●
●
●
●
●
●
●
● ●
● ●
●
●
●
● ●
● ●
●
● ● ● ● ● ● ● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
−0.5
●
●
●
●
● ●
● ● ●
●
●
●
0.0
(a) Classification of time constraints in simulation. Based on
[177, 239].
●
2.5
5.0
Time
7.5
10.0
(b) Approximate behavior trace of the mass-spring-damper system
computed by a Forward Euler solver. Parameters are: m = 1, p = 1,
s = 0, c1 = 1, d1 = 1, and Fe (t) = 0. x is the displacement and v is
the velocity. The dotted lines are the approximated behaviour traces
and the solid lines are the analytical behaviour traces.
Figure 3
trajectory over a continuum, and the finite representation of infinitely small quantities. Simulators
of discrete dynamical systems may also tolerate some inaccuracies in the behavior traces as well
(e.g., if that brings a performance benefit). fig. 3b shows an example approximation of the behavior
trace of the mass-spring-damper system, computed by the Forward Euler. The inaccuracies are
clear when compared to the analytical trace.
In order to define what an accurate simulator is, or even be able to talk about error, we need
to postulate that every dynamical system has an analytical behavior trace. The error can then
be defined as the absolute difference between the behavior trace computed by a simulator and the
analytical trace. An accurate simulator is one that produces traces that are very close9 to the
analytical behaviour. Even if it is not possible to obtain the analytical behavior of every dynamical
system, there are theoretical results that allow simulators to control the error they make. These
techniques are applied to co-simulation in section 4.3.4. For the mass-spring-damper, and linear
ODEs in general, the analytical trace follows a known structure [48]. For the traffic light, and timed
statemachine models in general, the analytical behavior trace can be obtained with a sequential
solver, that respects the causality of events. In short, validity is a property of a dynamical system
whereas accuracy is a property of a simulator [63]. It is perfectly possible to have an accurate
behaviour trace of a model that is invalid, and vice versa [62]. For continuous time systems, the
choice of an appropriate solver is important and should be made by domain experts [63, 167, 175].
9 The
meaning of “very close” depends on the numerical tolerance of the observer.
9
2.3
Simulation Units - Mock-ups of Reality
In strict terms, a simulator is not readily executable: it needs a dynamical system and the input
trajectories to that dynamical system, before being able to compute the behavior trace. We use the
term simulation unit for the composition of a simulator with a dynamical system. A simulation
unit is a replacement of the real system, ready to take inputs and produce a behavior trace. In the
FMI standard, the analogous term is the Functional Mock-up Unit (FMU) for co-simulation.
A simulation is the behavior trace obtained with a simulation unit. The correctness of a simulation unit is dictated by the correctness of the simulation, which depends on the accuracy of the
simulator and the validity of the dynamical system.
2.4
Compositional Co-simulation
As described in section 1, it useful to obtain correct simulations of complex, not yet existing,
systems. Since the constituent systems are developed independently by specialized teams/suppliers,
simulation units can be produced for these. The simulation units can be coupled via their inputs/outputs to produce a behavior trace of the coupled system. A co-simulation, a special kind of
simulation, is the set of simulations computed by the coupled simulation units.
The simulation units can be independent black boxes, possibly running in different computers.
Hence, an orchestrator is necessary to couple them. The orchestrator controls how the simulated
time progresses in each simulation unit and moves data from outputs to inputs according to a cosimulation scenario. A co-simulation scenario is the information necessary to ensure that a correct
co-simulation can be obtained. It includes how the inputs of each simulation unit are computed
from other outputs, their experimental frames, etc. In the FMI Standard, the orchestrator is called
a master algorithm.
Analogously to the simulator and simulation unit concepts, the composition of a specific orchestrator with a co-simulation scenario, yields a co-simulation unit, which is a special kind of simulation
unit, and a substitute of the real coupled system. It follows that a co-simulation is the simulation
trace computed by a co-simulation unit. This characterization enables hierarchical co-simulation
scenarios, where co-simulation units are coupled.
The main purpose of the orchestrator is to obtain a correct behavior trace of the coupled system,
assuming that each simulation unit in itself is correctly defined. We make this assumption because:
(1) the simulation units are often well known parts of the system, and (2) they are developed by
experienced and specialized teams/suppliers.
Co-simulation enables design decisions to be tried out in the model (what-if analysis), cheaply
10
, early in the process, and possibly automatically [87, 104].
In this survey, we focus on the coupling of black box simulation units, where limited knowledge
of the models and simulators is available. However, as will become clear in the sections below,
the black box restriction has to be relaxed so that certain properties related to correctness can
be ensured. Understanding what kind of information should be revealed and how IP can still be
protected is an active area of research in co-simulation.
Most challenges in co-simulation are related to compositionality: if every simulation unit Si in a
co-simulation scenario satisfies property P , then the co-simulation unit, with a suitable orchestrator,
must also satisfy P . The correctness is a property that should be compositional in co-simulation.
Other properties include validity, or accuracy. It is an open research question to ensure that a
10 Another
aspect to consider is the balance between insights gained and resources spent [85].
10
co-simulator is compositional for a given set of properties. The following three sections provide
an overview of the information and techniques being used throughout the state of the art, divided
into three main approaches: discrete event (section 3), continuous time (section 4), and hybrid
(section 5) co-simulation.
3
Discrete Event Based Co-simulation
The Discrete Event (DE) based co-simulation approach describes a family of orchestrators and
characteristics of simulation units that are borrowed from the discrete event system simulation
domain. We start with a description of DE systems, and then we extract the main concepts that
characterize DE based co-simulation.
The traffic light is a good example of a DE system. It can be one of the possible modes:
red, yellow, green or off. The off mode is often used by the police, which in some countries is
characterized by a blinking yellow. Initially, the traffic light can be red. Then, after 60 seconds, it
changes to green. Alternatively, before those 60 seconds pass, some external entity (e.g., a police
officer) may trigger a change from red to off. The output of this system can be an event signaling
its change to a new color. This example captures some of the essential characteristics of a DE
dynamical system: reactivity – instant reaction to external stimuli (turning off by an external
entity); and transiency – a DE system can change its state multiple times in the same simulated
time point, and receive simultaneous stimuli. In the traffic light, transiency would happen if the
light changes always after 0s (instead of 60s), or if the police officer would turn off and on the traffic
light in the same instant.
These characteristics are embraced in DE based co-simulation, where the orchestrator acknowledges that simulation units can evolve the internal state and exchange values despite the fact that
the simulated time is stopped.
3.1
DE Simulation units
A DE simulation unit is a black box that exhibits the characteristics of a DE dynamical system,
but the dynamical system it stands for does not need to be a DE one. Furthermore, it is typical to
assume that DE simulation units communicate with the environment via time-stamped events, as
opposed to signals. This means that the outputs of simulation units can be absent at times where
no event is produced.
We adapt the definition of DEVS11 in [242] (originally proposed in [259]) to formally define a
11 In the original DEVS definition, the initial state and the absent value in the output function are left implicit.
Here we make them explicit, to be consistent with section 4. Note also that there are many other variants of discreteevent formalisms. For instance, DE in hardware description languages (VHDL and Verilog) and actor based systems
(for instance the DE director in Ptolemy II [200]).
11
DE simulation unit Si , where i denotes the reference of the unit:
Si = Xi , Ui , Yi , δiext , δiint , λi , ta i , qi (0)
δiext : Qi × Ui → Xi
δiint : Xi → Xi
λi : Xi → Yi ∪ {φ}
ta i : Xi →
R+
0
(1)
∪∞
qi (0) ∈ Qi
Qi = {(x, e)|x ∈ Xi and 0 ≤ e ≤ ta i (x)}
where:
• Xi , Ui , and Yi are the set of possible discrete states, input events, and output events, respectively;
• δiext (qi , ui ) = x0i is the external transition function that computes a new total state (x0i , 0) ∈ Qi
based on the current total state qi and an input event ui ;
• δiint (xi ) = x0i is the internal transition function that computes a new total state (x0i , 0) ∈ Qi
when the current total state is (xi , ta i (xi )) ∈ Qi ;
• e denotes the elapsed units of time since the last transition (internal or external);
• λi (xi ) = yi ∈ Yi ∪ {φ} is the output event function, invoked right before an internal transition
takes place and φ encodes an absent value;
• ta i (xi ) ∈ R is the time advance function that indicates how much time passes until the next
state change occurs, assuming that no external events arrive;
• qi (0) is the initial state;
The execution of a DEVS simulation unit is described informally as follows. Suppose that the
simulation unit is at time ti ∈ R+
0 and marks the current discrete state as xi for e ≥ 0 elapsed
units of time. Since e ≤ ta i (xi ), the total state is (xi , e) ∈ Qi . Let tn = ti + ta i (xi ) − e. If
no input event happens until tn, then at time tn an output event is computed as yi := λi (xi )
and the new discrete state xi is computed as xi := (δiint (xi ), 0). If, on the other hand, there is
an event at time ts < tn, that is, ui is not absent at that time, then the solver changes to state
xi := (δiext ((xi , e + ts − ti ), ui ), 0) instead.
In the above description, if two events happen at the same time, both are processed before the
simulated time progresses. Due to the transiency and reactivity properties, the state and output
trajectories of a DE simulation unit can only be well identified if the time base, traditionally the
positive real numbers, includes a way to order simultaneous events, and simultaneous state changes.
An example of such a time base is the notion of superdense time [154, 162, 165], where each time
point is a pair (t, n) ∈ T × N , with T typically being the positive real numbers and N , called the
index, is the natural numbers. In this time base, a state trajectory is a function xi : T × N → Vxi ,
where Vxi is the set of values for the state, and an output/input trajectory is ui : T ×N → Vui ∪{φ}.
Simultaneous states and events can be formally represented with incrementing indexes. See [47] for
an introduction.
Equations 2 and 3 show examples of simulation units.
A DE simulation unit is passive: it expects some external coordinator to set the inputs and call
the transition functions. This passivity enables an easier composition of simulation units in a cosimulation, by means of a coordination algorithm, as will be shown later in section 3.2. Algorithm 1
shows a trivial orchestrator which computes the behavior trace of a single DE simulation unit, as
12
specified in eq. (1), that has no inputs. Remarks:
• tl holds the time of the last transition;
• the initial elapsed time satisfies 0 ≤ e ≤ ta i (xi (0));
If Algorithm 1 is used to coordinate the execution of the traffic light simulation unit in eq. (2),
then the resulting behavior trace is the piecewise constant traffic light state x1 (t), together with the
output events. The latter is represented as a trajectory yi (t) that is mostly undefined (or absent),
except for the single points where an output is produced, according to ta 1 .
Algorithm 1: Single autonomous DE simulation unit orchestration. Based on [242] and
originally proposed in [259].
Data: A simulation unit Si = Xi , ∅, Yi , δiext , δiint , λi , ta i , (xi (0), ei ) .
ti := 0 ;
xi := xi (0) ;
// Initial discrete state
tl := −ei ;
// Account for initial elapsed time
while true do
ti := tl + ta i (xi ) ;
// Compute time of the next transition
yi := λi (xi ) ;
// Output
xi := δiint (xi ) ;
// Take internal transition
tl := ti ;
end
3.2
DE Co-simulation Orchestration
DEVS simulation units communicate with their environment exclusively through inputs and outputs. DE co-simulation scenarios are comprised of multiple DE units Si (eq. (1)) coupled through
output to input connections, which map output events of one unit to external events in other unit.
Consider the following DE simulation units of a traffic light and a police office, respectively:
13
D
E
S1 = X1 , U1 , Y1 , δ1ext , δ1int , λ1 , ta 1 , q1 (0)
D
E
S2 = X2 , U2 , Y2 , δ2ext , δ2int , λ2 , ta 2 , q2 (0)
X1 = {red , yellow , green, off }
X2 = {working, idle}
U1 = {toAuto, toOff }
U2 = ∅
Y1 = X1
Y2 = {toWork , toIdle}
(
idle
if x2 = working
int
δ2 (x2 ) =
working if x2 = idle
(
toIdle
if x2 = working
λ2 (x2 ) =
toWork if x2 = idle
(
200 if x2 = working
ta 2 (x2 ) =
100 if x2 = idle
(
δ1ext ((x1 , e), u1 )
=
green
δ1int (x1 ) = yellow
red
green
λ1 (x1 ) = yellow
red
60 if
50 if
ta 1 (x1 ) =
10 if
∞ if
off
red
if u1 = toOff
if u1 = toAuto and x1 = off
if x1 = red
if x1 = green
if x1 = yellow
if x1 = red
if x1 = green
if x1 = yellow
x1
x1
x1
x1
(3)
q2 (0) = (idle, 0)
= red
= green
= yellow
= off
q1 (0) = (red , 0)
(2)
With the following remarks:
• The current state of the model in the definition of δ1ext is q1 = (x1 , e) with e being the elapsed
time since the last transition.
• The output event function λ1 is executed immediately before the internal transition takes
place. It must then publish the next state instead of the current.
To model a scenario where the police officer interacts with a traffic light, the output events Y2
have to be mapped into the external events U1 of the traffic light simulation unit (eq. (2)). In
this example, if U1 = {toAuto, toOff } are the external input events handled by the traffic light
simulation unit, the mapping Z2,1 : Y2 → U1 is defined by:
(
toAuto if y2 = toIdle
Z2,1 (y2 ) =
(4)
toOff
if y2 = toWork
This way, if the police officer changes to working state at time tn, then the output signal y2 :=
toWork will be translated by Z2,1 into an input event u1 := toOff of the traffic light simulation
unit.
Based on the idea of abstract simulation units [262], we formalize and illustrate the idea of a
DE co-simulation scenario with reference cs as follows:
hUcs , Ycs , D, {Sd : d ∈ D} , {Id : d ∈ D ∪ {cs}} , {Zi,d : d ∈ D ∧ i ∈ Id } , Selecti
where:
14
(5)
•
•
•
•
•
Ucs is the set of possible input events, external to the scenario;
Ycs is the set of possible output events from the scenario to the environment;
D is an ordered set of simulation unit references;
For each d ∈ D, Sd denotes a DE unit, as defined in eq. (1);
For each d ∈ D ∪ {cs}, Id ⊆ (D \ {d}) ∪ {cs} is the set of simulation units that can influence
simulation unit Sd , possibly including the environment external to the scenario (cs), but
excluding itself (d);
Zi,d :Ui → Ud , if i = cs
• For each i ∈ Id , Zi,d specifies the mapping of events: Zi,d :Yi → Yd , if d = cs
Zi,d :Yi → Ud , if i 6= cs and d 6= cs
• Select : 2D → D is used to deterministically select one simulation unit, among multiple
simulation units ready to produce output events simultaneously, i.e., when at time t, the set
of simulation units
IMM (t) = {d|d ∈ D ∧ qd (t) = (xd , ta d (xd ))}
(6)
has more than one simulation unit reference. This function is restricted to select one from
among the set IMM (t), i.e., Select(IMM (t)) ∈ IMM (t).
The following co-simulation scenario cs couples the traffic light simulation unit to the police
officer simulation unit:
h∅, Ycs , {1, 2} , {S1 , S2 } , {I1 , I2 , Ics } , {Z2,1 , Z1,cs } , Selecti
Ycs = Y1
I1 = {2}
(7)
I2 = ∅
Ics = {1}
Z1,cs (y1 ) = y1
where:
• S1 is the traffic light simulation unit (eq. (2)) and S2 the police officer simulation unit (eq. (3));
• Y1 is the output of S1 ;
• Z2,1 is defined in eq. (4); and
• The omitted Zi,d functions map anything to absent (φ).
The Select function is particularly important to ensure that the co-simulation trace is unique.
For example, consider the co-simulation scenario of eq. (7), and suppose that at time tn both
simulation units are ready to output an event and perform an internal transition. Should the traffic
light output the event and perform the internal transition first, or should it be the police office to
do it first? In general, the order in which these output/transition actions are performed matters.
The reason is that the way one simulation unit reacts to the other simulation unit’s output may be
different, depending on the internal state of the former. In the example co-simulation scenario, the
end result is always the same but this is not the general case.
Algorithm 2 illustrates the orchestrator of an autonomous (without inputs) DE co-simulation
scenario. It assumes that the co-simulation scenario does not expect external events, that is, all
events that can affect the simulation units are produced by other simulation units in the same
scenario. External output events are possible though. Remarks:
• tcs holds the most recent time of the last transition in the scenario;
• ed is the elapsed time of the current state qd = (xd , ed ) of simulation unit Sd ;
15
•
•
•
•
•
tn is the time of the next transition in the scenario;
i∗ denotes the chosen imminent simulation unit;
Ics is the set of simulation units that can produce output events to the environment;
ycs is the output event signal of the scenario to the environment; and
{d|d ∈ D ∧ i∗ ∈ Id } holds the simulation units that Si∗ can influence.
Algorithm 2: Autonomous DE co-simulation scenario orchestration. Based on [242].
Data: A co-simulation scenario cs = h∅, Ycs , D, {Sd } , {Id } , {Zi,d } , Selecti.
tcs := 0 ;
xi := xi (0) for all i ∈ D ;
// Store initial discrete state for each unit
while true do
ta cs := mind∈D {ta d (xd ) − ed } ;
// Time until the next internal transition
tn := tcs + ta cs ;
// Time of the next internal transition
i∗ := Select(IMM (tn)) ;
// Get next unit to execute
yi∗ := λi∗ (xi∗ ) ;
xi∗ := δiint
// Store new discrete state
∗ (xi∗ ) ;
ei∗ := 0 ;
// Reset elapsed time for the executed unit
if i∗ ∈ Ics then
ycs := Zi∗ ,cs (yi∗ ) ;
// Compute output of the scenario
end
for d ∈ {d|d ∈ D ∧ i∗ ∈ Id } do
ud := Zi∗ ,d (yi∗ ) ; // Trigger internal units that are influenced by unit i∗
xd := δdext ((xd , ed + ta cs ), ud ) ;
ed := 0 ;
end
for d ∈ {d|d ∈ D ∧ i∗ 6∈ Id } do
ed := ed + ta cs ;
// Update the elapsed time of the remaining units
end
tcs := tn ;
// Advance time
end
fig. 4 shows the behavior trace of the co-simulation scenario in eq. (7).
Algorithm 2 is similar to Algorithm 1:
• The time advance of the scenario ta cs corresponds to the time advance of a single simulation
unit.
• The output produced by the state transition is analogous to the λ function of a single simulation unit.
• The output and state transition of child Si∗ , together with the external transitions of the simulation units influenced by Si∗ , are analogous to the internal transition of a single simulation
unit.
It is natural then that a co-simulation scenario cs as specified in eq. (5), can be made to behave as
a single DE simulation unit Scs . Intuitively, the state of Scs is the set product of the total states
of each child DE unit; ta cs is the minimum time until one of the DE units executes an internal
transition; the internal transition of Scs gets the output event of the imminent unit, executes the
external transitions of all the affected units, updates the elapsed time of all unaffected units, and
16
computes the next state of the imminent unit; the external transition of Scs gets an event from
the environment, executes the external transition of all the affected units, and updates the elapsed
time of all the unaffected units [262]. Formally:
ext int
Scs = Xcs , Ucs , Ycs , δcs
, δcs , λcs , ta cs , qcs (0)
Xcs = ×d∈D Qd
qcs (0) = (×d∈D qi (0), min ed )
d∈D
ta cs ((. . . , (xd , ed ), . . .)) = min {ta d (xd ) − ed }
d∈D
∗
i = Select(IMM (t))
(
Zi∗ ,cs (yi∗ (tn))
λcs (xcs ) =
φ
if i∗ ∈ Ics
otherwise
int
δcs
(xcs ) = (. . . , (x0d , e0d ), . . .), for all d ∈ D, where:
(8)
xcs = (. . . , (xd , ed ), . . .)
int
∗
(δd (xd ), 0) if i = d
(x0d , e0d ) = (δdext ((xd , ed + ta cs (xcs )), Zi∗ ,d (λi∗ (xi∗ )), 0) if i∗ ∈ Id
(xd , ed + ta cs (xcs )) otherwise
ext
δcs
((xcs , ecs ) , ucs ) = (. . . , (x0d , e0d ), . . .), for all d ∈ D, where:
xcs = (. . . , (xd , ed ), . . .)
(
(δdext ((xd , ed + ecs ) , Zcs,d (ucs )), 0) if cs ∈ Id
(x0d , e0d ) =
(xd , ed + ecs ) otherwise
Remarks:
It is the Cartesian product of the total state of each child simulation unit that makes the discrete
state of the co-simulation unit;
The elapsed times of each child simulation unit are managed solely by the co-simulation unit,
whenever there is a transition (internal or external);
The external transition functions of each child are executed with the mapping of the events produced
by the current state of the imminent child, and not the next one computed by (δdint (xd ), 0);
An internal transition of a child simulation unit may cause an output event to the environment of
the co-simulation unit, if the child is connected to the output of the co-simulation unit.
The same internal transition causes not only a change in the child discrete state, but also, due
to its output event, may cause external transitions in other child simulation units. This is not a
recursive nor iterative process: at most one external transition will occur in all the affected child
simulation units; if any of the affected simulation units becomes ready for an internal transition, it
waits for the next internal transition invoked from the coordinator of the co-simulation unit;
The resulting co-simulation unit Scs behaves exactly as a DE unit specified in eq. (1). It
can thus be executed with Algorithm 1 (in case of no inputs), or composed with other units in
hierarchical co-simulation scenarios. Hierarchical co-simulation scenarios can elegantly correspond
to real hierarchical systems, a natural way to deal with their complexity [138].
In summary, DE based co-simulation exhibits the following characteristics:
17
Figure 4: Example co-simulation trace of the traffic light and police officer scenario.
reactivity: A DE simulation unit (analogously, a DE co-simulation unit) has to process an event
at the moment it occurs.
transiency: In both Algorithm 2 and in a DE co-simulation unit, the time advance ta cs to the
next imminent child internal transition can be zero for successive iterations, so an orchestrator
has to be able to tolerate the fact that simulated time may not advance for several iterations.
predictable step sizes: In a DE co-simulation scenario without inputs, the orchestrator, as shown
in Algorithm 2, can always predict the next simulated time step. In a scenario with inputs,
if the environment provides the time of the next event, then the next simulated time step
can be predicted too. For this to be possible, black box DE simulation units have to be able
to inform the orchestrator what their time advance is, not a trivial task for DE units that
simulate continuous systems whose future behavior trace, especially reacting to future inputs,
is not easily predicted without actually computing it.
In the next sub-section, the main challenges in DE based co-simulation, and the requirements
(or capabilities) their solutions impose in DE simulation units, are made explicit.
18
3.3
3.3.1
Challenges
Causality
For the sake of simplicity, Algorithm 2 is sequential. In a hierarchical co-simulation unit, the
imminent simulation unit (closest to performing an internal transition) will be the one to execute,
thus inducing that there is a global order in the events that are exchanged. This global order avoids
causality violations but is too pessimistic. If an event y1 (t1 ) causes another event —by changing the
internal state of some other simulation unit, which in turn changes its next output event— y2 (t2 ),
then t1 ≤ t2 , which is ok. However, the converse is not true: t1 ≤ t2 does not necessarily imply that
y1 (t1 ) has caused y2 (t2 ), which means that simulation unit S2 could execute before —in the wallclock time sense— y1 (t1 ) without violating causality, at least within a small window of simulated
time. To see why, suppose that S1 and S2 do not influence each other in the scenario. Then y2 (t2 )
would happen anyway, regardless of y1 (t1 ) occurring or not. Moreover, the co-simulation scenario
holds information —the dependencies {Id }— that can be used to determine who influences what
[64, 147].
A parallel optimistic orchestrator that takes {Id } into account is, in general, faster in the wall
clock time sense, than a pessimistic, sequential one. However, most of these, the Time-warp algorithm [123, 124] being a well known example, require rollback capabilities of simulation units. This
is because simulation units proceed to advance their own time optimistically, assuming that any
other simulation units will not affect them, until they are proven wrong by receiving an event which
occurs before their own internal time. When that happens, the simulation unit has to rollback to a
state prior to the time of timestamp of the event that just arrived. This may in turn cause a cascade
of rollbacks in other affected simulation units. Moreover, in parallel optimistic DE co-simulation,
any of the units in the scenario needs (theoretically) to support multiple rollbacks and have enough
memory to do so for an arbitrary distant point in the past [94]. This point in the past is limited in
Time-warp by the Global Virtual Time (GVT). The GVT represents the minimum internal time
of all simulation units. By definition, no event that is yet to be produced (in wall-clock time) can
have a timestamp smaller than the GVT.
We make the distinction between multiple rollback, from single rollback capabilities. To support
single rollback, a simulation unit needs to store only the last committed state, thereby saving
memory.
Causality is a compositionality property: if each child simulation unit does not violate causality,
then any orchestrator has to ensure that the causality is not violated when these units are coupled.
Optimistic orchestration algorithms do so by requiring rollback capabilities from child simulation
units, whereas pessimistic algorithms do so at the cost of performance.
3.3.2
Determinism and Confluence
Determinism is also a compositional property. The Select function, in the co-simulation scenario
definition of eq. (5), is paramount to ensure the compositionality of deterministic behavior. This
function is used to ensure that a unique behavior trace can be obtained when the co-simulation
scenario is executed by Algorithm 2 or when it is turned into a co-simulation unit, as in eq. (8). The
alternative to the Select function is to ensure that all possible interleavings of executions always
lead to the same behavior trace – this is known as “confluence”. Intuitively, if a co-simulation unit is
compositional with respect to confluence, then it is also compositional with respect to determinism.
Proving confluence is hard in general black box DE co-simulation because it depends on how
19
the child simulation units react to external events: potentially valuable IP. Parallel-DEVS [67] is
an approach, which leaves the confluence property to be satisfied by the modeler.
3.3.3
Dynamic Structure
Until now, the dependencies {Id }, in eq. (5), have been assumed to be fixed over time. From a
performance perspective, a static sequence of dependencies may be too conservative, especially if
used to ensure causality in optimistic parallel co-simulation. To see why, consider that in a large
scale simulation, there is a simulation unit S1 which may influence simulation unit S2 but only under
a very specific set of conditions, which may not be verified until a large amount of simulated time
has passed. A pessimistic co-simulation unit assumes that S1 may always affect S2 and hence, tries
to ensure that the simulated time of S2 is always smaller than S1 , to minimize possible rollbacks.
This incurs an unnecessary performance toll in the overall co-simulation because S1 does not affect
S2 most of the time. This is where making I2 dynamic can improve the performance of the cosimulation since the co-simulation unit will know that most of the time, S1 does not affect S2 .
Dynamic structure co-simulation allows for {Id } to change over time, depending on the behavior
trace of the simulation units. It can be used to study self-organizing systems [21, 235].
3.3.4
Distribution
Co-simulation units whose child simulation units are geographically distributed are common [94].
Interesting solutions like computation allocation [180, 240], bridging the hierarchical encapsulation
[241], and the use of dead-reckoning models [152], have been proposed to mitigate the additional
communication cost. Moreover, security becomes important, and solutions such as [187] address it.
4
Continuous Time Based Co-simulation
In the continuous time (CT) based co-simulation approach, the orchestrators’ and simulation units’
behavior and assumptions are borrowed from the continuous time system simulation domain. We
describe these below.
4.1
CT Simulation Units
A continuous time simulation unit is assumed to have a state that evolves continuously over time.
It is easier to get the intuitive idea of this by considering a simulation unit of a continuous time
dynamical system, such as a mass-spring-damper, depicted in the left hand side of fig. 5. The state
is given by the displacement x1 and velocity v1 of the mass, and the evolution by:
x˙1 = v1
m1 · v˙1 = −c1 · x1 − d1 · v1 + Fe
x1 (0) = p1
(9)
v1 (0) = s1
where ẋ denotes the time derivative of x; c1 is the spring stiffness constant and d1 the damping
coefficient; m1 is the mass; p1 and s1 the initial position and velocity; and Fe denotes an external
20
input force acting on the mass over time. The solutions x1 (t) and v1 (t) that satisfy eq. (9) constitute
the behavior trace of the dynamical system. fig. 3b shows an example of such trace.
eq. (9) can be generalized to the state space form:
ẋ = f (x, u) ; y = g(x, u) ; x(0) = x0
(10)
where x is the state vector, u the input and y the output vectors, and x0 is the initial state.
T
A solution [x(t), y(t)] that obeys eq. (10) is the behavior trace of the system. If f is linear and
time-invariant, an analytical form for x(t) can be obtained [17]. An analytical solution obtained
by the application of mathematical identities is an example of a behavior trace obtained via the
translational approach, described in section 2.2. Alternatively, the behavior trace can be computed.
If f (x, u) is sufficiently differentiable, x can be approximated with a truncated Taylor series
[63, 227]:
(11)
x(t + h) = x(t) + f (x(t), u(t)) · h + O h2
where
O hn+1 = max
i
xn+1 (ζ(t∗ )) n+1
h
h→0
(n + 1)!
lim
= const · hn+1
denotes the order of the truncated residual term; t∗ ∈ [t, t + h]; and h ≥ 0 is the micro-step size.
eq. (11) is the basis of a family of numerical solvers that iteratively compute an approximated
behavior trace x̃. For example, the Forward Euler method is given by:
x̃(t + h) := x̃(t) + f (x̃(t), u(t)) · h
(12)
x̃(0) := x(0)
A CT simulation unit is assumed to have a behavior that is similar to one of a numerical solver
computing a set of differential equations. We reinforce that this does not restrict CT simulation
units to being mockups of CT systems, even though it is easier to introduced them as such. In
the FMI Standard, a simulation unit is analogous to a Functional Mock-up Unit (FMU) for cosimulation. For example, a simulation unit S1 of the mass-spring-damper, using the Forward Euler
solver, can be written by embedding the solver (eq. (12)) into eq. (9):
x̃1 (t + h1 ) := x̃1 (t) + v1 (t) · h1
1
ṽ1 (t + h1 ) := ṽ1 (t) +
· (−c1 · x̃1 (t) − d1 · ṽ1 (t) + Fe (t)) · h1
m1
x̃1 (0) := p1
ṽ1 (0) := s1
T
where h1 is the micro-step size, Fe (t) is the input, and [x(t + h), v(t + h)] is the output.
21
(13)
4.2
CT Co-simulation Orchestration
Consider now a second system, depicted in the right hand side of fig. 5. It is governed by the
differential equations:
x˙2 = v2
m2 · v˙2 = −c2 · x2 − Fc
Fc = cc · (x2 − xc ) + dc · (v2 − x˙c )
(14)
x2 (0) = p2
v2 (0) = s2
where cc and dc denote the stiffness and damping coefficients of the spring and damper, respectively;
xc denotes the displacement of the left end of the spring-damper. Combining with the Forward
Euler solver, yields the following simulation unit:
x̃2 (t + h2 ) := x̃2 (t) + ṽ2 (t) · h2
1
· (−c2 · x̃2 (t) − Fc (t)) · h2
ṽ2 (t + h2 ) := ṽ2 (t) +
m2
Fc (t) = cc · (x̃2 (t) − xc (t)) + dc · ṽ2 (t) − xc˙(t)
(15)
x̃2 (0) := p2
ṽ2 (0) := s2
where h2 is the micro-step size, xc and x˙c are inputs, and Fc the output. Suppose S1 (eq. (13)) and
S2 are coupled, setting xc = x1 , x˙c = v1 and Fe = Fc , so that the resulting co-simulation scenario
represents the multi-body system depicted in fig. 5.
Figure 5: A multi-body system comprised of two mass-spring-damper subsystems.
In the co-modeling approach, the models in Equations 9 and 14 would be combined to get the
22
following coupled model:
x˙1 = v1
m1 · v˙1 = −c1 · x1 − d1 · v1 + Fc
x˙2 = v2
m2 · v˙2 = −c2 · x2 − Fc
Fc = cc · (x2 − x1 ) + dc · (v2 − v1 )
(16)
x1 (0) = p1
v1 (0) = s1
x2 (0) = p2
v2 (0) = s2
which can be written in the state space form (eq. (10)) as:
0
x˙1
v˙1 − c1m+cc
1
=
x˙2 0
cc
v˙2
m
2
p1
x1 (0)
v1 (0) s1
x2 (0) = p2
v2 (0)
s2
1
0
c
− d1m+d
1
0
cc
m1
dc
m2
0
c
− c2m+c
2
x1
dc
m1 v1
1 x2
dc
v2
−m
2
0
(17)
The behavior trace of eq. (17) can be obtained either analytically, or with Forward Euler solver
(eq. (12)).
In CT based co-simulation, to overcome the fact the each simulation unit’s micro-step sizes
are independent, a communication step size H (also known as macro-step size or communication
grid size) has to be defined. H marks the times at which the simulation units exchange values of
inputs/outputs.
Suppose a simulation unit Si is at time n · H, for some natural n, and it being asked by an
orchestrator to execute until time (n + 1) · H. If Si only gets its inputs valued at n · H, then
extrapolation must be used to associate a value with those inputs in any of the internal micro-steps
of the simulation unit. In other words, when time n · H + m · hi , for m ≤ hHi and micro-step size hi ,
an extrapolation function φui (m · hi , ui (n · H), ui ((n − 1) · H), . . .), built from input values known
at previous communication time points, is used to approximate the value of ui (n · H + m · hi ).
Notice that m = hHi is allowed, even though, theoretically, the value of ui ((n + 1) · H) can be
obtained from the environment. The reason for this becomes clear in section 4.3.2. Analogously,
interpolation techniques have to be used when the orchestrator makes the input value available at
time (n + 1) · H but the simulation unit is still at time n · H.
For example, the input Fe of the simulation unit described in eq. (13) can be defined as:
Fe (n · H + m · h1 ) := φFe (m · h1 , Fe (n · H), Fe ((n − 1) · H), . . .), for m ≤
23
H
h1
(18)
Similarly, the inputs xc and x˙c of the simulation unit described in eq. (15) can be defined as
xc (n · H + m · h2 ) := φxc (m · h2 , xc (n · H), xc ((n − 1) · H), . . .)
x˙c (n · H + m · h2 ) := φx˙c (m · h2 , x˙c (n · H), x˙c ((n − 1) · H), . . .)
H
for m ≤
h2
(19)
In the simplest case, the extrapolations can be constant. In the coupled mass-spring-dampers
example:
φFe (t, Fe (n · H))) = Fe (n · H))
φxc (t, xc (n · H)) = xc (n · H)
(20)
φx˙c (t, x˙c (n · H)) = x˙c (n · H)
In the state of the art, input extrapolation approaches can be classified by increasing degree
of complexity: Constant; Linear; Polynomial; Extrapolated-Interpolation [49, 50, 75]; Contextaware [134]; and Estimated Dead-Reckoning Model [43, 221]; These can, and often are, combined
in practical use cases. See Andersson [8], Arnold [10], Busch [50], Schweizer et al. [219] for an
overview of linear and higher order extrapolation techniques and how these affect the accuracy of
the co-simulation trace.
The orchestrator for this co-simulation scenario, at a time t = n · H, gets the outputs of both
simulation units and computes their inputs. Then, each simulation unit is instructed to compute
its behavior trace until the next communication step size, at t = (n + 1) · H, making use of the
extrapolating functions to get the inputs at each of the micro steps (Equations 18 and 19).
We are ready to formally define the behavior of a CT simulation unit Si :
Si = hXi , Ui , Yi , δi , λi , xi (0), φUi i
δi : R × Xi × Ui → Xi
λi : R × Xi × Ui → Yi or R × Xi → Yi
(21)
xi (0) ∈ Xi
φUi : R × Ui × . . . × Ui → Ui
where:
• Xi is the state set, typically Rn ;
• Ui is the input set, typically Rm ;
• Yi is the output set, typically Rp ;
• δi (t, xi (t), ui (t)) = xi (t + H) or δi (t, xi (t), ui (t + H)) = xi (t + H) is the function that instructs
the simulation unit to compute a behavior trace from t to t + H, making use of the input
extrapolation (or interpolation) function φUi ;
• λi (t, xi (t), ui (t)) = yi (t) or λi (t, xi (t)) = yi (t) is the output function; and
• xi (0) is the initial state.
24
For instance, the simulation unit in eq. (13) can be described as follows:
p
S1 = R2 , R, R2 , δ1 , λ1 , 1 , φFe
s1
x̃ (t)
x̃ (t + H)
δ1 (t, 1
, Fe (t)) = 1
ṽ1 (t)
ṽ1 (t + H)
x̃ (t)
x̃ (t)
λ1 (t, 1
)= 1
ṽ1 (t)
ṽ1 (t)
(22)
T
where [x̃1 (t + H), ṽ1 (t + H)] is obtained by the iterative application of the simulation unit in
eq. (13) over a finite number of micro-steps, making use of the extrapolation of Fe (defined in
eq. (18)):
x̃1 (t + H)
x̃1 (t)
x˙1 (t)
x˙1 (t + h)
=
+
·h+
· h + ...
ṽ1 (t + H)
ṽ1 (t)
v˙1 (t, φFe (t, Fe (t), . . .))
v˙1 (t + h, φFe (t + h, Fe (t), . . .))
A continuous time co-simulation scenario with reference cs includes at least the following information12 :
hUcs , Ycs , D, {Si : i ∈ D} , L, φUcs i
(23)
L : (Πi∈D Yi ) × Ycs × (Πi∈D Ui ) × Ucs → Rm
where D is an ordered set of simulation unit references, each Si is defined as in eq. (21), m ∈ N,
Ucs is the space of inputs external to the scenario, Ycs is the space of outputs of the scenario, φUcs a
set of input approximation functions, and L induces the simulation unit coupling constraints (e.g.,
if D = {1, . . . , n}, then the coupling is L(y1 , . . . , yn , ycs , u1 , . . . , un , ucs ) = 0̄).
As an example, the co-simulation scenario representing the system of fig. 5 is:
cs = h∅, ∅, {1, 2} , {S1 , S2 } , L, ∅i
x c − v1
L = x˙c − x1
Fe − Fc
(24)
where:
• S1 is the simulation unit for the constituent system on the left (eq. (22)), and S2 is the
simulation unit for the remaining constituent system;
• xc , x˙c are the inputs of S2 , and Fe is the input of S1 ; and
• x1 , v1 are outputs of S1 and Fc is the output of S2 .
Algorithm 3 summarizes in a generic way the tasks of the orchestrator for computing the cosimulation of a scenario cs with no external inputs. It represents the Jacobi communication approach: simulation units exchange values at time t and independently compute the trace until the
next communication time t + H. The way the system in eq. (25) is solved depends on how the
simulation units are coupled, that is, the definition of L. In the most trivial case, the system reduces
to an assignment of an output yj (t) to each input ui (t), and so the orchestrator just gets the output
of each simulation unit and copies it onto the input of some other simulation unit, in an appropriate
order. Concrete examples of Algorithm 3 are described in [25, 52, 79, 93, 96, 105, 140, 253].
12 Please note that this formalization is related to the formalization proposed by Broman et al. [46], with the main
differences: i) it is not designed to formalize a subset of the FMI Standard, ii) it accommodates algebraic coupling
conditions, and iii) it does not explicitly define port variables.
25
An alternative to the Jacobi communication approach is the Gauss-Seidel (a.k.a. sequential or
zig-zag) approach, where an order of the simulation units’ δ function is forced to ensure that, at
time t, they get inputs from a simulation unit that is already at time t + H. Gauss-Seidel approach
allows for interpolations of inputs, which is more accurate, but hinders the parallelization potential.
Examples are described in [10, 11, 25, 59, 236].
Algorithm 3: Generic Jacobi based orchestrator for autonomous CT co-simulation scenarios.
Data: An autonomous scenario cs = h∅, Ycs , D = {1, . . . , n} , {Si } , L, ∅i and a
communication step size H.
Result: A co-simulation trace.
t := 0 ;
xi := xi (0) for i = 1, . . . , n ;
while true do
Solve the following system for the unknowns:
y1 = λ1 (t, x1 , u1 )
...
(25)
y
=
λ
(t, xn , un )
n
n
L(y1 , . . . , yn , ycs , u1 , . . . , un ) = 0̄
xi := δi (t, xi , ui ), for i = 1, . . . , n ; // Instruct each simulation unit to advance
to the next communication step
t := t + H ;
// Advance time
end
Similarly to DE based co-simulation, a CT co-simulation scenario, together with an orchestrator,
should behave as a (co-)simulation unit of the form of eq. (21), and thus be coupled with other
simulation units, forming hierarchical co-simulation scenarios: the state of the co-simulation unit is
the set product of the states of the internal units; the inputs are given by Ucs and the outputs by
Ycs ; the transition and output functions are implemented by the orchestrator; the communication
step size H used by the orchestrator is analogous to a simulation unit’s micro-step sizes, and the
input extrapolation function is φUi .
Algorithm 3 makes it clear that the simulation units can be coupled with very limited information
about their internal details. In concrete:
• The output λi and state transition δi functions need to be executable but their internal details
can remain hidden;
• the inputs ui need to be accessible;
• the state variables can be hidden. These are represented merely to illustrate that the internal
state of the simulation unit changes when executing δi .
However, the blind coupling can lead to compositionality problems, as will be discussed in the sections below. The common trait in addressing these is to require more from the individual simulation
units: either more capabilities, or more information about the internal (hidden) dynamical system.
26
Figure 6: A multi-body system coupled by a mass-less link, based on the example provided in
Schweizer and Lu [214].
4.3
4.3.1
Challenges
Modular Composition – Algebraic Constraints
In the co-simulation scenario described in eq. (24), the coupling condition L translates into a set of
assignments from outputs to inputs. This is because the inputs of the simulation unit of the system
in the left hand side of fig. 5 and the outputs of the simulation unit of the system represented in
the right hand side of the same picture can be connected directly, and vice versa. In practice, the
simulation units’ models are not created with a specific coupling pattern in mind and L can be more
complex. As an example, consider the system coupled by a massless rigid link, depicted in fig. 6.
The first subsystem is the same as the one in the left hand side of fig. 5 and its simulation unit is
in eq. (13). The second constituent system is governed by the following differential equations:
x˙3 = v3
m3 · v˙2 = −c3 · x3 + Fc
x3 (0) = p3
(26)
v3 (0) = s3
And the following simulation unit:
x̃3 (t + h3 ) = x̃3 (t) + v3 (t) · h3
1
ṽ3 (t + h3 ) = ṽ3 (t) +
· (−c3 · x3 (t) + Fc (t)) · h3
m3
x̃3 (0) = p3
(27)
ṽ3 (0) = s3
The input to the simulation unit S3 is the coupling force Fc , and the output is the state of the mass
T
[x̃3 , ṽ3 ] . The input to the simulation unit S1 is the external force Fe and the outputs are the state
T
T
of the mass [x̃1 , ṽ1 ] . Recall eq. (13). There is clearly a mismatch. The outputs [x̃1 , ṽ1 ] of the first
simulation unit cannot be coupled directly to the input Fc of the second simulation unit, and vice
versa. However, the massless link restricts the states and inputs of the two units to be the same.
Whatever the input forces may be, they are equal and opposite in sign. Hence, any orchestration
27
algorithm has to find inputs that ensure the coupling constraints are satisfied:
x̃1 (n · H) − x̃3 (n · H)
L = ṽ1 (n · H) − ṽ3 (n · H) = 0̄
Fe (n · H) + Fc (n · H)
(28)
This problem has been addressed in Arnold [10], Arnold and Günther [11], Gu and Asada
[105, 106, 107], Schweizer and Lu [214, 215, 216], Schweizer et al. [218, 219], Sicklinger et al. [220].
The approach taken in Gu and Asada [105] is worth mentioning because it defines a Boundary
Condition Coordinator (BCC) which behaves as an extra simulation unit, whose inputs are the
outputs of the original two simulation units, and whose outputs are Fe and Fc . They show that the
initial co-simulation scenario with the non-trivial constraint can be translated into a co-simulation,
with a trivial constraint, by adding an extra simulation unit. This is illustrated in fig. 7.
Figure 7: Transforming a co-simulation scenario with a non-trivial constraint into a simpler scenario
by adding an extra simulation unit that induces a trivial constraint. This promotes separation of
concerns.
Transforming the co-simulation scenario to make it simpler marks an important step in separating the concerns of the orchestrator [102]. In fact, the newly created simulation unit can be run
with a smaller internal micro-step size, required to meet stability and accuracy criteria, as shown
in Gu and Asada [105].
In many of the solutions proposed (e.g., [10, 11, 214, 219, 220]), information about the rate of
change (or sensitivity) of outputs and states of each simulation unit, with respect to changes in
its inputs is required to solve the non-trivial coupling condition. This information can be either
provided directly as a Jacobian matrix of the system and output functions, or estimated by finite
differences, provided that the simulation units can be rolled back to previous states. A frequent
28
characteristic of co-simulation: the availability of certain capabilities from simulation units can
mitigate the lack of other capabilities.
To show why the sensitivity information is useful, one of the tasks of the BCC is to ensure that
x̃1 − x̃3 is as close to zero as possible, by finding appropriate inputs Fe and Fc . This is possible
since x̃1 and x̃3 are functions of the inputs Fe and Fc , and −Fe = Fc . So the constraint can be
written as
g(Fe ) = x̃1 (Fe ) − x̃3 (−Fe ) = 0
(29)
From one communication step to the next, g can be expanded with the Taylor series:
g(Fe ((n + 1) · H)) = g(Fe (n · H) + ∆Fe ) ≈ g(Fe (n · H)) +
∂g(Fe (n · H))
· ∆Fe
∂Fe
(30)
From a known input Fe (n · H), Equations 29 and 30 can be combined to obtain the input
Fe ((n + 1) · H) at the next communication step:
∂g(Fe (n · H))
· ∆Fe = 0 ↔
∂Fe
∂g(Fe (n · H))
g(Fe (n · H)) = −
· ∆Fe ↔
∂Fe
−1
∂g(Fe (n · H))
∆Fe = −
· g(Fe (n · H)) ↔
∂Fe
−1
∂g(Fe (n · H))
Fe ((n + 1) · H) = Fe (n · H) −
· g(Fe (n · H))
∂Fe
g(Fe (n · H) + ∆Fe ) ≈ g(Fe (n · H)) +
(31)
with
∂g(Fe (n · H))
∂ x̃1 (Fe (n · H)) ∂ x̃3 (−Fe (n · H))
=
+
∂Fe
∂Fe
∂Fc
(32)
A simple orchestration algorithm will then perform the following steps, at each co-simulation
step:
1. Let x̃1 (nH), x̃3 (nH) be the current position outputs of the two simulation units S1 and S3 ;
2. Perform a co-simulation step with a known Fe , obtaining x̃p1 (nH), x̃p3 (nH) as new outputs.
3. Rollback simulation units to state x̃1 (nH), x̃3 (nH);
4. Perform a co-simulation step with Fe + ∆Fe , obtaining x̃d1 (nH), x̃d3 (nH);
e (n·H))
5. Approximate ∂g(F∂F
by finite differences and eq. (32);
e
6. Obtain a corrected Fec by eq. (31);
7. Rollback simulation units to state x̃1 (nH), x̃3 (nH);
8. Perform the final co-simulation step with Fec ;
9. Commit states and advance time;
As can be seen in fig. 8, this coupling cannot be carried out without errors: the constraint
g(Fe ((n + 1) · H)) cannot be accurately forced to zero at first try. Furthermore, finding initial
conditions and initial inputs that satisfy Equations 9, 26, and 28 is very important and usually
requires a fixed point iteration. The above algorithm could be changed to perform an arbitrary
29
Coupling
20
10
0
−10
−20
Solution
x1
x3
x_analytical
Fe
1.0
Trajectory
0.5
0.0
−0.5
0.0
2.5
5.0
Time
7.5
10.0
Figure 8: Co-simulation of algebraically coupled masses. Parameters are: m2 = 2, m1 = c1 = c3 =
d1 = cc = 1, H = 0.1, x1 (0) = 1.0, x3 (0) = 1.1, v1 (0) = v3 (0) = 0. Notice the small disturbance at
the initial conditions.
number of iterations, repeating steps 1–7 until g(Fe ((n + 1) · H)) is close enough to zero. This would
increase the accuracy but also increase the amount of computation.
These examples show that rollback capabilities are important. If a simulation unit is a black
box, then the rollback capability has to be provided by the simulation unit itself and there is little
that the orchestrator can do to make up for the lack of the feature. See Broman et al. [46] for an
orchestrator that takes into account the existence of the rollback feature. If, on the other hand, the
simulation unit provides access to its state, and allows the state to be set, as in Blockwitz et al. [35],
then the orchestrator can implement the rollback by keeping track of the state of the simulation
unit. Rollback also plays a key role when dealing with algebraic loops in the co-simulation scenario.
Finally, to explain why this sub-section refers to modular composition of simulation units, the
example in fig. 6 makes explicit one of the problems in co-simulation: the “rigid” and protected
nature of simulation units can make their coupled simulation very difficult. To contrast, in a white
box approach where the equations of both constituent systems are available, the whole system is
simplified, with the two masses being lumped together, and their coupling forces canceling each
other out. The simplified system is a lumped mass-spring-damper, which is easily solvable. Such
an approach is common in a-causal modeling languages, such as Modelica [1]. To be concrete, the
coupled system is obtained by combining Equations 9, 26, and 28, and simplifying to:
x˙1 = v1
(m1 + m3 ) · v˙1 = −(c1 + c3 ) · x1 − d1 · v1
x1 (0) = p1
(33)
v1 (0) = s1
fig. 9 compares the behavior trace produced by Algorithm 3 when applied to the co-simulation
scenario described in eq. (24), with the analytical solution, obtained from the coupled model of
eq. (17) (co-modelling). It is obvious that there is an error due to the extrapolation functions and
the large communication step size H = 0.1.
30
1.0
0.5
Solution
x1 cosim.
x1 analytical
x2 cosim.
x2 analytical
0.0
−0.5
−1.0
0.0
2.5
5.0
Time
7.5
10.0
Figure 9: Comparison of co-simulation with co-modelling for the sample coupled system. Parameters are: m1 = m2 = c1 = c2 = d1 = cc = 1, H = 0.1.
This is more modular because (if the equations are made available) the same constituent system
can be coupled to other systems in many different contexts, without further changes. As this
sub-section shows, in co-simulation it is possible to get around the modularity aspect, but at a cost.
4.3.2
Algebraic loops
Algebraic loops occur whenever there is a variable that indirectly depends on itself. To see how
algebraic loops arise in co-simulation scenarios, recall (see eq. (21)) that the state evolution and
output of each simulation unit Si can be written as:
xi (t + H) = δi (t, xi (t), ui (t))
yi (t + H) = λi (t, xi (t + H), ui (t + H))
(34)
To simplify things, assume that the simulation units are coupled by a set of assignments from
outputs to inputs, i.e.,
ui (t) := yj (t)
(35)
where ui is the input of simulation unit Si and yj the output of a simulation unit Sj , in the same
co-simulation scenario.
With these definitions, it is easy to see that, depending on the coupling assignments of the
co-simulation scenario, the output of a simulation unit may depend on itself, that is,
yi (t + H) = λi (t, xi (t + H), ui (t + H))
ui (t + H) = yj (t + H)
yj (t + H) = λj (t, xj (t + H), uj (t + H))
uj (t + H) = yk (t + H)
...
uz (t + H) = yi (t + H)
31
(36)
We distinguish two kinds of algebraic loops in co-simulation [141]: the ones spanning just input
variables, and the ones that include state variables as well. The first kind can be avoided by
using the input extrapolations as parameters to the output functions. The second kind arises when
implicit numerical solvers are used, or when the input approximating functions are interpolations
instead of extrapolations. In the previous example, the first kind can be removed by replacing
ui (t + H) in eq. (34) by the corresponding extrapolation φui (H, ui (n · H), ui ((n − 1) · H), . . .) which
does not depend on ui ((n + 1) · H), thus breaking the algebraic loop. These methods just ignore
the algebraic loop though, and as is shown in Arnold et al. [14], Kübler and Schiehlen [141] (and
empirically in Bastian et al. [25]), neglecting an algebraic loop can lead to a prohibitively high error
in the co-simulation. A better way is to use a fixed point iteration technique. For algebraic loops
involving state variables, the same co-simulation step has to be repeated until convergence. If the
algebraic loop does not involve any state variable, then the iteration is just on the output functions.
To see how algebraic loops involving state variables arise, suppose that, in the example above,
φui is constructed from ui ((n + 1) · H):
ui (n · H + m · hi ) := φui (m · hi , ui ((n + 1) · H), ui (n · H), ui ((n − 1) · H), . . .)
(37)
If an order can be imposed in the evaluation of the simulation units that ensures ui ((n + 1) · H)
can be computed from some λj (t, xj ((n + 1) · H), uj ((n + 1) · H)) that does not indirectly depend on
ui ((n + 1) · H), then this approach —Gauss-Seidel— can improve the accuracy of the co-simulation,
as shown in Arnold [10], Arnold and Günther [11], Arnold et al. [14], Busch [50], Kalmar-Nagy and
Stanciulescu [128]. Obviously, the execution of simulation unit Si has to start after simulation unit
Sj has finished and its output λj (t, xj ((n + 1) · H), uj ((n + 1) · H)) can be evaluated. If the input
uj ((n + 1) depends indirectly on ui ((n + 1) · H), then an algebraic loop exists. The output function
λj (t, xj ((n + 1) · H), uj ((n + 1) · H)) depends on the state of the simulation unit at xj ((n + 1) · H),
which in turn can only be obtained by executing the simulation unit from time n · H to (n + 1) · H,
using the extrapolation of the input uj , φuj (m · hi , uj ((n + 1) · H, . . .)); any improvement in the
input uj ((n+1)·H, means that the whole co-simulation step has to be repeated, to get an improved
xj ((n + 1) · H) and by consequence, an improved output λj (t, xj ((n + 1) · H), uj ((n + 1) · H)).
A fixed point iteration technique that makes use of rollback to repeat the co-simulation step
with corrected inputs is called dynamic iteration, waveform iteration, and strong or onion coupling
[119, 231]. If the simulation units expose their outputs at every internal micro-step, then the
waveform iteration can be improved [155]. Strong coupling approaches are typically the best in terms
of accuracy, but worst in terms of performance. Approaches that do not perform any correction
steps are the best in terms of performance, but worst in accuracy. A variant that attempts to obtain
the middle-ground is the so-called semi-implicit method, where a fixed limited number of correction
steps is performed. See Schweizer and Lu [214, 215, 216], Schweizer et al. [218, 219] for examples
of this approach.
In the current FMI Standard for co-simulation, it is not possible to perform a fixed point iteration
on the output variables only, in the step mode. A work-around this is to rollback the simulation
units and repeat the co-simulation step, effectively treating the algebraic loop as involving the state
variables too.
Until here, we have assumed full knowledge of the models being simulated in each simulation
unit to explain how to identify, and deal with, algebraic loops. In practice, with general black-box
simulation units, extra information is required to identify algebraic loops. According to Arnold
et al. [13], Benveniste et al. [32], Broman et al. [46], a binary flag denoting whether an output
32
depends directly on an input is sufficient. A structural analysis, for example, with Tarjan’s strong
component algorithm [225], can then be performed to identify the loops.
4.3.3
Consistent Initialization of Simulators
The definition of a simulation unit in eq. (21) assumes that an initial condition is part of the
simulation unit. However, as seen in the example of section 4.3.1, the initial states of the simulation
units can be coupled by algebraic constraints, through the output functions, which implies that the
initial states of the simulation units cannot be set independently of the co-simulation in which they
are used. For example, the constraint in eq. (28) has to be satisfied for the initial states:
{x̃1 (0), ṽ1 (0), x̃3 (0), ṽ3 (0)}.
In general, for a co-simulation scenario as defined in eq. (23), there is an extra coupling function
L0 that at the time t = 0, has to be satisfied. For example:
L0 (x1 (0), . . . , xn (0), y1 (0), . . . , yn (0), ycs (0), u1 (0), . . . , un (0), ucs (0)) = 0̄
(38)
where:
• xi (0) denotes the initial state of simulation unit Si ; and
• L0 : X1 × . . . × Xn × Y1 × . . . × Yn × U1 × . . . × Un → Rm represents the initial constraint, not
necessarily equal to L in eq. (23).
eq. (38) may have an infinite number of solutions – as in the case of the example provided in
section 4.3.1 – or have algebraic loops. The initialization problem (or co-initialization) is identified
in Blockwitz et al. [35] and addressed in Galtier et al. [96].
In the FMI Standard, there is a dedicated mode for the (possible fixed point iteration based)
search of a consistent initial state in all simulation units.
4.3.4
Compositional Convergence – Error Control
The accuracy of a co-simulation trace is the degree to which it conforms to the real trace as described
in section 2.2. Obtaining the real trace can be a challenge. Error —the difference between the cosimulation trace and the real trace— is then a measure of accuracy.
In the context of continuous co-simulation, the most accurate trace is the analytical solution
to the coupled model that underlies the scenario. For example, the coupled model in eq. (17),
corresponding to the multi-body system in fig. 5, is implicitly created from the co-simulation scenario
described in eq. (24). Fortunately, the analytical solution can be obtained for this coupled model
because it forms a linear time invariant system. In practice, the analytical solution for a coupled
model cannot be found easily. Calculating the error precisely is impossible for most cases but
getting an estimate in how the error grows is a well understood procedure in numerical analysis.
In simulation, the factors that influence the error are [63]: model, solver, micro-step size, and,
naturally, the size of the time interval to be simulated. In co-simulation, the extrapolation functions introduce error in the inputs of the simulation units, which is translated into error in the
state/outputs of these, causing a feedback on the error that can increase over time. Intuitively, the
larger the co-simulation step size H, the larger is the error made by the extrapolation functions.
33
For example, when the Forward Euler solver (eq. (12)) is used to compute the approximated
behavior trace of the dynamical system in eq. (10), in a single micro step, it is making an error in
the order of
x(t) + f (x(t)) · h + O h2 − (x(t) + f (x(t)) · h) = O h2
|
{z
}
|
{z
}
by Forward Euler
by infinite Taylor series
Obviously, the order in the error made at one step O h2 , most commonly called the local error,
depends on:
• f having no unbounded derivatives – to see why, observe that if the derivative of f is infinite,
then the residual term cannot be bounded by a constant multiplied by h2 . Fortunately, since
most continuous time dynamic systems model some real system, this assumption is satisfied.
• The solver used – other solvers, such as the midpoint method, are derived by truncating
higher order
terms of the Taylor series. For the midpoint method, the local truncation error
is O h3 ;
• Naturally, the larger the micro step size h is, the larger the local error O h2 is.
The local error assumes that the solver only made one step, starting from an accurate point
x(t). To compute the approximate behavior trace, the only accurate point the solver starts from is
the initial value x(0). The rest of the trace is approximate and the error gets compounded over the
multiple steps. For the Forward Euler method, if there is a limit to how f reacts to deviations on
its parameter x̃(t) = x(t) + e(t) from the true parameter x(t), that is, if
kf (x(t)) − f (x(t) + e(t))k ≤ const · e(t)
and const < ∞, then the order of the total accumulation of error can be defined in terms of the
micro-step size. This condition is called global Lipschitz continuity [80]. For the Forward Euler
solver, the total (or global) error is O (h).
For a solver to be useful, it must be convergent, that is, the computed trace must coincide with
the accurate trace when h → 0 [252]. It means the error can be controlled by adjusting the micro
step size h. The same concept of convergence applies to co-simulation but does, as the intuition
suggests, decreasing the communication step size H lead to a more accurate co-simulation trace?
This cannot be answered yet in general co-simulation because the behavior of the coupled model
induced by the coupling of simulation units may not satisfy Lipschitz continuity.
In the context of CT co-simulation, according to Arnold and Günther [11], Arnold et al. [14],
Busch and Schweizer [52], Hafner et al. [114], Kübler and Schiehlen [141], if the simulation units are
convergent and the coupled model induced by the scenario coupling conditions can be written in
the state space form of eq. (10), then the co-simulation unit induced by any of the Jacobi, GaussSeidel, or Strong coupling methods, is convergent, with any polynomial extrapolation technique for
the inputs. Presence of algebraic loops, complex coupling constraints (such as the one shown in
section 4.3.1), are factors that may make it impossible to write the coupled model in state space
form. See Arnold [10] for more examples.
The local error vector, in a co-simulation, is defined as the deviation from the true trace after
34
one co-simulation step H, starting from an accurate point.
x1 (t + H) − x̃1 (t + H)
···
xn (t + H) − x̃n (t + H)
y1 (t + H) − ỹ1 (t + H)
···
yn (t + H) − ỹn (t + H)
(39)
where x̃i (t + H) = δi (t, xi (t), φui (t)), ỹi (t + H) = λi (t, x̃i (t + H), φui (t + H)), and xi (t + H) and
yi (t + H) are the true state vectors and outputs, respectively, for simulation unit Si .
For a convergent co-simulation unit, some of the techniques used traditionally in simulation to
estimate the error, have been applied in co-simulation:
Richardson extrapolation: This well-known technique is compatible with black-box simulation
units as long as these provide rollback and state saving/restore capabilities [13, 15, 96]. The
T
essential idea is to get an estimate of the local error by comparing [x̃i (t + H), ỹi (t + H)]
T
with a less accurate point [x̄i (t + H), ȳi (t + H)] . The less accurate point can be computed
by the same orchestrator but using a larger communication step size. We have seen that
larger communication step sizes affect the accuracy so if the two points are not too far apart,
it means the communication step H does not need to be changed. It is importance to notice
T
that the less accurate point [x̄i (t + H), ȳi (t + H)] has to be computed from the accurate
T
starting point [x̃i (t), ỹi (t)] .
Multi-Order Input Extrapolation: The outputs of two different order input approximation
methods are compared [52, 54].
Milne’s Device: Similar to the previous ones, but the extrapolation of the inputs is compared with
its actual value, at the end of the co-simulation step. Iterative approaches such as the ones
studied in Arnold [10], Arnold and Günther [11], Schweizer and Lu [214, 215, 216], Schweizer
et al. [217, 218, 219] can readily benefit from this technique.
Parallel Embedded Method: This technique runs a traditional adaptive step size numerical
method in parallel with the co-simulation [119]. The purpose is to piggy back in the auxiliary
method, the decisions on the step size. The derivatives being integrated in each simulation
unit have to be either provided, or estimated.
Conservation Laws: The local error is estimated based on the deviation from a known conservation law [207]. Extra domain knowledge about the coupling between simulation units is
required. For example, if the couplings form power bonds [193], then energy should be conserved across a co-simulation step. In practice there is always an error due to the usual factors.
The magnitude of the energy residual at a start and at end of a co-simulation step serves as
an estimate of the local error. This technique has been implemented and studied in Sadjina
et al. [207]. It has the advantage that it may not require rollback functionalities.
Embedded Solver Method: If the individual simulation units support adaptive step size, then
the decisions made internally can be made public to help the orchestrator decide on the
communication step size. To the best of our knowledge, there is no orchestrator proposed that
performs this, but the FMI Standard allows simulation units to reject too large communication
step sizes [35, 46].
After the error is deemed too large by one of the above methods, the correction can be applied
pessimistically (rollback and repeating the same step) or optimistically (adapt the next step). To
35
mitigate the overhead of a pessimistic approach, the corrections may be applied only to sensitive
simulation units, as done in Verhoeven et al. [249].
Finally, the traditional simulation techniques can be applied to chose the right communication
step size H: See Busch and Schweizer [52] for the PI-controller approach, and Gustafsson [112],
Gustafsson et al. [113] for other techniques that can potentially be applied to co-simulation.
4.3.5
Compositional Stability
In the previous section we have presented conditions in which an orchestration engine can reduce
the communication step size to an arbitrarily small value in order to meet arbitrary accuracy.
Theoretically, this is useful as it tells the orchestrator that by reducing the local error, it also reduces
the global error. In practice, the communication step size cannot be reduced to an arbitrarily small
value without facing performance and roundoff error problems. Performance because, for smaller
communication step sizes, it takes more steps to compute a behavior trace over a given interval
of time. Round-off accuracy because in a digital computer, real numbers can only be represented
approximately. Computations involving very small real numbers incur a non-negligible round-off
error. So that means that in practice convergence does not imply that arbitrary accuracy can be
achieved. A better question is to analyze what happens to the global error, as the co-simulation
trace is computed with a non-null communication step size H.
Suppose that the analytical solution to the coupled model induced by the co-simulation scenario
eventually goes to zero. This is the case for the coupled multi-body system of fig. 5, described in
eq. (17), provided that at least one of the constants d1 or d2 is positive non-zero. Intuitively, this
means that the system will lose energy over time, until it eventually comes to rest.
Let x1 (t) denote the analytical solution of the position the mass m1 in the system, and let x˜1 (t)
be the solution computed by a co-simulation unit. Then exi (t) = kx1 (t) − x˜1 (t)k denotes the global
error at time t made by the co-simulation unit. If limt→∞ x1 (t) = 0, then limt→∞ exi (t) = x˜1 (t).
If the co-simulation unit is convergent, then for an arbitrarily small H → 0, limt→∞ exi (t) → 0
will be arbitrarily small too. Since in practice we cannot take arbitrarily small H, we want to
know whether there is some non-zero H such that limt→∞ x˜1 (t) = 0, thus driving exi (t) to zero as
well. If that is the case, then it means that, assuming the system will eventually come to rest, the
co-simulation unit will too. This property is called numerical stability.
Contrarily to convergence, numerical stability is a property that depends on the characteristics
of the system being co-simulated. Numerical stability is always studied assuming that the system
being co-simulated is stable. It makes no sense to show that the co-simulation trace will grow
unbounded provided that the system does too. It is a comparison of two infinities. One of the ways
numerical stability in co-simulation can be studied is by calculating the spectral radius of the error
in the co-simulation unit, written as an autonomous linear discrete system [49, 51].
To give an example, recall that the coupled model induced by the co-simulation scenario de-
36
scribed in eq. (24) can be written as:
0
1
0
x˙1
x1
=
+ 1 u1
d1
c1
−
−m
v˙1
v
1
m
m1
1
| {z1 }
|
{z
}
B1
A1
1 0 x1
y1 =
0 1 v1
| {z }
C1
(40)
0
0
0
1
x2
x˙2
+ cc
=
dc
dc u2
c
−m
− c2m+c
v2
v˙2
m2
m2
2
2
{z
}
|
| {z }
A2
B2
x2
y2 = cc dc
+ −cc −dc u2
| {z } v2
{z
}
|
C2
D2
with the coupling conditions u1 = y2 and u2 = y1 .
In order to write the co-simulation model as an autonomous linear discrete system, we have
to write what happens at a single co-simulation step t ∈ [nH, (n + 1)H] when executed by the
orchestrator presented in Algorithm 3. Since the purpose is to analyze the stability of a co-simulation
unit, and not the stability of each of the simulation units in the co-simulation, it is common to
assume that the simulation units compute the analytical trace of the system. This enables the
study of the stability properties of the co-simulation unit, starting from stable simulation units.
From time t ∈ [nH, (n + 1)H], simulation unit S1 is computing the behavior trace of the following
Initial Value Problem Ordinary Differential Equation (IVP-ODE):
x˙1 (t)
x (t)
= A1 1
+ B1 u1 (nH)
(41)
v˙1 (t)
v1 (t)
T
with initial conditions x1 (nH) v1 (nH) given from the previous co-simulation step. The term
u1 (nH) denotes the fact that we are assuming a constant extrapolation of the input in the interval
t ∈ [nH, (n + 1)H].
x ((n + 1)H)
eq. (41) is linear and time invariant, so the value of 1
can be given analytically
v1 ((n + 1)H)
[49] as:
!
Z (n+1)H
x1 ((n + 1)H)
x
(nH)
1
= eA1 H
+
eA1 ((n+1)H−τ ) dτ B1 u1 (nH)
(42)
v1 ((n + 1)H)
v1 (nH)
nH
or, replacing the integration variable with s = τ − nH,
x1 ((n + 1)H)
A1 H x1 (nH)
=e
+
v1 ((n + 1)H)
v1 (nH)
Z
e
A1 (H−s)
ds B1 u1 (nH)
0
|
37
!
H
{z
K1
}
(43)
P∞ 1 k
where eX = k=0 k!
X is the matrix exponential.
Rewriting eq. (43) as a discrete time system gives us the computation performed by simulation
unit S1 in a single co-simulation step, that is, the state transition function δ1 :
"
#
"
#
(n+1)
(n)
x1
(n)
A 1 H x1
(44)
(n+1) = e
(n) + K1 B1 u1
v1
v1
where z (n) = z(nH).
At the end of the co-simulation step (t = (n + 1)H) the output of the first simulation unit, that
is, its output function λ1 , is given by plugging in eq. (44) to the output y1 in eq. (40):
"
#
(n)
(n+1)
(n)
A1 H x1
y1
= C1 e
(45)
(n) + C1 K1 B1 u1
v1
Repeating the same procedure for the second simulation unit, yields the state transition δ2 and
output functions λ2 :
#
#
"
"
(n)
(n+1)
x2
(n)
A 2 H x2
(n) + K2 B2 u2
(n+1) = e
v2
v2
#
"
(46)
(n)
(n)
(n+1)
A2 H x2
y2
= C2 e
(n) + (C2 K2 B2 + D2 ) u2
v2
RH
with K2 = 0 eA2 (H−u) du.
Since the coupling conditions are u1 = y2 and u2 = y1 , we can combine Equations 46, 45, and
41 into a single discrete time system:
" (n) #
" (n+1) #
x1
x1
(n)
(n+1) A H
v
v1
e 1
0̄
0̄
K1 B1
1(n)
(n+1)
C1 eA1 H
y
0̄
0̄
C1 K1 B1 "y1 #
" 1
#
(47)
A2 H
(n+1) = 0̄
x(n)
K 2 B2
e
0̄
x2
2
(n)
(n+1)
0̄
0̄
C2 K2 B2 + D2 C2 eA2 H
|
v2
{z
} v2
(n)
(n+1)
A
y2
y2
The above system is stable if the behavior traces remain bounded (e.g., by going to zero) as
n → ∞. This can be checked by observing whether the spectral radius ρ(A) < 1. For parameters
m1 = m2 = c1 = c2 = d1 = cc = dc = 1, d2 = 2, a communication step size of H = 0.001,
ρ(A) = 0.9992, which means that the co-simulation unit is stable. If the damping constant were
dk = 6.0E6, then the co-simulation unit would be unstable (ρ(A) ≈ 76.43). A stable co-simulation
is shown in fig. 10.
Different coupling methods, and different approximation functions yield different stability properties. See Busch [49, 50], Busch and Schweizer [51, 53] for the stability analysis of multiple coupling
approaches and approximating functions. Stability of various co-simulation units has been also studied in Arnold [10], Gu and Asada [105], Kalmar-Nagy and Stanciulescu [128], Kübler and Schiehlen
[142], Schweizer et al. [217]. The rules of thumb drawn from these papers can be summarized as:
38
1.0
0.5
Solution
x1
v1
x2
v2
0.0
−0.5
−1.0
0.0
2.5
5.0
Time
7.5
10.0
Figure 10: Behavior trace of co-simulator described in eq. (47). Parameters are: m1 = m2 = c1 =
c2 = d1 = cc = dc = 1, d2 = 2, H = 0.001.
• Co-simulators that employ fixed point iteration techniques typically have better stability
properties;
• Gauss-Seidel coupling approach has slightly better stability properties when the order in
which the simulation units compute is appropriate. For example, the simulation unit with the
highest mass should be computed first [10];
The main problem is that in co-simulation applied to industrial problems, the solvers and models
may be coupled in a black box to protect IP, so there is little knowledge about the kind of solver and
model being used and its stability properties. The best is then to always use iterative techniques
that have been shown to have better stability properties. However, these techniques require rollback
functionalities which can be difficult to support for certain simulation units. Even if those functionalities are available, the cost of computing a co-simulation trace can be prohibitively high when
compared with non-iterative approaches. This creates a paradox where industrial co-simulation
units should make use of iterative techniques but the performance toll may be too high.
4.3.6
Compositional Continuity
Let us for now assume that each CT simulation unit is a mock-up of a continuous system. A
prerequisite is that the physical laws of continuity are obeyed. When using extrapolation in the
inputs (e.g., constant extrapolation), these laws may not be obeyed, as discussed in Busch [50],
Sadjina et al. [207]. Consider the point of view of a simulation unit Si in co-simulation. Throughout
a co-simulation step t ∈ [nH, (n + 1)H] the input φui (t, ui (nH)) = ui (nH) is kept constant. At
the next co-simulation step t ∈ [(n + 1)H, (n + 2)H], the input φui (t, ui ((n + 1)H)) = ui ((n + 1)H)
may change radically if ui ((n + 1)H) is too far away from ui (nH).
The discontinuities in the inputs may wreak havoc in the performance of the simulation unit
Si , causing it to reduce inappropriately the micro step size, to reinitialize the solver [63], to discard
useful information about the past (in multi-step solvers [8, 9]), and/or produce inaccurate values in
its input extrapolation [191]. Furthermore, a discontinuity may be propagated to other simulation
units, aggravating the problem.
Most numerical methods assume that the input is a discretized version of a continuous trace.
39
That means that, when a discontinuity occurs, simulation unit Si cannot distinguish it from a very
steep change in the continuous trace. The way traditional solvers deal with this behavior is to reduce
the micro step size hi until the change is not so steep. This works with a continuous signal with
a steep change, but does not work with a discontinuity: even if the micro-step size hi is reduced,
the difference between limt→((n+1)H)− φui (t, ui (nH)) = ui (nH) and limt→((n+1)H)+ φui (t, ui ((n +
1)H)) = ui ((n + 1)H) is still the same, as it depends on the communication step size H and not
on the micro step size hi . The solver will reduce the micro step size until a minimum is reached, at
which point it gives up and finally advantages the micro step [63].
Most of the times this gives acceptable results but has a huge performance toll: when the solver
is repeatedly retrying a small micro-step size, it does not advance the simulated time. This means
that a huge computational effort goes to waste until the solver finally gives up [61].
We defer the discussion of the correct ways to deal with discontinuities to co-simulation scenario
where discontinuities are welcome, section 5. In continuous co-simulation scenarios, discontinuities
should not occur.
A solution to avoid discontinuities in the input approximations is to use the extrapolated interpolation methods instead of constant extrapolation [49, 50, 75]. These methods ensure at least
that limt→((n+1)H)− φui (t, ui (nH)) = limt→((n+1)H)+ φui (t, ui ((n + 1)H)).
To give an example, we derive one possible linear extrapolated interpolation method for φui
over the interval t ∈ [nH, (n + 1)H]. Since φui is linear, then φui (t, ui (nH), ui ((n − 1)H)) =
b + a(t − nH), for some constants a, b. Let ūi (nH) = φui (nH, ui ((n − 1)H), ui ((n − 2)H)). To
avoid discontinuities, we require that φui (nH, ui (nH), ui ((n − 1)H)) = ūi (nH). And we want that
φui ((n + 1)H, ui (nH), ui ((n − 1)H)) = ui (nH).
So putting these constraints together gives
φui (t, ui (nH), ui ((n − 1)H)) = b + a(t − nH)
ūi (nH) = φui (nH, ui ((n − 1)H), ui ((n − 2)H))
φui (nH, ui (nH), ui ((n − 1)H)) = ūi (nH)
(48)
φui ((n + 1)H, ui (nH), ui ((n − 1)H)) = ui (nH)
Solving this system for φui (t, ui (nH), ui ((n − 1)H)) gives:
φui (t, ui (nH), ui ((n − 1)H)) = ui ((n − 1)H) +
4.3.7
ui (nH) − ui ((n − 1)H)
(t − nH)
H
(49)
Real-time Constraints
As introduced in section 2, the major challenge in real-time simulation is to ensure that a simulation
unit is fast-enough to satisfy the timing constraint t = ατ . In real-time co-simulation, this challenge
gets aggravated due to the presence of multiple simulation units, with different capabilities [222]. In
order to enable real-time co-simulation, every simulation unit has to be fast enough. Furthermore,
real-time co-simulation is often needed because one of the simulation unit is actually the original
system, wrapped as a simulation unit. This means that measurements are performed to the state of
the system, and this means noise in the signals. Therefore, the extrapolation functions used in the
other simulation units have to be properly protected from the noise in the signal, using statistical
techniques such as Kalman filtering [127]. Finally, the quality of the network is important, as the
40
real-time simulation units needs to receive their inputs in a timely manner. To mitigate this, the
orchestration algorithm has to compensate for any delays in the receiving of data, and provide
inputs to the real-time simulation unit [221].
5
Hybrid Co-simulation Approach
Sections 3 and 4 described the essential characteristics and assumptions of simulation units for each
kind of co-simulation approach. When compared to a CT unit, whose state evolves continuously in
time and whose output may have to obey to physical laws of continuity, a DE unit state can assume
multiple values at the same time (transiency) and its output can dramatically change over a short
period of time. For an orchestrator, a CT unit has some flexibility (safe for algebraic loops and
ugly coupling conditions) in computing the co-simulation. In contrast, a DE simulation unit has to
get inputs and produce outputs at the precise time some event occurs. And due to the potentially
drastic change in the outputs, there is no Lipschitz continuous condition that allows predicting how
a delay in the output of the DE unit can affect the overall co-simulation trace.
For example, in the simulation unit of the mass-spring-damper system, eq. (22), with a constant
extrapolation function, and running under the orchestrator in Algorithm 3, the change in the input
can only affect the output after at least H units of time. For continuous time solvers in general, as
can be seen for the explicit solver in eq. (12), a delayed response to the inputs is normal.
These differences between CT and DE units are at the heart of many challenges in hybrid
co-simulation scenarios, mixing the two.
5.1
Hybrid Co-simulation Scenarios
We do not give a formal definition of a hybrid co-simulation scenarios because that is related
to finding an appropriate standard for hybrid co-simulation, which is a non trivial challenge (see
section 5.2.9) [47].
Instead, we define it broadly as mixing the characteristics and assumptions of both kinds of
simulation units. These scenarios, together with an adequate orchestrator, can be used as mockups of hybrid systems [6, 57, 60, 162]. A thermostat regulating the temperature in a room is a
classical example [161]. The continuous constituent system represents the temperature dynamics
of the room, accounting for a source of heat (radiator). The discrete event part is a controller that
turns on/off the radiator depending on the temperature.
The continuous time simulation unit S1 simulates the following dynamics:
ẋ = −α (x − 30q)
x(0) = x0
(50)
where x is the output temperature in the room, α > 0 denotes how fast the room can be heated
(or cooled) down, and q ∈ {0, 1} is the control input that turns on/off the radiator. The discrete
event simulation unit S2 simulates the statemachine shown in fig. 11, where one can think of the
input event tooHot as happening when x(t) ≥ 21 and tooCold when x(t) ≤ 19. The output events
off and on will assign the appropriate value to the input q of S1 . Therefore, the temperature x(t)
is kept within a comfort region.
Clearly, the two units cannot just be coupled together via input to output assignments. Any
orchestrator for this co-simulation scenario has to reconcile the different assumptions about the
inputs and output of each simulation unit.
41
Figure 11: Statemachine model of the controller constituent system.
• The CT simulation unit expects a continuous input, whereas the output of the DE simulation
unit is an event signal.
• The output of the CT simulation unit is a continuous signal, whereas the DE simulation units
expects an event signal as input.
The coupling of continuous time and discrete event black box simulation units has been studied
in the state of the art. In essence, two approaches are known, both based on creating a wrapper
component around a simulation unit to adapt its behavior:
Hybrid DE – wrap every CT unit as a DE simulation unit, and use a DE based orchestration;
Hybrid CT – wrap every DE unit to become a CT unit and use a CT based orchestrator.
According to the formalization that we have proposed for CT and DE simulation units, the
Hybrid DE approach, applied to the thermostat example may involve: wrapping S1 as a DE simulation unit, S10 , with a time advance that matches the size of the co-simulation step; and keeping
track of the output of S1 in order to produce an output event whenever it crosses the thresholds.
Conversely, any output event from S2 has to be converted into a continuous signal for the input
q(t) of S1 .
Other examples of Hybrid DE are described in Awais et al. [19], Bolduc and Vangheluwe [37,
38], Camus et al. [55, 56], Fey et al. [84], Kofman and Junco [137], Kounev et al. [139], Kuhr
et al. [146], Neema et al. [182], Nutaro [189], Quesnel et al. [202], Vangheluwe [244], Widl et al.
[254], Yılmaz et al. [257], Zeigler [260].
The Hybrid CT, in our example, can be followed by wrapping the DE unit S2 as a CT unit that
takes as input the temperature continuous signal, and internally reacts to an event caused by the
crossing of the threshold. Conversely, the output event of S2 can be converted into a continuous
signal q(t).
Examples of the Hybrid CT include Denil et al. [70], Feldman et al. [83], Garro and Falcone
[97], Lawrence et al. [150], Quaglia et al. [201], Tavella et al. [226], Tripakis [232].
Regardless of the approach taken, the properties of the constituent systems have to be retained:
the fact that an otherwise discontinuous signal becomes continuous as a result of a linear or higher
order extrapolation may not respect the properties of the coupled system. Knowledge of the domain
and the simulation units is paramount.
A third alternative, compared to only using Hybrid CT or Hybrid DE, is to have different
mechanisms of orchestrating the simulation units depending on the semantic domain. For instance,
in the actor modeling language Ptolemy II [200], an actor has many similarities to a simulation
unit. Instead of using either Hybrid CT or Hybrid DE, a so called Director block is used for a
particular set of connected actors. In this context, the notion of superdense time is fundamental,
as also discussed in [47].
In the subsection below, different issues that arise in hybrid co-simulation will be described.
These should be read in the light of hierarchical hybrid co-simulation scenarios, where composition42
ality is important.
5.2
5.2.1
Challenges
Semantic Adaptation
While a generic wrapper based on the underlying model of computation of the simulation unit can
be used, as done in [68, 200], the realization of any of the approaches Hybrid DE or Hybrid CT
depends on the concrete co-simulation scenario and the features of the simulation units [42, 178], as
shown with the thermostat example. There is simply no best choice of wrappers for all scenarios.
Even at the technical level, the manner in which the events or signals are sent to (or obtained from)
the unit may need to be adapted [232]. To be concrete, the simulation unit S2 can assume that
all events are communicated by encoding them in a single string signal, as opposed to having a
different signal signal to denote different events. To account for this variability, the most common
adaptations can be captured in a configuration language, as was done in Denil et al. [70], Meyers
et al. [169], or in a specialization of a model of computation, as done in Kuhr et al. [146], Muller
and Widl [176], Pedersen et al. [195]. These approaches require that a person with the domain
knowledge describes how the simulation units can be adapted.
Our choice of wrapper for the Hybrid DE approach is meant to highlight another problem with
the adaptations of simulation units: the wrapper incorporates information that will ultimately have
to be encoded in the software controller. As such, we argue that the need for sophisticated semantic
adaptations should be smaller in later stages of the development of the components so that, for
more refined models of the thermostat, the decision about when to turn off the radiator is not made
by a wrapper of S1 .
5.2.2
Predictive Step Sizes
In the Hybrid DE approach, the time advance has to be defined (recall eq. (1)). Setting it to
whatever co-simulation step size H the orchestrator decides will work, but the adapted simulation
unit may produce many absent output events. Better adaptations have been proposed. In the
thermostat example, S10 can propose a time advance that coincides with the moment that x(t) will
leave the comfort region, thereby always being simulated at the relevant times.
Naturally, these approaches rely in information that may expose the IP of simulation units.
Others try to adaptively guess the right time advance by monitoring other conditions of interest,
set over the own dynamics of the adapted simulation unit, the most common approach being the
quantization of the output space [38, 136, 137, 190, 261].
The capability to predict the time advance is also useful to enhance the performance/accuracy
of CT based co-simulation, as shown in Broman et al. [46].
5.2.3
Event Location
Locating the exact time at which a continuous signal crosses a threshold is a well known problem
[39, 41, 263] and intimately related to guessing the right time advance for predicting the step size
[56, 96]. To address this, solutions typically require derivative information of the signal that causes
the event, and/or the capability to perform rollbacks. In the thermostat example, a co-simulation
that shows the output q of the controller changing from 0 to 1 at time te while the temperature of
43
the room x actually crossed the confort zone at te − k, for k > 0, may not be accurate if k is too
large. Note that k is a consequence of the decisions made in the orchestrator.
5.2.4
Discontinuity Identification
Until here, we have based our discussion in the knowledge of what kind of simulation units comprise
a co-simulation. In a general hierarchical co-simulation, a simulation unit’s output may be an event
signal coming from a wrapper of a CT unit, or vice-versa. In any case, at runtime, a signal is often
represented as a set of time-stamped points. Observing this sequence of points alone does not make
it possible to discern a steep change in a continuous signal, from a true discontinuity, that occurs
in an event signal [47, 154, 172, 263]. Extra information is currently used: a) a formalization of
time which include the notion of absent signal, as proposed in Broman et al. [47], Lee and Zheng
[154], Tavella et al. [226]; or b) an extra signal can be used to discern when a discontinuity occurs,
as done in the FMI for Model Exchange [35], even facilitating the location of the exact time of the
discontinuity; or c) symbolic information (e.g., Dirac impulses [72]) that characterize a discontinuity
can be included, as done in Nilsson [186]
5.2.5
Discontinuity Handling
Once a discontinuity is located, how it is handled depends on the nature of the simulation units and
their capabilities. If the simulation unit is a mock-up of a continuous system then, traditionally,
discontinuities in the inputs should be handled by reinitializing the simulation unit [63]. This step
can incur a too high performance cost, especially with multi-step numerical methods, and Andersson
[8], Andersson et al. [9] proposes an improvement for these solvers. Furthermore, a discontinuity
can cause other discontinuities, producing a cascade of re-initializations. During this process, which
may not finish, care must be taken to ensure that physically meaningful properties such as energy
distribution, are respected [174].
5.2.6
Algebraic Loops, Legitimacy, and Zeno Behavior
Algebraic loops are non-causal dependencies between simulation units that can be detected using
feedthrough information, as explained in section 4.3.2. In CT based co-simulation, the solution to
algebraic loops can be attained by a fixed point iteration technique, as covered in section 4.3.2.
There is the possibility that the solution to an algebraic loop will fail to converge. The result is
that, if left unchecked, the orchestrator would move an infinite number of input and output values
between simulation units, at the same point in time.
In DE based co-simulation a related property is legitimacy [262], which is the undesirable version
of the transiency property, explained in section 3. A illegitimate co-simulation scenario will cause the
co-simulation orchestrator to move an infinite number of events with the same timestamp between
units, never advancing time. Distance matrices, used to optimize parallel optimistic approaches, as
explained in Fujimoto [94] and used in Ghosh et al. [101], can be leveraged to detect statically the
presence of some classes of illegitimacy.
A similar behavior, but more difficult to detect is Zeno behavior. It occurs when there is an
increasingly small interval of time between two consecutive events, up to the point that the the sum
of all these intervals is finite [238]. However, while illegitimate behaviors are not desired in pure
DE co-simulation, at least in the theoretical sense, Zenoness is a desired feature in some hybrid
co-simulation scenarios. We say in the theoretical sense because, for the purposes of co-simulation,
44
scenarios with Zenoness still have to be recognized and appropriate measures, such as regularization
[125], have to be taken.
5.2.7
Stability
If a hybrid co-simulation represents a hybrid or switched system [126, 238], then it is possible that
a particular sequence of events cases the the system to become unstable, even if all the individual
continuous modes of operation are stable. New analyses are required to identify whether the CT
units can yield unstable trajectories as a result of the events of wrapped DE simulation units, while
keeping the IP hidden.
5.2.8
Theory of DE Approximated States
In a pure DE based co-simulation, if round-off errors are neglected, the computed trajectories are
essentially exact. To the best of our knowledge, only Zeigler et al. [262] addresses theoretically how
the error in a discrete event system can be propagated. In CT based co-simulation however, error
is an accepted and well studied and techniques exist to control it.
In Hybrid co-simulation, there is a need for analysis techniques that provide bounds on the error
propagation in the DE simulation units, when these are coupled to sources of error.
In addition, based on these analyzes, it should be possible for a DE simulation unit to recognize
that its error has exceeded a given tolerance, and measures should be taken to reduce that error.
Having these techniques in place allows a hybrid co-simulation orchestrator to take appropriate
measures (e.g., adapt the communication step size, etc. . . ) the keep the error bounded in every
simulation unit.
5.2.9
Standards for Hybrid Co-simulation
While for CT co-simulation there is the Functional Mock-up Interface (FMI) standard [35], and for
DE co-simulation there is the High Level Architecture (HLA) [2] standard, as of the time of writing,
both standards have limitations for hybrid co-simulation. Bogomolov et al. [36], Garro and Falcone
[97], Tavella et al. [226] use/propose extensions to the FMI standard and Awais et al. [18] proposes
techniques to perform CT simulation conforming to HLA. Recognizing that hybrid co-simulation
is far from well studied, Broman et al. [47] proposes a set of idealized test cases that any hybrid
co-simulation unit, and underlying standard, should pass. In particular, it is important to have
correct handling and representation of time, to achieve a sound approach for simultaneity.
Finally, even with a standardized interface, simulation units are not all equal: a fact that makes
coding an orchestration algorithm a real challenge. A possible approach to deal with this heterogeneity, proposed in Gomes [102], is to assume that all units implement the same set of features,
code the orchestration algorithm for those features, and delegate to wrappers the responsibility
of leveraging extra features (or mitigating the lack of). In the section below, these features are
classified.
6
Classification
Having described the multiple facets of co-simulation, this section summarizes our classification and
methodology.
45
Non-Functional
Requirements
Platform
Independence
Configuration
Reusability
Scalability
Performance
Co-Simulation
Extensibility
IP Protection
Accuracy
Hierarchy
Non-Functional
Requirements
Simulator
Requirements
Framework
Requirements
Fault
Tolerance
Optional
Open-source
Legend
Feature
Mandatory
Figure 12: Top-level.
6.1
Distribution
Feature
Optional
Parallelism
Figure 13: Non-Functional Requirements.
Methodology
To find an initial set of papers related to co-simulation, we used Google Scholar with the keywords
“co-simulation”, “cosimulation”, “coupled simulation”, and collected the first 10 pages of papers.
Every paper was then filtered by the abstract, read in detail, and its references collected. To guide
our reading to the most influential papers, we gave higher priority to most cited (from the papers
that we have collected).
We read approximately 30 papers to create the initial version of the taxonomy. Then, as we
read new papers, we constantly revised the taxonomy and classified them.
After a while, new references did not cause revisions to the taxonomy, which prompted us to
classify the collected papers in a more systematic fashion: all the papers that we collected from
2011 (inclusive) up to, and including, 2016 were classified. Two main reasons justify the last 5
years interval: limited time; and most of the papers refer to, and are based on, prior work. As
a consequence, the classification would be very similar for many of the related references prior to
2011.
From the papers classified, those that report case studies where noted to create fig. 1.
6.2
Taxonomy
The taxonomy is represented as a feature model [129] structured in three main categories, shown
in fig. 12:
Non-Functional Requirements (NFRs): Groups concerns (e.g., performance, accuracy, and IP
Protection) that the reference addresses.
Simulation unit Requirements (SRs): Features required/assumed from the simulation units
by the orchestrator described in the paper. Examples: Information exposed, causality, local/remote availability, or rollback support.
Framework Requirements (FRs): Features provided by the orchestrator. Examples: dynamic
structure, adaptive communication step size, or strong coupling support.
Each main group is detailed in Figures 13, 14, and 15. Abstract features denote concepts that
can be easily detailed down but we chose not to, for the sake of brevity. Mandatory features are
required for the activity of co-simulation while optional are not.
46
6.3
State of the Art
To give an example on how the taxonomy is used, consider the work in Van Acker et al. [236], where
an FMI based multi-rate orchestration algorithm is generated from a description of the co-simulation
scenario. In the paper, the description language introduced can be reused in a tool-agnostic manner.
The orchestration code generator analyzes the co-simulation scenario, and: a) identifies algebraic
loops using I/O feedthrough information; b) separates the fast moving simulation units from the
slow moving ones, using the preferred step size information, and provides interpolation to the
fast ones (multi-rate); and c) finds the largest communication step size that divides all step sizes
suggested by simulation units and uses it throughout the whole co-simulation. We argue that the
generated orchestrator is fast because all the decisions are made at code generation stage. The
algebraic loops are solved via successive substitution of inputs, storing and restoring the state of
the simulation units.
Based on these facts, Van Acker et al. [236] can be classified as follows:
Non-Functional requirements: Performance, IP protection, and Configuration reusability;
Simulation unit requirements: Causal simulation units, Locally available simulation units, Preferred step size information about the solver, Feedthrough I/O causality information, State
values, No time constraints, and No rollback support;
Framework requirements: Continuous time domain, Multi-rate simulation, Fixed communication step size, Input/Output coupling, Fully implicit strong coupling support, Postmortem
visualization of results, Postmortem visualization of results, FMI as the underlying standard,
Gauss-seidel communication approach, and Support for three or more simulation units.
With similar reasoning, the FMI standard for co-simulation, version 2.0, can be classified according to the assumptions it makes about the participating simulation units. These are highlighted
in fig. 14.
The remaining state of the art is classified in Figures 16 – 19. The raw data is available online13 .
6.4
Discussion
Observing fig. 16, Accuracy is the most observed NFR, with 31 reports, followed by IP protection
and Performance. The least observed NFRs are Fault tolerance, Hierarchy and Extensibility.
Fault tolerance is especially important for long running co-simulations. One of the industrial
partners of the INTO-CPS project has running co-simulations that takes a minimum of two weeks
to complete.
We argue that Extensibility (the ability to easily accomodate new features) should be given
more importance: if an heterogeneous set of simulation units participate in the same co-simulation
scenario, the combination of capabilities provided (see fig. 14) can be huge. Thus, the orchestrator can either assume a common homogeneous set of capabilities, which is the most common
approach, or can leverage the capabilities provided by each one. The later approach can lead to
an extremely complex orchestration algorithm. In any case, extensibility is key to address new
semantic adaptations (recall section 5.2.1).
As fig. 18 suggests, we could not find approaches that make use of the nominal values of state
and output variables, even though these are capabilities supported in the FMI Standard (see fig. 14),
and are useful to detect co-simulations that are not valid. A-causal approaches are important for
modularity, as explained in section 4.3.1, but these are scarce too.
13 http://msdl.cs.mcgill.ca/people/claudio/pub/Gomes2016bClassificationRawData/raw_data.zip
47
As for the framework requirements, in fig. 19, the least observed features are dynamic structure
co-simulation, interactive visualization, multi-rate, algebraic coupling, and partial/full strong coupling support. This can be explained by the fact that these features depend upon the capabilities
of the simulation units, which may not be mature.
Figures 16 – 19 do not tell the full story because they isolate each feature. Feature interaction is
a common phenomenon, and among many possible interactions, we highlight the accuracy concern,
domain of the co-simulation, number of simulation units supported, and IP protection. As can be
seen from fig. 21, there is only one approach [146] that is both CT and DE based, up to any number
of simulation units. Note that this does not mean that the work addresses all challenges that were
identified in section 5. Accommodating the different CT and DE domains means that the approach
assumes that the simulation units can behave both as a CT and as a DE unit.
The concern with IP protection is evident in fig. 16 but the number of DE and CT based
approaches that provide some support for it is small, as shown in fig. 20. Similarly, as fig. 22 suggests,
accuracy does not show up a lot in the DE and CT approaches, for more than two simulation units.
Accuracy is particularly important in interactions between DE and CT simulation units.
In general, from the observed classification, there is a lack of research into approaches that are
both DE and CT based, and that leverage the extra the features from the simulation units.
7
Concluding Remarks
As this work shows, there are many interesting challenges to be explored in co-simulation, which
will play a key role in enabling the virtual development of complex heterogeneous systems in the
decades to come. The early success can be attributed to a large number of reported applications.
However, from the application references covered (see fig. 1), the large majority represent ad-hoc
couplings between two simulators of two different domains (e.g., a network simulator with a power
grid one, or a HVAC simulator with a building envelop one). This, however, excludes the (potentially
many) unreported applications of co-simulation. As systems become complex, the demand for cosimulation scenarios that are large, hierarchical, heterogeneous, accurate, IP protected, and so on,
will increase.
This survey presents an attempt at covering the main challenges in co-simulation. To tackle
such a broad topic, we have covered the two main domains —continuous time and discrete event
based co-simulation— separately and then surveyed the challenges that arise when the two domains
are combined. A taxonomy is proposed and a classification of the works related to co-simulation in
the last five years is carried out using that taxonomy.
From the challenges, we highlight semantic adaptation, modular coupling, stability and accuracy,
and finding a standard for hybrid co-simulation as being particular important.
For early system analysis, the adaptations required to combine simulators from different formalisms, even conforming to the same standard, are very difficult to generalize to any co-simulation
scenario. A possible work around this is to allow the system integrator to describe these, as proposed
in [70].
One of the main conclusions of the classification is that there is lack of research into modular,
stable and accurate coupling of simulators in dynamic structure scenarios. This is where a-causal
approaches for co-simulation can play a key role. The use of bi-directional effort/flow ports can be
a solution inspired by Bond-graphs [193], and there is some work in [207] already in this direction.
Finally, this document is an attempt to summarize, bridge, and enhance the future research in
co-simulation, wherever it may lead us to.
48
Acknowledgment
The authors wish to thank Yentl Van Tendeloo, for the in depth review of DE based co-simulation,
Kenneth Guldbrandt Lausdahl for providing valuable input and discussions throughout the making
of this survey, and TWT GmbH for the valuable input on everyday challenges faced by a cosimulation master. This research was partially supported by Flanders Make vzw, the strategic
research centre for the manufacturing industry, and a PhD fellowship grant from the Agency for
Innovation by Science and Technology in Flanders (IWT). In addition, the work presented here is
partially supported by the INTO-CPS project funded by the European Commission’s Horizon 2020
programme under grant agreement number 664047. This project is financially supported by the
Swedish Foundation for Strategic Research.
49
Simulator
Requirements
Causality
Causal
Rollback
Support
A-causal
None
Availability
Local
None
Scaled RT
Static
Model
Dynamic
Solver
Micro-step
Outputs
Deadreckoning
Model
Detailed
Model
Nominal Values
Discontinuity
Indicator
Preferred
Outputs
State
Order of
Step-size Accuracy
State
Frequency
Multiple
Remote
Information
Exposed
Outputs
Single
Time
Constraints
Next
Values
WCET
I/O Signal Kind
Serialization
Input
Extrapolation
I/O
Causality
State
Derivative
Time
Outputs
Jacobian
State
Outputs
Propagation Feedthrough
Delay
Dependency
Kind
State
Non-Linear
Linear
Abstract Feature
Feature
FMI CS 2.0
Mandatory
Exclusive Or
Optional
Or
Figure 14: Simulation Unit Requirements and features provided in the FMI Standard for cosimulation, version 2.0.
50
Framework
Requirements
Dynamic
Structure
No.
Simulators
Domain
CT
2
DE
Comm.
Approach
Co-simulation
Rate
Jacobi
Feature
Single
3+
Multiple
Comm.
Step Size
Fixed
Mandatory
Exclusive Or
Optional
Or
Standard
HLA
FDMU
FMI
Results
Visualization
Adaptive
Coupling
I/O
Assignment
Gauss-seidel
Live
Algebraic
Constraints
Postmortem
Interactive
Strong Coupling
Support
None
Semi-Implicit
Fully Implicit
Category
Figure 15: Framework Requirements.
Platform Independence
Performance
IP Protection
Hierarchy
Fault Tolerance
Open−source
Parallelism
Distribution
Accuracy
Extensibility
Config. Reusability
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
● ●
● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ● ●
●
●
●
2000
2005
Year
●
●
● ●
● ●
2010
●
●
●
●
●
●
●
●
●
●
●
●
Total
11
25
● 25
Reports
3
8
1
6
6
4
● 21
2
● 24
● 31
4
7
●
●
●
2015
Figure 16: Classification with respect to non-functional requirements.
51
A−Causal
Causal
Category
Locally Available
●
Remotely Available
No Time Constraints
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Fixed Real−Time Constraints
●
●●● ● ●●
●● ● ●●●
● ●
●
●
●
● ●● ● ●●
●
●
●
Dynamic Real−Time Constraints
●
●
●
No Rollback
Single Rollback
●
●
●
●●● ● ●●
●
●
●
Reports
12
9
6
3
●
●
●
● ●
●
●
Multi Rollback
2000
2005
2010
Year
2015
Category
Figure 17: Classification with respect to simulation unit requirements: execution capabilities.
State Nominal Values
Output Nominal Values
Input Extrapolation
Model
Kind of Signal
Frequency of Outputs
Preferred Step Size
Next Step Size
I/O Feedthrough
Worst Case Exec. Time
State Values
Serialized State
Micro−step Outputs
State Jacobian
Output Jacobian
State Derivatives
Output Derivatives
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
● ●
●
Reports
4
3
2
1
●
●
●
2000
●
2005
Year
● ●
2010
2015
Figure 18: Classification with respect to simulation unit requirements: information exposed.
52
Category
Dynamic
DE Domain
CT Domain
Single Rate
Multi−Rate
Fixed Comm. Step
Adaptive Comm. Step
I/O Coupling
Alg. Constraints Coupling
No Strong Coupling
Partial Strong Coupling
Full Strong Coupling
Post Mortem Visualization
Live Visualization
Interactive Visualization
FMI Based
HLA Based
FDMU Based
Jacobi Communication
Gauss−Seidel Communication
Two Simulators
More than Three Simulators
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●●●●●
● ● ● ● ● ●
●●● ● ●●
●
● ●
●
●
● ● ● ● ● ● Reports
12
●
9
●
●
6
●●● ●●●
3
●
●
●
●
●●● ●●●
●●● ● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
2000
●
●
●
●
●
2005
Year
● ● ● ● ●●
●●● ●
●●● ●
● ●
●
●●
● ● ●
2010
●
●
2015
CT
●
●
●
DE DE+CT
Domain
Reports
40
30
20
10
3+
2
●
●
CT
●
●
●
●
DE DE+CT
Domain
Reports
30
20
10
No. Simulators
●
●
3+
●
●
●
2
●
●
●
3+
●
●
●
●
●
2
●
CT
No Accuracy
No
●
No. Simulators
Yes
Accuracy
IP Protection
Figure 19: Classification with respect to framework requirements.
Reports
20
15
10
5
DE DE+CT
Domain
Figure 20: Formalisms vs IP Figure 21: Formalisms vs Sim- Figure 22: Accuracy vs ForProtection.
ulation units.
malisms vs Simulation units.
53
References
[1] Modelica - A Unified Object-Oriented Language for Physical Systems Modeling, 2007. URL
https://www.modelica.org/documents/ModelicaSpec30.pdf.
[2] IEEE Standard for Modeling and Simulation (M&S) High Level Architecture (HLA) - Federate Interface Specification, 2010. URL https://standards.ieee.org/findstds/standard/
1516-2010.html.
[3] An-jelo Gian C. Abad, Lady Mari Faeda G. Guerrero, Jasper Kendall M. Ignacio, Dianne C Magtibay, Mark Angelo C Purio, and Evelyn Q Raguindin. A simulation of a
power surge monitoring and suppression system using LabVIEW and multisim co-simulation
tool. In 2015 International Conference on Humanoid, Nanotechnology, Information Technology,Communication and Control, Environment and Management (HNICEM), pages 1–3.
IEEE, dec 2015. ISBN 978-1-5090-0360-0. doi: 10.1109/HNICEM.2015.7393204.
[4] Andreas Abel, Torsten Blochwitz, Alexander Eichberger, Peter Hamann, and Udo Rein. Functional mock-up interface in mechatronic gearshift simulation for commercial vehicles. In 9th
International Modelica Conference. Munich, 2012.
[5] Ahmad T. Al-Hammouri. A comprehensive co-simulation platform for cyber-physical systems.
Computer Communications, 36(1):8–19, dec 2012. doi: 10.1016/j.comcom.2012.01.003.
[6] R. Alur, C. Courcoubetis, N. Halbwachs, T. A. Henzinger, P. H. Ho, X. Nicollin, A. Olivero,
J. Sifakis, and S. Yovine. The algorithmic analysis of hybrid systems. Theoretical Computer
Science, 138(1):3–34, feb 1995. ISSN 03043975. doi: 10.1016/0304-3975(94)00202-T.
[7] Andrés A. Alvarez Cabrera, Krijn Woestenenk, and Tetsuo Tomiyama. An architecture model
to support cooperative design for mechatronic products: A control design case. Mechatronics,
21(3):534–547, apr 2011. ISSN 09574158. doi: 10.1016/j.mechatronics.2011.01.009.
[8] Christian Andersson. Methods and Tools for Co-Simulation of Dynamic Systems with the
Functional Mock-up Interface. PhD thesis, Lund University, 2016.
[9] Christian Andersson, Claus Führer, and Johan Åkesson.
Efficient Predictor for
Co-Simulation with Multistep Sub-System Solvers.
Technical Report in Mathematical Sciences, 2016(1), 2016. ISSN 1403-9338. URL http://lup.lub.lu.se/record/
dbaf9c49-b118-4ff9-af2e-e1e3102e5c22.
[10] Martin Arnold. Stability of Sequential Modular Time Integration Methods for Coupled Multibody System Models. Journal of Computational and Nonlinear Dynamics, 5(3):031003, may
2010. ISSN 15551423. doi: 10.1115/1.4001389.
[11] Martin Arnold and Michael Günther. Preconditioned Dynamic Iteration for Coupled
Differential-Algebraic Systems. BIT Numerical Mathematics, 41(1):1–25, 2001. doi: 10.1023/
A:1021909032551.
[12] Martin Arnold, Antonio Carrarini, Andreas Heckmann, and Gerhard Hippmann. Simulation
techniques for multidisciplinary problems in vehicle system dynamics. In Vehicle System
Dynamics Supplement 40, volume 40, pages 17–36, Vienna, Austria, 2003.
54
[13] Martin Arnold, Christoph Clauss, and Tom Schierz. Error Analysis and Error Estimates for
Co-Simulation in FMI for Model Exchange and Co-Simulation V2.0. Archive of Mechanical
Engineering, LX(1):75, jan 2013. ISSN 0004-0738. doi: 10.2478/meceng-2013-0005.
[14] Martin Arnold, Christoph Clauß, and Tom Schierz. Error Analysis and Error Estimates for
Co-simulation in FMI for Model Exchange and Co-Simulation v2.0. In Sebastian Schöps,
Andreas Bartel, Michael Günther, W E Jan ter Maten, and C Peter Müller, editors, Progress
in Differential-Algebraic Equations, pages 107–125. Springer Berlin Heidelberg, Berlin, Heidelberg, 2014. ISBN 978-3-662-44926-4. doi: 10.1007/978-3-662-44926-4 6.
[15] Martin Arnold, Stefan Hante, and Markus A Köbis. Error analysis for co-simulation with
force-displacement coupling. PAMM, 14(1):43–44, dec 2014. ISSN 16177061. doi: 10.1002/
pamm.201410014.
[16] Memduha Aslan, Umut Durak, and Koray Taylan. MOKA: An Object-Oriented Framework
for FMI Co-Simulation. 2015.
[17] Karl Johan Aström and Richard M Murray. Feedback systems: an introduction for scientists
and engineers. Princeton university press, 2010. ISBN 1400828732.
[18] Muhammad Usman Awais, Wolfgang Mueller, Atiyah Elsheikh, Peter Palensky, and Edmund Widl. Using the HLA for Distributed Continuous Simulations. In 2013 8th EUROSIM
Congress on Modelling and Simulation, pages 544–549, Washington, DC, USA, sep 2013.
IEEE. ISBN 978-0-7695-5073-2. doi: 10.1109/EUROSIM.2013.96.
[19] Muhammad Usman Awais, Peter Palensky, Atiyah Elsheikh, Edmund Widl, and Stifter
Matthias. The high level architecture RTI as a master to the functional mock-up interface
components. In 2013 International Conference on Computing, Networking and Communications (ICNC), pages 315–320, San Diego, USA, jan 2013. IEEE. ISBN 978-1-4673-5288-8.
doi: 10.1109/ICCNC.2013.6504102.
[20] Muhammad Usman Awais, Peter Palensky, Wolfgang Mueller, Edmund Widl, and Atiyah
Elsheikh. Distributed hybrid simulation using the HLA and the Functional Mock-up Interface.
In IECON 2013 - 39th Annual Conference of the IEEE Industrial Electronics Society, pages
7564–7569, Vienna, Austria, nov 2013. IEEE. ISBN 978-1-4799-0224-8. doi: 10.1109/IECON.
2013.6700393.
[21] Fernando J. Barros. Modeling formalisms for dynamic structure systems. ACM Transactions
on Modeling and Computer Simulation, 7(4):501–515, oct 1997. ISSN 10493301. doi: 10.
1145/268403.268423.
[22] Fernando J Barros. Dynamic structure multiparadigm modeling and simulation. ACM Transactions on Modeling and Computer Simulation, 13(3):259–275, jul 2003. ISSN 10493301. doi:
10.1145/937332.937335.
[23] Fernando J. Barros. Semantics of dynamic structure event-based systems. In Proceedings of
the second international conference on Distributed event-based systems - DEBS ’08, DEBS ’08,
page 245, New York, USA, 2008. ACM Press. ISBN 9781605580906. doi: 10.1145/1385989.
1386020.
55
[24] Paul I. Barton and C. C. Pantelides. Modeling of combined discrete/continuous processes.
AIChE Journal, 40(6):966–979, jun 1994. ISSN 0001-1541. doi: 10.1002/aic.690400608.
[25] Jens Bastian, Christoph Clauß, Susann Wolf, and Peter Schneider. Master for Co-Simulation
Using FMI. In 8th International Modelica Conference, pages 115–120, Dresden, Germany, jun
2011. Fraunhofer Institute for Integrated Circuits IIS. doi: 10.3384/ecp11063115.
[26] Lionel Belmon, Yujung Geng, and Huaqiang He. Virtual Integration for hybrid powertrain
development, using FMI and Modelica models. 10th International Modelica Conference, 2014.
[27] Giovanni Beltrame, Donatella Sciuto, and Cristina Silvano. Multi-Accuracy Power and
Performance Transaction-Level Modeling. IEEE Transactions on Computer-Aided Design
of Integrated Circuits and Systems, 26(10):1830–1842, oct 2007. ISSN 0278-0070. doi:
10.1109/TCAD.2007.895790.
[28] Abir Ben Khaled, Mongi Ben Gaid, Nicolas Pernet, and Daniel Simon. Fast multi-core cosimulation of Cyber-Physical Systems: Application to internal combustion engines. Simulation
Modelling Practice and Theory, 47:79–91, sep 2014. ISSN 1569190X. doi: 10.1016/j.simpat.
2014.05.002.
[29] M. Benedikt, D. Watzenig, J. Zehetner, and A. Hofer. Macro-step-size selection and monitoring of the coupoling error for weak coupled subsystems in the frequency-domain. V
International Conference on Computational Methods for Coupled Problems in Science and
Engineering, pages 1–12, 2013.
[30] Martin Benedikt and Anton Hofer. Guidelines for the Application of a Coupling Method for
Non-iterative Co-simulation. In 2013 8th EUROSIM Congress on Modelling and Simulation,
pages 244–249. IEEE, sep 2013. ISBN 978-0-7695-5073-2. doi: 10.1109/EUROSIM.2013.52.
[31] Martin Benedikt and Franz Rudolf Holzinger. Automated configuration for non-iterative
co-simulation. In 17th International Conference on Thermal, Mechanical and Multi-Physics
Simulation and Experiments in Microelectronics and Microsystems (EuroSimE), pages 1–7,
Montpellier, apr 2016. IEEE. ISBN 978-1-5090-2106-2. doi: 10.1109/EuroSimE.2016.7463355.
[32] Albert Benveniste, Benoı̂t Caillaud, and Paul Le Guernic. Compositionality in Dataflow
Synchronous Languages: Specification and Distributed Code Generation. Information and
Computation, 163(1):125–171, nov 2000. ISSN 08905401. doi: 10.1006/inco.2000.9999.
[33] D. Bian, M. Kuzlu, M. Pipattanasomporn, S. Rahman, and Y. Wu. Real-time co-simulation
platform using OPAL-RT and OPNET for analyzing smart grid performance. In 2015 IEEE
Power & Energy Society General Meeting, pages 1–5. IEEE, jul 2015. ISBN 978-1-4673-8040-9.
doi: 10.1109/PESGM.2015.7286238.
[34] Torsten Blochwitz, Martin Otter, Martin Arnold, C. Bausch, Christoph Clauss, Hilding
Elmqvist, Andreas Junghanns, Jakob Mauss, M. Monteiro, T. Neidhold, Dietmar Neumerkel,
Hans Olsson, J.-V. Peetz, and S. Wolf. The Functional Mockup Interface for Tool independent
Exchange of Simulation Models. In 8th International Modelica Conference, pages 105–114,
Dresden, Germany, jun 2011. Linköping University Electronic Press; Linköpings universitet.
doi: 10.3384/ecp11063105.
56
[35] Torsten Blockwitz, Martin Otter, Johan Akesson, Martin Arnold, Christoph Clauss, Hilding
Elmqvist, Markus Friedrich, Andreas Junghanns, Jakob Mauss, Dietmar Neumerkel, Hans
Olsson, and Antoine Viel. Functional Mockup Interface 2.0: The Standard for Tool independent Exchange of Simulation Models. In 9th International MODELICA Conference, pages
173–184, Munich, Germany, nov 2012. Linköping University Electronic Press; Linköpings
universitet. doi: 10.3384/ecp12076173.
[36] Sergiy Bogomolov, Marius Greitschus, Peter G. Jensen, Kim G. Larsen, Marius Mikucionis,
Thomas Strump, and Stavros Tripakis. Co-Simulation of Hybrid Systems with SpaceEx and
Uppaal. In 11th International Modelica Conference (MODELICA), pages 159–169, Paris,
France, sep 2015. Linköping University Electronic Press. doi: 10.3384/ecp15118159.
[37] Jean-Sébastien Bolduc and Hans Vangheluwe. Expressing ODE models as DEVS: Quantization approaches. In Proceedings of the AIS’2002 Conference (AI, Simulation and Planning in
High Autonomy Systems), April 2002, Lisboa, Portugal/F. Barros and N. Giambiasi (eds.),
pages 163–169, 2002.
[38] Jean-Sébastien Bolduc and Hans Vangheluwe. Mapping odes to devs: Adaptive quantization. In Summer Computer Simulation Conference, pages 401–407. Society for Computer
Simulation International; 1998, 2003. ISBN 0094-7474.
[39] Massimo Bombino and Patrizia Scandurra. A model-driven co-simulation environment for
heterogeneous systems. International Journal on Software Tools for Technology Transfer, 15
(4):363–374, aug 2013. ISSN 1433-2779. doi: 10.1007/s10009-012-0230-5.
[40] Spencer Borland. Transforming statechart models to DEVS. PhD thesis, 2003.
[41] F. Bouchhima, M. Briere, G Nicolescu, M Abid, and E. Aboulhamid. A SystemC/Simulink
Co-Simulation Framework for Continuous/Discrete-Events Simulation. In 2006 IEEE International Behavioral Modeling and Simulation Workshop, pages 1–6. IEEE, sep 2006. ISBN
0-7803-9742-8. doi: 10.1109/BMAS.2006.283461.
[42] Frédéric Boulanger, Cécile Hardebolle, Christophe Jacquet, and Dominique Marcadet. Semantic Adaptation for Models of Computation. In Application of Concurrency to System Design
(ACSD), 2011 11th International Conference on, pages 153–162, 2011. ISBN 1550-4808 VO
-. doi: 10.1109/ACSD.2011.17.
[43] Jonathan Brembeck, Andreas Pfeiffer, Michael Fleps-Dezasse, Martin Otter, Karl Wernersson,
and Hilding Elmqvist. Nonlinear State Estimation with an Extended FMI 2.0 Co-Simulation
Interface. In Proceedings of the 10th International Modelica Conference. Lund, Sweden, pages
53–62, 2014.
[44] T. Brezina, Z. Hadas, and J. Vetiska. Using of Co-simulation ADAMS-SIMULINK for development of mechatronic systems. In 14th International Conference Mechatronika, pages 59–64.
IEEE, jun 2011. ISBN 978-80-8075-477-8. doi: 10.1109/MECHATRON.2011.5961080.
[45] David Broman, Edward A. Lee, Stavros Tripakis, and Martin Törngren. Viewpoints, Formalisms, Languages, and Tools for Cyber-Physical Systems. In Proceedings of the 6th International Workshop on Multi-Paradigm Modeling, pages 49–54. ACM, 2012.
57
[46] David Broman, Christopher Brooks, Lev Greenberg, Edward A. Lee, Michael Masin, Stavros
Tripakis, and Michael Wetter. Determinate composition of FMUs for co-simulation. In
Eleventh ACM International Conference on Embedded Software, Montreal, Quebec, Canada,
2013. IEEE Press Piscataway, NJ, USA. ISBN 978-1-4799-1443-2.
[47] David Broman, Lev Greenberg, Edward A Lee, Michael Masin, Stavros Tripakis, and Michael
Wetter. Requirements for Hybrid Cosimulation Standards. In Proceedings of the 18th International Conference on Hybrid Systems: Computation and Control, HSCC ’15, pages 179–188,
New York, NY, USA, 2015. ACM. ISBN 978-1-4503-3433-4. doi: 10.1145/2728606.2728629.
[48] Richard L. Burden and John Douglas Faires. Numerical Analysis. Cengage Learning, 9
edition, 2010. ISBN 0538733519.
[49] Martin Busch. On the efficient coupling of simulation codes. kassel university press GmbH,
2012. ISBN 3862192962.
[50] Martin Busch. Continuous approximation techniques for co-simulation methods: Analysis of
numerical stability and local error. ZAMM - Journal of Applied Mathematics and Mechanics
/ Zeitschrift für Angewandte Mathematik und Mechanik, 96(9):1061–1081, sep 2016. ISSN
00442267. doi: 10.1002/zamm.201500196.
[51] Martin Busch and Bernhard Schweizer. Numerical stability and accuracy of different cosimulation techniques: analytical investigations based on a 2-DOF test model. In 1st Joint
International Conference on Multibody System Dynamics, pages 25–27, 2010.
[52] Martin Busch and Bernhard Schweizer. An explicit approach for controlling the macro-step
size of co-simulation methods. Proceedings of The 7th European Nonlinear Dynamics, ENOC,
pages 24–29, 2011.
[53] Martin Busch and Bernhard Schweizer. Stability of Co-Simulation Methods Using Hermite
and Lagrange Approximation Techniques. In ECCOMAS Thematic Conference on Multibody
Dynamics, pages 1–10, Brussels, Belgium, jul 2011.
[54] Martin Busch and Bernhard Schweizer. Coupled simulation of multibody and finite element
systems: an efficient and robust semi-implicit coupling approach. Archive of Applied Mechanics, 82(6):723–741, jun 2012. ISSN 0939-1533. doi: 10.1007/s00419-011-0586-0.
[55] Benjamin Camus, Christine Bourjot, and Vincent Chevrier. Combining DEVS with multiagent concepts to design and simulate multi-models of complex systems (WIP). In Proceedings
of the Symposium on Theory of Modeling & Simulation: DEVS Integrative M&S Symposium,
pages 85–90. Society for Computer Simulation International, 2015. ISBN 978-1-5108-0105-9.
[56] Benjamin Camus, Virginie Galtier, Mathieu Caujolle, Vincent Chevrier, Julien Vaubourg,
Laurent Ciarletta, and Christine Bourjot. Hybrid Co-simulation of FMUs using DEV&DESS
in MECSYCO. In Proceedings of the Symposium on Theory of Modeling & Simulation DEVS Integrative M&S Symposium (TMS/DEVS 16), Pasadena, CA, United States, jan
2016. Université de Lorraine, CNRS, Inria, LORIA, UMR 7503 ; CentraleSupélec UMI GTCNRS 2958 Université Paris-Saclay ; EDF - R&D MIRE/R44.
58
[57] Luca P. Carloni, Roberto Passerone, Alessandro Pinto, and Alberto L. Angiovanni-Vincentelli.
Languages and Tools for Hybrid Systems Design. Foundations and Trends R in Electronic
Design Automation, 1(1/2):1–193, 2006. ISSN 1551-3939. doi: 10.1561/1000000001.
[58] Christopher D. Carothers, Kalyan S. Perumalla, and Richard M. Fujimoto. Efficient optimistic
parallel simulations using reverse computation. ACM Transactions on Modeling and Computer
Simulation, 9(3):224–253, jul 1999. ISSN 10493301. doi: 10.1145/347823.347828.
[59] Volker Carstens, Ralf Kemme, and Stefan Schmitt. Coupled simulation of flow-structure
interaction in turbomachinery. Aerospace Science and Technology, 7(4):298–306, jun 2003.
ISSN 12709638. doi: 10.1016/S1270-9638(03)00016-6.
[60] François Edouard Cellier. Combined Continuous/Discrete System Simulation Languages:
Usefulness, Experiences and Future Development. SIGSIM Simul. Dig., 9(1):18–21, 1977.
ISSN 0163-6103. doi: 10.1145/1102505.1102514.
[61] François Edouard Cellier. Combined Continuous Discrete Simulation by use of Digital Computers: Techniques and Tools. PhD thesis, 1979.
[62] François Edouard Cellier. Continuous system modeling. Springer Science & Business Media,
1991.
[63] François Edouard Cellier and Ernesto Kofman. Continuous System Simulation. Springer
Science & Business Media, 2006. ISBN 9780387261027.
[64] K.M. Chandy and J Misra. Distributed Simulation: A Case Study in Design and Verification
of Distributed Programs. IEEE Transactions on Software Engineering, SE-5(5):440–452, sep
1979. ISSN 0098-5589. doi: 10.1109/TSE.1979.230182.
[65] W-T. Chang, A. Kalavade, and E. A. Lee. Effective Heterogenous Design and Co-Simulation.
In Giovanni De Micheli and Mariagiovanna Sami, editors, Hardware/Software Co-Design,
pages 187–212. Springer Netherlands, Dordrecht, 1996. ISBN 978-94-009-0187-2. doi: 10.
1007/978-94-009-0187-2 8.
[66] Yung-Wei Chen, Chein-Shan Liu, and Jiang-Ren Chang. A chaos detectable and time step-size
adaptive numerical scheme for nonlinear dynamical systems. Journal of Sound and Vibration,
299(4-5):977–989, feb 2007. ISSN 0022460X. doi: 10.1016/j.jsv.2006.08.028.
[67] Alex Chung Hen Chow and Bernard P. Zeigler. Parallel DEVS: A Parallel, Hierarchical,
Modular, Modeling Formalism. In Proceedings of the 26th Conference on Winter Simulation, WSC ’94, pages 716–722, San Diego, CA, USA, 1994. Society for Computer Simulation
International. ISBN 0-7803-2109-X.
[68] Fabio Cremona, Marten Lohstroh, Stavros Tripakis, Christopher Brooks, and Edward A Lee.
FIDE: an FMI integrated development environment. In Proceedings of the 31st Annual ACM
Symposium on Applied Computing - SAC ’16, SAC ’16, pages 1759–1766, New York, New
York, USA, 2016. ACM Press. ISBN 9781450337397. doi: 10.1145/2851613.2851677.
[69] Makarand Datar, Ilinca Stanciulescu, and Dan Negrut. A co-simulation environment for
high-fidelity virtual prototyping of vehicle systems. International Journal of Vehicle Systems
Modelling and Testing, 7(1):54, jan 2012. ISSN 1745-6436. doi: 10.1504/IJVSMT.2012.045308.
59
[70] Joachim Denil, Bart Meyers, Paul De Meulenaere, and Hans Vangheluwe. Explicit Semantic
Adaptation of Hybrid Formalisms for FMI Co-Simulation. In Society for Computer Simulation
International, editor, Proceedings of the Symposium on Theory of Modeling & Simulation:
DEVS Integrative M&S Symposium, pages 99–106, Alexandria, Virginia, 2015.
[71] Stefan Dietz, Gerhard HIPPMANN, and Gunter SCHUPP. Interaction of vehicles and flexible
tracks by co-simulation of multibody vehicle systems and finite element track models. Vehicle
system dynamics, 37:372–384, 2002. ISSN 0042-3114.
[72] Paul Adrien Maurice Dirac. The principles of quantum mechanics. Number 27. Oxford
university press, 1981. ISBN 0198520115.
[73] W. Stuart Dols, Steven J. Emmerich, and Brian J. Polidoro. Coupling the multizone airflow
and contaminant transport software CONTAM with EnergyPlus using co-simulation. Building
Simulation, 9(4):469–479, aug 2016. ISSN 1996-3599. doi: 10.1007/s12273-016-0279-2.
[74] Edo Drenth, Mikael Törmänen, Krister Johansson, Bengt-Arne Andersson, Daniel Andersson,
Ivar Torstensson, and Johan Åkesson. Consistent Simulation Environment with FMI based
Tool Chain. In Proceedings of 10th International Modelica Conference, Lund, Sweden, 2014.
[75] Sven Dronka and Jochen Rauh. Co-simulation-interface for user-force-elements. In Proceedings
of SIMPACK user meeting, 2006.
[76] John C. Eidson, Edward A. Lee, Slobodan Matic, Sanjit A. Seshia, and Jia Zou. Distributed
Real-Time Software for Cyber-Physical Systems. Proceedings of the IEEE, 100(1):45–59, jan
2012. ISSN 0018-9219. doi: 10.1109/JPROC.2011.2161237.
[77] H. El Tahawy, D. Rodriguez, S. Garcia-Sabiro, and J.-J. Mayol. VHD/sub e/LDO: A new
mixed mode simulation. In Proceedings of EURO-DAC 93 and EURO-VHDL 93- European
Design Automation Conference, pages 546–551. IEEE Comput. Soc. Press, 1993. ISBN 08186-4350-1. doi: 10.1109/EURDAC.1993.410690.
[78] Atiyah Elsheikh, Muhammed Usman Awais, Edmund Widl, and Peter Palensky. Modelicaenabled rapid prototyping of cyber-physical energy systems via the functional mockup interface. In 2013 Workshop on Modeling and Simulation of Cyber-Physical Energy Systems
(MSCPES), pages 1–6. IEEE, may 2013. ISBN 978-1-4799-1307-7. doi: 10.1109/MSCPES.
2013.6623315.
[79] Olaf Enge-Rosenblatt, Christoph Clauß, André Schneider, Peter Schneider, and Olaf Enge.
Functional Digital Mock-up and the Functional Mock-up Interface–Two Complementary Approaches for a Comprehensive Investigation of Heterogeneous Systems. In Proc. of the 8th
Int. Modelica Conference, 2011.
[80] Kenneth Eriksson, Donald Estep, and Claes Johnson. Applied Mathematics: Body and Soul.
Springer Berlin Heidelberg, Berlin, Heidelberg, 2004. ISBN 978-3-642-05659-8. doi: 10.1007/
978-3-662-05796-4.
[81] Emeka Eyisi, Jia Bai, Derek Riley, Jiannian Weng, Wei Yan, Yuan Xue, Xenofon Koutsoukos,
and Janos Sztipanovits. NCSWT: An integrated modeling and simulation tool for networked
control systems. Simulation Modelling Practice and Theory, 27:90–111, sep 2012. ISSN
1569190X. doi: 10.1016/j.simpat.2012.05.004.
60
[82] Cyril Faure, Mongi Ben Gaid, Nicolas Pernet, Morgan Fremovici, Gregory Font, and
Gilles Corde. Methods for real-time simulation of Cyber-Physical Systems: application
to automotive domain. In 2011 7th International Wireless Communications and Mobile
Computing Conference, pages 1105–1110. IEEE, jul 2011. ISBN 978-1-4244-9539-9. doi:
10.1109/IWCMC.2011.5982695.
[83] Yishai A. Feldman, Lev Greenberg, and Eldad Palachi. Simulating Rhapsody SysML Blocks
in Hybrid Models with FMI. In Proceedings of the 10th International Modelica Conference,
pages 43–52. Linköping University Electronic Press, mar 2014. doi: 10.3384/ecp1409643.
[84] P. Fey, H.W. Carter, and P.A. Wilsey. Parallel synchronization of continuous time discrete
event simulators. In Proceedings of the 1997 International Conference on Parallel Processing
(Cat. No.97TB100162), pages 227–231. IEEE Comput. Soc, 1997. ISBN 0-8186-8108-X. doi:
10.1109/ICPP.1997.622649.
[85] J. S. Fitzgerald and P. G. Larsen. Balancing Insight and Effort: the Industrial Uptake
of Formal Methods. In Cliff B. Jones, Zhiming Liu, and Jim Woodcock, editors, Formal
Methods and Hybrid Real-Time Systems, Essays in Honour of Dines Bjørner and Chaochen
Zhou on the Occasion of Their 70th Birthdays, pages 237–254, Volume 4700, September 2007.
Springer, Lecture Notes in Computer Science. ISBN 978-3-540-75220-2.
[86] John Fitzgerald, Peter Gorm Larsen, Ken Pierce, Marcel Verhoef, and Sune Wolff. Collaborative Modelling and Co-simulation in the Development of Dependable Embedded Systems.
pages 12–26. Springer Berlin Heidelberg, Berlin, Heidelberg, 2010. ISBN 978-3-642-16265-7.
doi: 10.1007/978-3-642-16265-7 2.
[87] John Fitzgerald, Peter Gorm Larsen, and Marcel Verhoef. Collaborative Design for Embedded
Systems. Springer Berlin Heidelberg, Berlin, Heidelberg, 2014. ISBN 978-3-642-54117-9. doi:
10.1007/978-3-642-54118-6.
[88] John Fitzgerald, Carl Gamble, Richard Payne, Peter Gorm Larsen, Stylianos Basagiannis,
and Alie El-Din Mady. Collaborative Model-based Systems Engineering for Cyber-Physical
Systems – a Case Study in Building Automation. In INCOSE 2016, Edinburgh, Scotland, jul
2016.
[89] John S. Fitzgerald, Peter Gorm Larsen, Ken G. Pierce, and Marcel Henri Gerard Verhoef. A
formal approach to collaborative modelling and co-simulation for embedded systems. Mathematical Structures in Computer Science, 23(04):726–750, aug 2013. ISSN 0960-1295. doi:
10.1017/S0960129512000242.
[90] Alain Fourmigue, Bruno Girodias, Gabriela Nicolescu, and E.-M. Aboulhamid. Co-simulation
based platform for wireless protocols design explorations. In 2009 Design, Automation & Test
in Europe Conference & Exhibition, pages 874–877. IEEE, apr 2009. ISBN 978-1-4244-3781-8.
doi: 10.1109/DATE.2009.5090785.
[91] P. Frey, R. Radhakrishnan, H. W. Carter, and P. A. Wilsey. Optimistic synchronization of
mixed-mode simulators. In 1998 First Merged International on Parallel Processing Symposium
and Symposium on Parallel and Distributed Processing, pages 694–699, 1998. ISBN 1063-7133
VO -. doi: 10.1109/IPPS.1998.670002.
61
[92] Jonathan Friedman and Jason Ghidella. Using model-based design for automotive systems
engineering-requirements analysis of the power window example. Technical report, 2006.
[93] Markus Friedrich. Parallel Co-Simulation for Mechatronic Systems. PhD thesis, 2011.
[94] Richard M. Fujimoto. Parallel and distributed simulation systems, volume 300. Wiley New
York, New York, USA, 1 edition, 2000. ISBN 0-7803-7307-3. doi: 10.1109/WSC.2001.977259.
[95] Jason C. Fuller, Selim Ciraci, Jeffrey A. Daily, Andrew R. Fisher, and M. Hauer. Communication simulations for power system applications. In 2013 Workshop on Modeling and
Simulation of Cyber-Physical Energy Systems (MSCPES), pages 1–6. IEEE, may 2013. ISBN
978-1-4799-1307-7. doi: 10.1109/MSCPES.2013.6623314.
[96] Virginie Galtier, Gilles Plessis, and Les Renardi. FMI-Based Distributed Multi-Simulation
with DACCOSIM. In Spring Simulation Multi-Conference, pages 804–811. Society for Computer Simulation International, 2015.
[97] Alfredo Garro and Alberto Falcone. On the integration of HLA and FMI for supporting interoperability and reusability in distributed simulation. In Spring Simulation Multi-Conference,
pages 774–781. Society for Computer Simulation International, 2015.
[98] C. W. Gear and D. R. Wells. Multirate linear multistep methods. BIT, 24(4):484–502, dec
1984. ISSN 0006-3835. doi: 10.1007/BF01934907.
[99] L. Gheorghe, F. Bouchhima, G. Nicolescu, and H. Boucheneb. Formal Definitions of Simulation Interfaces in a Continuous/Discrete Co-Simulation Tool. In Rapid System Prototyping,
2006. Seventeenth IEEE International Workshop on, pages 186–192, 2006. ISBN 1074-6005
VO -. doi: 10.1109/RSP.2006.18.
[100] L. Gheorghe, F. Bouchhima, G. Nicolescu, and H. Boucheneb. A Formalization of Global
Simulation Models for Continuous/Discrete Systems. In Proceedings of the 2007 Summer
Computer Simulation Conference, SCSC ’07, pages 559–566, San Diego, CA, USA, 2007.
Society for Computer Simulation International. ISBN 1-56555-316-0.
[101] A. Ghosh, M. Bershteyn, R. Casley, C. Chien, A. Jain, M. Lipsie, D. Tarrodaychik, and
O. Yamamoto. A hardware-software co-simulator for embedded system design and debugging.
In Design Automation Conference, pages 155–164, 1995. doi: 10.1109/ASPDAC.1995.486217.
[102] Cláudio Gomes. Foundations for Continuous Time Hierarchical Co-simulation. In ACM
Student Research Competition (ACM/IEEE 19th International Conference on Model Driven
Engineering Languages and Systems), page to appear, Saint Malo, Brittany, France, 2016.
[103] Francisco González, Miguel Ángel Naya, Alberto Luaces, and Manuel González. On the effect
of multirate co-simulation techniques in the efficiency and accuracy of multibody system
dynamics. Multibody System Dynamics, 25(4):461–483, apr 2011. ISSN 1384-5640. doi:
10.1007/s11044-010-9234-7.
[104] Matthias Gries. Methods for evaluating and covering the design space during early design
development. Integration, the VLSI Journal, 38(2):131–183, dec 2004. ISSN 01679260. doi:
10.1016/j.vlsi.2004.06.001.
62
[105] Bei Gu and H H Asada. Co-simulation of algebraically coupled dynamic subsystems. In
American Control Conference, 2001. Proceedings of the 2001, volume 3, pages 2273–2278
vol.3, 2001. ISBN 0743-1619 VO - 3. doi: 10.1109/ACC.2001.946089.
[106] Bei Gu and H. Harry Asada. Co-simulation of algebraically coupled dynamic subsystems. PhD
thesis, 2001.
[107] Bei Gu and H. Harry Asada. Co-Simulation of Algebraically Coupled Dynamic Subsystems
Without Disclosure of Proprietary Subsystem Models. Journal of Dynamic Systems, Measurement, and Control, 126(1):1, apr 2004. ISSN 00220434. doi: 10.1115/1.1648307.
[108] Felix Günther, Georg Mallebrein, and Heinz Ulbrich. A Modular Technique for Automotive
System Simulation. In Proc. 9th International Modelica Conference, Munich, Germany, 2012.
[109] M. Günther, A. Kværnø, and P. Rentrop. Multirate Partitioned Runge-Kutta Methods. Bit,
41(3):504–514, 2001. ISSN 00063835. doi: 10.1023/A:1021967112503.
[110] R. K. Gupta, C. N. Coelho Jr., and G. De Micheli. Synthesis and Simulation of Digital
Systems Containing Interacting Hardware and Software Components. In Proceedings of the
29th ACM/IEEE Design Automation Conference, DAC ’92, pages 225–230, Los Alamitos,
CA, USA, 1992. IEEE Computer Society Press. ISBN 0-89791-516-X.
[111] Dinesh Rangana Gurusinghe, Saranga Menike, A I Konara, Athula D Rajapakse, Pradeepa
Yahampath, U D Annakkage, Brian A Archer, and Tony Weekes. Co-simulation of Power System and Synchrophasor Communication Network on a Single Simulation Platform. Technology
and Economics of Smart Grids and Sustainable Energy, 1(1):6, dec 2016. ISSN 2199-4706. doi:
10.1007/s40866-015-0003-9. URL http://dx.doi.org/10.1007/s40866-015-0003-9http:
//link.springer.com/10.1007/s40866-015-0003-9.
[112] Kjell Gustafsson. Control of error and convergence in ODE solvers. PhD thesis, 1992.
[113] Kjell Gustafsson, Michael Lundh, and Gustaf Söderlind. API stepsize control for the numerical
solution of ordinary differential equations. BIT, 28(2):270–287, jun 1988. ISSN 0006-3835.
doi: 10.1007/BF01934091.
[114] Irene Hafner, Bernhard Heinzl, and Matthias Roessler. An Investigation on Loose Coupling
Co-Simulation with the BCVTB. SNE Simulation Notes Europe, 23(1), 2013. ISSN 2305-9974.
doi: 10.11128/sne.23.tn.10173.
[115] Walid Hassairi, Moncef Bousselmi, Mohamed Abid, and Carlos Valderrama. Matlab/SystemC
for the New Co-Simulation Environment by JPEG Algorithm. MATLAB–A Fundamental Tool
for Scientific Computing and Engineering Applications, 2:120–138, 2012.
[116] D. R. Hickey, P. A. Wilsey, R. J. Hoekstra, E. R. Keiter, S. A. Hutchinson, and T. V. Russo.
Mixed-signal simulation with the Simbus backplane. In 39th Annual Simulation Symposium,
2006, Huntsville, AL, 2006. doi: 10.1109/ANSS.2006.25.
[117] Ken Hines and Gaetano Borriello. Selective focus as a means of improving geographically
distributed embedded system co-simulation. In 8th IEEE International Workshop on Rapid
System Prototyping, 1997, pages 58–62, 1997. doi: 10.1109/IWRSP.1997.618825.
63
[118] Ken Hines and Gaetano Borriello. Dynamic Communication Models in Embedded System Cosimulation. In Proceedings of the 34th Annual Design Automation Conference, DAC ’97, pages
395–400, New York, NY, USA, 1997. ACM. ISBN 0-89791-920-3. doi: 10.1145/266021.266178.
[119] Matthias Hoepfer. Towards a Comprehensive Framework for Co- Simulation of Dynamic
Models With an Emphasis on Time Stepping. PhD thesis, 2011.
[120] Hua Lin, Santhoshkumar Sambamoorthy, Sandeep Shukla, James Thorp, and Lamine Mili.
Power system and communication network co-simulation for smart grid applications. In ISGT
2011, pages 1–6. IEEE, jan 2011. ISBN 978-1-61284-218-9. doi: 10.1109/ISGT.2011.5759166.
[121] Z. Jackiewicz and M. Kwapisz. Convergence of Waveform Relaxation Methods for DifferentialAlgebraic Systems. SIAM Journal on Numerical Analysis, 33(6):2303–2317, dec 1996. ISSN
0036-1429. doi: 10.1137/S0036142992233098.
[122] Kenneth R Jackson. A survey of parallel numerical methods for initial value problems for
ordinary differential equations. IEEE Transactions on Magnetics, 27(5):3792–3797, 1991.
ISSN 0018-9464.
[123] David Jefferson and Henry A. Sowizral. Fast concurrent simulation using the time warp
mechanism. Technical report, Rand Corporation, 1982.
[124] David R. Jefferson. Virtual time. ACM Transactions on Programming Languages and Systems,
7(3):404–425, jul 1985. ISSN 01640925. doi: 10.1145/3916.3988.
[125] Karl Henrik Johansson, Magnus Egerstedt, John Lygeros, and Shankar Sastry. On the regularization of Zeno hybrid automata. Systems & Control Letters, 38(3):141–150, oct 1999.
ISSN 01676911. doi: 10.1016/S0167-6911(99)00059-6.
[126] Raphaël Jungers. The joint spectral radius: theory and applications, volume 385. Springer
Science & Business Media, 2009. ISBN 3540959793.
[127] Rudolph Emil Kalman. A new approach to linear filtering and prediction problems. Journal
of basic Engineering, 82(1):35–45, 1960. ISSN 0021-9223.
[128] Tamas Kalmar-Nagy and Ilinca Stanciulescu. Can complex systems really be simulated?
Applied Mathematics and Computation, 227:199–211, jan 2014. ISSN 00963003. doi: 10.
1016/j.amc.2013.11.037.
[129] K. C. Kang, S. Cohen, J. Hess, W. Novak, and A. Peterson. Feature-Oriented Domain
Analysis. Feasibility study,. Technical report, Carnegie Mellon University, 1990.
[130] M. Karner, M. Krammer, S. Krug, E. Armengaud, C. Steger, and R. Weiss. Heterogeneous
co-simulation platform for the efficient analysis of FlexRay-based automotive distributed embedded systems. In 8th IEEE International Workshop on Factory Communication Systems
(WFCS), pages 231–240, 2010. doi: 10.1109/WFCS.2010.5548627.
[131] Michael Karner, Eric Armengaud, Christian Steger, and Reinhold Weiss. Holistic Simulation
of FlexRay Networks by Using Run-time Model Switching. In Proceedings of the Conference
on Design, Automation and Test in Europe, DATE ’10, pages 544–549, 3001 Leuven, Belgium,
Belgium, 2010. European Design and Automation Association. ISBN 978-3-9810801-6-2.
64
[132] Michael Karner, Martin Krammer, Markus Schratter, Peter Wimmer, Daniel Watzenig, and
ChristianMichael Gruber. A Comprehensive Approach for Modeling, Simulation and Virtual
Validation of Integrated Safety Systems. In Jan Fischer-Wolfarth and Gereon Meyer, editors,
Advanced Microsystems for Automotive Applications 2013 SE - 10, Lecture Notes in Mobility,
pages 101–110. Springer International Publishing, 2013. ISBN 978-3-319-00475-4. doi: 10.
1007/978-3-319-00476-1 10.
[133] Abir Ben Khaled, Mongi Ben Gaid, Daniel Simon, and Gregory Font. Multicore simulation of powertrains using weakly synchronized model partitioning. In IFAC Proceedings
Volumes, volume 45, pages 448–455, Rueil-Malmaison, France, oct 2012. doi: 10.3182/
20121023-3-FR-4025.00018.
[134] Abir Ben Khaled, Laurent Duval, Mohamed El Mongi Ben Gaı̈d, and Daniel Simon. Contextbased polynomial extrapolation and slackened synchronization for fast multi-core simulation
using FMI. In 10th International Modelica Conference 2014, pages 225–234. Link{ö}ping
University Electronic Press, 2014.
[135] James Ellis Kleckner. Advanced mixed-mode simulation techniques. PhD thesis, 1984.
[136] Ernesto Kofman. A Second-Order Approximation for DEVS Simulation of Continuous Systems. SIMULATION, 78(2):76–89, feb 2002. ISSN 0037-5497. doi: 10.1177/
0037549702078002206.
[137] Ernesto Kofman and Sergio Junco. Quantized-state systems: a DEVS Approach for continuous system simulation. Transactions of The Society for Modeling and Simulation International, 18(3):123–132, 2001. ISSN 0740-6797.
[138] Alexander Kossiakoff, William N. Sweet, Samuel J. Seymour, and Steven M. Biemer. Structure
of Complex Systems, pages 41–67. John Wiley & Sons, Inc., 2011. ISBN 9781118001028. doi:
10.1002/9781118001028.ch3. URL http://dx.doi.org/10.1002/9781118001028.ch3.
[139] Velin Kounev, David Tipper, Martin Levesque, Brandon M. Grainger, Thomas Mcdermott,
and Gregory F. Reed. A microgrid co-simulation framework. In 2015 Workshop on Modeling
and Simulation of Cyber-Physical Energy Systems (MSCPES), pages 1–6. IEEE, apr 2015.
ISBN 978-1-4799-7357-6. doi: 10.1109/MSCPES.2015.7115398.
[140] Martin Krammer, Johannes Fritz, and Michael Karner. Model-Based Configuration of Automotive Co-Simulation Scenarios. In Spring Simulation Multi-Conference, pages 246–253.
Society for Computer Simulation International, 2015.
[141] R. Kübler and W. Schiehlen. Two Methods of Simulator Coupling. Mathematical and
Computer Modelling of Dynamical Systems, 6(2):93–113, jun 2000. ISSN 1387-3954. doi:
10.1076/1387-3954(200006)6:2;1-M;FT093.
[142] R. Kübler and W. Schiehlen. Modular Simulation in Multibody System Dynamics. Multibody
System Dynamics, 4(2-3):107–127, 2000. ISSN 1384-5640. doi: 10.1023/A:1009810318420.
[143] Michal Kudelski, Luca M. Gambardella, and Gianni A. Di Caro. RoboNetSim: An integrated
framework for multi-robot and network simulation. Robotics and Autonomous Systems, 61
(5):483–496, may 2013. ISSN 09218890. doi: 10.1016/j.robot.2013.01.003.
65
[144] Frederick Kuhl, Richard Weatherly, and Judith Dahmann. Creating computer simulation
systems: an introduction to the high level architecture. Prentice Hall PTR, 1999. ISBN
0130225118.
[145] Thomas Kühne. What is a Model? In Jean Bezivin and Reiko Heckel, editors, Language Engineering for Model-Driven Software Development, volume 04101. Internationales Begegnungsund Forschungszentrum für Informatik (IBFI), 2005.
[146] T. Kuhr, T. Forster, T. Braun, and R. Gotzhein. FERAL - Framework for simulator coupling
on requirements and architecture level. In Eleventh IEEE/ACM International Conference on
Formal Methods and Models for Codesign (MEMOCODE), pages 11–22, 2013.
[147] Leslie Lamport. Time, clocks, and the ordering of events in a distributed system. Communications of the ACM, 21(7):558–565, jul 1978. ISSN 00010782. doi: 10.1145/359545.359563.
[148] Peter Gorm Larsen, Casper Thule, Kenneth Lausdahl, Victor Bardur, Carl Gamble, Etienne
Brosse, Andrey Sadovykh, Alessandra Bagnato, and Luis Diogo Couto. Integrated Tool Chain
for Model-Based Design of Cyber-Physical Systems. In Peter Gorm Larsen, Nico Plat, and
Nick Battle, editors, The 14th Overture Workshop: Towards Analytical Tool Chains, pages
63–78, Cyprus, November 2016. Aarhus University, Department of Engineering. ECE-TR-28.
[149] Kenneth Lausdahl, Peter Gorm Larsen, Sune Wolf, Victor Bandur, Anders Terkelsen, Miran
Hasanagić, Casper Thule Hansen, Ken Pierce, Oliver Kotte, Adrian Pop, Etienne Brosse, Jörg
Brauer, and Oliver Möller. Design of the INTO-CPS Platform. Technical report, INTO-CPS
Deliverable, D4.1d, December 2015.
[150] David P. Y. Lawrence, Cláudio Gomes, Joachim Denil, Hans Vangheluwe, and Didier Buchs.
Coupling Petri nets with Deterministic Formalisms Using Co-simulation. In Proceedings of
the Symposium on Theory of Modeling & Simulation: DEVS Integrative M&S Symposium,
pages 6:1—-6:8, Pasadena, CA, USA, 2016.
[151] P. Le Marrec, C. A. Valderrama, F. Hessel, A. A. Jerraya, M. Attia, and O. Cayrol. Hardware,
software and mechanical cosimulation for automotive applications. In 9th International Workshop on Rapid System Prototyping, pages 202–206, 1998. doi: 10.1109/IWRSP.1998.676692.
[152] Bu-Sung Lee, Wentong Cai, Stephen J. Turner, and L. Chen. Adaptive dead reckoning
algorithms for distributed interactive simulation. International Journal of Simulation, 1(1-2):
21–34, 2000.
[153] Edward A. Lee. Cyber Physical Systems: Design Challenges. In 11th IEEE International
Symposium on Object Oriented Real-Time Distributed Computing (ISORC), pages 363–369,
2008. doi: 10.1109/ISORC.2008.25.
[154] Edward A. Lee and Haiyang Zheng. Operational semantics of hybrid systems. In Manfred
Morari and Lothar Thiele, editors, Hybrid Systems: Computation and Control, volume 3414
of Lecture Notes in Computer Science, pages 25–53. Springer Berlin Heidelberg, 2005. ISBN
978-3-540-25108-8. doi: 10.1007/978-3-540-31954-2 2.
[155] E. Lelarasmee, Albert E. Ruehli, and A. L. Sangiovanni-Vincentelli. The Waveform Relaxation
Method for Time-Domain Analysis of Large Scale Integrated Circuits. In IEEE Transactions
66
on Computer-Aided Design of Integrated Circuits and Systems, volume 1, pages 131–145,
1982. ISBN 0278-00701. doi: 10.1109/TCAD.1982.1270004.
[156] Shengqin Li and Le He. Co-simulation Study of Vehicle ESP System Based on ADAMS and
MATLAB. Journal of Software, 6(5), 2011.
[157] W. Li, A. Monti, M. Luo, and R. A. Dougal. VPNET: A co-simulation framework for analyzing
communication channel effects on power systems. In 2011 IEEE Electric Ship Technologies
Symposium, pages 143–149. IEEE, apr 2011. ISBN 978-1-4244-9272-5. doi: 10.1109/ESTS.
2011.5770857.
[158] Weilin Li, Min Luo, Lin Zhu, Antonello Monti, and Ferdinanda Ponci. A co-simulation method
as an enabler for jointly analysis and design of MAS-based electrical power protection and
communication. SIMULATION, mar 2013.
[159] Weilin Li, Xiaobin Zhang, and Huimin Li. Co-simulation platforms for co-design of networked control systems: An overview. Control Engineering Practice, 23:44–56, feb 2014.
ISSN 09670661. doi: 10.1016/j.conengprac.2013.10.010.
[160] F. Liu, J. Cai, Y. Zhu, H. M. Tsai, and A. S. F. Wong. Calculation of Wing Flutter by
a Coupled Fluid-Structure Method. Journal of Aircraft, 38(2):334–342, mar 2001. ISSN
0021-8669. doi: 10.2514/2.2766.
[161] John Lygeros. Lecture notes on hybrid systems, 2004.
[162] Oded Maler, Zohar Manna, and Amir Pnueli. From timed to hybrid systems. Real-Time:
Theory in Practice, 600:447–484, 1992. doi: ˜10.1007/BFb0032003.
[163] M. Manbachi, A. Sadu, H. Farhangi, A. Monti, A. Palizban, F. Ponci, and S. Arzanpour.
Impact of EV penetration on Volt–VAR Optimization of distribution networks using real-time
co-simulation monitoring platform. Applied Energy, 169:28–39, may 2016. ISSN 03062619.
doi: 10.1016/j.apenergy.2016.01.084.
[164] Toni Mancini, Federico Mari, Annalisa Massini, Igor Melatti, Fabio Merli, and Enrico Tronci.
System Level Formal Verification via Model Checking Driven Simulation. In Natasha Sharygina and Helmut Veith, editors, Computer Aided Verification SE - 21, volume 8044 of Lecture
Notes in Computer Science, pages 296–312. Springer Berlin Heidelberg, 2013. ISBN 978-3642-39798-1. doi: 10.1007/978-3-642-39799-8 21.
[165] Zohar Manna and Amir Pnueli. Verifying hybrid systems. In RobertL. Grossman, Anil Nerode,
AndersP. Ravn, and Hans Rischel, editors, Hybrid Systems SE - 2, volume 736 of Lecture Notes
in Computer Science, pages 4–35. Springer Berlin Heidelberg, 1993. ISBN 978-3-540-57318-0.
doi: 10.1007/3-540-57318-6 22. URL http://dx.doi.org/10.1007/3-540-57318-6{_}22.
[166] J. A. Martinez, T. P. Kurzweg, S. P. Levitan, P. J. Marchand, and D. M. Chiarulli. MixedTechnology System-Level Simulation. Analog Integrated Circuits and Signal Processing, 29
(1-2):127–149, 2001. ISSN 0925-1030. doi: 10.1023/A:1011294616831.
[167] William J. McCalla. Fundamentals of computer-aided circuit simulation, volume 37. Springer
Science & Business Media, 1987. ISBN 1461320119.
67
[168] M Mews, J Svacina, and S Weissleder. From AUTOSAR Models to Co-simulation for MiLTesting in the Automotive Domain. In Software Testing, Verification and Validation (ICST),
2012 IEEE Fifth International Conference on, pages 519–528, 2012. ISBN VO -. doi: 10.
1109/ICST.2012.137.
[169] Bart Meyers, Joachim Denil, Frédéric Boulanger, Cécile Hardebolle, Christophe Jacquet,
and Hans Vangheluwe. A DSL for Explicit Semantic Adaptation. In Edward Jones
Tamás Mészáros Christophe Jacquet Daniel Balasubramanian, editor, MPM 2013, number
1112 in CEUR Workshop Proceedings, pages 47–56, Miami, United States, sep 2013.
[170] U. Miekkala and O. Nevanlinna. Convergence of Dynamic Iteration Methods for Initial Value
Problems. SIAM Journal on Scientific and Statistical Computing, 8(4):459–482, jul 1987.
ISSN 0196-5204. doi: 10.1137/0908046.
[171] Debasis Mitra. Asynchronous Relaxations for the Numerical Solution of Differential Equations
by Parallel Processors. SIAM Journal on Scientific and Statistical Computing, 8(1):s43–s58,
jan 1987. ISSN 0196-5204. doi: 10.1137/0908012.
[172] Pieter J. Mosterman. An Overview of Hybrid Simulation Phenomena and Their Support
by Simulation Packages. In FritsW. Vaandrager and JanH. van Schuppen, editors, Hybrid
Systems: Computation and Control SE - 17, volume 1569 of Lecture Notes in Computer
Science, pages 165–177. Springer Berlin Heidelberg, 1999. ISBN 978-3-540-65734-7. doi:
10.1007/3-540-48983-5 17.
[173] Pieter J. Mosterman. HYBRSIM—a modelling and simulation environment for hybrid bond
graphs. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems
and Control Engineering, 216(1), 2002. doi: 10.1243/0959651021541417.
[174] Pieter J. Mosterman and Gautam Biswas. A theory of discontinuities in physical system
models. Journal of the Franklin Institute, 335(3):401–439, apr 1998. ISSN 00160032. doi:
10.1016/S0016-0032(96)00126-3.
[175] Pieter J. Mosterman, Justyna Zander, Gregoire Hamon, and Ben Denckla. A computational
model of time for stiff hybrid systems applied to control synthesis. Control Engineering
Practice, 20(1):2–13, 2012. ISSN 09670661. doi: 10.1016/j.conengprac.2011.04.013.
[176] W Muller and E Widl. Using FMI components in discrete event systems. In Modeling and
Simulation of Cyber-Physical Energy Systems (MSCPES), 2015 Workshop on, pages 1–6,
2015. ISBN VO -. doi: 10.1109/MSCPES.2015.7115397.
[177] Sadaf Mustafiz and Hans Vangheluwe. Explicit Modelling of Statechart Simulation Environments. In Proceedings of the 2013 Summer Computer Simulation Conference, SCSC ’13,
pages 21:1—-21:8, Vista, CA, 2013. Society for Modeling & Simulation International. ISBN
978-1-62748-276-9.
[178] Sadaf Mustafiz, Bruno Barroca, Cláudio Gomes, and Hans Vangheluwe. Towards Modular
Language Design Using Language Fragments: The Hybrid Systems Case Study. In 13th
International Conference on Information Technology - New Generations (ITNG), pages 785–
797. 2016. doi: 10.1007/978-3-319-32467-8 68.
68
[179] Sadaf Mustafiz, Cláudio Gomes, Bruno Barroca, and Hans Vangheluwe. Modular Design
of Hybrid Languages by Explicit Modeling of Semantic Adaptation. In Proceedings of the
Symposium on Theory of Modeling & Simulation: DEVS Integrative M&S Symposium, DEVS
’16, pages 29:1—-29:8, San Diego, CA, USA, 2016.
[180] Alexandre Muzy, Luc Touraille, Hans Vangheluwe, Olivier Michel, Mamadou Kaba Traoré,
and David R. C. Hill. Activity Regions for the Specification of Discrete Event Systems. In
Proceedings of the 2010 Spring Simulation Multiconference, SpringSim ’10, pages 136:1—136:7, San Diego, CA, USA, 2010. Society for Computer Simulation International. ISBN
978-1-4503-0069-8. doi: 10.1145/1878537.1878679.
[181] Andreas Naderlinger. Multiple Real-time Semantics on Top of Synchronous Block Diagrams.
In Proceedings of the Symposium on Theory of Modeling & Simulation - DEVS Integrative
M&S Symposium, DEVS 13, pages 6:1—-6:7, San Diego, CA, USA, 2013. Society for Computer
Simulation International. ISBN 978-1-62748-032-1.
[182] Himanshu Neema, Jesse Gohl, Zsolt Lattmann, Janos Sztipanovits, Gabor Karsai, Sandeep
Neema, Ted Bapty, John Batteh, Hubertus Tummescheit, and Chandrasekar Sureshkumar.
Model-based integration platform for FMI co-simulation and heterogeneous simulations of
cyber-physical systems. In 10th International Modelica Conference, pages 10–12, 2014.
[183] Arthur Richard Newton and Alberto L. Sangiovanni-Vincentelli. Relaxation-Based Electrical
Simulation. SIAM Journal on Scientific and Statistical Computing, 4(3):485–524, sep 1983.
ISSN 0196-5204. doi: 10.1137/0904036.
[184] Y. Ni and J. F. Broenink. Hybrid systems modelling and simulation in DESTECS: a cosimulation approach. In M Klumpp, editor, 26th European Simulation and Modelling Conference, ESM 2012, Essen, Germany, pages 32–36, Ghent, Belgium, 2012. EUROSIS-ETI.
[185] Claus Ballegaard Nielsen, Peter Gorm Larsen, John Fitzgerald, Jim Woodcock, and Jan
Peleska. Systems of Systems Engineering: Basic Concepts, Model-Based Techniques, and
Research Directions. ACM Comput. Surv., 48(2):18:1—-18:41, sep 2015. ISSN 0360-0300.
doi: 10.1145/2794381.
[186] Henrik Nilsson. Functional automatic differentiation with dirac impulses. ACM SIGPLAN
Notices, 38(9):153–164, sep 2003. ISSN 03621340. doi: 10.1145/944746.944720.
[187] Kristoffer Norling, David Broman, Peter Fritzson, Alexander Siemers, and Dag Fritzson. Secure distributed co-simulation over wide area networks. In Proceedings of the 48th Conference
on Simulation and Modelling (SIMS 2007), pages 14–23. Citeseer, 2007.
[188] Thierry Nouidui, Michael Wetter, and Wangda Zuo. Functional mock-up unit for cosimulation import in EnergyPlus. Journal of Building Performance Simulation, 7(3):192–202,
may 2014. ISSN 1940-1493. doi: 10.1080/19401493.2013.808265.
[189] James Nutaro. Designing power system simulators for the smart grid: Combining controls,
communications, and electro-mechanical dynamics. In 2011 IEEE Power and Energy Society
General Meeting, pages 1–5. IEEE, jul 2011. ISBN 978-1-4577-1000-1. doi: 10.1109/PES.
2011.6039456.
69
[190] James J. Nutaro. A method for bounding error in multi-rate and federated simulations.
Proceeding of the 2016 Winter Simulation Conference, page to be accepted, 2016.
[191] Seaseung Oh and Suyong Chae. A Co-Simulation Framework for Power System Analysis.
Energies, 9(3):131, 2016.
[192] Xiufeng Pang, Michael Wetter, Prajesh Bhattacharya, and Philip Haves. A framework for
simulation-based real-time whole building performance assessment. Building and Environment, 54:100–108, aug 2012. ISSN 03601323. doi: 10.1016/j.buildenv.2012.02.003.
[193] Henry M. Paynter. Analysis and design of engineering systems. MIT press, 1961.
[194] Nicolai Pedersen, Jan Madsen, and Morten Vejlgaard-Laursen. Co-Simulation of Distributed
Engine Control System and Network Model using FMI & SCNSL. IFAC-PapersOnLine, 48
(16):261–266, 2015. ISSN 24058963. doi: 10.1016/j.ifacol.2015.10.290.
[195] Nicolai Pedersen, Tom Bojsen, Jan Madsen, and Morten Vejlgaard-Laursen. FMI for CoSimulation of Embedded Control Software. In The First Japanese Modelica Conferences,
May 23-24, Tokyo, Japan, number 124, pages 70–77. Linköping University Electronic Press,
may 2016. ISBN 1650-3740. doi: 10.3384/ecp1612470.
[196] Nikos C. Petrellis, Alexis N. Birbas, Michael K. Birbas, Evangelinos P. Mariatos, and
George D. Papadopoulos. Simulating Hardware, Software and Electromechanical Parts Using Communicating Simulators. Design Automation for Embedded Systems, 3(2/3):187–198,
1998. ISSN 09295585. doi: 10.1023/A:1008846508549.
[197] Ralf Uwe Pfau. A priori step size adaptation for the simulation of non-smooth systems. Communications in Numerical Methods in Engineering, 23(2):85–96, jun 2006. ISSN 10698299.
doi: 10.1002/cnm.884.
[198] K. Pierce, C. Gamble, Y. Ni, and J. F. Broenink. Collaborative Modelling and Co-simulation
with DESTECS: A Pilot Study. In IEEE 21st International Workshop on Enabling Technologies: Infrastructure for Collaborative Enterprises (WETICE), pages 280–285, 2012. ISBN
1524-4547 VO -. doi: 10.1109/WETICE.2012.69.
[199] Uwe Pohlmann, Wilhelm Schäfer, Hendrik Reddehase, Jens Röckemann, and Robert Wagner.
Generating functional mockup units from software specifications. In Proceedings of the 9th
International MODELICA Conference, September 3-5, 2012, Munich, Germany, number 078,
pages 765–774, 2012.
[200] Claudius Ptolemaeus. System Design, Modeling, and Simulation: Using Ptolemy II. Berkeley:
Ptolemy.org, 2014. ISBN 1304421066.
[201] Davide Quaglia, Riccardo Muradore, Roberto Bragantini, and Paolo Fiorini. A SystemC/Matlab co-simulation tool for networked control systems. Simulation Modelling Practice and
Theory, 23:71–86, apr 2012. ISSN 1569-190X. doi: 10.1016/j.simpat.2012.01.003.
[202] Gauthier Quesnel, Raphaël Duboz, David Versmisse, and E Ramat. DEVS coupling of spatial
and ordinary differential equations: VLE framework. OICIMS, 5:281–294, 2005.
70
[203] M. Radetzki and R. S. Khaligh. Accuracy-adaptive Simulation of Transaction Level Models.
In Proceedings of the Conference on Design, Automation and Test in Europe, DATE ’08, pages
788–791, New York, NY, USA, 2008. ACM. ISBN 978-3-9810801-3-1. doi: 10.1145/1403375.
1403566.
[204] Derek Riley, Emeka Eyisi, Jia Bai, Xenofon Koutsoukos, Yuan Xue, and Janos Sztipanovits.
Networked Control System Wind Tunnel (NCSWT): An Evaluation Tool for Networked Multiagent Systems. In Proceedings of the 4th International ICST Conference on Simulation Tools
and Techniques, SIMUTools ’11, pages 9–18, ICST, Brussels, Belgium, Belgium, 2011. ICST
(Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering).
ISBN 978-1-936968-00-8.
[205] Robin Roche, Sudarshan Natarajan, Ayan Bhattacharyya, and Siddharth Suryanarayanan.
A Framework for Co-simulation of AI Tools with Power Systems Analysis Software. In 23rd
International Workshop on Database and Expert Systems Applications, pages 350–354. IEEE,
sep 2012. ISBN 978-1-4673-2621-6. doi: 10.1109/DEXA.2012.9.
[206] J. A. Rowson. Hardware/Software Co-Simulation. In 31st Conference on Design Automation,
pages 439–440, 1994. ISBN 0738-100X VO -. doi: 10.1109/DAC.1994.204143.
[207] Severin Sadjina, Lars T. Kyllingstad, Eilif Pedersen, and Stian Skjong. Energy Conservation
and Power Bonds in Co-Simulations: Non-Iterative Adaptive Step Size Control and Error
Estimation. arXiv preprint arXiv:1602.06434, 2016.
[208] Salah Eddine Saidi, Nicolas Pernet, Yves Sorel, and Abir Ben Khaled. Acceleration of FMU
Co-Simulation On Multi-core Architectures. In The First Japanese Modelica Conferences,
May 23-24, Tokyo, Japan, number 124, pages 106–112. Linköping University Electronic Press,
may 2016. ISBN 1650-3740. doi: 10.3384/ecp16124106.
[209] Resve A Saleh, Shyh-Jye Jou, and A Richard Newton. Mixed-mode simulation and analog multilevel simulation, volume 279. Springer Science & Business Media, 2013. ISBN 1475758545.
[210] Tom Schierz and Martin Arnold. Stabilized overlapping modular time integration of coupled
differential-algebraic equations. Applied Numerical Mathematics, 62(10):1491–1502, oct 2012.
ISSN 01689274. doi: 10.1016/j.apnum.2012.06.020.
[211] Tom Schierz, Martin Arnold, and Christoph Clauss. Co-simulation with communication step
size control in an FMI compatible master algorithm. In 9th Int. Modelica Conf., Munich,
Germany, pages 205–214, nov 2012. doi: 10.3384/ecp12076205.
[212] Robert Schmoll and Bernhard Schweizer. Convergence Study of Explicit Co-Simulation Approaches with Respect to Subsystem Solver Settings. PAMM, 12(1):81–82, dec 2012. ISSN
1617-7061. doi: 10.1002/pamm.201210032.
[213] Stefan-Alexander Schneider, Johannes Frimberger, and Michael Folie. Significant Reduction
of Validation Efforts for Dynamic Light Functions with FMI for Multi-Domain Integration
and Test Platforms. In 10th International Modelica Conference, 2014.
[214] B. Schweizer and D. Lu. Predictor/corrector co-simulation approaches for solver coupling with
algebraic constraints. ZAMM - Journal of Applied Mathematics and Mechanics / Zeitschrift
71
für Angewandte Mathematik und Mechanik, 95(9):911–938, sep 2015. ISSN 00442267. doi:
10.1002/zamm.201300191.
[215] Bernhard Schweizer and Daixing Lu. Semi-implicit co-simulation approach for solver coupling.
Archive of Applied Mechanics, 84(12):1739–1769, dec 2014. ISSN 0939-1533. doi: 10.1007/
s00419-014-0883-5.
[216] Bernhard Schweizer and Daixing Lu. Stabilized index-2 co-simulation approach for solver
coupling with algebraic constraints. Multibody System Dynamics, 34(2):129–161, jun 2015.
ISSN 1384-5640. doi: 10.1007/s11044-014-9422-y.
[217] Bernhard Schweizer, Pu Li, and Daixing Lu. Explicit and Implicit Cosimulation Methods:
Stability and Convergence Analysis for Different Solver Coupling Approaches. Journal of
Computational and Nonlinear Dynamics, 10(5):051007, sep 2015. ISSN 1555-1415. doi: 10.
1115/1.4028503.
[218] Bernhard Schweizer, Pu Li, Daixing Lu, and Tobias Meyer. Stabilized implicit co-simulation
methods: solver coupling based on constitutive laws. Archive of Applied Mechanics, 85(11):
1559–1594, nov 2015. ISSN 0939-1533. doi: 10.1007/s00419-015-0999-2.
[219] Bernhard Schweizer, Daixing Lu, and Pu Li. Co-simulation method for solver coupling with
algebraic constraints incorporating relaxation techniques. Multibody System Dynamics, 36(1):
1–36, jan 2016. ISSN 1384-5640. doi: 10.1007/s11044-015-9464-9.
[220] S. Sicklinger, V. Belsky, B. Engelmann, H. Elmqvist, H. Olsson, R. Wüchner, and K. U.
Bletzinger. Interface Jacobian-based Co-Simulation. International Journal for Numerical
Methods in Engineering, 98(6):418–444, may 2014. ISSN 1097-0207. doi: 10.1002/nme.4637.
[221] G. Stettinger, M. Horn, M. Benedikt, and J. Zehetner. Model-based coupling approach for
non-iterative real-time co-simulation. In 2014 European Control Conference (ECC), pages
2084–2089, 2014. doi: 10.1109/ECC.2014.6862242.
[222] Georg Stettinger, Josef Zehetner, Martin Benedikt, and Norbert Thek. Extending CoSimulation to the Real-Time Domain. apr 2013. doi: 10.4271/2013-01-0421.
[223] H T Su, H. T. Su, K W Chan, K. W. Chan, L A Snider, and L. A. Snider. Parallel interaction
protocol for electromagnetic and electromechanical hybrid simulation, 2005.
[224] Yongqi Sun, Stephanie Vogel, and Haiko Steuer. Combining Advantages of Specialized Simulation Tools and Modelica Models using Functional Mock-up Interface (FMI). In Proceedings
of the 8th International Modelica Conference, 2011.
[225] Robert Tarjan. Depth-first search and linear graph algorithms. 12th Annual Symposium on
Switching and Automata Theory (swat 1971), 1(2), jun 1971. ISSN 0272-4847. doi: 10.1109/
SWAT.1971.10.
[226] Jean-Philippe Tavella, Mathieu Caujolle, Charles Tan, Gilles Plessis, Mathieu Schumann,
Stéphane Vialle, Cherifa Dad, Arnaud Cuccuru, and Sébastien Revol. Toward an Hybrid
Co-simulation with the FMI-CS Standard, apr 2016. URL https://hal-centralesupelec.
archives-ouvertes.fr/hal-01301183.
72
[227] Brook Taylor. Methodus Incrementorum Directa et Inversa. London, 1715.
[228] T Tomiyama, V. D’Amelio, J Urbanic, and W ElMaraghy. Complexity of Multi-Disciplinary
Design. CIRP Annals - Manufacturing Technology, 56(1):185–188, 2007. ISSN 00078506. doi:
10.1016/j.cirp.2007.05.044.
[229] Pawel Tomulik and Janusz Fra̧czek. Simulation of multibody systems with the use of coupling
techniques: a case study. Multibody System Dynamics, 25(2):145–165, feb 2011. ISSN 13845640. doi: 10.1007/s11044-010-9206-y.
[230] Mamadou K Traoré and Alexandre Muzy. Capturing the dual relationship between simulation
models and their context. Simulation Modelling Practice and Theory, 14(2):126–142, feb 2006.
ISSN 1569190X. doi: 10.1016/j.simpat.2005.03.002.
[231] Marija Trcka, Michael Wetter, and Jan Hensen. Comparison of co-simulation approaches
for building and HVAC/R system simulation. In Proceedings of the International IBPSA
Conference, Beijing, China, 2007.
[232] Stavros Tripakis. Bridging the semantic gap between heterogeneous modeling formalisms
and FMI. In 2015 International Conference on Embedded Computer Systems: Architectures,
Modeling, and Simulation (SAMOS), pages 60–69. IEEE, jul 2015. ISBN 978-1-4673-7311-1.
doi: 10.1109/SAMOS.2015.7363660. URL http://ieeexplore.ieee.org/lpdocs/epic03/
wrapper.htm?arnumber=7363660.
[233] F. Tseng and G. Hulbert. Network-distributed multibody dynamics simulation—gluing algorithm. Advances in Computational Multibody Dynamics, pages 521–540, 1999.
[234] A. M. Uhrmacher. Dynamic Structures in Modeling and Simulation: A Reflective Approach.
ACM Trans. Model. Comput. Simul., 11(2):206–232, apr 2001. ISSN 1049-3301. doi: 10.1145/
384169.384173.
[235] Adelinde M. Uhrmacher. Variable structure models: autonomy and control answers from two
different modeling approaches. In 4th Annual Conference on AI, Simulation and Planning in
High Autonomy Systems, pages 133–139. IEEE Comput. Soc. Press, 1993. ISBN 0-8186-40200. doi: 10.1109/AIHAS.1993.410588.
[236] Bert Van Acker, Joachim Denil, Paul De Meulenaere, Hans Vangheluwe, Bert Vanacker, and
Paul Demeulenaere. Generation of an Optimised Master Algorithm for FMI Co-simulation. In
Proceedings of the Symposium on Theory of Modeling & Simulation-DEVS Integrative, pages
946–953. Society for Computer Simulation International, 2015.
[237] Herman Van der Auweraer, Jan Anthonis, Stijn De Bruyne, and Jan Leuridan. Virtual
engineering at work: the challenges for designing mechatronic products. Engineering with
Computers, 29(3):389–408, 2013. ISSN 0177-0667. doi: 10.1007/s00366-012-0286-6.
[238] Arjan J. Van Der Schaft and Johannes Maria Schumacher. An introduction to hybrid dynamical systems, volume 251. Springer London, 2000.
[239] Simon Van Mierlo. Explicitly Modelling Model Debugging Environments. In ACM Student
Research Competition at MODELS 2015, pages 24–29. CEUR, 2015.
73
[240] Yentl Van Tendeloo and Hans Vangheluwe. Activity in PythonPDEVS. In Proceedings of
ACTIMS 2014, 2014.
[241] Yentl Van Tendeloo and Hans Vangheluwe. PythonPDEVS: a distributed Parallel DEVS
simulator. In Proceedings of the 2015 Spring Simulation Multiconference, SpringSim ’15,
pages 844–851. Society for Computer Simulation International, 2015.
[242] Yentl Van Tendeloo and Hans Vangheluwe. An Introduction to Classic DEVS. Technical
report, 2017. URL https://arxiv.org/pdf/1701.07697v1.pdf.
[243] Luigi Vanfretti, Tetiana Bogodorova, and Maxime Baudette. Power system model identification exploiting the Modelica language and FMI technologies. In 2014 IEEE International
Conference on Intelligent Energy and Power Systems (IEPS), pages 127–132. IEEE, jun 2014.
ISBN 978-1-4799-2266-6. doi: 10.1109/IEPS.2014.6874164.
[244] Hans Vangheluwe. DEVS as a common denominator for multi-formalism hybrid systems
modelling. In CACSD. Conference Proceedings. IEEE International Symposium on ComputerAided Control System Design (Cat. No.00TH8537), pages 129–134. IEEE, 2000. ISBN 0-78036566-6. doi: 10.1109/CACSD.2000.900199.
[245] Hans Vangheluwe. Foundations of Modelling and Simulation of Complex Systems. EASST,
10, 2008. doi: 10.14279/tuj.eceasst.10.162.148.
[246] Hans Vangheluwe, Juan De Lara, and Pieter J. Mosterman. An introduction to multiparadigm modelling and simulation. In Proceedings of AIS2002 (AI, Simulation & Planning),
pages 9–20. SCS, 2002.
[247] Ken Vanherpen, Joachim Denil, Hans Vangheluwe, and Paul De Meulenaere. Model Transformations for Round-trip Engineering in Control Deployment Co-Design. In Theory of Modeling
and Simulation - DEVS, TMS/DEVS ’15, part of the Spring Simulation Multi-Conference,
pages 820–827, Alexandria, Virginia, USA, apr 2015. Society for Computer Simulation International.
[248] Julien Vaubourg, Yannick Presse, Benjamin Camus, Christine Bourjot, Laurent Ciarletta,
Vincent Chevrier, Jean-Philippe Tavella, and Hugo Morais. Multi-agent Multi-Model Simulation of Smart Grids in the MS4SG Project. pages 240–251. Springer International Publishing,
Cham, 2015. ISBN 978-3-319-18944-4. doi: 10.1007/978-3-319-18944-4 20.
[249] A Verhoeven, B Tasic, T G J Beelen, E J W ter Maten, and R M M Mattheij. BDF compoundfast multirate transient analysis with adaptive stepsize control. J. Numer. Anal. Ind. Appl.
Math, 3(3-4):275–297, 2008.
[250] Antoine Viel. Implementing stabilized co-simulation of strongly coupled systems using the
Functional Mock-up Interface 2.0. 10th International Modelica Conference, 2014.
[251] B. Wang and J. S. Baras. HybridSim: A Modeling and Co-simulation Toolchain for Cyberphysical Systems. In Distributed Simulation and Real Time Applications (DS-RT), 2013
IEEE/ACM 17th International Symposium on, pages 33–40, 2013. ISBN 1550-6525. doi:
10.1109/DS-RT.2013.12.
74
[252] G. Wanner and E. Hairer. Solving ordinary differential equations I, volume 1. Springer-Verlag,
Berlin, 1991.
[253] Michael Wetter. Co-simulation of building energy and control systems with the Building
Controls Virtual Test Bed. Journal of Building Performance Simulation, 4(3):185–203, nov
2010. ISSN 1940-1493. doi: 10.1080/19401493.2010.518631.
[254] E. Widl, W. Muller, A. Elsheikh, M. Hortenhuber, and P. Palensky. The FMI++ library:
A high-level utility package for FMI for model exchange. In Modeling and Simulation of
Cyber-Physical Energy Systems (MSCPES), 2013 Workshop on, pages 1–6, 2013. doi: 10.
1109/MSCPES.2013.6623316.
[255] Xiaorong Xie, Chuanyu Zhang, Huakun Liu, Chao Liu, Dongxiang Jiang, and Baorong Zhou.
Continuous-Mass-Model-Based Mechanical and Electrical Co-Simulation of SSR and Its Application to a Practical Shaft Failure Event. IEEE Transactions on Power Systems, 31(6):
5172–5180, nov 2016. ISSN 0885-8950. doi: 10.1109/TPWRS.2016.2537001.
[256] Masahiro Yamaura, Nikos Arechiga, Shinichi Shiraishi, Scott Eisele, Joseph Hite2 Sandeep
Neema2 Jason Scott, and Theodore Bapty. ADAS Virtual Prototyping using Modelica and
Unity Co-simulation via OpenMETA. 2016.
[257] Faruk Yılmaz, Umut Durak, Koray Taylan, and Halit Oğuztüzün. Adapting Functional
Mockup Units for HLA-compliant Distributed Simulation. In 10th International Modelica
Conference, 2014.
[258] Josef Zehetner, Di Wenpu Lu, and Daniel Watzenig. Design of modern vehicle electrical
systems based on co-simulation and a model library. ATZelektronik worldwide, 8(4):20–24,
2013. doi: 10.1365/s38314-013-0182-x.
[259] Bernard P. Zeigler. Theory of modelling and simulation. New York, Wiley, 1976. ISBN
0471981524.
[260] Bernard P. Zeigler. Embedding dev&dess in devs. In Proc. DEVS Integrative M&S Symp,
volume 7, 2006.
[261] Bernard P. Zeigler and J. S. Lee. Theory of quantized systems: formal basis for DEVS/HLA
distributed simulation environment. volume 3369, pages 49–58, aug 1998. doi: 10.1117/12.
319354.
[262] Bernard P. Zeigler, Herbert Praehofer, and Tag Gon Kim. Theory of modeling and simulation:
integrating discrete event and continuous complex dynamic systems. Academic press, 2 edition,
2000.
[263] Fu Zhang, Murali Yeddanapudi, and Pieter Mosterman. Zero-crossing location and detection
algorithms for hybrid system simulation. In IFAC World Congress, pages 7967–7972, 2008.
[264] Saipeng Zhang, Jun Liu, and Wei Su. A Co-simulation Platform of Integrated Starter/Generator System Based on ANSYS, pages 35–43. Springer Berlin Heidelberg, Berlin, Heidelberg,
2016. ISBN 978-3-662-49367-0. doi: 10.1007/978-3-662-49367-0 5.
75
[265] X. Zhang and J. F. Broenink. A structuring mechanism for embedded control systems using co-modelling and co-simulation. In Proceedings of the 2nd International Conference on
Simulation and Modeling Methodologies, Technologies and Applications, SIMULTECH 2012,
Rome, Italy, pages 131–136, Portugal, 2012. SciTePress.
[266] Zhenkai Zhang, Emeka Eyisi, Xenofon Koutsoukos, Joseph Porter, Gabor Karsai, and Janos
Sztipanovits. A co-simulation framework for design of time-triggered automotive cyber physical systems. Simulation Modelling Practice and Theory, 43:16–33, 2014. ISSN 1569190X. doi:
10.1016/j.simpat.2014.01.001.
[267] Chenguang Zhao, Hongman Yan, Dong Liu, Hong Zhu, Yun Wang, and Yunhui Chen. Cosimulation research and application for Active Distribution Network based on Ptolemy II and
Simulink. In 2014 China International Conference on Electricity Distribution (CICED), pages
1230–1235. IEEE, sep 2014. ISBN 978-1-4799-4126-1. doi: 10.1109/CICED.2014.6991903.
[268] Vojin Živojnovic and Heinrich Meyr. Compiled HW/SW Co-simulation. In Proceedings of the
33rd Annual Design Automation Conference, DAC ’96, pages 690–695, New York, NY, USA,
1996. ACM. ISBN 0-89791-779-0. doi: 10.1145/240518.240649.
[269] M. Zwolinski, C. Garagate, Z. Mrcarica, T. J. Kazmierski, and A. D. Brown. Anatomy of
a simulation backplane. IEE Proceedings - Computers and Digital Techniques, 142(6):377–
385(8), nov 1995. ISSN 1350-2387.
76
Table 1: Historical Perspective of Co-simulation.
Time
Concept
Description
<80s
Single Formalism
The equations describing dynamic behavior are integrated
together.
80s
Dynamic Iteration
Large circuits are decomposed into coupled constituent systems and dynamic iteration techniques are used [98, 155,
167, 170, 171, 183].
90s
Multi-Formalism
Software and Hardware are developed and simulated concurrently [65, 110, 206, 268] at multiple levels of abstraction [77, 117, 118]. Orchestration methods are explored in
Carothers et al. [58], Fey et al. [84], Frey et al. [91], Tseng
and Hulbert [233].
Late 90s and
Early 2000s
Standard Interfaces
Recognized as key for co-simulation [116, 144, 196, 223, 269]
2010s
IP
Protection,
X-in-the-loop, and
Scale
Important to enhance industrial applicability of cosimulation [7, 18–20, 74, 96, 237].
A
Historical Perspective of Co-simulation
This section provides an historical perspective that relates the major concepts in co-simulation to
the time at which they are recognized in the studied state of the art, summarized in table 1.
A.1
One Formalism and Dynamic Iteration
Traditionally, the equations describing the dynamical behavior of large circuits were integrated together. These systems are sparsely coupled, reflecting the connections of the corresponding circuits,
and many techniques were developed that take advantage of this structure [167].
The crucial idea that improved the simulation speed in up to two orders of magnitude is to
decompose the large system into a set of coupled constituent systems and integrate them independently.
The decomposition of the circuit implies the definition of inputs and outputs for each of the
resulting constituent systems. The coupling is then the assignment of outputs to inputs.
For a subsystem Si , we call the subsystems, whose outputs are assigned to any of the inputs of
Si , for neighbor subsystems.
The essence of the dynamic iteration approach is to integrate each subsystem independently, for
a period of time Tn → Tn+1 , using the extrapolated outputs of the neighbor subsystems as inputs
[155, 170, 171, 183].
Naturally, the fact that outputs are extrapolated introduces inaccuracy in the solution of the
77
subsystem, so the integration can be repeated for the same period of time, with corrected outputs,
until some form of convergence criteria is met [121]. The extrapolated outputs of a subsystem Sj
can be corrected by collecting the outputs during the integration of Sj .
It is easy to see that this approach only requires communication between constituent systems
at times Tn and Tn+1 and that the integration of each subsystem can be done independently and in
parallel [122], using any numerical method with any step size control policy. The signals exchanged
are functions in the interval [Tn , Tn+1 ].
The advantages of independent step size control policy become evident when one observes that
many circuits have components that change at different rates. If the whole system were to be
simulated, the simulation unit would have to use the smallest time step that ensures sufficient
accuracy for the fastest changing component, which would be a huge waste of computational effort
for the slow components. This is the similarity to multi-rate numerical methods [98].
To the best of our knowledge, dynamic iteration techniques and multi-rate numerical are the
first to resemble co-simulation. The coordination software that implements these techniques expect
any number of subsystems but assumes that the subsystems are all specific in the same formalism:
differential equations.
A.2
Two Formalisms: Digital and Analog Co-simulation
Co-simulation, in its modern definition, was applied to enable the virtual development of coupled
software and hardware systems [65, 110, 206, 268]. In this application domain, co-simulation decreases the need to build prototype board circuits to validate the composition of the software and
the hardware part. It enables software and hardware to be developed and validated concurrently.
To the best of our knowledge, this was one of the first uses of co-simulation in the modern sense. The
co-simulation frameworks developed in this application domain typically assumed two simulation
units and two formalisms.
The hardware/software systems quickly became more complex and a new idea was introduced:
use multiple models at different levels of abstraction of each subsystem. Simulations could be made
arbitrarily faster in some intervals by solving the more abstract models, and arbitrarily accurate
in other intervals, by solving the more detailed ones [77, 135, 166, 209]. In the particular case of
analog-digital co-simulation, each level of abstraction was solved by a different tool: a continuous
time tool and a discrete event tool. The separation into continuous time and discrete event made
the abstract synchronization problem and synchronization methods between simulation units in
these two domains were developed [58, 84, 91, 233]. We could call these some of the first master
algorithms.
A.3
Multi-abstraction/Multi-Formalism Co-simulation
The heterogeneity aspect of co-simulation comes into play at this time: multiple formalisms can be
used to describe the same subsystem at multiple levels of abstraction: state machines can describe a
rough approximation of the modes, while differential equations can describe the detailed dynamics of
the electronic circuit. Depending on the purpose of the co-simulation, a subsystem and its neighbors
can be solved in detail, whereas subsystems that are “farther away” can be simulated with higher
levels of abstraction [117, 118]. For the domain of Hw/sw co-simulation, RTL and TLM classify
the multiple abstraction levels of models [27, 203] and switching between these multiple levels of
abstraction have been studied in [131].
78
As the number and heterogeneity of simulation tools to be coupled increases, the need to provide
a common interface to couple any number of tools is recognized in Hickey et al. [116], Kuhl et al.
[144], Petrellis et al. [196], Zwolinski et al. [269] and later in Blochwitz et al. [34].
In parallel with the previous advancements, co-simulation has also been in use for heterogeneous
physical systems, such as automotive [105, 142, 151, 213], railway [12, 71] and HVAC [107, 231],
to name just a few. The common motivation is the fact that co-simulation enables specialized
simulation units to cooperatively simulated the system, with huge savings in time and cost, when
compared to a monolithic modeling approach.
A.4
Black-box Co-simulation
Later, distributed and concurrent development processes, enabled by co-simulation, are studied and
IP protection is identified as a desired characteristic [7, 237] to enable suppliers and integrators to
exchange co-simulation units without having to disclose sensitive information and avoiding vendor
lock-in contracts.
A.5
Real-time Co-simulation
Furthermore, co-simulation is used at every stage of the development process, from early system validation, to X-in-the-Loop co-simulation, bringing hard real-time constraints to the set of challenges
[74].
A.6
Many simulation units: Large Scale Co-simulation
More recently, with the acknowledgment that there is a need to be able to simulate even larger
systems of systems, scale and distribution become inherent challenges in co-simulation [18–20, 96].
B
State of the Art in Co-simulation Frameworks
This section provides the detailed classification of each reference.
Ref 1. Pedersen2015: Co-Simulation of Distributed Engine Control System and Network Model
using FMI & SCNSL [194].
Summary. This work describes a co-simulation master in the context of the maritime industry.
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr : results visualization :post mortem
fr : alg loop : explicit
fr : sim step size : fixed
fr :sim rate: single
fr :domain:ct
fr :coupling model:io assignments
fr :standard:fmi
fr :communication model:jacobi
79
fr :num sim:three more
Ref 2. Lin2011: Power system and communication network co-simulation for smart grid applications [120].
Summary. This work describes a co-simulation between power system and network simulator.
sr : info : predict step sizes
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr : results visualization :post mortem
fr : alg loop : explicit
fr :sim rate: single
fr :coupling model:io assignments
fr :num sim:two
fr :domain:de
fr : sim step size : variable
fr :communication model:gauss seidel
Ref 3. Hoepfer2011: Towards a Comprehensive Framework for Co- Simulation of Dynamic
Models With an Emphasis on Time Stepping [119].
Summary. This work describes a co-simulation approach that finds an appropriate co-simulation
step size.
nfr :performance
nfr :accuracy
nfr : ip protection
sr : availability : local
sr : info : derivatives :out
sr : info : derivatives : state
sr : info : statevars
sr : causality : causal
sr : rollback :none
sr : rel time : analytic
fr :num sim:three more
fr :domain:ct
fr :sim rate: single
fr : sim step size : variable
fr :communication model:jacobi
fr :communication model:gauss seidel
fr : alg loop : explicit
fr : results visualization :post mortem
80
Ref 4. Faure2011: Methods for real-time simulation of Cyber-Physical Systems: application to
automotive domain [82].
Summary. This work addresses co-simulation with real-time simulators.
nfr :performance
nfr : parallelism
sr : availability : local
sr : rel time : fixed real scaled time simulation
sr : info :wcet
sr : rollback :none
fr : results visualization :post mortem
fr :num sim:three more
fr :coupling model:io assignments
fr :domain:ct
fr :communication model:jacobi
fr : alg loop : explicit
fr :sim rate: single
fr : sim step size : fixed
Ref 5. Tomulik2011: Simulation of multibody systems with the use of coupling techniques: a
case study [229].
Summary. This work discusses a co-simulation method for couplings with algebraic constraints. One
of the results is that this kind of coupling should be done with many derivatives of the coupling
variables.
nfr :accuracy
sr : rollback : single
sr : availability : local
sr : rel time : analytic
sr : causality : causal
sr : info : derivatives :out
sr : info :jacobian:out
fr : results visualization :post mortem
fr :communication model:jacobi
fr : alg loop : implicit
fr :coupling model:algebraic constraints
fr :domain:ct
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
Ref 6. Sun2011: Combining Advantages of Specialized Simulation Tools and Modelica Models
using Functional Mock-up Interface (FMI) [224].
81
Summary. This work describes the application of co-simulation to the power production domain.
nfr : ip protection
nfr :performance
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr : results visualization :post mortem
fr :sim rate: single
fr :coupling model:io assignments
fr :standard:fmi
fr :domain:ct
fr :communication model:gauss seidel
fr :num sim:two
fr : alg loop : explicit
fr : sim step size : variable
Ref 7. Bastian2011a: Master for Co-Simulation Using FMI [25].
Summary. This work describes a co-simulation approach.
nfr : ip protection
nfr :platform independence
nfr : parallelism
sr : causality : causal
sr : rollback :none
sr : info : stateserial
sr : info :jacobian:out
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr : sim step size : fixed
fr :sim rate: single
fr : alg loop : implicit
fr :domain:ct
fr :num sim:three more
fr :standard:fmi
fr :communication model:jacobi
fr :communication model:gauss seidel
Ref 8. Friedrich2011: Parallel Co-Simulation for Mechatronic Systems [93].
Summary. This work describes a co-simulation framework based on the Jacobi iteration scheme.
nfr : parallelism
nfr :performance
nfr : distribution
82
nfr : ip protection
sr : availability :remote
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
fr : results visualization :post mortem
fr : alg loop : explicit
fr :domain:ct
fr :coupling model:io assignments
fr :coupling model:algebraic constraints
fr :num sim:three more
fr :communication model:jacobi
fr :sim rate: single
fr : sim step size : fixed
Ref 9. Gonzalez2011: On the effect of multirate co-simulation techniques in the efficiency and
accuracy of multibody system dynamics [103].
Summary. This work deals with multi-rate co-simulation. Essentially, one of the simulators (the
fast one) drives the simulation, while the slow one provides extrapolated inputs, to avoid excessive
computation.
nfr :accuracy
nfr :performance
sr : causality : causal
sr : rollback :none
sr : availability : local
sr : rel time : analytic
fr : results visualization :post mortem
fr : alg loop : explicit
fr :communication model:gauss seidel
fr :coupling model:io assignments
fr :num sim:two
fr :domain:ct
fr :sim rate:multi
fr : sim step size : fixed
Ref 10. Nutaro2011: Designing power system simulators for the smart grid: Combining controls,
communications, and electro-mechanical dynamics [189].
Summary. This work describes a tool that is formed by the coupling of a DEVS simulator with
some other modules that wrap CT as DEVS simulators.
nfr : distribution
nfr :accuracy
sr : causality : causal
sr : rollback : single
83
sr : rel time : analytic
sr : availability : local
fr : alg loop : explicit
fr :domain:de
fr :communication model:gauss seidel
fr :num sim:two
fr :sim rate: single
fr : sim step size : variable
fr :coupling model:io assignments
fr : results visualization :post mortem
Ref 11. Busch2012: Asynchronous method for the coupled simulation of mechatronic systems [54].
Summary. This work describes co-simulation approaches between two simulation tools. The main
contribution is a semi-implicit method that applies a correction based on the jacobian of the subsystem’s coupling variables.
nfr :accuracy
nfr : distribution
sr : causality : causal
sr : rollback : single
sr : availability :remote
sr : info :jacobian:out
sr : rel time : analytic
fr :sim rate: single
fr : sim step size : fixed
fr : results visualization :post mortem
fr :communication model:gauss seidel
fr : alg loop : semi implicit
fr : alg loop : explicit
fr : alg loop : implicit
fr :coupling model:io assignments
fr :domain:ct
fr :num sim:two
Ref 12. Pohlmann2012: Generating functional mockup units from software specifications [199].
Summary. This work describes an application of co-simulation to robotics.
Ref 13. Schmoll2012: Convergence Study of Explicit Co-Simulation Approaches with Respect
to Subsystem Solver Settings [212].
Summary. This paper describes global error analysis for co-simulation, that takes into account
sub-system solvers (instead of analytical solvers, as more commonly done).
nfr :accuracy
sr : rollback :none
sr : info : full model
84
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr : alg loop : explicit
fr :domain:ct
fr :num sim:two
fr :coupling model:io assignments
fr :communication model:jacobi
fr :sim rate: single
fr : sim step size : fixed
fr : results visualization :post mortem
Ref 14. Ni2012: Hybrid systems modelling and simulation in DESTECS: a co-simulation approach [184].
Summary. This work present a coupling of the tools Crescendo and 20-sim.
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :communication model:gauss seidel
fr : results visualization :post mortem
fr :coupling model:io assignments
fr : alg loop : explicit
fr :domain:ct
fr :sim rate: single
fr : sim step size : fixed
fr :num sim:two
Ref 15. Hassairi2012: Matlab/SystemC for the New Co-Simulation Environment by JPEG Algorithm [115].
Summary. This work introduces guidelines for the implementation of co-simulation between Matlab
and SystemC. The case study is the JPEG Algorithm.
sr : info : full model
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :coupling model:io assignments
fr : sim step size : fixed
fr :sim rate: single
fr : alg loop : explicit
fr :communication model:gauss seidel
fr : results visualization : live
85
fr :num sim:two
Ref 16. Schierz2012: Stabilized overlapping modular time integration of coupled differentialalgebraic equations [210].
Summary. This work discusses co-simulation techniques for simulators coupled via algebraic constraints.
nfr :accuracy
sr : availability : local
sr : rel time : analytic
sr : causality : causal
sr : rollback :none
sr : info : full model
sr : info :jacobian:out
fr : results visualization :post mortem
fr :sim rate: single
fr :num sim:three more
fr : alg loop : explicit
fr : sim step size : fixed
fr :domain:ct
fr :coupling model:algebraic constraints
fr :communication model:gauss seidel
fr :communication model:jacobi
Ref 17. Gunther2012: A Modular Technique for Automotive System Simulation [108].
Summary. This work describes the MDPCosim framework.
nfr :performance
The decomposition of the system for co-simulation is done for performance reasons.
nfr : parallelism
nfr :accuracy
sr : availability : local
IPC communication is used.
sr : causality : causal
sr : info : derivatives :out
sr : rollback :none
sr : info : predict step sizes
fr : results visualization :post mortem
fr : alg loop : explicit
fr :coupling model:io assignments
fr :domain:ct
fr :communication model:jacobi
fr : sim step size : variable
The step size control approach is based on looking at the derivatives.
fr :sim rate: single
86
fr :num sim:three more
Ref 18. Quaglia2012: A SystemC/Matlab co-simulation tool for networked control systems [201].
Summary. Work describing another tool coupling.
nfr : distribution
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
fr : results visualization :post mortem
fr : alg loop : explicit
fr :coupling model:io assignments
fr :domain:ct
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
fr :communication model:gauss seidel
Ref 19. Al-Hammouri2012: A comprehensive co-simulation platform for cyber-physical systems [5].
Summary. The work describes the integration of two tools: Modelica, and NS-2.
sr : causality : causal
sr : rel time : analytic
sr : availability : local
Communication is done over named pipes.
sr : rollback :none
fr : results visualization :post mortem
fr :num sim:two
fr :coupling model:io assignments
fr : alg loop : explicit
fr :domain:ct
fr :communication model:gauss seidel
fr :domain:de
fr :sim rate: single
fr : sim step size : variable
Ref 20. Eyisi2012: NCSWT: An integrated modeling and simulation tool for networked control
systems [81].
Summary. This work describes the coupling of two tools: Matlab and NS-2. The coupling is done
through HLA standard. The preliminary version of the tool is described in [204].
nfr :platform independence
nfr :performance
87
nfr : distribution
sr : info : predict step sizes
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
fr : alg loop : explicit
fr :num sim:two
fr :coupling model:io assignments
fr :domain:de
fr :standard:hla
fr :communication model:gauss seidel
fr :sim rate: single
fr : sim step size : variable
fr : results visualization : live
Ref 21. Riley2011: Networked Control System Wind Tunnel (NCSWT): An Evaluation Tool for
Networked Multi-agent Systems [204].
Summary. This work describes the coupling of two tools: Matlab and NS-2. The coupling is done
through HLA standard.
nfr :platform independence
nfr :performance
nfr : distribution
sr : info : predict step sizes
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
fr : alg loop : explicit
fr :num sim:two
fr :coupling model:io assignments
fr :domain:de
fr :standard:hla
fr :communication model:gauss seidel
fr :sim rate: single
fr : sim step size : variable
fr : results visualization : live
Ref 22. Roche2012: A Framework for Co-simulation of AI Tools with Power Systems Analysis
Software [205].
Summary. This work describes a co-simulation between two tools in the power grid domain with
matlab running the co-simulation.
nfr : distribution
88
nfr :open source
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
fr :coupling model:io assignments
fr :domain:ct
fr :communication model:gauss seidel
fr : alg loop : explicit
fr : results visualization :post mortem
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
Ref 23. Fitzgerald2010: Collaborative Modelling and Co-simulation in the Development of
Dependable Embedded Systems [86].
Summary. This work describes the coupling between two tools: Overture and 20-sim.
nfr :accuracy
nfr : distribution
nfr :platform independence
sr : availability :remote
sr : causality : causal
sr : rel time : analytic
fr : sim step size : variable
fr :sim rate: single
fr :domain:de
fr :domain:ct
fr :num sim:two
fr :coupling model:io assignments
fr : alg loop : explicit
fr : results visualization : live
fr :communication model:gauss seidel
Ref 24. Fitzgerald2013: A formal approach to collaborative modelling and co-simulation for
embedded systems [89].
Summary. This work describes the coupling between two tools: Overture and 20-sim; already
described in [86].
nfr :accuracy
nfr : distribution
nfr :platform independence
sr : availability :remote
sr : causality : causal
sr : rel time : analytic
89
fr : sim step size : variable
fr :sim rate: single
fr :domain:de
fr :domain:ct
fr :num sim:two
fr :coupling model:io assignments
fr : alg loop : explicit
fr : results visualization : live
fr :communication model:gauss seidel
Ref 25. Kudelski2013: RoboNetSim: An integrated framework for multi-robot and network
simulation [143].
Summary. This work describes the integration of three simulators (ARGoS, NS-2 and NS-3) that
can be used in co-simulation scenarios with two simulators.
nfr : distribution
nfr :platform independence
sr : rollback :none
sr : causality : causal
sr : rel time : analytic
sr : availability :remote
fr : results visualization :post mortem
fr : alg loop : explicit
fr :communication model:jacobi
fr :domain:de
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
fr :coupling model:io assignments
Ref 26. Broman2013: Determinate Composition of FMUs for Co-simulation [46].
Summary. This work describes a master algorithm that ensures a determinate execution.
nfr : ip protection
sr : availability : local
sr : rollback :none
sr : rel time : analytic
sr : causality : causal
sr : info : causality :feedthrough
sr : info : predict step sizes
sr : info : stateserial
fr :coupling model:io assignments
fr :standard:fmi
fr : sim step size : variable
fr :domain:ct
90
fr :communication model:jacobi
fr :num sim:three more
fr :sim rate: single
fr : alg loop : explicit
fr : results visualization :post mortem
Ref 27. Benedikt2013: Guidelines for the Application of a Coupling Method for Non-iterative
Co-simulation [30].
Summary. This work describes a co-simulation approach where energy information about the signals
is used, and those errors are compensated in a corrector step.
nfr :accuracy
nfr :performance
sr : rollback :none
sr : rel time : analytic
sr : causality : causal
sr : info :record outputs
sr : availability : local
fr : results visualization :post mortem
fr : alg loop : explicit
fr :num sim:two
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr :domain:ct
fr :sim rate: single
fr : sim step size : fixed
Ref 28. Benedikt2013b: Macro-step-size selection and monitoring of the coupoling error for
weak coupled subsystems in the frequency-domain [29].
Summary. The work describes a method for finding appropriate communication step sizes in cosimulations between LTI systems. Essentially, it provides rules of thumb to chose a communication
step size based on the maximum instantaneous frequency of components.
nfr :accuracy
sr : availability : local
sr : rollback :none
sr : rel time : analytic
sr : causality : causal
sr : info :frequency outputs
fr : results visualization :post mortem
fr :sim rate: single
fr :coupling model:io assignments
fr : alg loop : explicit
fr :communication model:gauss seidel
fr :domain:ct
91
fr : sim step size : fixed
fr :num sim:two
Ref 29. Fuller2013: Communication simulations for power system applications [95].
Summary. This work describes a co-simulation between two co-simulation tools (ns-3 and GridLABD) for smart grid development.
nfr : scalability
nfr : faulttolerance
nfr : ip protection
nfr : distribution
The tools keeps track of messages in transit.
sr : causality : causal
sr : rel time : analytic
sr : availability :remote
sr : rollback :none
fr :coupling model:io assignments
fr : alg loop : explicit
fr :domain:de
fr :num sim:two
fr :sim rate: single
fr : sim step size : variable
fr : results visualization :post mortem
fr :communication model:gauss seidel
Ref 30. Bombino2013: A model-driven co-simulation environment for heterogeneous systems [39].
Summary. This work describes the coupling between two simulation tools.
nfr : distribution
sr : rollback :none
sr : causality : causal
sr : rollback : single
sr : rel time : dy real scaled time simulation
sr : availability :remote
fr : alg loop : explicit
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr :domain:ct
fr :num sim:two
fr : results visualization : interactive live
fr :sim rate: single
fr : sim step size : fixed
Ref 31. Wang2013: HybridSim: A Modeling and Co-simulation Toolchain for Cyber-physical
Systems [251].
92
Summary. The approach described in this reference allows to arrange and process co-simulation
units, Modelica models and TinyOS applications. SysML is used to configure the co-simulation
master. The coordination of simulators is done through the FMI standard.
nfr : ip protection
nfr : config reusability
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :coupling model:io assignments
fr : alg loop : explicit
fr :communication model:gauss seidel
fr :num sim:two
fr :domain:ct
fr :standard:fmi
fr :sim rate: single
fr : sim step size : variable
fr : results visualization :post mortem
Ref 32. Hafner2013: An Investigation on Loose Coupling Co-Simulation with the BCVTB [114].
Summary. This work discusses the consistency and stability of the Jacobi and Gauss-Seidel cosimulation methods. Later, it presents a case study in HVAC systems.
nfr :accuracy
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :coupling model:io assignments
fr : results visualization :post mortem
fr :communication model:jacobi
fr :communication model:gauss seidel
fr : alg loop : explicit
fr :domain:ct
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
Ref 33. Zhao2014: Co-simulation research and application for Active Distribution Network based
on Ptolemy II and Simulink [267].
Summary. This work describes the co-simulation between Ptolemy II and Simulink.
nfr : distribution
sr : causality : causal
sr : rel time : analytic
93
sr : rollback :none
sr : availability :remote
fr :coupling model:io assignments
fr :domain:ct
fr :communication model:gauss seidel
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
fr : results visualization :post mortem
Ref 34. Li2011c: VPNET: A co-simulation framework for analyzing communication channel
effects on power systems [157].
Summary. This work describes the coupling of two simulation tools (VTB and OPNET) to achieve
co-simulation.
sr : availability : local
sr : rollback :none
sr : causality : causal
sr : rel time : analytic
fr : results visualization :post mortem
fr : alg loop : explicit
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr :domain:ct
The coordination is a sample discrete time system.
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
Ref 35. Awais2013b: Distributed hybrid simulation using the HLA and the Functional Mock-up
Interface [20].
Summary. The main difference between this work and [18] is that this proposes a variable step size
wrapper around CT components. The approach taken to do this is quantization.
nfr : distribution
nfr : parallelism
sr : rollback :none
sr : causality : causal
sr : rel time : analytic
sr : availability :remote
sr : availability : local
fr : alg loop : explicit
fr :communication model:gauss seidel
fr :communication model:jacobi
94
fr :standard:fmi
fr :standard:hla
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : variable
fr :coupling model:io assignments
fr :domain:de
fr : results visualization :post mortem
Ref 36. Awais2013a: Using the HLA for Distributed Continuous Simulations [18].
Summary. This work addresses the need to adapt CT simulators as DE simulators, in order to be
used in a hybrid co-simulation scenario that is fundamentally DE oriented.
nfr : distribution
nfr : parallelism
sr : rollback :none
sr : causality : causal
sr : rel time : analytic
sr : availability :remote
sr : availability : local
fr : alg loop : explicit
fr :communication model:gauss seidel
fr :communication model:jacobi
fr :standard:fmi
fr :standard:hla
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr :coupling model:io assignments
fr :domain:de
fr : results visualization :post mortem
Ref 37. Kuhr2013: FERAL - Framework for simulator coupling on requirements and architecture
level [146].
Summary. They describe a framework that borrows many concepts from Ptolemy, but that is
fundamentally event based co-simulation. It allows for the specialization of basic directors for the
semantic adaptation of simulation units.
nfr : ip protection
nfr : extensibility
sr : info : signal
sr : causality : causal
sr : rollback :none
sr : availability : local
sr : rel time : analytic
95
fr :sim rate:multi
fr :communication model:gauss seidel
fr :standard:fmi
fr :domain:de
fr :domain:ct
fr :num sim:three more
fr : sim step size : variable
Ref 38. Viel2014: Implementing stabilized co-simulation of strongly coupled systems using the
Functional Mock-up Interface 2.0. [250].
Summary. This work describes the implementation of the method described in [10] in the context
of the FMI standard.
nfr :accuracy
nfr : ip protection
sr : info :jacobian:out
sr : info :input extrapolation
sr : info :record outputs
sr : info : stateserial
sr : causality : causal
sr : rel time : analytic
sr : availability : local
sr : rollback :none
fr :domain:ct
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : implicit
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:gauss seidel
fr :standard:fmi
Ref 39. Sicklinger2014: Interface Jacobian-based Co-Simulation [220].
Summary. Describes a co-simulation method that makes use of the Jacobian information for fixed
point computations.
nfr :performance
nfr :accuracy
sr : availability : local
sr : rel time : analytic
sr : rollback : single
sr : info :jacobian:out
sr : causality : causal
fr : results visualization :post mortem
96
fr :communication model:gauss seidel
fr :communication model:jacobi
fr :coupling model:algebraic constraints
fr :domain:ct
fr : alg loop : implicit
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
Ref 40. Zhang2014: A co-simulation framework for design of time-triggered automotive cyber
physical systems [266].
Summary. The work describes a co-simulation that integrates SystemC and CarSim.
sr : availability : local
sr : rollback :none
sr : causality : causal
sr : info : full model
sr : rel time : analytic
fr :coupling model:io assignments
fr :domain:de
fr :domain:ct
fr : alg loop : explicit
fr : results visualization :post mortem
fr : sim step size : fixed
fr :sim rate: single
fr :num sim:two
Ref 41. Kounev2015: A microgrid co-simulation framework [139].
Summary. Describes the coupling of two simulators written in MATLAB and OMNeT++.
nfr :performance
sr : availability : local
sr : rel time : analytic
sr : rollback :none
The DEV’s orchestration is conservative.
sr : causality : causal
sr : info : predict step sizes
fr :coupling model:io assignments
fr : results visualization :post mortem
fr :communication model:gauss seidel
fr :domain:de
fr :num sim:two
fr :sim rate: single
fr : sim step size : variable
fr : alg loop : explicit
97
Ref 42. Bogomolov2015: Co-Simulation of Hybrid Systems with SpaceEx and Uppaal [36].
Summary. The orchestration algorithm is the one described in [46]. The work exploits the standard
by allowing zero step transitions.
sr : info : causality :feedthrough
sr : rollback :none
sr : info : stateserial
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :coupling model:io assignments
fr :standard:fmi
fr :num sim:two
fr :communication model:jacobi
fr : alg loop : explicit
fr :sim rate: single
fr : sim step size : variable
This is due to the rejection of steps, not due to accuracy.
fr :domain:ct
fr :domain:de
They abuse the FMI standard to be able to support state transitions.
fr : results visualization :post mortem
Ref 43. Bian2015: Real-time co-simulation platform using OPAL-RT and OPNET for analyzing
smart grid performance [33].
Summary. Not many details are provided about the co-simulation orchestration. However, due to
the fact that it is real-time, we can infer certain features.
nfr :performance
sr : rollback :none
sr : causality : causal
sr : rel time : fixed real scaled time simulation
sr : availability :remote
Communication is done through UDP.
fr : sim step size : fixed
fr :sim rate: single
Two simulators do not give more than this.
fr :domain:ct
fr :coupling model:io assignments
fr :num sim:two
fr : alg loop : explicit
fr : results visualization :post mortem
Ref 44. Dols2016: Coupling the multizone airflow and contaminant transport software CONTAM
with EnergyPlus using co-simulation [73].
98
Summary. The work described the coupling of the CONTAM and EnergyPlus tools to achieve
HVAC simulation. The coupling is done through FMI. The coupling is done through the compiled
binaries. The case study highlights the problems with an explicit method for co-simulation, even if
the Gauss-seidel. Instabilities occur.
nfr : ip protection
sr : rollback :none
sr : rel time : analytic
sr : causality : causal
sr : availability :remote
fr : results visualization :post mortem
fr :domain:ct
fr :num sim:two
fr :standard:fmi
fr : alg loop : explicit
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr : sim step size : fixed
It uses a 5-minute synchronization step.
fr :sim rate: single
Ref 45. BenKhaled2012: Multicore simulation of powertrains using weakly synchronized model
partitioning [133].
Summary. According to [28], this work explores variable step solvers.
nfr : parallelism
nfr :performance
sr : info : causality :feedthrough
sr : info : full model
sr : rel time : fixed real scaled time simulation
sr : rollback :none
sr : availability : local
fr :standard:fmi
fr :coupling model:io assignments
fr :num sim:three more
fr :domain:ct
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
fr : results visualization :post mortem
fr :communication model:jacobi
fr :communication model:gauss seidel
Ref 46. BenKhaled2014: Fast multi-core co-simulation of Cyber-Physical Systems: Application
to internal combustion engines [28].
99
Summary. This paper focus on the parallelization of co-simulation. The approach is to start with
a single model and partition it into multiple models, which are then executed in separate FMUs in
parallel. The partitioning is important for accuracy reasons (e.g., break the algebraic loops at less
sensitive variables).
nfr : parallelism
nfr :accuracy
nfr :performance
nfr : scalability
sr : info : full model
sr : info :wcet
sr : info : causality :feedthrough
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :communication model:gauss seidel
fr : sim step size : fixed
fr :sim rate: single
fr :standard:fmi
fr :domain:ct
fr :num sim:three more
fr : alg loop : explicit
It breaks the loops by establishing an order and delaying one of the variables in the loop.
Ref 47. Saidi2016: Acceleration of FMU Co-Simulation On Multi-core Architectures [208].
Summary. The paper addresses the problem of performance in FMI co-simulation. The solution
proposed is to go parallel. The parallelization approach is the same as the one presented in [133].
Since FMI does not enforce thread safety across multiple instances of the same FMU, the work presented ensures that these do not execute concurrently by using mutexes or changing the scheduling
policy.
nfr : parallelism
nfr :performance
nfr : ip protection
nfr : scalability
sr : info :wcet
sr : info : causality :feedthrough
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :communication model:jacobi
fr :standard:fmi
fr :domain:ct
fr :num sim:three more
fr : alg loop : explicit
100
Ref 48. Yamaura2016: ADAS Virtual Prototyping using Modelica and Unity Co-simulation via
OpenMETA [256].
Summary. The co-simulation framework includes 4 tools. The communication between the tools is
realized using OpenMeta. The work uses Unity for the modelling and simulation of the environment,
allowing for live interaction. Communication is over UDP but there is no report on extra caution
due to network delays and failures.
nfr : parallelism
sr : info : full model
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
fr :num sim:three more
fr :domain:ct
fr : results visualization : live
fr : results visualization : interactive live
fr :coupling model:io assignments
fr :sim rate: single
fr : alg loop : explicit
Ref 49. Camus2015: Combining DEVS with multi-agent concepts to design and simulate multimodels of complex systems (WIP) [55].
Summary. This work is the preliminary description of [56].
nfr : parallelism
nfr : distribution
nfr : ip protection
nfr :accuracy
sr : rollback :none
sr : info : stateserial
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:de
fr : alg loop : explicit
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : variable
fr :standard:fmi
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr : results visualization :post mortem
101
Ref 50. Camus2016: Hybrid Co-simulation of FMUs using DEV & DESS in MECSYCO [56].
Summary. It proposes to use a FMU wrapper around DEV and DESS models, meaning that the
co-simulation proceeds using a DE approach. It handles black box FMUs and the algorithm used
to drive the co-simulation is the conservative parallel DEVS simulator. It requires that the FMU
is able to perform rollback (through the use of state set and get).
nfr : parallelism
nfr : distribution
nfr : ip protection
nfr :accuracy
sr : rollback :none
sr : info : stateserial
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:de
fr : alg loop : explicit
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : variable
fr :standard:fmi
fr :coupling model:io assignments
fr :communication model:gauss seidel
fr : results visualization :post mortem
Ref 51. Pedersen2016: FMI for Co-Simulation of Embedded Control Software [195].
Summary. The paper describes the adaptation of an embedded system to comply with FMI and
thus interface with other FMUs. To validate the implementation, they run a co-simulation.
nfr : distribution
nfr : parallelism
sr : rel time : fixed real scaled time simulation
fr :domain:ct
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
fr :communication model:gauss seidel
fr :standard:fmi
fr :coupling model:io assignments
fr : alg loop : explicit
fr : results visualization : live
Ref 52. Oh2016: A Co-Simulation Framework for Power System Analysis [191].
102
Summary. The paper proposes a co-simulation framework that takes into account network delays
and compensates for that. It proposes to use cubic spline extrapolation to compensate for the delay
but recognizes that if there are faults in the line (resulting in voltage drops), the derivatives used in
the extrapolation assume gigantic proportions, thus wreaking havoc in the simulation. To address
that, the framework employes an algorithm to detect discontinuities. The detection is simple: they
check the derivative of the signal to see whether it exceeds a pre-determined empirically threshold.
Basically, it looks for and Dirac delta. Figure 7 shows the effect of not handling a discontinuity.
nfr : distribution
nfr :accuracy
nfr : parallelism
fr :num sim:two
fr :sim rate: single
fr : sim step size : fixed
fr :communication model:gauss seidel
Due to the parallel interface protocol that they use.
fr :coupling model:io assignments
fr :domain:ct
fr : alg loop : explicit
sr : availability :remote
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
Ref 53. Xie2016: Continuous-Mass-Model-Based Mechanical and Electrical Co-Simulation of
SSR and Its Application to a Practical Shaft Failure Event [255].
Summary. Between two simulators. As it is explained in the paper, prior to co-simulation, the most
common approach would be to run two simulations: one complete for one sub-system, and then
another for the second sub-system, using the first as inputs. This is an open loop approach, whose
results can be misleading due to ignoring the feedback loops. Each simulator advances in parallel
and their communication is made with a barrier.
nfr : parallelism
fr :communication model:jacobi
fr :num sim:two
fr :domain:ct
fr : sim step size : fixed
fr : results visualization :post mortem
fr :sim rate: single
sr : availability : local
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
fr : alg loop : explicit
fr :coupling model:io assignments
103
Ref 54. Manbachi2016: Impact of EV penetration on Volt–VAR Optimization of distribution
networks using real-time co-simulation monitoring platform [163].
Summary. It describes an application of co-simulation in the distribution of energy in smart grids,
supported by a real-time co-simulation framework. The simulators involved are the RTDS, which
simulates the distribution network model, and the VVO Engine, coded in MATLAB.
sr : rel time : fixed real scaled time simulation
fr :num sim:two
sr : causality : causal
sr : rollback :none
sr : availability : local
fr :coupling model:io assignments
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
fr :communication model:jacobi
Ref 55. Schierz2012a: Co-simulation with communication step size control [211].
Summary. Describes a master algorithm. Does not allow for interpolation of inputs. Needs rollback.
It touches upon accuracy, as it suggests an adaptive step size control mechanism. It does not address
algebraic loops. It assumes that there is no feedthrough information.
nfr :performance
nfr :accuracy
sr : info : derivatives :out
sr : info : stateserial
sr : causality : causal
sr : rel time : analytic
sr : rollback : single
sr : availability : local
fr :standard:fmi
fr :coupling model:io assignments
fr :num sim:three more
fr :domain:ct
fr :sim rate: single
fr : sim step size : variable
fr : alg loop : explicit
fr : results visualization :post mortem
fr :communication model:jacobi
Ref 56. Fourmigue2009: Co-simulation based platform for wireless protocols design explorations [90].
104
Summary. Application of co-simulation to wireless network development. One of the simulators is
the actual Linux operating system, and the other is represents a wireless network protocol simulator.
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :coupling model:io assignments
fr :num sim:two
fr :domain:de
fr :communication model:jacobi
Ref 57. Liu2001: Calculation of Wing Flutter by a Coupled Fluid-Structure Method [160].
Summary. A fully implicit method, dealing with parallelism.
nfr : parallelism
nfr :performance
sr : causality : causal
sr : rel time : analytic
sr : rollback : single
sr : availability :remote
fr :coupling model:io assignments
fr :num sim:two
fr :domain:ct
fr : alg loop : implicit
fr : results visualization :post mortem
Ref 58. Carstens2003: Coupled simulation of flow-structure interaction in turbomachinery [59].
Summary. Relates to the application of a co-simulation algorithm to the simulation of the deformation in the blades of a transonic compressor rotor under airflow. One of the simulators calculates
deformation of the blades, while the other calculates the flow dynamics around the blades.
The communication of orchestration algorithm in use is shifted by half a step.
nfr :performance
They highlight the need for it, because the computation of a rotor is just too expensive.
sr : info : derivatives :out
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability :remote
It seems that they perform the computation in separate computers.
fr :coupling model:io assignments
fr :num sim:two
fr :domain:ct
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
105
fr : results visualization :post mortem
fr :communication model:gauss seidel
Although it is a gauss seidel shifted in time.
Ref 59. Stettinger2014: Model-based coupling approach for non-iterative real-time co-simulation [221].
Summary. Proposes to address the challenges in real-time co-simulation by using a model based
coupling approach. The master has to keep track of two values for each packet of data: receiving
time delay tr – the time it takes for a packet to reach the master from the simulator –, and sending
time delay ts – the time it takes for a packet to leave the master and reach the simulator. When a
sample is delayed, the master acts as a replacement for it. Basically, it is a dead reckoning model.
nfr :performance
nfr : parallelism
nfr :accuracy
sr : causality : causal
fr :domain:ct
fr :num sim:two
sr : availability : local
fr :coupling model:io assignments
sr : rollback :none
sr : rel time : fixed real scaled time simulation
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
fr : results visualization :post mortem
fr :communication model:jacobi
fr :communication model:gauss seidel
Ref 60. Benedikt2016: Automated configuration for non-iterative co-simulation [31].
Summary. Describes how a co-simulation master can configure some parameters throughout the
co-simulation execution. This is the idea behind adaptive master algorithms.
nfr : ip protection
nfr :accuracy
sr : info : causality :feedthrough
sr : causality : causal
sr : availability : local
sr : rollback :none
sr : rel time : analytic
fr :domain:ct
fr :num sim:three more
fr :coupling model:io assignments
fr :sim rate: single
fr : sim step size : variable
fr : alg loop : explicit
fr : results visualization :post mortem
106
Ref 61. Busch2011: An explicit approach for controlling the macro-step size of co-simulation
methods [52].
Summary. Presents an approach to estimate the local truncation error caused by the extrapolations
of inputs in a co-simulation. The sub-systems are assumed to make no error. It does not require
rollback or re-initialization.
Categories:
nfr :accuracy
Because they study the global error and control the local error.
nfr : ip protection
nfr :performance
They control the step size, which increases performance. And they study how to get an optimal
step size.
sr : info : causality :feedthrough
sr : rollback :none
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
fr : sim step size : variable
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
In theory, they seem to support any communication model. In the paper they studied assuming
the Jacobi.
Ref 62. Quesnel2005: DEVS coupling of spatial and ordinary differential equations: VLE framework [202].
Summary. Proposes a way to wrap a continuous time ODE simulator as a DEVS model. It requires
that the state variables, and derivatives are available.
Categories:
nfr :hierarchy
nfr :open source
sr : info : derivatives :out
sr : info : statevars
sr : info : predict step sizes
sr : causality : causal
fr :domain:de
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
Any discrete event framework is by definition multi-rate.
107
fr : sim step size : variable
Any discrete event framework is by definition in this category.
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:gauss seidel
A discrete event framework is in this category as there is no extrapolation of inputs. Also,
Gauss seidel does not violate the causality of inputs and outputs, because it sorts according to
these dependencies. Events are processed to retain their causality.
Ref 63. Arnold2014a: Error analysis for co-simulation with force-displacement coupling [15].
Summary. Describes an FMI based master called SNiMoWrapper.
Categories:
nfr :accuracy
Because they study the global error and control the local error.
nfr : ip protection
sr : info : causality :feedthrough
sr : rollback :none
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : explicit
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
fr :standard:fmi
Ref 64. Arnold2014: Error Analysis and Error Estimates for Co-simulation in FMI for Model
Exchange and Co-Simulation v2.0 [14].
Summary. Studies the error control method known as Richard’s extrapolation.
Categories:
nfr :accuracy
Because they study the global error and control the local error.
nfr : ip protection
sr : info : causality :feedthrough
sr : info : statevars
sr : rollback : single
sr : causality : causal
108
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
fr : sim step size : variable
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
fr :communication model:gauss seidel
fr :standard:fmi
Ref 65. Arnold2001: Preconditioned Dynamic Iteration for Coupled Differential-Algebraic Systems [11].
Summary. Studies the convergence of the Gauss-Seidel dynamic iteration method and proposes a
way to ensure it. The way to do it though, requires information from the model.
Categories:
nfr :accuracy
Because they study the global error.
sr : info :jacobian:out
sr : info :record outputs
sr : info : full model
sr : rollback : single
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : implicit
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:gauss seidel
Ref 66. Schweizer2014: Semi-implicit co-simulation approach for solver coupling [215].
Summary. Proposes a predictor corrector master that evaluates the macro step twice and uses a
perturbation on the inputs to get an estimate of the required partial derivatives. This approach is
then generalized to multiple kinds of joints in the mechanical domain. A double pendulum, double
mass-spring-damper and a slider crank mechanism are used as numerical examples.
Categories:
sr : rollback : single
109
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:two
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : semi implicit
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 67. Schweizer2015d: Stabilized implicit co-simulation methods: solver coupling based on
constitutive laws [218].
Summary. It presents an implicit and semi-explicit methods for the co-simulation of scenarios coupled via applied forces. The difference between this paper and the previous ones by the same author
seems to be in the fact that the coupling constraints are integrated and differentiated, to enrich the
information being used to ensure that the original coupling constraints are met.
Categories:
sr : rollback : single
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : semi implicit
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 68. Sadjina2016: Energy Conservation and Power Bonds in Co-Simulations: Non-Iterative
Adaptive Step Size Control and Error Estimation [207].
Summary. Proposes a master for co-simulation that requires the identification of power bonds
between sub-systems. It assumes that the scenario is energy conserving and thus calculate the
energy residual as an error to be minimized. The step size is adapted via a PI-Controller. When
the step size is reduced, it is only on the next co-simulation step, so the method is explicit.
nfr : ip protection
nfr :accuracy
Due to step size control.
sr : rollback :none
sr : causality : causal
110
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : variable
fr : alg loop : explicit
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
Ref 69. Busch2016: Continuous approximation techniques for co-simulation methods: Analysis
of numerical stability and local error [50].
Summary. Analyses the stability and local error of multiple co-simulation approaches with multiple
extrapolation approaches for the inputs. It considers Gauss-Seidel and Jacobi. It also talks about a
method called the extrapolated interpolation method, which ensure no discontinuities at the inputs
of the subsystems.
sr : rollback :none
The method is explicit.
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
fr :communication model:gauss seidel
Ref 70. Arnold2010: Stability of Sequential Modular Time Integration Methods for Coupled
Multibody System Models [10].
Summary. Studies stability of a gauss Seidel co-simulation method proposed in previous work:
[11]. Based on that analysis, it proposes an implicit stabilization technique that uses Gauss-Seidel
iteration. The resulting method is implicit but the equations that are being solved are linear.
Categories:
nfr :accuracy
Because they study the global error.
sr : info :jacobian:out
sr : info :record outputs
sr : info : full model
111
sr : rollback : single
sr : causality : causal
sr : rel time : analytic
sr : availability : local
fr :domain:ct
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
fr : alg loop : implicit
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:gauss seidel
Ref 71. Gu2001: Co-simulation of algebraically coupled dynamic subsystems [105].
Summary. Describes a technique to deal with algebraically coupled sub-systems using a control
theoretic approach. The highlights of this method are: it supports scenarios of arbitrary index; the
boundary condition coordinator is seen as a co-simulation unit (this is an elegant approach) and
the method is explicit. The beauty of making the BCC as a co-simulation unit, is that it can, just
like any other sub-system be run at a different rate and in the paper they show that by running it
at a higher rate, the stability of the co-simulation increases.
sr : rollback :none
sr : causality : causal
fr :domain:ct
fr :num sim:two
sr : rel time : analytic
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 72. Gu2004: Co-Simulation of Algebraically Coupled Dynamic Subsystems Without Disclosure of Proprietary Subsystem Models [107].
Summary. Describes a technique to solve causal conflicts using a Boundary Condition Coordinator
(BCC). Causal conflicts arise naturally from the coupling of different sub-systems and they are
a relevant challenge that needs to be overcome in order to perform correct co-simulation. While
in [105], the BCC requires the knowledge of the state variables of the simulations, in [107], some
modifications are made to ensure that this information is not required.
nfr : ip protection
nfr : distribution
sr : rollback :none
112
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : explicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 73. Schweizer2016: Co-simulation method for solver coupling with algebraic constraints
incorporating relaxation techniques [219].
Summary. A master algorithm capable of dealing with algebraic constraints is described. It requires
the derivatives of the coupled variables to be available. It executes each communication step twice,
being a semi-implicit method. It uses a predict step and a corrector step. The final corrected
coupling variables are obtained by polynomial extrapolation and relaxation (to avoid instabilities).
Categories:
sr : rollback : single
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
sr : info :jacobian:out
fr : sim step size : fixed
fr : alg loop : semi implicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 74. Schweizer2015: Predictor/corrector co-simulation approaches for solver coupling with
algebraic constraints [214].
Summary. Proposes a predictor corrector master that evaluates the macro step twice and uses a
perturbation on the inputs to get an estimate of the required partial derivatives.
Categories:
sr : rollback : single
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
113
fr :sim rate:multi
fr : sim step size : fixed
fr : alg loop : semi implicit
sr : info :jacobian:out
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 75. Schweizer2015a: Stabilized index-2 co-simulation approach for solver coupling with
algebraic constraints [216].
Summary. A master algorithm capable of dealing with algebraic constraints is described. It requires
the derivatives of the coupled variables to be available. It executes each communication step twice,
being a semi-implicit method. It uses a predict step and a corrector step. The predictor step
allows the method to estimate the sensitivity of the state variables with respect to the applied
forces/torques.
Categories:
sr : rollback : single
sr : causality : causal
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr :sim rate:multi
sr : info :jacobian:out
fr : sim step size : fixed
fr : alg loop : semi implicit
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:algebraic constraints
fr :communication model:jacobi
Ref 76. Andersson2016: Methods and Tools for Co-Simulation of Dynamic Systems with the
Functional Mock-up Interface [8].
Summary. This is a Phd Thesis. A linear extrapolation based master is proposed that is convergent
and does not require fixed point iterations. Then, a modification to multi-step methods is proposed
to increase their performance when executing in a co-simulation environment. This modification
avoids the need to restart when dealing with discontinuities.
Categories:
nfr : ip protection
nfr :platform independence
nfr :open source
sr : rollback :none
sr : causality : causal
114
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
fr : sim step size : fixed
sr : availability : local
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
fr :standard:fmi
Ref 77. Krammer2015: Model-Based Configuration of Automotive Co-Simulation Scenarios [140].
Summary. The language is further developed in [140] with the addition of three novel diagrams to
represent different aspects of the co-simulation configuration:
• Architectural – coupling of executable units;
• Tools – assignment of tools to models;
• Connections – connections (it was not clear what does this diagram do);
In addition, they define a couple of well formedness properties that can be checked more easily with
the model-based approach. They give a brief summary of the tool ICOS.
Categories:
nfr : config reusability
nfr : parallelism
nfr :hierarchy
nfr : extensibility
fr :domain:ct
fr :num sim:three more
sr : rel time : analytic
sr : availability : local
fr :coupling model:io assignments
fr : results visualization :post mortem
sr : info : full model
Ref 78. Galtier2015: FMI-Based Distributed Multi-Simulation with DACCOSIM [96].
Summary. DACCOSIM is able to perform Distributed simulations and multi-core simulations. The
term “computation node” is used for a collection of FMU wrappers (which include an FMU) and
a local master / global master. The FMU wrappers, and thereby not the masters, are responsible
for passing outputs to connected inputs. This is to avoid bottlenecks. A component node contains
a master and some FMUs, which are wrapped in so-called “FMU-wrappers”. The masters take
responsibility of coordinated step sizes in case an FMU needs to roll back.
nfr :performance
Because of the possibility of splitting the simulation over a cluster / multi-core and the focus
on performance in the article. Additionally because of their use of variable step size
nfr : config reusability
115
It is possible to create multiple co-simulation configuration files in the simulation configuration
GUI. These can be stored and therefore reused.
nfr : ip protection
Some level of IP protection because of FMI..
nfr : parallelism
Because of the possibility of splitting the simulation over a cluster
nfr : distribution
Because a co-simulation can be executed on a cluster of computers
sr : availability :remote
sr : availability : local
nfr :hierarchy
DACCOSIM is weak hierarchical because it has the notion of local and global masters.
nfr : scalability
The framework is considered to be scalable because of the multi-core and distributed architecture.
nfr :platform independence
There are two versions of the DACCOSIM library. A cross-platform version relying on JAVA
and a Windows version using C++ and QTronic SDK.
nfr :accuracy
The article provides an example where the result of the co-simulation using DACCOSIM is
compared to the simulation using Dymola and the results are very close to each other. Accuracy is
ensured by each FMU examining its outputs and estimating how far they are from the exact value.
nfr :open source
The framework is distributed under an open source license from January 2016.
sr : info : stateserial
The framework can perform a single rollback using the state variable serialization.
sr : causality : causal
Because the framework is based on FMI for co-simulation it is considered to be causal.
fr :domain:ct
The framework supports multiple formalisms because it is based on FMI for co-simulation.
fr :num sim:three more
The frameworks is capable of supporting many FMUs and thereby many simulation units. DACCOSIM offers its own algorithm depending on global/local masters.
fr :coupling model:io assignments
sr : rel time : analytic
There is no mentioning of any other time models than this in the article.
fr :sim rate: single
The simulation rate is the same for all FMUs.
fr : sim step size : variable
The framework uses Coordinated Variable Step
fr : alg loop : explicit
The framework uses Euler’s method and Richardson’s method. Whether this is default, parameterizable or fully customizable is unknown based on this article.
fr :communication model:jacobi
See Co-initialization bullet 2 in the article.
fr :standard:fmi
116
It is based on the FMI standard.
fr : results visualization :post mortem
Ref 79. Fey1997: Parallel synchronization of continuous time discrete event simulators [84].
Summary. Presents two synchronization approaches, detailed in three different synchronization
protocols, to coordinate simulation scenarios that include one discrete event simulator and one
continuous time simulator. The discrete event simulator can implement any parallel simulation
approach that we know, such as Time-Warp. This means that, even internally, the DE simulator
can be forced to rollback due to straggler messages. The focus is on parallel approaches.
Categories:
nfr :performance
nfr : ip protection
nfr : parallelism
nfr :accuracy
fr :domain:de
fr :domain:ct
fr :num sim:two
fr :sim rate:multi
fr : sim step size : fixed
fr : results visualization :post mortem
sr : rel time : analytic
sr : availability : local
sr : rollback : single
fr :coupling model:io assignments
sr : info : full model
Ref 80. Acker2015: Generation of an Optimised Master Algorithm for FMI Co-simulation [236].
Summary. Essentially, this paper shows how a compiled approach increases the performance of the
co-simulation. It also shows that, because there are so many decisions to be made when designing
the master, a compiled approach allows for a more elegant, and specifically tailored master, to be
generated.
nfr :performance
nfr : ip protection
nfr : config reusability
nfr :open source
sr : rel time : analytic
sr : availability : local
sr : info : causality :feedthrough
sr : info : statevars
sr : rollback :none
sr : causality : causal
sr : info : preferred step sizes
fr :domain:ct
117
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr : results visualization :post mortem
fr :standard:fmi
fr :communication model:gauss seidel
fr : alg loop : implicit
fr :coupling model:io assignments
Ref 81. Enge-Rosenblatt2011: Functional Digital Mock-up and the Functional Mock-up Interface - Two Complementary Approaches for a Comprehensive Investigation of Heterogeneous
Systems [79].
Summary. The paper describes and compares two approaches to performing co-simulation of heterogeneous systems, namely the Functional Digital Mock-up (FDMU) and the Functional Mock-up
Interface (FMI). Besides describing these approaches it also introduces the “FDMU framework”, a
framework that implements the Functional Digital Mock-up approach. Furthermore, proposals are
presented for combining FDMU and FMI approaches.
The FDMU approach is a tool-independent and web service-based framework build on the Web
Service standards. It is capable of coupling different simulation tools and provide visualization
based on CAD models.
FDMU consists of three main concepts: functional building blocks (FBB), wrappers, FDMU
master, and FDMU Console. The functional building block can wrap geometric information (CAD
Models), behavioral models, and a simulator tool. It is the responsibility of the wrappers to establish
a connection between the different simulation tools and the FDMU Master Simulator. Finally, the
FDMU master ensures correct communication between the simulators. The FDMU Console is the
user’s front-end.
nfr :performance
Communication overhead of a web service-based approach.
nfr : ip protection
Because of the the web service-based approach IP protection should be possible.
nfr : parallelism
Because of the web service-based approach it is parallel by nature. It uses thread-safe queues
and deadlock-free transmission of data.
nfr : distribution
Because of the web service-based approach it is easy distributable.
nfr : scalability
The distributed systems paradigm ensures scalability.
nfr :platform independence
web service-based approach.
nfr : extensibility
A new wrapper can be implemented.
sr : causality : causal
Every input of an FBB must have an appropriate output belonging to another FBB.
fr :domain:ct
fr :coupling model:io assignments
118
fr :num sim:three more
fr :sim rate:multi
sr : availability :remote
fr : results visualization : live
The framework provides an interactive 3D visualization based on CAD.
fr :standard:fdmu
Ref 82. Karner2010a: Heterogeneous co-simulation platform for the efficient analysis of FlexRaybased automotive distributed embedded systems [130].
Summary. Motivation: FlexRay is a wired network for automotive high-speed control applications
and no solutions exist that simulates all parts of the network.
What: a co-simulation platform called TEODACS FlexRayExprt.Sim. The simulation approach
used in the platform covers mechanics and all parts of the network from physical layer to application
layer, which is not done by other solutions. The framework CISC SyAD is used to perform the
microelectronics co-simulation, CarMaker/AVL InMotion for the mechanics, and they are bridged
by TEODACS FlexRayEprt.Sim. The platform uses a very interesting approach to faster cosimulations, namely the use of model switching, where a less detailed model replaces a more detailed
model in parts of the simulation.
The paper provides an overview of existing approaches such as transaction based modeling,
HDLs such as SystemC and Verilog, and cable harness and topology modeling along with why
these contain shortcomings to this domain. Furthermore, the paper provides some details of the
implementation of the models used in the co-simulation and showcases how the platform can analyse
a system with specific examples.
nfr :accuracy
Because of model switching.
nfr :performance
Because of model switching.
fr :dynamic structure
Because the structure of the co-simulation is changed via model switching.
fr :domain:ct
Because SyAD supprots multiple formalisms and CarMaker / AVL InMotion.
fr :domain:de
fr :coupling model:io assignments
fr :num sim:three more
Multiple FlexRay nodes can be added.
sr : rel time : analytic
From model switching and similar it is clear that analytic simulation is used.
fr : results visualization :post mortem
Ref 83. Aslan2015: MOKA: An Object-Oriented Framework for FMI Co-Simulation [16].
Summary. The paper describes MOKA, which is a framework for performing co-simulations and
creating FMUs using FMI 2.0 for co-simulation. The framework turns the creation of FMUs into
an object-oriented process by using C++. An FMU is created by inheriting one of the classes and
119
implementing virtual functions thereby avoiding writing boilerplate code. The implementation of
FMUs is realised by the concepts of FMUBlock, which is to be inherited, FMUPort, and FMUStateVariables. FMUBlock is to be extended by a concrete FMU slave and implements common
computation phase functions for slaves. It contains FMUPort for data exchange and FMUStateVariables for state tracking during the simulation. The FMUPort classes provides the data exchange
interface of a slave. It abstracts the value references by automatically assigning a value reference to
the variable. The BaseStateVariable class also functions as base that is to be extended. It provides
virtual functions for state variable services. The StateVariable inherits from BaseStateVariable
and represents state variables for the slave. The framework also provides a template for the FMU
Master so the master code changes minimally for different scenarios.
The article exemplifies an application of the MOKA framework where two FMUs are used: The
bouncing ball and integer counter example from the QTronic SDK, where the bouncing ball has
been re-developed with MOKA.
In future work it is stated that development of a DSL in a current study, so that different
scenarios can be executed without altering the master code.
nfr : ip protection
nfr : config reusability
sr : causality : causal
sr : rel time : analytic
sr : rollback :none
sr : availability : local
fr :coupling model:io assignments
fr :num sim:three more
fr :standard:fmi
fr :domain:ct
fr : alg loop : explicit
fr :sim rate: single
fr :communication model:gauss seidel
fr : results visualization :post mortem
Ref 84. Wetter2010: Co-simulation of building energy and control systems with the Building
Controls Virtual Test Bed [253].
Summary. Describes a co-simulation framework called Building Controls Virtual Test Bed (BCVTB)
that can be used for real-time simulation and co-simulation. It is a modular extensible open-source
platform to interface different simulation programs with each other. The intention is to give users
the option to use the best tools suited to model various aspects of building energy and control
systems, or use programs where they have expertise. The middleware to couple any number of
simulation programs also provides libraries such that it can be extended. Furthermore, the paper
describes how they gathered capabilities the framework should support. The framework is based
on Ptolemy II, which is extended by some java packages. The simulator package adds functionality
that allows an actor to perform system calls to start any executable on Windows, OSX or Linux.
It simply starts a simulation program, sends input tokens to the simulation program, receives new
values and sends them to its output port. Algorithms are also provided on how simulators are
coupled. These are also exemplified with specific simulators. It is also described how to connect
120
client programs. The article describes how the interfaces are created for simulink, matlab, modelica,
and system cals. Furthermore, a specific example is presented.
nfr : config reusability
nfr : distribution
nfr :platform independence
nfr : extensibility
nfr :open source
sr : causality : causal
sr : rel time : analytic
sr : rel time : fixed real scaled time simulation
fr :domain:ct
In the paper, their explanation is focused on the CT domain.
fr :num sim:three more
fr :sim rate: single
fr : sim step size : fixed
fr : results visualization :post mortem
fr :coupling model:io assignments
fr :communication model:jacobi
Ref 85. Neema2014: Model-based integration platform for FMI co-simulation and heterogeneous
simulations of cyber-physical systems [182].
Summary. The article concerns integrating FMI as an HLA federate and extending the Command
and Control Wind Tunnel (C2WT) metamodel to include FMI-specifics. This enables the C2WT
tool to use FMUs as part of a simulation. The C2WT tool is describes as a multi-model integration
platform that allows users to model and synthesize complex, heterogeneous, command and control
simulations. The tool therefore has support for multiple simulation engines and an introduction to
the tool is given in the paper. Furthermore, a case study on Vehicle Thermal Management using
FMUs are presented and The focompared to a simulation in a different environment. The work is
sponsored by the US DoD.
nfr :platform independence
nfr : distribution
nfr : config reusability
fr :domain:de
fr :num sim:three more
fr :sim rate:multi
fr : sim step size : fixed
fr : sim step size : variable
sr : rel time : fixed real scaled time simulation
sr : rel time : analytic
fr : results visualization : live
fr :standard:hla
fr :standard:fmi
fr :communication model:jacobi
121
Ref 86. Larsen16c: Integrated Tool Chain for Model-Based Design of Cyber-Physical Systems [148].
Summary. This article presents an overview of the INTO-CPS project and thereby a Co-Simulation
tool. The projects concerns production of a well-founded tool chain for model-based design of
CPSs, and therefore consists of a semantic foundation and several baseline tools such as Modelio,
Overture, 20-sim, OpenModelica and RT-Tester. Furthermore, an application called the INTO-CPS
application is the entry point for configuring co-simulations and uses the co-simulation orchestration
engine (COE) to perform the actual simulations. This COE is based on the FMI standard. The
entire tool chain and the semantic foundation are presented in this paper. This is related to [148].
nfr : config reusability
nfr : ip protection
sr : info :nominal values:output
sr : info :nominal values:state
sr : info : derivatives :out
sr : info : derivatives : state
sr : info :jacobian:out
sr : info :jacobian: state
sr : info : preferred step sizes
sr : info : causality :feedthrough
sr : causality : causal
sr : availability : local
sr : info : statevars
sr : info :record outputs
sr : rel time : analytic
sr : info : stateserial
fr :standard:fmi
sr : info : signal
fr :num sim:three more
fr :domain:de
fr :domain:ct
fr :sim rate: single
fr : results visualization :post mortem
fr : results visualization : live
Ref 87. INTOCPSD41d: Design of the INTO-CPS Platform [149].
Summary. This is an EU deliverable related to the INTO-CPS project. It contains the technical
documentation of the INTO-CPS platform at the end of 2015 (the first year of the project). Part
of this project is the Co-simulation Orchestration Engine (COE). This is related to [148].
sr : info : predict step sizes
Supports the FMI suggested extension fmi2getMaxStepSize
nfr :performance
The COE supports variable step size, which can increase performance.
sr : rollback : single
Supports rollback to last successful state.
122
nfr :performance
The COE supports variable step size, which can increase performance.
nfr :accuracy
Contains various constraints such as zero-crossing, bounded difference and sampling rate. Furthermore, support for allowed min and max values.
C
Co-Simulation Scenario Categorization
This section describes each category and lists the references that belong to that category, classified
in the previous section.
C.1
C.1.1
Non-Functional Requirements
Fault Tolerance
A co-simulation platform is fault tolerant if, for example, when one simulation unit fails, other can
take its place. This is particularly important for long running simulations. To be fault tolerant,
certain features need to available: periodically store the state of simulation units; record all inputs
to each simulation unit. If a simulation unit fails and the state is periodically stored, then the
simulation can be paused while the state is restored in a new instance of the simulation unit. The
history of input values passed to each simulation unit can be used to bring the simulation unit to
the current state.
References in this category:
Fuller2013
C.1.2
Configuration Reusability
This category refers to the fact that frameworks can provide a way to configure co-simulation
scenarios that can be reused. This means that the configuration is considered external to the
execution of the co-simulation. External in this context means that the configuration can reused
without altering the binaries for the co-simulation application.
If a tool/frame does not provide a way to reuse configurations for co-simulation, then it is a
time-consuming, error-prone and non-trivial process to set up co-simulations [140].
References in this category:
Wang2013
Krammer2015
Galtier2015
Acker2015
Aslan2015
Wetter2010
Neema2014
C.1.3
Performance
Performance is a relative measure: a co-simulation platform is performant when it is able to simulate
a great deal in a short amount of time while needing little resources. This can be achieved by using
123
variable step integration methods and signal extrapolation techniques. Parallelism also plays a role
but mostly on the time aspect of performance.
References in this category:
Hoepfer2011
Faure2011
Sun2011
Friedrich2011
Gonzalez2011
Gunther2012
Eyisi2012
Riley2011
Benedikt2013
Sicklinger2014
Kounev2015
Bian2015
BenKhaled2012
BenKhaled2014
Saidi2016
Schierz2012a
Liu2001
Carstens2003
Stettinger2014
Busch2011
Galtier2015
Fey1997
Acker2015
Enge−Rosenblatt2011
Karner2010a
C.1.4
IP Protection
IP Protection deals with not requiring the models participating in the co-simulation to provide detailed structure, variables, etc. . . There are multiple levels of protection ranging from fully protected
to not protected at all. A good IP protection enables component suppliers to provide the system
integrators with detailed simulations of their components avoiding expensive lock-in contracts.
There are multiple techniques can be be employed to ensure some degree of protection. For
instance, making the models (and corresponding simulation units) available as a web service is a
possible solution. Another example is any framework that implements the FMI Standard [34, 35],
which allows models and simulation units to be exported as a single functional unit, in binary
format, that can be imported into a co-simulation.
References in this category:
Hoepfer2011
Sun2011
Bastian2011a
Friedrich2011
Broman2013
124
Fuller2013
Wang2013
Kuhr2013
Viel2014
Dols2016
Saidi2016
Camus2015
Camus2016
Benedikt2016
Busch2011
Arnold2014a
Arnold2014
Sadjina2016
Gu2004
Andersson2016
Galtier2015
Fey1997
Acker2015
Enge−Rosenblatt2011
Aslan2015
C.1.5
Parallelism
A co-simulation framework is parallel when it makes use of multiple processes/threads to perform
the co-simulation. This is typically in the same computer or same local network.
Techniques such as signal extrapolation help improve the speed-up gained from parallelism.
Furthermore waveform relaxation techniques and the Jacobi iterations promote parallelism [167].
References in this category:
Faure2011
Bastian2011a
Friedrich2011
Gunther2012
Awais2013b
Awais2013a
BenKhaled2012
BenKhaled2014
Saidi2016
Yamaura2016
Camus2015
Camus2016
Pedersen2016
Oh2016
Xie2016
Liu2001
Stettinger2014
Krammer2015
125
Galtier2015
Fey1997
Enge−Rosenblatt2011
C.1.6
Distribution
A co-simulation framework is parallel and distributed when it allows each simulation unit to be
remote, across a wide area network.
This is very important since suppliers can, instead of transferring the simulation units in executable form across computers, can make them available over the web. This offers much more
control over how the simulation units are used.
The same techniques used in parallelism can be used to promote distribution, but fault tolerance
is also important.
References in this category:
Friedrich2011
Nutaro2011
Busch2012
Quaglia2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Fuller2013
Bombino2013
Zhao2014
Awais2013b
Awais2013a
Camus2015
Camus2016
Pedersen2016
Oh2016
Gu2004
Galtier2015
Enge−Rosenblatt2011
Wetter2010
Neema2014
C.1.7
Hierarchy
A hierarchical co-simulation framework is able to abstract a co-simulation scenario as a black box
simulation unit. This is very intuitive and promotes abstraction of complex systems.
References in this category:
Quesnel2005
Krammer2015
Galtier2015
126
C.1.8
Scalability
A co-simulation framework is scalable when it supports a large number of simulation units. It is
intimately related to performance and paralelism.
References in this category:
Fuller2013
BenKhaled2014
Saidi2016
Galtier2015
Enge−Rosenblatt2011
C.1.9
Platform Independence
A co-simulation framework is platform independent when it works on multiple computing platforms.
For this to be achieved, a platform independent language, such as Java, can be used to coordinate
the simulation.
References in this category:
Bastian2011a
Eyisi2012
Riley2011
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Andersson2016
Galtier2015
Enge−Rosenblatt2011
Wetter2010
Neema2014
C.1.10
Extensibility
A co-simulation framework is extensible when it can be easily extended to support new kinds of
simulation units, with new kinds of capabilities. A higher level, domain specific, language can
be used to specify the behaviour in a platform agnostic way. Code is then generated from this
description. The hypothesis is that the high level description can be more easily extended to
describe new behaviour and that the code generation process can be adapted accordingly.
References in this category:
Kuhr2013
Krammer2015
Enge−Rosenblatt2011
Wetter2010
C.1.11
Accuracy
A co-simulation is accurate when the error between the trace produced and the correct trace is
minimal. This can be achieved by error control mechanisms.
References in this category:
127
Hoepfer2011
Tomulik2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Schierz2012
Gunther2012
Fitzgerald2010
Fitzgerald2013
Benedikt2013
Benedikt2013b
Hafner2013
Viel2014
Sicklinger2014
BenKhaled2014
Camus2015
Camus2016
Oh2016
Schierz2012a
Stettinger2014
Benedikt2016
Busch2011
Arnold2014a
Arnold2014
Arnold2001
Sadjina2016
Arnold2010
Galtier2015
Fey1997
Karner2010a
C.1.12
Open source
We consider open source the frameworks that make available the source code under certain licenses
that are not paid for in any way.
References in this category:
Roche2012
Quesnel2005
Andersson2016
Galtier2015
Acker2015
Wetter2010
C.2
Simulator Requirements
This sub section covers the taxonomy that focuses on individual simulators’ capabilities.
128
C.2.1
Information Exposed
Frequency of State The instantaneous frequency of the state of the sub-system can be used to
adjust the communication step size.
Frequency of Outputs The instantaneous frequency of the output of the sub-system can be
used to adjust the communication step size, as is done in [29].
References in this category:
Benedikt2013b
Detailed Model Simulators that make the equations of the dynamic system available fall into
this category.
References in this category:
Schmoll2012
Hassairi2012
Schierz2012
Zhang2014
BenKhaled2012
BenKhaled2014
Yamaura2016
Arnold2001
Arnold2010
Krammer2015
Fey1997
Nominal Values of Outputs This information indicates the order of magnitude of output signals.
Nominal Values of State This information indicates the order of magnitude of state signals.
I/O Signal Kind The kind of output signal helps the master algorithm understand what assumptions are in a signal.
References in this category:
Kuhr2013
Time Derivative
Output References in this category:
Hoepfer2011
Tomulik2011
Gunther2012
Schierz2012a
Carstens2003
Quesnel2005
129
State References in this category:
Hoepfer2011
Jacobian
Output References in this category:
Tomulik2011
Bastian2011a
Busch2012
Schierz2012
Viel2014
Sicklinger2014
Arnold2001
Arnold2010
Schweizer2016
Schweizer2015
Schweizer2015a
State
Discontinuity Indicator A discontinuity indicator is a signal that indicates the presence of a
discontinuity in the output of the simulation unit.
Deadreckoning model A deadreckoning model is a function that can be used by other simulation
units to extrapolate the behavior of this simulation unit.
Preferred Step Size References in this category:
Acker2015
Next Step Size The next step size indicates the next communication time that is appropriate
for the current simulator.
References in this category:
Lin2011
Gunther2012
Eyisi2012
Riley2011
Broman2013
Kounev2015
Quesnel2005
Order of Accuracy The order of accuracy can be used to determine the appropriate input
extrapolation functions to be used in a co-simulation scenario.
130
I/O Causality Feedthrough
References in this category:
Broman2013
Bogomolov2015
BenKhaled2012
BenKhaled2014
Saidi2016
Benedikt2016
Busch2011
Arnold2014a
Arnold2014
Acker2015
Propagation Delay
The propagation delay indicates how many micro-steps have to be performed before a change in
the input affects an output. In Simulink, this is the number of delay blocks in chain from an input
to an output.
Input Extrapolation This information denotes the kind of input extrapolation being performed
by the simulation unit.
References in this category:
Viel2014
State Variables References in this category:
Hoepfer2011
Quesnel2005
Arnold2014
Acker2015
Serialized State References in this category:
Bastian2011a
Broman2013
Viel2014
Bogomolov2015
Camus2015
Camus2016
Schierz2012a
Galtier2015
Micro-Step Outputs This information denotes the output of the simulation unit, evaluated at
each of the micro-steps.
References in this category:
Benedikt2013
Viel2014
Arnold2001
Arnold2010
131
WCET This denotes the worst case excution time.
References in this category:
Faure2011
BenKhaled2014
Saidi2016
C.2.2
Causality
Causal References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Tomulik2011
Sun2011
Bastian2011a
Friedrich2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Sicklinger2014
132
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2014
Saidi2016
Yamaura2016
Camus2015
Camus2016
Oh2016
Xie2016
Manbachi2016
Schierz2012a
Fourmigue2009
Liu2001
Carstens2003
Stettinger2014
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Galtier2015
Acker2015
Enge−Rosenblatt2011
Aslan2015
Wetter2010
A-Causal
133
C.2.3
Time Constraints
Analytic Simulation References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Tomulik2011
Sun2011
Friedrich2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Sicklinger2014
Zhang2014
Kounev2015
Bogomolov2015
Dols2016
BenKhaled2014
Saidi2016
Yamaura2016
Camus2015
Camus2016
Oh2016
134
Xie2016
Schierz2012a
Fourmigue2009
Liu2001
Carstens2003
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Krammer2015
Galtier2015
Fey1997
Acker2015
Karner2010a
Aslan2015
Wetter2010
Neema2014
Scaled Real Time Simulation
Fixed
A simulator is fixed scaled real time when it simulated time progresses according to a fixed linear
relationship with the real time.
References in this category:
Faure2011
Bian2015
BenKhaled2012
Pedersen2016
Manbachi2016
Stettinger2014
Wetter2010
Neema2014
Dynamic
135
A simulator is dynamic scaled real time when the relation between the simulated time and the
real time can be changed throughout the simulation.
References in this category:
Bombino2013
C.2.4
Rollback Support
None References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Faure2011
Sun2011
Bastian2011a
Friedrich2011
Gonzalez2011
Schmoll2012
Ni2012
Schierz2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
BenKhaled2014
136
Saidi2016
Yamaura2016
Camus2015
Camus2016
Oh2016
Xie2016
Manbachi2016
Carstens2003
Stettinger2014
Benedikt2016
Busch2011
Arnold2014a
Sadjina2016
Busch2016
Gu2001
Gu2004
Andersson2016
Acker2015
Aslan2015
Single Single rollback means that the simulation unit is capable of saving a certain state in the
past (simulated time) and revert to that state. Once reverted, the simulation unit cannot revert
further in the past.
References in this category:
Tomulik2011
Nutaro2011
Busch2012
Bombino2013
Sicklinger2014
Schierz2012a
Liu2001
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Arnold2010
Schweizer2016
Schweizer2015
Schweizer2015a
Fey1997
Multiple Multiple rollback means that the simulation unit is capable of saving a certain state
in the past (simulated time) and revert to that state. Once reverted, the simulation unit revert
further into the past as many times as necessary.
137
C.2.5
Availability
Remote References in this category:
Friedrich2011
Busch2012
Quaglia2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Fuller2013
Bombino2013
Zhao2014
Awais2013b
Awais2013a
Bian2015
Dols2016
Yamaura2016
Oh2016
Liu2001
Carstens2003
Galtier2015
Enge−Rosenblatt2011
Local References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Faure2011
Tomulik2011
Sun2011
Bastian2011a
Gonzalez2011
Nutaro2011
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Gunther2012
Al−Hammouri2012
Broman2013
Benedikt2013
Benedikt2013b
Wang2013
138
Hafner2013
Li2011c
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Sicklinger2014
Zhang2014
Kounev2015
Bogomolov2015
BenKhaled2012
BenKhaled2014
Saidi2016
Camus2015
Camus2016
Xie2016
Manbachi2016
Schierz2012a
Fourmigue2009
Stettinger2014
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Krammer2015
Galtier2015
Fey1997
Acker2015
Aslan2015
139
C.3
C.3.1
Framework Requirements
Standard
High Level Architecture References in this category:
Eyisi2012
Riley2011
Awais2013b
Awais2013a
Neema2014
Functional Mock-up Interface
Pedersen2015
Sun2011
Bastian2011a
Broman2013
Wang2013
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Bogomolov2015
Dols2016
BenKhaled2012
BenKhaled2014
Saidi2016
Camus2015
Camus2016
Pedersen2016
Schierz2012a
Arnold2014a
Arnold2014
Andersson2016
Galtier2015
Acker2015
Aslan2015
Neema2014
References in this category:
Functional Digital Mock-up References in this category:
Enge−Rosenblatt2011
C.3.2
Coupling
Input/Output Assignments References in this category:
Pedersen2015
Lin2011
Faure2011
140
Sun2011
Bastian2011a
Friedrich2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
Yamaura2016
Camus2015
Camus2016
Pedersen2016
Oh2016
Xie2016
Manbachi2016
Schierz2012a
Fourmigue2009
Liu2001
Carstens2003
141
Stettinger2014
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Sadjina2016
Busch2016
Andersson2016
Krammer2015
Galtier2015
Fey1997
Acker2015
Enge−Rosenblatt2011
Karner2010a
Aslan2015
Wetter2010
Algebraic Constraints References in this category:
Tomulik2011
Friedrich2011
Schierz2012
Viel2014
Sicklinger2014
Arnold2001
Schweizer2014
Schweizer2015d
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
C.3.3
Number of Simulation Units
Two References in this category:
Lin2011
Sun2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Quaglia2012
142
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Zhao2014
Li2011c
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
Pedersen2016
Oh2016
Xie2016
Manbachi2016
Fourmigue2009
Liu2001
Carstens2003
Stettinger2014
Schweizer2014
Gu2001
Fey1997
Three or More
Pedersen2015
Hoepfer2011
Faure2011
Tomulik2011
Bastian2011a
Friedrich2011
Schierz2012
Gunther2012
Broman2013
Hafner2013
Awais2013b
Awais2013a
Kuhr2013
Viel2014
References in this category:
143
Sicklinger2014
BenKhaled2012
BenKhaled2014
Saidi2016
Yamaura2016
Camus2015
Camus2016
Schierz2012a
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Krammer2015
Galtier2015
Acker2015
Enge−Rosenblatt2011
Karner2010a
Aslan2015
Wetter2010
Neema2014
C.3.4
Domain
CT References in this category:
Pedersen2015
Hoepfer2011
Faure2011
Tomulik2011
Sun2011
Bastian2011a
Friedrich2011
Gonzalez2011
Busch2012
Schmoll2012
Ni2012
144
Hassairi2012
Schierz2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Roche2012
Fitzgerald2010
Fitzgerald2013
Broman2013
Benedikt2013
Benedikt2013b
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Kuhr2013
Viel2014
Sicklinger2014
Zhang2014
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
BenKhaled2014
Saidi2016
Yamaura2016
Pedersen2016
Oh2016
Xie2016
Schierz2012a
Liu2001
Carstens2003
Stettinger2014
Benedikt2016
Busch2011
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2001
Gu2004
145
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Krammer2015
Galtier2015
Fey1997
Acker2015
Enge−Rosenblatt2011
Karner2010a
Aslan2015
Wetter2010
DE References in this category:
Lin2011
Nutaro2011
Al−Hammouri2012
Eyisi2012
Riley2011
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Fuller2013
Awais2013b
Awais2013a
Kuhr2013
Zhang2014
Kounev2015
Bogomolov2015
Camus2015
Camus2016
Fourmigue2009
Quesnel2005
Fey1997
Karner2010a
Neema2014
C.3.5
Dynamic structure
References in this category:
Karner2010a
C.3.6
Co-simulation Rate
Single References in this category:
Pedersen2015
Lin2011
146
Hoepfer2011
Faure2011
Tomulik2011
Sun2011
Bastian2011a
Friedrich2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Viel2014
Sicklinger2014
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
BenKhaled2014
Yamaura2016
Pedersen2016
Oh2016
Xie2016
Manbachi2016
Schierz2012a
Carstens2003
147
Stettinger2014
Benedikt2016
Busch2016
Arnold2010
Galtier2015
Aslan2015
Wetter2010
Multi Multi-rate co-simulation denotes that the framework distinguishes between slow and fast
sub-systems and dimensions the communication step size accordingly, providing for interpolation/extrapolation of the slow systems.
References in this category:
Gonzalez2011
Awais2013b
Awais2013a
Kuhr2013
Camus2015
Camus2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Fey1997
Acker2015
Enge−Rosenblatt2011
Neema2014
C.3.7
Communication Step Size
Fixed References in this category:
Pedersen2015
Faure2011
Tomulik2011
Bastian2011a
Friedrich2011
Gonzalez2011
Busch2012
148
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Quaglia2012
Roche2012
Kudelski2013
Benedikt2013
Benedikt2013b
Bombino2013
Hafner2013
Zhao2014
Li2011c
Awais2013a
Viel2014
Sicklinger2014
Zhang2014
Bian2015
Dols2016
BenKhaled2012
BenKhaled2014
Pedersen2016
Oh2016
Xie2016
Manbachi2016
Carstens2003
Stettinger2014
Arnold2014a
Arnold2001
Schweizer2014
Schweizer2015d
Busch2016
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Fey1997
Acker2015
Wetter2010
Neema2014
149
Variable References in this category:
Lin2011
Hoepfer2011
Sun2011
Nutaro2011
Gunther2012
Al−Hammouri2012
Eyisi2012
Riley2011
Fitzgerald2010
Fitzgerald2013
Broman2013
Fuller2013
Wang2013
Awais2013b
Kuhr2013
Kounev2015
Bogomolov2015
Camus2015
Camus2016
Schierz2012a
Benedikt2016
Busch2011
Quesnel2005
Arnold2014
Sadjina2016
Galtier2015
Neema2014
C.3.8
Strong Coupling Support
None – Explicit Method References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Faure2011
Sun2011
Friedrich2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Hassairi2012
Schierz2012
Gunther2012
150
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Kudelski2013
Broman2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
BenKhaled2014
Saidi2016
Yamaura2016
Camus2015
Camus2016
Pedersen2016
Oh2016
Xie2016
Manbachi2016
Schierz2012a
Carstens2003
Stettinger2014
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Sadjina2016
Busch2016
Gu2001
Gu2004
151
Galtier2015
Aslan2015
Partial – Semi-Implicit Method References in this category:
Busch2012
Schweizer2014
Schweizer2015d
Schweizer2016
Schweizer2015
Schweizer2015a
Full – Implicit Method References in this category:
Tomulik2011
Bastian2011a
Busch2012
Viel2014
Sicklinger2014
Liu2001
Arnold2001
Arnold2010
Acker2015
C.3.9
Results Visualization
Postmortem The results are available after the simulation. References in this category:
Pedersen2015
Lin2011
Hoepfer2011
Faure2011
Tomulik2011
Sun2011
Bastian2011a
Friedrich2011
Gonzalez2011
Nutaro2011
Busch2012
Schmoll2012
Ni2012
Schierz2012
Gunther2012
Quaglia2012
Al−Hammouri2012
Roche2012
Kudelski2013
Broman2013
Benedikt2013
152
Benedikt2013b
Fuller2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Viel2014
Sicklinger2014
Zhang2014
Kounev2015
Bogomolov2015
Bian2015
Dols2016
BenKhaled2012
Camus2015
Camus2016
Xie2016
Schierz2012a
Liu2001
Carstens2003
Stettinger2014
Benedikt2016
Busch2011
Quesnel2005
Arnold2014a
Arnold2014
Arnold2001
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Arnold2010
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Krammer2015
Galtier2015
Fey1997
Acker2015
Karner2010a
Aslan2015
153
Wetter2010
Live References in this category:
Hassairi2012
Eyisi2012
Riley2011
Fitzgerald2010
Fitzgerald2013
Yamaura2016
Pedersen2016
Enge−Rosenblatt2011
Neema2014
Interactive References in this category:
Bombino2013
Yamaura2016
C.3.10
Communication Approach
Jacobi References in this category:
Pedersen2015
Hoepfer2011
Faure2011
Tomulik2011
Bastian2011a
Friedrich2011
Schmoll2012
Schierz2012
Gunther2012
Kudelski2013
Broman2013
Hafner2013
Awais2013b
Awais2013a
Sicklinger2014
Bogomolov2015
BenKhaled2012
Saidi2016
Xie2016
Manbachi2016
Schierz2012a
Fourmigue2009
Stettinger2014
Busch2011
Arnold2014a
Arnold2014
154
Schweizer2014
Schweizer2015d
Sadjina2016
Busch2016
Gu2001
Gu2004
Schweizer2016
Schweizer2015
Schweizer2015a
Andersson2016
Galtier2015
Wetter2010
Neema2014
Gauss-Seidel References in this category:
Lin2011
Hoepfer2011
Sun2011
Bastian2011a
Gonzalez2011
Nutaro2011
Busch2012
Ni2012
Hassairi2012
Schierz2012
Quaglia2012
Al−Hammouri2012
Eyisi2012
Riley2011
Roche2012
Fitzgerald2010
Fitzgerald2013
Benedikt2013
Benedikt2013b
Fuller2013
Bombino2013
Wang2013
Hafner2013
Zhao2014
Li2011c
Awais2013b
Awais2013a
Kuhr2013
Viel2014
Sicklinger2014
Kounev2015
155
Dols2016
BenKhaled2012
BenKhaled2014
Camus2015
Camus2016
Pedersen2016
Oh2016
Carstens2003
Stettinger2014
Quesnel2005
Arnold2014
Arnold2001
Busch2016
Arnold2010
Acker2015
Aslan2015
156
D
List of Acronyms
CPS
CT
DE
DEVS
DTSS
FMI
FR
GVT
IP
IVP
NFR
ODE
SR
Cyber-Physical System
Continuous Time
Discrete Event
Discrete Event System Specification
Discrete Time System Specification
Functional Mock-up Interface
Framework Requirement
Global Virtual Time
Intellectual Property
Initial Value Problem
Non-Functional Requirement
Ordinary Differential Equation
Simulator Requirement
157
| 3cs.SY
|
arXiv:cs/0401016v3 [cs.LO] 14 Mar 2006
Generalized Strong Preservation
by Abstract Interpretation
F RANCESCO R ANZATO F RANCESCO TAPPARO
Dipartimento di Matematica Pura ed Applicata, Università di Padova
Via Belzoni 7, 35131 Padova, Italy
[email protected] [email protected]
Abstract
Standard abstract model checking relies on abstract Kripke structures which approximate concrete
models by gluing together indistinguishable states, namely by a partition of the concrete state space.
Strong preservation for a specification language L encodes the equivalence of concrete and abstract
model checking of formulas in L . We show how abstract interpretation can be used to design abstract
models that are more general than abstract Kripke structures. Accordingly, strong preservation is generalized to abstract interpretation-based models and precisely related to the concept of completeness in
abstract interpretation. The problem of minimally refining an abstract model in order to make it strongly
preserving for some language L can be formulated as a minimal domain refinement in abstract interpretation in order to get completeness w.r.t. the logical/temporal operators of L . It turns out that this refined
strongly preserving abstract model always exists and can be characterized as a greatest fixed point. As
a consequence, some well-known behavioural equivalences, like bisimulation, simulation and stuttering,
and their corresponding partition refinement algorithms can be elegantly characterized in abstract interpretation as completeness properties and refinements.
Keywords: Abstract interpretation, abstract model checking, strong preservation, completeness, refinement, behavioural equivalence.
1 Introduction
The design of an abstract model checking framework always includes a preservation result, roughly stating
that for any formula ϕ specified in some temporal language L , if ϕ holds on an abstract model then ϕ
also holds on the concrete model. On the other hand, strong preservation means that a formula of L holds
on an abstract model if and only if it holds on the concrete model. Strong preservation is highly desirable
since it allows to draw consequences from negative answers on the abstract side [10].
Generalized Strong Preservation. The relationship between abstract interpretation and abstract model
checking has been the subject of a number of works (see e.g. [9, 11, 16, 17, 19, 20, 28, 39, 40, 41, 43, 47]).
This paper follows the standard abstract interpretation approach [13, 14] where abstract domains are specified by Galois connections, namely pairs of abstraction and concretization maps α/γ. We deal with generic
(temporal) languages L of state formulae that are inductively generated by some given sets of atomic
propositions and operators. The interpretation p of atomic propositions p ∈ AP as subsets of States and
of operators f ∈ Op as mappings f on ℘(States) is determined by a suitable semantic structure S, e.g. a
Kripke structure, so that the concrete semantics JϕKS ∈ ℘(States) of a formula ϕ ∈ L is the set of states
making ϕ true w.r.t. S. Abstract semantics can be systematically defined by standard abstract interpretation. The powerset ℘(States) plays the role of concrete semantic domain so that abstract domains range
in AbsDom(℘(States)). Any abstract domain A ∈ AbsDom(℘(States)) induces an abstract semantic
structure S A where atoms p are abstracted to α(p) and operators f are interpreted as best correct approximations on A, that is α◦ f ◦ γ. Thus, A determines an abstract semantics JϕKA
S ∈ A that evaluates formulae
ϕ ∈ L in the abstract domain A.
It turns out that this approach generalizes standard abstract model checking [9, 10]. Given a Kripke
1
structure K = (States, →) (for simplicity we omit here a labeling function for atomic propositions), a
standard abstract model is specified as an abstract Kripke structure A = (AStates, →♯ ) where the set
AStates of abstract states is defined by a surjective map h : States → AStates. Thus, AStates determines a partition of States and vice versa. It turns out that state partitions are particular abstract domains.
In fact, the lattice of partitions of States is an abstract interpretation of the lattice of abstract domains
AbsDom(℘(States)) so that the abstract state space AStates corresponds to a particular abstract domain
ad(AStates) ∈ AbsDom(℘(States)). Abstract domains that can be derived from a state partition are
called partitioning. The interpretation of the language L w.r.t. the abstract Kripke structure A determines
an abstract semantic function JϕKA ∈ AStates. The abstract Kripke structure A strongly preserves L
when for any ϕ ∈ L and s ∈ States, it turns out that h(s) ∈ JϕKA ⇔ s ∈ JϕKK .
Strong preservation can then be generalized from standard abstract models to abstract interpretationbased models. Given a generalized abstract model A ∈ AbsDom(℘(States)), the induced abstract semanA
tics J·KA
S is strongly preserving for L when for any ϕ ∈ L and S ∈ ℘(States), α(S) ≤A JϕKS ⇔ S ⊆
♯
JϕKS . It turns out that this is an abstract domain property, because any abstract semantics J·K : L → A that
evaluates formulae in the abstract domain A is strongly preserving for L if and only if J·KA
S is. Standard
strong preservation becomes a particular instance, namely an abstract Kripke structure strongly preserves
L if and only if the corresponding partitioning abstract model strongly preserves L . On the other hand,
generalized strong preservation may work where standard strong preservation may fail, namely it may happen that although a strongly preserving abstract semantics on a partitioning abstract model ad(AStates)
exists this cannot be derived from a strongly preserving abstract Kripke structure on AStates.
Generalized Strong Preservation and Complete Abstract Interpretations. Given a language L and
a Kripke structure K = (States, →), a well-known key problem is to compute the smallest abstract
state space AStates L , when this exists, such that one can define an abstract Kripke structure AL =
(AStates L , →♯ ) that strongly preserves L . This problem admits solution for a number of well-known
temporal languages like CTL (or, equivalently, the µ-calculus), ACTL and CTL-X (i.e. CTL without the
next-time operator X). A number of algorithms for solving this problem exist, like those by Paige and
Tarjan [42] for CTL, by Henzinger et al. [35], Bustan and Grumberg [5] and Tan and Cleaveland [48]
for ACTL, and Groote and Vaandrager [32] for CTL-X. These are coarsest partition refinement algorithms: given a language L and a partition P of States, which is determined by a state labeling, these
algorithms can be viewed as computing the coarsest partition PL that refines P and strongly preserves
L . It is worth remarking that most of these algorithms have been designed for computing well-known
behavioural equivalences used in process algebra like bisimulation (for CTL), simulation (for ACTL) and
divergence-blind stuttering (for CTL-X) equivalence. Our abstract interpretation-based framework allows
to give a generalized view of the above partition refinement algorithms. We show that the most abstract domain ADL ∈ AbsDom(℘(States)) that strongly preserves a given language L always exists. It turns out
that ADL is a partitioning abstract domain if and only if L includes full propositional logic, that is when
L is closed under logical conjunction and negation. Otherwise, a proper loss of information occurs when
abstracting ADL to the corresponding partition PL . Moreover, for some languages L , it may happen that
one cannot define an abstract Kripke structure on the abstract state space PL that strongly preserves L
whereas the most abstract strongly preserving semantics in AbsDom(℘(States)) instead exists.
The concept of complete abstract interpretation is well known [14, 31]. This encodes an ideal situation where the abstract semantics coincides with the abstraction of the concrete semantics. We establish a precise correspondence between generalized strong preservation of abstract models and completeness in abstract interpretation. Our results are based on the notion of forward complete abstract domain.
An abstract domain A is forward complete for a concrete semantic function f when for any a ∈ A,
f (γ(a)) = γ(α(f (γ(a)))), namely when no loss of precision occurs by approximating in A a computation f (γ(a)). This notion of forward completeness is dual and orthogonal to the standard definition of
completeness in abstract interpretation. Giacobazzi et al. [31] showed how complete abstract domains can
be systematically and constructively derived from noncomplete abstract domains by minimal refinements.
This can be done for forward completeness as well. Given any domain A, the most abstract domain that
refines A and is forward complete for f does exist and can be characterized as a greatest fixpoint. Such a
domain is called the forward complete shell of A for f . It turns out that strong preservation is related to forward completeness as follows. As described above, the most abstract domain ADL that strongly preserves
2
L always exists. It turns out that ADL coincides with the forward complete shell for the operators of L
of a basic abstract domain determined by the state labeling. This characterization provides an elegant generalization of partition refinement algorithms used in standard abstract model checking. As a consequence
of these results, we derive a novel characterization of the corresponding behavioural equivalences in terms
of forward completeness of abstract domains. For example, it turns out that a partition P is a bisimulation
on some Kripke structure K if and only if the corresponding partitioning abstract domain ad(P ) is forward
complete for the standard predecessor transformer pre→ in K.
2 Basic Notions
2.1 Notation and Preliminaries
Let X be any set. Fun(X) denotes the set of functions f : X n → X, for some n ≥ 0, called arity of
f . Following a standard convention, when n = 0, f is meant to be a specific object of X. The arity
of f is also denoted by ♯(f ) ≥ 0. id denotes the identity map. If F ⊆ Fun(X) and Y ⊆ X then
def
F (Y ) = {f (~y) | f ∈ F, ~
y ∈ Y ♯(f ) }, namely F (Y ) is the set of images of Y for each function in F . If
f : X → Y then the image of f is also denoted by img(f ) = {f (x) ∈ Y | x ∈ X}. If f : X → Y
and g : Y → Z then g ◦ f : X → Z denotes the composition of f and g, i.e. g ◦ f = λx.g(f (x)). The
complement operator for the universe set X is ∁ : ℘(X) → ℘(X), where ∁(S) = X r S. When writing
a set S of subsets of a given set, like a partition, we often write S in a compact form like {1, 12, 13} or
{[1], [12], [13]} that stand for {{1}, {1, 2}, {1, 3}}. Ord denotes the proper class of ordinals and ω ∈ Ord
denotes the first infinite ordinal.
Let hP, ≤i be a poset. Posets are often denoted also by P≤ . We use the symbol ⊑ to denote pointwise
ordering between functions: If X is any set and f, g : X → P then f ⊑ g if for all x ∈ X, f (x) ≤ g(x).
A mapping f : P → Q on posets is continuous when f preserves least upper bounds (lub’s) of countable
chains in P , while, dually, it is co-continuous when f preserves greatest lower bounds (glb’s) of countable
chains in P . A complete lattice C≤ is also denoted by hC, ≤, ∨, ∧, ⊤, ⊥i where ∨, ∧, ⊤ and ⊥ denote,
respectively, lub, glb, greatest element and least element in C. A mapping f : C → D between complete
lattices is additive (co-additive) when for any Y ⊆ C, f (∨C Y ) = ∨D f (Y ) (f (∧C Y ) = ∧D f (Y )). We
denote by lfp(f ) and gfp(f ), respectively, the least and greatest fixpoint, when they exist, of an operator f
on a poset. The well-known Knaster-Tarski’s theorem states that any monotone operator f : C → C on a
complete lattice C admits a least fixpoint and the following characterization holds:
lfp(f ) = ∧{x ∈ C | f (x) ≤ x} = ∨α∈Ord f α,↑ (⊥)
where the upper iteration sequence {f α,↑ (x)}α∈Ord of f in x ∈ C is defined by transfinite induction on α
as usual:
– α = 0: f 0,↑ (x) = x;
– successor ordinal α = β + 1: f β+1,↑ (x) = f (f β,↑ (x));
– limit ordinal α: f α,↑ (x) = ∨β<α f β,↑ (x).
It is well known that if f is continuous then lfp(f ) = ∨α∈ω f α,↑ (⊥). Dually, f also admits a greatest
fixpoint and the following characterization holds:
gfp(f ) = ∨{x ∈ C | x ≤ f (x)} = ∧α∈Ord f α,↓ (⊤),
where the lower iteration sequence {f α,↓ (x)}α∈Ord of f in x ∈ C is defined as the upper iteration sequence
but for the case of limit ordinals: f α,↓ (x) = ∧β<α f β,↓ (x).
Let Σ be any set. PreOrd(Σ) denotes the set of preorder relations on Σ, that is R ⊆ Σ × Σ is a
preorder on Σ if R is reflexive and transitive. Part(Σ) denotes the set of partitions of Σ. Sets in a partition
P are called blocks of P . If ≡ ⊆ Σ × Σ is an equivalence relation then we denote by P≡ ∈ Part(Σ) the
corresponding partition of Σ. Vice versa, if P ∈ Part(Σ) then ≡P ⊆ Σ × Σ denotes the corresponding
3
equivalence relation on Σ. Part(Σ) is endowed with the following standard partial order 4: P1 4 P2 , i.e.
P2 is coarser than P1 (or P1 refines P2 ) iff ∀B ∈ P1 .∃B ′ ∈ P2 .B ⊆ B ′ . It is well known that hPart(Σ), 4i
is a complete lattice.
A transition system T = (Σ, ) consists of a (possibly infinite) set Σ of states and a transition relation
⊆ Σ × Σ. As usual [10], we assume that the relation is total, i.e., for any s ∈ Σ there exists some
t ∈ Σ such that st, so that any maximal path in T is necessarily infinite. T is finitely branching when for
any s ∈ Σ, {t ∈ Σ | st} is a finite set. The pre/post transformers on ℘(Σ) are defined as usual:
def
– pre = λY.{a ∈ Σ | ∃b ∈ Y. ab};
def
– pre
f = ∁ ◦ pre ◦∁ = λY.{a ∈ Σ | ∀b ∈ Σ.(ab ⇒ b ∈ Y )};
def
– post = λY.{b ∈ Σ | ∃a ∈ Y. ab};
g def
– post
= ∁ ◦ post ◦∁ = λY.{b ∈ Σ | ∀a ∈ Σ.(ab ⇒ a ∈ Y )}.
g are co-additive.
Let us observe that pre and post are additive operators on ℘(Σ)⊆ while pre
f and post
∃∃
∀∃
If R ⊆ Σ1 × Σ2 is any relation then the relations R , R ⊆ ℘(Σ1 ) × ℘(Σ2 ) are defined as follows:
– (S1 , S2 ) ∈ R∃∃ iff ∃s1 ∈ S1 .∃s2 ∈ S2 . (s1 , s2 ) ∈ R;
– (S1 , S2 ) ∈ R∀∃ iff ∀s1 ∈ S1 .∃s2 ∈ S2 . (s1 , s2 ) ∈ R.
2.2 Abstract Interpretation and Completeness
2.2.1 Abstract Domains
In standard Cousot and Cousot’s abstract interpretation, abstract domains can be equivalently specified
either by Galois connections, i.e. adjunctions, or by upper closure operators (uco’s) [13, 14]. Let us recall
these standard notions.
Galois Connections and Insertions. If A and C are posets and α : C → A and γ : A → C are
monotone functions such that ∀c ∈ C. c ≤C γ(α(c)) and α(γ(a)) ≤A a then the quadruple (α, C, A, γ) is
called a Galois connection (GC for short) between C and A. If in addition α ◦ γ = λx.x then (α, C, A, γ)
is a Galois insertion (GI for short) of A in C. In a GI, γ is 1-1 and α is onto. Let us also recall that the
notion of GC is equivalent to that of adjunction: if α : C → A and γ : A → C then (α, C, A, γ) is a GC
iff ∀c ∈ C.∀a ∈ A. α(c) ≤A a ⇔ c ≤C γ(a). The map α (γ) is called the left- (right-) adjoint to γ (α).
It turns out that one adjoint map α/γ uniquely determines the other adjoint map γ/α as follows. On the
one hand, a map α : C → A admits a necessarily unique right-adjoint map γ : A → C iff α preserves
def
arbitrary lub’s; in this case, we have that γ = λa. ∨C {c ∈ C | α(c) ≤A a}. On the other hand, a map
γ : A → C admits a necessarily unique left-adjoint map α : C → A iff γ preserves arbitrary glb’s; in
def
this case, α = λc. ∧A {a ∈ A | c ≤C γ(a)}. In particular, we have that in any GC (α, C, A, γ) between
complete lattices it turns out that α is additive and γ is co-additive. Also, if (α, C, A, γ) is a GI and C is a
complete lattice then A is a complete lattice as well and hA, ≤A i is order-isomorphic to himg(γ), ≤C i.
We assume the standard abstract interpretation framework, where concrete and abstract domains, C and
A, are complete lattices related by abstraction and concretization maps α and γ forming a GC (α, C, A, γ).
A is called an abstraction of C and C a concretization of A. The ordering relations on concrete and abstract
domains describe the relative precision of domain values: x ≤ y means that y is an approximation of x or,
equivalently, x is more precise than y. Galois connections allow to relate the concrete and abstract notions
of relative precision: an abstract value a ∈ A approximates a concrete value c ∈ C when α(c) ≤A a, or,
equivalently (by adjunction), c ≤C γ(a). As a key consequence of requiring a Galois connection, it turns
out that α(c) is the best possible approximation in A of c, that is α(c) = ∧{a ∈ A | c ≤C γ(a)} holds. If
(α, C, A, γ) is a GI then each value of the abstract domain A is useful in representing C, because all the
values in A represent distinct members of C, being γ 1-1. Any GC can be lifted to a GI by identifying in
an equivalence class those values of the abstract domain with the same concretization. Abs(C) denotes the
set of abstract domains of C and we write A ∈ Abs(C) to mean that the abstract domain A is related to
4
C through a GI (α, C, A, γ). An abstract domain A is disjunctive when the corresponding concretization
map γ is additive.
Closure Operators. An (upper) closure operator, or simply a closure, on a poset P≤ is an operator
µ : P → P that is monotone, idempotent and extensive, i.e., ∀x ∈ P. x ≤ µ(x). Dually, lower closure
operators are monotone, idempotent, and restrictive, i.e., ∀x ∈ P. µ(x) ≤ x. uco(P ) denotes the set of
closure operators on P . Let hC, ≤, ∨, ∧, ⊤, ⊥i be a complete lattice. A closure µ ∈ uco(C) is uniquely
determined by its image img(µ), which coincides with its set of fixpoints, as follows: µ = λy. ∧ {x ∈
img(µ) | y ≤ x}. Also, X ⊆ C is the image of some closure operator µX on C iff X is a Moore-family of
def
C, i.e., X = M(X) = {∧S | S ⊆ X} — where ∧∅ = ⊤ ∈ M(X). In other terms, X is a Moore-family
of C when X is meet-closed. In this case, µX = λy. ∧ {x ∈ X | y ≤ x} is the corresponding closure
operator on C. For any X ⊆ C, M(X) is called the Moore-closure of X in C, i.e., M(X) is the least
(w.r.t. set inclusion) subset of C which contains X and is a Moore-family of C. Moreover, it turns out
that for any µ ∈ uco(C) and any Moore-family X ⊆ C, µimg(µ) = µ and img(µX ) = X. Thus, closure
operators on C are in bijection with Moore-families of C. This allows us to consider a closure operator
µ ∈ uco(C) both as a function µ : C → C and as a Moore-family img(µ) ⊆ C. This is particularly
useful and does not give rise to ambiguity since one can distinguish the use of a closure µ as function or
set according to the context.
It turns out that hµ, ≤i is a complete meet subsemilattice of C, i.e. ∧ is its glb, but, in general, it is not
a complete sublattice of C, since the lub in µ — defined by λY ⊆ µ. µ(∨Y ) — might be different from
that in C. In fact, it turns out that µ is a complete sublattice of C (namely, img(µ) is also join-closed) iff µ
is additive.
If C is a complete lattice then uco(C) endowed with the pointwise ordering ⊑ is a complete lattice
denoted by huco(C), ⊑, ⊔, ⊓, λx.⊤, λx.xi, where for every µ, η ∈ uco(C), {µi }i∈I ⊆ uco(C) and x ∈ C:
– µ ⊑ η iff ∀y ∈ C. µ(y) ≤ η(y) iff img(η) ⊆ img(µ);
– (⊓i∈I µi )(x) = ∧i∈I µi (x);
– x ∈ ⊔i∈I µi ⇔ ∀i ∈ I. x ∈ µi ;
– λx.⊤ is the greatest element, whereas λx.x is the least element.
Thus, the glb in uco(C) is defined pointwise, while the lub of a set of closures {µi }i∈I ⊆ uco(C) is the
closure whose image is given by the set-intersection ∩i∈I µi .
The Lattice of Abstract Domains. It is well known since [14] that abstract domains can be equivalently
specified either as Galois insertions or as closures. These two approaches are completely equivalent. On the
one hand, if µ ∈ uco(C) and A is a complete lattice which is isomorphic to img(µ), where ι : img(µ) → A
and ι−1 : A → img(µ) provide the isomorphism, then (ι ◦ µ, C, A, ι−1 ) is a GI. On the other hand, if
def
(α, C, A, γ) is a GI then µA = γ ◦ α ∈ uco(C) is the closure associated with A such that himg(µA ), ≤C i
is a complete lattice which is isomorphic to hA, ≤A i. Furthermore, these two constructions are inverse of
each other. Let us also remark that an abstract domain A is disjunctive iff µA is additive. Given an abstract
domain A specified by a GI (α, C, A, γ), its associated closure γ ◦ α on C can be thought of as the “logical
meaning” of A in C, since this is shared by any other abstract representation for the objects of A. Thus, the
closure operator approach is particularly convenient when reasoning about properties of abstract domains
independently from the representation of their objects.
Abstract domains specified by GIs can be pre-ordered w.r.t. precision as follows: if A1 , A2 ∈ Abs(C)
then A1 is more precise (or concrete) than A2 (or A2 is an abstraction of A1 ), denoted by A1 A2 , when
µA1 ⊑ µA2 . The pointwise ordering ⊑ between uco’s corresponds therefore to the standard ordering used
to compare abstract domains with respect to their precision. Also, A1 and A2 are equivalent, denoted by
A1 ≃ A2 , when their associated closures coincide, i.e. µA1 = µA2 . Hence, the quotient Abs(C)/≃ gives
rise to a poset that, by a slight abuse of notation, is simply denoted by hAbs(C), ⊑i. Thus, when we write
A ∈ Abs(C) we mean that A is any representative of an equivalence class in Abs(C)/≃ and is specified by
5
a Galois insertition (α, C, A, γ). It turns out that hAbs(C), ⊑i is a complete lattice, called the lattice of abstract interpretations of C [13, 14], because it is isomorphic to the complete lattice huco(C), ⊑i. Lub’s and
glb’s in Abs(C) have therefore the following reading as operators on domains. Let {Ai }i∈I ⊆ Abs(C):
(i) ⊔i∈I Ai is the most concrete among the domains which are abstractions of all the Ai ’s; (ii) ⊓i∈I Ai is
the most abstract among the domains which are more concrete than every Ai — this latter domain is also
known as reduced product of all the Ai ’s.
2.2.2 Completeness
Let C be a concrete domain, f : C → C be a concrete semantic function1 and let f ♯ : A → A be
a corresponding abstract function on an abstract domain A ∈ Abs(C) specified by a GI (α, C, A, γ).
Then, hA, f ♯ i is a sound abstract interpretation when α ◦ f ⊑ f ♯ ◦ α holds. The abstract function f ♯ is
called a correct approximation on A of f . This means that a concrete computation f (c) can be correctly
approximated in A by f ♯ (α(c)), namely α(f (c)) ≤A f ♯ (α(c)). An abstract function f1♯ : A → A is more
precise than f2♯ : A → A when f1♯ ⊑ f2♯ . Since α ◦ f ⊑ f ♯ ◦ α holds iff α ◦ f ◦ γ ⊑ f ♯ holds, the abstract
def
function f A = α ◦ f ◦ γ : A → A is called the best correct approximation of f in A.
Completeness in abstract interpretation corresponds to requiring that, in addition to soundness, no loss
of precision occurs when f (c) is approximated in A by f ♯ (α(c)). Thus, completeness of f ♯ for f is encoded
by the equation α ◦ f = f ♯ ◦ α. This is also called backward completeness because a dual form of forward
completeness may be considered. As a very simple example, let us consider the abstract domain Sign
representing the sign of an integer variable, namely Sign = {⊥, Z<0 , 0, Z>0 , ⊤} ∈ Abs(℘(Z)⊆ ). Let us
def
consider the binary concrete operation of integer addition on sets of integers, that is X + Y = {x + y | x ∈
2
2 def
X, y ∈ Y }, and the square operator on sets of integers, that is X = {x | x ∈ X}. It turns out that
the best correct approximation +Sign of integer addition in Sign is sound but not complete — because
α({−1} + {1}) = 0 <Sign ⊤ = α({−1})+Sign α({1}) — while it is easy to check that the best correct
approximation of the square operation in Sign is instead complete.
A dual form of completeness may be considered. The soundness condition α ◦ f ⊑ f ♯ ◦ α can be
equivalently formulated as f ◦ γ ⊑ γ ◦ f ♯ . Forward completeness for f ♯ corresponds to requiring that
the equation f ◦ γ = γ ◦ f ♯ holds, and therefore means that no loss of precision occurs when a concrete
computation f (γ(a)), for some abstract value a ∈ A, is approximated in A by f ♯ (a). Let us notice
that backward and forward completeness are orthogonal concepts. In fact: (1) as observed above, we
have that +Sign is not backward complete while it is forward complete because for any a1 , a2 ∈ Sign,
γ(a1 ) + γ(a2 ) = γ(a1 +Sign a2 ); (2) the best correct approximation (·)2Sign of the square operator on Sign
is not forward complete because γ(Z>0 )2 ( γ(Z>0 ) = γ((Z>0 )2Sign ) while, as observed above, it is
instead backward complete.
Giacobazzi et al. [31] observed that completeness uniquely depends upon the abstraction map, i.e. upon
the abstract domain: this means that if f ♯ is backward complete for f then the best correct approximation
f A of f in A is backward complete as well, and, in this case, f ♯ indeed coincides with f A . Hence, for any
abstract domain A, one can define a backward complete abstract operation f ♯ on A if and only if f A is
backward complete. Thus, an abstract domain A ∈ Abs(C) is defined to be backward complete for f iff the
equation α ◦ f = f A ◦ α holds. This simple observation makes backward completeness an abstract domain
property, namely an intrinsic characteristic of the abstract domain. Let us observe that α ◦ f = f A ◦ α
holds iff γ ◦ α ◦ f = γ ◦ f A ◦ α = γ ◦ α ◦ f ◦ γ ◦ α holds, so that A is backward complete for f when
µA ◦ f = µA ◦ f ◦ µA . Thus, a closure µ ∈ uco(C), that defines some abstract domain, is backward
complete for f when µ ◦ f = µ ◦ f ◦ µ holds. Analogous observations apply to forward completeness,
which is also an abstract domain property: A ∈ Abs(C) is forward complete for f (or forward f -complete)
when f ◦ µA = µA ◦ f ◦ µA , while a closure µ ∈ uco(C) is forward complete for f when f ◦ µ = µ ◦ f ◦ µ
holds.
Let us also recall that, by a well-known result (see, e.g., [14, Theorem 7.1.0.4], [1, Fact 2.3] and [21,
Lemma 4.3]), backward complete abstract domains are “fixpoint complete” as well. This means that if
A ∈ Abs(C) is backward complete for a concrete monotone function f : C → C then α(lfp(f )) =
lfp(f A ). Moreover, if α and f are both co-continuous then this also holds for greatest fixpoints, namely
1 For
simplicity of notation we consider here unary functions since the extension to generic n-ary functions is straightforward.
6
α(gfp(f )) = gfp(f A ). As far as forward completeness is concerned, the following result holds.
Lemma 2.1. If A ∈ Abs(C) is forward complete for a monotone f then α(gfp(f )) = gfp(f A ). Moreover,
if γ and f are both continuous and γ(⊥A ) = ⊥C then α(lfp(f )) = lfp(f A ).
Proof. Let us show that α(gfp(f )) = gfp(f A ). On the one hand, since gfp(f ) ≤ γ(α(gfp(f ))), we
have that gfp(f ) = f (gfp(f )) ≤ f (γ(α(gfp(f )))), therefore, by using forward completeness, gfp(f ) ≤
γ(f A (α(gfp(f )))). Thus, α(gfp(f )) ≤ f A (α(gfp(f ))), from which follows that α(gfp(f )) ≤ gfp(f A ).
On the other hand, by using forward completeness, f (γ(gfp(f A ))) = γ(f A (gfp(f A ))) = γ(gfp(f A )),
so that γ(gfp(f A )) ≤ gfp(f ), and therefore, by applying α, we obtain that gfp(f A ) = α(γ(gfp(f A ))) ≤
α(gfp(f )).
Assume now that γ and f are both continuous and γ(⊥A ) = ⊥C . Let us show by induction on k that for
any k ∈ N, γ((f A )k,↑ (⊥A )) = f k,↑ (⊥C ).
(k = 0): By hypothesis, γ((f A )0,↑ (⊥A )) = γ(⊥A ) = ⊥C = f 0,↑ (⊥C ).
(k + 1):
γ((f A )k+1,↑ (⊥A )) =
γ(f A ((f A )k,↑ (⊥A ))) =
A k,↑
f (γ((f )
f (f
(⊥A ))) =
k,↑
[by forward completeness]
[by inductive hypothesis]
(⊥C )) =
f k+1,↑ (⊥C )).
Thus, by applying α, we obtain that for any k ∈ N,
(f A )k,↑ (⊥A ) = α(f k,↑ (⊥C )).
(†)
Since γ and f are continuous and α is always additive, we have that f A = α ◦ f ◦ γ is continuous because
it is a composition of continuous functions. Hence:
lfp(f A ) =
∨k∈N (f A )k,↑ (⊥A ) =
[by Knaster-Tarski’s theorem]
[by (†)]
∨k∈N α(f
k,↑
(⊥C )) =
[as α is additive]
α(∨k∈N f
k,↑
(⊥C )) =
[by Knaster-Tarski’s theorem]
α(lfp(f ))
and this concludes the proof.
It is worth noting that concretization maps of abstract domains which satisfies the ascending chain
conditions (i.e., every ascending chain is eventually stationary) are always trivially continuous.
2.2.3 Shells
Refinements of abstract domains have been studied from the beginning of abstract interpretation [13, 14]
and led to the notion of shell of an abstract domain [26, 29, 31]. Given a generic poset P≤ of semantic
objects — where x ≤ y intuitively means that x is a “refinement” of y — and a property P ⊆ P of these
objects, the generic notion of shell goes as follows: the P-shell of an object x ∈ P is defined to be an
object sx ∈ P such that:
(i) sx satisties the property P,
(ii) sx is a refinement of x, and
(iii) sx is the greatest among the objects satisfying (i) and (ii).
7
Note that if a P-shell exists then it is unique. Moreover, if the P-shell exists for any object in P then it
turns out that the operator mapping x ∈ P to its P-shell is a lower closure operator on P, being monotone,
idempotent and reductive: this operator will be called the P-shell refinement. We will be particularly
interested in shells of abstract domains and partitions, namely shells in the complete lattices of abstract
domains and partitions. Given a state space Σ and a partition property P ⊆ Part(Σ), the P-shell of
P ∈ Part(Σ) is the coarsest refinement of P satisfying P, when this exists. Also, given a concrete
domain C and a domain property P ⊆ Abs(C), the P-shell of A ∈ Abs(C), when this exists, is the most
abstract domain that satisfies P and refines A. Giacobazzi et al. [31] gave a constructive characterization of
backward complete abstract domains, under the assumption of dealing with continuous concrete functions.
As a consequence, they showed that backward complete shells always exist when the concrete functions
are continuous. In Section 6 we will follow this same idea for forward completeness and this will provide
the link between strongly preserving abstract models and complete abstract interpretations.
2.3 Abstract Model Checking and Strong Preservation
Standard temporal languages like CTL, CTL∗ , ACTL, the µ-calculus, LTL, etc., are interpreted on models specified as Kripke structures. Given a set AP of atomic propositions (of some language), a Kripke
structure K = (Σ, , ℓ) over AP consists of a transition system (Σ, ) together with a state labeling funcdef
tion ℓ : Σ → ℘(AP ). We use the following notation: for any s ∈ Σ, [s]ℓ = {s′ ∈ Σ | ℓ(s) = ℓ(s′ )},
def
while Pℓ = {[s]ℓ | s ∈ Σ} ∈ Part(Σ) denotes the state partition that is induced by ℓ. The notation s|=K ϕ
means that a state s ∈ Σ satisfies in K a state formula ϕ of some language L , where the specific definition
of the satisfaction relation |=K depends on the language L (interpretations of standard logical/temporal
operators can be found in [10]).
Standard abstract model checking [9, 10] relies on abstract Kripke structures that are defined over
partitions of the concrete state space Σ. A set A of abstract states is related to Σ by a surjective abstraction h : Σ → A that maps concrete states into abstract states and thus gives rise to a state partition
def
Ph = {h−1 (a) | a ∈ A} ∈ Part(Σ). Thus, in standard abstract model checking, formulae are interpreted
on an abstract Kripke structure A = (A, ♯ , ℓ♯ ) whose states are an abstract representation in A of some
block of the partition Ph . Given a specification language L of state formulae, a weak preservation result
for L guarantees that if a formula in L holds on an abstract Ktipke structure A then it also holds on the
corresponding concrete structure K: for any ϕ ∈ L , a ∈ A and s ∈ Σ such that h(s) = a, if a|=A ϕ
then s|=K ϕ. Moreover, strong preservation (s.p. for short) for L encodes the equivalence of abstract and
concrete validity for formulae in L : for any ϕ ∈ L , a ∈ A and s ∈ Σ such that h(s) = a, a|=A ϕ if and
only if s|=K ϕ.
The definition of weakly/strongly preserving abstract Kripke structures depends on the language L .
Let us recall some well-known examples [9, 10, 33]. Let K = (Σ, , ℓ) be a concrete Kripke structure
h : Σ → A be a surjection.
(i) Consider the language ACTL∗ . If Ph Pℓ then the abstract Kripke structure A = (A, ∃∃
h , ℓh )
weakly preserves ACTL∗ , where ℓh (a) = ∪{ℓ(s) | s ∈ Σ, h(s) = a} and ∃∃
⊆
A
×
A
is
defined
h
′
′
′
′
′
′
as: h(s1 ) ∃∃
h h(s2 ) ⇔ ∃s1 , s2 . h(s1 ) = h(s1 ) & h(s2 ) = h(s2 ) & s1 s2 .
(ii) Let Psim ∈ Part(Σ) be the partition induced by simulation equivalence on K. If Ph = Psim (this
also holds when Ph Psim ) then the abstract Kripke structure A = (A, ∀∃
h , ℓh ) strongly preserves
′
′
′
′
′
′
ACTL∗ , where h(s1 ) ∀∃
h(s
)
⇔
∀s
.
h(s
)
=
h(s
).
∃s
.
h(s
)
=
h(s
2
1
2 ) & s1 s2 .
1
1
2
2
h
(iii) Let Pbis ∈ Part(Σ) be the partition induced by bisimulation equivalence on K. If Ph = Pbis (this
also holds when Ph Pbis ) then the abstract Kripke structure A = (A, ∃∃
h , ℓh ) strongly preserves
∗
CTL .
Following Dams [19, Section 6.1] and Henzinger et al. [36, Section 2.2], the notion of strong preservation can be also given w.r.t. a mere state partition rather than w.r.t. an abstract Kripke structure. Let
J·KK : L → ℘(Σ) be the semantic function of state formulae in L w.r.t. a Kripke structure K = (Σ, , ℓ),
def
i.e., JϕKK = {s ∈ Σ | s|=K ϕ}. Then, the semantic interpretation of L on K induces the following logical
8
w
GFED
@ABC
R
stop
stop
ONML
/ HIJK
RY
GFED
/ @ABC
G
go
go
GFED
/ @ABC
Y
Figure 1: A U.K. traffic light.
equivalence ≡K
L ⊆ Σ × Σ:
′
′
s ≡K
L s iff ∀ϕ ∈ L . s ∈ JϕKK ⇔ s ∈ JϕKK .
Let PL ∈ Part(Σ) be the partition induced by ≡K
L (the index K denoting the Kripke structure is omitted). Then, a partition P ∈ Part(Σ) is strongly preserving2 for L (when interpreted on K) if P 4 PL .
Thus, PL is the coarsest partition that is strongly preserving for L . For a number of well known temporal
languages, like ACTL∗ , CTL∗ (see, respectively, the above points (ii) and (iii)), CTL∗ -X and the fragments of the µ-calculus described by Henzinger et al. [36], it turns out that if P is strongly preserving for
L then the abstract Kripke structure (P, ∃∃ , ℓL ) is strongly preserving for L , where, for any B ∈ P ,
ℓL (B) = ∪s∈B ℓ(s). In particular, (PL , ∃∃ , ℓL ) is strongly preserving for L and, additionally, PL is
the smallest possible abstract state space, namely if A = (A, ♯ , ℓ♯ ) is an abstract Kripke structure that
strongly preserves L then |PL | ≤ |A|.
However, given a language L and a Kripke structure K where formulae of L are interpreted, the
following example shows that it is not always possible to define an abstract Kripke structure A on the
partition PL such that A strongly preserves L .
Example 2.2. Consider the following simple language L :
L ∋ ϕ ::= stop | go | AXXϕ
and the Kripke structure K depicted in Figure 1, where superscripts determine the labeling function. K
models a four-state traffic light controller (like in the U.K. and in Germany): Red → RedYellow →
Green → Yellow. According to the standard semantics of AXX, we have that s|=K AXXϕ iff for any
path s0 s1 s2 . . . starting from s0 = s, it happens that s2 |=K ϕ. It turns out that JAXXstopKK = {G, Y }
and JAXXgoKK = {R, RY }. Thus, we have that PL = {{R, RY }, {G, Y }}. However, let us show
that there exists no abstract transition relation ♯ ⊆ PL × PL such that the abstract Kripke structure
A = (PL , ♯ , ℓL ) strongly preserves L . Assume by contradiction that such an abstract Kripke structure
A exists. Let B1 = {R, RY } ∈ PL and B2 = {G, Y } ∈ PL . Since R|=K AXXgo and G|=K AXXstop,
by strong preservation, it must be that B1 |=A AXXgo and B2 |=A AXXstop. Hence, necessarily, B1 ♯ B2
A
and B2 ♯ B1 . This leads to the contradiction B1 6|= AXXgo. In fact, if ♯ = {(B1 , B2 ), (B2 , B1 )} then
A
we would have that B1 6|= AXXgo. On the other hand, if, instead, B1 ♯ B1 (the case B2 ♯ B2 is analogous), then we would still have that B1 6|=A AXXgo. Even more, along the same lines it is not hard to show
that no proper abstract Kripke structure that strongly preserves L can be defined, because even if either
B1 or B2 is split we still cannot define an abstract transition relation that is strongly preserving for L .
3 Partitions as Abstract Domains
Let Σ be any (possibly infinite) set of states. Following [15, Section 5], a partition P ∈ Part(Σ) can be
viewed as an abstraction of ℘(Σ)⊆ as follows: any S ⊆ Σ is over approximated by the unique minimal
cover of S in P , namely by the union of all the blocks B ∈ P such that B ∩ S 6= ∅. A graphical example is
depicted on the left-hand side of Figure 2. This abstraction is formalized by a GI (αP , ℘(Σ)⊆ , ℘(P )⊆ , γP )
where:
def
def
αP (S) = {B ∈ P | B ∩ S 6= ∅}
γP (B) = ∪B∈B B.
Hence, any partition P ∈ Part(Σ) induces an abstract domain adp (P ) ∈ Abs(℘(Σ)), and an abstract
domain A ∈ Abs(℘(Σ)) is called partitioning when A is equivalent to adp (P ) for some partition P .
Observe that the closure adp (P ) = γP ◦ αP associated to a partitioning abstract domain is defined as
2 Dams
[19] uses the term “fine” instead of “strongly preserving”.
9
Figure 2: Partitions as abstract domains: over-approximation on the left and under-approximation on the
right.
adp (P ) = λS. ∪ {B ∈ P | B ∩ S 6= ∅}. Accordingly, a closure µ ∈ uco(℘(Σ)) that coincides with
γP ◦ αP , for some partition P , is called partitioning. We denote by Abspar (℘(Σ)) and ucopar (℘(Σ))
the sets of, respectively, partitioning abstract domains and closures on ℘(Σ). As noted in [16], a surjective abstraction h : Σ → A used in standard abstract model checking that maps concrete states into
abstract states (cf. Section 2.3) gives rise to a partitioning Galois insertion (αh , ℘(Σ)⊆ , ℘(A)⊆ , γh ) where
def
def
αh = λS ⊆ Σ.{h(s) ∈ A | s ∈ S} and γh = λX ⊆ A.{s ∈ Σ | h(s) ∈ X}.
Partitions can be also viewed as dual abstractions when a set S is under approximated by the union of
all the blocks B ∈ P such that B ⊆ S. A graphical example of this under approximation is depicted on the
right-hand side of Figure 2. This dual abstraction is formalized by the GI (e
αP , ℘(Σ)⊇ , ℘(P )⊇ , e
γP ) where
the ordering on the concrete domain ℘(Σ) is given by the subset relation and
def
α
eP (S) = {B ∈ P | B ⊆ S}
def
γ
eP (B) = ∪B∈B B.
In the following, we will be interested in viewing partitions as over approximations, that is partitions as
abstract domains of ℘(Σ)⊆ .
Thus, partitions can be viewed as representations of abstract domains. On the other hand, it turns out
that abstract domains can be abstracted to partitions. An abstract domain A ∈ Abs(℘(Σ)⊆ ) induces a state
equivalence ≡A on Σ by identifying those states that cannot be distinguished by A:
s ≡A s′ iff α({s}) = α({s′ }).
def
For any s ∈ Σ, [s]A = {s′ ∈ Σ | α({s}) = α({s′ })} is a block of the state partition par(A) induced by A:
def
par(A) = {[s]A | s ∈ Σ}.
Thus, par : Abs(℘(Σ)) → Part(Σ) is a mapping from abstract domains to partitions.
Example 3.1. Let Σ = {1, 2, 3, 4} and let us specify abstract domains as uco’s on ℘(Σ). The uco’s
µ1 = {∅, 12, 3, 4, 1234}, µ2 = {∅, 12, 3, 4, 34, 1234}, µ3 = {∅, 12, 3, 4, 34, 123, 124, 1234}, µ4 =
{12, 123, 124, 1234} and µ5 = {∅, 12, 123, 124, 1234} all induce the same partition P = par(µi ) =
{12, 3, 4} ∈ Part(Σ). For example, µ5 ({1}) = µ5 ({2}) = {1, 2}, µ5 ({3}) = {1, 2, 3} and µ5 ({4}) =
{1, 2, 3, 4} so that par(µ5 ) = P . Observe that µ3 is the only partitioning abstract domain because
adp (P ) = µ3 .
Abstract domains of ℘(Σ) carry additional information other than the underlying state partition and
this additional information allows us to distinguish them. It turns out that this can be precisely stated by
abstract interpretation since the above mappings par and adp allows us to show that the whole lattice of
partitions of Σ can be viewed as a (“higher-order”) abstraction of the lattice of abstract domains of ℘(Σ).
Theorem 3.2. (par, Abs(℘(Σ))⊒ , Part(Σ) , adp ) is a Galois insertion.
Proof. Let A ∈ Abs(℘(Σ)) and P ∈ Part(Σ) and let µA ∈ uco(℘(Σ)) be the closure associated with the
abstract domain A. Let us prove that P par(A) ⇔ adp (P ) ⊑ µA .
10
(⇒) For S ∈ ℘(Σ) we have to prove that adp (P )(S) ⊆ µA (S). Consider s ∈ adp (P )(S). Hence, there
exists some B ∈ P such that s ∈ B and B ∩ S 6= ∅. Let q ∈ B ∩ S. Since P par(A), there exists
some block [r]A ∈ par(A) such that B ⊆ [r]A . Thus, for any x, y ∈ B, α({x}) = α({r}) = α({y}), in
particular, α({s}) = α({q}). Consequently, since q ∈ S and therefore µA ({q}) ⊆ µA (S), we have that
µA ({s}) = µA ({q}) ⊆ µA (S), so that s ∈ µA (S).
(⇐) Consider a block B ∈ P and some s ∈ B. We show that B ⊆ [s]A , namely if s′ , s′′ ∈ B
then α({s′ }) = α({s′′ }). Since adp (P ) ⊑ µA , if s′ , s′′ ∈ B then adp (P )({s′ }) = B ⊆ µA ({s′ })
so that s′′ ∈ µA ({s′ }) and therefore µA ({s′′ }) ⊆ µA ({s′ }). Likewise, µA ({s′ }) ⊆ µA ({s′′ }) so that
µA ({s′ }) = µA ({s′′ }) and in turn α({s′ }) = α({s′′ }).
Finally, observe that adp is 1-1 so that the above adjunction is indeed a Galois insertion.
Let us observe that, as recalled in Section 2.2, the adjoint maps par and adp give rise to an order
isomorphism between the lattices hPart(Σ), i and hAbspar (℘(Σ)), ⊑i.
Corollary 3.3. Let A ∈ Abs(℘(Σ)). The following statements are equivalent:
(1) A is partitioning.
(2) γ is additive and {γ(α({s}))}s∈Σ is a partition of Σ. In this case, par(A) = {γ(α({s}))}s∈Σ .
(3) A is forward complete for the complement operator ∁.
Proof. Let A ∈ Abs(℘(Σ)) and let µA = γ ◦ α ∈ uco(℘(Σ)) be the corresponding uco.
(1) ⇒ (2) By Theorem 3.2, A ∈ Abspar (℘(Σ)) iff adp (par(A)) = A. Thus, if adp (par(A)) = A then
µA = γ ◦ α is obviously additive. Moreover, s ≡A s′ iff α({s}) = α({s′ }) iff γ(α({s})) = γ(α({s′ })),
so that, for any s ∈ Σ, [s]A = γ(α({s})) and therefore par(A) = {γ(α({s}))}s∈Σ .
(2) ⇒ (1) Since {γ(α({s}))}s∈Σ = P ∈ Part(Σ) we have that for any s ∈ Σ, [s]A = γ(α({s})): in fact,
if s′ ∈ γ(α({s})) then α({s′ }) ≤ α({s}), hence γ(α({s′ })) ⊆ γ(α({s})) and therefore γ(α({s′ })) =
γ(α({s})). Thus, par(A) = P . Moreover, since γ is additive, for any S ⊆ Σ, ∪s∈S γ(α({s})) =
γ(∨s∈S α({s})) = γ(α(S)) ∈ µA . Hence, since adp (P ) = {∪s∈S γ(α({s})) | S ⊆ Σ} we have that
adp (par(A)) = A.
(1) ⇒ (3) Assume that A ∈ Abspar (℘(Σ)). It is enough to prove that for any s ∈ Σ, ∁(µA ({s})) ∈ µA : in
fact, by (1) ⇒ (2), γ is additive and therefore µA is additive (because it is a composition of additive maps)
and therefore if S ∈ µA then S = ∪s∈S µA ({s}) so that ∁(S) = ∩s∈S ∁(µA ({s})). Let us observe the
following fact (∗): for any s, s′ ∈ Σ, s 6∈ µA ({s′ }) ⇔ µA ({s}) ∩ µA ({s′ }) = ∅; this is a consequence of
the fact that, by (1) ⇒ (2), {µA ({s})}s∈Σ is a partition. For any s ∈ Σ, we have that ∁(µA ({s})) ∈ µA
because:
µA (∁(µA ({s}))) = µA ({s′ ∈ Σ | s′ 6∈ µA ({s})}
= ∪{µA ({s′ }) | s′ 6∈ µA ({s})}
[by additivity of µA ]
[by the above fact (∗)]
= ∪{µA ({s′ }) | µA ({s′ }) ∩ µA ({s}) = ∅}
= ∪{µA ({s′ }) | µA ({s′ }) ⊆ ∁(µA ({s}))}
⊆ ∁(µA ({s}))
(3) ⇒ (1) Assume that µA is forward complete for ∁, i.e. µA is closed under complements. By (2) ⇒ (1),
it is enough to prove that γ is additive and that {µA ({s})}s∈Σ ∈ Part(Σ).
(i) γ is additive. Observe that γ is additive iff µA is additive iff µA is closed under arbitrary unions. If
{Si }i∈I ⊆ µA then ∪i Si = ∁(∩i ∁(Si )) ∈ µA , because, µA is closed under complements (and arbitrary
intersections).
(ii) {µA ({s})}s∈Σ ∈ Part(Σ). Clearly, we have that ∪s∈Σ µA ({s}) = Σ. Consider now s, r ∈ Σ such that
µA ({s}) ∩ µA ({r}) 6= ∅. Let us show that µA ({s}) = µA ({r}). In order to show this, let us prove that
s ∈ µA ({r}). Notice that µA ({s})rµA ({r}) = µA ({s})∩∁(µA ({r})) ∈ µA , because µA is closed under
complements. If s 6∈ µA ({r}) then we would have that s ∈ µA ({s}) r µA ({r}) ∈ µA , and this would
imply µA ({s}) ⊆ µA ({s}) r µA ({r}) ⊆ µA ({s}), namely µA ({s}) = µA ({s}) r µA ({r}). Thus, we
would obtain the contradiction µA ({s}) ∩ µA ({r}) = ∅. Hence, we have that s ∈ µA ({r}) and therefore
µA ({s}) ⊆ µA ({r}). By swapping the roles of s and r, we also obtain that µA ({r}) ⊆ µA ({s}), so that
µA ({s}) = µA ({r}).
11
def
Let us remark that P = adp ◦ par is a lower closure operator on hAbs(℘(Σ)), ⊑i and that for any
A ∈ Abs(℘(Σ)), A is partitioning iff P(A) = A. Hence, P is exactly the partitioning-shell refinement,
namely P(A) is the most abstract refinement of A that is partitioning.
4 Abstract Semantics of Languages
4.1 Concrete Semantics
We consider temporal specification languages L whose state formulae ϕ are inductively defined by:
L ∋ ϕ ::= p | f (ϕ1 , ..., ϕn )
where p ranges over a (typically finite) set of atomic propositions AP, while f ranges over a finite set Op
of operators. AP and Op are also denoted, respectively, by AP L and Op L . Each operator f ∈ Op has
an arity3 ♯(f ) > 0.
Formulae in L are interpreted on a semantic structure S = (Σ, I) where Σ is any (possibly infinite) set
of states and I is an interpretation function I : AP ∪ Op → Fun(℘(Σ)) that maps p ∈ AP to I(p) ∈ ℘(Σ)
and f ∈ Op to I(f ) : ℘(Σ)♯(f ) → ℘(Σ). I(p) and I(f ) are also denoted by, respectively, p and f .
def
def
Moreover, AP = {p ∈ ℘(Σ) | p ∈ AP } and Op = {f : ℘(Σ)♯(f ) → ℘(Σ) | f ∈ Op}. Note that the
def
interpretation I induces a state labeling ℓI : Σ → ℘(AP) by ℓI (s) = {p ∈ AP | s ∈ I(p)}. The concrete
state semantic function J·KS : L → ℘(Σ) evaluates a formula ϕ ∈ L to the set of states making ϕ true
w.r.t. the semantic structure S:
JpKS = p and Jf (ϕ1 , ..., ϕn )KS = f (Jϕ1 KS , ..., Jϕn KS ).
Semantic structures generalize the role of Kripke structures. In fact, in standard model checking a semantic
structure is usually defined through a Kripke structure K so that the interpretation of logical/temporal
operators is defined in terms of standard logical operators and paths in K. In the following, we freely
use standard logical and temporal operators together with their corresponding usual interpretations: for
example, I(∧) = ∩, I(∨) = ∪, I(¬) = ∁, I(EX) = preR , I(AX) = pre
f R , etc. As an example, consider
the standard semantics of CTL:
CTL ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | AXϕ | EXϕ | AU(ϕ1 , ϕ2 ) | EU(ϕ1 , ϕ2 ) | AR(ϕ1 , ϕ2 ) | ER(ϕ1 , ϕ2 )
with respect to a Kripke structure K = (Σ, R, ℓ). Hence, K determines a corresponding interpretation I
for atoms in AP and operators of Op CTL , namely I(AX) = pre
f R , I(EX) = preR , etc., and this defines
the concrete semantic function J·KK : CTL → ℘(Σ).
If g is any operator with arity ♯(g) = n > 0 whose interpretation is given by g : ℘(Σ)n → ℘(Σ) and
S = (Σ, I) is a semantic structure then we say that a language L is closed under g for S when for any
ϕ1 , ..., ϕn ∈ L there exists some ψ ∈ L such that g(Jϕ1 KS , ..., Jϕn KS ) = JψKS . For instance, if Op L
includes EX and negation with their standard interpretations, i.e. I(EX) = preR and I(¬) = ∁, then L
is closed under AX with its standard interpretation pre
f R because pre
f R = ∁ ◦ preR ◦∁. This notion can be
extended in a straightforward way to infinitary operators: for instance, T
L is closed under infinite logical
conjunction for S iff for any Φ ⊆ L , there exists some ψ ∈ L such that ϕ∈Φ JϕKS = JψKS . In particular,
let us remark that if L is closed under infinite logical conjunction then it must exist some ψ ∈ L such
that ∩∅ = Σ = JψKS , namely L is able to express the tautology true. Let us remark that if the state space
Σ is finite and L is closed under logical conjunction then we always mean that there exists some ψ ∈ L
such that ∩∅ = Σ = JψKS . Finally, note that L is closed under negation and infinite logical conjunction
if and only if L includes propositional logic.
3 It would be possible to consider generic operators whose arity is any possibly infinite ordinal, thus allowing, for example, infinite
conjunctions or disjunctions.
12
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p,q
?>=<
89:;
3
>=<
/ ?89:;
4 j
r
⊤
@@
~~~ @
a5 A
a4 A
A
A
}}} A }}} A
a3
a2
a1 OO
OOO
oo
O oooo
* ?>=<
89:;
5
q
⊥
Figure 3: A Kripke structre on the left and an abstract domain on the right.
4.2 Abstract Semantics
In the following, we apply the standard abstract interpretation approach for defining abstract semantics
[13, 14]. Let L be a language and S = (Σ, I) be a semantic structure for L . An abstract semantic
structure S ♯ = (A, I ♯ ) is given by an abstract domain A ∈ Abs(℘(Σ)⊆ ) and by an abstract interpretation
function I ♯ : AP ∪ Op → Fun(A). An abstract semantic structure S ♯ therefore induces an abstract
semantic function J·KS ♯ : L → A that evaluates formulae in L to abstract values in A. The abstract
interpretation I ♯ is a correct over-approximation (respectively, under-approximation) of I on A when for
any p ∈ AP , γ(I ♯ (p)) ⊇ I(p) (respectively, γ(I ♯ (p)) ⊆ I(p)) and for any f ∈ Op, γ ◦ I ♯ (f ) ⊒
I(f ) ◦ γ (respectively, γ ◦ I ♯ (f ) ⊑ I(f ) ◦ γ). If I ♯ is a correct over-approximation (respectively, underapproximation) of I and the semantic operations in Op are monotone then the abstract semantics is an
over-approximation (respectively, under-approximation) of the concrete semantics, namely for any ϕ ∈ L ,
γ(JϕKS ♯ ) ⊇ JϕKS (respectively, γ(JϕKS ♯ ) ⊆ JϕKS ).
In particular, the abstract domain A always induces an abstract semantic structure S A = (A, I A ) where
I A is the best correct approximation of I on A, i.e. I A interprets atoms p and operators f as best correct
approximations on A of, respectively, p and f : for any p ∈ AP and f ∈ Op,
I A (p) = α(p) and I A (f ) = f A .
def
def
Thus, the abstract domain A systematically induces an abstract semantic function J·KS A : L → A, also
denoted by J·KA
S , which is therefore defined by:
A
A
A
A
JpKA
S = α(p) and Jf (ϕ1 , ..., ϕn )KS = f (Jϕ1 KS , ..., Jϕn KS ).
As usual in abstract interpretation, observe that the concrete semantics is a particular abstract semantics,
namely it is the abstract semantics induced by the “identical abstraction” (id, ℘(Σ), ℘(Σ), id).
Example 4.1. Let L ∋ ϕ ::= p | q | r | ϕ1 ∧ ϕ2 | EXϕ. Let us consider the Kripke structure K = (Σ, , ℓ)
and the lattice A both depicted in Figure 3. Let S be the semantic structure induced by the Kripke structure
K so that EX = pre . Let us consider the formulae EXr and EX(p ∧ q), whose concrete semantics are as
follows: JEXrKS = {3, 5} and JEX(p ∧ q)KS = {1, 2}. A is an abstract domain of ℘(Σ) where the Galois
insertion (α, ℘(Σ), A, γ) is determined by the following concretization map:
γ(⊥) = ∅; γ(a1 ) = {1, 2}; γ(a2 ) = {3}; γ(a3 ) = {3, 4};
γ(a4 ) = {1, 2, 3}; γ(a5 ) = {3, 4, 5}; γ(⊤) = {1, 2, 3, 4, 5}.
Note that, by Corollary 3.3, A is not partitioning because γ is not additive: γ(a2 ) ∪ γ(a3 ) = {3, 4} (
{3, 4, 5} = γ(a2 ∨ a3 ). It turns out that:
JEXrKA
S
= α(pre (γ(JrKA
S )) = α(pre (γ(α(r)))) = α(pre (γ(a3 )))
= α(pre ({3, 4})) = α({1, 2, 3, 5}) = ⊤;
JEX(p ∧ q)KA
S
A
= α(pre (γ(JpKA
S ∧ JqKS ))) = α(pre (γ(α(p) ∧ α(q))))
= α(pre (γ(a4 ∧ a5 ))) = α(pre (γ(a2 ))) = α(pre (3)) = α({1, 2}) = a1 .
Observe that the abstract semantics JEXrKA
S is a proper over-approximation of JEXrKS because JEXrKS (
).
On
the
other
hand,
the
concrete
semantics JEX(p ∧ q)KS is precisely represented in A because
γ(JEXrKA
S
)
=
JEX(p
∧
q)K
.
γ(JEX(p ∧ q)KA
S
S
13
5 Generalized Strong Preservation
We showed in Section 3 how a state partition P can be viewed as a partitioning abstract domain adp (P )
specified by the GI (αP , ℘(Σ)⊆ , ℘(P )⊆ , γP ). Thus, given a language L and a corresponding semantic
structure S = (Σ, I), it turns out that
any partition P ∈ Part(Σ) systematically induces a corresponddef
adp (P )
: L → adp (P ) that evaluates a formula in L to a (possibly
ing abstract semantics J·KP
S = J·KS
empty) union of blocks of P . Strong preservation for a partition P can be characterized in terms of the
corresponding abstract domain adp (P ) as follows.
Lemma 5.1. P ∈ Part(Σ) is s.p. for L iff ∀ϕ ∈ L and S ⊆ Σ, αP (S) ⊆ JϕKP
S ⇔ S ⊆ JϕKS .
Proof. (⇒): Let us first observe that for any ϕ ∈ L , γP (αP (JϕKS )) = JϕKS : in fact, for any s ∈ JϕKS ,
αP ({s}) is the block of P containing s; since P PL , we have that αP ({s}) ⊆ JϕKS , and from this
αP (JϕKS ) ⊆ JϕKS and in turn γP (αP (JϕKS )) = JϕKS .
Let us now prove by structural induction on ϕ ∈ L that JϕKS = γP (JϕKP
S ):
– ϕ ≡ p ∈ AP L : by using the above observation, JpKS = γP (αP (JpKS )) = γP (JpKP
S ).
– ϕ ≡ f (ϕ1 , . . . , ϕn ):
Jf (ϕ1 , . . . , ϕn )KS = [by the above observation]
γP (αP (Jf (ϕ1 , . . . , ϕn )KS )) = [by definition]
γP (αP (f (Jϕ1 KS , . . . , Jϕn KS ))) = [by inductive hypothesis]
P
γP (αP (f (γP (Jϕ1 KP
S ), . . . , γP (Jϕn KS )))) = [by definition]
γP (Jf (ϕ1 , . . . , ϕn )KP
S ).
P
Now, consider any ϕ ∈ L . If S ⊆ JϕKS then αP (S) ⊆ αP (JϕKS ) = αP (γP (JϕKP
S )) = JϕKS . Conversely,
P
if αP (S) ⊆ JϕKP
then
S
⊆
γ
(JϕK
)
=
JϕK
.
P
S
S
S
(⇐): Consider a block B ∈ P and s, s′ ∈ B so that αP ({s}) = B = αP ({s′ }). By hypothesis, for
′
P
′
any ϕ ∈ L , we have that s ∈ JϕKS iff αP ({s}) ⊆ JϕKP
S iff αP ({s }) ⊆ JϕKS iff s ∈ JϕKS . Thus,
′
s ≡L s .
This states that a partition P ∈ Part(Σ) is s.p. for L if and only if to check whether some set S of
states satisfies some formula ϕ ∈ L , i.e. S ⊆ JϕKS , is equivalent to check whether the abstract state
P
αP (S) is more precise than the abstract semantics JϕKP
S , that is S is over-approximated by JϕKS . The
key observation here is that in our abstract interpretation-based framework partitions are particular abstract
domains. This allows us to generalize the notion of strong preservation from partitions to generic abstract
semantic functions as follows.
Definition 5.2. Let L be a language, S = (Σ, I) be a semantic structure for L and S ♯ = (A, I ♯ ) be a
corresponding abstract semantic structure. The abstract semantics J·KS ♯ is strongly preserving for L (w.r.t.
S) if for any ϕ ∈ L and S ⊆ Σ,
α(S) ≤A JϕKS ♯ ⇔ S ⊆ JϕKS .
Definition 5.2 generalizes standard strong preservation from partitions, as characterized by Lemma 5.1,
both to an arbitrary abstract domain A ∈ Abs(℘(Σ)) and to a corresponding abstract interpretation function
I ♯ . Likewise, standard weak preservation can be generalized as follows. Let K = (Σ, R, ℓ) be a concrete
Kripke structure that induces the concrete semantics JϕKK = {s ∈ Σ | s|=K ϕ}. Let h : Σ → A be a
surjective abstraction and let (αh , ℘(Σ), ℘(A), γh ) be the corresponding partitioning abstract domain. Let
A = (A, R♯ , ℓ♯ ) be an abstract Kripke structure on A that gives rise to the abstract semantics JϕKA = {a ∈
A | a|=A ϕ}. Then, A weakly preserves L when
∀φ ∈ L .∀S ⊆ Σ. αh (S) ⊆ JϕKA ⇒ S ⊆ JϕKK .
Hence, weak preservation can be generalized to generic abstract domains and abstract semantics accordingly to Definition 5.2.
14
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p
>=<
/ ?89:;
3
p
ONML
HIJK
[12]
h
p
HIJK
/ ONML
[3]
r
Figure 4: A Kripke structure K on the left and an abstract Kripke structure A on the right.
5.1 Strong Preservation is an Abstract Domain Property
Definition 5.2 is a direct and natural generalization of the standard notion of strong preservation in abstract
model checking. It can be equivalently stated as follows.
Lemma 5.3. J·KS ♯ is s.p. for L iff for any ϕ ∈ L , JϕKS = γ(JϕKS ♯ ).
Proof. (⇒) On the one hand, γ(JϕKS ♯ ) ⊆ JϕKS iff α(γ(JϕKS ♯ )) ≤ JϕKS ♯ iff JϕKS ♯ ≤ JϕKS ♯ , which is
trivially true. On the other hand, JϕKS ⊆ γ(JϕKS ♯ ) iff α(JϕKS ) ≤ JϕKS ♯ iff JϕKS ⊆ JϕKS , that is trivially
true.
(⇐) We have that S ⊆ JϕKS iff S ⊆ γ(JϕKS ♯ ) iff α(S) ≤ JϕKS ♯ .
In particular, it is worth noting that if J·KS ♯ is s.p. for L then J·KS ♯ = α ◦ J·KS holds.
Lemma 5.4. Let A ∈ Abs(℘(Σ)).
(1) Let S1♯ = (A, I1♯ ) and S2♯ = (A, I2♯ ) be abstract semantic structures on A. If J·KS ♯ and J·KS ♯ are both
1
2
s.p. for L then J·KS ♯ = J·KS ♯ .
1
2
(2) Let S ♯ = (A, I ♯ ) be an abstract semantic structure on A. If J·KS ♯ is s.p. for L then J·KA
S is s.p. for L .
Proof. (1) By Lemma 5.3, for any ϕ ∈ L , γ(JϕKS ♯ ) = JϕKS = γ(JϕKS ♯ ), so that, by applying α,
1
2
JϕKS ♯ = α(γ(JϕKS ♯ )) = α(JϕKS ) = α(γ(JϕKS ♯ )) = JϕKS ♯ .
1
1
2
2
(2) Let us first observe that for any ϕ ∈ L , γ(α(JϕKS )) = JϕKS . In fact, γ(α(JϕKS )) ⊆ JϕKS ⇔
α(γ(α(JϕKS ))) ≤ JϕKS ♯ ⇔ α(JϕKS ) ≤ JϕKS ♯ ⇔ JϕKS ⊆ JϕKS . As a consequence of this fact, by
structural induction on ϕ ∈ L , analogously to the proof of Lemma 5.1, it is easy to prove that γ(JϕKA
S) =
JϕKS . Thus, by Lemma 5.3, J·KA
is
s.p.
for
L
.
S
Thus, it turns out that strong preservation is an abstract domain property. This means that given any
abstract domain A ∈ Abs(℘(Σ)), it is possible to define an abstract semantic structure S ♯ = (A, I ♯ ) on
A such that the corresponding abstract semantics J·KS ♯ is s.p. for L if and only if the induced abstract
semantics J·KA
S : L → A is s.p. for L . In particular, this also holds for the standard approach: if A =
(A, R♯ , ℓ♯ ) is an abstract Kripke structure for L , where h : Σ → A is the corresponding surjection, then
the standard abstract semantics J·KA strongly preserves L if and only if the abstract semantics induced by
the partitioning abstract domain (αh , ℘(Σ), ℘(A), γh ) strongly preserves L , and in this case this abstract
semantics coincides with the standard abstract semantics J·KA . Strong preservation is an abstract domain
property and therefore can be defined without loss of generality as follows.
Definition 5.5. An abstract domain A ∈ Abs(℘(Σ)) is strongly preserving for L (w.r.t. a semantic structure S) when J·KA
S is s.p. for L (w.r.t. S). We denote by SPL ⊆ Abs(℘(Σ)) the set of abstract domains
that are s.p. for L .
Example 5.6. Let us consider Example 4.1. It turns out that the abstract domain A is not s.p. for L
because, by Lemma 5.3, JEXrKS = {3, 5} ( {1, 2, 3, 4, 5} = γ(⊤) = γ(JEXrKA
S ).
Example 5.7. Let us consider the simple language L ∋ ϕ ::= p | EXϕ and the Kripke structure K
depicted in Figure 4. The Kripke structure K induces the semantic structure S = ({1, 2, 3}, I) such that
I(p) = {1, 2, 3} and I(EX) = pre . Hence, we have that JpKS = {1, 2, 3}, JEXpKS = {1, 2, 3} and, for
k > 1, JEXk pKS = {1, 2, 3}. Let us consider the partitioning abstract domain A induced by the partition
P = {[12], [3]} and related to ℘(Σ) by α and γ. Let us consider two different abstract semantic structures
on A.
15
– The abstract semantic structure S A = (A, I A ) is induced as best correct approximation of I by A.
– The abstract semantic structure S A = (A, I A ) is instead induced by the abstract Kripke structure
A = (A, ♯ , ℓ♯ ) in Figure 4. Hence, I A (p) = {[12], [3]} and I A (EX) = pre♯ .
S A is different from S A because I A (EX) 6= I A (EX). In fact, I A (EX)({[12]}) = α(pre (γ({[12]}))) =
α(pre ({1, 2})) = α({1}) = {[12]}, while I A (EX)({[12]}) = pre♯ ({[12]}) = ∅.
Let us show that both the abstract semantics J·KA
S and J·KS ♯ are s.p. for L .
A
– We have that JpKA
S = {[12], [3]}, JEXpKS = α(pre ({1, 2, 3})) = α({1, 2, 3}) = {[12], [3]} and,
k A
for k > 1, JEX pKS = {[12], [3]}. Thus, for any ϕ ∈ L, JϕKS = γ(JϕKA
S ).
– We have that JpKS A = {[12], [3]}, JEXpKS A = pre♯ ({[12], [3]}) = {[12], [3]} and, for k > 1,
JEXk pKS A = {[12], [3]}. Thus, for any ϕ ∈ L, JϕKS = γ(JϕKS A ).
Consequently, by Lemma 5.3, both abstract semantics are s.p. for L .
5.2 The Most Abstract Strongly Preserving Domain
As recalled in Section 2.3, a language L and a semantic structure S for L induce a corresponding logical
partition PL ∈ Part(Σ). By Lemma 5.1, it turns out that PL is the coarsest strongly preserving partitioning abstract domain for L . This can be generalized to arbitrary abstract domains as follows. Let us define
ADL by:
def
ADL = M({JϕKS | ϕ ∈ L }).
Hence, ADL is the closure under arbitrary intersections of the set of concrete semantics of formulae in L .
Observe that ADL ∈ Abs(℘(Σ)) because it is a Moore-family of ℘(Σ).
Theorem 5.8. For any A ∈ Abs(℘(Σ)), A ∈ SPL iff A ⊑ ADL .
Proof. Let µ = γ ◦ α ∈ uco(℘(Σ)) and let µL ∈ uco(℘(Σ)) be the uco associated to ADL , that is
µL (S) = ∩{JϕKS | ϕ ∈ L , S ⊆ JϕKS }. Recall that A ⊑ ADL iff for any ϕ ∈ L , JϕKS ∈ µ.
(⇒) For any ϕ ∈ L , we have that γ(α(JϕKS )) = JϕKS because, by Lemma 5.3, γ(α(JϕKS )) =
A
γ(α(γ(JϕKA
S ))) = γ(JϕKS ) = JϕKS .
(⇐) By hypothesis, γ(α(JϕKS )) = JϕKS for any ϕ. Let us show by structural induction on ϕ ∈ L that
JϕKS = γ(JϕKA
S ).
– ϕ ≡ p ∈ AP L : by using the hypothesis, JpKS = γP (αP (JpKS )) = γP (JpKA
S ).
– ϕ ≡ f (ϕ1 , . . . , ϕn ):
Jf (ϕ1 , . . . , ϕn )KS = [by hypothesis]
γ(α(Jf (ϕ1 , . . . , ϕn )KS )) = [by definition]
γ(α(f (Jϕ1 KS , . . . , Jϕn KS ))) = [by inductive hypothesis]
A
γ(α(f (γ(Jϕ1 KA
S ), . . . , γ(Jϕn KS )))) = [by definition]
γ(Jf (ϕ1 , . . . , ϕn )KA
S ).
Thus, by Lemma 5.3, A ∈ SPL .
Thus, ADL is the most abstract domain that is s.p. for L w.r.t. S. As a consequence, it turns out that
A is s.p. for L if and only if A represents with no loss of precision the concrete semantics of any formula
in L , that is ∀ϕ ∈ L . γ(α(JϕKS )) = JϕKS . Lemma 5.4 states that if a s.p. abstract semantics on a given
abstract domain exists then this is unique. Nevertheless, Example 5.7 shows that this unique s.p. abstract
semantics may be induced from different abstract semantic structures, i.e. different abstract interpretation
functions. However, when L is closed under conjunction, it turns out that on the most abstract s.p. domain
ADL , the abstract interpretation function is unique and is given by the best correct approximation I ADL .
16
Theorem 5.9. Let L be closed under infinite logical conjunction and let S ♯ = (ADL , I ♯ ) be an abstract
semantic structure on ADL . If J·KS ♯ is s.p. for L then I ♯ = I ADL .
Proof. Since L is closed under arbritrary logical conjunctions we have that ADL = {JϕKS | ϕ ∈ L }.
L
Thus, for any a ∈ ADL , there exists some ϕ ∈ L such that a = JϕKS ♯ = JϕKAD
. In fact, if a ∈ ADL
S
then a = JϕKS , for some ϕ ∈ L , so that, by Lemmata 5.3 and 5.4, a = JϕKS = γ(JϕKS ♯ ) = JϕKS ♯ =
L
JϕKAD
.
S
L
Let p ∈ AP. Then, by Lemma 5.4, JpKS ♯ = JpKAD
so that I ♯ (p) = I ADL (p).
S
Let f ∈ Op. Then,
I ♯ (f )(a1 , ..., an ) =
♯
I (f )(Jϕ1 KS ♯ , ..., Jϕn KS ♯ ) =
Jf (ϕ1 , ..., ϕn )KS ♯ =
L
Jf (ϕ1 , ..., ϕn )KAD
=
S
L
L
I ADL (f )(Jϕ1 KAD
, ..., Jϕn KAD
)=
S
S
ADL
I
(f )(a1 , ..., an ).
[by the observation above]
[by definition]
[by Lemma 5.4]
[by definition]
[by the observation above]
Thus, I ♯ = I ADL .
Hence, in the most abstract s.p. domain ADL there is a unique choice for interpreting atoms and operations
of L .
In our generalized framework, strong preservation for partitions becomes a particular instance through
the Galois insertion par/adp . Moreover, when L is closed under infinite conjunction, it turns out that the
most abstract s.p. domain ADL is partitioning if and only if L is also closed under negation.
Proposition 5.10.
(1) PL = par(ADL ) and adp (PL ) = P(ADL ).
(2) P is strongly preserving for L iff P 4 par(ADL ) iff adp (P ) ⊑ ADL .
(3) Let L be closed under conjunction. Then, ADL is partitioning iff L is closed under logical negation.
Proof. (1) Let µL ∈ uco(℘(Σ)) be the uco associated to ADL . We have that par(ADL ) = {[s]ADL | s ∈
Σ}, where [s]ADL = {s′ ∈ Σ | µL ({s′ }) = µL ({s})}. We also have that s ≡L s′ iff ∀ϕ ∈ L .s ∈
JϕKS ⇔ s′ ∈ JϕKS iff µL ({s}) = µL ({s′ }), so that PL = par(ADL ). Moreover, adp (PL ) =
adp (par(ADL )) = P(ADL ).
(2) P is s.p. for L iff P 4 PL iff, by Point (1), P 4 par(AL ) iff, by Theorem 3.2, adp (P ) ⊑ ADL .
(3) Since L is closed under infinite logical conjunction, ADL = {JϕKS | ϕ ∈ L }. Thus, L is closed
under logical negation iff ADL is closed under complementation ∁ and this exactly means that ADL is
forward complete for the complement ∁. By Corollary 3.3, this latter fact happens iff ADL is partitioning.
In particular, when L is closed under conjunction but not under negation, it turns out that adp (PL ) ⊏
ADL , i.e. a proper loss of information occurs when the domain ADL is abstracted to the partition
par(ADL ) = PL . On the other hand, when L is closed under conjunction and negation, we have that
adp (PL ) = ADL and therefore, by Theorem 5.9, the abstract interpretation function on the partitioning
abstract domain adp (PL ) is uniquely determined.
Example 5.11. Let us consider the traffic light controller K in Example 2.2. As already observed, formulae
of L have the following semantics in K:
JstopKK = {R, RY }; JgoKK = {G, Y }; JAXXstopKK = {G, Y }; JAXXgoKK = {R, RY }
so that
ADL = M({JϕKK | ϕ ∈ L }) = {∅, {R, RY }, {G, Y }, {R, RY, G, Y }}
and PL = par(ADL ) = {{R, RY }, {G, Y }}. We denote by µL the uco associated to ADL . As
shown in Example 2.2, it turns out that no abstract Kripke structure that properly abstracts K and strongly
17
p
?>=<
89:;
1
p
>=<
/ ?89:;
2
p
?>=<
89:;
3
>=<
/ ?89:;
4 j
p
GFED
@ABC
[12]
p
p
GFED
@ABC
[3]
* ?>=<
89:;
5
q
p
GFED
/ @ABC
[4] k
+ @ABC
GFED
[5]
q
Figure 5: Concrete (on the left) and abstract (on the right) Kripke structures.
preserves L can be defined. In our approach, the abstract domain ADL induces a corresponding strongly
L
preserving abstract semantics J·KAD
: L → ADL , where the best correct approximation of the operator
K
AXX : ℘(Σ) → ℘(Σ) on ADL is:
µL ◦ AXX = {∅ 7→ ∅, {R, RY } 7→ {G, Y }, {G, Y } 7→ {R, RY },
{R, RY, G, Y } 7→ {R, RY, G, Y }}.
Example 5.12. Consider the language CTL and the Kripke structure K = (Σ, R, ℓ) depicted in Figure 5,
where the interpretation of temporal operators of CTL on K is standard. It is well known that the coarsest
s.p. partition PCTL can be obtained by refining the initial partition P = {1234, 5} induced by the labeling
ℓ through the Paige-Tarjan [42] algorithm, since PCTL coincides with bisimulation equivalence on K. It
is easy to check that PCTL = {12, 3, 4, 5}. This partition determines (see point (2) in Section 2.3) the
s.p. abstract Kripke structure depicted in Figure 5. Since CTL is closed under conjunction and negation,
by Proposition 5.10 (1) and (3), it turns out that the most abstract s.p. domain ACTL is partitioning and
coincides with the following partitioning closure:
adp (PCTL ) = {∅, 12, 3, 4, 5, 34, 35, 45, 122, 124, 125, 345, 1234, 1235, 1245, 12345}.
Let us now consider the following language L ∋ ϕ ::= p | q | ϕ1 ∧ ϕ2 | EF[0,2] ϕ, where EF[0,2] is
a time bounded reachability operator that is useful for quantitative temporal analysis [24], e.g., of discrete
real-time systems [10, Chapter 16]. The standard interpretation of EF[0,2] is as follows: s|=K EF[0,2] ϕ iff
there exists a path s0 s1 s2 s3 . . . in K starting from s = s0 and some n ∈ [0, 2] such that sn |=K ϕ. Let us
characterize the semantics of formulae in L :
JpKK = {1, 2, 3, 4}; JqKK = {5}; JEF[0,2] pKK = {1, 2, 3, 4, 5};
JEF[0,2] qKK = {3, 4, 5}; JEF[0,2] (EF[0,2] q)KK = {1, 2, 3, 4, 5};
Jp ∧ EF[0,2] qKK = {3, 4}; JEF[0,2] (p ∧ EF[0,2] q)KK = {1, 2, 3, 4, 5}.
Thus, ADL = M({JϕKK | ϕ ∈ L }) = {∅, 5, 34, 345, 1234, 12345}. On the other hand, by Proposition 5.10 (1), PL = par(ADL ) = {12, 34, 5}. In this case, it turns out that adp (PL ) ⊏ ADL . Moreover,
analogously to Example 2.2, let us show that there exists no abstract transition relation ♯ ⊆ PL ×PL that
determines an abstract Kripke structure A = (PL , ♯ , ℓL ) which strongly preserves L . Let B = {1, 2},
B ′ = {3, 4} and B ′′ = {5} be the blocks in PL . Assume by contradiction that such an abstract Kripke
structure A exists.
(i) On the concrete model K we have that 3|=K EF[0,2] q. Thus, by strong preservation, it must be that
B ′ |=A EF[0,2] q. On the other hand, if B ′ ♯ B and B♯ B ′′ then B|=A EF[0,2] q and therefore, by
weak preservation, we would have that 1|=K EF[0,2] q, which is a contradiction. Thus, necessarily,
B ′ ♯ B ′′ .
(ii) Let us observe that 1|=K EF[0,2] EF[0,2] q. Hence, by strong preservation, B|=A EF[0,2] EF[0,2] q. If
B♯ B ′′ then, as in point (i), we would still have that 1|=K EF[0,2] q, i.e. a contradiction. Hence,
necessarily, B♯ B ′ .
18
(iii) From B♯ B ′ and B ′ ♯ B ′′ , we would obtain that B|=A EF[0,2] q that, as observed in point (ii), is a
contradiction.
Thus, this shows that it is not possible to define an abstract Kripke structure on the abstract state space PL
L
that strongly preserves L . The abstract domain ADL induces a corresponding abstract semantics J·KAD
K
that instead strongly preserves L . In this case, the best correct approximation of the operator EF[0,2] on
ADL is:
µL ◦ EF[0,2] = {∅ 7→ ∅, 5 7→ 345, 34 7→ 12345, 345 7→ 12345,
1234 7→ 12345, 12345 7→ 12345}.
6 Strong Preservation and Completeness
In this section we establish a precise correspondence between generalized strong preservation of abstract
models and completeness of abstract interpretations, so that the problem of minimally refining an abstract
model in order to get strong preservation can be formulated as a complete domain refinement in abstract
interpretation.
6.1 Forward Complete Shells
Let us consider forward completeness of abstract domains A ∈ Abs(C) for generic n-ary concrete operations f : C n → C, with n ≥ 0. Hence, A is forward complete for f , or simply f -complete,
when f ◦ hµA , ..., µA i = µA ◦ f ◦ hµA , ..., µA i, that is, for any ~x ∈ C n , f (µA (x1 ), ..., µA (xn )) =
µA (f (µA (x1 ), ..., µA (xn ))). Equivalently, A is f -complete when for any ~a ∈ An , f (γ(a1 ), ..., γ(an )) =
γ(α(f (γ(a1 ), ..., γ(an )))). For a set of operations F ⊆ Fun(C), A is F -complete when A is f -complete
for each f ∈ F . Observe that F -completeness for an abstract domain A means that the associated closure
µA is closed under the image of functions in F , namely F (µA ) ⊆ µA . Also note that when k : C 0 → C,
i.e. k ∈ C is a constant, A is k-complete iff k is precisely represented in A, i.e. γ(α(k)) = k. Let
us also note that an abstract domain A ∈ Abs(C) is always forward meet-complete because any uco is
Moore-closed.
Let us first note that forward F -complete shells always exist. Let SF : Abs(C) → Abs(C) be defined
def
as SF (A) = ⊔ {X ∈ Abs(C) | X ⊑ A, X is F -complete}.
Lemma 6.1. SF (A) is the F -complete shell of A.
Proof. Let η = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete} = ∩{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete}.
Let f ∈ F , with ♯(f ) = n > 0 (if ♯(f ) = 0 then, trivially, f ∈ η) and ~c ∈ η n . Consider any ρ ∈ uco(C)
that is F -complete and such that ρ ⊑ µ. Since η ⊆ ρ, we have that ~c ∈ ρn and therefore f (~c) ∈ ρ because
ρ is F -complete. Thus, f (~c) ∈ η, i.e., η is F -complete.
A forward complete shell SF (A) is a more concrete abstraction than A. How to characterize SF (A)?
It is here useful to view abstract domains as closure operators on the concrete domain, i.e. as subsets
of C. Hence, A is viewed as the subset img(µA ) = γ(A) of the concrete domain C so that SF (A)
can be characterized as the least Moore-closed subset of C that contains img(µA ) and is forward F complete. We need to characterize the least amount of concrete information that must be added to γ(A) in
order to get forward completeness. It turns out that forward complete shells admit a constructive fixpoint
def
characterization. Let F uco : uco(C) → uco(C) be defined as follows: F uco (ρ) = M(F (ρ)), namely
uco
F (ρ) is the most abstract domain that contains the image of F on ρ. Observe that the operator λρ.µA ⊓
F uco (ρ) : uco(C) → uco(C) is monotone.
Lemma 6.2. SF (A) = gfp(λρ.µA ⊓ F uco (ρ)).
Proof. Observe that a uco ρ is F -complete iff F (ρ) ⊆ ρ iff M(F (ρ)) = F uco (ρ) ⊆ ρ iff ρ ⊑ F uco (ρ).
Thus, we have that SF (A) = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ is F -complete} = ⊔{ρ ∈ uco(C) | ρ ⊑ µA , ρ ⊑
F uco (ρ)} = ⊔{ρ ∈ uco(C) | ρ ⊑ µA ⊓ F uco (ρ)} = gfp(λρ.µA ⊓ F uco (ρ)).
19
Thus, it turns out that the lower iteration sequence of λρ.µA ⊓F uco (ρ) in uco(C) converges to the complete
shell SF (µA ).
Example 6.3. Let us consider the square operator on sets of integers sq : ℘(Z) → ℘(Z), i.e. sq(X) =
X 2 = {x2 | x ∈ X}, and the abstract domain Sign = {∅, Z<0 , {0}, Z>0 , Z}. As observed in Section 2.2.2, Sign is not forward complete for the square operator. Let us apply Lemma 6.2 in order to
compute the forward complete shell Ssq (Sign). Observe that
∅2 = ∅ ∈ Sign; {0}2 = {0} ∈ Sign; Z2<0 = Z2>0 = Z2 6∈ Sign.
Thus, the first step of iteration refines Sign to Sign ∪ {Z2 } (notice that this is an abstract domain because
2
it is Moore-closed). Then, (Z2 )2 = Z2 6∈ Sign ∪ {Z2 }, so that on the second step of iteration we obtain
2
k
Sign ∪ {Z2 , Z2 }. In general, for n ≥ 1, the n-th step of iteration provides Sign ∪ {Z2 | k ∈ [1, n]}, so
n
that the complete shell Ssq (Sign) coincides with the least fixpoint Sign ∪ {Z2 | n ≥ 1}.
Finally, the following easy observation will be useful later on.
Lemma 6.4. Let F, G ⊆ Fun(C). Then, SF = SG if and only if for any A ∈ Abs(C), A is F -complete
⇔ A is G-complete.
Proof. (⇒) If A is F -complete then A = SF (A) = SG (A) and therefore A is G-complete as well.
(⇐) This follows from SF (A) = ⊔{X ∈ Abs(C) | X ⊑ A, X is F -complete} = ⊔{X ∈ Abs(C) | X ⊑
A, X is G-complete} = SG (A).
6.2 Strong Preservation and Complete Shells
Let L be a language with atoms in AP L and operators in Op L and let S = (Σ, I) be a semantic structure
for L so that AP L and OpL denote, respectively, the corresponding sets of semantic interpretations of
atoms and operators. It turns out that forward completeness for AP L and OpL implies strong preservation for L .
Lemma 6.5. If A ∈ Abs(℘(Σ)) is forward complete for AP L and OpL then A is s.p. for L .
Proof. By Theorem 5.8, we show that A ⊑ ADL . Let us show by induction that for any ϕ ∈ L ,
JϕKS = γ(α(JϕKS )).
– ϕ ≡ p ∈ AP L : since A is forward complete for p, JpKS = p = γ(α(p)) = γ(α(JpKS )).
– ϕ ≡ f (ϕ1 , . . . , ϕn ) with f ∈ Op L :
Jf (ϕ1 , ..., ϕn )KS = [by definition]
f (Jϕ1 KS , ..., Jϕn KS ) = [by inductive hypothesis]
f (γ(α(Jϕ1 KS )), ..., γ(α(Jϕn KS ))) = [since A is forward complete for f ]
γ(α(f (γ(α(Jϕ1 KS )), ..., γ(α(Jϕn KS ))))) = [by inductive hypothesis and by definition]
γ(α(Jf (ϕ1 , ..., ϕn )KS )).
On the other hand, the converse is not true, that is strong preservation does not imply forward completeness, as shown by the following example.
Example 6.6. Let us consider again Example 5.7 where we showed that the partitioning abstract domain A = ℘(P )⊆ is s.p. for L . However, A is not forward complete for OpL = {pre }. In fact:
γ(α(pre (γ(α({3}))))) = γ(α(pre ({3}))) = γ(α({2, 3})) = {1, 2, 3} while pre (γ(α({3}))) =
pre ({3}) = {2, 3}.
Instead, it turns out that most abstract s.p. domains can be characterized as forward complete shells.
20
6.2.1 Complete Shells as Strongly Preserving Abstract Domains
Partition refinement algorithms for computing behavioural equivalences like bisimulation [42], simulation
equivalence [5, 35, 48] and (divergence blind) stuttering equivalence [32] are used in standard abstract
model checking to compute the coarsest strongly preserving partition of temporal languages like CTL∗ or
the µ-calculus for the case of bisimulation equivalence, ACTL∗ for simulation equivalence and CTL∗ -X
for stuttering equivalence.
Given a language L and a concrete state space Σ, these partition refinement algorithms work by iteratively refining an initial partition P within the lattice of partitions Part(Σ) until the fixpoint PL is
reached. The input partition P determines the set AP P of atoms and their interpretation IP as foldef
def
lows: AP P = {pB | B ∈ P } and IP (pB ) = B. More in general, any X ⊆ ℘(Σ) determines a set
{pX }X∈X of atoms with interpretation IX (pX ) = X. In particular, this can be done for an abstract domain A ∈ Abs(℘(Σ)) by considering its concretization γ(A) ⊆ Σ, namely A is viewed as a set of atoms
with interpretation IA (a) = γ(a). Thus, an abstract domain A ∈ Abs(℘(Σ)) together with a set of functions F ⊆ Fun(℘(Σ)) determine a language LA,F , with atoms in A, operations in F and endowed with
a semantic structure SA,F = (Σ, IA ∪ IF ) such that for any a ∈ A, IA (a) = γ(a) and for any f ∈ F ,
IF (f ) = f . Therefore, the most abstract s.p. domain ADLA,F generalizes in our framework the output
of a partition refinement algorithm for some language. Accordingly, we aim at characterizing ADLA,F
as the output of a refinement process of the initial domain A within the lattice Abs(℘(Σ)) of abstract domains. The following result shows that forward completeness for the operations in F is the right notion of
refinement to be used for the case of abstract domains.
Theorem 6.7. Let A ∈ Abs(℘(Σ)), F ⊆ Fun(℘(Σ)) and assume that LA,F is closed under infinite
logical conjunction. Then, ADLA,F = SF (A).
Proof. Since LA,F is closed under conjunction we have that ADLA,F = {JϕKSA,F | ϕ ∈ LA,F }. Let us
first prove that {JϕKSA,F | ϕ ∈ LA,F } ⊆ SF (A) by structural induction on ϕ ∈ LA,F :
– ϕ ≡ a ∈ A: JaKSA,F = IA (a) = γ(a) ∈ γ(A) ⊆ SF (A).
– ϕ ≡ f (ϕ1 , ..., ϕn ) with f ∈ F : Jf (ϕ1 , ..., ϕn )KSA,F = f (Jϕ1 KSA,F , ..., Jϕn KSA,F ), where, by
inductive hypothesis, Jϕi KSA,F ∈ SF (A). Therefore, since SF (A) is forward f -complete, we have
that f (Jϕ1 KSA,F , ..., Jϕn KSA,F ) ∈ SF (A).
Let us now prove the opposite inclusion. Let us first observe that ADLA,F is forward F -complete. For
simplicity of notation, consider f ∈ F with ♯(f ) = 1. If JϕKSA,F ∈ ADLA,F , where ϕ ∈ LA,F , then,
f (ϕ) ∈ LA,F and f (JϕKSA,F ) = Jf (ϕ)KSA,F ∈ ADLA,F . By Lemma 6.2, we know that SA (A) =
⊓α∈Ord (λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ), so that it is sufficient to prove by transfinite induction on
α ∈ Ord that
(λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F .
– α = 0: (λρ.µA ⊓ M(F (ρ)))0,↓ (⊤uco(℘(Σ)) ) = ⊤uco(℘(Σ)) = {Σ} ∈ γ(A) ⊆ ADLA,F .
– α + 1: By inductive hypothesis, (λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F . Moreover,
ADLA,F is Moore-closed and forward F -complete (hence closed under F ). Thus, M(F ((λρ.µA ⊓
M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ))) ⊆ ADLA,F , namely (λρ.µA ⊓M(F (ρ)))α+1,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F .
– limit ordinal α: This follows from
(λρ.µA ⊓ M(F (ρ)))α,↓ (⊤uco(℘(Σ)) ) = ⊓β<α (λρ.µA ⊓ M(F (ρ)))β,↓ (⊤uco(℘(Σ)) )
because, by inductive hypothesis, (λρ.µA ⊓ M(F (ρ)))β,↓ (⊤uco(℘(Σ)) ) ⊆ ADLA,F , for any β < α.
21
6.2.2 Strongly Preserving Abstract Domains as Complete Shells
Let us consider a language L , with atoms in AP L and operators in Op L , and a semantic structure
S = (Σ, I). As an immediate consequence of Theorem 6.7, the most abstract s.p. domain ADL for L
w.r.t. S can be characterized as the forward AP L ∪ OpL -complete shell of the most abstract domain
{Σ}.
Corollary 6.8. Let L be closed under infinite logical conjunction. Then, ADL = SAP L ∪OpL ({Σ}).
Let us also observe that ADL can be equivalently characterized as the forward OpL -complete shell
of an initial abstract domain M(AP L ) induced by atoms: ADL = SOpL (M(AP L )).
6.2.3 Strongly Preserving Partitions
Theorem 6.7 and Corollary 6.8 provide an elegant generalization of partition refinement algorithms for
strong preservation from an abstract interpretation perspective.
Given a language L with operators in Op L and a corresponding semantic structure S = (Σ, I), as
recalled in Section 6.2.1, an input partition P ∈ Part(Σ) for a partition refinement algorithm determines
the set AP L = {pB | B ∈ P } of atoms of L and their interpretation I(pB ) = B. Thus, M(AP L ) =
M(P ) = P ∪ {∅, Σ}. It turns out that the coarsest s.p. partition PL for L can be characterized in our
abstract domain-based approach as follows.
Corollary 6.9. Let L be closed under infinite logical conjunction.
(1) PL = par(SOpL (M(P ))).
(2) Let L be closed under logical negation. Then, adp (PL ) = SOpL (M(P )).
Proof. (1) By Corollary 6.8, ADL = SOpL (M(P )) and by Proposition 5.10 (1), PL = par(ADL ) =
par(SOpL (M(P ))).
(2) By Proposition 5.10 (1) and (3), Corollary 6.8 and point (1), adp (PL ) = adp (par(ADL )) = ADL =
SOpL (M(P )).
It is worth remarking that when L is not closed under negation, by Proposition 5.10 (3) and Corollary 6.9 (2), it turns out that adp (PL ) ⊏ SOpL (M(P )). This means that when L is not closed under
negation the output partition PL of any partition refinement algorithm for achieving strong preservation
for L is not optimal within the lattice of abstract domains.
Example 6.10. Let us consider the language L and the concrete Kripke structure K in Example 5.12.
The labeling determines the initial partition P = {p = 1234, q = 5} ∈ Part(Σ), so that M(P ) =
{∅, 1234, 5, 12345} ∈ Abs(℘(Σ)). Here, Op L = {∧, EF[0,2] }. Abstract domains are Moore-closed so
that SOpL = SEF[0,2] . Let us compute SEF[0,2] (M(P )).
A0 = M(P ) = {∅, 1234, 5, 12345}
A1 = A0 ⊓ M(EF[0,2] (A0 )) = M(A0 ∪ EF[0,2] (A0 ))
= M({∅, 1234, 5, 12345} ∪ {EF[0,2] ({5}) = 345}) = {∅, 5, 34, 1234, 12345}
A2 = A1
(fixpoint)
As already observed in Example 5.12, PL = {12, 34, 5} is such that adp (PL ) ⊏ µL and it is not possible
to define a strongly preserving abstract Kripke structure on the abstract space PL .
7 An Application to some Behavioural Equivalences
It is well known that some temporal languages like CTL, ACTL and CTL-X induce state logical equivalences that coincide with standard behavioural equivalences like bisimulation equivalence for CTL, (divergence blind) stuttering equivalence for CTL-X and simulation equivalence for ACTL. We derive here
a novel characterization of these behavioural equivalences in terms of forward completeness of abstract
interpretations.
22
7.1 Bisimulation Equivalence
Let K = (Σ, , ℓ) be a Kripke structure over some set AP of atomic propositions. A relation R ⊆ Σ × Σ
is a bisimulation on K if for any s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s) = ℓ(s′ );
(2) For any t ∈ Σ such that st, there exists t′ ∈ Σ such that s′ t′ and tRt′ ;
(3) s′ Rs, i.e. R is symmetric.
Since the empty relation is a bisimulation and bisimulations are closed under union, it turns out that the
largest (as a set) bisimulation relation exists. This largest bisimulation is an equivalence relation called
bisimulation equivalence and is denoted by ∼bis while Pbis ∈ Part(Σ) denotes the corresponding partition.
Thus, a partition P ∈ Part(Σ) is a bisimulation on K when P Pbis .
It is well known [4] that when K is finitely branching, bisimulation equivalence coincides with the
state equivalence induced by CTL, i.e., Pbis = PCTL (the same holds for CTL∗ and the µ-calculus, see
e.g. [19, Lemma 6.2.0.5]). Moreover, it is known (see e.g. [49, Section 12]) that it is enough to consider
finitary Hennessy-Milner logic [34], i.e. a language L1 including propositional logic and the existential
next operator in order to have that PL1 = Pbis :
L1 ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | EXϕ
where, as usual, the interpretation EX of EX in K is pre . A number of algorithms for computing
bisimulation equivalence exists [3, 23, 38, 42]. The Paige-Tarjan algorithm [42] runs in O(|| log(|Σ|))time and is the most time-efficient algorithm that computes bisimulation equivalence.
We recalled above that PL1 = PCTL . In our framework, this can be obtained as a consequence of the
fact that the most abstract s.p. domains for CTL and L1 coincide.
Lemma 7.1. Let K be finitely branching. Then, ADCTL = ADL1 = adp (Pbis ).
Proof. Let OpCTL = {∩, ∁, AX, EX, AU, EU, AR, ER} be the set of standard interpretations of the
operators of CTL on K, so that AX = pre
f and EX = pre . We show that µ ∈ uco(℘(Σ)) is forward
complete for OpCTL iff µ is forward complete for {∁, pre }. Assume that µ is forward complete for
{∁, pre }. Let us first prove that µ is forward complete for pre
f = AX:
µ ◦ pre
f ◦µ =
µ ◦ ∁ ◦ pre ◦∁ ◦ µ =
[by definition of pre
f ]
[as µ is complete for ∁]
µ ◦ ∁ ◦ pre ◦µ ◦ ∁ ◦ µ =
µ ◦ ∁ ◦ µ ◦ pre ◦µ ◦ ∁ ◦ µ =
[as µ is complete for pre ]
[as µ is complete for ∁]
∁ ◦ µ ◦ pre ◦µ ◦ ∁ ◦ µ =
∁ ◦ pre ◦µ ◦ ∁ ◦ µ =
[as µ is complete for pre ]
[as µ is complete for ∁]
∁ ◦ pre ◦∁ ◦ µ =
pre
f ◦µ
[by definition of pre
f ]
The following fixpoint characterizations are well known [10]:
– AU(S1 , S2 ) = lfp(λZ.S2 ∪ (S1 ∩ pre
f (Z)));
– EU(S1 , S2 ) = lfp(λZ.S2 ∪ (S1 ∩ pre (Z)));
– AR(S1 , S2 ) = gfp(λZ.S2 ∩ (S1 ∪ pre
f (Z)));
– ER(S1 , S2 ) = gfp(λZ.S2 ∩ (S1 ∪ pre (Z))).
23
Let us show that µ is forward complete for AU. The proofs for the remaining operators in OpCTL are
analogous. We need to show that µ(lfp(λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)))) = lfp(λZ.µ(S2 ) ∪ (µ(S1 ) ∩
pre
f (Z))). Let us show that µ is forward complete for the function λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)):
µ(µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (µ(Z)))) =
µ(µ(S2 ) ∪ (µ(S1 ) ∩ µ(pre
f (µ(Z))))) =
µ(µ(S2 ) ∪ µ(µ(S1 ) ∩ µ(pre
f (µ(Z))))) =
µ(S2 ) ∪ µ(µ(S1 ) ∩ µ(pre
f (µ(Z)))) =
µ(S2 ) ∪ (µ(S1 ) ∩ µ(pre
f (µ(Z)))) =
µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (µ(Z))).
[as µ is complete for pre
f ]
[as µ is complete for ∩]
[as µ is complete for ∪]
[as µ is complete for ∩]
[as µ is complete for pre
f ]
Observe that since µ is additive (and therefore continuous) we have that µ(∅) = ∅. Moreover, let us
show that from the hypothesis that K is finitely branching it follows that pre
f is continuous. First, notice
that pre
f is continuous iff pre is co-continuous. Hence, let us check that pre is co-continuous. Let
{Xi }i∈N be a decreasing chain of subsets of Σ and let x ∈ ∩i∈N pre (Xi ). Since K is finitely branching,
post ({x}) is finite so that there exists some k ∈ N such that for any j > 0, post ({x}) ∩ Xk =
post ({x}) ∩ Xk+j . Hence, there exists some z ∈ ∩i∈N Xi ∩ post ({x}), so that x ∈ pre (∩i∈N Xi ).
Therefore, since pre
f is continuous we also have that λZ.µ(S2 ) ∪ (µ(S1 ) ∩ pre
f (Z)) is continuous. We
can therefore apply Lemma 2.1 so that µ(lfp(λZ.µ(S2 )∪(µ(S1 )∩ pre
f (Z)))) = lfp(λZ.µ(S2 )∪(µ(S1 )∩
pre
f (Z))).
Thus, by Lemma 6.4, S{∁,pre } = SOpCTL , so that, by Corollary 6.8, ADL1 = ADCTL . Finally, since
K is finitely branching and L1 is closed under conjunction and negation, adp (PL1 ) = adp (Pbis ) =
adp (PL1 ) = ADL1 .
As a consequence of this and of the results in Section 6 (in particular of Corollary 6.9), any partition refinement algorithm Algbis for computing bisimulation equivalence on a finitely branching Kripke structure,
like those in [3, 23, 38, 42], can be characterized as a complete shell refinement as follows:
Algbis (P ) = par(S{∁,pre } (M(P ))).
Thus, Algbis is viewed as an algorithm for computing a particular abstraction, that is par, of a particular complete shell, that is S{∁,pre } . In particular, this holds for the Paige-Tarjan algorithm [42] and
leads to design a generalized Paige-Tarjan-like procedure for computing most abstract strongly preserving
domains [45].
Finally, our abstract intepretation-based approach allows us to give the following nice characterization of bisimulation for a partition P in terms of forward completeness for the corresponding partitioning
abstract domain adp (P ).
Theorem 7.2. Let P ∈ Part(Σ). Then, P is a bisimulation on K iff adp (P ) is forward complete for
{p | p ∈ AP } ∪ {pre }.
Proof. We view adp (P ) as a uco so that adp (P ) = {∪i Bi ∈ ℘(Σ) | {Bi } ⊆ P }. Let us first observe that P Pℓ iff adp (P ) is forward complete for {p ⊆ Σ | p ∈ AP }. On the one hand, since
p = {s ∈ Σ | p ∈ ℓ(s)}, if s ∈ p and s ∈ B, for some B ∈ P , then B ⊆ [s]ℓ ⊆ p. Hence, p is a union of
some blocks of P and therefore p ∈ adp (P ). On the other hand, if adp (P ) contains {p ⊆ Σ | p ∈ AP }
then, for any p ∈ AP , p is a union of some blocks in P . Thus, for any B ∈ P , either B ⊆ p or B ∩ p = ∅.
Consequently, if s ∈ B then B ⊆ [s]ℓ ∈ Pℓ .
Let us now note that adp (P ) is forward complete for pre iff for any block B ∈ P , pre (B) is a (possibly empty) union of blocks of P : this holds because pre is additive, and therefore if {Bi } ⊆ P
then pre (∪i Bi ) = ∪i pre (Bi ). The fact that, for some B ∈ P , pre (B) = ∪i Bi , for some blocks
{Bi } ⊆ P , implies that if s ∈ pre (B), i.e., st for some t ∈ B, then s ∈ Bj , for some j, and if s′ ∈ Bj
then s′ ∈ pre (B), i.e., s′ t′ for some t′ ∈ B, namely condition (2) of bisimulation for P holds. On
the other hand, if condition (2) of bisimulation for P holds then if s, s′ ∈ B ′ and s ∈ pre (B), for some
B, B ′ ∈ P , then s′ t′ for some t ∈ B, i.e., s′ ∈ pre (B), and therefore pre (B) is a union of blocks of
P . This closes the proof.
24
7.1.1 On the Smallest Abstract Transition Relation
As recalled in Section 2.3, the abstract Kripke structure A = (Pbis , ∃∃ , ℓ∃ ) strongly preserves CTL,
where B1 ∃∃ B2 iff there exist s1 ∈ B1 and s2 ∈ B2 such that s1 s2 , and ℓ∃ (B) = ∪s∈B ℓ(s). As a
simple and elegant consequence of our approach, it is easy to show that ∃∃ is the unique (and therefore
the smallest) abstract transition relation on Pbis that induces strong preservation for CTL.
Let K = (Σ, , ℓ) be finitely branching so that, by Lemma 7.1, ADL1 = adp (Pbis ) = ℘(Pbis ). Recall
that the concrete interpretation I induced by K is such that I(EX) = pre . By Theorem 5.9, the unique
interpretation of atoms and operations in L1 on the abstract domain ℘(Pbis ) that gives rise to a s.p. abstract
semantics is the best correct approximation I ℘(Pbis ) . Hence, if A = (Pbis , ♯ , ℓ♯ ) is strongly preserving for
CTL then the interpretation pre♯ of EX induced by A must coincide with I ℘(Pbis ) (EX). Consequently,
pre♯ = α ◦ pre ◦γ so that for any B1 , B2 ∈ Pbis , we have that B1 ♯ B2 iff B1 ∈ α(pre (γ({B2 }))).
Therefore, we conclude by observing that B1 ∈ α(pre (γ({B2 }))) iff B1 ∃∃ B2 .
We believe that a similar reasoning could be also useful for other languages L in order to prove that
the smallest abstract transition relation on PL that induces strong preservation exists. For example, this
has been proved for the case of ACTL by Bustan and Grumberg [5].
7.2 Stuttering Equivalence
Lamport’s criticism [37] of the next-time operator X in CTL/CTL∗ is well known. This motivated the
study of temporal logics CTL-X/CTL∗ -X obtained from CTL/CTL∗ by removing the next-time operator
and this led to study notions of behavioural stuttering-based equivalences [4, 22, 32]. We are interested
here in divergence blind stuttering (dbs for short) equivalence. Let K = (Σ, , ℓ) be a Kripke structure
over a set AP of atoms. A relation R ⊆ Σ × Σ is a divergence blind stuttering relation on K if for any
s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s) = ℓ(s′ );
(2) If st then there exist t0 , ..., tk ∈ Σ, with k ≥ 0, such that: (i) t0 = s′ ; (ii) for all i ∈ [0, k − 1],
ti ti+i and sRti ; (iii) tRtk ;
(3) s′ Rs, i.e. R is symmetric.
Observe that condition (2) allows the case k = 0 and this simply boils down to requiring that tRs′ . Since
the empty relation is a dbs relation and dbs relations are closed under union, it turns out that the largest
dbs relation relation exists. It turns out that this largest dbs relation is an equivalence relation called
dbs equivalence and is denoted by ∼dbs while Pdbs ∈ Part(Σ) denotes the corresponding partition. In
particular, a partition P ∈ Part(Σ) is a dbs relation on K when when P Pdbs .
De Nicola and Vaandrager [22, Theorem 3.2.5] showed that for finite Kripke structures and for an
interpretation of universal/existential path quantifiers over all the, possibly finite, prefixes, dbs equivalence
coincides with the state equivalence induced from the language CTL-X (this also holds for CTL∗ -X),
that is Pdbs = PCTL-X . This is not true with the standard interpretation of path quantifiers over infinite
paths, since this requires a divergence sensitive notion of stuttering (see the details in [22]). Groote and
Vaandrager [32] presented a partition refinement algorithm that computes the partition Pdbs in O(|Σ|||)time.
We provide a characterization of divergence blind stuttering equivalence as the state equivalence induced by the following language L2 that includes propositional logic and the existential until operator EU,
where the interpretation of the existential path quantifier is standard, i.e. over infinite paths:
L2 ∋ ϕ ::= p | ϕ1 ∧ ϕ2 | ¬ϕ | EU(ϕ1 , ϕ2 )
Since the transition relation is assumed to be total, let us recall that the standard semantics EU :
℘(Σ)2 → ℘(Σ) of the existential until operator is as follows:
EU (S1 , S2 ) = S2 ∪ {s ∈ S1 | ∃s0 , ..., sn ∈ Σ, with n ≥ 0, such that (i) s0 = s,
(ii) ∀i ∈ [0, n − 1]. si ∈ S1 and si si+1 , (iii) sn ∈ S2 }.
25
The following result characterizes a dbs partition P in terms of forward completeness for the corresponding
partitioning abstract domain adp (P ).
Theorem 7.3. Let P ∈ Part(Σ). Then, P ∈ Part(Σ) is a dbs partition on K iff adp (P ) is forward
complete for {p | p ∈ AP } ∪ {EU }.
Proof. As already shown in the proof of Theorem 7.2, it turns out that P Pℓ iff adp (P ) is forward
complete for {p ⊆ Σ | p ∈ AP }. Thus, it remains to show P ∈ Part(Σ) satisfies condition (2) of the
definition of dbs relation iff adp (P ) is forward complete for EU . Let us first observe that P ∈ Part(Σ)
satisfies this condition (2) iff for any B1 , B2 ∈ P , EU (B1 , B2 ) = B1 ∪ B2 .
(⇒) If B1 = B2 then EU (B1 , B1 ) = B1 . Otherwise, assume that B1 6= B2 . If B2 ( EU (B1 , B2 ) ⊆
B1 ∪ B2 then there exists s ∈ EU (B1 , B2 ) such that s ∈ B1 . Thus, if s′ ∈ B1 then, by condition (2),
s′ ∈ EU (B1 , B2 ). This implies that EU (B1 , B2 ) = B1 ∪ B2 .
(⇐) Let B ∈ P , s, s′ ∈ B and st. If t ∈ B then condition (2) is satisfied. Otherwise, t ∈ B ′ , for some
B ′ ∈ P , with B 6= B ′ . Thus, s ∈ EU (B, B ′ ) and therefore EU (B, B ′ ) = B ∪ B ′ . This means that
condition (2) is satisfied for P .
To complete the proof it is now sufficient to show that if, for any B1 , B2 ∈ P , EU (B1 , B2 ) = B1 ∪ B2
then adp (P ) is forward complete for EU , i.e., for any {Bi }i∈I , {Bj }j∈J ⊆ P , EU (∪i Bi , ∪j Bj ) =
∪k Bk , for some {Bk }k∈K ⊆ P . The function EU is additive in its second argument, thus we only need
to show that, for any B ∈ P , EU (∪i Bi , B) = ∪k Bk , namely if s ∈ EU (∪i Bi , B) and s ∈ B ′ , for
some B ′ ∈ P , then B ′ ⊆ EU (∪i Bi , B). If s ∈ EU (∪i Bi , B) and s ∈ B ′ , for some B ′ ∈ {Bi }i ,
then there exist n ≥ 0 and s0 , ..., sn ∈ Σ such that s0 = s, ∀j ∈ [0, n − 1].sj ∈ ∪i Bi and sj sj+1 , and
sn ∈ B. Let us prove by induction on n ∈ N that if s′ ∈ B ′ then s′ ∈ EU (∪i Bi , B).
(n = 0): In this case s ∈ ∪i Bi and s ∈ B = B ′ . Hence, for some k, s ∈ Bk = B = B ′ and therefore
s ∈ EU (B, B). By hypothesis, EU (B, B) = B. Moreover, EU is monotone on its first component
and therefore B ′ = B = EU (B, B) ⊆ EU (∪i Bi , B).
(n+1): Suppose that there exist s0 , ..., sn+1 ∈ Σ such that s0 = s, ∀j ∈ [0, n].sj ∈ ∪i Bi and sj sj+1 , and
sn+1 ∈ B. Let sn ∈ Bk , for some Bk ∈ {Bi }i∈I . Then, s ∈ EU (∪i Bi , Bk ) and s = s0 s1 ...sn .
Since this finite path has length n, by inductive hypothesis, s′ ∈ EU (∪i Bi , Bk ). Hence, there exist
r0 , ..., rm ∈ Σ, with m ≥ 0, such that s′ = r0 , ∀j ∈ [0, m − 1].rj ∈ ∪i Bi and rj rj+1 , and rm ∈ Bk .
Moreover, since sn sn+1 , we have that sn ∈ EU (Bk , B). By hypothesis, EU (Bk , B) = Bk ∪ B,
and therefore rm ∈ EU (Bk , B). Thus, there exist q0 , ..., ql ∈ Σ, with l ≥ 0, such that rm = q0 ,
∀j ∈ [0, l − 1].qj ∈ Bk and qj qj+1 , and ql ∈ B. We have thus found the following finite path:
s′ = r0 r1 ...rm = q0 q1 ...ql , where all the states in the sequence but the last one ql belong to
∪i Bi , while ql ∈ B. This means that s′ ∈ EU (∪i Bi , B).
As a consequence, we obtain a characterization of dbs equivalence as the state equivalence induced by
the standard interpretation of the language L2 .
Corollary 7.4. Let Σ be finite. Then, Pdbs = PL2 .
Proof. By definition, Pdbs = gPart(Σ) {P ∈ Part(Σ) | P is a dbs relation on K}. By Theorem 7.3,
Pdbs = gPart(Σ) {P ∈ Part(Σ) | adp (P ) is complete for {p | p ∈ AP } ∪ {EU }}. By Theorem 3.2, adp is co-additive on Part(Σ) , that is adp preserves lub’s in Part(Σ) . Hence, adp (Pdbs ) =
⊔Abs(℘(Σ)) {adp (P ) ∈ Abs(℘(Σ)) | P ∈ Part(Σ), adp (P ) is complete for {p | p ∈ AP} ∪ {EU }}.
By Theorem 3.2, Abspar (℘(Σ)) = {adp (P ) | P ∈ Part(Σ)} so that adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈
Abspar (℘(Σ)) | A is complete for {p | p ∈ AP} ∪ {EU }}. By Corollary 3.3, A ∈ Abspar (℘(Σ))
iff A is forward complete for ∁, so that adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈ Abs(℘(Σ)) | A is complete
for {p | p ∈ AP} ∪ {∁, EU }}. Then, we note that A is forward complete for {p | p ∈ AP } iff
A ⊑ M({p | p ∈ AP }). Hence, adp (Pdbs ) = ⊔Abs(℘(Σ)) {A ∈ Abs(℘(Σ)) | A ⊑ M({p | p ∈ AP}), A
is complete for {∁, EU }} = S{∁,EU } (M({p | p ∈ AP})). Finally, since Σ is finite and therefore closure under infinite conjunction boils down to closure under finite conjunction, by Corollary 6.8,
S{∁,EU } (M({p | p ∈ AP })) = ADL2 . Thus, by Proposition 5.10 (1), adp (Pdbs ) = ADL2 , so that
Pdbs = par(adp (Pdbs )) = par(ADL2 ) = PL2 .
26
As a consequence of Corollary 6.9, the Groote-Vaandrager algorithm [32] GV for computing dsb equivalence on a finite Kripke structure can be characterized as a complete shell refinement as follows:
GV(P ) = par(S{∁,EU } (M(P ))).
7.3 Simulation Preorder and Equivalence
Simulations are possibly nonsymmetric bisimulations, that is R ⊆ Σ × Σ is a simulation on a Kripke
structure K = (Σ, , ℓ) if for any s, s′ ∈ Σ such that sRs′ :
(1) ℓ(s′ ) ⊆ ℓ(s);
(2) For any t ∈ Σ such that st, there exists t′ ∈ Σ such that s′ t′ and tRt′ .
The empty relation is a simulation and simulation relations are closed under union, so that the largest simulation relation exists. It turns out that the largest simulation is a preorder relation called similarity preorder
and denoted by Rsim ∈ PreOrd(Σ). Therefore, a preorder relation R ∈ PreOrd(Σ) is a simulation on K
when R ⊆ Rsim . Simulation equivalence ∼simeq ⊆ Σ × Σ is the symmetric closure of Rsim : s ∼simeq s′
iff there exist two simulation relations R1 and R2 such that sR1 s′ and s′ R2 s. Psimeq ∈ Part(Σ) denotes
the partition corresponding to ∼simeq .
A number of algorithms for computing simulation equivalence have been proposed [2, 5, 12, 27, 35]
and some of them like [2, 35] first compute the similarity preorder and then from it they obtain simulation
equivalence. The problem of computing simulation equivalence is important in model checking because,
as recalled in Section 2.3, simulation equivalence strongly preserves ACTL so that Psimeq = PACTL (see
[33, Section 4]). Recall that ACTL is obtained by restricting CTL, as defined in Section 4.1, to universal
quantifiers and by allowing negation on atomic propositions only:
ACTL ∋ ϕ ::= p | ¬p | ϕ1 ∧ ϕ2 | ϕ1 ∨ ϕ2 | AXϕ | AU(ϕ1 , ϕ2 ) | AR(ϕ1 , ϕ2 )
It turns out that the most abstract s.p. domain for ACTL can be obtained as the most abstract s.p.
domain for the following sublanguage L3 :
L3 ∋ ϕ ::= p | ¬p | ϕ1 ∧ ϕ2 | ϕ1 ∨ ϕ2 | AXϕ
Lemma 7.5. Let K be finitely branching. Then, ADACTL = ADL3 .
Proof. Let OpACTL = {∩, ∪, AX, AU, AR} be the set of standard interpretations of the operators of
ACTL on K, so that AX = pre
f . Analogously to the proof of Lemma 7.1, as a consequence of
the least/greatest fixpoint characterizations of AU and AR, it turns out that for any A ∈ Abs(℘(Σ)),
A is forward complete for OpACTL iff A is forward complete for {∪, pre
f }. Thus, by Lemma 6.4,
=
AD
.
,
so
that,
by
Corollary
6.8,
AD
S{∪,pre
=
S
ACTL
L
Op
f }
3
ACTL
Thus, by Proposition 5.10 (1), PACTL = par(ADACTL ) = par(ADL3 ) = PL3 , so that Psimeq = PL3 .
As a further consequence, by Corollary 6.9, any algorithm Algsimeq that computes simulation equivalence
can be viewed as a partitioning abstraction of the {∪, pre
f }-complete shell refinement:
Algsimeq (P ) = par(S{∪,pre } (M(P ))).
An instantiation of the generalized Paige-Tarjan-like procedure in [45] for the complete shell S{∪,pre }
allows to design a new efficient abstract intepretation-based algorithm for computing simulation equivalence [46] whose space and time complexity is comparable with that of state-of-the-art algorithms like
[5, 27].
27
7.3.1 Preorders as Abstract Domains
Simulations give rise to preorders rather than equivalences like in the case of bisimulations and dbs relations. Thus, in order to characterize simulation for preorders as forward completeness for abstract domains
we need to view preorders as abstract domains. This can be obtained by generalizing the abstraction in
Section 3 from partitions to preorders.
def
Let R ∈ PreOrd(Σ) and for any x ∈ Σ let us define Rpre = {preR ({x}) ⊆ Σ | x ∈ Σ}. The preorder
R gives rise to an abstract domain ℘(Rpre )⊆ which is related to ℘(Σ)⊆ through the following abstraction
and concretization maps:
def
αR (S) = {preR ({x}) ⊆ Σ | x ∈ S}
def
γR (X ) = ∪X∈X X.
It is easy to check that from the hypothesis that R is a preorder it follows that (αR , ℘(Σ)⊆ , ℘(Rpre )⊆ , γR )
is indeed a GI. Hence, any R ∈ PreOrd(Σ) induces an abstract domain denoted by add (R) ∈ Abs(℘(Σ)).
Also, note that γR ◦ αR = preR , namely preR is the closure associated to add (R). The notation add comes
from the fact that an abstract domain A is equivalent to some add (R) if and only if A is disjunctive.
Lemma 7.6. {add (R) ∈ Abs(℘(Σ)) | R ∈ PreOrd(Σ)} = {A ∈ Abs(℘(Σ)) | A is disjunctive}.
Proof. Observe that γR is trivially additive, so that any add (R) is disjunctive. On the other hand, let
A ∈ Abs(℘(Σ)) be disjunctive and consider the relation RA = {(x, y) | α({x}) ≤A α({y})} which is
trivially a preorder. Thus, add (RA ) is disjunctive so that in order to conclude that add (RA ) is equivalent to
A it is enough to observe that for any y ∈ Σ, preRA ({y}) = γ(α({y})): this is true because γ(α({y})) =
{x ∈ Σ | α({x}) ≤A α({y})} = preRA ({y}).
Let us observe that add indeed generalizes adp from partitions to preorders because for any P ∈
Part(Σ), adp (P ) = add (R): this is a simple consequence of the fact that for a partition P viewed as an
equivalence relation and for x ∈ Σ, Px is exactly a block of P so that αP (S) = {preP ({x}) | x ∈ S}. On
the other hand, an abstract domain A ∈ Abs(℘(Σ)) induces a preorder relation preord(A) ∈ PreOrd(Σ)
as follows:
(x, y) ∈ preord(A) iff α({x}) ≤A α({y}).
It turns out that the maps add and preord allows to view the lattice of preorder relations as an abstraction
of the lattice of abstract domains.
Theorem 7.7. (preord, Abs(℘(Σ))⊒ , PreOrd(Σ)⊇ , add ).
Proof. Let A ∈ Abs(℘(Σ)) and R ∈ PreOrd(Σ). Let us prove that R ⊆ preord(A) ⇔ add (R) ⊑ γ ◦ α.
(⇒) Let S ⊆ Σ and let us show that add (R)(S) = preR (S) ⊆ γ(α(S)). If x ∈ preR (S) then xRy for
some y ∈ S, so that (x, y) ∈ preord(A), i.e. α({x}) ≤A α({y}). Thus, by applying γ, x ∈ γ(α({x})) ⊆
γ(α({y})) ⊆ γ(α(S)).
(⇐) Let (x, y) ∈ R and let us show that α({x}) ≤ α({y}). Note that x ∈ preR ({y}) = add (R)({y}) ⊆
γ(α({y})), so that α({x}) ≤A α({y}), namely (x, y) ∈ preord(A).
def
Let us remark that D = add ◦ preord is a lower closure operator on hAbs(℘(Σ)), ⊑i and that, by
Lemma 7.6, for any A ∈ Abs(℘(Σ)), A is disjunctive iff D(A) = A. Hence, D coincides with the
disjunctive-shell refinement, also known as disjunctive completion [14], namely D(A) is the most abstract
disjunctive refinement of A.
We can now provide a characterization of simulation preorders in terms of forward completeness.
Theorem 7.8. Let R ∈ PreOrd(Σ). Then, R is a simulation on K iff add (R) is forward complete for
{p | p ∈ AP } ∪ {pre
f }.
Proof. Recall that preR is the closure associated to add (R). We first observe that (sRs′ ⇒ ℓ(s′ ) ⊆ ℓ(s))
iff preR is forward complete for AP. On the one hand, if p ∈ AP and s ∈ preR (p) then sRs′ for some
s′ ∈ p, so that, from ℓ(s′ ) ⊆ ℓ(s), we obtain s ∈ p, and therefore preR (p) = p. On the other hand, if
sRs′ and s′ ∈ p, for some p ∈ AP , then s′ ∈ p = preR (p) so that preR ({s′ }) ⊆ preR (preR (p)) =
preR (p) = p and therefore from s ∈ preR ({s′ }) we obtain s ∈ p.
28
Thus, it remains to show that R satisfies condition (2) of the definition of simulation iff preR is forward
complete for pre .
(⇒) We prove that for any S, preR (pre
f (preR (S))) ⊆ pre
f (preR (S)). Let x ∈ preR (pre
f (preR (S)))
so that there exists some y ∈ pre
f (preR (S)) such that xRy. If xx′ , for some x′ , then, by simulation,
there exists some y ′ such that yy ′ and x′ Ry ′ . Hence, y ′ ∈ preR (S) and this together with x′ Ry ′ , as R is
transitive, gives x′ ∈ preR (S). Therefore, x ∈ pre
f (preR (S)).
(⇐) Observe that in order to show that R is a simulation it is enough to show that if xRy then x ∈
pre
f (preR (post ({y}))). The following implications hold, where post ({y}) ⊆ preR (post ({y}))
holds because preR is a uco:
post ({y}) ⊆ preR (post ({y})) ⇒
pre
f (post ({y})) ⊆ pre
f (preR (post ({y}))) ⇒
{y} ⊆ pre
f (preR (post ({y}))) ⇒
preR ({y}) ⊆ preR (pre
f (preR (post ({y})))) ⇒
preR ({y}) ⊆ pre
f (preR (post ({y}))) ⇒
x ∈ pre
f (preR (post ({y})))
[as pre
f is monotone]
[as y ∈ pre
f (post ({y}))]
[as preR is monotone]
[as preR is forward complete for pre
f ]
[as x ∈ preR ({y})]
and this closes the proof.
8 Related work
Loiseaux et al. [39] generalized the standard approach to abstract model checking to more general abstract models where an abstraction relation σ ⊆ States × A is used instead of a surjective function
h : States → A. However, the results of strong preservation given there (cf. [39, Theorems 3 and 4])
require the hypothesis that the relation σ is difunctional, i.e. σ = σσ −1 σ. In this case the abstraction
relation σ can indeed be derived from a function, so that the class of strongly preserving abstract models
in Loiseaux et al.’s framework is not really larger than the class of standard partition-based abstract models
(see the detailed discussion by Dams et al. [20, Section 8.1]).
Giacobazzi and Quintarelli [28] first noted that strong preservation is related to completeness in abstract interpretation by studying the relationship between complete abstract interpretations and Clarke et
al.’s [6, 7, 8] spurious counterexamples. Given a formula ϕ of ACTL, a model checker running on a standard abstract Kripke structure defined over a state partition P may provide a spurious counterexample π ♯
for ϕ, namely a path of abstract states, namely blocks of P , which does not correspond to a real concrete
counterexample. In this case, by exploiting the spurious counterexample π ♯ , the partition P is refined to P ′
by splitting a single block of P . As a result, this refined partition P ′ does not admit the spurious counterexample π ♯ for ϕ so that P ′ is given as a new refined abstract model for ϕ to the model checker. Giacobazzi
and Quintarelli [28] cast spurious counterexamples for a partition P as a lack of (standard) completeness
in the abstract interpretation sense for the corresponding partitioning abstract domain adp (P ). Then, by
applying the results in [31] they put forward a method for systematically refining abstract domains in order
to eliminate spurious counterexamples. The relationship between completeness and spurious counterexamples was further studied in [18], where it is also shown that a block splitting operation in Paige and
Tarjan [42] partition refinement algorithm can be characterized in terms of complete abstract interpretations. More in general, the idea of systematically enhancing the precision of abstract interpretations by
refining the underlying abstract domains dates back to the early works by Cousot and Cousot [14], and
evolved to the systematic design of abstract interpretations by abstract domain refinements [26, 29, 31].
9 Conclusion
This work shows how the abstract interpretation technique allows to generalize the notion of strong preservation from standard abstract models specified as abstract Kripke structures to generic domains in abstract
interpretation. For any inductively defined language L , it turns out that strong preservation of L in a
standard abstract model checking framework based on partitions of the space state Σ becomes a particular
29
instance of the property of forward completeness of abstract domains w.r.t. the semantic operators of the
language L . In particular, a generalized abstract model can always be refined through a fixpoint iteration
to the most abstract domain that strongly preserves L . This generalizes in our framework the idea of
partition refinement algorithms that reduce the state space Σ in order to obtain a minimal abstract Kripke
structure that is strongly preserving for some temporal language.
This work deals with generic temporal languages consisting of state formulae only. As future work,
it would be interesting to study whether the ideas of our abstract interpretation-based approach can be
applied to linear languages like LTL consisting of formulae that are interpreted as sets of paths of a Kripke
structure. The idea here is to investigate whether standard strong preservation of LTL can be generalized to
abstract interpretations of the powerset of traces and to the corresponding completeness properties. Fairness
can be also an interesting topic of investigation, namely to study whether our abstract interpretation-based
framework allows to handle fair semantics and fairness constraints [10].
Finally, let us mention that the results presented in this paper led to design a generalized Paige-Tarjan
refinement algorithm based on abstract interpretation for computing most abstract strongly preserving domains [45]. As shown in Section 6, a most abstract strongly preserving domain can be characterized as
a greatest fixpoint computation in Abs(℘(Σ)). It is shown in [45] that the Paige-Tarjan algorithm [42]
can be viewed exactly as a corresponding abstract greatest fixpoint computation in Part(Σ). This leads to
an abstract interpretation-based Paige-Tarjan-like refinement algorithm that is parameteric on any abstract
interpretation of the lattice Abs(℘(Σ)) of abstract domains of ℘(Σ) and on any generic inductive language
L.
Acknowledgements. We wish to thank Mila Dalla Preda and Roberto Giacobazzi who contributed to the
early stage of this work. This paper is an extended and revised version of [44]. This work was partially
supported by the FIRB Project “Abstract interpretation and model checking for the verification of embedded
systems” and by the COFIN2004 Project “AIDA: Abstract Interpretation Design and Applications”.
References
[1] K.R. Apt and G.D. Plotkin. Countable nondeterminism and random assignment. J. ACM, 33(4):724–767, 1986.
[2] B. Bloom and R. Paige. Transformational design and implementation of a new efficient solution to the ready
simulation problem. Sci. Comp. Program., 24(3):189–220, 1995.
[3] A. Bouajjani, J.-C. Fernandez and N. Halbwachs. Minimal model generation. In Proc. of the 2nd Internat. Conf.
on Computer Aided Verification (CAV’90), LNCS 531, pp. 197–203, Springer, 1990.
[4] M.C. Browne, E.M. Clarke and O. Grumberg. Characterizing finite Kripke structures in propositional temporal
logic. Theoret. Comp. Sci., 59:115–131, 1988.
[5] D. Bustan and O. Grumberg. Simulation-based minimization. ACM Trans. Comput. Log., 4(2):181–204, 2003.
[6] E.M. Clarke, O. Grumberg, S. Jha, Y. Lu and H. Veith. Counterexample-guided abstraction refinement. In Proc.
of the 12th Internat. Conf. on Computer Aided Verification (CAV’00), LNCS 1855, pp. 154–169, Springer, 2000.
[7] E.M. Clarke, O. Grumberg, S. Jha, Y. Lu and H. Veith. Counterexample-guided abstraction refinement for
symbolic model checking. J. ACM, 50(5):752–794, 2003.
[8] E.M. Clarke, S. Jha, Y. Lu and H. Veith. Tree-like counterexamples in model checking. In Proc. of the 17th IEEE
Symp. on Logic in Computer Science (LICS’02), pp. 19–29, IEEE Press, 2002.
[9] E.M. Clarke, O. Grumberg and D. Long. Model checking and abstraction. ACM Trans. Program. Lang. Syst.,
16(5):1512–1542, 1994.
[10] E.M. Clarke, O. Grumberg and D.A. Peled. Model checking. The MIT Press, 1999.
[11] R. Cleaveland, S.P. Iyer, D. Yankelevich. Optimality in abstractions of model checking. In Proc. 2nd Intern.
Static Analysis Symposium (SAS’95), LNCS 983, pp. 51–63, Springer, 1995.
[12] R. Cleaveland, J. Parrow and B. Steffen. The Concurrency Workbench: a semantics based tool for the verification
of concurrent systems. ACM Trans. Program. Lang. Syst., 15(1):36–72, 1993.
[13] P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by
construction or approximation of fixpoints. In Proc. 4th ACM POPL, pp. 238–252, 1977.
30
[14] P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In Proc. 6th ACM POPL, pp. 269–
282, 1979.
[15] P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination, projection and PER analysis of functional languages). In Proc. IEEE Int. Conf. on
Computer Languages (ICCL’94), pp. 95–112, 1994.
[16] P. Cousot and R. Cousot. Refining model checking by abstract interpretation. Automated Software Engineering
Journal, 6(1):69–95, 1999.
[17] P. Cousot and R. Cousot. Temporal abstract interpretation. In Proc. 27th ACM POPL, pp. 12–25, 2000.
[18] M. Dalla Preda. Completeness and stability in abstract model checking. Laurea Thesis (in Italian), Univ. of
Verona, Italy, 2003.
[19] D. Dams. Abstract interpretation and partition refinement for model checking. Ph.D. Thesis, Eindhoven University of Technology, The Netherlands, 1996.
[20] D. Dams, O. Grumberg and R. Gerth. Abstract interpretation of reactive systems. ACM Trans. Program. Lang.
Syst., 16(5):1512–1542, 1997.
[21] J.W. De Bakker, J.-J.C. Meyer and J.I. Zucker. On infinite computations in denotational semantics. Theoret.
Comp. Sci., 26(1-2):53–82, 1983.
[22] R. De Nicola and F. Vaandrager. Three logics for branching bisimulation. J. ACM, 42(2):458–487, 1995
[23] A. Dovier, C. Piazza and A. Policriti. An efficient algorithm for computing bisimulation equivalence. Theoret.
Comp. Sci., 311(1-3):221–256, 2004.
[24] E.A. Emerson, A.K. Mok, A.P. Sistla and J. Srinivasen. Quantitative temporal reasoning. In Proc. of the 2nd
Internat. Conf. on Computer Aided Verification (CAV’90), LNCS 531, pp. 136–145, Springer, 1990.
[25] E.A. Emerson and E.M. Clarke. Characterizing correctness properties of parallel programs using fixpoints. In
Proc. ICALP’80, LNCS 85, pp. 169–181, Springer, 1980.
[26] G. Filé, R. Giacobazzi and F. Ranzato. A unifying view of abstract domain design. ACM Comput. Surv.,
28(2):333–336, 1996.
[27] R. Gentilini, C. Piazza and A. Policriti. From bisimulation to simulation: coarsest partition problems. J. Automated Reasoning, 31(1):73-103, 2003.
[28] R. Giacobazzi and E. Quintarelli. Incompleteness, counterexamples and refinements in abstract model checking.
In Proc. 8th Intern. Static Analysis Symposium (SAS’01), LNCS 2126, pp. 356–373, Springer, 2001.
[29] R. Giacobazzi and F. Ranzato. Refining and compressing abstract domains. In Proc. 24th ICALP, LNCS 1256,
pp. 771–781, Springer, 1997.
[30] R. Giacobazzi and F. Ranzato. Optimal domains for disjunctive abstract interpretation. Sci. Comp. Program.,
32:177–210, 1998.
[31] R. Giacobazzi, F. Ranzato and F. Scozzari. Making abstract interpretations complete. J. ACM, 47(2):361–416,
2000.
[32] J.F. Groote and F. Vaandrager. An efficient algorithm for branching bisimulation and stuttering equivalence. In
Proc. ICALP’90, LNCS 443, pp. 626-638, Springer, 1990.
[33] O. Grumberg and D.E. Long. Model checking and modular verification. ACM Trans. Program. Lang. Syst.,
16(3):843–871, 1994.
[34] M. Hennessy and R. Milner. Algebraic laws for nondeterminism and concurrency. J. ACM, 32(1):137–161, 1985.
[35] M.R. Henzinger, T.A. Henzinger and P.W. Kopke. Computing simulations on finite and infinite graphs. In Proc.
36th FOCS, pp. 453–462, IEEE Press, 1995.
[36] T.A. Henzinger, R. Maujumdar and J.-F. Raskin. A classification of symbolic transition systems. ACM Trans.
Comput. Log., 6(1), 2005.
[37] L. Lamport. What good is temporal logic? In Information Processing ’83, pp. 657–668, IFIP North-Holland,
1983.
[38] D. Lee and M. Yannakakis. Online minimization of transition systems. In Proc. 24th ACM STOC, pp. 264–274,
1992.
[39] C. Loiseaux, S. Graf, J. Sifakis, A. Bouajjani and S. Bensalem. Property preserving abstractions for the verification of concurrent systems. Formal Methods in System Design, 6:1–36, 1995.
31
[40] D. Massé. Semantics for abstract interpretation-based static analyzes of temporal properties. In Proc. 9th Intern.
Static Analysis Symposium (SAS’02), LNCS 2477, pp. 428–443, Springer, 2002.
[41] D. Massé. Abstract domains for property checking driven analysis of temporal properties. In Proc. 10th Intern.
Conf. on Algebraic Methodology and Software Technology (AMAST’04), LNCS 3116, pp. 349–363, Springer,
2004.
[42] R. Paige and R.E. Tarjan. Three partition refinement algorithms. SIAM J. Comput., 16(6):973–989, 1987
[43] F. Ranzato and F. Tapparo. Making abstract model checking strongly preserving. In Proc. 9th Intern. Static
Analysis Symposium (SAS’02), LNCS 2477, pp. 411–427, Springer, 2002.
[44] F. Ranzato and F. Tapparo. Strong preservation as completeness in abstract interpretation. In Proc. 13th European
Symposium on Programming (ESOP’04), LNCS. 2986, pp. 18–32, Springer, 2004.
[45] F. Ranzato and F. Tapparo. An abstract interpretation-based refinement algorithm for strong preservation. In
Proc. 11th Intern. Conf. on Tools and Algorithms for the Construction and Analysis of Systems (TACAS’05),
LNCS 3440, pp. 140–156, Springer, 2005.
[46] F. Ranzato and F. Tapparo. An efficient algorithm for computing simulation equivalence based on abstract interpretation. In preparation, 2006.
[47] D.A. Schmidt. Closed and logical relations for over- and under-approximation of powersets. In Proc. 11th Intern.
Static Analysis Symposium (SAS’04), LNCS 3148, pp. 22–37, Springer, 2004.
[48] L. Tan and R. Cleaveland. Simulation revisited. In In Proc. 7th Intern. Conf. on Tools and Algorithms for the
Construction and Analysis of Systems (TACAS’01), LNCS 2031, pp. 480-495, Springer, 2001.
[49] R.J. van Glabbeek. The linear time - branching time spectrum. In Handbook of Process Algebra, pp. 3–99,
Elsevier, 2001.
32
| 6cs.PL
|
arXiv:cs/0104003v1 [cs.LO] 2 Apr 2001
Under consideration for publication in Theory and Practice of Logic Programming
1
Chain Programs for Writing
Deterministic Metainterpreters
DAVID A. ROSENBLUETH
Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas
Universidad Nacional Autónoma de México
Apdo. 20-726, 01000 México D.F.
Abstract
Many metainterpreters found in the logic programming literature are nondeterministic in
the sense that the selection of program clauses is not determined. Examples are the familiar
“demo” and “vanilla” metainterpreters. For some applications this nondeterminism is
convenient. In some cases, however, a deterministic metainterpreter, having an explicit
selection of clauses, is needed. Such cases include (1) conversion of or parallelism into
and parallelism for “committed-choice” processors, (2) logic-based, imperative-language
implementation of search strategies, and (3) simulation of bounded-resource reasoning.
Deterministic metainterpreters are difficult to write because the programmer must be
concerned about the set of unifiers of the children of a node in the derivation tree. We argue
that it is both possible and advantageous to write these metainterpreters by reasoning
in terms of object programs converted into a syntactically restricted form that we call
“chain” form, where we can forget about unification, except for unit clauses. We give two
transformations converting logic programs into chain form, one for “moded” programs
(implicit in two existing exhaustive-traversal methods for committed-choice execution),
and one for arbitrary definite programs. As illustrations of our approach we show examples
of the three applications mentioned above.
1 Introduction
Perhaps the most common use of metalogic is the definition and implementation of
metainterpreters (Safra and Shapiro 1986, Abramson and Rogers 1988, Kowalski
1990, Apt and Turini 1995). Many applications of metainterpreters are based on
concise definitions, like that of the “vanilla” metainterpreter, which can be easily
elaborated as required. Other applications, however, have been neglected, possibly
because of employing convoluted definitions. Examples are deterministic metainterpreters exhaustively traversing search spaces. Our purpose will be to present a
technique simplifying the design of deterministic metainterpreters. This technique
converts the object program into a form severely restricted in its syntax, thereby
facilitating reasoning about its search space.
Early works exploiting metainterpreters to great advantage are for example Bowen
and Kowalski’s amalgamation of language and metalanguage (Bowen and Kowalski
1982), Sergot’s “query-the-user” facility (Sergot 1982), and Shapiro’s “algorithmic
debugger” (Shapiro 1982). These metainterpreters, just as the familiar vanilla and
2
David A. Rosenblueth
demo metainterpreters, have nondeterministic definitions. Consider for example the
following demo predicate (Kowalski 1990, Kowalski 1995):
demo(T , P ) ← axiom(T , P ← Q ), demo(T , Q )
demo(T , P ∧ Q ) ← demo(T , P ), demo(T , Q )
demo(T , true) ←
This definition is nondeterministic because it is not determined, in the first clause of
the definition, what axiom in the theory T , having conclusion P , might be needed
to demonstrate P (Kowalski 1995, p. 229).
For some applications this nondeterminism is convenient. For others, however,
a deterministic metainterpreter, having an explicit selection of axioms, is desired.
A problem amenable to deterministic metainterpretation is that of exhaustively
traversing, using a committed-choice processor, the search space generated by a
logic program and a goal (Ueda 1987, Tamaki 1987). Another problem is that
of describing search strategies with logic programs equivalent to flowcharts, like
the programs of (Clark and van Emden 1981). Yet another application is Kowalski’s approach for reconciling reactive and rational agents with bounded-resource
metainterpreters (Kowalski 1995).
It is of course possible to write deterministic metainterpreters for logic programs.
Clark and Gregory were perhaps among the first to publish (Clark and Gregory
1985) one such metainterpreter. We show a slightly modified version of their metainterpreter in Fig. 1. The intended meanings of some of the predicates in this metainterpreter are as follows. Assume that the set of answer substitutions to the goal having A as its only subgoal is {θ1 , . . . , θn }. Then the predicate single call set(A, Aθs)
is intended to hold when Aθs is a list of the form [Aθ1 , . . . , Aθn ].
The predicate set(A, Bs, Aθs) is a generalisation of single call set. In this case the
goal Bs, which may have more than one subgoal, is of the form B1 ∧ . . . ∧ Bm ∧
true, and Aθs is of the form [Aθ1 , . . . , Aθn ], where {θ1 , . . . , θn } is the set of answer
substitutions for Bs. (Bs terminates by true to simplify the code.)
The predicate all set(ABs, Aθs), in turn, can be viewed as a generalisation of set,
where ABs is a list of clauses [A1 ← Bs1 , . . . , Am ← Bsm ] and Aθs is of the form
[A1 θ1,1 , . . . , A1 θ1,n1 , . . . , Am θm,1 , . . . , Am θm,nm ], where {θi,1 , . . . , θi,ni } is the set of
answer substitutions for Bsi . We refer the reader to (Clark and Gregory 1985) for
a thorough discussion of this metainterpreter.
As exhibited in Fig. 1, the programmer must be concerned about the set of unifiers
of the children of a node in the derivation tree (cf. the term instances predicate).
The reason for this concern, as we will see, is that variables in a logic program
can appear anywhere in a clause. This is an additional difficulty, absent in usual
nondeterministic metainterpreters.
The study of deterministic metainterpreters can be viewed as an attempt to
narrow the “gap” between the “don’t-know” form of logic programming, needed
for user-level applications, and the “don’t-care” form, useful for controlling the
execution of logic programs. Kowalski’s observation (Kowalski 1993) that the FGCS
project had experienced such a gap suggests looking at the work done in connection
with this project.
Chain Programs for Writing Deterministic Metainterpreters
3
Example of object-program representation:
definition(ap( , , ), [ap([ ], L, L) ← true,
ap([A|L′ ], M , [A|N ]) ← (ap(L′ , M , N ) ∧ true)
]
)←
Example of goal: ← single call set(ap(X , Y , [a, b]), Aθs)
Metainterpreter:
single call set(A, Aθs) ← definition(A, ABs),
all matches(A, ABs, ABθs ′ ),
all set(ABθs ′ , Aθs)
all set([ ], [ ]) ←
all set([A ← Bs|ABs], Aθs) ← set(A, Bs, Aθs 1 ),
all set(ABs, Aθs 2 ),
append(Aθs 1 , Aθs 2 , Aθs)
set(A, true, [A]) ←
set(A, B ∧ Bs, Aθs) ← single call set(B , Bθs),
term instances(B , Bθs, A ← Bs, ABθs),
all set(ABθs, Aθs)
all matches(A, [ ], [ ]) ←
all matches(A, [Aθ ← Bθs|ABs], [Aθ ← Bθs|ABθs ′ ]) ← copy(A, Aθ′ ),
unify(Aθ, Aθ′ ), !,
all matches(A, ABs, ABθs ′ )
all matches(A, [A′ ← Bs|ABs], ABs ′ ) ← all matches(A, ABs, ABs ′ )
term instances(B , [ ], C , [ ]) ←
term instances(B , [Bθ|Bθs], C , [Cθ|Cθs]) ← copy(f (B , C ), f (Bθ ′ , Cθ)),
unify(Bθ, Bθ ′ ),
term instances(B , Bθs, C , Cθs)
Figure 1. A deterministic, exhaustive-traversal metainterpreter for arbitrary definite programs. For readability, we use a string Xθ as the name of a variable taking as value an
instance of the value of X . Similarly, Xθs is the name of a variable taking as value a list
of instances of the value of X .
First Ueda (Ueda 1987), and then Tamaki (Tamaki 1987), published methods
converting a nondeterministic logic program into a deterministic version. The motivation for developing such methods was that of allowing the execution of or-parallel
programs by committed-choice processors (which are and parallel). Understanding
how these methods work might lead to key ideas for obtaining other deterministicevaluation methods. However, the considerable intricacy of these methods is an
obstacle for giving a clear and concise explanation of their central mechanisms.
4
David A. Rosenblueth
Trying to elucidate the principles on which these methods are based, we found an
important common characteristic. Both tend to hide, as it were, certain occurrences
of variables, in a list behaving like a stack. In particular, such occurrences are
those of variables that (1) receive a substitution before a subgoal A is selected
and (2) occur in a subgoal selected after A has succeeded. Variables having such
occurrences are called “pass-on” variables in (Tamaki 1987). Hence, if we consider
programs lacking pass-on variables, a fortiori, these transformations get simplified.
A kind of program lacking pass-on variables is that of “chain” programs, having
clauses of the form:
p(X0 , Xn ) ← q1 (X0 , X1 ), q2 (X1 , X2 ), . . . , qn (Xn−1 , Xn )
(as well as other clauses; we formally define chain programs in Sect. 2). Apparently,
concentrating on chain programs should simplify the task of devising deterministictraversal methods in general, and defining deterministic metainterpreters in particular. We confirm this possibility by deriving in Sect. 2 a deterministic metainterpreter
by reasoning in terms of relational union and composition.
This metainterpreter is more useful once we provide ways of transforming logic
programs that do not have chain form into such a form. The methods of (Ueda
1987) and (Tamaki 1987) can handle “moded” programs, that do not necessarily have chain form. In these programs, each argument place of each predicate is
used either as input (instantiated) or as output (uninstantiated). By comparing the
original methods with their versions simplified to chain programs, we have uncovered a transformation converting moded programs into chain form. In Sect. 3 we
give such a transformation, which is in fact implicit in (Ueda 1987) and (Tamaki
1987). As in the methods of Ueda and Tamaki, we hide pass-on variables in a list
behaving like a stack. (We have previously used this transformation for adapting
parsers for context-free languages obtaining inference systems for moded logic programs (Rosenblueth 1996, Rosenblueth and Peralta 1998).)
Next, in Sect. 4 we give another transformation, converting arbitrary definite
programs into chain form, inspired by the previous one.
Once we have this more general transformation, we can easily extend, in Sect. 5,
the existing methods for deterministic, exhaustive traversal (Ueda 1987, Tamaki
1987) to handle arbitrary definite programs.
Section 6 shows how to write deterministic metainterpreters for other applications. First we exhibit a metainterpreter having the same behaviour as that of
Prolog systems. Next we give a bounded-resource metainterpreter.
We will assume some familiarity with logic program transformation through the
unfold/fold rules (Pettorossi and Proietti 1994).
Code appearing throughout the sequel, as well as further examples of programs
converted into chain form, may be found at:
http://leibniz.iimas.unam.mx/~drosenbl/detmeta.
5
Chain Programs for Writing Deterministic Metainterpreters
2 A Deterministic, Exhaustive-Traversal Metainterpreter for Chain
Programs
In this section, we will write a deterministic, exhaustive-traversal metainterpreter
for chain programs by reasoning in terms of relational union and composition.
A chain program can be viewed as defining a system of equations of relational
expressions. Such systems of relational equations have been studied for example
in (Engelfriet 1974, de Bakker and Meertens 1975, Blikle 1977). Our translation
of chain programs into relational equations enables us to regard logical inference
from such programs as the evaluation of relational expressions built from union
and composition. Our chain programs are similar to, but different from, certain
programs occurring in the deductive-database literature under the same name; a
difference is that we allow function symbols other than constants.
For clarity, we will discuss now this metainterpreter assuming that only ground
terms are constructed. However, in Sect. 4 we will observe that with the addition
of variable renaming and full unification (as opposed to matching, i.e. one-way
unification), this metainterpreter is also valid for constructing terms with variables.
2.1 Chain programs as systems of relational equations
We define a chain program as a logic program consisting only of clauses of the form:
p(X0 , Xn ) ← q1 (X0 , X1 ), q2 (X1 , X2 ), . . . , qn (Xn−1 , Xn ) n > 0
′
or p(t , t ) ←
(1)
(2)
where the Xi ’s are distinct variables, and t and t ′ are any term. The first argument
place of a predicate will be called its input, and the second argument place its
output. It will be useful to single out a kind of chain program where all answers for
such a program and a goal with a leftmost ground input are ground, assuming a
leftmost computation rule. If var(t ′ ) ⊆ var(t ) in every clause of the form (2), then
the program is called a G-chain program. Here, and throughout the sequel, we use
var(t ) to denote the set of variables in expression t .
Clearly, the clause (1) denotes the inclusion:
P ⊇ (Q1 ; Q2 ; . . . ; Qn )
(3)
where P , Qi name the relations denoted by p, qi , respectively, and “;” denotes
relational composition.1
Let us define
··
Pj = Q1 ; Q2 ; . . . ; Qn
if the j -th clause defining the predicate with symbol p has the form (1). If, on the
other hand, the j -th clause defining the predicate with symbol p has the form (2),
then
··
Pj = {(x, x′ ) : p(x, x′ ) ← is a ground instance of p(t , t ′ ) ←}
1
··
The composition of relations P and Q is defined as: P ; Q = {(x , z ) : ∃y[(x , y) ∈ P & (y, z ) ∈
Q]}.
6
David A. Rosenblueth
Hence, a chain program denotes a system of relational expressions having an
inclusion of the form
P ⊇ (P1 ∪ P2 ∪ . . . ∪ Pm )
for each largest set of clauses defining a predicate with symbol p. Just as sometimes
the meaning of a logic program is defined as its least Herbrand model, here we are
interested in the least solution of this system. It can be shown that such a solution
is equal to the unique solution of the system obtained by replacing the inclusions
by equalities.
For notational convenience, we will extend relational composition to the case
where the first argument is a set. Let S ⊆ D and R ⊆ D × D .
··
S ; R = {z : ∃y[y ∈ S & (y, z ) ∈ R]}
i.e. S ; R denotes the image of S under R. Also, we will sometimes omit the curly
braces in singletons. By I we will denote the identity relation and by ∅ the empty
relation as well as the empty set.
Given an object chain program, we will use X as a metavariable taking as value
a ground term, Xs a set of ground terms, Q a relation denoted by an (objectlevel) predicate, Qs compositions Q1 ; . . . ; Qn of relations denoted by predicates,
Pj a relation denoted by a single clause, and Pjs unions P1 ∪ . . . ∪ Pm of relations
denoted each by a single clause.
Computing all answers for a chain program and a goal ← q(x, Z ), where x is
a ground term, translates to evaluating the expression x ; Q , where Q names the
relation denoted by q. During this evaluation, we will have to evaluate relational
expressions of the form:
Xs ; Qs
(4)
which represent and branches of the SLD tree.
2.2 A metainterpreter as a relational-expression evaluator
Let us now establish an object-program representation. For simplicity, we use the
ambivalent syntax of (Jiang 1994). If the j -th clause defining a predicate with
symbol p is of the form (1) we will have the following clause at the metalevel:
nonunit(pj , (q1 ; q2 ; . . . ; qn ; I)) ←
where “;” can be interpreted as a right-associative infix list constructor and I as a
constant. If the j -th clause defining a predicate with symbol p is of the form (2) we
will have:
is unit(pj ) ←
and
′
unit(pj , t , t ) ←
(The is unit predicate is clearly unnecessary, as could be defined using the unit
predicate, but we will use it for readability.)
Chain Programs for Writing Deterministic Metainterpreters
7
Also, for each largest set of clauses defining a predicate with symbol p we will
have:
defn(p, (p1 ∪ p2 ∪ . . . ∪ pm ∪ ∅)) ←
where ∪ can be interpreted as a right-associative infix list constructor denoting set
union and ∅ as a constant.
In general, we could have two main evaluation strategies for (4): “termwise” (i.e.
decomposing Xs) and “relationwise” (i.e. decomposing Qs). Here we will concentrate on a termwise definition (but see (Rosenblueth 1998) for an application of a
metainterpreter using a relationwise definition).
Consider (4). The predicate a(Xs, Qs, Zs) is intended to hold if Zs is Xs;Qs. Then
the next two clauses, which constitute a termwise definition of this predicate, follow
from the distributivity of composition over union:
a(∅, Qs, ∅) ←
a({X } ∪ Xs, Qs, YsZs) ← a ′ (X , Qs, Ys), a(Xs \ {X }, Qs, Zs),
union(Ys, Zs, YsZs)
where a ′ (X , Qs, Ys) is meant to hold when Ys is X ; Qs and union(Ys, Zs, YsZs) is
meant to hold when YsZs is Ys ∪ Zs.
We can decompose Qs in the definition of a ′ , which follows from the definition
of composition:
a ′ (X , I , {X }) ←
a ′ (X , (Q ; Qs), Zs) ← b ′ (X , Q , Ys), a(Ys, Qs, Zs)
where b ′ (X , Q , Ys) represents the composition Ys of a single term X with a single
relation Q .
The following clause translates such a composition into the composition of X
with a union Pjs of relations:
b ′ (X , Q , Ys) ← defn(Q , Pjs), c ′ (X , Pjs, Ys)
where c ′ (X , Pjs, Ys) is intended to hold if Ys is X ; Pjs.
Now we inductively define c ′ by decomposing Pjs. This definition of c ′ , as that
of a, follows from the distributivity of composition over union:
c ′ (X , ∅, ∅) ←
c ′ (X , {Pj} ∪ Pjs, YsZs) ← d ′ (X , Pj, Ys), c ′ (X , Pjs \ {Pj}, Zs),
union(Ys, Zs, YsZs)
where the predicate d ′ (X , Pj, Ys) is assumed to hold when Ys is X ; Pj, and the
predicate union(Ys, Zs, YsZs) is assumed to hold when YsZs is Ys ∪ Zs.
Next we write a definition of d ′ (X , Pj, Ys). This definition uses an auxiliary
predicate e ′ (X , Pj, Ys) in case Pj represents a unit clause. If, on the other hand, Pj
represents a nonunit clause, then the d ′ predicate uses the (object-level) definition
8
David A. Rosenblueth
of Pj to translate X ; Pj into X ; Qs, so that the previously defined predicate
a ′ (X , Qs, Zs) can be used.
d ′ (X , Pj, Ys) ← is unit(Pj), e ′ (X , Pj, Ys)
d ′ (X , Pj, Zs) ← nonunit(Pj, Qs), a ′ (X , Qs, Zs)
It only remains to define e ′ (X , Pj, Ys), intended to hold if Ys is X ; Pj and Pj
represents a unit clause.
e ′ (X , Pj, {Y }) ← unit(Pj, X , Y )
′
e (X , Pj, ∅) ← not(unit(Pj, X , ))
(5)
(6)
The clause (5) covers the case where X matches the input of the unit clause named
Pj, whereas the clause (6) covers the case where X does not match the input of the
unit clause named Pj.
Finally, we give the complete metainterpreter in a more standard Prolog notation.
In addition, we have approximated set union with list concatenation. (A more
efficiently executable metainterpreter would use difference lists, but for clarity we
prefer ordinary lists.) We will call this the abcde metainterpreter.
a([ ], Qs, [ ]) ←
a([X |Xs], Qs, YsZs) ← a ′ (X , Qs, Ys), a(Xs, Qs, Zs),
append(Ys, Zs, YsZs)
(7)
a ′ (X , [ ], [X ]) ←
a ′ (X , [Q |Qs], Zs) ← b ′ (X , Q , Ys), a(Ys, Qs, Zs)
(8)
b ′ (X , Q , Ys) ← defn(Q , Pjs), c ′ (X , Pjs, Ys)
c ′ (X , [ ], [ ]) ←
c ′ (X , [Pj|Pjs], YsZs) ← d ′ (X , Pj, Ys), c ′ (X , Pjs, Zs),
append(Ys, Zs, YsZs)
(9)
d ′ (X , Pj, Ys) ← is unit(Pj), e ′ (X , Pj, Ys)
d ′ (X , Pj, Zs) ← nonunit(Pj, Qs), a ′ (X , Qs, Zs)
(10)
e ′ (X , Pj, [Y ]) ← unit(Pj, X , Y )
e ′ (X , Pj, [ ]) ← not(unit(Pj, X , ))
3 Conversion of Moded Programs into Chain Form
Having written a deterministic metainterpreter for chain programs, our aim now is
to develop a transformation converting “moded” programs (Apt 1997) into chain
Chain Programs for Writing Deterministic Metainterpreters
9
form. We will derive such a transformation with unfold/fold rules (Pettorossi and
Proietti 1994). This transformation is in fact implicit in two existing methods for
deterministic, exhaustive traversal: the continuation-based (Ueda 1987) and the
stream-based (Tamaki 1987) methods. We have previously used such a transformation for adapting parsers for context-free grammars obtaining inference systems for
moded logic programs (Rosenblueth 1996, Rosenblueth and Peralta 1998).
A clause
′
p(t0 , tn′ ) ← q1 (t0′ , t1 ), q2 (t1′ , t2 ), . . . , qn (tn−1
, tn )
n≥0
(11)
is called moded if:
1. var(ti′ ) ⊆ var(t0 ) ∪ · · · ∪ var(ti ), for i = 0, . . . , n; and
2. var(ti ) ∩ var(tj ) = ∅, for i, j = 0, . . . , n and i 6= j .
A program is called moded if it consists only of moded clauses.
Condition 1 causes every input to be ground if the input of the initial goal is
also ground and we use a leftmost computation rule. When a subgoal succeeds,
condition 2 causes the constructed term to have an effect only on the input of other
subgoals, thus avoiding speculative bindings. In (Rosenblueth 1996, Rosenblueth
and Peralta 1998) we had a third condition (that each variable occurring in ti′
occurs only once in ti′ , for i = 0, . . . , n, if n > 0) meant only for simplifying the
transformation. However, as we observe below, it is possible to eliminate such a
condition without excessively elaborating the transformation.
We will use the standard equality theory. Given a program P , this theory consists
of the following axioms:
X =X ←
X =Y ←Y =X
X = Z ← X = Y, Y = Z
{f (X1 , . . . , Xnf ) = f (Y1 , . . . , Ynf ) ← X1 = Y1 , . . . , Xnf = Ynf :
f is a function symbol occurring in P }
{p(X1 , . . . , Xnp ) ← X1 = Y1 , . . . , Xnp = Ynp , p(Y1 , . . . , Ynp ) :
p is a predicate symbol occurring in P }
which are called, reflexivity, symmetry, transitivity, function substitutivity, and
predicate substitutivity, respectively.
One way of obtaining a chain clause from a moded clause would be to resolve first
the moded clause with predicate substitutivity so as to replace each argument of a
subgoal by a variable and then to fold the resulting clause using some new predicates
so as to remove the introduced equations. We will see, however, that such a folding
operation may not always be sound (Gardner and Shepherdson 1992, Tamaki and
Sato 1984).
Consider the following append program used for splitting lists.
s(hLi, h[ ], Li) ←
(12)
10
David A. Rosenblueth
s (h[A|N ]i, h[A|L], M i) ← s (hN i , hL, M i)
|{z} |{z} | {z }
|{z} | {z } | {z }
q1 t ′
p
t0
t1
t′
1
(13)
0
We employ angled brackets h i instead of ordinary brackets [ ] for grouping input
and output arguments. We do this for clarity.
By predicate substitutivity and symmetry, it is possible to derive from (13):
s(X , Y ) ← X = h[A|N ]i , Y = h[A|L], M i, X ′ = hN i , Y ′ = hL, M i,
s(X ′ , Y ′ )
(14)
Next, we could fold (14) using the following definitions:
naive h0 (X , X ′ ) ↔ ∃A ∃N (X = h[A|N ]i & X ′ = hN i)
naive h1 (Y ′ , Y ) ↔ ∃A ∃L ∃M (Y ′ = hL, M i & Y = h[A|L], M i)
In the case of naive h0 , this folding operation would replace the subgoals enclosed
in rectangles by the definiendum naive h0 (X , X ′ ). However, as observed in (Tamaki
and Sato 1984), in general it is incorrect to fold a clause such as (14) using a definition such as that of naive h0 , where a variable like A: (a) appears in atoms replaced
by the definiendum (i.e. X = h[A|N ]i), (b) appears in atoms not replaced by the
definiendum (i.e. Y = h[A|L], M i), and (c) does not unify with any variable appearing in the definiendum. To see this incorrectness, fold and subsequently unfold (14)
with naive h0 , and a generalisation of (14) is obtained.
Note that unlike A in (13), N (which does not cause the incorrectness of Tamaki
and Sato) occurs only in t0 and t0′ . This suggests that a minimal strengthening of
the syntactic conditions defining moded clauses so as to avoid the situation above
would be requiring that all variables of ti′ occur in ti (for i = 0, . . . , n). Hence, we
define a clause in prechain form as:
′
p̂(s0 , sn′ ) ← q̂1 (s0′ , s1 ), q̂2 (s1′ , s2 ), . . . , q̂n (sn−1
, sn )
n≥0
where:
1. var(si′ ) ⊆ var(si ) for i = 0, . . . , n; and
2. var(si ) ∩ var(sj ) = ∅, for i, j = 0, . . . , n and i 6= j .
From a clause in prechain form it is possible to arrive at chain form by first
applying predicate substitutivity and then folding with the completed definitions
of predicates of the form:
hi (si , si′ ) ←
Such foldings satisfy also Gardner and Shepherdson’s stronger condition (Gardner
and Shepherdson 1992) for folding.
We face now the problem of converting a moded clause into prechain form. Recall
first the pass-on variables of the deterministic, exhaustive-traversal methods (Ueda
1987, Tamaki 1987), i.e. variables that receive a substitution before a subgoal B is
selected and occur in a subgoal selected after B has succeeded. Observe next that
Chain Programs for Writing Deterministic Metainterpreters
11
prechain form clauses lack pass-on variables. Finally, note that the fact that these
methods use a stack to hide pass-on variables, suggests that we also use a stack to
achieve prechain form.
Since both chain form and prechain form lack pass-on variables, we could have
chosen prechain form for the object programs of our deterministic metainterpreter.
The resulting metainterpreter, however, would not have been as concise.
In converting a moded clause into prechain form, we will often apply the following
sequence of unfolding steps, which we group in a lemma.
Lemma 1 (Equation introduction)
Let C be a definite clause having an occurrence of a variable X . Then the clause
obtained by replacing that occurrence of X by X ′ and adding the equation X = X ′
is logically implied by C and the standard equality theory for C .
Proof
First, we resolve C with predicate substitutivity as follows. If the occurrence of X
is in the head of C , then we select the subgoal of predicate substitutivity which is
not an equation. Next, we apply symmetry to all equations of the resolvent. The
resulting clause has the form:
p(Y1 , . . . , Yd ) ← t1 = Y1 , . . . , td = Yd , A1 , . . . , An
If, on the other hand, the occurrence of X is in a subgoal Ai of C , then we select
such a subgoal. The resulting clause has the form:
A0 ← A1 , . . . , Ai−1 , t1 = Y1 , . . . , td = Yd , p(Y1 , . . . , Yd ), Ai+1 , . . . , An
In either case, the occurrence of X is in some equation tj = Yj .
Next, we apply function substitutivity to tj = Yj as many times as it is necessary
to make such an occurrence appear at the top level of an equation X = X ′ .
Finally, we apply reflexivity to all equations except X = X ′ , thus disposing of all
unwanted equations. The resulting clause has the claimed form. Hence the lemma
holds.
Often, we will use equation introduction followed by symmetry, to which we will
also refer as equation introduction.
Example 1
Let us convert (13) first into prechain form and then into chain form. To achieve
prechain form, we will use an auxiliary stack, which we introduce in the following
predicate:
ŝ(hSt, X i, hSt, Y , Z i) ← s(hX i, hY , Z i)
We first apply equation introduction to the definition of ŝ:
ŝ(hSt, X i, hSt′ , Y , Z i) ← St = St′ , s(hX i, hY , Z i)
(15)
Next, we apply equation introduction to (13).
s(h[A|N ]i, h[A′ |L], M i) ← A = A′ , s(hN i, hL, M i)
(16)
12
David A. Rosenblueth
Subsequently, we resolve (15) and (16) unifying the two underlined atoms.
ŝ(hSt, [A|N ]i, hSt′ , [A′ |L], M i) ← St = St′ , A = A′ ,
s(hN i, hL, M i)
Next, we fold using function substitutivity.
ŝ(hSt, [A|N ]i, hSt′ , [A′ |L], M i) ← [A|St] = [A′ |St′ ] ,
s(hN i, hL, M i)
Finally, we fold using (15) and systematically rename variables.
ŝ(hSt0 , [A0 |N0 ]i, hSt1 , [A1 |L1 ], M1 i) ← ŝ(h[A0 |St0 ], N0 i, h[A1 |St1 ], L1 , M1 i)
Once we have prechain form, we can apply predicate substitutivity and symmetry,
and then fold w.r.t. the completed definitions of:
h0 (hSt0 , [A0 |N0 ]i, h[A0 |St0 ], N0 i) ←
h1 (h[A1 |St1 ], L1 , M1 i, hSt1 , [A1 |L1 ], M1 i) ←
arriving at:
ŝ(X0 , X3 ) ← h0 (X0 , X1 ), ŝ(X1 , X2 ), h1 (X2 , X3 )
(17)
End of example
In (Rosenblueth 1996, Rosenblueth and Peralta 1998) we required that each variable occurring in ti′ occurs only once in ti′ . Note that if there is more than one
occurrence of a variable in t1′ , like in:
r (h[A|N ]i, h[A|L], M , A i) ← s(hN i, hL, M i)
thus violating such a condition, we would have an extra equation after applying
equation introduction twice:
r (h[A|N ]i, h[A′ |L], M , A′′ i) ← A = A′ , A = A′′ , s(hN i, hL, M i)
Note that we do not wish to eliminate A = A′′ with reflexivity, since we would
obtain a clause with a variable occurring both in the input and the output of the
head, preventing us from folding the h’s. However, equations such as this one can
be eliminated by factoring all equations with the same left-hand side. Hence we
have withdrawn such a condition.
This derivation suggests a proof of a theorem relating a moded clause with its
chain form.
Theorem 1
Let C be a moded clause:
′
, tn )
p(t0 , tn′ ) ← q1 (t0′ , t1 ), q2 (t1′ , t2 ), . . . , qn (tn−1
n≥0
and let
Πj = var(t0 ) ∪ · · · ∪ var(tj −1 ) ∩ var(tj′ ) ∪ · · · ∪ var(tn′ )
j = 0, . . . , n + 1
13
Chain Programs for Writing Deterministic Metainterpreters
Then the clause Ĉ :
p̂(U0 , Un′ ) ← h0 (U0 , U0′ ), q̂1 (U0′ , U1 ), h1 (U1 , U1′ ), q̂2 (U1′ , U2 ), . . . ,
′
q̂n (Un−1
, Un ), hn (Un , Un′ )
is logically implied by C , the standard equality theory for C , the “iff” version of
the function substitutivity axiom for the list-constructor function symbol:
[X |Y ] = [X ′ |Y ′ ] ↔ X = X ′ & Y = Y ′
and the completed definitions of:
p̂(hSt|X i, hSt|Y i) ← p(X , Y )
q̂j (hSt|X i, hSt|Y i) ← qj (X , Y )
j = 1, . . . , n
hj (hΣj |tj i, hΣj +1 |tj′ i)
j = 0, . . . , n
←
where Σj is any list of the form [X1,j , . . . , Xdj ,j |St], such that {X1,j , . . . , Xdj ,j } = Πj ,
if Πj 6= ∅, and Σj is St if Πj = ∅, for j = 0, . . . , n + 1, and the hj ’s are predicate
symbols not occurring in C .
Proof
First, we apply equation introduction to the definitions of p̂ and the q̂i ’s.
p̂(hSt|X i, hSt′ |Y i) ← St = St′ , p(X , Y )
′
(18)
′
q̂j (hSt|X i, hSt |Y i) ← St = St , qj (X , Y )
j = 1, . . . , n
(19)
Let σi , i = 0, . . . , n, be renaming substitutions (Lloyd 1987, p. 22) for C such
that:
[
··
σi = {X /Z : X ∈ var(C ) & Z 6∈ var(C ) ∪
rhs(σj ) }
j 6=i
where rhs(σj ) = {Z1 , . . . , Zk } if σj = {X1 /Z1 , . . . , Xk /Zk }.
Since the variables in rhs(σi ) do not occur in C or in any other σj (i 6= j ), the
application of σi to a variable X renames X uniquely. Hence, we can think of such
an application as the addition of the subscript i to X .
Next, we apply equation introduction to C and rename variables, obtaining:
′
p(t0 σ0 , tn′ σn ) ← E , q1 (t0′ σ0 , t1 σ1 ), q2 (t1′ σ1 , t2 σ2 ), . . . , qn (tn−1
σn−1 , tn σn )
where
E = {X σi = X σk : X ∈ var(ti ) ∩ var(tk′ ) & i < k }
Subsequently, we resolve with (18) and rename St and St′ .
p̂(hSt0 |t0 σ0 i, hStn |tn′ σn i) ← St0 = Stn , E ,
′
q1 (t0′ σ0 , t1 σ1 ), q2 (t1′ σ1 , t2 σ2 ), . . . , qn (tn−1
σn−1 , tn σn )
·· S
Let Π = j Πj and X ∈ Π (i.e. X occurs in some (and only one) ti and some
tk′ , for i < k ). We now define
φ(X ) = i, where X ∈ var(ti )
γ(X ) = max{k : X ∈ var(tk′ )}
14
David A. Rosenblueth
and apply transitivity and factoring, replacing the equations by (up to variable
renaming):
[
{St0 = St1 , St1 = St2 , . . . , Stn−1 = Stn } ∪
EX
X ∈Π
where
EX = {X σφ(X ) = X σφ(X )+1 , X σφ(X )+1 = X σφ(X )+2 , . . . , X σγ(X )−1 = X σγ(X ) }
Note that for all j and all X :
X ∈ var(ti ) ∩ var(tk′ ) for some i ≤ j − 1 and some k ≥ j
iff
(X σj −1 = X σj ) ∈ EX
Equivalently, for all j and all X :
[
var(ti ) ∩ var(tk′ ) iff (X σj −1 = X σj ) ∈ EX
X ∈
i≤(j −1), k ≥j
Hence,
Πj = {X : (X σj −1 = X σj ) ∈ EX }
for j = 0, . . . , n + 1, so that we have exactly all equations for constructing the Σj ’s.
Next, we repetitively fold using function substitutivity, and arrive at:
p̂(hSt0 |t0 θ0 i, hStn |tn′ θn i) ← Σ1 θ0 = Σ1 θ1 , Σ2 θ1 = Σ2 θ2 , . . . , Σn θn−1 = Σn θn ,
′
q1 (t0′ θ0 , t1 θ1 ), q2 (t1′ θ1 , t2 θ2 ), . . . , qn (tn−1
θn−1 , tn θn )
Prechain form is now obtained by folding with (19):
p̂(hSt0 |t0 θ0 i, hStn |tn′ θn i) ← q̂1 (hΣ1 θ0 |t0′ θ0 i, hΣ1 θ1 |t1 θ1 i),
q̂2 (hΣ2 θ1 |t1′ θ1 i, hΣ2 θ2 |t2 θ2 i), . . . ,
′
q̂n (hΣn θn−1 |tn−1
θn−1 i, hΣn θn |tn θn i)
Finally, we apply predicate substitutivity and symmetry, and then fold with the
completed definitions of the hi predicates. The resulting clause has the desired
form; hence we conclude that the theorem holds.
Example 2
Let us apply Theorem 1 to the clause (13), of Example 1.
s (h[A|N ]i, h[A|L], M i) ← s (hN i , hL, M i)
|{z} | {z } | {z }
|{z} |{z} | {z }
p
q1 t ′
t0
t1
t′
1
Π0
Π1
Π2
0
=
=
=
∅ ∩ var(t0′ ) ∪ var(t1′ )
′
var(t0 ) ∩ var(t
1)
var(t0 ) ∪ var(t1 ) ∩ ∅
=
=
=
∅
{A}
∅
So, Σ0 = St, Σ1 = [A|St], and Σ2 = St. The resulting clause in chain form is (17).
The definitions of the hi ’s are as before, up to variable renaming.
End of example
Chain Programs for Writing Deterministic Metainterpreters
15
Let P be a moded program. We define P̂ as the program resulting from applying
Theorem 1 to every clause in P in such a way that the hi predicate symbols of a
clause do not occur in any other clause of P̂.
Theorem 1 associates a chain program P̂ with a moded program P in such a way
that P̂ is a logical consequence of a conservative extension of P . The implication
in the other direction also holds. That P is logically implied by a conservative
extension of P̂ can be seen by first resolving each clause in P̂ having a head with
predicate symbol p̂ with the only-if part of the definition of p̂ and unfolding the
definitions of the hi ’s and the q̂i ’s.
Finally, note that the chain program P̂ of a moded program P is a G-chain
program (i.e. a program such that in every unit clause p(t , t ′ ) ←, var(t ′ ) ⊆ var(t ),
so that all answers for a subgoal with a ground input are ground using a leftmost
computation rule (cf. Sect. 2)).
Example 3
As an example explicitly linking this transformation with the abcde metainterpreter, we give the object-program representation of the chain form of the clauses (12)
and (13).
defn(ŝ, [ŝ1 , ŝ2 ]) ←
defn(h0 , [h0′ ]) ←
defn(h1 , [h1′ ]) ←
nonunit(ŝ2 , [h0 , ŝ, h1 ]) ←
unit(ŝ1 , hSt, Li, hSt, [ ], Li) ←
unit(h0′ , hSt, [A|N ]i, h[A|St], N i) ←
unit(h1′ , h[A|St], L, M i, hSt, [A|L], M i) ←
is unit(ŝ1 ) ←
is unit(h0′ ) ←
is unit(h1′ ) ←
End of example
4 Conversion of Definite Programs into Chain Form
In this section, we will first give a transformation inspired by the previous one,
converting an arbitrary definite program into chain form. Next, we will explain
how to couple the abcde metainterpreter to this “unmoded” transformation.
4.1 Transformation
Roughly, the moded transformation takes a clause with predicates having one input
argument place and one output argument place, disposes of the pass-on variables
(the Πj ’s) by adding a stack, and replaces both arguments of each subgoal by
variables.
Hence, the first apparent obstacle we find in trying to convert an arbitrary, unmoded clause into chain form, is that arguments of predicates in such a clause do
16
David A. Rosenblueth
not have predetermined input/output roles. The fact that any argument of a predicate may play the role of either input or output suggests treating all arguments
uniformly. One way of doing so and yet have binary predicates could be to replicate
the arguments of each predicate so as to have two copies of each set of arguments:
one copy behaving as a single input (possibly having variables when the subgoal is
selected), and the other copy behaving as a single output. Naturally, we now have
to give up the groundness property of runtime terms, which translates to having to
use full unification instead of matching.
Thus, we can associate with each predicate p(X1 , . . . , Xn ), another predicate,
defined as: p̂(hX1 , . . . , Xn i, hX1 , . . . , Xn i) ← p(X1 , . . . , Xn ), which denotes a subset
of the identity relation of the Herbrand universe. But we also have to add the stack,
so that we have: p̂(hSt, X1 , . . . , Xn i, hSt, X1 , . . . , Xn i) ← p(X1 , . . . , Xn ).
Another decision we have to make is which variables to push onto the stack. In
fact, we could push all variables of the clause, but we will be more economical by
pushing only the variables not occurring in all atoms of the clause.
Example 4
Consider the usual append program:
a([ ], L, L) ←
(20)
a([A|L], M , [A|N ]) ← a(L, M , N )
(21)
First we write the if-part of the definition of â:
â(hSt, X , Y , Z i, hSt, X , Y , Z i) ← a(X , Y , Z )
(22)
We now apply equation introduction to (22), getting:
â(hSt, X , Y , Z i, hSt′ , X ′ , Y ′ , Z ′ i) ←
St = St′ , X = X ′ , Y = Y ′ , Z = Z ′ ,
a(X ′ , Y ′ , Z ′ )
(23)
which we will need for a later folding application.
Next, we obtain an instance of (22), to which we apply equation introduction:
â(hSt, [A|L], M , [A|N ]i, hSt′ , [A′ |L′ ], M ′ , [A′ |N ′ ]i) ←
St = St′ , A = A′ , L = L′ , M = M ′ , N = N ′ ,
a([A′ |L′ ], M ′ , [A′ |N ′ ])
Now we resolve (21) with (24) unifying the two underlined atoms, and get:
â(hSt, [A|L], M , [A|N ]i, hSt′ , [A′ |L′ ], M ′ , [A′ |N ′ ]i) ←
St = St′ , A = A′ , L = L′ , M = M ′ , N = N ′ ,
a(L′ , M ′ , N ′ )
Subsequently, we fold using function substitutivity.
â(hSt, [A|L], M , [A|N ]i, hSt′ , [A′ |L′ ], M ′ , [A′ |N ′ ]i) ←
[A|St] = [A′ |St′ ] , L = L′ , M = M ′ , N = N ′ ,
a(L′ , M ′ , N ′ )
(24)
Chain Programs for Writing Deterministic Metainterpreters
17
Finally, we fold using (23) and systematically rename variables.
â(hSt0 , [A0 |L0 ], M0 , [A0 |N0 ]i, hSt1 , [A1 |L1 ], M1 , [A1 |N1 ]i) ←
â(h[A0 |St0 ], L0 , M0 , N0 i, h[A1 |St1 ], L1 , M1 , N1 i)
We can now apply predicate substitutivity and symmetry, and then fold w.r.t.
the completed definitions of the predicates manipulating the stack, as in the moded
transformation, thus arriving at chain form.
End of example
An arbitrary definite program can be converted into chain form with the following
theorem.
Theorem 2
Let C be a definite clause:
p(t̃0 ) ← q1 (t̃1 ), q2 (t̃2 ), . . . , qn (t̃n )
n ≥0
and let:
Π = var(t̃0 ) ∪ · · · ∪ var(t̃n ) \ var(t̃0 ) ∩ · · · ∩ var(t̃n )
Then the clause C ′ :
p̂(U0 , Un′ ) ← h0 (U0 , U0′ ), q̂1 (U0′ , U1 ), h1 (U1 , U1′ ), q̂2 (U1′ , U2 ), . . . ,
′
q̂n (Un−1
, Un ), hn (Un , Un′ )
is logically implied by C , the standard equality theory for C , the “iff” version of
the function substitutivity axiom for the list-constructor function symbol:
[X |Y ] = [X ′ |Y ′ ] ↔ X = X ′ & Y = Y ′
and the completed definitions of:
p̂(hSt, X1,0 , . . . , Xr0 ,0 i, hSt, X1,0 , . . . , Xr0 ,0 i) ← p(X1,0 , . . . , Xr0 ,0 )
q̂i (hSt, X1,i , . . . , Xri ,i i, hSt, X1,i , . . . , Xri ,i i) ← qi (X1,i , . . . , Xri ,i )
h0 (hSt|t̃0 i, hΣ|t̃1 i) ←
hi (hΣ|t̃i i, hΣ|t̃i+1 i) ←
hn (hΣ|t̃n i, hSt|t̃0 i) ←
i = 1, . . . , n
(i = 0)
i = 1, . . . , n − 1
(i = n)
where Σ is any list of the form [X1 , . . . , Xd |St], such that {X1 , . . . , Xd } = Π, if
Π 6= ∅, and Σ is St if Π = ∅.
Proof
First, we use equation introduction on the definitions of the q̂i ’s.
′
q̂i (hSt, X1,i , . . . , Xri ,i i, hSt′ , X1,i
, . . . , Xr′i ,i i) ← St = St′ ,
′
X1,i = X1,i
, . . . , Xri ,i = Xr′i ,i ,
′
qi (X1,i
, . . . , Xr′i ,i )
i = 1, . . . , n
(25)
18
David A. Rosenblueth
Next, we apply equation introduction to C in such a way that no two t̃i ’s have
variables in common, except for t̃0 and t̃n . We obtain:
p(t̃0 σn ) ← E , q1 (t̃1 σ1 ), q2 (t̃2 σ2 ), . . . , qn (t̃n σn )
(26)
where the σj ’s are as in Theorem 1 and
E = {X σj = X σk : X ∈ var(t̃j ) ∩ var(t̃k ) & 0 < j < k < n}
∪ {X σj = X σn : X ∈ var(t̃j ) ∩ var(t̃0 ) ∪ var(t̃n ) & 0 < j < n}
Now, we apply θ to the if-part of the definition of p̂, where θ is the mgu of
p(X1,0 , . . . , Xr0 ,0 ) and p(t̃0 ) and use equation introduction in the resulting instance,
getting (up to variable renaming):
p̂(hSt0 |t̃0 σ0 i, hStn |t̃0 σn i) ← St0 = Stn , F , p(t̃0 σn )
(27)
where
F = {X σ0 = X σn : X ∈ var(t̃0 )}
Subsequently, we resolve (26) with (27):
p̂(hSt0 |t̃0 σ0 i, hStn |t̃0 σn i) ← St0 = Stn , F , E ,
q1 (t̃1 σ1 ), q2 (t̃2 σ2 ), . . . , qn (t̃n σn )
Before folding, we add the following equations, recalling that subgoal addition
preserves soundness:
{X σ0 = X σn : X ∈ Π \ var(t̃0 ) }
(Such equations, after applying transitivity, will enable us to have the same Σ in
each subgoal of the resulting clause in prechain form.)
Now we add the equations:
{X σi−1 = X σi : X ∈ var(t̃i )}
0<i ≤n
(Such equations will enable us to fold w.r.t. (25).)
We now apply transitivity and factoring in such a way that the equations are
replaced by (up to variable renaming):
[
[
{St0 = St1 , St1 = St2 , . . . , Stn−1 = Stn } ∪
Gi ∪
Hi
0<i≤n
0<i≤n
where
Gi = {X σi−1 = X σi : X ∈ Π}
Hi = {X σi−1 = X σi : X ∈ var(t̃i )}
Next, we repetitively fold using function substitutivity and arrive at:
p̂(hSt0 |t̃0 σ0 i, hStn |t̃0 σn i) ← Σσ0 = Σσ1 , Σσ1 = Σσ2 , . . . , Σσn−1 = Σσn ,
t̃1 σ0 ∼
= t̃n σn ,
= t̃2 σ2 , . . . , t̃n σn−1 ∼
= t̃1 σ1 , t̃2 σ1 ∼
q1 (t̃1 σ1 ),
q2 (t̃2 σ2 ),
. . . , qn (t̃n σn )
19
Chain Programs for Writing Deterministic Metainterpreters
where
r̃ ∼
= s̃ = {r i = s i : r i is the i th term of r̃ and s i is the i th term of s̃}
Prechain form is now obtained by folding with (25):
p̂(hSt0 |t̃0 σ0 i, hStn |t̃0 σn i) ← q̂1 (hΣσ0 |t̃1 σ0 i, hΣσ1 |t̃1 σ1 i),
q̂2 (hΣσ1 |t̃2 σ1 i, hΣσ2 |t̃2 σ2 i), . . . ,
q̂n (hΣσn−1 |t̃n σn−1 i, hΣσn |t̃n σn i)
Finally, we apply predicate substitutivity and symmetry, and then fold with the
completed definitions of the hi predicates. The resulting clause has the desired
form; hence we conclude that the theorem holds.
As in the moded transformation, that P is logically implied by a conservative
extension of P̂ can be seen by unfolding the definitions of the hi ’s and the q̂i ’s.
4.2 A deterministic metainterpreter for arbitrary chain programs
In Sect. 2 we wrote a deterministic metainterpreter assuming that the leftmost
input in every goal of the LD tree (Apt 1997) (i.e. an SLD tree with a leftmost
computation rule) was ground. We will now modify such a metainterpreter so that
it is also correct for LD trees that do not necessarily have this groundness property.
It is possible to handle terms with variables by generalising matching to full unification. A well-known metainterpreter explicitly using unification is Bowen and
Kowalski’s demo metainterpreter (Bowen and Kowalski 1982). In our case, unification for clauses of the form (1) reduces to argument passing so that we need only
incorporate it to the clauses of the form (2). Consider for instance our previous
definition of e ′ , which included the clause:
e ′ (X , Pj, [Y ]) ← unit(Pj, X , Y )
Following the demo metainterpreter, we would replace this clause by:
e ′ (X , Pj, [Y ]) ← unit(Pj, X ′ , Y ′ ),
rename(f (X ′ , Y ′ ), X , f (X ′′ , Y ′′ )),
match(X , X ′′ , Sub),
apply(Y ′′ , Sub, Y )
(28)
where
1. rename(Z , X , Z ′ ) holds when Z ′ is the result of renaming the variables in Z
so that they are distinct from the variables in X ,
2. match(X , X ′′ , Sub) holds when Sub is the mgu of X and X ′′ , and
3. apply(Y ′′ , Sub, Y ) holds when Y is the result of applying Sub to Y ′′ .
Prolog provides a way to approximate this effect through the extralogical copy
“predicate”:
e ′ (X , Pj, [Y ]) ← copy(X , Xθ), unit(Pj, Xθ, Y )
(29)
20
David A. Rosenblueth
Note that our addition of unification to the abcde metainterpreter occurs at a
single point, unlike the unifiers of the metainterpreter in Fig. 1, which are pervasive
(cf. the variables with a θ in their name).
5 Extending Existing Committed-Choice Traversal Methods to
Arbitrary Definite Programs
This section deals first with a reconstruction and then with an extension of the existing stream-based (Tamaki 1987) and continuation-based (Ueda 1987) deterministic,
exhaustive-traversal methods. An objective of these methods is that of executing
or-parallel programs in committed-choice processors (which are and parallel).
The existing versions of such methods are restricted to moded programs. Hence,
our reconstruction uses our transformation of moded programs into G-chain form.
The extension modifies such methods so as to make them applicable to arbitrary
definite programs essentially by replacing the moded transformation of Sect. 3 by
the definite transformation of Sect. 4.
We fall short of proposing practical methods because we do not eliminate the
layer of interpretation. One way of eliminating such a layer would be to feed the
metainterpreter and the object program to a general-purpose partial evaluator such
as Mixtus (Sahlin 1993). However, the resulting residual program may be enormous.
Another possibility would be to compile away the layer of interpretation “by hand,”
but we have not done so in the present work.
5.1 Reconstruction
The derivations of both methods start from the abcde metainterpreter. For brevity,
we will omit detailed derivations, and will only indicate how such derivations could
be obtained. Also, instead of using difference lists as the original methods do, we
will employ ordinary lists for clarity.
5.1.1 A chain-program reconstruction of the stream-based method
As observed by Tamaki, programs originally having some degree of and parallelism
may lose such a parallelism if we only capture their or parallelism. He thus treats
clauses with and parallelism in a special way. For simplicity we will not be concerned
with such a special treatment here, and will concentrate on the main component of
this method, that converts or parallelism into and parallelism.
We can obtain programs produced by the stream-based method if we unfold (7)
using (8):
a([X |Xs], [Q |Qs], YsZs) ← b ′ (X , Q , Ys), a(Ys, Qs, Ys′ ),
a(Xs, [Q |Qs], Zs), append(Ys′ , Zs, YsZs)
(30)
Let us consider first how a chain program is transformed by the stream-based
method. Perhaps the most interesting clause in the program produced by this
Chain Programs for Writing Deterministic Metainterpreters
21
method is a clause associated with each subgoal qi+1 (Xi , Xi+1 ), which is of the
form:
ki ([X |Xs], YsZs) ← all qi+1 (X , Ys), ki+1 (Ys, Ys′ ),
ki (Xs, Zs), append(Ys′ , Zs, YsZs)
(31)
where all qi+1 (x, ys) is intended to hold when ys is the set of answers to
← qi+1 (x, Y ). Observe first that all qi+1 (x, ys) has the same intended meaning
as b ′ (x, qi+1 , ys). Next, it is easy to obtain (31) from (30) by identifying the ki
predicate with the a predicate. The rest of the clauses resulting from transforming
a chain program are readily obtainable from the abcde metainterpreter.
Consider now an arbitrary moded program. The clause corresponding to (31) in
this case is:
ki (πi , [ti |Xs], YsZs) ← all qi+1 (ti′ , Ys), ki+1 (πi+1 , Ys, Ys′ ),
ki (πi , Xs, Zs), append(Ys′ , Zs, YsZs)
(32)
where πj is any term such that var(πj ) = Πj , and Πj is as defined in Theorem 1.
Recall now (30) and an object clause transformed by our moded transformation.
A subgoal with an hi predicate symbol, when partially evaluated with the abcde
metainterpreter, results in a subgoal of the form b ′ (X , hi , Ys), which can be easily
unfolded away producing the following clause, similar to (32):
a([hΣi |ti i|Xs], [Q |Qs], YsZs) ← b ′ (hΣi+1 |ti′ i, Q , Ys), a(Ys, Qs, Ys′ ),
a(Xs, [Q |Qs], Zs), append(Ys′ , Zs, YsZs) (33)
A difference between (32) and (33) is that the stream-based method uses a parameter πi in the ki predicates for recording the values of the Πi variables, whereas (33)
keeps the Πi variables as part of each term hΣi |ti i. Even with this difference, the
stream-based compiled program and the abcde metainterpreter follow the same
search strategy. By using a separate parameter πi , however, the stream-based
method is more economical because of exploiting the fact that only the hi predicates
may modify the stack: all other predicates hold for relations with an output stack
equal to the input stack. (To see this, observe the stack in the definitions of p̂, q̂i in
Theorem 1.) Hence, we can compute all answers to a goal ← b ′ (hΣi+1 |ti′ i, Q , Ys),
by first computing all answers to ← b ′ (ti′ , Q , Ws) and then affixing Σi+1 in front of
each such answer, if Q is not an hi predicate:
hSt|t i ; Q = St ⋄ (t ; Q )
(34)
where X ⋄ Xs is the list obtained by concatenating all lists having X affixed in front
of every list in Xs:
X ⋄ [] = []
X ⋄ [Y |Ys] = [[X |Y ]|X ⋄ Ys]
Let us first rewrite the stream-based clause (32) as:
astream (Σi , [ti |Xs], [Q |Qs], YsZs) ← b ′ (ti′ , Q , Ys), astream (Σi+1 , Ys, Qs, Ys′ ),
astream (Σi , Xs, [Q |Qs], Zs),
append(Ys′ , Zs, YsZs)
(35)
22
David A. Rosenblueth
where the predicate astream (St , Xs, Qs, YsZs) is intended to hold iff
(St ⋄ Xs) ; Qs = St ⋄ YsZs
Hence, (35) asserts that:
Σi ⋄ [ti |Xs] ; (Q ; Qs) ⊇ Σi+1 ⋄ (ti′ ; Q ) ; Qs ++ Σi ⋄ Xs ; (Q ; Qs)
(36)
Here, and throughout this section, ⋄ binds stronger than “;”, and ++ denotes list
concatenation.
Let us now consider the abcde metainterpreter. For clarity, we rename variables
in (33):
a([hΣi |ti i|StXs], [Q |Qs], StYsZs) ← b ′ (hΣi+1 |ti′ i, Q , StYs), a(StYs, Qs, StYs′ ),
a(StXs, [Q |Qs], StZs),
append(StYs′ , StZs, StYsZs)
(37)
where a(StXs, Qs, StYsZs) holds when StXs ; Qs = StYsZs and the same stack Σi
occurs in front of every element of StXs and StYsZs. Hence, (37) asserts that:
[hΣi |ti i|Σi ⋄ Xs] ; (Q ; Qs) ⊇ (hΣi+1 |ti′ i ; Q ) ; Qs ++ Σi ⋄ Xs ; (Q ; Qs) (38)
By the definition of ⋄ and (34), we obtain (36) from (38). (Specifically, by the
recursive equation in the definition of ⋄ we obtain the left-hand side and by (34)
we obtain the right-hand side.)
5.1.2 A chain-program reconstruction of the continuation-based method
Let us now turn our attention to the continuation-based method. Starting also from
the abcde metainterpreter, we use the completed definitions of:
c ′ a(X , Pjs, Qs, M ) ← c ′ (X , Pjs, L), a(L, Qs, M )
′
′
d a(X , Pj, Qs, M ) ← d (X , Pj, L), a(L, Qs, M )
(39)
(40)
where Qs acts like a list of “continuations,” and the underlined subgoals indicate a
forthcoming unfolding application.
First we unfold (39) using (9):
c ′ a(X , [Pj|Pjs], Qs, M ) ← d ′ (X , Pj, L1 ), c ′ (X , Pjs, L2 ),
append(L1 , L2 , L), a(L, Qs, M )
(41)
Using now the identity
(X ; P1 ) ∪ . . . ∪ (X ; Pm ) ; Qs = (X ; P1 ; Qs) ∪ . . . ∪ (X ; Pm ; Qs)
which follows from the right distributivity of composition over union, we rewrite (41)
as:
c ′ a(X , [Pj|Pjs], Qs, M ) ← d ′ (X , Pj, L1 ), a(L1 , Qs, M1 ) ,
c ′ (X , Pjs, L2 ), a(L2 , Qs, M2 ) ,
append(M1 , M2 , M )
(42)
23
Chain Programs for Writing Deterministic Metainterpreters
where each rectangle indicates a forthcoming folding application. We can now
fold (42) using the definitions of the d ′ a and c ′ a predicates:
c ′ a(X , [Pj|Pjs], Qs, M ) ← d ′ a(X , Pj, Qs, M1 ),
c ′ a(X , Pjs, Qs, M2 ),
append(M1 , M2 , M )
(43)
arriving at a clause of the continuation-based metainterpreter.
Another interesting clause is obtained by unfolding (40) using (10):
d ′ a(X , Pj, Qs′ , M ) ← nonunit(Pj, Qs), a ′ (X , Qs, L), a(L, Qs′ , M )
which we rewrite as:
d ′ a(X , Pj, Qs′ , M ) ← nonunit(Pj, Qs),
append(Qs, Qs′ , QsQs′ ), a ′ (X , QsQs′ , M )
This step can be justified using the associativity of composition:
(X ; Qs) ; Qs′ = X ; (Qs ; Qs′ )
Figure 2 shows the resulting metainterpreter, where we have applied an unfolding
step using the definition of e ′ .
a ′ (X , [ ], [X ]) ←
a ′ (X , [Q|Qs], Zs) ← defn(Q, Pjs), c ′ a(X , Pjs, Qs, Zs)
c ′ a(X , [ ], Qs, [ ]) ←
c ′ a(X , [Pj|Pjs], Qs, YsZs) ← d ′ a(X , Pj, Qs, Ys), c ′ a(X , Pjs, Qs, Zs),
append(Ys, Zs, YsZs)
d ′ a(X , Pj, Qs, Zs) ← is unit(Pj), unit(Pj, X , Y ), a ′ (Y , Qs, Zs)
d ′ a(X , Pj, Qs, [ ]) ← is unit(Pj), not(unit(Pj, X , Y ))
d ′ a(X , Pj, Qs′ , Zs) ← nonunit(Pj, Qs), append(Qs, Qs′ , QsQs′ ), a ′ (X , QsQs′ , Zs)
Figure 2. A continuation-based, deterministic, exhaustive-traversal metainterpreter.
5.2 Unmoded versions of the stream-based and the continuation-based
methods
Having reconstructed the stream- and continuation-based methods through chain
programs, we can now replace the moded transformation by the definite transformation. However, as in Sect. 4, we must also rename variables when using the unit
predicate with either (28) or (29).
24
David A. Rosenblueth
6 Other Applications
6.1 Prolog as the continuation-based metainterpreter together with the
definite transformation
So far we have designed metainterpreters for performing traversals in committedchoice processors. Observe that just as committed-choice processors have deterministic bindings, so do standard (deterministic) imperative languages. This suggests
the possibility of using our metainterpreters for describing search strategies in one
such imperative language. In particular, we will see how to obtain an imperative
implementation of Prolog’s search strategy by slightly modifying the continuationbased metainterpreter of Sect. 5 (Fig. 2).
A difference between our previous metainterpreters and the standard implementations of Prolog is that whereas we perform exhaustive traversals, Prolog systems may or may not do so. However, we can easily modify the continuation-based
metainterpreter so as to ask the user whether or not more answers are requested.
Also, note that Prolog systems do not usually remember the answers to a query,
allowing us to eliminate the answer list.
a ′ (X , [ ], halt) ← write(X ), write(’ more? ’), read(n)
a ′ (X , [ ], cont) ←
a ′ (X , [Q |Qs], HaltCont) ← defn(Q , Pjs), c ′ a(X , Pjs, Qs, HaltCont)
c ′ a(X , [ ], Qs, cont) ←
c ′ a(X , [Pj|Pjs], Qs, HaltCont′ ) ← d ′ a(X , Pj, Qs, HaltCont),
halt cont(X , Pjs, Qs, HaltCont, HaltCont′ )
d ′ a(X , Pj, Qs, HaltCont) ← is unit(Pj), unit(Pj, X , Y ), a ′ (Y , Qs, HaltCont)
d ′ a(X , Pj, Qs, cont) ← is unit(Pj), not(unit(Pj, X , Y ))
d ′ a(X , Pj, Qs′ , HaltCont) ← nonunit(Pj, Qs),
append(Qs, Qs′ , QsQs′ ), a ′ (X , QsQs′ , HaltCont)
halt cont(X , Pjs, Qs, halt, halt) ←
halt cont(X , Pjs, Qs, cont, HaltCont) ← c ′ a(X , Pjs, Qs, HaltCont)
This metainterpreter is meant for constructing ground terms. To obtain a true
(pure) Prolog system, handling terms with variables, we would have to include
variable renaming and unification in a manner similar to that of either (28) or (29).
Deterministic metainterpreters for arbitrary definite programs can be written
directly, without using our transformations, into chain form. (An example is the
metainterpreter in Fig. 1.) The programmer, however, has to be aware of the set
of unifiers of the children of a node in the derivation tree. By contrast, in our
approach the programmer can write a metainterpreter without considering such
Chain Programs for Writing Deterministic Metainterpreters
25
unifiers, except when using the unit predicate, in which case the object clause has
no body, simplifying the treatment of unifiers.
6.2 A bounded-resource metainterpreter
As a final application, we exhibit a bounded-resource metainterpreter. As usual, we
will give a metainterpreter for chain programs. Our transformations of moded and
definite programs into chain form make this metainterpreter applicable to programs
that do not necessarily have chain form.
The next variant of the abcde metainterpreter constructs at most one proof, and
has an extra argument to indicate the amount of resources needed to construct such
a proof.
a ′ (X , [ ], ans(X ), 0) ←
a ′ (X , [Q |Qs], Z , R + 1) ← defn(Q , Pjs), c ′ a(X , Pjs, Qs, Z , R)
c ′ a(X , [ ], Qs, no ans, 0) ←
c ′ a(X , [Pj|Pjs], Qs, Z , R + S ) ← d ′ a(X , Pj, Qs, Y , R),
halt cont(X , Pjs, Qs, Y , Z , S )
d ′ a(X , Pj, Qs, Z , R) ← is unit(Pj), unit(Pj, X , Y ), a ′ (Y , Qs, Z , R)
d ′ a(X , Pj, Qs, no ans, 0) ← is unit(Pj), not(unit(Pj, X , Y ))
d ′ a(X , Pj, Qs′ , Z , R) ← nonunit(Pj, Qs),
append(Qs, Qs′ , QsQs′ ), a ′ (X , QsQs′ , Z , R)
halt cont(X , Pjs, Qs, ans(Y ), ans(Y ), 0) ←
halt cont(X , Pjs, Qs, no ans, Z , S ) ← c ′ a(X , Pjs, Qs, Z , S )
This style of writing bounded-resource metainterpreters may be viewed as an alternative to that appearing in (Kowalski 1995, Kowalski and Sadri 1996).
7 Concluding Remarks
7.1 Contributions
Some applications of metainterpreters have been neglected, perhaps because of
being based on convoluted definitions. By comparison with the demo predicate,
deterministic metainterpreters, for example, result especially elaborate, since the
programmer must consider the set of unifiers of the children of a node in the derivation tree. Thus, metainterpreters for (1) converting or parallelism into and parallelism (Ueda 1987, Tamaki 1987), (2) describing search-strategies in logic-based,
26
David A. Rosenblueth
state-oriented languages (Clark and van Emden 1981), and (3) simulating boundedresource reasoning (Kowalski 1995, Kowalski and Sadri 1996), have not received due
attention.
Compilation methods converting or parallelism into and parallelism have been
developed first by Ueda (Ueda 1987) and then by Tamaki (Tamaki 1987). By studying these methods, we have identified chain programs as important for exhibiting
the essence of such techniques. If we use chain programs as a stepping stone, then
the methods of (Ueda 1987, Tamaki 1987) can be viewed as comprising two parts:
a. conversion of a moded program into chain form, and
b. application of partial deduction to a deterministic metainterpreter for chain
programs.
Our contribution to part (a) consisted first in having extracted from (Ueda 1987,
Tamaki 1987) the implicit transformation that converts a moded program into an
equivalent chain form. Next, by using a generalisation of this transformation, we
have given another, “unmoded” transformation, that converts arbitrary definite
programs into chain form.
To part (b) we contributed by showing how to write deterministic metainterpreters for chain programs. One such metainterpreter served us first to reconstruct
and then to extend to arbitrary (unmoded) definite programs the existing methods
of (Ueda 1987, Tamaki 1987).
Finally, we observed that deterministic metainterpreters have applications other
than exhaustive traversals. We gave a metainterpreter that follows Prolog’s search
strategy and another one that counts the number of steps in the search for a refutation (as opposed to the number of steps in the refutation).
Our methodology for designing deterministic-traversal methods is then as follows:
1. Write a deterministic metainterpreter for chain programs ignoring unification.
2. Incorporate to the metainterpreter one of the transformations converting either moded or unmoded programs into chain form.
3. In case the unmoded transformation was selected, add renaming and unification to the metaclauses dealing with the object unit clauses.
We observed that even after adding unification, we need only be concerned about
substitutions at a single point of the chain-program metainterpreter, whereas in a
metainterpreter written directly the unifiers are pervasive (cf. Fig. 1).
7.2 Performance study
We have made a study illustrating how the performance of some programs is degraded as a result of transforming such programs into chain form. For this study, we
used SICStus Prolog version 3.7.1, which we ran under RedHat Linux version 6.0.
In this table and the next, the columns labeled A show the data for the source program and the columns labeled B show the data for the corresponding transformed
program.
First we exhibit the number of clauses and the program size (measured in bytes
for compiled code).
Chain Programs for Writing Deterministic Metainterpreters
program
num. clauses
A
B
split (moded append)
append (for splitting)
quicksort, ord. lists
quicksort, diff. lists
2
2
7
5
4
4
22
17
27
program size
A
B
445
440
1,405
1,066
1,051
1,168
5,833
5,871
Now we give the relative execution times (according to SICStus’ profile_data/4)
and the memory requirements for the local and global stacks (according to SICStus’
statistics/0) for splitting a 100-element list into all its prefixes and suffixes and
sorting the reverse of a sorted 100-element list with quicksort programs.
program
split (moded append)
append (for splitting)
quicksort, ord. lists
quicksort, diff. lists
execution time
A
B
global+local stacks
A
B
26
26
2,360
2,007
32,760
32,760
147,720
81,900
571
636
7,539
5,788
32,760
32,760
1,310,400
1,179,360
7.3 Related work
Our work stemmed from the continuation-based and the stream-based exhaustivetraversal methods. There are, however, various other publications studying deterministic traversals of search spaces within logic programming (Hirakawa, Chikayama
and Furukawa 1984, Bansal and Sterling 1987, Codish and Shapiro 1987, Lichtenstein, Codish and Shapiro 1987, Shapiro 1987, Sato and Tamaki 1989b, Mariën and
Demoen 1993). Of these contributions, (Mariën and Demoen 1993) has perhaps the
closest motivation to ours: The authors sketch a reconstruction of the continuationbased method and give a metainterpreter of their exhaustive-search method (based
on recomputation).
Similarly, there are a number of transformations converting logic programs into
a syntactically restricted form (Sato and Tamaki 1989a, Tarau and Boyer 1990, Tarau 1991). The one in (Sato and Tamaki 1989a) has in common with our work a
connection with the continuation-based exhaustive-traversal method. These transformations differ from ours, however, in producing programs in which every clause
is binary (i.e. has only one atom in the body).
7.4 Future work
We have argued that writing deterministic metainterpreters is advantageous with
our approach because this task amounts to that of describing an evaluation strategy
28
David A. Rosenblueth
for relational expressions of the form:
P = P1 ∪ P2 ∪ . . . ∪ Pm
where each Pi is defined as:
Pi = Q1 ; Q2 ; . . . ; Qn
Evaluation strategies for expressions have a close connection with the implementation of functional programming languages (Peyton Jones 1987) and rewrite
systems (Dershowitz and Jouannaud 1990). Investigating how different evaluation strategies for these relational expressions lead to different search strategies
for spaces determined by chain programs would be one way of extending our contributions.
During the presentation of our results, we came across the need for eliminating
the layer of interpretation. Our work would have a greater practical impact if it
were combined with an algorithmic elimination of such a layer, without producing
an excessively large residual program.
Chain programs have also proved to be useful in devising (Rosenblueth 1996,
Rosenblueth and Peralta 1998) inference systems derived from context-free parsers,
because we need only consider unification in the treatment of unit clauses and
hence we need only modify the treatment of terminals. Studying other applications
of chain programs where it might be helpful to relegate the role played by unifiers
would be another avenue of research.
Acknowledgments
This work owes much to Carlos Velarde. He contributed with motivating discussions,
he carefully read previous versions of this paper, and he spotted errors in the
theorem proofs. We also thank the referees, whose comments substantially improved
the presentation of these results. We gratefully acknowledge the facilities provided
by IIMAS, UNAM.
References
Abramson, H. and Rogers, M. (eds) (1988). Meta-Programming in Logic Programming,
MIT Press.
Apt, K. and Turini, F. (eds) (1995). Meta-Logics and Logic Programming, MIT Press.
Apt, K. R. (1997). From Logic Programming to Prolog, Prentice Hall.
Bansal, A. K. and Sterling, L. (1987). Compiling generate-and-test programs to committedchoice AND-parallelism, Technical Report CES-87-13, Case Western Reserve University,
Cleveland, Ohio, U.S.A.
Blikle, A. (1977). A comparative review of some program verification methods, in J. Gruska
(ed.), Mathematical Foundations of Computer Science 1977, Springer-Verlag, pp. 17–33.
Lecture Notes in Computer Science No. 53.
Bowen, K. A. and Kowalski, R. A. (1982). Amalgamating language and metalanguage in
logic programming, in K. Clark and S.-A. Tärnlund (eds), Logic Programming, Academic Press, pp. 153–172.
Chain Programs for Writing Deterministic Metainterpreters
29
Clark, K. and Gregory, S. (1985). Notes on the implementation of Parlog, The Journal of
Logic Programming 1: 17–42.
Clark, K. L. and van Emden, M. (1981). Consequence verification of flowcharts, IEEE
Transactions on Software Engineering SE-7(1): 52–60.
Codish, M. and Shapiro, E. (1987). Compiling or-parallelism into and-parallelism, in
E. Shapiro (ed.), Concurrent Prolog. Collected Papers, Vol. 2, MIT Press, pp. 351–
382.
de Bakker, J. and Meertens, L. (1975). On the completeness of the inductive assertion
method, Journal of Computer and System Sciences 11: 323–357.
Dershowitz, N. and Jouannaud, J.-P. (1990). Rewrite systems, in J. van Leeuwen (ed.),
Handbook of Theoretical Computer Science, Vol. B, Elsevier Science Publishers B.V.,
pp. 243–320.
Engelfriet, J. (1974). Simple Program Schemes and Formal Languages, Springer-Verlag.
Lecture Notes in Computer Science No. 20.
Gardner, P. and Shepherdson, J. (1992). Unfold/fold transformations of logic programs,
in J.-L. Lassez and G. Plotkin (eds), Computational Logic. Essays in Honor of Alan
Robinson, The MIT Press, pp. 565–583.
Hirakawa, H., Chikayama, T. and Furukawa, K. (1984). Eager and lazy enumerations in
Concurrent Prolog, in S.-Å. Tärnlund (ed.), Proc. Second International Logic Programming Conference, Uppsala, Sweden, pp. 89–100.
Jiang, Y. (1994). Ambivalent logic as the semantic basis of metalogic programming: I, in
P. V. Hentenryck (ed.), Proc. Eleventh International Conference on Logic Programming,
pp. 387–401.
Kowalski, R. A. (1990). Problems and promises of computational logic, in J. Lloyd (ed.),
Computational Logic. Symposium Proceedings, Springer-Verlag, pp. 1–36.
Kowalski, R. A. (1993). A springboard for information processing in the 21st century,
ICOT Journal 38: 17–41. Pannel Discussion at the Fifth Generation Computer Systems
Conference 1992.
Kowalski, R. A. (1995). Using meta-logic to reconcile reactive with rational agents, in
K. Apt and F. Turini (eds), Meta-Logics and Logic Programming, MIT Press, pp. 227–
242.
Kowalski, R. and Sadri, F. (1996). Towards a unified agent architecture that combines
rationality with reactivity, in D. Pedreschi and C. Zaniolo (eds), Proc. International
Workshop on Logic in Databases, Springer-Verlag, San Miniato, Italy, pp. 137–149.
Lecture Notes in Computer Science No. 1154.
Lichtenstein, Y., Codish, M. and Shapiro, E. (1987). Representation and enumeration of
Flat Concurrent Prolog computations, in E. Shapiro (ed.), Concurrent Prolog. Collected
Papers, Vol. 2, MIT Press, pp. 197–210.
Lloyd, J. (1987). Foundations of Logic Programming, 2nd edn, Springer-Verlag.
Mariën, A. and Demoen, B. (1993). Findall without findall/3, in D. S. Warren (ed.), Proc.
Tenth International Conference on Logic Programming, MIT Press, pp. 408–423.
Pettorossi, A. and Proietti, M. (1994). Transformation of logic programs: Foundations and
techniques, The Journal of Logic Programming 19, 20: 261–320.
Peyton Jones, S. L. (1987). The Implementation of Functional Languages, Prentice Hall.
Rosenblueth, D. A. (1996). Chart parsers as inference systems for fixed-mode logic programs, New Generation Computing 14(4): 429–458.
Rosenblueth, D. A. (1998). An exhaustive-search method using layered streams obtained
through a meta-interpreter for chain programs, Extended Abstracts of LOPSTR’98,
Eighth International Workshop on Logic-based Program Synthesis and Transforma-
30
David A. Rosenblueth
tion, 15–19 June 1998, Manchester, UK. Technical Report Series, Department of Computer Science, University of Manchester, ISSN 1361-6161. Report number UMCS-98-61. http://www.cs.man.ac.uk/cstechrep/titles98.html. A summary appeared in: Lecture
Notes in Computer Science No. 1559, pp 322–324, Springer-Verlag, 1999.
Rosenblueth, D. A. and Peralta, J. C. (1998). SLR inference: An inference system for
fixed-mode logic programs, based on SLR parsing, The Journal of Logic Programming
34(3): 227–260.
Safra, S. and Shapiro, E. (1986). Meta interpreters for real, in H. Kugler (ed.), Information
Processing 86 IFIP, Elsevier Science Publishers B.V., North-Holland, pp. 532–557.
Sahlin, D. (1993). Mixtus: An automatic partial evaluator for full Prolog, New Generation
Computing 12: 7–51.
Sato, T. and Tamaki, H. (1989a). Existential continuation, New Generation Computing
6(4): 421–438.
Sato, T. and Tamaki, H. (1989b). First order compiler: A deterministic logic program
synthesis algorithm, Symbolic Computation 8: 605–627.
Sergot, M. (1982). A query-the-user facility for logic programming, in P. Degano and
E. Sandewall (eds), Proc. European Conference on Integrated Interactive Computing
Systems, North Holland, pp. 27–44.
Shapiro, E. (1987). Or-parallel Prolog in Flat Concurrent Prolog, in E. Shapiro (ed.),
Concurrent Prolog. Collected Papers, Vol. 2, MIT Press, pp. 415–441.
Shapiro, E. Y. (1982). Algorithmic Program Debugging, MIT Press.
Tamaki, H. (1987). Stream-based compilation of ground I/O Prolog into committed-choice
languages, Proc. Fourth International Conference on Logic Programming, Melbourne,
Australia, pp. 376–393.
Tamaki, H. and Sato, T. (1984). Unfold/fold transformation of logic programs, Proc. Second
International Logic Programming Conference, pp. 127–138.
Tarau, P. (1991). Program transformations and WAM-support for the compilation of definite metaprograms, in A. Voronkov (ed.), Proc. Russian Conference on Logic Programming, Springer-Verlag, pp. 462–473. Lecture Notes in Artificial Intelligence No. 592.
Tarau, P. and Boyer, M. (1990). Elementary logic programs, in P. Deransart and
J. Maluszyński (eds), Proc. Programming Language Implementation and Logic Programming, Springer-Verlag, Linköping, Sweden, pp. 159–173. Lecture Notes in Computer Science No. 456.
Ueda, K. (1987). Making exhaustive search programs deterministic, New Generation Computing 5: 29–44.
| 6cs.PL
|
MAP moving horizon state estimation with binary measurements
arXiv:1804.02167v1 [cs.SY] 6 Apr 2018
Giorgio Battistelli, Luigi Chisci, Nicola Forti, Stefano Gherardini
Abstract— The paper addresses state estimation for discretetime systems with binary (threshold) measurements by following a Maximum A posteriori Probability (MAP) approach
and exploiting a Moving Horizon (MH) approximation of the
MAP cost-function. It is shown that, for a linear system
and noise distributions with log-concave probability density
function, the proposed MH-MAP state estimator involves the
solution, at each sampling interval, of a convex optimization
problem. Application of the MH-MAP estimator to dynamic
estimation of a diffusion field given pointwise-in-time-and-space
binary measurements of the field is also illustrated and, finally,
simulation results relative to this application are shown to
demonstrate the effectiveness of the proposed approach.
I. I NTRODUCTION
Binary sensors, whose output just indicates whether the
noisy measurement of the sensed variable (analog measurement) exceeds or not a given threshold, are frequently
employed in monitoring and control applications [1]-[14].
The idea is that by a multitude of low-cost and low-resolution
sensing devices it is possible to achieve the same estimation
accuracy that a few (possibly a single one) expensive highresolution sensors could provide, with significant practical benefits in terms of ease of sensor deployment and
minimization of communication requirements. The fact that
a binary (threshold) measurement just conveys a minimal
amount (i.e. a single bit) of information, while implying
communication bandwidth savings and consequently greater
energy efficiency, makes of paramount importance to fully
exploit the little available information by means of smart
estimation algorithms. In this respect, some work has recently addressed system identification [1]-[2], parameter [3][6] or state estimation [7]-[14] with binary measurements
by following either a deterministic [1]-[2], [7]-[9] or a
probabilistic [3]-[6], [10]-[14] approach.
In a deterministic context, the available information is
essentially concentrated at the sampling instants in which
some binary measurement signal has switched value [7],
[8]. As shown in [8], some additional information can be
exploited in the other (non switching) sampling instants by
penalizing values of the estimated quantity such that the
corresponding predicted measurement is on the opposite side,
with respect to a binary sensor reading, far away from the
threshold. Nevertheless, it is clear that there is no or very
little information available for estimation purposes whenever
no or very few binary sensor switchings occur. Hence, a
The authors are with Dipartimento di Ingegneria dell’Informazione
(DINFO), Università degli Studi di Firenze, via di Santa Marta
3, {giorgio.battistelli,luigi.chisci,nicola.forti,stefano.gherardini}@unifi.it.
S. Gherardini is also with Dipartimento di Fisica e Astronomia, Università
degli Studi di Firenze, via G. Sansone 1, and Istituto Nazionale di Fisica
Nucleare (INFN), Sezione di Firenze.
possible way to achieve high estimation accuracy is to have
many binary sensors measuring the same variable with different thresholds as this would clearly increase the number of
switchings, actually emulating, when the number of sensors
tends to infinity, the availability of a single continuous-valued
(analog) measurement.
Conversely, following a probabilistic approach, binary sensor readings could be exploited to infer information about the
probability distribution of the variable of interest. To clarify
this point, let us assume that a very large number of binary
sensors of the same type (i.e. measuring the same variable
with the same threshold) be available and the distribution
of their measurement noise (e.g. Gaussian with zero mean
and given standard deviation) be known. Then, thanks to
the numerosity of measurements, the relative frequency of
1 (or 0) values occurring in the sensor readings could be
considered as a reasonable estimate of the probability that
the sensed variable is above (or below) the threshold and
this, in turn, exploiting the knowledge of the measurement
noise distribution allows to extract information about the
location of the value of the sensed variable with respect to
the threshold. If, for example, it is found that the binary
measurement is equal to 1 for 70% of the sensors and
Gaussian measurement noise is hypothesized, it turns out that
the expected measurement of the sensed variable is above
the threshold of an amount equal to 0.525 times the standard
deviation of the measurement noise. Notice that if the sensors
are noiseless, they all provide either 0 or 1 output and,
paradoxically, in this case minimal information, i.e. that the
sensed variable takes values in a semi-infinite interval (either
below or above the threshold), is extracted from the set
of binary measurements. The above arguments suggest that,
adopting a probabilistic approach to estimation using binary
measurements, the presence of measurement noise can be a
helpful source of information. In other words, it can be said
that noise-aided procedures can be devised for estimation
with binary measurements by exploiting the fact that the
measurement noise randomly shifts the analog measurement
thus making possible to infer statistical information on the
sensed variable.
Relying on the above stated noise-aided paradigm, this
paper presents a novel approach to recursive estimation of
the state of a discrete-time dynamical system given binary
measurements. The proposed approach is based on a movinghorizon (MH) approximation of the Maximum A-posteriori
Probability (MAP) estimation and extends previous work
[3]-[4] concerning parameter estimation to recursive state
estimation. A further contribution of the paper is to show
that for a linear system the optimization problem arising from
the MH-MAP formulation turns out to be convex and, hence,
practically feasible for real-time implementation.
The rest of the paper is organized as follows. Section
II introduces the MAP problem formulation of state estimation with binary measurements. Section III presents a
MH approximation of MAP estimation, referred to as MHMAP algorithm, and analyzes the properties of the resulting
optimization problem. Section IV discusses a possible application of the proposed approach to the dynamic estimation
of a diffusion field from binary pointwise-in-space-and-time
field measurements. Section V presents simulation results
relative to the dynamic field estimation case-study. Finally,
section VI concludes the paper with perspectives for future
work.
II. M AXIMUM A POSTERIORI STATE ESTIMATION WITH
BINARY SENSORS
The following notation will be used throughout the paper:
col(·) denotes the matrix obtained by stacking its arguments
one on top of the other; diag(m1 , . . . , mq ) denotes the diagonal matrix with diagonal entries m1 , . . . , mq ; 0n , 1n indicate
the n−dimensional vectors, respectively, with all zero and
unit entries.
Let us consider the problem of recursively estimating the
state of the discrete-time nonlinear dynamical system
xt+1
zti
=
f (xt , ut ) + wt
= hi (xt ) + vti , i = 1, . . . , l
(1)
(2)
from a set of measurements provided by binary sensors
1, if zti ≥ τ i
yti = gi (zti ) =
(3)
0, if zti < τ i
where xt ∈ Rn is the state to be estimated, ut ∈ Rm is a
known input, and τ i is the threshold of the i−th binary
l
sensor. For the sake of simplicity, we define zt = col zti i=1 ∈
l
Rl and yt = col yti i=1 ∈ Rl . The vector wt ∈ Rn is an
additive disturbance affecting the system dynamics which
accounts for uncertainties in the mathematical model, while
l
vt = col vti i=1 ∈ Rl is the measurement noise vector.
Let N (µ, Σ) denote as usual the normal distribution with
mean µ and variance Σ. The statistical behaviour of the
system is characterized by
x0 ∼ N (x0 , P−1 ), wt ∼ N (0, Q−1 ), vt ∼ N (0, R)
(4)
where: R = diag(r1 , . . . , r p ); E[w j w0k ] = 0 and E[v j v0k ] = 0
if j 6= k; E[w j v0k ] = 0, E[w j x00 ] = 0, E[v j x00 ] = 0 for any
j, k. Notice from (2)-(3) that sensor i produces a binary
measurements yti ∈ {0, 1} depending on whether the noisy
system output zti is below or above the threshold τ i .
According to the available probabilistic description (4),
the problem of estimating the state of system (1) under the
binary measurement model (2)-(3) is formulated hereafter
in a Bayesian framework by resorting to a maximum a
posteriori probability (MAP) criterion. In the remainder of
this section, as a preliminary step, the full-information MAP
state estimation problem is formulated.
To this end, notice that each binary measurement yti
provides intrinsically relevant information on the state xt
which can be taken into account by means of the a posteriori
probabilities p(yti |xt ). In particular, the binary measurement
yti is a Bernoulli random variable such that, for any binary
sensor i and any time instant t, the a posteriori probability
p(yti |xt ) is given by
i
i
p(yti |xt ) = p(yti = 1|xt )yt p(yti = 0|xt )1−yt
(5)
p(yti = 1|xt ) = F i (τ i − hi (xt ))
(6)
where
and p(yti = 0|xt ) = 1 − p(yti = 1|xt ) , Φi (τ i − hi (xt )). The
function F i (τ i − hi (xt )) is the complementary cumulative
distribution function (CDF) of the random variable τ i −
hi (xt ). Since vti ∼ N (0, ri ), the conditional probability p(yti =
1|xt ) = F i (τ i − hi (xt )) can be written in terms of the Qfunction as follows
i
Z ∞
2
τ − hi (xt )
1
−ui
i i
i
2r
√
√
du = Q
.
e
F (τ −h (xt )) =
2πri τ i −hi (xt )
ri
(7)
Let us now denote by Yt = col(y0 , . . . , yt ) the vector of
all binary measurements collected up to time t and by Xt ,
col(x0 , . . . , xt ) the vector of the state trajectory. Further, let
us denote by X̂t|t , col(x̂0|t , . . . , x̂t|t ) the estimates of Xt to be
made at any stage t. Then, at each time instant t, given the
a posteriori probability p(XN |YN ), the estimate of the state
trajectory can be obtained by solving the following MAP
estimation problem:
X̂t|t = arg max p(Xt |Yt ) = arg min − ln p(Xt |Yt ).
(8)
Xt
Xt
From the Bayes rule
p(Xt |Yt ) ∝ p(Yt |Xt ) p(Xt ),
(9)
where
t−1
p(Xt ) = ∏ p(xt−k |xt−k−1 , . . . , x0 ) p(x0 )
k=0
t−1
(10)
= ∏ p(xt−k |xt−k−1 ) p(x0 ).
k=0
Notice that in the latter equation we have considered the
Markov property for the dynamical system state. As x0 and
wt are normally distributed vectors, we have
1
2
p(x0 ) ∝ e− 2 kx0 −x0 kP
(11)
− 12 kxk+1 − f (xk ,uk )k2Q
p(xk |xk−1 ) ∝ e
.
(12)
Moreover, the likelihood function p(Yt |Xt ) of the binary
measurement vector Yt can be written as
t
t
p(Yt |Xt ) = ∏ p(yk |xk ) = ∏
k=0
t
=∏
l
∏ p(yik |xk )
k=0 i=1
l
i
∏ F i (τ i − hi (xk ))yk
i
Φi (τ i − hi (xk ))1−yk
k=0 i=1
(13)
where in the latter equality we have exploited the statistical
independence of the binary sensors. Accordingly, the loglikelihood is
t
l
ln p(Yt |Xt ) = ∑
∑
i
yk ln F i (τ i − hi (xk ))
(14)
k=0 i=1
+(1 − yik ) ln Φi (τ i − hi (xk ))
,
and the cost function − ln p(Xt |Yt ) = − ln p(Yt |Xt ) − ln p(Xt )
to be minimized in the MAP estimation problem (8) turns
out to be, up to additive constant terms,
t
Jt (Xt ) = kx0 − x0 k2P + ∑ kxk+1 − f (xk , uk )k2Q +
k=0
t
−∑
l
∑ yik ln F i (τ i − hi (xk )) + (1 − yik ) ln Φi (τ i − hi (xk )) .
k=0 i=1
(15)
Unfortunately, a closed-form expression for the global
minimum of (15) does not exist and, hence, the optimal
MAP estimate X̂t|t has to be determined by resorting to some
numerical optimization routine. With this respect, the main
drawback is that the number of optimization variables grows
linearly with time, since the vector Xt has size (t + 1) n. As
a consequence, as t grows the solution of the full information MAP state estimation problem (8) becomes eventually
unfeasible, and some approximation has to be introduced.
III. M OVING - HORIZON APPROXIMATION
In this section, an approximate solution to the MAP state
estimation problem is proposed by resorting to the MHE approach [15]-[21]. Accordingly, by defining a sliding window
Wt = {t − N,t − N + 1, . . . ,t}, the goal is to find an estimate
of the partial state trajectory Xt−N:t , col(xt−N , . . . , xt ) by
using the information available in Wt . Then, in place of
the full information cost Jt (Xt ), at each time instant t
the minimization of the following moving-horizon cost is
addressed:
t
JtMH (Xt−N:t ) = Γt−N (xt−N ) +
∑
kxk+1 − f (xk , uk )k2Q +
k=t−N
t
−
∑
the present work, consists of assigning to the arrival cost a
fixed structure penalizing the distance of the state xt−N at
the beginning of the sliding window from some prediction
x̄t−N computed at the previous time instant, thus making
the estimation scheme recursive. A natural choice is then
a quadratic arrival cost of the form
l
∑ yik ln F i (τ i − hi (xk )) + (1 − yik ) ln Φi (τ i − hi (xk ))
k=t−N i=1
(16)
where the non-negative initial penalty function Γt−N (xt−N ),
known in the MHE literature as arrival cost [16], [19], is
introduced so as to summarize the past data y0 , . . . , yt−N−1
not explicitly accounted for in the objective function.
As a matter of fact, the form of the arrival cost plays an
important role in the behavior and performance of the overall
estimation scheme. While in principle Γt−N (xt−N ) could be
chosen so that minimization of (16) yields the same estimate
that would be obtained by minimizing (15), an algebraic
expression for such a true arrival cost seldom exists, even
when the sensors provide continuous (non-binary) measurements [16]. Hence, some approximation must be used. With
this respect, a common choice [18], [19], also followed in
Γt−N (xt−N ) = kxt−N − x̄t−N k2Ψ ,
(17)
which, from the Bayesian point of view, corresponds to
approximating the PDF of the state xt−N conditioned to all
the measurements collected up to time t − 1 with a Gaussian
having mean x̄t−N and covariance Ψ−1 . As for the choice of
the weight matrix Ψ, in the case of continuous measurements
it has been shown that stability of the estimation error dynamics can be ensured provided that Ψ is not too large (so as
to avoid an overconfidence on the available estimates) [18],
[19]. Recently [8], similar results have been proven to hold
also in the case of binary sensors in a deterministic context.
In practice, Ψ can be seen as a design parameter which has
to be tuned by pursuing a suitable tradeoff between such
stability considerations and the necessity of not neglecting
the already available information (since in the limit for Ψ
going to zero the approach becomes a finite memory one).
Summing up, at any stage t = N, N + 1, . . ., the following
problem has to be solved.
Problem Et : Given the prediction x̄t−N ,
sequence {ut−N , . . . , ut−1 }, the measurement
i
{yt−N
, . . . , yti , i = 1, . . . , l}, find the optimal
x̂t−N|t , . . . , x̂t|t that minimize the cost function
arrival cost (17).
the input
sequences
estimates
(16) with
Concerning the propagation of the estimation procedure
from Problem Et−1 to Problem Et , the prediction x̄t−N is
set equal to the value of the estimate of xt−N made at time
instant t − 1, i.e., x̄t−N = x̂t−N|t−1 . Clearly, the recursion is
initialized with the a priori expected value x̄0 of the initial
state vector.
In general, solving Problem Et entails the solution of
a non-trivial optimization problem. However, when both
equations (1) and (2) are linear, the resulting optimization
problem turns out to be convex so that standard optimization
routines can be used in order to find the global minimum.
To see this, let us consider the following assumption.
A1 The functions f (·) and hi (·), i = 1, . . . , l, are linear, i.e.,
f (xt , ut ) = Axt + But and hi (xt ) = Ci xt , i = 1, . . . , l, where
A, B, Ci are constant matrices of suitable dimensions.
Proposition 1: If assumption A1 holds, the CDF Φi (τ i −
and its complementary function F i (τ i −Ci xt ) are logconcave. Hence, the cost function (16) with arrival cost (17)
is convex.
Ci xt )
Remark 1: Under assumption A1, the convexity of the
cost function (16) is guaranteed also in the more general case
in which the statistical behaviour of the random variables
x0 , wt , vt is described by logarithmically concave distribution functions. Indeed, if a PDF is log-concave, also its
cumulative distribution function is log-concave; hence the
contribution related to the binary measurements in (16) turns
out to be convex.
In the next section we will focus on the case of a discretetime linear system, in particular considering a diffusion
process governed by a partial differential equation (PDE)
and spatially discretized by means of the finite element
method (FEM).
IV. DYNAMIC FIELD ESTIMATION
In this section, we consider the problem of reconstructing
a two-dimensional diffusion field, sampled with a network of
binary sensors arbitrarily deployed over the spatial domain
of interest Ω. The diffusion process is governed by the
following parabolic PDE:
∂c
− λ ∇2 c = 0 in Ω
∂t
(18)
which models various physical phenomena such as the spread
of a pollutant in a fluid. In this case, c(ξ , η,t) represents
the space-time dependent substance concentration, λ denotes
the constant diffusivity of the medium, and ∇2 = ∂ 2 /∂ ξ 2 +
∂ 2 /∂ η 2 is the Laplace operator, (ξ , η) ∈ Ω being the 2D
spatial variables. Furthermore, let us assume mixed boundary
conditions, i.e. a non-homogeneous Dirichlet condition
c=ψ
on ∂ ΩD ,
Accordingly, the unknown function c(ξ η,t) is approximated as
c(ξ , η,t) ≈
where υ is the outward pointing unit normal vector of ∂ ΩN .
The objective is to estimate the values of the dynamic
field of interest c(ξ , η,t), given the binary measurements
(3). The PDE system (18)-(20) is simulated with a mesh of
finite elements over Ω via the Finite Element (FE) approximation described in [22]-[23]. Specifically, the domain Ω
is subdivided into a suitable set of non overlapping regions,
or elements, and a suitable set of basis functions φ j (ξ , η),
j = 1, . . . , nφ is defined on such elements. The choices of
the basis functions φ j and of the elements are key points
of the FE method. In the specific case under investigation,
the elements are triangle in 2D and define a FE mesh with
vertices (ξ j , η j ) ∈ Ω, j = 1, . . . , nφ . Then each basis function
φ j is a piece-wise affine function which vanishes outside
the FEs around (ξ j , η j ) and such that φ j (ξ j , η j ) = δi j , δi j
denoting the Kronecker delta. In order to account for the
mixed boundary conditions, the basis functions are supposed
to be ordered so that the first n correspond to vertices of the
mesh which lie either in the interior of Ω or on ∂ ΩN , while
the last nφ − n correspond to the vertices lying on ∂ ΩD .
∑ φ j (ξ , η) c j (t) + ∑
φ j (ξ , η) ψ j (t)
(21)
j=n+1
j=1
where c j (t) is the unknown expansion coefficient of the function c(ξ η,t) relative to time t and basis function φ j (ξ , η),
and ψ j is the known expansion coefficient of the function
ψ(ξ η) relative to to the basis function φ j (ξ , η). Notice that
the second summation in (21) is needed so as to impose the
non-homogeneous Dirichlet condition (19) on the boundary
∂ ΩD .
The PDE (18) can be recast into the following integral
form:
Z
Z
∂c
ϕ dξ dη − λ
∇2 c ϕ dξ dη = 0
(22)
Ω
Ω ∂t
where ϕ(ξ , φ ) is a generic space-dependent weight function.
By applying Green’s identity, one obtains:
Z
Z
Z
∂c
∂c
ϕ dξ dη +λ ∇T c ∇ϕ dξ dη −λ
ϕ dξ dη = 0 .
Ω ∂t
Ω
∂Ω ∂ υ
(23)
By choosing the test function ϕ equal to the selected basis
functions and exploiting the approximation (21), the Galerkin
weighted residual method is applied and the following equation is obtained
n
n
Z
∑
i=1 Ω
φi φ j dξ dη ċi (t) + λ ∑
+λ
Z
∇T φi ∇φ j dξ dη ci (t)
i=1 Ω
nφ
(19)
which specifies a constant-in-time value of concentration on
the boundary ∂ ΩD , and a homogeneous Neumann condition
on ∂ ΩN = ∂ Ω \ ∂ ΩD , assumed impermeable to the contaminant, so that
∂ c/∂ υ = 0 on ∂ ΩN ,
(20)
nφ
n
∑
Z
∇T φi ∇φ j dξ dη ψi (t) = 0
(24)
i=n+1 Ω
for j = 1, . . . , n. Notice that in the latter equation the boundary integral of equation (23) is omitted since it is equal to
0 thanks to the homogeneous Neumann condition (20) on
∂ ΩN and to the fact that, by construction, the basis functions
φ j , j = 1, . . . , n vanish on ∂ ΩD . The interested reader is
referred to [24] for further details on the FEM theory, and
in particular on how to convert the case of inhomogeneous
boundary conditions to the homogeneous one.
By defining the state vector x = col(c1 , . . . , cn ) and the vector of boundary conditions γ = col(ψn+1 , . . . , ψnφ ), equation
(24) can be written in the more compact form
M ẋ(t) + Sx(t) + SD γ = 0
where S is the so-called stiffness matrix representing diffusion, M is the mass matrix, and SD captures the physical
interconnections among the vertices affected by boundary
condition (19) and the remaining nodes of the mesh.
By applying for example the implicit Euler method, the
latter equation can be discretized in time, thus obtaining the
the linear discrete-time model
xt+1 = A xt + B u + wt
where
−1
A = I + δt M −1 S
−1 −1
B = I + δt M −1 S
M δt
u = −SD γ
(25)
δt is the time integration interval, and wt is the process
disturbance taking into account also the space-time discretization errors.
Notice that the linear system (25) has dimension n equal
to the number of vertices of the mesh not lying on ∂ ΩD . The
linear system (25) is assumed to be monitored by a network
of l threshold sensors. Each sensor, before binary quantization is applied, directly measure the pointwise-in-time-andspace concentration of the contaminant in a point of the
spatial domain Ω. By exploiting (21), such a concentration
can be written as a linear combination of the concentrations
on the grid points, so that the resulting output function takes
the form
zti = Ci xt + vti , i = 1, . . . , l
(26)
3
2.5
η
2
1.5
1
0.5
0
0
1
2
3
ξ
and assumption A1 is fulfilled.
Fig. 2. Mesh used by the MAP-MHE estimator (152 elements, 97 nodes).
V. N UMERICAL RESULTS
RMSE in concentration (g/m2)
6
In this section, we present the simulation results of the
proposed approach applied to the problem of state estimation
of spatially distributed processes, discussed in the previous
section. We consider the simulated system (25)-(26) with
1695 triangular elements, 915 vertices, λ = 0.01 [m2 /s], fixed
integration step length δt = 1 [s], γ = 30 [g/m2 ], and initial
condition of the field vector x0 = 0n [g/m2 ]. The field of
interest is defined over a bounded 2D spatial domain Ω
which covers an area of 7.44 [m2 ] (see Fig. 1), with boundary
condition (19) on the bottom edge and no-flux condition
(20) on the remaining portions of ∂ Ω. Compared to the
5
4
3
2
1
0
0
200
400
600
800
1000
1200
Time (s)
Fig. 3. RMSE in concentration of the MAP-MHE state estimator as a
function of time, for a random network of 5 threshold sensors.
Fig. 1. Concentration field at time t = 100 [s] monitored by a random
network of 20 binary sensors (red ◦).
ground truth simulator, the proposed MH-MAP estimator
implements a coarser mesh (see Fig. 2) of 97 nodes (n = 89,
d = 8), and runs at a slower sample rate (0.1 [Hz]), so that
model uncertainty is taken into account. The initial condition
of the estimated dynamic field is set to x0 = 5 1n [g/m2 ], the
moving window has size N = 5, and the weight matrices
in (4) are chosen as Ψ = 103 In and Q = 102 In . The true
concentrations from (25) are first corrupted with a Gaussian
noise with variance ri , then binary observations are obtained
by applying a different threshold τ i for each sensor i of the
network. Note that, in order to receive informative binary
measurements, τ i , i = 1, ..., l are generated as uniformly
distributed random numbers in the interval (0.05, 29.95),
being (0, 30) the range of nominal concentration values
throughout each experiment. The duration of each simulation
experiment is fixed to 1200 [s] (120 samples).
Fig. 3 shows the performance of the novel MH-MAP state
estimator implemented in MATLAB, in terms of Root Mean
Square Error (RMSE) of the estimated concentration field,
i.e.:
!1
2
α
ket, j k2
RMSE(t) = ∑
,
(27)
α
j=1
where ket, j k is the norm of the estimation error at time t in
the j−th simulation run, averaged over 304 sampling points
(evenly spread within Ω) and α = 100 independent Monte
Carlo realizations. The estimation error is computed at time
t on the basis of the estimate x̂t−N|t . It can be observed that
the proposed estimator successfully estimates the dynamic
field, even when observed by a network of l = 5 randomly
deployed binary sensors, with ri = 0.25 [g/m2 ] ∀i = 1, ..., l.
The effect of measurement noise on the mean value of the
RMSE can be seen in Fig. 4, in which it becomes apparent
0.78
A PPENDIX
RMSE in concentration
0.77
0.76
0.75
0.74
0.73
0.72
0.71
0.7
10-4
10-3
10-2
10-1
100
ri
Fig. 4. RMSE in concentration as a function of the measurement noise
variance, for a fixed constellation of 20 binary sensors. It is shown here
that operating in a noisy environment turns out to be beneficial, for certain
values of ri , to the state estimation problem.
f (b(x)) ∂ b(x)
where a(x) and b(x) are arbitrary functions
∂x
of x, the first and the second derivatives of the function
F i (τ i −Ci xt ) with respect to xt are, respectively, equal to
0.72
RMSE in concentration
Proof of Proposition 1: Under assumption A1, the cost
function (16) is convex if and only if F i (τ i − Ci xt ) and
Φi (τ i − Ci xt ) are log-concave functions, ∀i = 1, . . . , p. A
function f : Rn → R is log-concave if f (x) > 0 for all x
in its domain and ln f (x) is concave [25], namely
2
1
∂ f (x) 0 ∂ f (x)
∂ f (x)
2
f (x) −
< 0.
∇ ln f (x) = 2
f (x)
∂ x2
∂x
∂x
(28)
Let us now consider the CDF Φi (τ i − Ci xt ) and its complementary function F i (τ i −Ci xt ), that are positive functions
for all dti , τ i −Ci xt , i = 1, . .. , l. From thefundamental theoR a(x)
rem of calculus, namely ∂∂x b(x) f (x)dx = f (a(x)) ∂ a(x)
∂x −
0.715
Ci − (τ i −Cii xt )2
∂ F i (τ i −Ci xt )
2r
e
=√
∂ xt
2πri
0.71
(29)
0.705
and
0.7
0.695
5
(τ i −Ci xt )2
∂ 2 F i (τ i −Ci xt )
(Ci )0Ci i
−
i
2ri
√
(τ
−C
x
)e
.
=
t
∂ xt2
ri 2πri
10
20
30
(30)
50
Number of sensors
Fig. 5. RMSE of the concentration estimates as a function of the number
of sensors deployed over the monitoring area.
how for certain values of ri , including an observation noise
with higher variance, can actually improve the quality of the
overall estimates. The results in Fig. 4 numerically demonstrates the validity of the above stated noise-aided paradigm
in the recursive state estimation with binary measurements
and, thus, represents an interesting contribution of this work.
Finally, Fig. 5 shows the evolution of the RMSE as a function
of the number of binary observations available at the fusion
center.
If τ i − Ci xt ≤ 0, then
State estimation with binary sensors has been formulated
as a Moving Horizon (MH) Maximum A posteriori Probability (MAP) optimization problem and it has been shown how
such a problem turns out to be convex in the linear system
case. Simulation results relative to a dynamic field estimation
case-study have exhibited the conjectured noise-aided feature
of the proposed estimator in that the estimation accuracy
improves, starting from a null measurement noise, until the
variance of the latter achieves an optimal value beyond which
estimation performance deteriorates.
Future work on the topic will concern stability properties
of the MH-MAP state estimator and its application to target
tracking with binary proximity sensors.
≤ 0. Hence
∂ 2Fi i
F ≤
∂ x2
i
F is log-
0 and, from (28), it follows that the Q-function
concave. Conversely, if τ i −Ci xt > 0, the log-concavity of F i
depends on the sign of the term
i 0 i
∂ 2Fi i
∂F
∂F
F
−
=
2
∂x
∂x
∂x
Z ∞
i −Ci x )2
2
(τ
(τ i −Ci xt )2
t
−ui
−
(Ci )0Ci −
τ i −Ci xt
i
i
2r
2r
2r
e
e
du − e
.
2πri
ri
τ i −Ci xt
From the convexity properties of the function f (x) = x2 /2,
it can be easily verified for any variable s, k that s2 /2 ≥
2
2
−k2 /2+sk, and hence e−s /2 ≤ e−sk+k /2 [25]. Then, if k > 0,
it holds that
Z ∞
VI. C ONCLUSIONS AND FUTURE WORK
∂ 2 F i (τ i −Ci xt )
∂ xt2
2
e
− s2
ds ≤
Z ∞
2
e
−sk+ k2
k
k
k2
e− 2
ds =
.
k
Since τ i −Ci xt > 0, with a simple change of variable, it can
be stated that
Z ∞
2
(τ i −Ci xt )2
τ i −Ci xt
−ui
−
2r du
2ri
e
≤
e
,
(31)
ri
τ i −Ci xt
proving, as a consequence, the log-concavity of the Qfunction F i (τ i −Ci xt ).
By using the complement rule, the cumulative
distribution function can be written as Φi (τ i − Ci xt ) =
2 i i
ix )
2 i i
ix )
t
t
1 − F i (τ i − Ci xt ) ≥ 0 and ∂ Φ (τ∂ x−C
= − ∂ F (τ∂ x−C
. If
2
2
2
i
t
t
τ i −Ci xt > 0, then ∂∂ xΦ2 Φi < 0 such that Φi is log-concave.
In the remaining case, i.e. τ i − Ci xt ≤ 0, noting that
i
i
2
2
τ −C (xt )
∞
1
1
−u
−u
Φi = √
e 2ri du = √
e 2ri du,
2πri −∞
2πri −(τ i −Ci xt )
it can be observed that the sign of the term
i 0 i
(τ i −Ci xt )2
(Ci )0Ci −
∂Φ
∂ 2 Φi i
∂Φ
2ri
Φ
−
=
e
i
2
∂x
∂x
2πr
∂x
Z ∞
2
(τ i −Ci xt )2
−ui
−
−(τ i −Ci xt )
i
2r
×
e 2r du − e
ri
Z
Z
−(τ i −Ci xt )
is negative, thus proving the log-concavity of the CDF
Φi (τ i −Ci xt ) and the convexity of the whole cost function.
R EFERENCES
[1] L.Y. Wang, J.F. Zhang, and G.G. Yin, “System identification using
binary sensors”, IEEE Transactions on Automatic Control, vol. 48,
no. 11, pp. 1892–1907, 2003.
[2] L.Y. Wang, G.G. Yin, and J.F. Zhang, “Joint identification of plant
rational models and noise distribution functions using binary-valued
observations”, Automatica, vol. 42, no. 4, pp. 543-547, 2006.
[3] B. Ristic, A. Gunatilaka and R. Gailis, “Achievable accuracy in
Gaussian plume parameter estimation using a network of binary
sensors”, Information Fusion, vol. 25, pp. 42-48, 2015.
[4] S. Vijayakumaran, Y. Levinbook and T.F. Wong, “Maximum likelihood
localization of a diffusive point source using binary observations”,
IEEE Transactions on Signal Processing, vol. 55, no. 2, pp. 665–676,
2007.
[5] A. Ribeiro and G. B. Giannakis, “Bandwidth-constrained distributed
estimation for wireless sensor networks - part I: Gaussian case”, IEEE
Transactions on Signal Processing, vol. 54, no. 3, pp. 1131–1143,
2006.
[6] A. Ribeiro and G. B. Giannakis, “Bandwidth-constrained distributed
estimation for wireless sensor networks - part II: unknown probability
density function”, IEEE Transactions on Signal Processing, vol. 54,
no. 7, pp. 2784–2796, 2006.
[7] L.Y. Wang, C. Li, G.G. Yin, L. Guo, and C.-Z. Xu, “State observability and observers of linear-time-invariant systems under irregular
sampling and sensor limitations”, IEEE Transactions on Automatic
Control, vol. 56, no. 11, pp. 2639–2654, 2011.
[8] G. Battistelli, L. Chisci and S. Gherardini, “Moving horizon state
estimation for discrete-time linear systems with binary sensors”,
accepted at the 54th IEEE Conference on Decision and Control, Osaka,
Japan, 2015.
[9] Er-wei Bai, H. E. Baidoo-Williams, R. Mudumbai and S. Dasgupta
“Robust tracking of piecewise linear trajectories with binary sensor
networks”, Automatica, vol. 61, pp. 134-145, 2015.
[10] J. Aslam, Z. Butler, F. Constantin, V. Crespi, G. Cybenko, and D. Rus,
“Tracking a moving object with a binary sensor network”, Proc. 1st
ACM Conf. on Embedded Networked Sensor Systems, pp. 150–161,
Los Angeles, USA, 2003.
[11] P.M. Djuric, M. Vemula and M.F. Bugallo, “Signal processing by
particle filtering for binary sensor networks”, Proc. Digital Signal
Processing Workshop, pp. 263–267, 2004.
[12] P.M. Djuric, M. Vemula and M.F. Bugallo, “Target tracking by particle
filtering in binary sensor networks”, IEEE Transactions on Signal
Processing, vol. 56, no. 6, pp. 2229–2238, 2008.
[13] J. Teng, H. Snoussi and C. Richard, “Decentralized variational filtering
for simultaneous sensor localization and target tracking in binary
sensor networks”, IEEE Transactions on Mobile Computing, vol. 9,
no. 10, pp. 1465–1477, 2010.
[14] A. Capponi, I. Fatkullin and Ling Shi, “Stochastic filtering for
diffusion processes with level crossings”, IEEE Transactions on
Automatic Control, vol. 56, no. 9, pp. 2201-2206, 2011.
[15] G. Ferrari-Trecate, D. Mignone and M. Morari, “Moving horizon
estimation for hybrid systems”, IEEE Transactions on Automatic
Control, vol. 47, no. 10, pp. 1663–1676, 2002.
[16] C.V. Rao, J.B. Rawlings, and D.Q. Mayne, “Constrained state
estimation for nonlinear discrete-time systems: stability and moving
horizon approximations”, IEEE Transactions on Automatic Control,
vol. 48, no. 2, pp. 246-258, 2003.
[17] A. Alessandri, M. Baglietto, and G. Battistelli, “Receding horizon
estimation for discrete-time linear systems”, IEEE Transactions on
Automatic Control, vol. 48, no. 3, pp. 473–478, 2003.
[18] A. Alessandri, M. Baglietto, and G. Battistelli, “Moving horizon state
estimation for nonlinear discrete-time systems: new stability results
and approximation schemes”, Automatica, vol. 44, no. 7, pp. 1753–
1765, 2008.
[19] A. Alessandri, M. Baglietto, G. Battistelli, and V. Zavala, “Advances
in moving horizon estimation for nonlinear systems,” Proc. 49th IEEE
Conference on Decision and Control (CDC), pp. 5681-5688, 2010.
[20] A. Liu, L. Yu, W.-A. Zhang and M.Z.Q. Chen, “Moving horizon
estimation for networked systems with quantized measurements and
packet dropouts”, IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 60, no. 7, pp. 1823–1834, 2013.
[21] R. A. Delgado and G. C. Goodwin, “A combined MAP and Bayesian
scheme for finite data and/or moving horizon estimation”, Automatica,
vol. 50, no. 4, pp. 1116-1121, 2014.
[22] G. Battistelli, L. Chisci, N. Forti, G. Pelosi, and S. Selleri, “Point
source estimation via finite element multiple-model Kalman filtering”,
accepted at 54th IEEE Conference on Decision and Control, Osaka,
Japan, 2015.
[23] G. Battistelli, L. Chisci, N. Forti, G. Pelosi, and S. Selleri, “Distributed finite element Kalman filter”, 14th IEEE European Control
Conference, Linz, Austria, 2015.
[24] S.C. Brenner and L.R. Scott, The mathematical theory of finite element
methods, Springer–Verlag, New York, NY, 1996.
[25] S. Boyd and L. Vandenderghe, Convex Optimization, Cambridge
University Press, Cambridge, UK, 2004.
| 3cs.SY
|
Toward a Robust Sparse Data Representation for
Wireless Sensor Networks
Mohammad Abu Alsheikh∗† , Shaowei Lin† , Hwee-Pink Tan‡ , and Dusit Niyato∗
∗ School
of Computer Engineering, Nanyang Technological University, Singapore 639798
and Sense-abilities Programme, Institute for Infocomm Research, Singapore 138632
‡ School of Information Systems, Singapore Management University, Singapore 188065
arXiv:1508.00230v1 [cs.NI] 2 Aug 2015
† Sense
Abstract—Compressive sensing has been successfully used for
optimized operations in wireless sensor networks. However, raw
data collected by sensors may be neither originally sparse nor
easily transformed into a sparse data representation. This paper
addresses the problem of transforming source data collected by
sensor nodes into a sparse representation with a few nonzero
elements. Our contributions that address three major issues
include: 1) an effective method that extracts population sparsity
of the data, 2) a sparsity ratio guarantee scheme, and 3) a
customized learning algorithm of the sparsifying dictionary. We
introduce an unsupervised neural network to extract an intrinsic
sparse coding of the data. The sparse codes are generated at
the activation of the hidden layer using a sparsity nomination
constraint and a shrinking mechanism. Our analysis using real
data samples shows that the proposed method outperforms
conventional sparsity-inducing methods.
Abstract—Sparse coding, compressive sensing, sparse autoencoders, wireless sensor networks.
I. I NTRODUCTION
A sparsely-activated data (a few nonzero elements in a
sample vector) may naturally exist for compressive sensing (CS) applications in wireless sensor networks (WSNs)
such as the path reconstruction problem [1], indoor localization [2], and sparse event detection [3]. On the other
hand, a sparse data representation cannot be easily induced
in many other real-world contexts (e.g., in meteorological
applications and environmental data gathering). In particular,
noise patterns are usually presented in collected data from
WSNs which greatly affect the performance of conventional
sparsity-inducing (transformation) algorithms such as the Haar
wavelet and discrete cosine transforms [4]. This motivates the
quest for noise-robust and effective sparsity-inducing methods
for WSNs.
One of the breakthroughs in recent deep learning paradigms
for finding high level data abstractions is achieved by introducing sparsity constraints on data representations, e.g., the
Kullback–Leibler divergence [5], rectifier function [6], and
topographic coding [7]. These methods are introduced for
extracting intrinsic features from the data in a similar way
that the human brain does while encoding sensory organ data,
e.g., the low percentage of spikes in a visual cortex [8].
In particular, sparse deep learning methods generate sparse
representations across training data for each single unit (i.e.,
lifetime sparsity), and they neither guarantee sparsity for each
input signal nor assert on the number of nonzero values in
the sparse codes. However, a practical CS implementation in
WSNs requires a sparse representation for each input signal
(i.e., population sparsity) with a sparsity ratio guarantee.
Specifically, the CS solution to the underdetermined system
(more number of unknowns than the number of equations)
is dependent on the sparsity ratio of the signal, and the
sparsity-inducing mechanism must assert an upper limit for the
sparsity ratio. This sparsity bounding is necessary in WSNs
as it enables using only one flat acquisition matrix for data
encoding in the node. Therefore, it reduces the CS overhead
in terms of memory for storing many measurement matrices
in transmitting node and data control exchange as there is no
need to send out rate control messages.
The main contributions of this paper can be summarized
into three folds as follows.
1) This paper introduces an effective, population sparsityinducing algorithm with sparsity ratio guarantee. The
algorithm is based on a customized unsupervised neural
network model of three layers (also called an autoencoder network) that generates the required, sparse coding
at the second (hidden) layer. In the proposed shrinking
sparse autoencoder (SSAE), the sparsity is achieved by
introducing a regularization term to the cost function of
the basic autoencoder.
2) We customize the learning algorithm to meet WSN
characteristics. For example, the activations of the hidden layer during parameter learning stage are rounded
to only three places to consider limited computational
precision of the node. The rounding considers the low
precision computations of sensor nodes, and it reduces
the compressed data size and data transmission load.
3) We present a customized learning method that optimizes
the SSAE cost function. Basically, the back propagated
error is only used to update the nonzero and active
neurons with dominant output values for each input pattern. Moreover, a shrinking mechanism that guarantees
the sparsity bound is also used during the learning of
the SSAE’s parameters. Therefore, an SSAE asserts on
the number of nonzero elements generated at any time
instant.
The literature is rich with sparsity-based methods that are
designed for optimized WSN operations [1]–[4], [9]–[14].
Nonetheless, much less attention is given to the sparsityinducing stage, and using straightforward methods to extract
Short distance
connection
Region of interest (RoI)
A backhaul
connection
Base station (BS)
Sensor nodes
(motes)
Gateway
(energy constrained)
Fig. 1: Compressive sensing (CS) based data aggregation
model: The RoI is assumed to be relatively far from the BS.
Therefore, a gateway is designed to transmit compressed data
over a costly long distance wireless connection.
sparsity basis is common in previous studies such as using
principal component analysis (PCA) [10], discrete cosine
transform (DCT) [4], [11], [12], discrete Fourier transform
(DFT) [9], discrete wavelet transforms [4], [13], and difference
matrices [4], [15]. However, the sparse coding discipline has
evolved considerable advances that significantly enhance the
sparsity-inducing and hence overall WSN operations. Therefore, this paper is intended to introduce a robust and more
effective sparsity-inducing method. The proposed method consists of three steps: (i) data collection, (ii) offline training and
modeling, and (iii) online sparse code generation. An example
of the online sparse code generation for a CS application is
shown in Figure 2 which will be described in details later.
The rest of the paper is organized as follows. In Section II,
the problem formulation is presented. Section III describes the
proposed algorithm and the SSAE structure. Then, Section IV
discusses important practical issues of training and fitting the
proposed model. In Section V, numerical results using realworld data set are presented. Finally, Section VI summarizes
this paper.
II. P ROBLEM F ORMULATION
Consider a dense wireless sensor network consisting of N
nodes, as in Figure 1, that collects data about a region of
interest (RoI). Each sensor i (where i = 1, . . . , N ) collects
a real-valued sample xi (e.g., temperature measurements)
at a predefined sampling period and transmits packets at a
configured transmission power that is not sufficient to reach the
base station (BS) due to long distance propagation. Therefore,
a gateway (GW) is used to collect a data vector x ∈ RN from
all sensor nodes and relay it to the BS for further analysis and
processing. Thereafter, a historical data matrix X ∈ RT ×N is
formulated at the BS containing the collected data vectors as
its rows, where T is the number of collected vectors. Here,
the sensors’ oscillators are assumed to be synchronized to the
GW’s clock.
After collecting sufficient historical samples (details of data
collection are elucidated in Section IV-A), and as the GW
is energy and bandwidth constrained, the GW employs CS
to spatially compress the data into a smaller data size. The
radio transceiver is the most energy consuming unit in an
ordinary sensor node [16]. Thereby, the energy consumption
becomes more critical in the GW unit as it transmits huge
data over the backhaul connection, while sensor nodes are
assumed to transmit for short distances. It is important to
note that our algorithm can be also temporally applied at
each individual sensor node. However, data delivery latency is
provoked as temporal samples must be collected at the node
before being transmitted as one compressed chunk. Next, we
give an overview of the CS framework and its implementation
at the GW device, and the data reconstruction at the BS unit.
A. Compressive Sensing (CS)
CS is a signal processing method for effective data recovery
from a few data samples than the Nyquist rate [17]. Assuming
a sparse signal s ∈ RL that has only K nonzero elements;
therefore, s is called a K-sparse signal, and the sparsity ratio η
is equal to K
L . Moreover, suppose a measurement (or sensing)
matrix Φ ∈ RM ×L that obeys the restricted isometry property
(RIP) [18]. Here, M is assumed to be much smaller than L;
therefore, Φ is a flat matrix with more columns than rows.
The sensing system under consideration that is executed by
the GW to compress data can be expressed as
y = Φs
(1)
M
where y ∈ R is the resulted measurement vector. Φ can be
sampled from different distributions to meet the RIP such as
the Gaussian distribution [19]. Moreover, for high probability
recovery, M must also meet the following constraint [20]:
L
(2)
M ≥ ρK log2
K
where ρ is a constant, and M N . At the BS unit, the
reconstruction of s from y can be achieved by minimizing the
following relaxed problem [21]:
s∗ = arg
min
ky−Φsk2 ≤
ksk1
(3)
where is a small constant. The optimization problem (3) can
be solved using a regularized least square method called least
absolute shrinking and selection operator (LASSO) [22].
B. Sparsity-inducing
Clearly, the whole CS framework is based on the sparsity
assumption. Natural signal such as sound and images can
be transformed into a sparse form by projecting them into a
suitable basis [17]. However, this is not the case when dealing
with WSN data. More precisely, sensor nodes produce noisy
readings of the form
x = x∗ + z
(4)
where x∗ ∈ RN is the noiseless data vector of the physical
phenomenon, and z ∈ RN is the added noise vector. Noise
values are assumed to be independent Gaussian variables with
zero mean and variance σz2 such that z ∼ N 0, σz2 IN .
Therefore, even through the neighbor sensors are spatially
correlated and hence compressible, the noise existence hampers the accurate approximation of source signal x using
linear projection methods. In particular, smooth signal are
representable using linear combinations of Fourier bases, and
Source signal (xÎÂ128 )
Reconstructed signal(xˆ ÎÂ128)
hidden layer
Input layer
h1
s1
d2
Sparsity-recovery
method
h3
d3
Sparse signal (sÎÂ128 )
h4
d4
5 non - zero values
Compressive
sensing decoding
Compressive
sensing encoding
Transmitted signal (yÎÂ35)
.
.
.
.
.
.
dˆ 2
dˆ 3
dˆ 4
. A neuron
(a computational
unit)
.
.
dN
hL
Base station
(BS)
s2
Shrinking mechanism
h2
Sparsity-inducing
method
Output layer
dˆ1
d1
sL
dˆ N
Fig. 3: Illustration of the SSAE structure.
Gateway
Fig. 2: Example of data compression, transmission, and recovery operations using CS and sparsity-inducing models.
smooth piecewise signals are linearly representable in wavelet
bases [11]. Nonetheless, the smoothness condition is not
guaranteed in sensor data as data samples are usually affected
by noise patterns, and commercial sensors sense phenomenon
with finite precision and are not noise robust. For example, a
few noise readings can destroy the sparsity pattern of a DCT
transformed data [12].
The main aim of any robust sparsity-inducing mechanism is
to transform the source signal x ∈ RN into the sparse signal
s ∈ RL . An upper bound guarantee on the sparsity ratio of the
generated signal s is a “must-have” feature in most sparsitybased applications such as in CS. In particular, this guarantee enables designing low memory and low communication
overhead applications for WSNs as a single sensing matrix
Φ is used by the GW unit to compress data. Then, the BS
does not require any information from the GW to recover the
reconstruction signal x other than the measurement vector y,
where x is a reconstruction of the noiseless data vector x∗ .
An example of the system online operational procedure is
shown in Figure 2 which includes the sparsity-inducing and CS
components. The next section presents the proposed sparsityinducing mechanism.
III. S HRINKING S PARSE AUTOENCODER (SSAE)
In this section, we introduce an autoencoder’s variant which
we call shrinking sparse autoencoder (SSAE) as shown in
Figure 3. The SSAE network is specially designed to transform
sensory data from its original domain into a sparse domain.
The SSAE structure consists of three neural (or computational
unit) layers. Firstly, an input layer that is connected to the
input signal d ∈ RN , where N is the number of sensor nodes
in the network. Briefly, d is a sphered version of the raw
sensor data x, where {di ∈ R| − 1 < di < 1} as described in
Section IV-B. Secondly, a hidden layer is used to generate an
intrinsic code h ∈ RL at its activation. Thirdly, an output layer
that includes the same number of neurons as the input layer
and generates a recovery of the input data d̂ ∈ RN . The layers
are connected to each other using the following formulations:
h = f W(1) d + b(1)
(5)
d̂ = f W(2) s + b(2)
(6)
where W(1) is the weight matrix connecting the input and
hidden layers, W(2) is the weight matrix connecting the
hidden and output layers, and b(1) and b(2) are the biases
of the input and hidden layers, respectively. Additionally, s
is the sparse data representation that is obtained by applying
the shrinking operation as described in Section III-A. For
simplicity, we define θ to contain all the SSAE’s parameters
.
such that θ = W(1) , W(2) , b(1) , b(2) . Moreover, f (·) is the
non-linear hyperbolic tangent function.
The SSAE’s cost function Γ (·) includes two terms as
follows:
!
T
2
1 X 1 b (u)
(u)
Γ (θ; D) =
d −d
+
T u=1 2
!
T
2
γ X
(u)
log
1+ h
(7)
T u=1 10
where D ∈ RT ×N is the training matrix of historical data
T
such that each input vector d(u) u=1 is stored in a row of
this matrix, and h(u) is the hidden layer activation of d(u) .
Moreover, T is the training set size configured at the offline
training algorithm (the details are given in Section III-B). As
with any other autoencoder, the first term is the average sum of
the difference between input vectors and their reconstructions
at the output layer. This term is used to encourage the neural
network to reconstruct its input data at the output layer. The
second term is used to encourage sparsity at the generated
coding in the hidden layer. The sparsity penalty γ is a hyper-
parameter to manage the weights of each term in the optimization problem. In other words, using a big value for γ results
in highly sparse representation, but with poor reconstruction
capability. Then, the well-known delta rule can be used to
update the SSAE’s weights and biases as follows [23]:
(q)
Wij
(q)
bi
.
(q)
= Wij − α
∂
Γ (θ; D) ,
(q)
∂Wij
∂
.
(q)
= bi − α (q) Γ (θ; D)
∂bi
(8)
(9)
where α is the learning rate, and q ∈ {1, 2} is the layer
number within the SSAE network. These update rules are
executed at each iteration of a gradient descent method. The
partial derivative is given by
∂
Γ (θ; D) =
(q)
∂Wij
T
1 X ∂
(u)
Γ
θ;
d
T u=1 ∂W (q)
ij
Algorithm 1 Pseudo-code for the shrinking operation of
hidden layer’s neurons.
1: Input h ∈ RL : hidden layer activation before shrinking
2: Input K: maximum nonzero activations
3: s = h
. copy operation
4: for i = 0 to L − K do
5:
p=0
6:
for j = 0 to (L − 1) do
7:
if |sp | > |sj | and |sj | > 0 then
8:
p=j
9:
end if
10:
end for
11:
sp = 0
. zero-out smallest value
12: end for
13: Output s ∈ RL
(10)
where Γ θ; d(u) is the cost function defined for a single
sample d(u) ∈ D. This means that the overall partial derivative
of (7) is found by averaging the partial derivatives of all
input samples. The second term of (7) only affects the partial
derivative of the hidden layer (q = 2) which is computed as
follows:
T
2
γ X ∂
(u)
γ
log
1
+
h
=
10
T u=1 ∂W (2)
ij
!
T
X
2h(u)
γ
0
(2) (u)
(2)
2 f W d + b
loge (10) × T u=1 1 + h(u)
(11)
2
where f 0 (·) = 1 − (f (·)) is the element-wise derivative
of f (·). Thereby, the SSAE is designed to generate many
zeros at the hidden layer. One can think of a neuron as being
active when its output is not equal to zero, and an inactive
neuron does not participate in forwarding the input data to the
output (because it does not generates signals). To this end, two
important issues of the second term of (7) must be noted as
follows:
• The second term minimizes the hidden layer activation,
but it still does not ensure exactly zero activations.
• It does not guarantee sparsity ratio at the generated codes.
Accordingly, a shrinking mechanism must be applied at the
hidden layer activation and before propagating them to the
output layer to reconstruct the input. In particular, one can
think of the second term as only being used as a mechanism
of nominating the most promising neurons to be deactivated
by the shrinking mechanism as described in the next section.
A. Shrinking (Pruning) Scheme
Even though the cost function of the SSAE is designed
to generate a sparse data coding at the hidden layer, it
does still neither guarantee a coding with population sparsity
(sparsity at each input vector) nor assert on the maximum
nonzeros for each input. Equally important, it will most likely
generate values close to, but not absolutely zero. Therefore, we
propose a simple shrinking mechanism that can complete the
design cycle. For each input vector, the proposed shrinking
mechanism “zero out” the least dominant neurons from the
hidden layer, and therefore switching them to the deactivation
mode. The least dominant neurons are the ones with the
least effect on the data reconstruction at the output layer,
and hence the minimum activation values that result from the
sparsity restrictions. Therefore, only K active neurons at the
hidden layer forward propagate the input through the SSAE
network, and the remaining L − K neurons are switched off.
An optimized implementation of the shrinking scheme is given
by the pseudo-code in Algorithm 1, where |·| is the absolute
value function.
B. Offline Training
The SSAE’s parameter adjustment is done during an offline
training stage. As a resource demanding process, the training
must be performed at the BS unit, and then the SSAE’s
parameters (θ) are disseminated for online data compression at GW. The learning stage and SSAE’s parameters are
tuned using a resourceful BS with relatively high precision
operations. However, GW is usually constrained in terms of
computational resources and computational precision (i.e., the
machine epsilon value). Therefore, rounding the activation at
the hidden layer is needed during the learning stage to match
the GW’s low precision. Moreover, with rounding, less data
is transmitted from GW to BS.
To learn the SSAE’s parameters (θ), we minimize
(7) by using a non-linear quasi-Newton method called
the limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) method [24]. However, firstly the collected historical
data X ∈ RT ×N must be randomly shuffled. This is because
sensors’ readings are highly correlated over time, and a nonshuffled data causes the SSAE to dominantly learn the training
data’ patterns in training data only. Therefore, the shuffling
step ensures that the training and testing data sets contain
all possible data patterns. Moreover, the cross validation
technique [25] is an effective method for testing the model
Algorithm 2 The offline training algorithm.
1: Input X ∈ RT ×N : historical sensor data
2: Input K: maximum nonzero activations
3: Input γ: sparsity hyper-parameter
4: Input ϕ: number of folds for cross validation
5: Randomly shuffle X
6: Divide X into ϕ folds X1 , . . . , Xϕ
7: for all Xi , i = 1, . . . , ϕ do
8:
for all x ∈ (X \ Xi ) do
. held out Xi for testing
9:
Sphere input x to get d using (12)
10:
Append d to D
11:
end for
12:
repeat
13:
for all d ∈ D do
14:
Forwardly propagate d to compute h using (5)
15:
Shrink h to get s as in Algorithm 1
16:
Compute d̂ using (6)
17:
end for
18:
Compute the cost value using (7)
19:
Compute the gradient vector as in (10)
20:
Update θ using the L-BFGS method
21:
until learning converges
22:
Compute accuracy using Xi
23: end for
24: Compute average
accuracy of the ϕ folds
.
25: Output θ = W(1) , W(2) , b(1) , b(2)
to the input layer. In other words, this ensures that any neuron
in the hidden layer will be active for some input patterns,
and hence no “always-off” neuron exists. This increases the
model performance of generalizing to non local data, and
hence it performs well on extremely non linear data, as all
neurons participating increases the possible code formulations
(i.e., the number of distinct combinations is increased when
having more active neurons). Figure 4b shows hidden layer
activations over time. Here, two main desirable properties can
be observed
1) Population sparsity is achieved, and the maximum number of active neurons at any time instant is guaranteed
by the SSAE network. This upper bound of nonzeros in
a generated sparse code considers the tradeoff between
the recovery error and compression ratio of the data
aggregation model. Therefore, only a single sensing
matrix is needed when using CS to create a measurement
vector at the GW node.
2) All neurons are participating in the sparse code generation, and without any “always-off” neuron. Moreover,
the activation values of the active neurons are not concentrated around very small values near zero. This feature cannot be achieved in conventional average activity
ratio sparse autoencoders, such as the Kullback–Leibler
divergence, as they are designed for lifetime sparsity
only.
IV. D ISCUSSION AND P RACTICAL C ONSIDERATIONS
generalization capabilities, while benefiting from all available
samples for training. Cross validation divides the training data
into ϕ groups (e.g., 10 groups) then at each time, one group
is held out for testing while using the remaining for model
fitting. Then, the model performance is found by averaging
errors of all cross validation’s groups. The offline learning is
described in Algorithm 2.
The learning algorithm is computationally intensive for
sensor nodes and must be performed at the BS. Moreover,
if the statistical parameters of the underlying phenomenon
change,
offline
training must be re-executed and an updated
(1) the
W , b(1) should be disseminated to the nodes.
C. Computational Complexity
The online encoding and decoding of sparse codes are
lightweight. In particular, the GW (or a sensor
node) can
generate sparse codes by only using W(1) , b(1) as in (5) and
Algorithm 1 with O(L × N ) of overall time complexity.
The
data recovery is performed at the BS by using W(2) , b(2)
as in (6) with a similar time complexity of O(L × N ).
D. Sparse Codes
For the verification and analysis in the following sections,
a meteorological data set from the Sensorscope project [26] is
used. The data set contains surface temperature samples of 23
sensors. The learning curve of the SSAE is shown in Figure 4a.
An important indication of successful SSAE training is ensuring that hidden neurons are not connected with zero weights
In this section, some practical issues of the SSAE training
and fitting are discussed.
A. Data Collection
A crucial aspect of machine leaning-based approaches, such
as the SSAE network, is the training data requirement. A
system designer may have access to a large historical data
set that is collected in the past. This historical data can be
used to train the SSAE’s model. However, this is not the
case for new WSN’s deployments, and the lack of sensor data
hinders the accurate fitting of the SSAE’s parameters (i.e.,
θ). Clearly, the SSAE’s model needs to globally generalize to
unseen data samples. In any machine learning method, having
more training data can improve generalization performance,
but having more data is not the only solution [27]. In WSNs,
the following issues must be considered when using an SSAE
as a sparsity inducing method.
1) It is assumed that sensor nodes are densely deployed
and hence spatially correlated with each other (e.g., as
in Figure 5 for the Sensorscope project’s data). SSAE
learns these spatial correlation and redundant patterns
in the nodes’ collected data. Therefore, if the underlying phenomenon becomes different in the way that it
changes the nodes’ spatial correlation, then new data
collection and offline model fitting must be performed.
2) The amount of data required to fit the SSAE’s model
depends on the underlying sensed phenomenon, and for
(a)
(a)
(b)
Fig. 6: Data sphering and its effects on data by showing
histograms and basic statistical values. (a) Raw data in the
range of [−16.15, 47.91]. (b) Sphered data that is scaled to a
new range of (−1, 1) with a Gaussian-like distribution.
stage. Data sphering is simply achieved by applying the
following operation on each sensors’ raw input vector x ∈ RN
max (min ((x − x) , 3σ) , −3σ)
(12)
3σ
where σ is the standard
deviation of the historical training
PN
matrix X, x = N1 i=1 xi is the arithmetic mean of each input
vector, and again d ∈ RN is the SSAE’s input vector which
is the resulting data vector after sphering. Unlike the standard element-wise standardization, this subtracts the arithmetic
mean of each input vector and not the whole training matrix’s
mean value. The effect of data sphering on training data is
shown in Figure 6. Clearly, the data is transformed into a
smoother Gaussian-like curve with zero mean (other statistical
parameters are also shown). Equally important, the resulting
scale of sphered data is in the (−1, 1) interval, which makes it
suitable for the operation of the hyperbolic tangent function. In
particular, the hyperbolic tangent function generates an output
in the range of (−1, 1) and without data pre-processing to this
range, the SSAE cannot produce outputs similar to input data.
The reverse operation of data sphering is required at BS to
reconstruct the original raw sensors’ vector x̂ ∈ RN from the
SSAE’s output values d̂ ∈ RN . The reverse operation is given
as
x̂ = desphere(d̂, x, σ) = 3σ d̂ + x.
(13)
d = sphere(x, σ) =
(b)
Fig. 4: An SSAE which is designed to produce a maximum
of 5 nonzero values at each time instant (η = 0.2) for 23
sensors (i.e., x ∈ R23 ). (a) A learning curve that shows the
convergence of the offline learning algorithm, and (b) activation values of the hidden layer’s neurons.
Fig. 5: Surface temperature readings of 4 neighbor sensors
from the Sensorscope deployment over 1 day (1 sample every
2 minutes). This shows the spatial correlation among sensors’
measurements, and hence data compressibility.
Here, σ is constant for all recovered vectors, and therefore
can be stored at the BS. However, x must be sent from the
GW to the BS along with the compressed data. Therefore, the
transmitted data size when using CS is M + 1.
V. N UMERICAL R ESULTS
In this section, we evaluate the performance of the SSAEbased sparsity inducing method.
A. SSAE Tunning
more complex correlation patterns among sensors, more
data samples are needed.
B. Data Sphering
Before using historical sensor data to train the SSAE, a
pre-processing operation is required, namely the data sphering
One of the main difficulties of applying neural networkbased methods is the numerical tuning of the network hyperparameters. Hyper-parameter setting of autoencoder’s variants
can be facilitated by searching over a scale of values in the logdomain (e.g., values such as 10−1 , 10−2 , 10−3 , . . .), and then
the value that minimizes the cross validation error is selected
1.3
1.2
RMSE
1.1
1.0
0.9
0.8
° =0:05
° =0:1
° =0:2
° =0:5
Sparsity penalty
° =1:0
Fig. 7: Sparsity parameter setting for η = 0.217. Bars
represent root mean square error (RMSE) values over 10 runs.
The maximum performance achieved at γ = 0.2.
2.5
SSAE-CS
PCA-CS
DCT-CS
DFT-CS
DL-CS
2.0
RMSE
1.5
1.0
0.5
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Sparsity ratio
Fig. 8: Root
mean square error (RMSE) versus sparsity ratio
η= K
.
L
comparison as they are widely used in the CS literature [4],
[9]–[12]. Two important observations can be made.
1) Most sparsity inducing algorithms will achieve a relatively similar recovery error at high values of η. However, these high sparsity ratio values (e.g., η > 0.7)
are not typical in practical applications as the reduction
in data size is not noticeable. Therefore, these values
cannot be used for CS’s applications as the measurement
vector size will be similar to the source signal size (i.e.,
N ≈ M ). On the other hand, SSAE significantly outperforms conventional methods for practical low sparsity
ratios and when the nonzero values in the generated
sparse codes are required to be minimized.
2) Conventional DL methods (e.g., [29], [30]) use the `1
minimization to model the raw data as linear combinations of sparse bases. In this paper, we used the scikitlearn library [31] for testing the dictionary learning
method in which the coordinate descent method is
used to find the LASSO problem solution. Similar to
our algorithm, the scikit-learn’s implementation enables
setting the required sparsity ratio by defining the number
of nonzero coefficients in the sparse code, while we
set the remaining parameters to their default values. We
normalize the data to a zero mean and a unit variance
before learning the dictionary model. In addition to the
slightly better performance, we also noticed that the
learning time of the SSAE method is also shorter than
the DL method. This is significant for large-scale WSNs.
C. Noisy Data
accordingly [28]. Figure 7 shows the setting of the sparsity
hyper-parameter γ for sparsity ratio η = 0.39. The sparsity
term in (7) can be interpreted as sparsity nomination term, that
is fed to the shrinking mechanism to generate sparse codes.
Therefore, trying different values of γ is useful to achieve
maximum signal reconstruction performance. For SSAE in
the next experiments, the following function is used for the
sparsity penalty γ settings
γ (η) = 0.26 − 0.26η,
(14)
which is found by manually fitting the hyper-parameter γ for
two values of η, as described above, and then finding the line
connecting these two manually fitted points.
B. Comparing to Benchmarks
Using a difference matrix that captures the difference between adjacent and correlated values as a sparse basis was used
in [4], [15]. Similar to [4], we noted the difference matrix’s
poor performance in sparsifying the data, and hence it is not
included in our comparison analysis.
Figure 8 shows a comparison between the SSAE recovery performance and other conventional methods including
principal component analysis (PCA), discrete Fourier transform (DFT), discrete cosine transform (DCT), and dictionary
learning (DL). These conventional methods are chosen for
Sensors may report imprecise measurements due to external
noise sources, inaccurate sensor calibration, unstable power
supply, and imperfect node design [32]. In this section, we
assume that noise values are independent Gaussian variables
with zero mean and variance σz2 such that z ∼ N 0, σz2 IN ,
where z ∈ RN is an added noise vector. We noticed that the
SSAE method does not only allows the compression of the
sensors’ data, but it also helps in estimating the noiseless data
vector of the physical phenomenon x∗ ∈ RN .
An overcomplete sparse representation is achieved when
the number of hidden layer’s neurons (sparse code’s size) is
greater than the input layer’s neurons (i.e., L > N ). However,
the measurement vector’s size M of CS is proportional to the
sparse code’s size as in (3). Therefore, the number of nonzero
items must be minimized, and less nonzero coefficients are
defined in the overcomplete sparse code. On the other hand,
using more neurons in the hidden layers can result in the
overfitting problem [33]. Overfitting degrades the neural network’s reconstruction performance
and increases the learning
.
time of the parameters θ = W(1) , W(2) , b(1) , b(2) . Table I
summarizes the experiments of using overcomplete sparse representation. The results also include the case of adding external
noise z ∼ N (0, IN ) to sensors’ measurements. This shows that
the overcomplete case is useful in unreliable network to reduce
the noise effects while producing sparse codes. However, in
noise-free networks, using overcomplete codes can degrade the
TABLE I: System performance with different numbers of
hidden neurons.
L
K
γ
M
RMSE (no
external noise)
23
5
0.2
12
0.987
25
5
0.25
12
0.930 (best)
30
4
0.5
12
32
4
0.6
12
0.982
(overfitting)
1.027
(overfitting)
RMSE (noise
σz2 = 1)
1.522
(unreliable)
1.512
(unreliable)
1.259 (best)
1.338
sparsity-inducing algorithm performance due to the overfitting
problem.
VI. S UMMARY
In this paper, we have introduced a sparsity-inducing algorithm for data aggregation of non-sparse signal in wireless
sensor networks. The proposed method consists of three steps:
data collection, offline training and modeling, and online
sparse code generation. The modeling scheme is based on
a neural network with three layers, where the sparse codes
are exposed at the hidden layer’s neurons. A cost function is
introduced as a sparsity nomination scheme. Then, a shrinking
mechanism is used to switch off the least dominant neurons in
the hidden layer, while asserting on the number of generated
nonzero values in the sparse code. The resulting scheme can
be used in many applications such as in compressive sensingbased data aggregation schemes.
For future research, we will analytically study the energy consumption and computational burdens of the proposed
scheme.
R EFERENCES
[1] Z. Liu, Z. Li, M. Li, W. Xing, and D. Lu, “Path reconstruction in
dynamic wireless sensor networks using compressive sensing,” in Proc.
15th ACM Int. Symp. on Mobile Ad Hoc Networking and Computing.
ACM, 2014, pp. 297–306.
[2] S.-Y. Fu, X.-K. Kuai, R. Zheng, G.-S. Yang, and Z.-G. Hou, “Compressive sensing approach based mapping and localization for mobile robot
in an indoor wireless sensor network,” in Proc. Int. Conf. Networking,
Sensing and Control. IEEE, 2010, pp. 122–127.
[3] J. Meng, H. Li, and Z. Han, “Sparse event detection in wireless sensor
networks using compressive sensing,” in Proc. 43rd Annu. Conf. Inform.
Sciences and Syst. IEEE, 2009, pp. 181–185.
[4] G. Quer, R. Masiero, D. Munaretto, M. Rossi, J. Widmer, and M. Zorzi,
“On the interplay between routing and signal representation for compressive sensing in wireless sensor networks,” in Proc. Inform. Theory
and Applicat. Workshop. IEEE, 2009, pp. 206–215.
[5] H. Lee, C. Ekanadham, and A. Y. Ng, “Sparse deep belief net model
for visual area V2,” in Proc. 21st Conf. Advances in Neural Inform.
Process. Syst., 2008, pp. 873–880.
[6] X. Glorot, A. Bordes, and Y. Bengio, “Deep sparse rectifier networks,”
in Proc. 14th Int. Conf. Artificial Intell. and Stat., vol. 15, 2011, pp.
315–323.
[7] Y. LeCun, “Learning invariant feature hierarchies,” in Proc. 12th European Conf. Comput. Vision. Springer, 2012, pp. 496–505.
[8] P. Lennie, “The cost of cortical computation,” Current Biology, vol. 13,
no. 6, pp. 493–497, 2003.
[9] F. Fazel, M. Fazel, and M. Stojanovic, “Random access compressed
sensing over fading and noisy communication channels,” IEEE Trans.
Wireless Commun., vol. 12, no. 5, pp. 2114–2125, 2013.
[10] R. Masiero, G. Quer, D. Munaretto, M. Rossi, J. Widmer, and M. Zorzi,
“Data acquisition through joint compressive sensing and principal component analysis,” in Proc. Global Telecommun. Conf. IEEE, 2009, pp.
1–6.
[11] W. Bajwa, J. Haupt, A. Sayeed, and R. Nowak, “Compressive wireless
sensing,” in Proc. 5th Int. Conf. Inform. Process. in Sensor Networks.
ACM, 2006, pp. 134–142.
[12] C. Luo, F. Wu, J. Sun, and C. W. Chen, “Compressive data gathering
for large-scale wireless sensor networks,” in Proc. 15th Annu. Int. Conf.
Mobile Computing and Networking. ACM, 2009, pp. 145–156.
[13] A. Griffin and P. Tsakalides, “Compressed sensing of audio signals using
multiple sensors,” in Proc. 16th European Signal Process. Conf., 2008,
pp. 1–5.
[14] P. Misra, W. Hu, M. Yang, and S. Jha, “Efficient cross-correlation via
sparse representation in sensor networks,” in Proc. 11th Int. Conf. on
Inf. Processing in Sensor Networks. ACM, 2012, pp. 13–24.
[15] X. Wu and M. Liu, “In-situ soil moisture sensing: Measurement scheduling and estimation using compressive sensing,” in Proc. 11th Int. Conf.
Inform. Process. in Sensor Networks. ACM, 2012, pp. 1–12.
[16] G. J. Pottie and W. J. Kaiser, “Wireless integrated network sensors,”
Commun. of the ACM, vol. 43, no. 5, pp. 51–58, 2000.
[17] S. Qaisar, R. M. Bilal, W. Iqbal, M. Naureen, and S. Lee, “Compressive
sensing: From theory to applications, a survey,” J. of Commun. Networks,
vol. 15, no. 5, pp. 443–456, 2013.
[18] E. J. Candes, “The restricted isometry property and its implications for
compressed sensing,” Comptes Rendus Mathematique, vol. 346, no. 9,
pp. 589–592, 2008.
[19] Z. Chen and J. J. Dongarra, “Condition numbers of gaussian random
matrices,” SIAM J. on Matrix Analysis and Applicat., vol. 27, no. 3, pp.
603–620, 2005.
[20] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles:
Exact signal reconstruction from highly incomplete frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, 2006.
[21] E. J. Candes, J. K. Romberg, and T. Tao, “Stable signal recovery
from incomplete and inaccurate measurements,” Commun. on Pure and
Applied Mathematics, vol. 59, no. 8, pp. 1207–1223, 2006.
[22] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. of
the Royal Statistical Soc., pp. 267–288, 1996.
[23] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Cognitive modeling, 1988.
[24] R. H. Byrd, P. Lu, J. Nocedal, and C. Zhu, “A limited memory algorithm
for bound constrained optimization,” SIAM J. on Scientific Computing,
vol. 16, no. 5, pp. 1190–1208, 1995.
[25] R. Kohavi et al., “A study of cross-validation and bootstrap for accuracy
estimation and model selection,” in Proc. Int. Joint Conf. Artificial Intell.,
vol. 14, no. 2, 1995, pp. 1137–1145.
[26] “Sensorscope: Sensor networks for environmental monitoring,” 2007.
[Online]. Available: http://lcav.epfl.ch/op/edit/sensorscope-en
[27] P. Domingos, “A few useful things to know about machine learning,”
Commun. of the ACM, vol. 55, no. 10, pp. 78–87, 2012.
[28] Y. Bengio, “Practical recommendations for gradient-based training of
deep architectures,” in Neural Networks: Tricks of the Trade. Springer,
2012, pp. 437–478.
[29] H. Lee, A. Battle, R. Raina, and A. Y. Ng, “Efficient sparse coding
algorithms,” in Proc. 19th Advances in Neural Inform. Process. Syst.,
2006, pp. 801–808.
[30] J. Mairal, F. Bach, J. Ponce, and G. Sapiro, “Online dictionary learning
for sparse coding,” in Proc. 26th Annu. Int. Conf. Machine Learning.
ACM, 2009, pp. 689–696.
[31] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg et al.,
“Scikit-learn: Machine learning in python,” The J. of Machine Learning
Research, vol. 12, pp. 2825–2830, 2011.
[32] K. Ni, N. Ramanathan, M. N. H. Chehade, L. Balzano, S. Nair,
S. Zahedi, E. Kohler, G. Pottie, M. Hansen, and M. Srivastava, “Sensor
network data fault types,” ACM Trans. Sensor Networks, vol. 5, no. 3,
p. 25, 2009.
[33] Y. Liu, J. A. Starzyk, and Z. Zhu, “Optimized approximation algorithm
in neural networks without overfitting,” IEEE Trans. Neural Netw.,
vol. 19, no. 6, pp. 983–995, 2008.
| 9cs.NE
|
Deep Residual Learning and PDEs on Manifold
arXiv:1708.05115v3 [cs.IT] 25 Jan 2018
Zhen Li
∗
Zuoqiang Shi†
Abstract
In this paper, we formulate the deep residual network (ResNet) as a control problem
of transport equation. In ResNet, the transport equation is solved along the characteristics. Based on this observation, deep neural network is closely related to the control
problem of PDEs on manifold. We propose several models based on transport equation, Hamilton-Jacobi equation and Fokker-Planck equation. The discretization of these
PDEs on point cloud is also discussed.
keywords: Deep residual network; control problem; manifold learning; point cloud; transport equation; Hamilton-Jacobi equation
1
Deep Residual Network
Deep convolution neural networks have achieved great successes in image classification.
Recently, an approach of deep residual learning is proposed to tackle the degradation in the
classical deep neural network [7, 8]. The deep residual network can be realized by adding
shortcut connections in the classical CNN. A building block is shown in Fig. 1. Formally,
a building block is defined as:
y = F (x, {Wi }) + x.
Here x and y are the input and output vectors of the layers. The function F (x, {Wi })
represents the residual mapping to be learned. In Fig. 1, F = W2 · σ(W1 · σ(x)) in which
σ = ReLU ◦ BN denotes composition of ReLU and Batch-Normalization.
∗
Department of Mathematics, Hong Kong University of Science & Technology, Hong Kong. Email:
[email protected].
†
Yau Mathematical Sciences Center, Tsinghua University, Beijing, China, 100084.
Email:
[email protected].
1
xl
BN
ReLU
Conv.
BN
ReLU
Conv.
+
xl+1
Figure 1: Building block of residual learning [8].
2
Transport Equation and ResNet
Consider the the terminal value problem of linear transport equation in Rd :
∂u − v(x, t) · ∇u = 0, x ∈ Rd , t ≥ 0,
∂t
u(x, 1) = f (x), x ∈ Rd .
(1)
where v(x, t) is a given velocity field, f is the composition of the output function and the
fully connected layer. If we use the softmax activation function,
f (x) = softmax(WF C · x),
(2)
where WF C is the weight in the fully connected layer, softmax function is given by
exp(xi )
.
softmax(x)i = P
j exp(xj )
which models posterior probability of the instance belonging to each class.
It is well-known that the solution at t = 0 can be approximately solved along the
characteristics:
dx(t; x)
= v (x(t; x), t) , x(0; x) = x.
(3)
dt
We know that along the characteristics, u is a constant:
u(x, 0) = u (x(1; x), T ) = f (x(1; x)).
Let {tk }L
k=0 with t0 = 0 and tL = 1 be a partition of [0, 1]. The characteristic of the
transport equation (3) can be solved by using simple forward Euler discretization from
x0 (x) = x:
Xk+1 (x) = Xk (x) + ∆tv(Xk (x), tk ),
(4)
2
where ∆t is the time step. If we choose the velocity field such that
∆t v(x, t) = W (2) (t) · σ W (1) (t) · σ(x) ,
(5)
where W (1) (t) and W (2) (t) corresponds to the ’weight’ layers in the residual block, σ =
ReLU ◦ BN, one step in the forward Euler discretization (4) actually recovers one layer in
the deep ResNets, Fig. 1. Then the numerical solution of the transport equation (1) at
t = 0 is given by
u(x, 0) = f (XL (x))
(6)
This is exactly the output we get from the ResNets.
If x is a point from the training set, we already have a labeled value g(x) on it. Then
we want to match the output value given in (6) and g(x) to train the parameters in the
velocity filed (5) and the terminal value.
Based on above analysis, we see that the training process of ResNet can be formulated
as an control problem of a transport equation in Rd .
∂u
d
∂t (x, t) − v(x, t) · ∇u(x, t) = 0, x ∈ R , t ≥ 0,
(7)
u(x, 1) = f (x), x ∈ Rd ,
u(xi , 0) = g(xi ), xi ∈ T.
where T denotes the training set. g(xi ) is the labeled value at sample xi . Function u may
be scalar or vector value in different applications.
So far, we just formulate ResNet as a control problem of transport equation. This model
will inspire us to get new models by replacing different component in the control problem.
3
Modified PDE model
From the PDE point of view, ResNet consists of five components:
• PDE: transport equation;
• Numerical method: characteristic+forward Euler;
• Velocity filed model: v(x, t) = W (2) (t) · σ W (1) (t) · σ(x) ;
• Terminal value: f (x) = softmax(WF C · x);
• Initial value: label on the training set g(xi ), xi ∈ T .
In five components listed above, the last one, ”initial value”, is given by the data and
we have no other choice. All other four components, we can consider to replace them by
other options. Recently, there are many works in replacing forward Euler by other ODE
solver. Forward Euler is the simplest ODE solver. By replacing it to other solver, we
3
may get different network. In some sense, in DenseNet [9], forward Euler is replaced by
some multi-step scheme. There are many numerical schemes to solve the ODE (3). Some
numerical scheme may be too complicated to be used in DNN. Constructing good ODE
solver in DNN is an interesting problem worth to exploit in our future work.
3.1
Terminal Value
From the control problem point of view, the softmax output function is not a good choice
for terminal vaule, since they are pre-determined, may be far from the real value. Semisupevised learning (SSL) seems provides a good way to get terminal value instead of the
pre-determined softmax function as shown in Fig. 2(b).
DNN
DNN
DNN
X(x; Θ)
X(x; Θ)
X(x; Θ)
Softmax
SSL
WNLL
Output
Output
Output
(a)
(b)
(c)
Figure 2: (a): standard DNN; (b): DNN with semi-supervised learning; (c): DNN with last
layer replaced by WNLL.
Recently, we try to use weighted nonlocal Laplacian [11] to replace softmax and the
results is pretty encouraging [12].
3.2
Velocity Field Model
In transport equation, we need to model a high dimensional velocity filed. In general, it is
very difficult to compute a high dimensional vector field. In the application associated to
images, the successes of CNN and ResNet have proved that the velocity filed model based
on convolutional operators, (8), is effective and powerful.
v(x, t) = W (2) (t) · σ W (1) (t) · σ(x) .
(8)
However, this is not the only way to model the velocity field. Moreover, for the applications
in which convolutional operator makes no sense, we have to propose alternative velocity
model. Here, we propose a model based on Hamilton-Jacobi equation to reduce the degree
of freedom in the velocity field.
Notice that even though the velocity v(x, t) is a high dimensional vector field, in the
tranport equation, only the component along ∇u is useful. Based on this observation, one
4
idea is only model the component along ∇u by introducing v̄(x, t) = v(x, t) · n(x, t) and
∇M u
n(x, t) = |∇
. Then the transport equation in (7) becomes a Hamilton-Jacobi equation.
M u|
We get a control problem of Hamilton-Jacobi equation.
∂u
d
∂t − v̄(x, t) · |∇u| = 0, x ∈ R , t ≥ 0,
(9)
u(x, 0) = f (x), x ∈ M,
u(xi , 1) = g(xi ), xi ∈ T.
In (7), the velocity field, v(x, t), is a d-dimensional vector field. Meanwhile, in the
Hamilton-Jacobi model, we only need to model a scalar function v̄(x, t). The number of
parameters can be reduced tremendously. To model a scalar function is much easier than
model a high dimensional vector field. There are already many ways to approximate v̄(x, t).
1. The simplest way is to model v̄(x, t) as a linear function with respect to x, i.e.
v̄(x, t) = w(t) · x + b(t),
w(t) ∈ Rd , b(t) ∈ R.
(10)
Although v̄(x, t) is a linear function to x, the whole model (9) is not a linear model
since Hamilton-Jacobi equation is a nonlinear equation.
2. Neural network is consider to be an effective way to approximate scalar function in
high dimensioanl space. One option is a simple MLP with one hidden layer as shown
in Fig. 3.
Weight
Weight
x ∈ Rd
d×m
ReLU
m×1
v̄(x) ∈ R
Figure 3: MLP model with one hidden layer for v̄(x, t) in (9).
There are many other neural networks in the literature to approximate scalar function
in high dimensioanl space.
3. Radial basis function is another way to approximate v̄(x, t). Radial functions centered
at each sample point are used as basis function to approximate v̄(x, t), i.e.
!
X
|x − xj |2
v̄(x, t) =
cj (t)R
(11)
σj2
xj ∈P
where cj (t), j = 1, · · · , |P | are coefficients of the basis function. One often used basis
function is Gaussian function, R(r) = exp(−r).
Another difficulty in solving model (9) is efficient numerical solver of Hamilton-Jacobi
equation in high dimensional space. Hamilton-Jacobi equation is a nonlinear equation
which is more difficult to solve than linear transport equation in (7). Recently, fast solver
of Hamilton-Jacobi equation in high dimensional space attracts lots of attentions and many
efficient methods have been developed [1, 2, 3, 4, 5, 6].
5
3.3
PDEs on point cloud
Another thing we can consider to change is the characteristic method in solving the transport
equation. There are many other numerical method for transport equation based on Eulerian
grid. However, in these methods, we need to discretize the whole space Rd by Eulerian grid,
which is impossible when d is high. So in high dimensional space Rd , characteristic method
seems to be the only practical numerical method to solve the transport equation. On the
other hand, we only need to solve the transport equation in the dataset instead of the whole
Rd space. Usually, we can assume that the data set sample a low dimensional manifold
M ⊂ Rd , then the PDE can be confined on this manifold. The manifold is sampled by the
point cloud consists of the data set including the training set and the test set. Then the
numerical methods in point cloud can be used to solve the PDE.
Transport equation In the manifold, the transport equation model (7) can be rewritten
as follows:
∂u
d
∂t − v(x, t) · ∇M u = 0, x ∈ M ⊂ R , t ≥ 0,
(12)
u(x, 0) = f (x), x ∈ M,
u(xi , 1) = g(xi ), xi ∈ T.
∇M denotes the gradient on manifold M. Let X : V ⊂ Rm → M ⊂ Rd be a local
parametrization of M and θ ∈ V . For any differentiable function u : M → R, let U (θ) =
f (X(θ)), define
Dk u(X(θ)) =
m
X
g ij (θ)
i,j=1
(g ij )i,j=1,··· ,m
where
is defined by
=
G−1
∂U
∂Xk
(θ)
(θ),
∂θi
∂θj
k = 1, · · · , d.
(13)
and G(θ) = (gij )i,j=1,··· ,m is the first fundamental form which
gij (θ) =
d
X
∂Xk
k=1
∂θi
(θ)
∂Xk
(θ),
∂θj
i, j = 1, · · · , m.
(14)
∇M is defined as
∇M u = (D1 u, D2 u, · · · , Dd u).
(15)
In (12), the manifold M is sampled by the data set P and P is a collection of unstructured high dimensional points. Unlike the classical numerical methods which solve PDE on
regular grids (or meshes), in this case, we need to discretize PDE on unstructured high dimensional point cloud P . To handle this kind of problems, recently, point integral method
(PIM) was developed to solve PDE on point cloud. In PIM, gradient on point cloud is
approximate by an integral formula [10].
Z
1
(u(x) − u(y)) (xk − y k )Rδ (x, y)dy.
(16)
Dk u(x) =
wδ (x) M
6
where wδ (x) =
R
M Rδ (x, y)dy,
Rδ (x, y) = R
kx − yk2
δ2
.
(17)
The kernel function R(r) : R+ → R+ is assumed to be a C 2 function with compact support.
Corresponding discretization is
Dk u(x) =
1 X
(u(x) − u(y)) (xk − y k )Rδ (x, y)V (y).
w̄δ (x)
(18)
y∈P
P
w̄δ (x) = y∈P Rδ (x, y)V (y) and V (y) is the volume weight of y depends on the distribution of the point cloud in the manifold.
Hamilton-Jacobi equation We can also confine the Hamilton-Jacobi equation in (9) in
the manifold.
∂u
d
∂t − v̄(x, t) · |∇M u| = 0, x ∈ M ⊂ R , t ≥ 0,
(19)
u(x, 0) = f (x), x ∈ M,
u(xi , 1) = g(xi ), xi ∈ T.
On the point cloud, one possible choice to discretize |∇M u| is
|∇M u(x)| =
Z
M
2
1/2
w(x, y)(u(x) − u(y)) dy
.
(20)
v̄(x, t) can be modeled in the way discussed in the previous section.
PDEs with dissipation We can also consider to add dissipation to stabilize the PDEs.
∂u
− v̄(x, t) · |∇M u| = ∆M u, x ∈ M ⊂ Rd , t ≥ 0,
∂t
u(x, 0) = f (x), x ∈ M,
(21)
u(xi , t) = g(xi ), xi ∈ S,
u(xi , 1) = g(xi ), xi ∈ T \S,
∂u
(x, t) = 0,
x ∈ ∂M.
∂n
In the model with dissipation, we need to add constraints, u(xi , t) = g(xi ), xi ∈ S ⊂ T
in a subset S of the training set T . Otherwise, the solution will be too smooth to fit the
data due to the viscosity. The choice of S is an issue. The simplest way is to choose S at
random. In some sense, the viscous term maintains the regularity of the solution and the
convection term is used to fit the data.
On the point cloud, the Laplace-Beltrami operator along with the constraints u(xi , t) =
g(xi ), xi ∈ S ⊂ T can be discretized by the weighted nonlocal Laplacian [11].
7
4
Discussion
In this paper, we establish the connection between the deep residual network (ResNet)
and the transport equation. ResNet can be formulated as solving a control problem of
transport equation along the characteristics. Based on this observation, we propose several
PDE models on the manifold sampled by the data set. We consider the control problem of
transport equation, Hamilton-Jacobi equation and viscous Hamilton-Jacobi equation.
This is a very preliminary discussion on the relation between deep learning and PDEs.
There are many important issues remaining unresolved including the model of the velocity
field, numerical solver of the control problem and so on. This is just the first step in
exploring the relation between deep learning and control problems of PDEs.
References
[1] Y. T. Chow, J. Darbon, S. Osher, and W. Yin. Algorithm for overcoming the curse
of dimensionality for certain non-convex hamilton-jacobi equations, projections and
differential games. UCLA CAM-Report 16-27, 2016.
[2] Y. T. Chow, J. Darbon, S. Osher, and W. Yin. Algorithm for overcoming the curse of
dimensionality for time-dependent non-convex hamilton-jacobi equations arising from
optimal control and differential games problems. UCLA CAM-Report 16-62, 2016.
[3] Y. T. Chow, J. Darbon, S. Osher, and W. Yin. Algorithm for overcoming the curse
of dimensionality for state-dependent hamilton-jacobi equations. UCLA CAM-Report
17-16, 2017.
[4] J. Darbon and S. Osher. Algorithms for overcoming the curse of dimensionality for
certain hamilton-jacobi equations arising in control theory and elsewhere. UCLA CAMReport 15-50, 2015.
[5] J. Darbon and S. Osher. Splitting enables overcoming the curse of dimensionality.
UCLA CAM-Report 15-69, 2015.
[6] J. Han, A. Jentzen, and W. E. Overcoming the curse of dimensionality: Solving highdimensional partial differential equations using deep learning. ArXiv:1707.02568, July
2017.
[7] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages
770–778, 2016.
[8] K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks.
arXiv:1603.05027, 2016.
8
[9] G. Huang, Z. Liu, L. V. Maaten, and K. Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261–2269, 2017.
[10] Z. Li, Z. Shi, and J. Sun. Point integral method for elliptic equations with variable
coefficients on point cloud. mathscidoc:1708.25001, 2017.
[11] Z. Shi, S. Osher, and W. Zhu. Weighted nonlocal laplacian on interpolation from sparse
data. Journal of Scientific Computing, 73:1164–1177, 2017.
[12] B. Wang, X. Luo, Z. Li, W. Zhu, Z. Shi, and S. Osher. Deep neural networks with
data dependent implicit activation function. in preparation, 2018.
9
| 7cs.IT
|
1
Domain-Specific Face Synthesis for Video Face
Recognition from a Single Sample Per Person
arXiv:1801.01974v1 [cs.CV] 6 Jan 2018
Fania Mokhayeri, Student Member, IEEE, Eric Granger, Member, IEEE,
and Guillaume-Alexandre Bilodeau, Member, IEEE
Abstract—In video surveillance, face recognition (FR) systems
are employed to detect individuals of interest appearing over
a distributed network of cameras. The performance of still-tovideo FR systems can decline significantly because faces captured
in the unconstrained operational domain (OD) over multiple
video cameras have a different underlying data distribution
compared to faces captured under controlled conditions in the
enrollment domain (ED) with a still camera. This is particularly
true when individuals are enrolled to the system using a single
reference still. To improve the robustness of these systems, it
is possible to augment the reference set by generating synthetic
faces based on the original still. However, without knowledge of
the OD, many synthetic images must be generated to account
for all possible capture conditions. FR systems may therefore
require complex implementations and yield lower accuracy when
training on many less relevant images. This paper introduces an
algorithm for domain-specific face synthesis (DSFS) that exploits
the representative intra-class variation information available
from the OD. Prior to operation (during camera calibration), a
compact set of faces from unknown persons appearing in the OD
is selected through affinity propagation clustering in the captured
condition space (defined by pose and illumination estimation).
The domain-specific variations of these face images are then
projected onto the reference still of each individual by integrating
an image-based face relighting technique inside the 3D morphable
model framework. A compact set of synthetic faces is generated
that resemble individuals of interest under capture conditions
relevant to the OD. In a particular implementation based on
sparse representation classification, the synthetic faces generated
with the DSFS are employed to form a cross-domain dictionary
that accounts for structured sparsity where the dictionary blocks
combine the original and synthetic faces of each individual.
Experimental results obtained with videos from the Chokepoint
and COX-S2V datasets reveal that augmenting the reference
gallery set of still-to-video FR systems using the proposed DSFS
approach can provide a significantly higher level of accuracy
compared to state-of-the-art approaches, with only a moderate
increase in its computational complexity.
Index Terms—Face Recognition, Single Sample Per Person,
Face Synthesis, 3D Face Reconstruction, Illumination Transferring, Sparse Representation-Based Classification, Video Surveillance.
I. I NTRODUCTION
Still-to-video (S2V) face recognition (FR) is an important
function in several video surveillance applications, particularly
for watch-list screening. Given one or more reference still
images of a target individual of interest, S2V FR systems
Fania Mokhayeri and Eric Granger are with the Laboratory for Imagery, Vision and
Artificial Intelligence (LIVIA), École de technologie supérieure, Université du Québec,
Montréal, Canada, (e-mail: [email protected], [email protected]).
Guillaume-Alexandre Bilodeau is with the Image and Video Processing Laboratory
(LITIV), Ṕolytechnique Montréal, Montréal, Canada, (e-mail: [email protected]).
seeks to accurately detect their presence in videos captured
over multiple distributed surveillance cameras [1].
Despite the recent progress in computer vision and machine
learning, designing a robust system for S2V FR remains a
challenging problem in real-world surveillance applications.
One key issue is the visual domain shift between faces from
the enrollment domain (ED), where reference still images
are typically captured under controlled conditions, and those
from the operational domain (OD), where video frames are
captured under uncontrolled conditions with variations in
pose, illumination, blurriness, etc. The appearance of faces
captured in videos corresponds to multiple non-stationary data
distributions that can differ considerably from faces captured
during enrollment [2]. Another key issue is the limited number
of reference stills that are available per target individual to
design facial models. Although still faces from the cohort
or other non-target persons, and trajectories of video frames
from unknown individuals are typically available. In many
surveillance applications (e.g., watch-list screening), only a
single reference still per person is available for design, which
corresponds to the so-called Single Sample Per Person (SSPP)
problem. The performance of S2V FR systems can decline
significantly due to the limited information available to represent the intra-class variations seen in video frames. Many
discriminant subspaces and manifold learning algorithms cannot be directly employed with a SSPP problem. It is also
difficult to apply representation-based FR methods such as
sparse representation-based classification (SRC) [3], [4].
Different techniques for SSPP problems have been proposed to improve the robustness of FR systems, such as
using multiple face representations [2], face frontalization [5],
generating synthetic faces from the original reference stills
[6], [7], and incorporating generic auxiliary set1 [8], [9]. This
paper will focus on methods to augmenting the reference
gallery set either through synthetic set generated based on
the original reference still, or through intra-class variation
information transferred from a generic set. A challenge with
strategies for augmenting the reference gallery set is selecting
a sufficient number of synthetic or generic faces to cover intraclass variations in the OD. Many synthetic faces or generic
auxiliary faces may be generated or collected, respectively, to
account for all possible capture conditions. In this case, FR
systems would therefore require complex implementations and
1 A generic set is defined as an auxiliary set that contains multiple video
frames per person from other unknown individuals. It provides an abundance
of information on intra-class variations of the capture conditions in an OD.
2
1
...
3
2
m
...
Characterizing Capture
Condition
Unlabeled ROIs from Video Trajectories
(Operation Domain)
Pose and Illumination
Representative
1
2
3
4
2
3
...
q
1
+
...
...
2
+
...
3
+
...
4
+
...
...
n
q
...
...
+
Reference Faces
...
...
n
...
Labeled Reference Still ROIs
(Enrollment Domain)
...
Generating Synthetic Faces
based on OD knowledge
...
...
...
...
1
Extended Reference Set
1
2
3
4
Synthetic Faces
Faces
Synthetic
Figure 1: Overview of the proposed DSFS algorithm to augment the reference gallery set. We assume that the gallery set initially contains
only one reference still ROI per individual of interest.
may yield lower accuracy when training on many facial images
that provide less relevant information for FR in the OD.
In this paper, a new approach is proposed that exploits the
discriminant information of the generic set for the face synthesis process. The new algorithm called domain-specific face
synthesis (DSFS) maps representative variation information
from the generic set in the OD to the original facial regions
of interest (ROIs) isolated in reference stills. In this way,
a compact set of synthetic faces is generated that represent
still reference ROIs and probe video ROIs under a common
capture condition. As depicted in Fig. 1, the DSFS technique
involves two main steps: (1) characterizing capture condition
information from the OD, (2) generating synthetic face ROIs
based on the information obtained in the first step. Prior to
operation (during some camera calibration process), a generic
set of facial ROIs is collected from video captured in the OD.
A compact and representative subset of ROIs is selected by
clustering this generic set in a capture condition space defined
by pose, illumination, blur and contrast estimation measures.
The 3D model of each reference still ROI is reconstructed
via a 3D morphable model and rendered based on pose
representatives. Finally, the illumination-dependent layers of
the lighting representatives are extracted and projected on the
rendered reference ROIs with the same pose. In this manner,
domain-specific variations are effectively transferred onto the
reference still ROIs.
In a particular implementation for S2V FR, the original and
synthetic ROIs are employed to design a structural dictionary
with powerful variation representation ability for SRC. The
dictionary blocks represent intra-class variations computed
from either the reference faces themselves or the synthetic
faces [10]. The cooperation of SRC with the proposed DSFS
improves the robustness of SRC for video-based FR in a
SSPP scenario to domain variations. In order to validate the
performance of the proposed DSFS algorithm for S2V FR with
a SSPP, this SRC implementation is evaluated and compared
on two public face databases.
The main advantage of the proposed approach is the ability
to provide a compact set that can accurately represent the
original reference face with relevant of intra-class variations in
pose, illumination, motion blur, etc., corresponding to capture
condition in the OD. For instance, in the context of SRC
implementations, this set can prevent over-fitting and refines
more informative classes during the sparse coding process, and
accordingly does not require a traditional dictionary learning
process.
The rest of the paper is organized as follows. Section
II provides an overview of related works for FR with a
SSPP. Section III describes the proposed face synthesizing
algorithm. Section IV presents a particular implementation of
the DSFS for S2V FR system. In Section V, the experimental
methodology (dataset, protocol, and performance metrics) for
validation of FR systems is described and the experimental
results is presented. Finally, Section VII concludes the paper
and discusses some future research directions.
II. R ELATED W ORK – S2V FACE R ECOGNITION FROM A
S INGLE S TILL
Several techniques have been proposed in the literature to
improve the robustness of S2V FR systems designed using a
SSPP. They can be categorized into techniques for (1) multiple
face representation, (2) generic learning, and (3) generation of
synthetic faces. An overview of the techniques is presented as
bellow.
A. Multiple Face Representation
One effective approach to address the SSPP problem in FR
is to extract discriminant features from face images. Bashbaghi
et al. [2], [11] proposed a robust S2V FR system based on diverse face representations. They applied multiple appearanceinvariant feature extraction techniques to patches isolated in
the reference still images in order to produce multiple face
representations and generate a pool of diverse exemplarSVMs. This pool provides robustness to common nuisance
factors encountered in surveillance applications. Jiwen et al.
[12] proposed a discriminative multi-manifold analysis method
by learning discriminative features from image patches. In this
technique, the patches of each individual are considered to
3
form a manifold for each sample per person and a projection
matrix is learned by maximizing the manifold margin of different persons. In [13] a face image is processed by several posespecific deep convolutional neural network (CNN) models to
generate multiple pose-specific features. The multiple face
representation techniques are, however, able to compensate
only the small variations and consequently are not effective to
tackle with variations in practical applications (e.g., extreme
illumination, pose and expression variations).
B. Generic Learning
An early finding to compensate visual domain shift in FR
systems is to employ a generic set to enrich the diversity of
the reference gallery set that is the so-called generic learning
concept [14]. Generic learning has been widely discussed
by many researchers [12], [15], [16]. Su et al. proposed
an adaptive generic learning method for FR which utilized
external data to estimate the within-class scatter matrix for
each individual and applies this information to the reference
set [14]. In recent years, integration of sparse representationbased classification (SRC) with generic learning for FR has
attracted significant attention. Deng et al. [8] added generic
leaning into the SRC framework and proposed the extended
SRC (ESRC), which provide additional information from other
face datasets to construct an intra-class variation dictionary to
represent the changes between the training and probe images.
With further research, Deng et al. [9] proposed a generic
learning method which projects the generic samples data to
null space in order to reduce the negative influence of the naïve
variation information. Yang et al. [17] introduced a sparse
variation dictionary learning (SVDL) technique by taking the
relationship between the reference set and the external generic
set into account and obtained a projection by learning from
both generic and reference set. In [18], the intra-class variation
information from the OD is integrated with the reference
set through domain adaptation to enhance the facial models.
Authors in [19] proposed a robust auxiliary dictionary learning
(RADL) technique that extracts representative information
from generic dataset via dictionary learning without assuming prior knowledge of occlusion in probe images. Zhu et
al. proposed a local generic representation-based framework
(LGR) for FR with SSPP [20]. It builds a gallery dictionary
by extracting the neighboring patches from the gallery dataset,
while an intra-class variation dictionary is constructed by using
an external generic training dataset to predict the intra-class
variations. Authors in [21] proposed a robust S2V FR using
a multi-classifier system in which each classifier is trained
by a reference face still versus many lower-quality faces of
non-target individuals captured in videos. In this system, the
auxiliary set collected from videos of unknown people in
the OD are employed to select discriminant feature sets and
ensemble fusion functions. In [22], a supervised autoencoder
network is proposed for S2V FR system to generate canonical
face representations from unknown video ROIs that are robust
to appearance variations commonly found in the operational
video scene. Despite the significant improvements reported
with generic learning, several critical issues remain to be
addressed. The generic intra-class variation may not be similar
to that of gallery individuals, so the extraction of discriminative information from the generic set may not be guaranteed.
Moreover, the large number of images collected from external
data may contain redundant information which could lead
to complex implementations and degrade the capability in
covering intra-class variations.
C. Synthetic Face Generation
Augmenting the reference gallery set synthetically is another strategy to compensate the appearance variations in FR
with SSPP. Shao et al. in [23] presented a SRC-based FR
algorithm that extends the dictionary using a set of synthetic
faces generated by calculating the image difference of a pair
of faces. Authors in [24] augmented the reference gallery set
by generating a set of synthetic face images under cameraspecific lighting conditions to design a robust S2V FR system
under surveillance conditions. Blanz and Vetter [25] proposed
3D Morphable Model (3DMM) to reconstruct a 3D face from
a single 2D face image and accordingly synthesize new face
images. Authors in [26] employed a CNN to regress 3DMM
shape and texture parameters directly from an input image
without an optimization process which renders the face and
compares it to the image. Zhang et al. proposed a 3D Spherical Harmonic Basis Morphable Model (SHBMM) that is an
integration of spherical harmonics into the 3DMM framework
[27]. Richardson et al. presented a face reconstruction technique from a single image by introducing an end-to-end CNN
framework which derives the shape in a coarse-to-fine fashion
[28]. In [29] a model-based deep convolutional autoencoder
for 3D face reconstructing from a single is proposed, where a
convolutional encoder network is combined with an expertdesigned generative model that serves as a decoder. Apart
from 3D reconstruction techniques, some 2D-based techniques
generate synthetic images under various illumination conditions by transferring the illumination of target images to the
reference face images. Authors in [30] proposed an imagebased framework to relight the face through adaptive layer
decomposition. Although synthetic images can improve the
robustness of FR systems designed with a SSPP, they may
not be covering the range intra-class variations in practical
scenarios because of redundancy in the learned discriminative
subspace, since they are highly correlated with the original
face images. Many synthetic images should be generated to
account for all possible capture conditions in ODs. Without the
selection of representative face images from both the reference
gallery and external data, generating the synthetic faces may
require complex implementations and yield lower accuracy
when training on many less relevant images.
To overcome the challenges discussed above, this paper
presents a framework that exploits both face synthesis and
generic learning. The technique proposed in Section III generates a compact set of synthetic facial images per individual
of interest that corresponds to relevant OD capture conditions,
by mapping the intra-class variations from a representative set
of video ROIs selected from the OD into the original reference
still ROIs.
4
III. D OMAIN -S PECIFIC FACE S YNTHESIS
This paper focuses on augmenting reference face set to
cover the intra-class variations of individual appearing in ODs
with a compact set of synthetic ROIs. A new Domain-Specific
Face Synthesis (DSFS) technique is proposed that employs
knowledge of the OD to generate a compact set of synthetic
ROIs for the design of FR systems.
Prior to operation, e.g., during a camera calibration process,
DSFS selects video ROIs with representative pose angles and
illumination conditions from facial trajectories of unknown
persons captured in the OD. These video ROIs are selected
via clustering facial trajectories in the captured condition space
defined by pose and illumination conditions. Next, the DSFS
exploits a 3D shape reconstruction method and an image-based
illumination transferring technique to generate synthetic ROIs
under representative pose angles and illumination conditions
from the reference still ROIs. To do so, the 3D models of the
reference still ROIs are reconstructed and rendered w.r.t. the
representative pose angles. The illumination-dependent layers
of the representative illumination conditions are then extracted
and projected onto rendered images with the same view by
applying a morphing between the layers. In other words,
illumination-dependent layers of video ROIs from the OD are
replaced with that of the still reference ROI from the ED. Fig.
2 shows the pipeline of the DSFS technique.
A. Characterizing the Capture Conditions
An important concern for the reference set augmentation
is the selection of representative pose angles and illumination
conditions to represent relevant capture conditions in the OD.
As mentioned, adding a large number of potentially redundant
images to the reference set can significantly increase the time
and memory complexity, and may degrade the recognition
performance due to over-fitting.
With the DSFS technique, the representative pose angles
and illumination conditions to cover relevant intra-class variations is approximated by characterizing the capture conditions from a large generic set of video ROIs. This set is
formed with multiple ROIs isolated in several facial trajectories
of unknown persons captured in the OD. Let define
R = ri ∈ Rd×d |i = 1, . . . , n as the set of ROIs of still
reference individuals, and G = gi ∈ Rd×d |i = 1, . . . , m as
the video ROIs in the generic set, where n denotes the number
of individuals in the reference gallery set, and m total number
of ROIs in the generic set, respectively.
In the proposed technique (see Fig. 3), an estimation of
luminance, contrast and pose are measured for each video
ROIs in the generic set gi . Next, a two-step clustering process
is applied on video ROIs in the measurement space defined
by pose, luminance and contrast. The first step is applied on
all ROIs in the 3D metric space defined by pose (tilt, yaw and
roll), while the second step is applied on ROIs of each pose
cluster in the 2D space defined by luminance and contrast
metrics. The prototype of each cluster is considered as an
exemplar. The generic variational information obtained during
this step is then transferred to the reference still ROIs during
the face synthesizing step (see Section III B). Although many
algorithms are also suitable to implement DSFS, the following
subsections describe DSFS with specific algorithms.
1) Estimation of Head Pose:
The estimate of head pose for the ith video ROI (gi ) in the
generic set is defined as pi = (θipitch , θiyaw , θroll )i . Euler
angles θipitch , θiyaw , and θiroll are used to represent pitch, yaw
and roll rotation around X axis, Y axis, and Z axis of the
global coordinate system, respectively. In order to estimate
the head pose, the discriminative response map fitting method
[31] is employed. It is the current state-of-the-art method in
terms of fitting accuracy and efficiency suitable for handling
occlusions and changing illumination conditions.
2) Luminance-Contrast Distortion:
Luminance and contrast distortion measures estimate the distortion between a video ROI and the corresponding reference
still ROI. Components of the structural similarity index measure presented [32] are employed to measure the proximity of
the average luminance and contrast locally by utilizing sliding
window. The global luminance distortion in image quality
(GLQ) factor between ri and gj is calculated by sliding a
window of B×B pixels from the top-left corner to the bottomright corner of the image, for a total of M sliding steps:
li,j =
M
1 X 2.µk (ri ).µk (gj ) + Cl
,
2
2
M
µ (r ) + µk (gj ) + Cl
k=1 k i
(1)
where k is the sliding step and µk denotes mean values of
the k th image window. Cl is a constant value defined as Cl =
(Kl L)2 where L is the dynamic range of the pixel values and
Kl 1. Similarly, the contrast distortion between ri and gi
is estimated using global contrast distortion in image quality
(GCQ) factor defined as:
ci,j =
M
1 X 2.σk (ri ).σk (gj ) + Cc
,
2
2
M
σ (r ) + σk (gj ) + Cc
k=1 k i
(2)
where σk denotes the standard deviation of the k th image
window. Cc is a constant value defined as Cc = (Kc L)2 where
L is the dynamic range of the pixel values and Kc 1.
3) Representative Selection:
Affinity propagation (AP) [33] is applied to cluster video ROIs
from the generic set defined in the normalized space defined
by pj = (θipitch , θiyaw , θiroll ) and ui = (li , ci ) measures. This
clustering algorithm aims to maximize the net similarity (average distortion between ROIs and pose angles), and produce
a set exemplars. Two types of messages: responsibility and
availability are exchanged between data points until a highquality set of exemplars and corresponding clusters emerges.
AP is a suitable clustering technique for DSFS because: (1)
it can automatically determine the number of clusters based
on the data distribution, and (2) it produced exemplars that
correspond to actual ROIs. Indeed, cluster centroids produced
by many prototype-based clustering methods are not necessarily actual ROIs with a real-world interpretation. Given
that clustering samples simultaneously in terms of both pj
and ui may favor certain common pose angles, a two-step
clustering algorithm is proposed to preserves diversity in pose
angles and illumination effects. In the first step, clustering is
5
Reference ROI
Generic Set
ri
gi
Shading-Albedo
Decomposition
Characterization of
Capture Conditions
Iuij
Mi
pi
Shading-Albedo
Decomposition
Rendering
Lij
ri
Si
Texture Mapping onto
3D Face Shape
Vi
Morphing
Synthetic face images under
representative pose angles
Illumination-dependant layers of
Lighting representative samples
ai
3D Shape Recovery
Statistical Deformable
Model Trained from the
3D Shape Vectors
Synthetic Face Images under
Representative Pose & Illumination
Domain-specific
Synthetic Set
Figure 2: Block diagram of the DSFS technique applied to a reference still ROI.
gi
Estimation of
Pose Angles
Roll
Yaw
pi
AP Clustering
Pitch
Generic Set
Estimation of
Luminance Distortion
Pose Clusters
li
gi
AP Clustering
inside pose clusters
Estimation of
Contrast Distortion
ci
Pose and Lighting
Exemplars
Figure 3: Pipeline for characterizing capture conditions of video
ROIs in the operational domain.
performed on the pose angle vector, and then the population
of each pose cluster is clustered according to GLQ and
GCQ metrics to find the representative luminance and contrast
samples. Representative luminance and contrast samples –
called lighting exemplar – are found along with representative
pose angles – called "pose exemplar" (Fig. 4).
P4
Pj
×× ××× × × ×
×××××××××××××××××××× ××××××××
××××××××××××
××××××××××
× P3 ×××××××××××××× ×××××××××××
××××××× ×××××××××××××××××××× ×××××××
×
×××××××××××××× ××××××P×1××××××
××××××
× × × × × P2
Pitch
×
××××××××××××××××××××××
××××× ×××× ×
Yaw
×
Illumination
Roll
UjNj
Uj2
++ +++ ++++++++
++ + +
++
+++++++ +++ +++ +++++
+ ++ U 3+ +
Uj1
++ j +
++++ +++ ++++ + ++
++ ++ +++++ +++
++++++ +++
Contrast
Figure 4: An illustration of the AP clustering process.
The clustering algorithm inputs a set of pose similarities
sp (i, k) = − k pi − pk k2 indicating how well the sample pk
with index k is similar to the sample pi from the generic set.
The pose responsibility rp (i, k) is defined as the accumulated
evidence for how well-suited sample pk is to serve as the
exemplar for sample pi , taking into account other potential
exemplars for the sample pi . Evidence about whether each
pose candidate exemplar would be a good exemplar is obtained
from the application of the pose availability ap (i, k). The
availability reflects the accumulated evidence for how appropriate it would be for sample pi to choose sample pk as its
exemplar, taking into account the support from other samples
that sample pk should be an exemplar. The availabilities
are initialized to zero, and the pose responsibilities are then
computed iteratively using the rule of Eq.3. The availabilities
are updated in each iteration using Eq.4.
rp (i, k) = sp (i, k) − max
{ap (i, k 0 ) + sp (i, k 0 )} ,
0 0
k |k 6=k
(3)
n
o
X
ap (i, k) = min 0, rp (k, k) +
max{0, rp (i0 , k)} . (4)
i0 |i0 6∈{i,k}
For pi , the value of pk that maximizes ap (i, k) + rp (i, k)
either identifies sample pi as an exemplar if k = i, or
identifies the sample that is the exemplar for sample pi . The
message-passing procedure is terminated after a fixed number
of iterations when the local cost functions remain constant for
some number of iterations. At the end of pose clustering, K
pose clusters P = {P1 , P2 , . . . , Pj , . . . , PK } are determined,
where pi = [θipitch , θiyaw , θiroll ].
The second clustering is then applied for each pose cluster
in the li,j and ci,j measure space to find lighting exemplars. The first step computes illumination-contrast similarities
su (i, k) = − k (ui − uk ) k2 . The corresponding responsibility
and availability are obtained according to:
ru (i, k) = su (i, k) − max
{au (i, k 0 ) + su (i, k 0 )} ,
0 0
k |k 6=k
(5)
n
o
X
au (i, k) = min 0, ru (k, k) +
max{0, ru (i0 , k)} . (6)
i0 |i0 6∈{i,k}
The estimated rcl (i, k) and acl (i, k) are combined to mon-
6
itor the exemplar decisions and the algorithm is terminated
when these decisions do not change for several iterations.
At the end of the illumination-contrast clustering for each
pose cluster Pj , a number of Nj lighting clusters Pj =
{Uj1 , Uj2 , . . . , UjNj } are obtained. The central representative samples of illumination-contrast clusters in j th pose cluster
are considered as the pose and lighting exemplars for j th pose
as uij = (lji , cij ), 1 ≤ i ≤ Nj where l and c are illumination
and contrast of center of ith illumination-contrast cluster Uji
in the j th pose cluster Pj .
Larger clusters represent a greater number of generic samples, they should have more influence for the classification.
Therefore, a weight is assigned to each exemplar uij to
indicate its importance, approximated based on its cluster size,
Wij = nij /n, where nij is the number of samples in the
cluster Uij and n is the total number of generic samples. This
selection strengthens those classes that are more representative
in reconstructing a probe sample.
the shape vector (S) is defined as Eq.8 [25]. A principal
components analysis is performed to estimate the statistics of
the 3D shape of the faces.
S = S̄ +
mX
S −1
S = [X1 , Y1 , Z1 , X2 , . . . , Xns , Yns , Zns ]T .
ri (x, y) = Bi (x, y).Ti (x, y) = Li (x, y).Mi (x, y).Ti (x, y) ,
(7)
where B(x, y) = L(x, y).M(x, y) is a base layer, and L(x, y),
M(x, y) and T(x, y) are shading, material, and texture components at a pixel (x,y), respectively.
2) 3D Face Reconstruction:
3D face model of reference ROIs, ri , are reconstructed using
the 3D Morphable Model (3DMM) technique [25][35]. In this
study, a customized version of the 3DMM is employed in
which the texture fitting of the original 3DMM is replaced
with image mapping. By replacing the texture fitting in the
original 3DMM with 2D image mapping, an efficient method
is implemented for 3D face reconstruction from one frontal
face image. Basically, the shape model is defined as a convex
combination of shape vectors of a set of examples in which
(9)
Here, for each reference ROI, ri , we reconstruct the 3D shape.
B. Face Synthesis
1) Intrinsic Image Decomposition:
Each still reference image, ri , is decomposed and its materialdependent layer (albedo), Mi , is extracted based on the a
texture-aware image model defined in [34]. This image decomposition method explicitly models a separate texture layer
in addition to the shading layer and material layer in order
to avoid ambiguity caused by textures, and employ simple
constraints on shading and materials layers effectively. This
model is presented as follows:
(8)
where, the 3D shape is represented by the probability distribution of faces around the averages of shape S̄ are calculated
and the basis vectors S̃j , 1 ≤ j ≤ ms in Eq.8 where ms is
the number of the basis vectors.
Each vector S stores the reconstructed 3D shape in terms
of x, y, z-coordinates of all vertices {1, . . . , ns } of a highresolution 3D mesh as
Si = S̄ +
For generating synthetic ROIs based on the representative pose and lighting conditions, 3D models of reference
ROIs are reconstructed and their material-dependent layers
are extracted. In the rendering process, the extracted material
layers are employed as a texture of the 3D model. This
model is rendered w.r.t. the pose exemplars. Following this,
the illumination-dependent layers of the lighting exemplars
are extracted. Finally, the lighting layers are projected on the
rendered images with the same view by applying a morphing
between the layers. The following subsections describe the
steps proposed for the face synthesizing with DSFS.
αk .S̃k ,
k=1
mX
S −1
αji .S̃j ,
(10)
j=1
where αji ∈ [0, 1], 1 ≤ j ≤ ms are the shape parameters and
Si is the reconstructed shape of the ith reference still ROI
ri . The optimization algorithm presented in [25] is employed
to find optimal αji , 1 ≤ j ≤ ms , for each reference still
ROI ri . In the next step, the extracted material layers, Mi ,
are projected to the 3D geometry of the reference gallery set.
Given the 3D facial shape and texture, novel poses can be
rendered under various forms of the pose by adjusting the
parameters of a camera model. In the rendering procedure,
the 3D face is projected onto the image plane with Weak
Perspective Projection which is a linear approximation of the
full perspective projection.
mX
S −1
Vij = f ∗ Λ ∗ Rij ∗ (S̄ +
αji ) + ti2d ,
(11)
j=1
ij
where V is the 2D positions of model vertexes of the
j th reconstructed pose of ri , f is the scale
factor, Λ is the
1
0
0
orthographic projection matrix 0 1 0 , Rij is the rotation
matrix constructed from rotation angles pitch, yaw, roll and t2d
is the translation vector [36].
Since the 2D image is directly mapped to the 3D model,
no corresponding color information is available for some
vertices because they are occluded in the frontal face image.
Consequently, it is possible that there are still some blank areas
on the generated texture map. In order to correct these blank
space areas, a bilinear interpolation algorithm is utilized to fill
in areas of unknown texture using the known colors in the
vicinity [37].
3) Illumination Transferring:
d×d
For each pose exemplar, pj , a set of samples {Ijk
|1 ≤
u ∈R
k ≤ Nj } corresponding to the ujk , for k = 1, 2, ..., Nj ,
are selected as lighting exemplars. The illumination-dependent
layer of each Ijk
u are extracted using the same process described in section III-B1. Then, guided filter with the guidance
of the input illumination layer is applied to preserve the
7
structure of the input face. For each pose exemplar pj , Nj
the illumination layers, Ljk , for k = 1, 2, ..., Nj , are then
projected to the rendered reference, Vij . This is performed
by morphing between their filtered version according to the
following steps:
i) detect the landmark points of Ljk and Vij using active
shape model to locate corresponding feature points. The
landmark points of Ljk and Vij are denoted as ljk and
vij , respectively;
ii) define a triangular mesh over ljk and vij via the Delauij
nay triangulation technique and obtain djk
l and dv ;
jk
ij
iii) coordinate transformations between dl and dv with
affine projections on points;
iv) warp each triangle separately from the source to destination using mesh warping technique which moves
triangular patches to the newly established location to
align two ROIs;
v) cross-dissolve the triangulated layers considering
warped pixel locations.
IV. D OMAIN - INVARIANT S TILL - TO -V IDEO FACE
R ECOGNITION WITH DSFS
Several studies have revealed the limit robustness of S2V
FR system where the number of training images per class
is limited and the underlying distribution between the still
reference and probe video ROIs differ. In this section, a
particular S2V FR implementation is considered (see Fig. 5)
to assess the impact of using DSFS to generate synthetic ROIs
to address these limitations.
Design/Enrollment Phase
Compact set of synthetic ROIs
gi
DSFS
Reference Still ROIs
Synthetic Set
Reference Set
ri
Generic Set
Reference Still ROIs
Video ROIs
Synthetic ROIs
ri
ai
Face Detection
Design a Cross-Domain
Dictionary
Video Frames
Dc
Operational Phase
PK
In this way, a number q = j=1 Nj of synthetic ROIs are
generated for each reference still ROI ri . Therefore, the total
number of synthetic ROIs are qtotal = nq. The synthetic
set of ROIs for the ith reference still ROI are presented
2
by Ai = [ai1 , ai2 , . . . , aiq ] ∈ IRd ×q where aij is the j th
concatenated synthetic ROI for the ith reference still ROI.
The overall process of DSFS face generation technique is
formalized in Algorithm 1.
Probe Video
Frame
y
Face Detection
Probe Video
ROI
Structured-SRC
Sparse
Coefficients
Compute the
Reconstruction Error
Output
Identity
Stream of frames captured
by video cameras
Cross-Domain Dictionary
Probe ROI
×
Sparse coefficients associated
with reference+synthetic ROIs
Figure 5: Block diagram of the proposed domain-invariant SRCAlgorithm 1: The DSFS Algorithm
d×d |i = 1, . . . , n , and generic set
Input: Reference
set R = ri ∈ R
G = gi ∈ Rd×d |i = 1, . . . , m .
Estimate pose angles.
Calculate luminance and contrast distortion measures. // Eq.1,
Eq.2
AP clustering on pose space to obtain P = {P1 , P2 , . . . , PK }.
// SectionIII-A3
for j = 1 to K do
AP clustering on Illumination and contrast space for the Pj to
obtain {uji |1 ≤ i ≤ Nj }. // SectionIII-A3
end
for i = 1 to n do
Extract material-dependent layer of ri (Mi ).
// SectionIII-B1
Recover 3D face model of ri using 3DMM (Si ).
// SectionIII-B2
Map the texture of Mi to Si .
for j = 1 to K do
Render under pj pose to obtain Vij .
for each uji , i = 1 to Nj do
Extract illumination-dependent layers (Ljk ).
// SectionIII-B1
Apply guided filter on Ljk .
Morphing between the filtered Ljk and Vij to obtain
Ai . // SectionIII-B3
end
end
end
Output: all sets of synthetic face ROIs under representative pose and
illumination conditions.
2
Ai = [ai1 , ai2 , . . . , aiq ] ∈ IRd ×q , i = 1, 2, . . . , n .
based S2V FR system.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
An augmented dictionary is constructed by employing the
synthetic ROIs generated via DSFS technique, and classification is performed via a structured SRC approach. Since the
synthetic ROIs for each individual (including the synthetic
poses, illuminations, and etc.) form a block in this dictionary,
the SRC is considered as a structured sparse recovery problem.
The main steps of the proposed domain-invariant S2V FR with
dictionary augmentation are summarized as follows:
•
Step 1: Generation of Synthetic Facial ROIs
In the first step, q synthetic ROIs Ai = [ai1 , . . . , aiq ] ∈
2
Rd ×q are generated for each ri of the reference gallery
set using DSFS technique, where q is the number of
synthetic ROIs for each class.
•
Step 2: Augmentation of Dictionary
The synthetic ROIs generated through the DSFS technique are added to the reference dictionary to design a
cross-domain dictionary. let DR = [Ir 1 , Ir 2 , . . . , Ir n ] ∈
2
Rd ×n be the reference gallery dictionary, where Ir i is
the concatenated result of ri . The cross-domain dictionary
2
DC = [Ir 1 , A1 , . . . , Ir n , An ] ∈ Rd ×n(q+1) integrates
the original and synthetic ROIs in a linear model where
Aj is the j th set of synthetic ROIs added to the j th class.
Since q synthetic ROIs are added to each class, the total
8
The class label of the probe ROIs y is then determined
based on the block sparse reconstruction error as follows:
number of synthetic ROIs in the cross-domain dictionary
are nc = n(q + 1).
The presented dictionary design in this work enables
SRC to perform recognition with only one reference still
ROI and makes it robust to the visual domain shift.
•
label(y) = arg min k y − D[i]x[i] k2 .
i
In order to solve the SRC problem of equation 15,
the classical alternating direction method (ADM)
is considered [38]. ADM is an efficient first-order
algorithm with global convergence.
Step 3: Classification
Given a probe video ROI y, general SRC represents y
as a sparse linear combination of the codebook DC , and
derives the sparse coefficients of y by solving the `0 minimization problem as follows:
min kxk0
A`0 :
s.t. y = DC x .
In order to apply ADM to the A`2 /`1 problem, we first
define an auxiliary variable and transform it into an
equivalent problem:
(12)
Since the generated synthetic ROIs for each individual
form a block of the dictionary, a better classification can
arise from a representation of the probe ROI produced
from the minimum number of blocks from the dictionary
instead of looking for the representation of a probe
ROI in the dictionary of all the training data using the
so-called structured SRC which its goal is to find a
representation of a probe ROI that uses the minimum
number of blocks from the dictionary.
For a dictionary
DC = DC [1], DC [2], . . . , DC [n] with blocks DC [i],
i = 1, . . . , n, the block sparsity is formulated in terms of
mixed `2 /`0 norm as
n
X
I(k x[i] k2 > 0) s.t. y = DC x ,
A`2 /`0 : min
x
min k z k2,1 =
x
zgi
0
0
0
..
.
vi := Wi x[i] +
0
···
wi(q+1)
for i = 1, . . . , m ,
1
(λ1 )gi ,
β
(21)
(22)
Finally, the multipliers λ1 and λ2 are updated in the
standard way:
(
λ1 ← λ1 − γ1 β1 (z − x)
(23)
λ2 ← λ2 − γ2 β2 (DC x − y)
!
√
5+1
, the sequence
For β1 , β2 > 0 and γ1 , γ2 ∈ 0,
2
(k) (k)
(x , z ) generated by ADM algorithm from any
initial point (x(0) , z(0) ) converges to (x∗ , z∗ ), where
(x∗ , z∗ ) is a solution of Eq. 19.
1
vi
= max k vi k2 − , 0
,
β1
k vi k2
where
(15)
···
···
..
.
β1
β1
k z − x k22 − λT2 (DC x − y) + k DC x − y k22 .
2
2
Minimizing (Eq. 19) with respect to z is obtained by the
one-dimensional shrinkage formula as follows:
where
0
wi2
..
.
y = DC x .
(β1 I+β2 DTC DC )x = β1 z−λ1 +β2 DTC y+DTC λ2 . (20)
Finally, the weighted matrix obtained in III-A3 which
shows cluster weights is multiplied to the `1 -minimization
term.
n
X
A`2 /`1 : min
kWi x[i] k2 s.t. y = DC x .
wi1
0
Wi = .
..
s.t. z = x,
(19)
where λ1 ∈ Rn , λ2 ∈ Rm are multipliers and β1 , β2 > 0
are penalty parameters. Then we apply the ADM approach to minimize the augmented Lagrangian problem
with respect to x and z alternately. The x-subproblem,
namely minimizing (Eq. 19) with respect to x, is given
by the following linear system:
i=1
i=1
k zgi k2
i=1
minx k z k2,1 − λT1 (z − x) +
(13)
where I(.) is the indicator function, and x[i] is the ith
block in the sparse coefficient vector x corresponding to
the dictionary block DC [i]. Since each dictionary block
corresponds to a specific class, i represents the class index
ranging from 1 to n as well. This optimization problem
seeks the minimum number of non-zero coefficient blocks
that reconstruct the probe ROI.
Note that the optimization program A`2 /`0 is NP-hard
since it requires searching over all possible few blocks
of x and checking whether they span the given y. A
relaxation of this problem is obtained by replacing the `0
with the `1 norm and solving the Eq.14.
n
X
A`2 /`1 : min
k x[i] k2 s.t. y = DC x . (14)
x
m
X
(18)
Note that problem (Eq. 18) has two blocks of variables
(x and z) and its objective function is separable in the
form of f (x) + g(z) since it only involves z, thus ADM
is applicable. The augmented Lagrangian problem is of
the form
i=1
x
(17)
(16)
•
Step 4: Validation
In practical FR systems, it is important to detect and then
9
reject outlier invalid probe ROIs. We use the sparsity
concentration index (SCI) criteria defined in [3]:
. k. maxi k δi (x) k1 / k x k1 −1
SCI(x) =
k−1
∈ (0, 1) .
(24)
where k is the total number of classes in the probe set and
x ∈ Rn is a coefficient vector. A probe ROI is accepted
as valid if SCI ≥ τ and otherwise rejected as invalid,
where τ ∈ (0, 1) is a threshold.
The S2V face recognition process through dictionary augmentation is formalized in Algorithm 2.
Algorithm 2: A SRC-based still-to-video FR system
1
2
Input: Reference
face models of n classes enlisted in the gallery
R = ri ∈ Rd×d |i = 1, . . . , n , generic set
G = gi ∈ Rd×d |i = 1, . . . , m , threshold τ , and a probe
ROI y.
Generate nq synthetic ROIs for each class using the DSFS method.
Build the cross domain dictionary DC by adding synthetic ROIs to the
reference gallery set. Solve the A`2 /`1 problem using ADM
technique.
// Eq. 15
3
4
5
if SCI ≥ τ then
Find the class label y.
// Eq. 17
6
7
else
8
Reject as invalid.
Output: Class label of y.
V. E XPERIMENTAL M ETHODOLOGY
A. Databases
In order to validate the proposed DSFS for S2V FR under
real-world surveillance conditions, extensive experiments were
conducted on two publicly available datasets – COX-S2V [39]
and Chokepoint [40]. These datasets were selected because
they are the most representative for watch-list screening applications. They contain a high-quality reference image per subject captured under controlled condition (with a still camera),
and lower-quality surveillance videos for each subject captured
under uncontrolled conditions (with surveillance cameras).
Videos are captured over a distributed network of cameras that
covers a range of variations (e.g., pose, illumination, scale).
COX-S2V dataset [39] contains 1000 individuals, with 1
high-quality still image and 4 low-resolution video sequences
per individual simulating video surveillance scenario. In each
video, an individual walk through a designed S-shape route
with changes in illumination, expression, scale, pose, and
blur.The Chokepoint dataset [40] consists of 25 individuals
walking trough portal 1, and 29 individuals (23 male and 6
female) walking trough portal 2. The recording of portal 1 and
portal 2 are one month apart. A camera rig with 3 cameras
is placed just above a door and is used for simultaneously
recording the entry of a person during four sessions. In total,
the dataset consists of 54 video sequences and 64, 204 face
images. The appearance of the face captured have variations
in terms of illumination conditions, pose, and misalignment.
B. Experimental protocol
With the Chokepoint database, 5 individuals are randomly
chosen as watch-list individuals that each individual includes a
high-quality frontal captured image. Prior to each experiment,
the video data is split into 3 parts. ROIs are extracted from
the video sequences of 10 other individuals selected at random
as a generic set to represent capture conditions. ROIs of the
video sequences of the remaining individuals along with video
sequences of the 5 already selected watch-list individuals are
employed for testing. In order to obtain representative results,
this process is repeated 5 times with a different random
selection of watch-list and generic set individuals, and the
average accuracy is reported with mean and standard deviation
over all the runs. With COX-S2V, 30 individuals are randomly
considered as watch-list individuals including a high-quality
captured image per each individual. Their corresponding lowquality video sequences along with ROIs of the video sequences of 100 other individuals are employed for testing.
The ROIs extracted from the video sequences of 100 other
individuals are selected at random as a generic set to represent
capture conditions. This process is replicated 5 times with
different stills and videos of watch-list individuals, and the
average accuracy is reported with mean and standard deviation
over all the runs. During the enrollment, a generic set of faces
is captured from video trajectories across all ODs (i.e., global
modeling), and their ROIs are extracted using the Viola-Jones
face detection algorithm [41]. Face detection is also applied to
still images prior to face synthesis. AP clustering is applied to
the generic set, where q representative video ROIs are selected
under various pose, illumination and contrast conditions, and
a weight is assigned to each exemplar according to the
cluster size. Then, q synthetic face images are generated for
each individual based on the information obtained from these
selected exemplars. Recall that AP clustering seeks exemplars
(samples that are representative of clusters), and automatically
determines k and q, the number of clusters, for each independent replication. The domain-invariant dictionary is then
designed using the reference still and synthetic ROIs. During
the operational phase, recognition is performed by coding the
probe image over the domain-invariant dictionary regarding
the weights obtained in the enrollment domain. Throughout
the experiments, the sparsity parameter Λ is fixed to 0.005.
For reference, the S2V FR system based on individual-specific
SVMs is also evaluated. During the enrollment, a non-linear
SVM classifier with RBF kernel is trained for each individual
using target ROIs (reference still of the individual plus the
related synthetically face images) versus non-target ROIs (reference still of cohort persons plus their synthetic face images).
C. Performance Measures
To assess the ability of face synthesizing techniques to
address shifts in domain, a domain shift quantification (DSQ)
measure is employed. With this measure, the similarity between a dictionary designed using synthetic ROIs (DA ) is
compared with a dictionary formed with images collected from
the OD (DR ) by measuring the mean pixel error between
the dictionaries. Given two dictionaries DA and DR with
the same number of images, the DSQ measure is defined
10
VI. R ESULTS AND D ISCUSSION
This section first presents some examples of synthetic
faces generated using the DSFS technique, and compares
them with synthetic faces generated using state-of-the-art face
synthesizing methods: 3DMM [25], 3DMM-CNN [26], and
SHBMM [27]. Then, the performance of S2V FR systems
based on individual-specific SVMs and on SRC is presented
when using these synthetic facial ROIs for system design.
FR performance is assessed when increasing the number of
synthetic ROIs per each individual according to pose angles
and lighting effects. To characterize the impact on performance, these systems are tested with a growing number of
synthetic ROIs and generic training set, and compared with
several relevant state-of-the-art S2V FR systems: ESRC [8],
RADL [19], SVDL [17]), LGR [20], and Flow-based face
frontalization [5]. The final experiment compares the performance of a system designed with synthetic ROIs obtained
with DSFS, to a system designed with a growing number of
randomly-selected synthetic ROIs.
A. Face Synthesis
This subsection presents examples of pose and lighting
exemplars obtained by clustering of facial trajectories in the
captured condition space. Figs. 6a and 6b show an example of
pose clusters obtained with Chokepoint video trajectories of
10 individuals, and with COX-S2V video trajectories of 100
individuals. In this experiment, k = 9 and k = 7 pose clusters
(exemplars) are typically determined with the Chokepoint and
COX-S2V videos, respectively. The second level of clustering
is then applied in the illumination and contrast measure space
on each pose clusters. Figs. 6c and 6d show the exemplars
selected based on both pose and lighting with the proposed
representative selection of DSFS (see section III-A). Overall,
q = 22 and q = 18 exemplars were typically selected based
on both pose and lighting clusters determined in Chokepoint
and COX-S2V videos, respectively.
Figs. 7a to 7d show examples of synthetic ROIs generated
under different pose, illumination and contrast conditions
(b)
(a)
Pose Cluster 3
Pose Cluster 7
1
0.7
0.9
0.6
0.8
0.7
Illumination
0.5
Illumination
as Qdsq = kDTR DA kF where a higher value indicates less
domain shift [42].
The accuracy of the S2V FR system is assessed per individual of interest at the transaction level, using the receiver
operating characteristic (ROC) space, where the true positive
rates (TPRs) are plotted as a function of false positive rates
(FPRs) over all threshold values. TPR is the proportion of
target ROIs that correctly classified as individuals of interest
over the total number of target ROIs in the sequence, while
FPR is the proportion of non-target ROIs incorrectly classified
as individuals of interest over the total number of non-target
ROIs. The area under ROC curve is a global scalar measure of
accuracy that can be interpreted as the probability of correct
classification over the range of TPR and FPR. Accordingly, accuracy of FR systems is estimated using the partial area under
ROC curve pAUC(20%) (using the AUC at 0 <FPR≤ 20%).
Since the number of target and non-target data are imbalanced,
the area under precision-recall curves (AUPR) is also used to
estimate the performance of FR systems.
0.4
0.3
0.6
0.5
0.4
0.3
0.2
0.2
0.1
0.4
0.5
0.6
0.7
Contrast
(c)
0.8
0.9
1
0.1
0
0.2
0.4
0.6
0.8
1
Contrast
(d)
Figure 6: Examples of representative selection results using AP
clustering technique in terms of (a,b) pose angles, and (c,d) luminance
and contrast with Chokepoint dataset on video sequences of 10 individuals and COX-S2V dataset on video sequences of 100 individuals,
respectively, where the center of clusters show exemplars.
using the DSFS face synthesizing technique on the Chokepoint
and COX-S2V datasets, respectively, where Basel Face Model
(BFM) are used as generative 3D shape model [35]. In
Appendix A, Fig.12 compares the quality of synthetic face
generated via DSFS, 3DMM [25], and 3DMM-CNN [26]
techniques..
The synthetic ROIs generated using the DSFS are also
evaluated quantitatively. Table I shows the DSQ values of
the DSFS and other state-of-the-art face synthesizing methods
including 3DMM [25] 3DMM-CNN [26], and SHBMM [27]
on Chokepoint and COX-S2V datasets. Higher DSQ values
indicate a smaller domain shift, and potentially higher recognition rate between the corresponding two domains. The results
are provided under the two following scenarios.
1) Frontal View: In the first experiment, 5 individuals in ED
are randomly selected. A set of synthetic face are generated
with a frontal view under various lighting effects from the
still ROI of each individual to design DA . The corresponding
video ROIs in the OD under the frontal view are then collected
to form DR . Finally, the DSQ is measured for the DA and
Dr dictionaries.
2) Profile View: In the second experiment, 5 individuals
in ED are again randomly selected. Their synthetic ROIs
are generated with profile view and different illumination
conditions to form DA . The corresponding video ROIs in OD
under profile view are collected to construct DR . Finally, the
DSQ is estimated for the dictionaries.
As shown in Table I, DSQ values of DSFS method are
higher followed most closely by SHBMM in both scenarios.
Accordingly, the cross-domain dictionary designed by the
synthetic ROIs generated via the DSFS method is most suitable
to reduce visual domain shifts and potentially achieve a higher
level of accuracy. These results are in line with the recognition
performance results.
11
(c)
(b)
(a)
(d)
Figure 7: Examples of synthetic ROIs generated under different capture conditions using the DSFS technique with Chokepoint (a,b) and
COX-S2V (c,d) datasets.
Table I: Average DSQ value for frontal and profile views on
Chokepoint and COX-S2V datasets.
0.7
0.54
SVM
SRC
0.52
0.65
0.5
DSFQ
COX-S2V database
Frontal ViewProfile View
Frontal ViewProfile View
0.55
0.5
8.27
7.38
8.24
7.63
0.45
3DMM-CNN[26]
7.61
7.03
7.57
7.29
0.4
9.16
7.26
9.34
7.71
Proposed DSFS
9.53
8.17
9.58
8.39
0.46
0.44
0.42
0.4
3DMM[25]
SHBMM[27]
0.48
AUPR
Chokepoint database
pAUC
Technique
0.6
0.38
0.36
Baseline p
1
p2
p3
p4
p5
p6
p7
p8
Baseline p
p9
Number of Synthetic Pose ROIs per each Individual
1
p2
p3
p4
p5
p6
p7
p8
p9
Number of Synthetic Pose ROIs per each Individual
(a)
(b)
0.7
0.58
0.56
0.65
0.54
B. Face Recognition
0.52
AUPR
pAUC
In this subsection, the performance achieved using the S2V
FR system based on SRC and DSFS (see Section IV) is
assessed experimentally. For reference, the S2V FR system
based on individual-specific SVMs is also evaluated.
1) Pose Variations: The S2V FR system is evaluated versus
the number of synthetic ROIs that incorporate growing facial
pose. Figs. 8a and 8d show the average AUC and AUPR
obtained by increasing the number of synthetic ROIs generated
using DSFS from k representative pose angles (pi , i = 1 . . . k)
and with fixed lighting condition. Results indicate that by
adding extra synthetic ROIs generated under representative
pose angles allows to outperform baseline systems designed
with a original reference still ROI alone. AUC and AUPR
accuracy increases by about 10%, typically with only k = 9
and k = 7 synthetic pose ROIs for Chokepoint and COX-S2V
datasets, respectively.
2) Mixed Pose and Illumination Variations: The performance of S2V FR systems is assessed versus the number of
synthetic ROIs generated under both pose and lighting effects.
Figs. 9a and 9d show average AUC and AUPR obtained
by increasing the number of synthetic ROIs used to design
SRC and SVM classifiers on the Chokepoint and COX-S2V
databases, where Ai is a set of synthetic ROIs generated
using DSFS technique under various pose and illumination
conditions. Adding synthetic ROIs generated under various
pose, illumination and contrast conditions allows to significantly outperform the baseline system designed with the
original reference still ROI alone. AUC and AUPR accuracy
increases by about 40%, typically with only q = 24 and
q = 18 synthetic ROIs for Chokepoint and COX-S2V datasets,
0.6
0.55
0.5
0.5
0.48
0.46
0.44
0.45
0.42
0.4
0.4
Baseline p
1
p2
p3
p4
p5
p6
p7
Number of Pose Samples per each Individual
(c)
Baseline p
1
p2
p3
p4
p5
p6
p7
Number of Pose Samples per each Individual
(d)
Figure 8: Average AUC and AUPR versus the number of synthetic
ROIs generated with DSFS according to various pose and fixed
illumination. The S2V FR system employs either SVM and SRC
classifiers on Chokepoint (a,b) and COX-S2V (c,d) databases.
respectively. As shown in Figs. 9a and 9d, accuracy for
DSFS+SRC trends to stabilize to its maximum value when
the size of the generic set is greater than q in DSFS. To
view performance stabilizing with more than q synthetic ROIs,
additional samples were selected randomly among AP clusters.
Note that the Chokepoint dataset contains faces captured for
a range illumination conditions with various densities. Hence,
some exemplars may represent many video ROIs. Our method
assigns higher weights to such distributions, and may yield a
higher level of performance. The results obtained with DSFS
are also compared to S2V FR systems that exploit stateof-the-art face synthesis techniques including 3DMM [25]
(with randomly selected images), and SHBMM [27] (with 9
spherical harmonic basis images). As shown in Fig. 9, DSFS
always outperforms these other techniques.
3) Impact of Representative Selection: Without prior
knowledge of the OD, synthetic faces are generated according
to a uniform distribution. Adding a large number of synthetic
ROIs to the dictionary as needed to cover all possible cases can
12
1
0.7
0.6
0.7
0.5
0.6
0.4
0.5
0.4
0.3
Baseline
5
10
15
20
25
Baseline
5
10
15
20
1
25
Number of Synthetic Pose and Illumination ROIs per each Individual
Number of Synthetic Pose and Illumination ROIs per each Individual
0.9
(a)
(b)
0.8
0.6
0.75
0.5
Baseline
0.4
0.7
0.4
Baseline
0.3
0.3
0.2
AUPR
0.8
pAUC
0.5
0.6
0.8
0.9
0.7
0.7
pAUC
1
0.8
3DMM(with all possible cond.)+SRC
DSFS(with all capture cond.)+SRC
DSFS(with rep. capture cond.)+SRC
AUPR
pAUC
0.8
conditions. The third dictionary employs 180 synthetic pose
ROIs generated under different pose and illumination by
3DMM. The results in Fig. 10 suggest that augmenting the
dictionary using representative synthetic ROIs with the DSFS
technique yields a higher level of accuracy, particularly under
both pose and illumination conditions.
0.8
3DMM+SVM
3DMM-CNN+SVM
SHBMM+SVM
DSFS+SVM
3DMM+SRC
3DMM-CNN+SRC
SHBMM+SRC
DSFS+SRC
AUPR
0.9
0.7
0.65
0.2
0.6
0.1
0.6
0.1
0
0
Synthetic Pose ROIs
0.55
Synthetic Pose ROIs
(a)
0.5
0.5
Synthetic Pose and Illuination ROIs
Synthetic Pose and Illuination ROIs
(b)
0.45
0.8
1
0.4
0.4
Baseline
5
10
15
20
Baseline
5
10
15
0.9
20
Number of Synthetic Pose and Illumination ROIs per each Individual
Number of Synthetic Pose and Illumination ROIs per each Individual
0.8
(c)
(d)
0.7
0.7
0.6
ROIs generated with DSFS, 3DMM, and SHBMM according to pose
and lighting effects where S2V FR system employs either SVM and
SRC classifiers on Chokepoint (a,b) and COX-S2V (c,d) databases.
AUPR
pAUC
Figure 9: Average AUC and AUPR versus the number of synthetic
0.5
0.6
Baseline
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
Synthetic Pose ROIs
Synthetic Pose and Illuination ROIs
(c)
significantly increase the time and memory complexity of a FR
systems and, more importantly, may cause over-fitting. The
proposed DSFS technique extracts representative information
from the OD to produce a compact set of synthetic ROIs
that are robust to intra-class variations in the OD. In order
to evaluate the impact of generating synthetic ROIs based
on representative information (i.e., pose and lighting cluster
instances), three dictionaries are designed for SRC: (1) a
dictionary designed with representative synthetic ROIs (DSFS
technique); (2) a dictionary designed with the synthetic ROIs
under all capture conditions (DSFS without AP clustering);
and (3) a dictionary designed under all possible conditions.
Figs. 10a and 10d compares the AUC and AUPR of the
SRC-based S2V FR system according to different number of
synthetic ROIs in terms of pose and illumination conditions.
The first scenario evaluates the impact of representative selection in terms of pose with three dictionaries are designed
for SRC. The first dictionary typically employs k = 9 and
k = 7 representative synthetic pose ROIs generated with
the DSFS technique for Chokepoint and COX-S2V datasets,
respectively. The second dictionary employs 100 synthetic
pose ROIs generated by DSFS technique under all OD capture
conditions. The third dictionary employs 180 synthetic pose
ROIs generated by 3DMM in a set of rotation angles ranging
from to −60 to +60. The second scenario evaluates the
impact of representative selection in terms of both pose and
illumination conditions with three dictionaries are designed
for SRC. The first dictionary typically employs k = 9 and
k = 7 representative synthetic ROIs generated under different
pose and illumination with the DSFS technique for Chokepoint
and COX-S2V datasets, respectively. The second dictionary
employs 100 synthetic pose ROIs generated under different
pose and illumination by DSFS technique under all OD capture
Baseline
0.4
Synthetic Pose ROIs
Synthetic Pose and Illuination ROIs
(d)
Figure 10: Average AUC and AUPR of a S2VFR system designed
with representative synthetic ROIs vs a system designed with randomly generated synthetic ROIs on Chokepoint (a,b) and COX-S2V
(c,d) datasets.
C. Comparison with Reference Techniques
With the above experimental setting, we compare the
recognition rate of the DSFS technique with the recent face
synthesizing methods (3DMM[25], SHBMM[27]) in a S2V FR
framework. We also present the impact of using face synthesizing along with KSVD dictionary learning [43]. Following
this, the recognition rate of the DSFS technique with existing
generic learning techniques (ESRC[8], RADL[19], SVDL[17],
LGR[20]) is compared that regularization parameter Λ is set to
0.005. Note that the performance of the face synthesizing techniques is evaluated w/o dictionary learning. We also compared
the DSFS results with the results obtained by Flow-based face
frontalization method [5]. The quality of frontalization via
the flow-based face frontalization technique is shown in Fig.
13 (Appendix B). Table II lists and compares the recognition
performance where the results (recognition rate) are illustrated
by the mean and standard deviation of 5 runs.
Next, we compare the computational complexity in terms
of average running time for each individual as well as number
of inner products needed per each iteration. Figs. 11a and 11b
shows the computational complexity in terms of number of
inner products with a growing number of synthetic ROIs.
Table. III compares the complexity of the proposed
DSFS-SRC algorithm with RADL[19], LGR[20], and flowbased frontalization techniques on Chokepoint and COX-S2V
datasets per each iteration. The experiments are conducted in
MATLAB R2016b (64bit) Linux version on a PC workstation
with an INTEL CPU (3.41-GHz) and 16GB RAM.
13
Table II: Comparative transaction level analysis of the proposed FR approach and related state-of-the art FR methods with
Chokepoint and COX-S2V databases.
Category
pAUC
AUPR
pAUC
AUPR
Baseline
SRC
0.524±0.041
0.415±0.035
0.563±0.034
0.457±0.036
ESRC[8]
SRC
SRC-KSVD
0.802±0.028
0.813±0.027
0.651±0.032
0.672±0.020
0.835±0.031
0.842±0.017
0.695±0.032
0.715±0.020
RADL[19]
SRC
0.852±0.035
0.724±0.031
0.892±0.025
0.753±0.027
LGR[20]
SRC
0.849±0.031
0.717±0.024
0.881±0.028
0.744±0.025
SVDL[17]
SRC
0.825±0.022
0.703±0.031
0.843±0.025
0.724±0.031
3DMM[25]
SRC
SRC-KSVD
0.681±0.037
0.732±0.035
0.523±0.037
0.605±0.032
0.706±0.031
0.757±0.032
0.562±0.032
0.641±0.028
3DMM-CNN[26]
SRC
SRC-KSVD
0.652±0.028
0.714±0.027
0.516±0.026
0.585±0.025
0.702±0.027
0.721±0.027
0.552±0.025
0.603±0.025
SHBMM[27]
SRC
SRC-KSVD
0.725±0.034
0.786±0.028
0.593±0.040
0.672±0.022
0.739±0.035
0.791±0.030
0.607±0.041
0.681±0.028
Face Frontalization
Flow-based[5]
SRC
0.824±0.023
0.711±0.024
0.852±0.028
0.719±0.027
Face Synthesizing
+
Generic Learning
Proposed DSFS
SRC
0.903±0.032
0.752±0.037
0.925±0.26
0.774±0.31
Face Synthesizing
# 108
18
LGR
RADL
DSFS+SRC
4.5
cost than other state-of-the-art methods. The main reason is
that the dictionary designed by DSFS technique is able to
represent real-world capture conditions and does not require a
traditional dictionary learning process.
# 108
16
4
Number of Inner Products
Number of Inner Products
COX-S2V database
Classifier
Generic Learning
5
Chokepoint database
Technique
3.5
3
2.5
2
1.5
1
14
12
10
8
6
VII. C ONCLUSIONS
4
2
0.5
0
0
10
50
100
150
200
Number of Synthetic ROIs in the Gallery Dictionary
10
200
400
600
800
1000
Number of Synthetic ROIs in the Gallery Dictionary
(a)
(b)
Figure 11: Time complexity versus the number of synthetic ROIs
on Chokepoint (a) and COX-S2V (b) data.
Table III: Average computational complexity and time of the
DSFS and state-of-the-art techniques on Chokepoint and COXS2V datasets.
Technique
Chokepoint database
COX-S2V database
No.inner productsRun time(s)No.inner productsRun time(s)
RADL[19]
52,650,000
3.12
350,500,000
6.17
LGR[20]
273,100,000
7.13
1,147,200,000
11.35
Face Frontalization+SRC[5]
40,510,000
3.14
245,320,000
4.08
Proposed DSFS+SRC
32,450,000
1.55
190,250,000
2.61
The results show our proposed method that is a joint
use of generic learning and face synthesizing achieves superior recognition results compared to the other methods
under the same configuration which verifies that our face
synthesizing technique better preserves identity information.
Although RADL, LGR and the flow-base face frontalization
techniques can achieve comparable accuracy to our approach,
they are computationally expensive. It can be concluded that
augmenting the SRC with synthetic ROIs generated by DSFS
technique has a good recognition rate with less computational
This paper proposes a domain-specific face synthesizing
(DSFS) technique to improve the performance of still-tovideo FR systems when surveillance videos are captured under
various uncontrolled controlled conditions, and individuals are
recognized based on a single facial image. The proposed
approach takes advantage of operational domain information
from the generic set that can effectively represent probe ROIs.
A compact set of synthetic faces is generated that resemble
individuals of interest under capture conditions relevant to
the operational domain. For proof-of-concept validation, augmented dictionary with a block structure is designed based
on DSFS, and face classification is performed within a SRC
framework. Our experiments on the Chokepoint and COX-S2V
datasets show that augmenting the reference discretionary of
still-to-video FR systems using the proposed DSFS approach
can provide a significantly higher level of accuracy compared
to state-of-the-art approaches, with only a moderate increase
in its computational complexity. Result indicated that face
synthesis alone cannot effectively resolve the challenges of
the SSPP and visual domain shift problems. With DSFS,
generic learning and face synthesis operate complementarity.
The proposed DSFS technique could be improved to generate
synthetic faces with expression variations for a robust FR. In
addition, to improve performance, the representative synthetic
ROIs generated using DSFS could be applied to generate local
camera-specific ROIs. DSFS is general in that synthetic ROIs
could be applied to train or fine-tune a multitude of face
recognition systems like deep CNNs, with information that
robust models to specific operational domains.
14
R EFERENCES
[1] M. A. A. Dewan, E. Granger, G.-L. Marcialis, R. Sabourin, F. Roli,
Adaptive appearance model tracking for still-to-video face recognition,
Pattern Recognition 49 (2016) 129–151.
[2] S. Bashbaghi, E. Granger, R. Sabourin, G.-A. Bilodeau, Dynamic
ensembles of exemplar-svms for still-to-video face recognition, Pattern
Recognition 69 (2017) 61–81.
[3] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, Y. Ma, Robust face
recognition via sparse representation, IEEE Transactions on Pattern
Analysis and Machine Intelligence 31 (2) (2009) 210–227.
[4] A. Wanger, J. Wright, A. Ganesh, Z. Zhou, Y. Ma, Towards a practical
face recognition system: robust registration and illumination by sparse
representation, IIEEE Transactions on Pattern Analysis and Machine
Intelligence 34 (2) (2009) 597–604.
[5] T. Hassner, S. Harel, E. Paz, R. Enbar, Effective face frontalization in
unconstrained images, in: CVPR, 2015, pp. 4295–4304.
[6] Y. Liu, I. J. Wassell, A new face recognition algorithm based on
dictionary learning for a single training sample per person, in: BMVC,
2015.
[7] I. Masi, A. T. Tran, T. Hassner, J. T. Leksut, G. Medioni, Do we really
need to collect millions of faces for effective face recognition?, in:
ECCV, 2016.
[8] W. Deng, J. Hu, J. Guo, Extended SRC: Undersampled face recognition
via intraclass variant dictionary, IEEE Transactions on Pattern Analysis
and Machine Intelligence 34 (9).
[9] W. Deng, J. Hu, X. Zhou, J. Guo, Equidistant prototypes embedding
for single sample based face recognition with generic learning and
incremental learning, Pattern Recognition 47 (12) (2014) 3738–3749.
[10] E. Elhamifar, R. Vidal, Robust classification using structured sparse
representation, in: CVPR, 2011, pp. 1873–1879.
[11] S. Bashbaghi, E. Granger, R. Sabourin, G.-A. Bilodeau, Robust watchlist screening using dynamic ensembles of svms based on multiple face
representations, Machine Vision and Applications 28.
[12] J. Lu, Y.-P. Tan, G. Wang, Discriminative multimanifold analysis for face
recognition from a single training sample per person, IEEE Transactions
on Pattern Analysis and Machine Intelligence 35 (1) (2013) 39–51.
[13] W. AbdAlmageed, Y. Wu, S. Rawls, S. Harel, T. Hassner, I. Masi,
J. Choi, J. Lekust, J. Kim, P. Natarajan, et al., Face recognition using
deep multi-pose representations, in: WACV, 2016, pp. 1–9.
[14] Y. Su, S. Shan, X. Chen, W. Gao, Adaptive generic learning for face
recognition from a single sample per person, in: CVPR, 2010, pp. 2699–
2706.
[15] Y. Gao, J. Ma, A. L. Yuille, Semi-supervised sparse representation based
classification for face recognition with insufficient labeled samples,
IEEE Transactions on Image Processing 26 (5) (2017) 2545–2560.
[16] Z.-M. Li, Z.-H. Huang, K. Shang, A customized sparse representation
model with mixed norm for undersampled face recognition, IEEE
Transactions on Information Forensics and Security 11 (10) (2016)
2203–2214.
[17] M. Yang, L. Van Gool, L. Zhang, Sparse variation dictionary learning
for face recognition with a single training sample per person, in: ICCV,
2013, pp. 689–696.
[18] F. Nourbakhsh, E. Granger, G. Fumera, An extended sparse classification
framework for domain adaptation in video surveillance, in: ACCV, 2016,
pp. 360–376.
[19] C.-P. Wei, Y.-C. F. Wang, Undersampled face recognition via robust
auxiliary dictionary learning, IEEE Transactions on Image Processing
24 (6) (2015) 1722–1734.
[20] P. Zhu, M. Yang, L. Zhang, I.-Y. Lee, Local generic representation for
face recognition with single sample per person, in: ACCV, 2014, pp.
34–50.
[21] S. Bashbaghi, E. Granger, R. Sabourin, G.-A. Bilodeau, Ensembles of
exemplar-svms for video face recognition from a single sample per
person, in: AVSS, 2015, pp. 1–6.
[22] M. Parchami, S. Bashbaghi, E. Granger, S. Sayed, Using deep autoencoders to learn robust domain-invariant representations for still-to-video
face recognition, in: AVSS, 2017, pp. 1–6.
[23] C. Shao, X. Song, Z.-H. Feng, X.-J. Wu, Y. Zheng, Dynamic dictionary
optimization for sparse-representation-based face classification using
local difference images, Information Sciences 393 (2017) 1–14.
[24] F. Mokhayeri, E. Granger, G.-A. Bilodeau, Synthetic face generation
under various operational conditions in video surveillance, in: ICIP,
2015, pp. 4052–4056.
[25] V. Blanz, T. Vetter, Face recognition based on fitting a 3D morphable
model, IEEE Transactions on Pattern Analysis and Machine Intelligence
25 (9) (2003) 1063–1074.
[26] A. T. Tran, T. Hassner, I. Masi, G. Medioni, Regressing robust and
discriminative 3D morphable models with a very deep neural network,
in: CVPR, 2017, pp. 1493–1502.
[27] L. Zhang, D. Samaras, Face recognition from a single training image
under arbitrary unknown lighting using spherical harmonics, IEEE
Transactions on Pattern Analysis and Machine Intelligence 28 (3) (2006)
351–363.
[28] E. Richardson, M. Sela, R. Or-El, R. Kimmel, Learning detailed face
reconstruction from a single image, in: CVPR, 2017, pp. 5553–5562.
[29] A. Tewari, M. Zollhöfer, H. Kim, P. Garrido, F. Bernard, P. Pérez,
C. Theobalt, Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction, arXiv preprint
arXiv:1703.10580.
[30] X. Chen, H. Wu, X. Jin, Q. Zhao, Face illumination manipulation
using a single reference image by adaptive layer decomposition, IEEE
Transactions on Image Processing 22 (11) (2013) 4249–4259.
[31] A. Asthana, S. Zafeiriou, S. Cheng, M. Pantic, Robust discriminative
response map fitting with constrained local models, in: CVPR, 2013,
pp. 3444–3451.
[32] Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, Image quality assessment: from error visibility to structural similarity, IEEE Transactions
on Image Processing 13 (4) (2004) 600–612.
[33] B. J. Frey, D. Dueck, Clustering by passing messages between data
points, science 315 (5814) (2007) 972–976.
[34] J. Jeon, S. Cho, X. Tong, S. Lee, Intrinsic image decomposition using
structure-texture separation and surface normals, in: ECCV, 2014, pp.
218–233.
[35] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, T. Vetter, A 3D face
model for pose and illumination invariant face recognition, in: AVSS,
2009, pp. 296–301.
[36] X. Zhu, Z. Lei, X. Liu, H. Shi, S. Z. Li, Face alignment across large
poses: A 3D solution, in: CVPR, 2016, pp. 146–155.
[37] D. Jiang, Y. Hu, S. Yan, L. Zhang, H. Zhang, W. Gao, Efficient 3d
reconstruction for face recognition, Pattern Recognition 38 (6) (2005)
787–798.
[38] W. Deng, W. Yin, Y. Zhang, Group sparse optimization by alternating
direction method, in: SPIE, Vol. 8858, 2013, p. 88580R.
[39] Z. Huang, S. Shan, R. Wang, H. Zhang, S. Lao, A. Kuerban, X. Chen,
A benchmark and comparative study of video-based face recognition
on cox face database, IEEE Transactions on Image Processing 24 (12)
(2015) 5967–5981.
[40] Y. Wong, S. Chen, S. Mau, C. Sanderson, B. C. Lovell, Patch-based
probabilistic image quality assessment for face selection and improved
video-based face recognition, in: CVPRW, 2011, pp. 74–81.
[41] P. Viola, M. J. Jones, Robust real-time face detection, International
Journal of Computer Vision 57 (2) (2004) 137–154.
[42] J. Ni, Q. Qiu, R. Chellappa, Subspace interpolation via dictionary
learning for unsupervised domain adaptation, in: CVPR, 2013, pp. 692–
699.
[43] M. Aharon, M. Elad, A. Bruckstein, rmk-svd: An algorithm for
designing overcomplete dictionaries for sparse representation, IEEE
Transactions on Signal Processing 54 (11) (2006) 4311–4322.
A PPENDIX
A. Face Synthesizing Results
Examples of the synthetic face images generated via DSFS,
3DMM[25] and 3DMM-CNN[26] techniques for individual
ID#23 of the Chokepoint dataset in Fig. 12 shown.
B. Face Frontalization Results
Fig. 13 shows an example of the face frontalized for the
facial probe ROI for individual ID#23 of the Chokepoint
dataset via the flow-based face frontalization technique [5].
15
(i)
(ii)
(iii)
(i)
(iv)
(v)
(ii)
(iii)
(i)
(iv)
(v)
(ii)
(iii)
Figure 12: Synthetic face images generated via (a) DSFS, (b) 3DMM, (c) 3DMM-CNN with Chockpoint data.
Figure 13: Face frontalization via flow-based technique: (a) probe
face image under profile view, (b) frontalized face image.
(iv)
(v)
| 1cs.CV
|
Differentially Private Ordinary Least Squares
Or Sheffet 1
arXiv:1507.02482v4 [cs.DS] 21 Aug 2017
Abstract
Linear regression is one of the most prevalent
techniques in machine learning; however, it is
also common to use linear regression for its explanatory capabilities rather than label prediction. Ordinary Least Squares (OLS) is often used
in statistics to establish a correlation between
an attribute (e.g. gender) and a label (e.g. income) in the presence of other (potentially correlated) features. OLS assumes a particular model
that randomly generates the data, and derives tvalues — representing the likelihood of each real
value to be the true correlation. Using t-values,
OLS can release a confidence interval, which is
an interval on the reals that is likely to contain
the true correlation; and when this interval does
not intersect the origin, we can reject the null hypothesis as it is likely that the true correlation
is non-zero. Our work aims at achieving similar guarantees on data under differentially private estimators. First, we show that for wellspread data, the Gaussian Johnson-Lindenstrauss
Transform (JLT) gives a very good approximation of t-values; secondly, when JLT approximates Ridge regression (linear regression with
l2 -regularization) we derive, under certain conditions, confidence intervals using the projected
data; lastly, we derive, under different conditions,
confidence intervals for the “Analyze Gauss” algorithm (Dwork et al., 2014).
1. Introduction
Since the early days of differential privacy, its main goal
was to design privacy preserving versions of existing techniques for data analysis. It is therefore no surprise that several of the first differentially private algorithms were machine learning algorithms, with a special emphasis on the
ubiquitous problem of linear regression (Kasiviswanathan
et al., 2008; Chaudhuri et al., 2011; Kifer et al., 2012; Bass1
Computing Science Dept., University of Alberta, Edmonton
AB, Canada. This work was done when the author was at Harvard University, supported by NSF grant CNS-123723. Correspondence to: Or Sheffet <[email protected]>.
ily et al., 2014). However, all existing body of work on
differentially private linear regression measures utility by
bounding the distance between the linear regressor found
by the standard non-private algorithm and the regressor
found by the privacy-preserving algorithm. This is motivated from a machine-learning perspective, since bounds
on the difference in the estimators translate to error bounds
on prediction (or on the loss function). Such bounds are
(highly) interesting and non-trivial, yet they are of little use
in situations where one uses linear regression to establish
correlations rather than predict labels.
In the statistics literature, Ordinary Least Squares (OLS)
is a technique that uses linear regression in order to infer
the correlation between a variable and an outcome, especially in the presence of other factors. And so, in this paper, we draw a distinction between “linear regression,” by
which we refer to the machine learning technique of finding
a specific estimator for a specific loss function; and “Ordinary Least Squares,” by which we refer to the statistical inference done assuming a specific model for generating the
data and that uses linear regression. Many argue that OLS
is the most prevalent technique in social sciences (Agresti
& Finlay, 2009). Such works make no claim as to the labels of a new unlabeled batch of samples. Rather they aim
to establish the existence of a strong correlation between
the label and some feature. Needless to say, in such works,
the privacy of individuals’ data is a concern.
In order to determine that a certain variable xj is positively
(resp. negatively) correlated with an outcome y, OLS assumes a model where the outcome y is a noisy version of
a linear mapping of all variables: y = β · x + e (with e
denoting random Gaussian noise) for some predetermined
xi , yi ) OLS
and unknown β . Then, given many samples (x
establishes two things: (i) when fitting a linear function
to best predict y from x over the sample (via computing
P
P
T −1
β=
β̂
( i yix i )) the coefficient β̂j is positive
i x ix i
(resp. negative); and (ii) inferring, based on βˆj , that the
true βj is likely to reside in R>0 (resp. R<0 ). In fact, the
crux in OLS is by describing βj using a probability distribution over the reals, indicating where βj is likely to fall,
derived by computing t-values. These values take into account both the variance in the data as well as the variance of
the noise e.1 Based on this probability distribution one can
1
For example, imagine we run linear regression on a certain
β with coordinates β̂1 = β̂2 =
(X, y ) which results in a vector β̂
Differentially Private Ordinary Least Squares
define the α-confidence interval — an interval I centered
at β̂j whose likelihood to contain βj is 1 − α. Of particular
importance is the notion of rejecting the null-hypothesis,
where the interval I does not contain the origin, and so
one is able to say with high confidence that βj is positive
(resp. negative). Further details regarding OLS appear in
Section 2.
In this work we give the first analysis of statistical inference for OLS using differentially private estimators. We
emphasize that the novelty of our work does not lie in the
differentially-private algorithms, which are, as we discuss
next, based on the Johnson-Lindenstrauss Transform (JLT)
and on additive Gaussian noise and are already known to
be differentially private (Blocki et al., 2012; Dwork et al.,
2014). Instead, the novelty of our work lies in the analyses of the algorithms and in proving that the output of the
algorithms is useful for statistical inference.
The Algorithms. Our first algorithm (Algorithm 1) is an
adaptation of Gaussian JLT. Proving that this adaptation
remains (, δ)-differentially private is straightforward (the
proof appears in Appendix A.1). As described, the algorithm takes as input a parameter r (in addition to the
other parameters of the problem) that indicates the number
of rows in the JL-matrix. Later, we analyze what should
one set as the value of r. Our second algorithm is taken
Algorithm 1 Outputting a private Johnson-Lindenstrauss
projection of a matrix.
n×d
Input: A matrix A ∈ R
and a bound B > 0 on the
l2 -norm of any row in A.
Privacy parameters: , δ > 0.
Parameter r indicating the number of rows in the resulting matrix.
p
2
Set w s.t. w2 = 8B
2r ln(8/δ) + 2 ln(8/δ) .
Sample Z ∼ Lap(4B 2 /) and let σmin (A) denote the
smallest singular value of A.
2
if σmin (A)2 > w2 + Z + 4B ln(1/δ)
then
Sample a (r×n)-matrix R whose entries are i.i.d samples from a normal Gaussian.
return RA and “matrix unaltered”.
else
Let A0 denote the result of appending A with the d×dmatrix wId×d .
Sample a (r × (n + d))-matrix R whose entries are
i.i.d samples from a normal Gaussian.
returnRA0 and “matrix altered”.
end if
verbatim from the work of Dwork et al (2014). We de0.1. Yet while the column X1 contains many 1s and (−1)s, the
column X2 is mostly populated with zeros. In such a setting,
OLS gives that it is likely to have β1 ≈ 0.1, whereas no such
guarantees can be given for β2 .
liberately focus on algorithms that approximate the 2nd moment matrix of the data and then run hypothesis-testing
by post-processing the output, for two reasons. First, they
enable sharing of data2 and running unboundedly many
hypothesis-tests. Since, we do not deal with OLS based
on the private single-regression ERM algorithms (Chaudhuri et al., 2011; Bassily et al., 2014) as such inference requires us to use the Fisher-information matrix of the loss
function — but these algorithms do not minimize a private
loss-function but rather prove that outputting the minimizer
of the perturbed loss-function is private. This means that
differentially-private OLS based on these ERM algorithms
requires us to devise new versions of these algorithms,
making this a second step in this line of work... (After first
understanding what we can do using existing algorithms.)
We leave this approach — as well as performing private hypothesis testing using a PTR-type algorithm (Dwork & Lei,
2009) (output merely reject / don’t-reject decision without justification), or releasing only relevant tests judging
by their p-values (Dwork et al., 2015) — for future work.
Our Contribution and Organization. We analyze the
performances of our algorithms on a matrix A of the form
A = [X; y ], where each coordinate yi is generated according to the homoscedastic model with Gaussian noise, which
is a classical model in statistics. We assume the existence
of a vector β s.t. for every i we have yi = β Tx i + ei and
ei is sampled i.i.d from N (0, σ 2 ).3
We study the result of running Algorithm 1 on such data
in the two cases: where A wasn’t altered by the algorithm
and when A was appended by the algorithm. In the former
case, Algorithm 1 boils down to projecting the data under
a Gaussian JLT. Sarlos (2006) has already shown that the
JLT is useful for linear regression, yet his work bounds
the l2 -norm of the difference between the estimated regression before and after the projection. Following Sarlos’
work, other works in statistics have analyzed compressed
2
Researcher A collects the data and uses the approximation of
the 2nd -moment matrix to test some OLS hypothesis; but once
the approximation is published researcher B can use it to test for
a completely different hypothesis.
3
This model may seem objectionable. Assumptions like the
noise independence, 0-meaned or sampled from a Gaussian distribution have all been called into question in the past. Yet due to
the prevalence of this model we see fit to initiate the line of work
on differentially private Least Squares with this Ordinary model.
Algorithm 2 “Analyze Gauss” Algorithm of Dwork et
al (2014).
Input: A matrix A ∈ Rn×d and a bound B > 0 on the
l2 -norm of any row in A.
Privacy parameters: , δ > 0.
N ← symmetric (d × d)-matrix
with upper
triangle entries sampled i.i.d from N 0, 2B
T
return A A + N .
4
ln(2/δ)
2
.
Differentially Private Ordinary Least Squares
linear regression (Zhou et al., 2007; Pilanci & Wainwright,
2014a;b). However, none of these works give confidence
intervals based on the projected data, presumably for three
reasons. Firstly, these works are motivated by computational speedups, and so they use fast JLT as opposed to our
analysis which leverages on the fact that our JL-matrix is
composed of i.i.d Gaussians. Secondly, the focus of these
works is not on OLS but rather on newer versions of linear
regression, such as Lasso or when β lies in some convex
set. Lastly, it is evident that the smallest confidence interval is derived from the data itself. Since these works do
not consider privacy applications, (actually, (Zhou et al.,
2007; Pilanci & Wainwright, 2014a) do consider privacy
applications of the JLT, but quite different than differential
privacy) they assume the analyst has access to the data itself, and so there was no need to give confidence intervals
for the projected data. Our analysis is therefore the first, to
the best of our knowledge, to derive t-values — and therefore achieve all of the rich expressivity one infers from tvalues, such as confidence bounds and null-hypotheses rejection — for OLS estimations without having access to X
itself. We also show that, under certain conditions, the sample complexity for correctly rejecting the null-hypothesis
increases from a certain
bound N0 (without privacy) to a
√
bound of N0 + Õ( N0 · κ( n1 AT A)/) with privacy (where
κ(M ) denotes the condition number of the matrix M .) This
appears in Section 3.
In Section 4 we analyze the case Algorithm 1 does append
the data and the JLT is applied to A0 . In this case, solving
the linear regression problem on the projected A0 approximates the solution for Ridge Regression (Tikhonov, 1963;
Hoerl & Kennard,
we aim
P 1970).T In 2Ridge2 Regression
2
z
to solve minz
(y
−
z
x
)
+
w
kz
k
,
which
means
i
i
i
we penalize vectors whose l2 -norm is large. In general, it
is not known how to derive t-values from Ridge regression,
and the literature on deriving confidence intervals solely
from Ridge regression is virtually non-existent. Indeed,
prior to our work there was no need for such calculations,
as access to the data was (in general) freely given, and so
deriving confidence intervals could be done by appealing
back to OLS. We too are unable to derive approximated
t-values in the general case, but under additional assumptions about the data — which admittedly depend in part on
β k and so cannot be verified solely from the data — we
kβ
show that solving the linear regression problem on RA0 allows us to give confidence intervals for βj , thus correctly
determining the correlation’s sign.
In Section 5 we discuss the “Analyze Gauss” algorithm (Dwork et al., 2014) that outputs a noisy version of
a covariance of a given matrix using additive noise rather
than multiplicative noise. Empirical work (Xi et al., 2011)
shows that Analyze Gauss’s output might be non-PSD if
the input has small singular values, and this results in truly
bad regressors. Nonetheless, under additional conditions
(that imply that the output is PSD), we derive confidence
bounds for Dwork et al’s “Analyze Gauss” algorithm. Finally, in Section 6 we experiment with the heuristic of
computing the t-values directly from the outputs of Algorithms 1 and 2. We show that Algorithm 1 is more “conservative” than Algorithm 2 in the sense that it tends to
not reject the null-hypothesis until the number of examples is large enough to give a very strong indication of rejection. In contrast, Algorithm 2 may wrongly rejects the
null-hypothesis even when it is true.
Discussion. Some works have already looked at the intersection of differentially privacy and statistics (Dwork &
Lei, 2009; Smith, 2011; Chaudhuri & Hsu, 2012; Duchi
et al., 2013; Dwork et al., 2015) (especially focusing on robust statistics and rate of convergence). But only a handful
of works studied the significance and power of hypotheses
testing under differential privacy, without arguing that the
noise introduced by differential privacy vanishes asymptotically (Vu & Slavkovic, 2009; Uhler et al., 2013; Wang
et al., 2015; Rogers et al., 2016). These works are experimentally promising, yet they (i) focus on different statistical tests (mostly Goodness-of-Fit and Independence testing), (ii) are only able to prove results for the case of simple
hypothesis-testing (a single hypothesis) with an efficient
data-generation procedure through repeated simulations —
a cumbersome and time consuming approach. In contrast,
we deal with a composite hypothesis (we simultaneously
reject all β s with sign(βj ) 6= sign(β̂j )) by altering the
confidence interval (or the critical region).
One potential reason for avoiding confidence-interval analysis for differentially private hypotheses testing is that it
does involve re-visiting existing results. Typically, in statistical inference the sole source of randomness lies in the
underlying model of data generation, whereas the estimators themselves are a deterministic function of the dataset.
In contrast, differentially private estimators are inherently
random in their computation. Statistical inference that considers both the randomness in the data and the randomness
in the computation is highly uncommon, and this work, to
the best of our knowledge, is the first to deal with randomness in OLS hypothesis testing. We therefore strive in our
analysis to separate the two sources of randomness — as
in classic hypothesis testing, we use α to denote the bound
on any bad event that depends solely on the homoscedastic model, and use ν to bound any bad event that depends
on the randomized algorithm.4 (Thus, any result which is
originally of the form “α-reject the null-hypothesis” is now
converted into a result “(α+ν)-reject the null hypothesis”.)
4
Or any randomness in generating the feature matrix X which
standard OLS theory assumes to be fixed, see Theorems 2.2
and 3.3.
Differentially Private Ordinary Least Squares
2. Preliminaries and OLS Background
kζk2 ∼ χ2k , and taking the quantity Z/
p
kζk2 /k. It is a
k→∞
Notation. Throughout this paper, we use lower-case letters to denote scalars (e.g., yi or ei ); bold characters to
denote vectors; and UPPER-case letters to denote matrices. The l-dimensional all zero vector is denoted 0l , and
the l × m-matrix of all zeros is denoted 0l×m . We use e
β in our model; and
to denote the specific vector y − Xβ
though the reader may find it a bit confusing but hopefully
clear from the context — we also use e j and e k to denote
elements of the natural basis (unit length vector in the direction of coordinate j or k). We use , δ to denote the privacy parameters of Algorithms 1 and 2, and use α and ν to
denote confidence parameters (referring to bad events that
hold w.p. ≤ α and ≤ ν resp.) based on the homoscedastic
model or the randomized algorithm resp. We also stick to
the notation from Algorithm 1 and
use w to denote the posip
8B 2
2
tive scalar for which w =
2r ln(8/δ) + ln(8/δ)
throughout this paper. We use standard notation for SVD
composition of a matrix (M = U ΣV T ), its singular values
and its Moore-Penrose inverse (M + ).
The Gaussian distribution.
A univariate Gaussian
N (µ, σ 2 ) denotes the Gaussian distribution whose mean
is µ and variance σ 2 . Standard concentration
bounds on
p
Gaussians give that Pr[x > µ + 2σ ln(2/ν)] < ν
µ, Σ)
for any ν ∈ (0, 1e ). A multivariate Gaussian N (µ
for some positive semi-definite Σ denotes the multivariate Gaussian distribution where the mean of the j-th coordinate is the µj and the covariance between coordinates
j and k is Σj,k . The PDF of such Gaussian is defined
only on the subspace colspan(Σ). A matrix Gaussian distribution, denoted N (Ma×b , Ia×a , V ) has mean M , independence among its rows and variance V for each of
its columns. We also require the following property of
Gaussian random variables: Let X and Y be two random
Gaussians s.t. X ∼ N (0, σ 2 ) and Y ∼ N (0, λ2 ) where
2
1 ≤ σλ2 ≤ c2 for some c, then for any S ⊂ R we have
1
c Prx←Y [x ∈ S] ≤ Prx←X [x ∈ S] ≤ cPrx←Y [x ∈ S/c]
(see Proposition A.2).
Additional Distributions. We denote by Lap(σ) the
Laplace distribution whose mean is 0 and variance is 2σ 2 .
The χ2k -distribution, where k is referred to as the degrees of freedom of the distribution, is the distribution
over the l2 -norm squared of the sum of k independent
normal Gaussians. That is, given i.i.d X1 , . . . , Xk ∼
def
N (0, 1) it holds that ζ = (X1 , X2 , . . . , Xk ) ∼
2
N (00k , Ik×k ), and kζζ k ∼ χ2k . Existing tail bounds on
the hχ2k distribution (Laurent & Massart,
2000) give that
i
p
√
2
2
Pr kζζ k ∈ ( k ± 2 ln(2/ν)) ≥ 1 − ν. The Tk distribution, where k is referred to as the degrees of freedom of the distribution, denotes the distribution over the
reals created by independently sampling Z ∼ N (0, 1) and
known fact that Tk → N (0, 1), thus it is a common practice to apply Gaussian tail bounds to the Tk -distribution
when k is sufficiently large.
Differential Privacy. In this work, we deal with input in
the form of a n × d-matrix with each row bounded by a
l2 -norm of B. Two inputs A and A0 are called neighbors if
they differ on a single row.
Definition 2.1 ((Dwork et al., 2006a)). An algorithm ALG
which maps (n × d)-matrices into some range R is (, δ)differential privacy it holds that Pr[ALG(A) ∈ S] ≤
e Pr[ALG(A0 ) ∈ S] + δ for all neighboring inputs A and
A0 and all subsets S ⊂ R.
Background on OLS. For the unfamiliar reader, we give
here a very brief overview of the main points in OLS. Further details, explanations and proofs appear in Section A.3.
xi , yi )}ni=1 where ∀i, x i ∈
We are given n observations {(x
p
R and yi ∈ R. We assume the existence of β ∈ Rp s.t.
the label yi was derived by yi = β Tx i + ei where ei ∼
N (0, σ 2 ) independently (also known as the homoscedastic
Gaussian model). We use the matrix notation where X denotes the (n × p)- feature matrix and y denotes the labels.
We assume X has full rank.
The parameters of the model are therefore β and σ 2 , which
we set to discover. To that end, we minimize minz kyy −
Xzz k2 and have
β = (X T X)−1 X Ty = (X T X)−1 X T (Xββ + e) = β + X +e (1)
β̂
ζ = y − Xβ̂β = (Xββ + e ) − X(ββ + X +e ) = (I − XX + )ee (2)
And then for any coordinate j the t-value, which
def
β̂j −βj
q
is the quantity t(βj )
=
, is
ζk
kζ
−1
T
(X X)j,j · √n−p
distributed
according to Tn−p
I.e.,
h
i -distribution.
R
β and ζ satisfying t(βj ) ∈ S =
Pr β̂
PDFTn−p (x)dx
S
for any measurable S ⊂ R. Thus t(βj ) describes the
likelihood of any βj — for any z ∈ R we can now give
an estimation of how likely it is to have βj = z (which
is PDFTn−p (t(z))), and this is known as t-test for the
value z. In particular, given 0 < α < 1, we denote cα
as the number for which the interval (−cα , cα ) contains
a probability mass of 1 − α from the Tn−p -distribution.
And so we derive a corresponding confidence interval Iα
centered at β̂j where βj ∈ Iα with confidence of level of
1 − α.
def
Of particular importance is the quantity t0 = t(0) =
√
β̂j n−p
q
,since
(X T X)−1
j,j
kζζ k
if there is no correlation between xj
and y then the likelihood of seeing βˆj depends on the ratio
of its magnitude to its standard deviation. As mentioned
k→∞
earlier, since Tk → N (0, 1), then rather than viewing
Differentially Private Ordinary Least Squares
this t0 as sampled from a Tn−p -distribution, it is common
to think of t0 as a sample from a normal Gaussian N (0, 1).
This allows us to associate t0 with a p-value, estimating the
event “βj and β̂j have different signs.” Specifically, given
α ∈ (0, 1/2), we α-reject the null hypothesis if p0 < α.
R∞
2
Let τα be the number s.t. Φ(τα ) = τα √12π e−x /2 dx = α.
This means we α-reject the null hypothesis when |t0 | > τα .
We now lower bound the number of i.i.d sample points
needed in order to α-reject the null hypothesis. This bound
is our basis for comparison between standard OLS and the
differentially private version.5
Theorem 2.2. Fix any positive definite matrix Σ ∈ Rp×p
and any ν ∈ (0, 21 ). Fix parameters β ∈ Rp and σ 2
and a coordinate j s.t. βj 6= 0. Let X be a matrix whose n rows are i.i.d samples from N (00, Σ), and
β )i is sampled i.i.d from
y be a vector where yi − (Xβ
N (0, σ 2 ). Fix α ∈ (0, 1). Then w.p. ≥ 1 − α − ν we
have p
that OLS’s (1 − α)-confidence interval has length
O(cα σ 2 /(nσmin (Σ))) provided n ≥ C1 (p + ln(1/ν))
for some sufficiently large constant C1 . Furthermore,
there exists a constant C2 such that w.p. ≥ 1 − α − ν
OLS n
(correctly) rejects the null hypothesis provided
n ≥
o
2
c2α +τα
σ2
max C1 (p + ln(1/ν)), p + C2 β 2 · σmin (Σ) , where cα
j
Rc
is the number for which −cαα PDFTn−p (x)dx = 1 − α.
3. OLS over Projected Data
In this section we deal with the output of Algorithm 1
in the special case where Algorithm 1 outputs matrix
unaltered and so we work with RA.
To clarify, the setting is as follows. We denote A = [X; y ]
the column-wise concatenation of the (n × (d − 1))-matrix
X with the n-length vector y . (Clearly, we can denote
any column of A as y and any subset of the remaining
columns as the matrix X.) We therefore denote the output
RA = [RX; Ryy ] and for simplicity we denote M = RX
and p = d − 1. We denote the SVD decomposition of
X = U ΣV T . So U is an orthonormal basis for the columnspan of X and as X is full-rank V is an orthonormal basis
for Rp . Finally, in our work we examine the linear regression problem derived from the projected data. That is, we
denote
β = (X T RT RX)−1 (RX)T (Ryy ) = β + (RX)+ Ree
β̃
σ̃ 2 =
r
kζ̃ζ k2 , with ζ̃ζ =
r−p
√1 Ry
y
r
−
√1 (RX)β̃
β
r
(3)
(4)
We now give our main theorem, for estimating the t-values
β and σ̃.
based on β̃
5
Theorem 2.2 also illustrates how we “separate” the two
sources of privacy. In this case, ν bounds the probability of bad
events that depend to sampling the rows of X, and α bounds the
probability of a bad event that depends on the sampling of the y
coordinates.
Theorem 3.1. Let X be a (n × p)-matrix, and parameters β ∈ Rp and σ 2 are such that we generate the vector
β + e with each coordinate of e sampled indepeny = Xβ
dently from N (0, σ 2 ). Assume Algorithm 1 projects the
matrix A = [X; y ] without altering it. Fix ν ∈ (0, 1/2)
and r = p + Ω(ln(1/ν)). Fix coordinate j. Then we have
β and σ̃ 2 as in Equations (3)
that w.p. ≥ 1 − ν deriving β̃
β̃ −β
and (4), the pivot quantity t̃(βj ) = q Tj T j −1 has a
σ̃
(X R RX)j,j
distribution D satisfying e−a PDFTr−p (x) ≤ PDFD (x) ≤
ea PDFTr−p (e−a x) for any x ∈ R, where we denote a =
r−p
n−p .
The implications of Theorem 3.1 are immediate: all estimations one can do based on the t-values from the true data
X, y , we can now do based on t̃ modulo an approximation
r−p
factor of exp( n−p
). In particular, Theorem 3.1 enables us
β.
to deduce a corresponding confidence interval based on β̃
Corollary 3.2. In the same setting as in Theorem 3.1, w.p.
≥ 1 − ν we have the following. Fix any α ∈ (0, 12 ). Let c̃α
denote the number s.t. the interval (c̃α , ∞) contains α2 e−a
probability
mass q
of the Tr−p -distribution.
Then Pr[βj ∈
β̃j ± ea · c̃α · σ̃
(X T RT RX)−1
j,j ] ≥ 1 − α.
6
We compare the confidence interval of Corollary 3.2
to the confidence intervalqof the standard OLS model,
ζk
(X T X)−1
As R is a JLwhose length is cα √kζ
j,j .
n−p
matrix, known results regarding
the
JL
transform give
q
that kζ̃ζ k = Θ (kζζ k), and that (r − p)(X T RT RX)−1
j,j =
q
(X T X)−1
We therefore have that
Θ
j,j .
q
√ q
kζ̃ζ k r
√
σ̃ (X T RT RX)−1
=
(X T RT RX)−1
=
j,j
j,j
r−p
q
q
kζζ k
r·(n−p)
√
(X T X)−1
j,j . So for values of r
(r−p)2 · Θ
n−p
r
r−p
= Θ(1) we get that the confidence interval
q
of Theorem 3.1 is a factor of Θ cc̃αα n−p
r−p -larger than
the standard OLS confidence interval. Observe that when
α = Θ(1),
p which is the common case, the dominating
factor is (n − p)/(r − p). This bound intuitively makes
sense: we have contracted n observations to r observations, hence our model is based on confidence intervals
derived from Tr−p rather than Tn−p .
for which
In the supplementary material we give further discussion,
in which we compare our work to the more straight-forward
bounds one gets by “plugging in” Sarlos’ work (2006); and
we also compare ourselves to the bounds derived from alternative works in differentially private linear regression.
6
Moreover,
this interval is essentially optimal:
r−p
de-
note d˜α s.t the interval (d˜α , ∞) contains α2 e n−p probability
mass qof the Tr−p -distribution.
Then Pr[βj ∈
−1
T
T
˜
β̃j ± dα · σ̃ (X R RX)
] ≤ 1 − α.
j,j
Differentially Private Ordinary Least Squares
Rejecting the Null Hypothesis. Due to Theorem 3.1,
we can mimic OLS’ technique for rejecting the null hyβ̃
pothesis. I.e., we denote t̃0 = q T jT
and re−1
σ̃
(X R RX)j,j
ject the null-hypothesis if indeed the associated p̃0 , denotr−p
− n−p
ing p-value of the slightly truncated e
r−p
− n−p
t̃0 , is below
α·e
. Much like Theorem 2.2 we now establish a
lower bound on n so that w.h.p we end up (correctly) rejecting the null-hypothesis.
Theorem 3.3. Fix a positive definite matrix Σ ∈ Rp×p .
Fix parameters β ∈ Rp and σ 2 > 0 and a coordinate j
s.t. βj 6= 0. Let X be a matrix whose n rows are sampled
β )i
i.i.d from N (00p , Σ). Let y be a vector s.t. yi − (Xβ
is sampled i.i.d from N (0, σ 2 ). Fix ν ∈ (0, 1/2) and
α ∈ (0, 1/2). Then there exist constants C1 , C2 , C3
and C4 such that when we run Algorithm 1 over [X; y ]
with parameter r w.p. ≥ 1 − α − ν we (correctly)
reject the null hypothesis using p̃0 (i.e., Algorithm 1
returns matrix unaltered and we can estimate t̃0
−
r−p
and verify thatnindeed p̃0 < α · e n−p
o ) provided
2
σ 2 (c̃2α +τ̃α
)
r ≥ p + max C1 β 2 σmin (Σ) , C2 ln(1/ν) , and n ≥
j
n
o
w2
max r, C3 min{σmin
where
(Σ),σ 2 } , C4 p ln(1/ν)
r−p
c̃α , τ̃α defined s.t.
PrX∼Tr−p [X > c̃α /e n−p ] =
r−p
PrX∼N (0,1) [X > τ̃α /e n−p ] =
r−p
α − n−p
e
.
2
n 2 2
o
σ
(ΣA )
2
min n, B 4min
(n
−
ln(1/δ))
. This discussion culln(1/δ)
minates in the following corollary.
2
2
+τ̃α )
, we
= σβ 2(c̃σαmin
(Σ)
j
^
thus conclude that if n − p ≥ Ω LB
and n =
2.2
q
B 2 ln(1/δ)
^
Ω σ
· LB
2.2 , then the result of Theorem 3.3
min (ΣA )
n 2 2
o
σ
(ΣA )
2
holds by setting r = min n, B 4min
(n
−
ln(1/δ))
.
ln(1/δ)
^
Corollary 3.4. Denoting LB
2.2
It is interesting to note that when we know ΣA , we also
have a bound on B. Recall ΣA , the variance of the
x ◦ y). Since every sample is an independent
Gaussian (x
draw from N (00p+1 , ΣA ) then we have an upper bound of
B 2 ≤ log(np)σmax (ΣA ). So our lower bound on n (using
κ(ΣA ) to denote
number of ΣAq
) is given
by
the condition
κ(ΣA ) ln(1/δ)
^
^
n ≥ max Ω LB
· LB
.
2.2 , Ω̃
2.2
Observe, overall this result is similar in nature to many
other results in differentially private learning (Bassily et al.,
2014) which are of the form “without privacy, in order to
achieve a total loss of ≤ η we have a sample complexity
bound of some Nη ; and with differential
p privacy the sample complexity increases to Nη + Ω( Nη /).” However,
^
there’s a subtlety here worth noting. LB
2.2 is proportional
(ΣA )
to σmin1(ΣA ) but not to κ(ΣA ) = σσmax
. The additional
min (ΣA )
dependence on σmax follows from the fact that differential
privacy adds noise proportional to the upper bound on the
norm of each row.
3.1. Setting the Value of r, Deriving a Bound on n
Comparing the lower bound on n given by Theorem 3.3 to
the bound ofTheorem 2.2,
we have that the data-dependent
(c̃α +τ̃α )2 σ 2
bound of Ω β 2 σmin (Σ) should now hold for r rather than
j
n. Yet, Theorem 3.3 also introduces an additional depen2
w2
dency between n and r: we require n = Ω( w
σ 2 + σmin (Σ) )
(since otherwise we do not have σmin (A) w and Algorithm 1 might alter A before
projecting it) and by definition
p
w2 is proportional to r ln(1/δ)/. This is precisely the
focus of our discussion in this subsection. We would like
to set r’s value as high as possible — the larger r is, the
more observations we have in RA and the better our confidence bounds (that
depend on Tr−p ) are — while satisfying
√
n = Ω( ·min{σ2 ,σr min (Σ)} ).
Recall that if each sample point is drawn i.i.d x ∼
xi ◦ yi ) is sampled from
N (00p , Σ), then each sample (x
N (00p+1 , ΣA ) for Σ
defined
in
the proof of
A
Theorem 3.3,
β
Σ
Σβ
that is: ΣA =
. So, Theoβ T Σβ
β
β TΣ
σ 2 +β
2
2
+τ̃α )
rem 3.3 gives the lower bound r − p = Ω σβ 2(c̃σαmin
(Σ)
j
and the following
lower
bounds
on
n:
n
≥
r
and
√
B 2 ( r ln(1/δ)+ln(1/δ))
n = Ω
, which means r =
σmin (ΣA )
4. Projected Ridge Regression
We now turn to deal with the case that our matrix does not
pass the if-condition of Algorithm 1. In this case, the matrix is appended
with a d ×d-matrix which is wId×d . DeA
noting A0 =
we have that the algorithm’s
w · Id×d
output is RA0 . Similarly to before, we are going to denote
d = p + 1 and decompose A = [X; y ] with X ∈ Rn×p and
β + e and
y ∈ Rn , with the standard assumption of y = Xβ
ei sampled i.i.d from N (0, σ 2 ). We now need to introduce
some additional notation. We denote the appended matrix
and vectors X 0 and y 0 s.t. A0 = [X 0 ; y 0 ]. And so, using the
output RA0 of Algorithm 1, we solve the linear regression
problem derived from √1r RX 0 and √1r Ryy 0 . I.e., we set
β 0 = (X 0T RT RX 0 )−1 (RX 0 )T (Ryy 0 )
ζ 0 = √1r (Ryy 0 − RX 0β 0 )
(5)
Sarlos’ results (2006) regarding the Johnson Lindenstrauss transform give that, when R has sufficiently
many rows, solving the latter optimization problem gives
a good approximation for the solution of the optimization problem β R = argminz kyy 0 − X 0z k2 =
arg minz kyy − Xzz k2 + w2 kzz k2 . The latter problem is
Differentially Private Ordinary Least Squares
known as the Ridge Regression problem. Invented in the
60s (Tikhonov, 1963; Hoerl & Kennard, 1970), Ridge Regression is often motivated from the perspective of penalizing linear vectors whose coefficients are too large. It is also
often applied in the case where X doesn’t have full rank
or is close to not having full-rank: one can show that the
minimizer β R = (X T X + w2 Ip×p )−1 X Ty is the unique
solution of the Ridge Regression problem and that the RHS
is always well-defined.
While the solution of the Ridge Regression problem might
have smaller risk than the OLS solution, it is not known
how to derive t-values and/or reject the null hypothesis under Ridge Regression (except for using X to manipulate
β = (X T X)−1 X Ty and relying on OLS).
β R back into β̂
In fact, prior to our work there was no need for such analysis! For confidence intervals one could just use the standard
OLS, because access to X and y was given.
Therefore, much for the same reason, we are unable to derive t-values under projected Ridge Regression.7 Clearly,
there are situations where such confidence bounds simply
cannot be derived.Nonetheless, under additional assumptions about the data, our work can give confidence intervals
for βj , and in the case where the interval doesn’t intersect
the origin — assure us that sign(βj0 ) = sign(βj ) w.h.p.
This is detailed in the supplementary material.
To give an overview of our analysis, we first discuss a
β is fixed (i.e., the data is fixed
model where e = y − Xβ
and the algorithm is the sole source of randomness), and
β.
prove that in this model β 0 is as an approximation to β̂
so, in summary, in Section C we give conditions under
which the
q length of the interval I is dominated by the
ζ0k
c0α √kζr−p
r(M 0T M 0 )−1
j,j factor derived from Theorem 4.1.
5. Confidence Intervals for “Analyze Gauss”
In this section we analyze the “Analyze Gauss” algorithm
of Dwork et al (2014). Algorithm 2 works by adding random Gaussian noise to AT A, where the noise is symmetric
with each coordinate above
the diagonal
sampled i.i.d from
2
2
4 log(1/δ)
N (0, ∆ ) with ∆ = O B
. Using the same no2
tation for a sub-matrix of A as
[X; y ] as before, we
denote
the output of Algorithm 2 as
where c0α denotes the number such that (−c0α , c0α ) contains
1 − α mass of the Tr−p -distribution.
However, our goal remains to argue that βj0 serves as a good
approximation for βj . To that end, we combine the standard OLS confidence interval — which says that w.p. ≥
1 − α over the randomness of picking er
in the homoscedastic model we have |βj − β̂j | ≤ cα kζζ k
(X T X)−1
j,j
n−p
^
TX
X
TX
yg
we approximate β and kζζ k by βe =
g
Ty
X
.
Ty
yg
^
TX
X
−1
Thus,
g
Ty and
X
g
^
]
e + βe T X
e resp. We now argue
Ty − 2 y
TXβ
TX β
ζ k2 = yg
kζ
g
that it is possible to use βej and kζζ k2 to get a confidence
interval for βj under certain conditions.
Theorem 5.1. Fix α, ν ∈ (0, 12 ). Assume that there exists
p
η ∈ (0, 21 ) s.t. σmin (X T X) > ∆ p ln(1/ν)/η. Under the
homoscedastic model, given β and σ 2 , if we assume also
β k ≤ B and kβ̂
β k = k(X T X)−1 X Ty k ≤ B, then
that kβ
w.p. ≥ 1 − α − ν it holds that βj − βej is at most
s
−1
−2
p
^
^
TX
TX
O ρ·
+ ∆ p ln(1/ν) · X
X
ln(1/α)
j,j
n×p
β =
Theorem 4.1. Fix X ∈ R
and y ∈ R. Define β̂
X +y and ζ = (I − XX + )yy . Let RX 0 = M 0 and Ryy 0
denote the result of applying Algorithm 1 to the matrix A =
[X; y ] when the algorithm appends the data with a w · I
matrix. Fix a coordinate j and any α ∈ (0, 1/2). When
computing β 0 and ζ 0 as
have that w.p. ≥ 1 −
in (5), we q
r
0
0 ζ0
α it holds that β̂j ∈ βj ± cα kζ k r−p
· (M 0T M 0 )−1
j,j
j,j
r
+∆
−2
√
^
TX
X
j,j · ln(1/ν) · (B p + 1)
where ρ is w.h.p an upper bound on σ (details appear in
the Supplementary material).
β k ≤ B and kβ̂
βk ≤ B
Note that the assumptions that kβ
are fairly benign once we assume each row has bounded
l2 -norm. The key assumption is that X T X is well-spread.
Yet in the model where each row in X is sampled i.i.d
from N (00, Σ), this assumption merely
√ means that n is large
enough — namely, that n = Ω̃(
∆ p ln(1/ν)
η·σmin (Σ) ).
6. Experiment: t-Values of Output
— with
the confidence interval
q of Theorem 4.10 above,
q and denotkζζ k
kζζ k
−1
0
T
ing I = cα √n−p (X X)j,j + cα √r−p r(M 0T M 0 )−1
j,j
we have that Pr[|βj0 − βj | = O(I)] ≥ 1 − α. And
7
Note: The naı̈ve approach of using RX 0 and Ryy 0 to interpolate RX and Ryy and then apply Theorem 3.1 using these estimations of RX and Ryy ignores the noise added from appending the
matrix A into A0 , and therefore leads to inaccurate estimations of
the t-values.
Goal. We set to experiment with the outputs of Algorithms 1 and 2. While Theorem 3.1 guarantees that computing the t-value from the output of Algorithm 1 in the
matrix unaltered case does give a good approximation of the t-value – we were wondering if by computing
the t-value directly from the output we can (a) get a good
approximation of the true (non-private) t-value and (b) get
the same “higher-level conclusion” of rejecting the nullhypothesis. The answers are, as ever, mixed. The two main
Differentially Private Ordinary Least Squares
observations we do notice is that both algorithms improve
as the number of examples increases, and that Algorithm 1
is more conservative then Algorithm 2.
Setting. We tested both algorithms in two settings. The
first is over synthetic data. Much like the setting in Theorems 2.2 and 3.3, X was generated using p = 3 independent normal Gaussian features, and y was generated using
the homoscedastic model. We chose β = (0.5, −0.25, 0)
so the first coordinate is twice as big a the second but
of opposite sign, and moreover, y is independent of the
3rd feature. The variance of the label is also set to 1,
and so the variance of the homosedastic noise equals to
σ 2 = 1 − (0.5)2 − (−0.25)2 . The number of observations
n ranges from n = 1000 to n = 100000.
The second setting is over real-life data. We ran the two
algorithms over diabetes dataset collected over ten years
(1999-2008) taken from the UCI repository (Strack et al.,
2014). We truncated the data to 4 attributes: sex (binary),
age (in buckets of 10 years), number medications (numeric,
0-100), and a diagnosis (numeric, 0-1000). Naturally, we
added a 5th column of all-1 (intercept). Omitting any entry with missing or non-numeric values on these nine attributes we were left with N = 91842 entries, which we
shuffled and fed to the algorithm in varying sizes — from
n = 30, 000 to n = 90, 000. Running OLS over the entire
N observation yields β ≈ (14.07, 0.54, −0.22, 482.59),
and t-Values of (10.48, 1.25, −2.66, 157.55).
get fairly high (in magnitude) t-values. As the result, we
falsely reject the null-hypothesis based on the t-value of
Analyze Gauss quite often, even for large values of n. This
is shown in Figure 1b. Additional figures (including plotting the distribution of the t-value approximations) appear
in the supplementary material.
The results show that t-value approximations that do not
take into account the inherent randomness in the DPalgorithms lead to erroneous conclusions. One approach
would be to follow the more conservative approach we advocate in this paper, where Algorithm 1 may allow you to
get true approximation of the t-values and otherwise reject the null-hypothesis only based on the confidence interval (of Algorithm 1 or 2) not intersecting the origin. Another approach, which we leave as future work, is to replace the T -distribution with a new distribution, one that
takes into account the randomness in the estimator as well.
This, however, has been an open and long-standing challenge since the first works on DP and statistics (see (Vu
& Slavkovic, 2009; Dwork & Lei, 2009)) and requires we
move into non-asymptotic hypothesis testing.
The Algorithms. We ran a version of Algorithm 1 that uses
a DP-estimation of σmin , and finds the largest r the we can
use without altering the input, yet if this r is below 25 then
it does alter the input and approximates Ridge regression.
We ran Algorithm 2 verbatim. We set = 0.25 and δ =
10−6 . We repeated each algorithm 100 times.
(a) Synthetic data, coordinate β1
Results. We plot the t-values we get from Algorithms 1
and 2 and decide to reject the null-hypothesis based on tvalue larger than 2.8 (which corresponds to a fairly conservative p-value of 0.005). Not surprisingly, as n increases,
the t-values become closer to their expected value – the tvalue of Analyze Gauss is close to the non-private
p t-value
and the t-value from Algorithm 1 is a factor of nr smaller
as detailed above (see after Corollary 3.2). As a result,
when the null-hypothesis is false, Analyze Gauss tends to
produce larger t-values (and thus reject the null-hypothesis)
for values of n under which Algorithm 1 still does not reject, as shown in Figure 1a. This is exacerbated in real
data setting, where its actual least singular value (≈ 500) is
fairly small in comparison to its size (N = 91842).
However, what is fairly surprising is the case where the
null-hypothesis should not be rejected — since βj = 0
(in the synthetic case) or its non-private t-value is close
to 0 (in the real-data case). Here, the Analyze Gauss’ tvalue approximation has fairly large variance, and we still
(b) Synthetic data, coordinate β3
Figure 1. Correctly and Wrongly Rejecting the Null-Hypothesis
Differentially Private Ordinary Least Squares
Acknowledgements
The bulk of this work was done when the author was a
postdoctoral fellow at Harvard University, supported by
NSF grant CNS-123723; and also an unpaid collaborator
on NSF grant 1565387. The author wishes to wholeheartedly thank Prof. Salil Vadhan, for his tremendous help in
shaping this paper. The author would also like to thank
Prof. Jelani Nelson and the members of the “Privacy Tools
for Sharing Research Data” project at Harvard University (especially James Honaker, Vito D’Orazio, Vishesh
Karwa, Prof. Kobbi Nissim and Prof. Gary King) for many
helpful discussions and suggestions; as well as Abhradeep
Thakurta for clarifying the similarity between our result.
Lastly the author thanks the anonymous referees for many
helpful suggestions in general and for a reference to (Ullman, 2015) in particular.
References
Agresti, A. and Finlay, B. Statistical Methods for the Social
Sciences. Pearson P. Hall, 2009.
Bassily, R., Smith, A., and Thakurta, A. Private empirical
risk minimization: Efficient algorithms and tight error
bounds. In FOCS, 2014.
Blocki, J., Blum, A., Datta, A., and Sheffet, O. The
Johnson-Lindenstrauss transform itself preserves differential privacy. In FOCS, 2012.
Chaudhuri, Kamalika and Hsu, Daniel J. Convergence rates
for differentially private statistical estimation. In ICML,
2012.
Dwork, Cynthia, Su, Weijie, and Zhang, Li. Private false
discovery rate control. CoRR, abs/1511.03803, 2015.
Hoerl, A. E. and Kennard, R. W. Ridge regression: Biased
estimation for nonorthogonal problems. Technometrics,
12:55–67, 1970.
Kasiviswanathan, S., Lee, H., Nissim, K., Raskhodnikova,
S., and Smith, A. What can we learn privately? In FOCS,
2008.
Kifer, Daniel, Smith, Adam D., and Thakurta, Abhradeep.
Private convex optimization for empirical risk minimization with applications to high-dimensional regression. In
COLT, 2012.
Laurent, B. and Massart, P. Adaptive estimation of a
quadratic functional by model selection. The Annals of
Statistics, 28(5), 10 2000.
Ma, E. M. and Zarowski, Christopher J. On lower bounds
for the smallest eigenvalue of a hermitian positivedefinite matrix. IEEE Transactions on Information Theory, 41(2), 1995.
Muller, Keith E. and Stewart, Paul W. Linear Model Theory: Univariate, Multivariate, and Mixed Models. John
Wiley & Sons, Inc., 2006.
Pilanci, M. and Wainwright, M. Randomized sketches of
convex programs with sharp guarantees. In ISIT, 2014a.
Pilanci, Mert and Wainwright, Martin J. Iterative hessian
sketch: Fast and accurate solution approximation for
constrained least-squares. CoRR, abs/1411.0347, 2014b.
Chaudhuri, Kamalika, Monteleoni, Claire, and Sarwate,
Anand D. Differentially private empirical risk minimization. Journal of Machine Learning Research, 12, 2011.
Rao, C. Radhakrishna. Linear statistical inference and its
applications. Wiley, 1973.
Duchi, John C., Jordan, Michael I., and Wainwright, Martin J. Local privacy and statistical minimax rates. In
FOCS, pp. 429–438, 2013.
Rogers, Ryan M., Vadhan, Salil P., Lim, Hyun-Woo, and
Gaboardi, Marco. Differentially private chi-squared hypothesis testing: Goodness of fit and independence testing. In ICML, pp. 2111–2120, 2016.
Dwork, C. and Lei, J. Differential privacy and robust statistics. In STOC, 2009.
Dwork, Cynthia, Kenthapadi, Krishnaram, McSherry,
Frank, Mironov, Ilya, and Naor, Moni. Our data, ourselves: Privacy via distributed noise generation. In EUROCRYPT, 2006a.
Dwork, Cynthia, Mcsherry, Frank, Nissim, Kobbi, and
Smith, Adam. Calibrating noise to sensitivity in private
data analysis. In TCC, 2006b.
Dwork, Cynthia, Talwar, Kunal, Thakurta, Abhradeep, and
Zhang, Li. Analyze gauss - optimal bounds for privacy preserving principal component analysis. In STOC,
2014.
Rudelson, Mark and Vershynin, Roman. Smallest singular
value of a random rectangular matrix. Comm. Pure Appl.
Math, pp. 1707–1739, 2009.
Sarlós, T. Improved approx. algs for large matrices via random projections. In FOCS, 2006.
Sheffet, O. Private approximations of the 2nd-moment matrix using existing techniques in linear regression. CoRR,
abs/1507.00056, 2015. URL http://arxiv.org/
abs/1507.00056.
Smith, Adam D. Privacy-preserving statistical estimation
with optimal convergence rates. In STOC, pp. 813–822,
2011.
Differentially Private Ordinary Least Squares
Strack, B., DeShazo, J., Gennings, C., Olmo, J., Ventura,
S., Cios, K., and Clore, J. Impact of HbA1c measurement on hospital readmission rates: Analysis of 70,000
clinical database patient records. BioMed Research International, 2014:11 pages, 2014.
Tao, T. Topics in Random Matrix Theory. American Mathematical Soc., 2012.
Thakurta, Abhradeep and Smith, Adam. Differentially private feature selection via stability arguments, and the robustness of the lasso. In COLT, 2013.
Tikhonov, A. N. Solution of incorrectly formulated problems and the regularization method. Soviet Math. Dokl.,
4, 1963.
Uhler, Caroline, Slavkovic, Aleksandra B., and Fienberg, Stephen E. Privacy-preserving data sharing for
genome-wide association studies. Journal of Privacy
and Confidentiality, 2013. Available at: http://
repository.cmu.edu/jpc/vol5/iss1/6.
Ullman, J. Private multiplicative weights beyond linear
queries. In PODS, 2015.
Vu, D. and Slavkovic, A. Differential privacy for clinical
trial data: Preliminary evaluations. In ICDM, 2009.
Wang, Yue, Lee, Jaewoo, and Kifer, Daniel.
entially private hypothesis testing, revisited.
abs/1511.03376, 2015.
DifferCoRR,
Xi, B., Kantarcioglu, M., and Inan, A. Mixture of gaussian models and bayes error under differential privacy.
In CODASPY. ACM, 2011.
Zhou, S., Lafferty, J., and Wasserman, L. Compressed regression. In NIPS, 2007.
Differentially Private Ordinary Least Squares
A. Extended Introductory Discussion
A.2. Omitted Preliminary Details
Due to space constraint, a few details from the introductory parts (Sections 1,2) were omitted. We bring them in
this appendix. We especially recommend the uninformed
reader to go over the extended OLS background we provide in Appendix A.3.
Linear Algebra and Pseudo-Inverses. Given a matrix M
we denote its SVD as M = U SV T with U and V being
orthonormal matrices and S being a non-negative diagonal
matrix whose entries are the singular values of M . We use
σmax (M ) and σmin (M ) to denote the largest and smallest
singular value resp. Despite the risk of confusion, we stick
to the standard notation of using σ 2 to denote the variance
of a Gaussian, and use σj (M ) to denote the j-th singular
value of M . We use M + to denote the Moore-Penrose inverse of M , defined as M + = V S −1 U T where S −1 is a
−1
matrix with Sj,j
= 1/Sj,j for any j s.t. Sj,j > 0.
A.1. Proof Of Privacy of Algorithm 1
Theorem A.1. Algorithm 1 is (, δ)-differentially private.
Proof. The proof of the theorem is based on the fact
the Algorithm 1 is the result of composing the differentially private Propose-Test-Release algorithm of (Dwork &
Lei, 2009) with the differentially private analysis of the
Johnson-Lindenstrauss transform of (Sheffet, 2015).
More specifically, we use Theorem B.1 from (Sheffet,
2015) that states that given a matrix A whose all of its
2
singular
p values at greater than T (, δ) where T (, δ) =
2B 2
2r ln(4/δ) + 2 ln(4/δ) , publishing RA is (, δ)
differentially private for a r-row matrix R whose entries
sampled are i.i.d normal Gaussians. Since we have that all
of the singular values of A0 are greater than w (as specified
in Algorithm 1), outputting RA0 is (/2, δ/2)-differentially
private. The rest of the proof boils down to showing that
(i) the if-else-condition is (/2, 0)-differentially private and
that (ii) w.p. ≤ δ/2 any matrix A whose smallest singular value is smaller than w passes the if-condition (step 3).
If both these facts hold, then knowing whether we pass
the if-condition or not is (/2)-differentially private and
the output of the algorithm is (/2, δ)-differentially private,
hence basic composition gives the overall bound of (, δ)differential privacy.
To prove (i) we have that for any pair of neighboring matrices A and B that differ only on the i-th row, denoted a i and
T
T
b i resp., we have B T B − b ib T
i = A A − a ia i . Applying
Weyl’s inequality we have
b ib T
σmin (B T B) ≤ σmin (B T B − b ib T
i ) + σmax (b
i)
a ia T
b ib T
≤ σmin (AT A) + σmax (a
i ) + σmax (b
i)
≤ σmin (AT A) + 2B 2
2
hence |σmin (A)2 −σmin (B)2 | ≤ 2B 2 , so adding Lap( 4B )
is (/2)-differentially private.
To prove (ii), note that by standard tail-bounds on the
Laplace distribution we have that Pr[Z
<
−
4B 2 ln(1/δ)
δ
]
≤
.
Therefore,
w.p.
1
−
δ/2
it
holds
that
2
any matrix A that passes the if-test of the algorithm must
have σmin (A)2 > w2 . Also note that a similar argument shows that for any 0 < β < 1, any matrix A s.t.
2
passes the if-condition of
σmin (A)2 > w2 + 4B ln(1/β)
the algorithm w.p. 1 − β.
The Gaussian Distribution.
A univariate Gaussian N (µ, σ 2 ) denotes the Gaussian distribution
whose
mean is µ and variance σ 2 , with PDF(x) =
√
2
( 2πσ )−1 exp(− x−µ
bounds
2σ 2 ). Standard concentration
p
on Gaussians give that Pr[x > µ + 2σ ln(1/ν)] < ν
µ, Σ)
for any ν ∈ (0, 1e ). A multivariate Gaussian N (µ
for some positive semi-definite Σ denotes the multivariate Gaussian distribution where the mean of the
j-th coordinate is the µj and the co-variance between
coordinates j and k is Σj,k . The PDF of such Gaussian is defined only on the subspace colspan(Σ),
x) =
where for every x ∈ colspan(Σ) we have PDF(x
−1/2
1
rank(Σ)
T
+
˜
x − µ ) Σ (x
x − µ)
(2π)
· det(Σ)
exp − 2 (x
˜
and det(Σ)
is the multiplication of all non-zero singular values of Σ.
A matrix Gaussian distribution
denoted N (Ma×b , U, V ) has mean M , variance U
on its rows and variance V on its columns. For full
rank U and V it holds that PDFN (M,U,V ) (X) =
(2π)−ab/2 (det(U ))−b/2 (det(V ))−a/2
·
exp(− 21 trace V −1 (X − M )T U −1 (X − M ) ).
In
our case, we will only use matrix Gaussian distributions
with N (Ma×b , Ia×a , V ) and so each row in this matrix is
an i.i.d sample from a b-dimensional multivariate Gaussian
N ((M )j→ , V ).
We will repeatedly use the rules regarding linear operations
on Gaussians. That in, for any c, it holds that cN (µ, σ 2 ) =
µ, Σ) =
N (c · µ, c2 σ 2 ). For any C it holds that C · N (µ
µ, CΣC T ). And for any C is holds that N (M, U, V ) ·
N (Cµ
C = N (M C, U, C T V C). In particular, for any c (which
can be viewed as a b × 1-matrix) it holds that N (M, U, V ) ·
c = N (Mcc, U, c T V c ) = N (Mcc, c T V c · U ).
We will also require the following proposition.
2
Proposition A.2. Given σ 2 , λ2 s.t. 1 ≤ σλ2 ≤ c2 for some
constant c, let X and Y be two random Gaussians s.t. X ∼
N (0, σ 2 ) and Y ∼ N (0, λ2 ). It follows that 1c PDFY (x) ≤
PDFX (x) ≤ cPDFcY (x) for any x.
Corollary A.3. Under the same notation as in Proposition A.2, for any set S ⊂ R it holds that 1c Prx←Y [x ∈
S] ≤ Prx←X [x ∈ S] ≤ cPrx←cY [x ∈ S] =
Differentially Private Ordinary Least Squares
cPrx←Y [x ∈ S/c]
haustive account of OLS and we refer the interested reader
to (Rao, 1973; Muller & Stewart, 2006).
Proof. The proof is mere calculation.
xi , yi )}ni=1 where for all i we have
Given n observations {(x
p
x i ∈ R and yi ∈ R, we assume the existence of a pdimensional vector β ∈ Rp s.t. the label yi was derived by
yi = β Tx i + ei where ei ∼ N (0, σ 2 ) independently (also
known as the homoscedastic Gaussian model). We use the
matrix notation where X denotes the (n × p)-matrix whose
rows are x i , and use y , e ∈ Rn to denote the vectors whose
i-th entry is yi and ei resp. To simplify the discussion, we
assume X has full rank.
PDFX (x)
=
PDFcY (x)
r
2
x
c2 λ2 exp(− 2σ
2)
·
2
2
σ
exp(− 2cx2 λ2 )
x2 1
1
≤ c · exp( ( 2 2 − 2 )) ≤ c · exp(0) = c
2 c λ
σ
r
x2
)
PDFX (x)
λ2 exp(− 2σ
2
·
=
PDFY (x)
σ 2 exp(− x22 )
≥c
−1
2λ
x2 1
exp( 2 ( λ2 − σ12 ))
≥
exp(0)
c
= c−1
The parameters of the model are therefore β and σ 2 , which
we set to discover. To that end, we minimize minz kyy −
Xzz k2 and solve
β = (X T X)−1 X Ty = (X T X)−1 X T (Xβ
β +ee) = β +X +e
β̂
The Tk -Distribution.
The Tk -distribution, where k
is referred to as the degrees of freedom of the distribution, denotes the distribution over the reals created by independently sampling Z ∼ N (0, 1) and kζk2 ∼ χ2k ,
and taking the quantity √ Z 2 . Its PDF is given by
kζk /k
− k+1
2
2
x
. It
k
PDFTk (x) ∝ 1 +
is a known fact that
as k increases, Tk becomes closer and closer to a normal
Gaussian. The T -distribution is often used to determine
suitable bounds on the rate of converges, as we illustrate
in Section A.3. As the T -distribution is heavy-tailed, existing tail boundsp
on the T -distribution (which are of the
form:R if τν = C k((1/ν)2/k − 1) for some constant C
∞
then τν PDFTk (x)dx < ν) are often cumbersome to work
with. Therefore, in many cases in practice, it common to
assume ν = Θ(1) (most commonly, ν = 0.05) and use
existing tail-bounds on normal Gaussians.
Differential Privacy facts. It is known (Dwork et al.,
2006b) that if ALG outputs a vector in Rd such that for
any A and A0 it holds that kALG(A) − ALG(A0 )k1 ≤ B,
then adding Laplace noise Lap(1/) to each coordinate of
the output of ALG(A) satisfies -differential privacy. Similarly, (2006b) showed that if for any neighboring A and
A0 it holds that kALG(A) − ALG(A0 )k22 ≤ ∆2 then adding
Gaussian noise N (0, ∆2 · 2 ln(2/δ)
) to each coordinate of
2
the output of ALG(A) satisfies (, δ)-differential privacy.
Another standard result (Dwork et al., 2006a) gives that the
composition of the output of a (1 , δ1 )-differentially private
algorithm with the output of a (2 , δ2 )-differentially private
algorithm results in a (1 +2 , δ1 +δ2 )-differentially private
algorithm.
A.3. Detailed Background on Ordinary Least Squares
For the unfamiliar reader, we give a short description of the
model under which OLS operates as well as the confidence
bounds one derives using OLS. This is by no means an ex-
β
As e ∼ N (00n , σ 2 In×n ), it holds that β̂
∼
β , σ 2 (X T X)−1 ), or alternatively, that for every coorN (β
β ∼ N (βj , σ 2 (X T X)−1
dinate j it holds that β̂j = e T
j β̂
j,j ).
Hence we get
σ
β̂ −βj
q j
(X T X)−1
j,j
∼ N (0, 1). In addition, we de-
note the vector
β = (Xβ
β + e ) − X(β
β + X +e ) = (I − XX + )ee
ζ = y − X β̂
and since XX + is a rank-p (symmetric) projection matrix,
we have ζ ∼ N (0, σ 2 (I − XX + )). Therefore, kζζ k2 is
equivalent to summing the squares of (n − p) i.i.d samples
from N (0, σ 2 ). In other words, the quantity kζζ k2 /σ 2 is
sampled from a χ2 -distribution with (n − p) degrees of
freedom.
We sidetrack from the OLS discussion to give the following
β , as the next
bounds on the l2 -distance between β and β̂
claim shows.
Claim A.4. For any 0 < ν < 1/2, the following holds w.p.
≥ 1−ν over the randomness of the model (the randomness
over e )
β − β̂
β k2 = kX +e k2
kβ
= O σ 2 log(p/ν) · kX + k2F
2
+
(6)
2
β k = kβ
β + X ek
kβ̂
2
p
β k + σ · kX + kF · log(p/ν) )
= O( kβ
q
ln(1/ν)
2
2
1
ζ
kζ
k
−
σ
=
O(
n−p
n−p )
Proof. Since e ∼ N (00n , σ 2 In×n ) then X +e ∼
N (00n , σ 2 (X T X)−1 ). Denoting the SVD decomposition
(X T X)−1 = V SV T with S denoting the diagonal ma−2
−2
trix whose entries are σmax
(X), . . . , σmin
(X), we have
T +
2
that V X e ∼ N (00n , σ S). And so, each coordinate of V T X +e is distributed like an i.i.d Gaussian. So
Differentially Private Ordinary Least Squares
w.p. p≥ 1 − ν/2 non of these Gaussians is a factor of
O(σ ln(p/ν)) greater than its standard deviation. And so
w.p. ≥ 1 − ν/2P
it holds that kX +e k2 =PkV T X +e k2 ≤
−2
−2
2
Since
O(σ log(p/ν)
i σi (X) ).
i σi (X) =
T
−1
+
+ T
trace((X X) ) = trace(X (X ) ) = kX + k2F , the
bound of (6) is proven.
β k2 is an immediate corollary of (6) using
The bound on kβ̂
the triangle inequality.8 The bound on kζζ k2 follows from
tail bounds on the χ2n−p distribution, as detailed in Section 2.
β and ζ are
Returning to OLS, it is important to note that β̂
β depends solely on
independent of one another. (Note, β̂
X +e = (X + X)X +e = X + PU e , whereas ζ depends on
(I − XX + )ee = PU ⊥ e . As e is spherically symmetric, the
β is
two projections are independent of one another and so β̂
independent of ζ .) As a result of the above two calculations, we have that the quantity
def
tβ̂j (βj ) =
β̂j −βj
q
ζk
kζ
√
(X T X)−1
j,j · n−p
=
σ
β̂ −βj
q j
(X T X)−1
j,j
.
kζζ k
√
σ n−p
is distributed like a T -distribution with (n − p) degrees of
freedom. Therefore, we can compute an exact probability
estimation for this quantity. That is, for any measurable
S ⊂ R we have
i Z
h
β and ζ satisfying tβ̂j (βj ) ∈ S =
PDFTn−p (x)dx
Pr β̂
S
The importance of the t-value t(βj ) lies in the fact that it
can be fully estimated from the observed data X and y (for
any value of βj ), which makes it a pivotal quantity. Therefore, given X and y , we can use t(βj ) to describe the likelihood of any βj — for any z ∈ R we can now give an
estimation of how likely it is to have βj = z (which is
PDFTn−p (t(z))). The t-values enable us to perform multitude of statistical inferences. For example, we can say
which of two hypotheses is more likely and by how much
(e.g., we are 5-times more likely that the hypothesis βj = 3
is true than the hypothesis βj = 14 is true); we can compare between two coordinates j and j 0 and report we are
more confident that βj > 0 than βj 0 > 0; or even compare
among the t-values we get across multiple datasets (such
as the datasets we get from subsampling rows from a single dataset).
In particular, we can use t(βj ) to α-reject unlikely values
of βj . Given 0 < α < 1, we denote cα as the number for
which the interval (−cα , cα ) contains a probability mass
of 1 − α from the Tn−p -distribution. And so we derive a
8
Observe, though e is spherically symmetric, and is likely to
be approximately-orthogonal to β , this does not necessarily hold
for X +e which isn’t spherically symmetric. Therefore, we result
β using the triangle bound.
to bounding the l2 -norm of β̂
corresponding confidence interval Iα centered at β̂j where
βj ∈ Iα with confidence of level of 1 − α.
We comment as to the actual meaning of this confidence
interval. Our analysis thus far applied w.h.p to a vector y
derived according to this model. Such X and y will result in the quantity tβ̂j (βj ) being distributed like a Tn−p distribution — where βj is given as the model parameters
and β̂j is the random variable. We therefore have that guarantee that for X and
y derived according to this model,
the
q
2
def
ζ
kζ
k
−1
event Eα = β̂j ∈ βj ± cα · (X T X)j,j · n−p happens w.p. 1 − α. However, the analysis done over a given
dataset X and y (once y has been drawn) views the quantity tβ̂j (βj ) with β̂j given and βj unknown. Therefore the
event Eα either holds or does not hold. That is why the
alternative terms of likelihood or confidence are used, instead of probability. We have
q a confidence 2level of 1 − α
ζ
ˆ
that indeed βj ∈ βj ± cα · (X T X)−1 · kζ k , because this
j,j
n−p
event does happen in 1−α fraction of all datasets generated
according to our model.
Rejecting the Null Hypothesis. One important implication of the quantity t(βj ) is that we can refer specifically to
the hypothesis that βj = 0, called the null hypothesis. This
def
quantity, t0 = tβ̂j (0) =
√
β̂j n−p
q
,
kζζ k (X T X)−1
j,j
represents how
large is β̂j relatively to the empirical estimation of standard
deviation σ. Since it is known that as the number of degrees
of freedom of a T -distribution tends to infinity then the T distribution becomes a normal Gaussian, it is common to
think of t0 as a sample from a normal Gaussian N (0, 1).
This allows us to associate t0 with a p-value, estimating the
event “βj and β̂j have different signs.” Formally, we define
R∞
2
p0 = |t0 | √12π e−x /2 dx. It is common to reject the null
hypothesis when p0 is sufficiently small (typically, below
0.05).9
Specifically, given α ∈ (0, 1/2), we say we α-reject the
null hypothesis
R ∞ if p0 <2 α. Let τα be the number s.t.
Φ(τα ) = τα √12π e−x /2 dx = α. (Standard bounds
p
give that τα < 2 ln(1/α).) This means we α-reject
the null hypothesis
if t0 > τα or t0 < −τα , meaning if
q
|β̂j | > τα
kζζ k
√
(X T X)−1
j,j n−p .
ζ
We can now lower bound the number of i.i.d sample points
needed in order to α-reject the null hypothesis. This bound
will be our basis for comparison — between standard OLS
and the differentially private version.10
9
R ∞ Indeed, it is more accurate to associate with t0 the value
PDFTn−p (x)dx and check that this value is < α. However,
|t0 |
as most uses take α to be a constant (often α = 0.05), asymptotically the threshold we get for rejecting the null hypothesis are the
same.
10
This theorem is far from being new (except for maybe fo-
Differentially Private Ordinary Least Squares
Theorem A.5 (Theorem 2.2 restated.). Fix any positive
definite matrix Σ ∈ Rp×p and any ν ∈ (0, 12 ). Fix parameters β ∈ Rp and σ 2 and a coordinate j s.t. βj 6= 0.
Let X be a matrix whose n rows are i.i.d samples from
β )i is sampled
N (00, Σ), and y be a vector where yi − (Xβ
i.i.d from N (0, σ 2 ). Fix α ∈ (0, 1). Then w.p. ≥ 1 − ν
we have
p that the (1 − α)-confidence interval is of length
O(cα σ 2 /(nσmin (Σ))) provided n ≥ C1 (p + ln(1/ν))
for some sufficiently large constant C1 . Furthermore, there
exists a constant C2 such that w.p. ≥ 1 − α − ν we (correctly) reject the null hypothesis provided
(
)
σ 2 c2α + τα2
n ≥ max C1 (p + ln(1/ν)), C2 2 ·
βj σmin (Σ)
cα
denotes
the
number
for
which
PDF
(x)dx
=
1
−
α.
(If
we
are
content
T
n−p
−cα
with approximating Tn−p
p with a normal Gaussian than
one can set cα ≈ τα < 2 ln(1/α).)
Here
R cα
Proof. The discussion above
≥ 1−α
q shows that w.p.
2
−1 kζζ k
T
we have |βj − β̂j | ≤ cα (X X)j,j n−p ; and in orderqto α-reject the null hypothesis we must have |β̂j | >
kζζ k2
τα (X T X)−1
j,j n−p . Therefore, a sufficient condition for
OLS to α-reject the null-hypothesis
is to have n large
q
2
kζζ k
enough s.t. |βj | > (cα + τα ) (X T X)−1
j,j n−p . We therefore argue that w.p.≥ 1 − ν this inequality indeed holds.
ζ
We assume each row of X i.i.d vector x i ∼ N (00p , Σ), and
recall that according to the model kζζ k2 ∼ σ 2 χ2 (n − p).
Straightforward concentration bounds on Gaussians and on
the χ2 -distribution give:
√
(i) W.p. ≤ α it holds that kζζ k > σ ( n − p + 2 ln(2/α))).
(This is part of the standard OLS analysis.)
√
(ii) W.p.p≤ ν it holds that σmin (X T X) ≤ σmin (Σ)( n −
√
( p + 2 ln(2/ν)))2 . (Rudelson & Vershynin, 2009)
Therefore, due to the lower bound n = Ω(p +
ln(1/ν)), w.p.≥ 1 − ν − α we have that none
of
hold.
In such a case we have
q these events p
−1
T
(X X)j,j ≤
σmax ((X T X)−1 ) = O( √ 1
)
nσmin (Σ)
√
and kζζ k = O(σ n − p).
This implies that the
confidence
interval
of
level
1
− α has length of
q
q
2
−1 kζζ k
σ2
T
cα (X X)j,j · n−p = O cα nσmin
(Σ) ; and that in
order to α-reject
that
null-hypothesis
it
suffices to have
q
2
σ
|βj | = Ω (cα + τα ) nσmin
(Σ) . Plugging in the lower
bound on n, we see that this inequality holds.
We comment that for sufficiently large constants C1 , C2 ,
cusing on the setting where every row in X is sampled from an
i.i.d multivariate Gaussians), it is just stated in a non-standard
way, discussing solely the power of the t-test in OLS. For further
discussions on sample size calculations see (Muller & Stewart,
2006).
it holds that all the constants hidden in the O- and Ωnotations of the proof are close to 1. I.e., they are all
within the interval (1 ± η) for some small η > 0 given
C1 , C2 ∈ Ω(η −2 ).
B. Projecting the Data using Gaussian
Johnson-Lindenstrauss Transform
B.1. Main Theorem Restated and Further Discussion
Theorem B.1 (Theorem 3.1 restated.). Let X be a n × p
matrix, and parameters β ∈ Rp and σ 2 are such that
β + e with each coordiwe generate the vector y = Xβ
nate of e sampled independently from N (0, σ 2 ). Assume
σmin (X) ≥ C · w and that n is sufficiently large s.t.
all of the singular values of the matrix [X; y ] are greater
than C · w for some large constant C, and so Algorithm 1
projects the matrix A = [X; y ] without altering it, and publishes [RX; Ryy ].
Fix ν ∈ (0, 1/2) and r = p + Ω(ln(1/ν)). Fix coordinate
β , ζ̃ζ and σ̃ 2 as
j. Then w.p. ≥ 1 − ν we have that deriving β̃
follows
β
β̃
= (X T RT RX)−1 (RX)T (Ryy ) = β + (RX)+ Ree
ζ̃ζ
=
=
σ̃ 2
=
√1 Ry
y − √1r (RX)β̃
β
r
T
1
√
I − (RX)(X RT RX)−1 (RX)T ) Ree
r
r
kζ̃ζ k2
r−p
then the pivot quantity
β̃j − βj
t̃(βj ) =
σ̃
q
(X T RT RX)−1
j,j
has a distribution D satisfying e−a PDFTr−p (x) ≤
PDFD (x) ≤ ea PDFTr−p (e−a x) for any x ∈ R, where we
r−p
denote a = n−p
.
Comparison with Existing Bounds. Sarlos’ work (2006)
utilizes the fact that when r, the numbers of rows in R,
is large enough, then √1r R is a Johnson-Lindenstrauss
matrix. q
Specifically, given r and ν ∈ (0, 1) we denote
η = Ω(
p ln(p) ln(1/ν)
),
r
and so r = O( p ln(p)ηln(1/ν)
).
2
β = arg minz 1r kRXzz − Ryy k2 . In
Let us denote β̃
this setting, Sarlos’ work (Sarlós, 2006) (Theorem 12(3)) guarantees that w.p.q≥ 1 − ν we have
p log(p) log(1/ν)
β − β̃
β k2 ≤ ηkζζ k/σmin (X) = O
ζ
kβ̂
kζ
k
.
rσmin (X T X)
β − β̃
β k and using the
Naı̈vely bounding |β̂j − β̃j | ≤ kβ̂
β j − β j from Section A.311
confidence interval for β̂
11
Where we approximate cα , the tail bound of the Tn−p distribution with the tail
p bound on a Gaussian, i.e., use the approximation cα ≈ O( ln(1/α)).
Differentially Private Ordinary Least Squares
gives a confidence interval of q
level 1 − (α + ν)
cenp ln(p) log(1/ν)
tered at β̃j with length of O
kζζ k +
rσmin (X T X)
q
log(1/α)
ζk
O
(X T X)−1
=
j,j n−p kζ
q
p ln(p) log(1/ν)+log(1/α)
O
kζζ k .
This implies that
rσmin (X T X)
our confidence interval has decreased its degrees of
freedom from n − p to roughly r/p ln(p), and furthermore,
that it no longer depends on (X T X)−1
j,j but rather on
T
1/σmin (X X). It is only due to the fact that we rely on
Gaussians and by mimicking carefully the original proof
that we can deduce that the t̃-value has (roughly) r − p
degrees of freedom and depends solely on (X T X)−1
j,j .
(In the worst case, we have that (X T X)−1
j,j is proportional
to σmin (X T X)−1 , but it is not uncommon to have matrices
where the former is much larger than the latter.) As mentioned in the introduction, alternative techniques ((Chaudhuri et al., 2011; Bassily et al., 2014; Ullman, 2015)) for
finding a DP estimator β dp of the linear regression give a
β dp − β̂
β k = Õ(p/). Such
data-independent12 bound of kβ
bounds are harder to compare with the interval length given
by Corollary 3.2. Indeed, as we discuss in Section 3 under “Rejecting the null-hypothesis,” enough samples from
a multivariate Gaussian whose covariance-matrix is well
conditioned give a bound which is well below the worstupper bound of O(p/). (Yet, it is possible that these techniques also do much better on such “well-behaved” data.)
What the works of Sarlos and alternative works regrading
differentially private linear regression do not take into account are questions such as generating a likelihood for βj
nor do they discuss rejecting the null hypothesis.
B.2. Proof of Theorem 3.1
β and ζ̃ζ , where our goal
We now turn to our analysis of β̃
is to show that the distribution of the t̃-values as specified in Theorem 3.1 is well-approximated by the Tr−p distribution. For now, we assume the existence of fixed
β + e . (Later,
vectors β ∈ Rp and e ∈ Rn s.t. y = Xβ
we will return to the homoscedastic model where each coordinate of e is sampled i.i.d from N (0, σ 2 ) for some σ 2 .)
In other words, we first examine the case where R is the
sole source of randomness in our estimation. Based on the
assumption that e is fixed, we argue the following.
Claim B.2. In our model,
given X and
β
the output M
= RX, we have that β̃
∼
+
2
T
−1
N β + X e , kPU ⊥ e k (M M )
and
ζ̃ζ
∼
kPU ⊥ e k2
T
−1
T
N 0n,
(Ir×r − M (M M ) M ) .
Where
r
PU ⊥ denotes the projection operator onto the subspace
orthogonal to colspan(X); i.e., PU = XX + and
PU ⊥ = (Ir×r − XX + ).
12
In other words, independent of X, ζ .
Proof. The
matrix
R
is
sampled
from
N (0r×p , Ir×r , Ip×p ). Given X and RX = M , we
learn the projection of each row in R onto the subspace
spanned by the columns of X. That is, denoting u T as
the i-th row of R and v T as the i-th row of M , we have
that X Tu = v . Recall, initially u ∼ N (00n , In×n ) –
a spherically symmetric Gaussian. As a result, we can
denote u = PU u × PU ⊥ u where the two projections are
independent samples from N (00n , PU ) and N (00n , PU ⊥ )
resp. However, once we know that v = X Tu we have that
PU u = X(X T X)−1 X Tu = X(X T X)−1v so we learn
PU u exactly, whereas we get no information about PU ⊥
so PU ⊥ u is still sampled from a Gaussian N (00n , PU ⊥ ).
As we know for each row of R that u T PU = v T X + , we
therefore have that
R = RPU + RPU ⊥ = M X + + RPU ⊥
where RPU ⊥ ∼ N (0r×n , Ir×r , PU ⊥ ). From here on, we
just rely on the existing results about the linearity of Gaussians.
R ∼ N (M X + , Ir×r , PU ⊥ )
⇒ Ree ∼ N (M X +e , kPU ⊥ e k2 Ir×r )
⇒ M + Ree ∼ N (X +e , kPU ⊥ e k2 (M T M )−1 )
β = β + M + Ree implies β̃
β ∼ N (β
β +
so β̃
+
X e , kPU ⊥ e k2 (M T M )−1 ).
And as ζ̃ζ
=
T
−1
T
√1 (Ir×r
e
−
M
(M
M
)
M
)Re
then
we
r
kP
have ζ̃ζ
∼
N (00r , U r⊥
+
(Ir×r − M M )M = 0r×p .
e k2
(Ir×r − M M + )) as
Claim B.2 was based on the assumption that e is fixed.
However, given X and y there are many different ways to
β + e . However, the
assign vectors β and e s.t. y = Xβ
distributions we get in Claim B.2 are unique. To see that,
β and
recall Equations (1) and (2): β + X +e = X +y = β̂
PU ⊥ e = PU ⊥ y = (I − XX + )yy = ζ . We therefore have
2
β ∼ N (β̂
β , kζζ k2 (M T M )−1 ) and ζ̃ζ ∼ N (00n , kζζrk (I −
β̃
M M + )). We will discuss this further, in Section 4, where
we will not be able to better analyze the explicit distributions of our estimators. But in this section, we are able to
β and ζ̃ζ .
argue more about the distributions of β̃
So far we have considered the case that e is fixed, whereas
our goal is to argue about the case where each coordinate
of e is sampled i.i.d from N (0, σ 2 ). To that end, we now
switch to an intermediate model, in which PU e is sampled from a multivariate Gaussian while PU ⊥ e is fixed as
some arbitrary vector of length l. Formally, let Dl denote
the distribution where PU e ∼ N (0, σ 2 PU ) and PU ⊥ e is
fixed as some specific vector whose length is denoted by
kPU ⊥ ek = l.
Claim B.3. Under the same assumptions as in
Claim B.2, given that e ∼ Dl , we have that
Differentially Private Ordinary Least Squares
∼
N β , σ 2 (X T
X)−1 + l2 (M T M )−1
2
ζ̃ζ ∼ N 0 n , lr (I − M M + ) .
β
β̃
β = β + M + Ree = β + M + (M X + +
Proof. Recall, β̃
RPU ⊥ )ee = β + X +e + M + R(PU ⊥ e ). Now, under the
assumption e ∼ Dl we have that β is the sum of two independent Gaussians:
β , σ 2 X + · PU · (X + )T )
β + X +e ∼ N (β
β , σ 2 (X T X)−1 )
= N (β
∼ N (00r , kPU ⊥ e k2 Ir×r )
RPU ⊥ e
+ e
⇒ M Re ∼ N (00p , kPU ⊥ e k2 (M T M )−1 )
Summing the two independent Gaussians’ means and
β . Furthermore, in
variances gives the distribution of β̃
Claim B.2 we have
already
established
that
for any fixed
kPU ⊥ e k2
+
e we have ζ̃ζ ∼ N 0 n ,
(I − M M ) . Hence, for
r
2
e ∼ Dl we still have ζ̃ζ ∼ N 0 n , lr (I − M M + ) . (It is
easy to verify that the same chain of derivations is applicable when e ∼ Dl .)
Corollary B.4. Given that e ∼ Dl we have that β̃j ∼
−1
2
T
N (βj , σ 2 (X T X)−1
j,j + l (M M )j,j ) for any coordinate
j, and that kζ̃ζ k2 ∼
l2
r
· χ2r−p .
Proof. The corollary follows immediately from the fact
β , and from the definition of the χ2 that βj = e T
j β̃
distribution, as ζ̃ζ is a spherically symmetric Gaussian defined on the subspace colspan(M )⊥ of dimension r −
p.
To continue, we need the following claim.
Claim B.5. Given X and M = RX, and given that e ∼ Dl
β and ζ̃ζ are independent.
we have that β̃
β = β + X +e + M + R(PU ⊥ e ). And
Proof. Recall, β̃
so, given X, M and a specific vector PU ⊥ e we have
β depends on (i) the projection
that the distribution of β̃
of e on U = colspan(X) and on (ii) the projection of
each row in R onto Ũ = colspan(M ). The distribution of ζ̃ζ = √1r PŨ ⊥ Ree = √1r PŨ ⊥ (M X + + RPU ⊥ )ee =
√1 P ⊥ RPU ⊥ e depends on (i) the projection of e onto U ⊥
r Ũ
(which for the time being is fix to some specific vector of
length l) and on (ii) the projection of each row in R onto
Ũ ⊥ . Since PU e is independent from PU ⊥ e , and since for
any row u T of R we have that PŨ u is independent of PŨ ⊥ u ,
and since e and R are chosen independently, we have that
β and ζ̃ζ are independent.
β̃
Formally, consider any pair of coordinates β̃j and ζ̃k , and
we have
β̃j − βj
+
T
+
= eT
j X e + e j M (RPU ⊥ e )
ζ̃k
and
= eT
k PŨ ⊥ (RPU ⊥ e )
Recall, we are given X and M = RX. Therefore, we know
PU and PŨ . And so
Cov[β̃j , ζ̃k ]
= E[(β̃j − βj )(ζ̃k − 0)]
+
T
= E[eeT
j X e (RPU ⊥ e ) PŨ ⊥ e k ]
T
+
+ E[eeT
j M (RPU ⊥ e )(RPU ⊥ e ) PŨ ⊥ e k ]
+
ee T PU ⊥ ]E[RT ]PŨ ⊥ e k
= eT
j X E[e
T
+
+ eT
j M E[(RPU ⊥ e )(RPU ⊥ e ) ]PŨ ⊥ e k
+
eeT PU ⊥ ] (M X + )T + E[(RPU ⊥ )T ] PŨ ⊥ ek
= eT
j X E[e
+
+ eT
kPU ⊥ e k2 Ir×r PŨ ⊥ e k
jM
+
ee T PU ⊥ ](X + )T M T PŨ ⊥ e k + 0
= eT
j X E[e
+
+ l2 · e T
j M PŨ ⊥ e k
=0+0+0=0
β and ζ̃ζ are Gaussians, having their covariance = 0
And as β̃
implies independence.
β and ζ̃ζ are independent Gaussians
Having established that β̃
and specified their distributions, we continue with the proof
of Theorem 3.1. We assume for now that there exists some
small a > 0 s.t.
−1
−1
2
T
2
T
l2 (M T M )−1
j,j ≤ σ (X X)j,j + l (M M )j,j
2a
2
T
≤ e · l (M M )−1
j,j
(7)
Then, due to Corollary A.3, denoting the distributions N1 = N (0, l2 (M T M )−1
=
j,j ) and N2
−1
2
T
N (0, σ 2 (X T X)−1
+
l
(M
M
)
),
we
have
that
for
any
j,j
j,j
S ⊂ R it holds that13
e−a Prβ̃j ∼N1 [S] ≤ Prβ̃j ∼N2 [S] ≤ ea Prβ̃j ∼N1 [S/ea ]
(8)
More specifically, denote the function
t̃(ψ, kξξ k, βj )
ψ − βj
q
r
kξξ k r−p
(M T M )−1
j,j
q
r
ξ
kξ
k
.
r−p
ψ − βj
= q
l
l (M T M )−1
=
j,j
and observe that when we sample ψ, ξ independently s.t.
2
ξ k2 ∼ lr χ2r−p then
ψ ∼ N (βj , l2 (M T M )−1
j,j ) and kξ
t̃(ψ, kξξ k, βj ) is distributed like a T -distribution with r − p
13
In fact, it is possible to use standard techniques from differential privacy, and argue a similar result — that the probabilities
of any event that depends on some function f (βj ) under βj ∼ N1
and under βj ∼ N2 are close in the differential privacy sense.
Differentially Private Ordinary Least Squares
degrees of freedom. And so, for any τ > 0 we have that
under such way to sample ψ, ξ we have Pr[t̃(ψ, kξξ k, βj ) >
τ ] = 1 − CDFTr−p (τ ).
for any I = (τ1 , τ2 ) with τ1 < τ2 ≤ 0, and deduce
that the PDF of the function t̃(ψ, kξξ k, βj ) at x — where
−1
2
T
we sample ψ ∼ N (βj , l2 (M T M )−1
j,j + σ (X X)j,j )
For any τ ≥ 0 and for any non-negative real value z let Szτ
denote the suitable set of values s.t.
and kξξ k2 ∼ lr χ2r−p independently — lies in the range
e−a PDFTr−p (x), ea PDFTr−p (x/ea ) . And so, using
Corollary B.4 and Claim B.5, we have that when e ∼ Dl ,
the distributions of β̃j and kζ̃ζ k2 are precisely as stated
ξ k, βj ) > τ ]
Pr
ψ∼N (βj , l2 (M T M )−1 ) [t̃(ψ, kξ
j,j
l2
kξξ k2 ∼ r χ2r−p
Z∞
=
PDF l2
2
r χr−p
0
(z) ·
2
def
Pr
[Szτ ] dz
{ψ−βj ∼N (0, l2 (M T M )−1
j,j )}
above, and so we have that the distribution of t̃(βj ) =
t̃(β̃j , kζ̃ζ k, βj ) has a PDF that at the point x is “sandwiched”
between e−a PDFTr−p (x) and ea PDFTr−p (x/ea ).
Next, we aim to argue that this characterization of the
PDF of t̃(βj ) still holds when e ∼ N (00n , σ 2 In×n ).
r
That is, Szτ = τ · z r−p
(M T M )−1
,
∞
.
j,j
It would be convenient to think of e as a sample in
N (00n , σ 2 PU ) × N (00n , σ 2 PU ⊥ ). (So while in Dl we have
We now use Equation (8) (Since N (0, l2 (M T M )−1
)
is
j,j
PU e ∼ N (00n , σ 2 PU ) but PU ⊥ e is fixed, now both PU e and
precisely N1 ) to deduce that
PU ⊥ e are sampled from spherical Gaussians.) The reason
why the above still holds lies in the fact that t̃(βj ) does not
Prψ∼N (βj , l2 (M T M )−1 +σ2 (X T X)−1 ) [t̃(ψ, kξξ k, βj ) > τ ]
depend on l. In more details:
j,j
j,j
l2
kξξ k2 ∼ r χ2r−p
Pre ∼N (00n ,σ2 In×n ) t̃(βj ) ∈ I
Z ∞
Z
=
PDF l2 2 (z)
[Szτ ]dz = Pr
Pr
v )dvv
−1
−1
2
T
2
T
0n ,σ 2 In×n ) t̃(βj ) ∈ I | PU ⊥ e = v PDFPU ⊥ e (v
e ∼N (0
ψ − βj ∼ N (0, l (M M )j,j + σ (X X)j,j )
0
r χr−p
v
Z ∞
Z
[Szτ /ea ]dz =
≤ ea
Pr
PDF l2 2 (z)
Pr t̃(βj ) ∈ I | l = kvv k PDFPU ⊥ e (vv )dvv
2 (M T M )−1 )
χ
ψ−β
∼N
(0,
l
e
∼D
0
j
l
r r−p
j,j
v
!
Z ∞
Z
Z
(∗)
= ea
PDF l2
0
q
r
χ2r−p
(z)
Pr
ψ−βj ∼N (0,
a
[Szτ /e ]dz
l2 (M T M )−1
j,j )
j,j
= ea 1 −
l2
kξξ k2 ∼ r χ2r−p
CDFTr−p (τ /ea )
= ea
ξ k, βj ) > τ ]
Pr
ψ∼N (βj , l2 (M T M )−1 +σ 2 (X T X)−1 ) [t̃(ψ, kξ
j,j
> τ]
2
= e−a
l
kξξ k2 ∼ r χ2r−p
1 − CDFTr−p (τ )
In other words, we have just shown that for any interval I = (τ, ∞) with τ ≥ 0 we have that
ξ k, βj ) ∈ I]
Pr
ψ∼N (βj , l2 (M T M )−1 +σ 2 (X T X)−1 ) [t̃(ψ, kξ
j,j
PDFTr−p (z)dz
v
PDFPU ⊥ e (vv )dvv
Z
PDFTr−p (z)dz
I/ea
where the last transition is possible precisely because t̃ is
independent of l (or kvv k) — which is precisely what makes
this t-value a pivot quantity. The proof of the lower bound
is symmetric.
j,j
l2
kξξ k2 ∼ r χ2r−p
−a
[t̃(ψ, kξ
ξ k, βj )
e Pr
ψ∼N (βj , l2 (M T M )−1
j,j )
!Z
Z
I/ea
=
where the equality (∗) follows from the fact that
τ /c
Sz for any c > 0, since it is a non-negative interval.
Analogously, we can also show that
≥
e
a
Szτ /c
=
PDFPU ⊥ e (vv )dvv
PDFTr−p (z)dz
I/ea
v
ξ k, βj ) > τ /ea ]
= ea Pr
ψ∼N (βj , l2 (M T M )−1 ) [t̃(ψ, kξ
ea
≤
l2
kξξ k2 ∼ r χ2r−p
a
j,j
R
is lower bounded by e PDFTr−p (z)dz and upper
I
R
bounded by ea
PDFTr−p (z)dz. We can now repeat
I/ea
the same argument for I = (τ1 , τ2 ) with 0 ≤ τ1 <
τ2 (using an analogous definition of Szτ1 ,τ2 ), and again
To conclude, we have shown that if Equation (7)
holds, then for every
interval
I ⊂ R we have that
Pre ∼N (00n ,σ2 In×n ) t̃(βj ) ∈ I
is
lower
bounded
by e−a Prz∼Tr−p [z ∈ I] and upper bounded by
ea Prz∼Tr−p [z ∈ (I/ea )].
So to conclude the proof
of Theorem 3.1, we need to show that w.h.p such a as in
Equation (7) exists.
Claim B.6. In the homoscedastic model with Gaussian
noise, if both n and r satisfy n, r ≥ p + Ω(log(1/ν)), then
−1
−1
2
T
2
T
we have that σ 2 (X T X)−1
j,j +l (M M )j,j ≥ l (M M )j,j
and
2(r−p)
−1
−1
2
T
2
T
σ 2 (X T X)−1
j,j +l (M M )j,j ≤ (1+ n−p )·l (M M )j,j
2(r−p)
n−p , Theorem 3.1 now follows
Using (1 + 2(r−p)
n−p ) ≤ e
r−p
from plugging a = n−p to our above discussion.
Differentially Private Ordinary Least Squares
Proof. The lower bound is immediate from non-negativity
T
−1/2
of σ 2 and of (X T X)−1
e j k2 . We therej,j = k(X X)
fore prove the upper bound.
parameter r w.p. ≥ 1 − ν we correctly α-reject the null hypothesis using p̃0 (i.e., w.p. ≥ 1 − ν Algorithm 1 returns
matrix unaltered and we can estimate t̃0 and verify that
First, observe that l2 = kPU ⊥ e k2 is sampled from σ 2 ·χ2n−p
as U ⊥ is of dimension n − p. Therefore, it holds that w.p.
≥ 1 − ν/2 that
indeed p̃0 < α · e
σ2
√
n−p−
p
r−p
− n−p
) provided
(
σ 2 (c̃2 + τ̃α2 )
r ≥ p + max C1 2 α
, C2 ln(1/ν)
βj σmin (Σ)
2
2 ln(2/ν) ≤ l2
)
and
2
and assuming n > p+100 ln(2/ν) we therefore have σ ≤
4
2
3(n−p) l .
Secondly, we argue that when r > p + 300 ln(4/ν)
we have that w.p.
≥ 1 − ν/2 it holds that
−1
3
T
T T
(X
X)
≤
(r
−
p)(X
R RX)−1
j,j
j,j . To see this, first
4
observe that by picking R ∼ N (0r×n , Ir×r , In×n ) the
distribution of the product RX ∼ N (0r×d , Ir×r , X T X)
is identical to picking Q ∼ N (0r×d , Ir×r , Id×d )
and taking the product Q(X T X)1/2 .
Therefore,
the
distribution
of
(X T RT RX)−1
is
T
1/2 T
T
1/2 −1
identical to
(X X) Q Q(X X)
=
(X T X)−1/2 (QT Q)−1 (X T X)−1/2 .
Denoting
v = (X T X)−1/2e j we have kvv k2 = (X T X)−1
j,j .
Claim A.1 from (Sheffet, 2015) gives that w.p. ≥ 1 − ν/2
we have
−1
T
1/2 T
(r − p) · e T
Q Q(X T X)1/2
ej
j (X X)
1
= v T ( r−p
QT Q)−1v ≥ 43 v Tv = 34 (X T X)−1
j,j
w2
n ≥ max r, C3
, C4 (p + ln(1/ν))
min{σmin (Σ), σ 2 }
16l2 (r−p)
(X T RT RX)−1
j,j
n−p
−1
2
T T
≤ 2(r−p)
l
(X
R
RX)
j,j
n−p
and as we denote M = RX we are done.
We comment that our analysis in the proof of Claim B.6
implicitly assumes r n (as we do think of the projection R as dimensionality reduction), and so the ratio
r−p
n−p is small. However, a similar analysis holds for r
which is comparable to n — in which we would argue that
−1
2
T
σ 2 (X T X)−1
j,j +l (M M )j,j
σ 2 (X T X)−1
∈ [1, 1 + η] for some small η.
PDFN (0,1) (x)dx =
the
numbers
r−p
α − n−p
e
2
=
r−p
α − n−p
2e
s.t.
and
resp.
Proof. First we need to use the lower bound on n to show
that indeed Algorithm 1 does not alter A, and that various
quantities are not far from their expected values. Formally,
we claim the following.
Proposition B.8. Under the same lower bounds on n and r
as in Theorem 3.3, w.p. 1−α−ν we have that Theorem 3.1
holds and also that
r−p
2
2
ζ̃ζ k2 = Θ( r−p
r kPU ⊥ e k ) = Θ( r (n − p)σ )
and
−1
T
1
(X T RT RX)−1
j,j = Θ( r−p (X X)j,j )
Proof of Proposition B.8. First, we need to argue that we
2
have enough samples as to have the gap σmin
([X; y]) − w2
sufficiently large.
Since x i ∼ N (0, Σ), and yi = β Tx i + ei with ei ∼
xi ◦ yi ) is also
N (0, σ 2 ), we have that the concatenation (x
xi ] +
sampled from a Gaussian. Clearly, E[yi ] = β T E[x
β Txi +
E[ei ] = 0. Similarly, E[xi,j yi ] = E[xi,j · (β
β )j and E[yi2 ] = E[e2i ] + E[kXβ
β k2 ] = σ 2 +
ei )] = (Σβ
β T X T Xβ
β ] = σ 2 + β T Σβ
β . Therefore, each row of A is
E[β
an i.i.d sample of N (00p+1 , ΣA ), with
B.3. Proof of Theorem 3.3
Theorem B.7 (Theorem 3.3 restated.). Fix a positive definite matrix Σ ∈ Rp×p . Fix parameters β ∈ Rp and σ 2 > 0
and a coordinate j s.t. βj 6= 0. Let X be a matrix whose n
rows are sampled i.i.d from N (00p , Σ). Let y be a vector s.t.
β )i is sampled i.i.d from N (0, σ 2 ). Fix ν ∈ (0, 1/2)
yi −(Xβ
and α ∈ (0, 1/2). Then there exist constants C1 , C2 , C3
and C4 such that when we run Algorithm 1 over [X; y ] with
denote
r−p
τ̃α /e n−p
Combining the two inequalities we get:
≤
τ̃α
PDFTr−p (x)dx
r−p
c̃α /e n−p
R∞
which implies the required.
σ 2 (X T X)−1
j,j
c̃α ,
where
R∞
ΣA =
Σ
β TΣ
β
Σβ
β T Σβ
β
σ 2 +β
Denote λ2 = σmin (Σ). Then, to argue that σmin (ΣA )
is large we use the lower bound from (Ma & Zarowski,
1995) (Theorem 3.1) combining with some simple
arithmetic manipulations to deduce that σmin (ΣA ) ≥
min{σmin (Σ), σ 2 }.
Differentially Private Ordinary Least Squares
r−p
Having established a lower bound on σmin (ΣA ), it follows that with n = Ω(p ln(1/ν)) i.i.d draws from
N (00p+1 , ΣA ) we have w.p. ≤ ν/4 that σmin (AT A) =
o(n) · min{σmin (Σ), σ 2 }. Conditioned on σmin (AT A) =
Ω(nσmin (ΣA )) = Ω(w2 ) being large enough, we have that
w.p. ≤ ν/4 over the randomness of Algorithm 1 the matrix
A does not pass the if-condition and the output of the algorithm is not RA. Conditioned on Algorithm 1 outputting
RA, and due to the lower bound r = p + Ω(ln(1/ν)),
we have that the result of Theorem 3.1 does not hold w.p.
≤ α + ν/4. All in all we deduce that w.p. ≥ 1 − α − 3ν/4
the result of Theorem 3.1 holds. And since we argue Theorem 3.1 holds, then the following two bounds that are used
in the proof14 also hold:
= Ω(e n−p (c̃α + τ̃α )σ̃
(c̃α +τ̃α )2 σ 2
βj2 σmin (Σ)
C. Projected Ridge Regression
In this section we deal with the case that our matrix does
not pass the if-condition of Algorithm 1. In this case, the
matrix is appended
with a d× d-matrix which is wId×d .
A
0
Denoting A =
we have that the algorithm’s
w · Id×d
output is RA0 .
Similarly to before, we are going to denote d = p + 1 and
decompose A = [X; y ] with X ∈ Rn×p and y ∈ Rn , with
β + e and ei sampled
the standard assumption of y = Xβ
i.i.d from N (0, σ 2 ).15 We now need to introduce some additional notation. We denote the appended matrix and vectors X 0 and y 0 s.t. A0 = [X 0 ; y 0 ]. Meaning:
X
X 0 = wIp×p
0T
p
2
kPU ⊥ e k = Θ((n − p)σ )
Lastly, in the proof of Theorem 3.1 we argue that
for a given PU ⊥ e the length kζ̃ζ k2 is distributed like
kPU ⊥ e k2 2
χr−p .
r
Appealing again to the fact that r = p +
Ω(ln(1/ν) we have that w.p. ≥ ν/4 it holds that kζ̃ζ k2 >
kP
(X T RT RX)−1
j,j ))
which, given the lower bound r = p + Ω
indeed holds.
−1
T
1
(X T RT RX)−1
j,j = Θ( r−p (X X)j,j )
2
q
e k2
2(r − p) U r⊥ . Plugging in the value of kPU ⊥ e k2 concludes the proof of the proposition.
and
Based on Proposition B.8, we now show that we indeed
reject the null-hypothesis (as we should). When Theorem 3.1 holds, reject the null-hypothesis iff p̃0 < α ·
e
r−p
− n−p
y
e
def
β = X 0β + e 0
y 0 = 0 p = X 0β + −wβ
w
w
r−p
which holds iff |t̃0 | > e n−p τ̃α . This implies
r−p
And so we respectively denote R = [R1 ; R2 ; R3 ] with
R1 ∈ Rr×n , R2 ∈ Rr×p and R3 ∈ Rr×1 (so R3 is a
vector denoted as a matrix). Hence:
n−p τ̃ ·
we
α
q reject that null-hypothesis when |β̃j | > e
−1
σ̃ (X T RT RX)j,j ). Note that this bound is based
onCorollary 3.2 that determines
that |β̃j − βj | =
q
r−p
−1
T
T
O e n−p c̃α · σ̃ (X R RX)j,j ) . And so we have that
M 0 = RX 0 = R1 X + wR2
and
w.p. ≥ 1 − ν we α-reject the null hypothesis when it holds
q
r−p
n−p (c̃ +
that |βj | > 3(c̃α + τ̃α )· σ̃ (X T RT RX)−1
)
≥
e
α
j,j
q
τ̃α )σ̃ (X T RT RX)−1
j,j ) (due to the lower bound n ≥ r).
β +R1e +wR3
Ryy 0 = RX 0β +Ree0 = R1y +wR3 = R1 Xβ
Based on the bounds stated above we have that
q
q
q
√
√
r
r
σ̃ = kζ̃ζ k r−p
= Θ(σ n − p r−p
r
r−p ) = Θ(σ n − p)
1
r−p
·
1
nσmin (Σ)
And so, a sufficient condition for rejecting the nullhypothesis is to have
r
n−p q
1
|βj |
= Ω (c̃α + τ̃α )σ
· nσmin (Σ)
r−p
14
β0
= arg min 1r kRyy 0 − RX 0z k2
z
= (X 0T RT RX 0 )−1 (RX 0 )T (Ryy 0 )
and that
−1
T
1
(X T RT RX)−1
j,j = Θ( r−p (X X)j,j ) = O
And so, using the output RA0 of Algorithm 1, we solve
the linear regression problem derived from √1r RX 0 and
√1 Ry
y 0 . I.e., we set
r
More accurately, both are bounds shown in Claim B.6.
Sarlos’ results (2006) regarding the Johnson Lindenstrauss
transform give that, when R has sufficiently many rows,
solving the latter optimization problem gives a good approximation for the solution of the optimization problem
β R = arg minz kyy 0 − X 0z k2 = arg minz kyy − Xzz k2 + w2 kzz k2
15
Just as before, it is possible to denote any single column as y
and any subset of the remaining columns as X.
Differentially Private Ordinary Least Squares
The latter problem is known as the Ridge Regression problem. Invented in the 60s (Tikhonov, 1963; Hoerl & Kennard, 1970), Ridge Regression is often motivated from
the perspective of penalizing linear vectors whose coefficients are too large. It is also often applied in the case
where X doesn’t have full rank or is close to not having full-rank. That is because the Ridge Regression problem is always solvable. One can show that the minimizer
β R = (X T X + w2 Ip×p )−1 X Ty is the unique solution of
the Ridge Regression problem and that the RHS is always
defined (even when X is singular).
The original focus of Ridge Regression is on penalizing
β R for having large coefficients. Therefore, Ridge Regression actually poses a family of linear regression problems: minz ky − Xzz k + λkzz k2 , where one may set λ to be
any non-negative scalar. And so, much of the literature on
Ridge Regression is devoted to the art of fine-tuning this
penalty term — either empirically or based on the λ that
β R ] − β k2 + Var(β
β R ).16 Here we
yields the best risk: kE[β
propose a fundamentally different approach for the choice
of the normalization factor — we set it so that solution of
the regression problem would satisfy (, δ)-differential privacy (by projecting the problem onto a lower dimension).
While the solution of the Ridge Regression problem might
have smaller risk than the OLS solution, it is not known
how to derive t-values and/or reject the null hypothesis under Ridge Regression (except for using X to manipulate
β = (X T X)−1 X Ty and relying on OLS).
β R back into β̂
In fact, prior to our work there was no need for such analysis! For confidence intervals one could just use the standard
OLS, because access to X and y was given.
Therefore, much for the same reason, we are unable to derive t-values under projected Ridge Regression.17 Clearly,
there are situations where such confidence bounds simply
cannot be derived. (Consider for example the case where
X = 0n×p and y is just i.i.d draws from N (0, σ 2 ), so
obviously [X; y] gives no information about β .) Nonetheless, under additional assumptions about the data, our work
can give confidence intervals for βj , and in the case where
the interval doesn’t intersect the origin — assure us that
sign(βj0 ) = sign(βj ) w.h.p.
Clearly, Sarlos’ work (2006) gives an upper bound on the
β 0 −β
β R k. However, such distance bound doesn’t
distance kβ
come with the coordinate by coordinate confidence guarantee we would like to have. In fact, it is not even clear
β 0 ] = β R (though it is obvious
from Sarlos’ work that E[β
0T T
0 βR
to see that E[(X R RX )]β = E[(RX 0 )T Ryy 0 ]). Here,
β 0 ] = β̂
β which, more often than not, does
we show that E[β
R
not equal β .
Comment about notation. Throughout this section we assume X is of full rank and so (X T X)−1 is well-defined. If
X isn’t full-rank, then one can simply replace any occurrence of (X T X)−1 with X + (X + )T . This makes all our
formulas well-defined in the general case.
C.1. Running OLS on the Projected Data
In this section, we analyze the projected Ridge Regression,
under the assumption (for now) that e is fixed. That is, for
now we assume that the only source of randomness comes
from picking the matrix R = [R1 ; R2 ; R3 ]. As before, we
analyze the distribution over β 0 (see Equation (9)), and the
value of the function we optimize at β 0 . Denoting M 0 =
RX 0 , we can formally express the estimators:
β0
ζ0
= (M 0T M 0 )−1 M 0T Ryy 0
= √1r (Ryy 0 − RX 0β 0 )
(9)
(10)
β +ee for a fixed e , and given
Claim C.1. Given that y = Xβ
X and M 0 = RX 0 = R1 X + wR2 we have that
β 0 ∼ N β + X +e ,
β + X +e k2 + 1) + kPU ⊥ e k2 )(M 0T M 0 )−1
(w2 (kβ
ζ 0 ∼ N 0r ,
β +X +e k2 +1)+kPU ⊥ e k2
w2 (kβ
0
0+
(I
−
M
M
)
r×r
r
and furthermore, β 0 and ζ 0 are independent of one another.
Proof. First, we write β 0 and ζ 0 explicitly, based on e and
projection matrices:
β0
ζ0
= (M 0T M 0 )−1 M 0T Ryy 0
β + M 0+ (R1e + wR3 )
= M 0+ (R1 X)β
= √1r (Ryy 0 − RX 0β 0 )
= √1r (Ir×r − M 0 M 0+ )Ree0
= √1r PU 0⊥ (R1e − wR2β + wR3 )
with U 0 denoting colspan(M 0 ) and PU 0⊥ denoting the projection onto the subspace U 0⊥ .
16
Ridge Regression, as opposed to OLS, does not yield an unβ R ] 6= β .
biased estimator. I.e., E[β
17
Note: The naı̈ve approach of using RX 0 and Ryy 0 to interpolate RX and Ryy and then apply Theorem 3.1 using these estimations of RX and Ryy ignores the noise added from appending the
matrix A into A0 , and it is therefore bound to produce inaccurate
estimations of the t-values.
Again, we break e into an orthogonal composition: e =
PU e + PU ⊥ e with U = colspan(X) (hence PU = XX + )
and U ⊥ = colspan(X)⊥ . Therefore,
β0
β + M 0+ (R1 XX +e + R1 PU ⊥ e + wR3 )
= M 0+ (R1 X)β
0+
β + X +e ) + M 0+ (R1 PU ⊥ e + wR3 )
= M (R1 X)(β
Differentially Private Ordinary Least Squares
whereas ζ 0 is essentially
√1 (Ir×r
r
(∗)
− M 0 M 0+ )(R1 XX +e + R1 PU ⊥ e − wR2β + wR3 )
√1 (Ir×r
r
=
0
−M M
0+
)·
β + wR3 )
(R1 XX e + R1 PU ⊥ e + (M 0 − wR2 )β
+
=
√1 (Ir×r
r
− M 0 M 0+ )·
β + X +e k2 + 1) + kPU ⊥ e k2 )PU 0⊥
I.e., N 0 r , 1r (w2 (kβ
0
as PU 0⊥ M = 0r×r .
Finally, observe that β 0 and ζ 0 are independent as the former depends on the projection of the spherical Gaussian
R1 X(β + X +e ) + R1 PU ⊥ e + wR3 on U 0 , and the latter
depends on the projection of the same multivariate Gaussian on U 0⊥ .
β + X +e ) + R1 PU ⊥ e + wR3 )
(R1 X(β
where equality (∗) holds because (I − M 0 M 0+ )M 0v = 0
for any v .
We now aim to describe the distribution of R given that we
know X 0 and M 0 = RX 0 . Since
M 0 = R1 X + wR2 + 0 · R3 = R1 X(X + X) + wR2
= (R1 PU )X + wR2
then M 0 is independent of R3 and independent of R1 PU ⊥ .
Therefore, given X and M 0 the induced distribution over
R3 remains R3 ∼ N (00r , Ir×r ), and similarly, given X and
M 0 we have R1 PU ⊥ ∼ N (0r×n , Ir×r , PU ⊥ ) (rows remain
independent from one another, and each row is distributed
like a spherical Gaussian in colspan(X)⊥ ). And so, we
have that R1 X = R1 PU X = M 0 − wR2 , which in turn
implies:
R1 X ∼ N M 0 , Ir×r , w2 · Ip×p
Observe that Claim C.1 assumes e is given. This may seem
somewhat strange, since without assuming anything about
e there can be many combinations of β and e for which
β + e . However, we always have that β + X +e =
y = Xβ
+
β . Similarly, it is always the case the PU ⊥ e =
X y = β̂
β and ζ in
(I − XX + )yy = ζ . (Recall OLS definitions of β̂
Equation (1) and (2).) Therefore, the distribution of β 0 and
ζ 0 is unique (once y is set):
β , (w2 (kβ̂
β k2 + 1) + kζζ k2 )(M 0T M 0 )−1
β 0 ∼ N β̂
!
2 β 2
2
ζ
w
(k
β̂
k
+
1)
+
kζ
k
ζ 0 ∼ N 0r ,
(Ir×r − M 0 M 0+ )
r
And so for a given dataset [X; y ] we have that β 0 serves as
β.
an approximation for β̂
multiplying this random matrix with a vector, we get
An immediate corollary of Claim C.1 is that for
any fixed e it holds that the quantity t0 (βj ) =
β +X +e ) ∼ N (M 0β + M 0 X +e , w2 kββ + X +e k2 Ir×r )
R1 X(β
kζζ 0 k
and multiplying this random vector with a matrix we get
β +X +e ) ∼ N (β + X +e ,
M 0+ R1 X(β
β + X +e k2 (M 0T M )−1 )
w2 kβ
I.e.,
β +X +e ) ∼ kβ
β +X +e k·N (u
u, w2 (M 0T M )−1 )
M 0+ R1 X(β
where u denotes a unit-length vector in the direction of β +
X +e .
Similar to before we have
RPU ⊥ ∼ N (0r×n , Ir×r , PU ⊥ )
⇒ M 0+ (RPU ⊥ e) ∼ N (00d , kPU ⊥ ek2 (M 0T M 0 )−1 )
wR3 ∼ N (00r , w2 Ir×r )
⇒ M 0+ (wR3 ) ∼ N (00d , w2 (M 0+ M 0 )−1 )
Therefore, the distribution of β 0 , which is the sum of the 3
independent Gaussians, is as required.
β + X +e ) + R1 PU ⊥ e + wR3 )
Also, ζ 0 = √1r PU 0⊥ (R1 X(β
is the sum of 3 independent Gaussians, which implies its
distribution is
β + X +e ),
N √1r PU 0⊥ M 0 (β
2
+ 2
2
1
β
(w
(kβ
+
X
e
k
+
1)
+
kP
⊥ e k )PU 0⊥
U
r
βj0 −(βj +(X +e )j )
q r
0T
0 −1
r−p ·(M M )j,j
=
kζζ 0 k
β 0 −β̂j
q rj
0T
0 −1
r−p ·(M M )j,j
is dis-
tributed like a Tr−p -distribution. Therefore, the following
theorem follows immediately.
β =
Theorem C.2. Fix X ∈ Rn×p and y ∈ R. Define β̂
X +y and ζ = (I − XX + )yy . Let RX 0 and Ryy 0 denote
the result of applying Algorithm 1 to the matrix A = [X; y ]
when the algorithm appends the data with a w · I matrix.
Fix a coordinate j and any α ∈ (0, 1/2). When computing
β 0 and ζ 0 as in Equations (9) it and (10), we have that w.p.
≥ 1 − α it holds that
q
r
β̂j ∈ βj0 ± c0α kζζ 0 k r−p
· (M 0T M 0 )−1
j,j
where c0α denotes the number such that (−c0α , c0α ) contains
1 − α mass of the Tr−p -distribution.
Note that Theorem C.2, much like the rest of the discussion in this Section, builds on y being fixed, which means
βj0 serves as an approximation for β̂j . Yet our goal is to
argue about similarity (or proximity) between βj0 and βj .
To that end, we combine the standard OLS confidence interval — which says that w.p. ≥ 1 − α over the randomness of picking r
e in the homoscedastic model we have
|βj − β̂j | ≤ cα kζζ k
(X T X)−1
j,j
n−p
— with the confidence in-
terval of Theorem C.2 above, and deduce that w.p. ≥ 1 − α
Differentially Private Ordinary Least Squares
we have that |βj0 − βj | is at most
1 − ν over the randomness of R we have (r −
kζζ 0 k2
−1
2
T
p)(M 0T M )−1
j,j = Θ (w Ip×p + X X)j,j and r−p =
q
q
kζζ k (X T X)−1
kζζ 0 k r(M 0T M 0 )−1
j,j
j,j
√
√
O cα
+ c0α
n−p
r−p
Θ( w
(11)
And so, in the next section, our goal is to give conditions under which the interval of Equation (11) isn’t
much larger
q in comparison to the interval length of
18
ζ0
ζ k
c0α √kζr−p
r(M 0T M 0 )−1
j,j we get from Theorem C.2; and
more importantly — conditions that make the interval of
Theorem C.2 useful q
and not too large. (Note, in expecta-
tion
kζζ 0 k
√
r−p
is about
β k2 + kζζ k2 )/r. So, for
(w2 + w2 kβ̂
β k is very large, this interval
example, in situations where kβ̂
isn’t likely to inform us as to the sign of βj .)
Motivating Example. A good motivating example for the
discussion in the following section is when [X; y ] is a strict
submatrix of the dataset A. That is, our data contains many
variables for each entry (i.e., the dimensionality d of each
entry is large), yet our regression is made only over a modest subset of variables out of the d. In this case, the least
singular value of A might be too small, causing the algorithm to alter A; however, σmin (X T X) could be sufficiently large so that had we run Algorithm 1 only on [X; y ]
we would not alter the input. (Indeed, a differentially private way for finding a subset of the variables that induce a
submatrix with high σmin is an interesting open question,
partially answered — for a single regression — in the work
of Thakurta and Smith (Thakurta & Smith, 2013).) Indeed,
the conditions we specify in the following section depend
on σmin ( n1 X T X), which, for a zero-mean data, the minimal variance of the data in any direction. For this motivating example, indeed such variance isn’t necessarily small.
C.2. Conditions for Deriving a Confidence Interval for
Ridge Regression
2
β k2 +kζζ k2
+w2 kβ̂
).
r
Proof. The former bound follows from known results on
the Johnson-Lindenstrauss transform (as were shown in the
proof of Claim B.6). The latter bound follows from standard concentration bounds of the χ2 -distribution.
Plugging in the result of Proposition C.3 to Equation (11)
we get that w.p. ≥ 1 − ν the difference |βj0 − βj | is at most
q
kζζ k
O cα √
(X T X)−1
j,j
n−p
s
β k2 + kζζ k2 q 2
w2 + w2 kβ̂
+ c0α
(w Ip×p + X T X)−1
j,j
r−p
(12)
We will also use the following proposition.
Proposition C.4.
(X T X)−1
j,j ≤
1+
w2
σmin (X T X)
(w2 Ip×p + X T X)−1
j,j
Proof. We have that
(X T X)−1
= (X T X)−1 (X T X + w2 Ip×p )(X T X + w2 Ip×p )−1
= (X T X + w2 Ip×p )−1 + w2 (X T X)−1 (X T X + w2 Ip×p )−1
= (Ip×p + w2 (X T X)−1 )(X T X + w2 Ip×p )−1
Looking at the interval specified in Equation (11), we now
give an upper bound on the the random quantities in this
interval: kζζ k, kζζ 0 k, and (M 0T M 0 )−1
j,j . First, we give bound
that are dependent on the randomness in R (i.e., we continue to view e as fixed).
= (X T X + w2 Ip×p )−1/2 ·
Proposition C.3. For any ν ∈ (0, 1/2), if we
have r = p + Ω(ln(1/ν)) then with probability ≥
where the latter holds because (Ip×p + w2 (X T X)−1 ) and
(X T X + w2 Ip×p )−1 are diagonalizable by the same matrix V (the same matrix for which (X T X) = V S −1 V T ).
2
Since we have kIp×p + w2 (X T X)−1 k = 1 + σ2 w (X) , it is
18
Observe that w.p.
≥ 1 − α r
over the randomness of e
we have that |βj − β̂j | ≤ cα kζζ k
(X T X)−1
j,j
n−p
, and w.p.
≥
1 − α over the randomness of R we have that |βj0 − β̂j | ≤
q
r
c0α kζζ 0 k r−p
· (M 0T M 0 )−1
j,j . So technically, to give a (1 − α)βj0
confidence interval around
that contains βj w.p. ≥ 1 − α, we
need to use cα/2 and c0α/2 instead of cα and c0α resp. To avoid
overburdening the reader with what we already see as too many
parameters, we switch to asymptotic notation.
(Ip×p + w2 (X T X)−1 )·
(X T X + w2 Ip×p )−1/2
min
w2
)I .
2
σmin
(X) p×p
T
T
−1
e j (X X) e j ≤ (1 +
clear that (Ip×p + w2 (X T X)−1 ) (1 +
We deduce that (X T X)−1
j,j =
w2
)(X T X
2
σmin
(X)
+ w2 Ip×p )−1
j,j .
Based on Proposition C.4 we get from Equation (12) that
Differentially Private Ordinary Least Squares
|βj0 − βj | is at most
v
u
w2
2
u
t kζζ k (1 + σmin (X T X) )
O( cα
+
n−p
s
β k2 + kζζ k2 q 2
w2 + w2 kβ̂
(w Ip×p + X T X)−1
c0α
j,j )
r−p
(13)
And so, if it happens to be the case that exists some small
β , ζ and w2 satisfy
η > 0 for which β̂
!
2
kζζ k2 (1 + σminw(X T X) )
β k2 + kζζ k2
w2 + w2 kβ̂
2
≤η
n−p
r−p
(14)
then
we
have
that
Pr[β
∈
j
q
βj0
± O((1 + η) ·
1 − α.
c0α kζζ 0 k
r
r−p
·
(M 0T M 0 )−1
j,j )
]
≥
19
Moreover, if in this case |βj | >
q
q
2
2 kβ̂
β k2 +kζζ k2
+ η) w +w r−p
(w2 Ip×p + X T X)−1
j,j
then Pr[sign(βj0 ) = sign(βj )] ≥ 1 − α. This is precisely
what Claims C.5 and C.6 below do.
Claim C.5. If there
exists η
>
0
2
2
s.t.
n − p
≥
(r
−
p)
and
n
=
η2
2
1
3/2 B ln(1/δ)
Ω r
·
·
, then Pr[βj ∈
1
η 2 σmin ( n X T X)
q
r
βj0 ± O((1 + η) · c0α kζζ 0 k r−p
· (M 0T M 0 )−1
)
]
≥
j,j
1 − α.
c0α (1
Proof. Based on the above discussion, it is enough to argue that under the conditions of the claim, the constraint
2
r−p
of Equation (14) holds. Since we require η2 ≥ n−p
then
it is evident that
kζζ k2
n−p
2
kζζ k2
n−p
η 2 kζζ k2
2(r−p) . So we now show that
η 2 kζζ k2
2(r−p) under the conditions of the
≤
· σminw(X T X) ≤
claim, and this will show the required. All that is left is
some algebraic manipulations. It suffices to have:
η2
2
·
n−p
T
r−p σmin (X X)
η2
2
≥
32B
≥
which holds for n2 ≥ r3/2 ·
as we assume to hold.
·
T
n2
1
r σmin ( n X X)
√
2
r ln(8/δ)
≥ w2
64B 2 ln(1/δ)
σmin ( n1 X T X)−1 ,
η 2
Claim C.6. Fix ν ∈ (0, 12 ). If (i) n = p + Ω(ln(1/ν)),
β k2 = Ω(σ 2 kX + k2F ln( νp )) and (iii) r − p =
(ii) kβ
Ω
(c0α )2 (1+η)2
βj2
β k2 +
1 + kβ
σ2
1
σmin ( n X T X)
, then in the
homoscedastic model, with probability ≥ 1−ν −α we have
that sign(βj ) = sign(βj0 ).
19
We assume n ≥ r so cα < c0α as the Tn−p -distribution is
closer to a normal Gaussian than the Tr−p -distribution.
Proof. Based on the above discussion, we aim to show that
in the homoscedastic model (where each coordinate ei ∼
N (0, σ 2 ) independently) w.p. ≥ 1 − ν it holds that the
magnitude of βj is greater than
s
c0α (1 + η)
β k2 + kζζ k2 q 2
w2 + w2 kβ̂
(w Ip×p + X T X)−1
j,j
r−p
To show this, we invoke Claim A.4 to argue that w.p. ≥
1 − ν we have (i) kζζ k2 ≤ 2σ 2 (n − p) (since n = p +
β k2 ≤ 2kβ
β k2 (since kβ
β − β̂
β k2 ≤
Ω(ln(1/ν))), and (ii) kβ̂
p
2
+ 2
2
2
+
2
β k = Ω(σ kX kF ln( νp ))).
σ kX kF ln( ν ) whereas kβ
2
We also use the fact that (w2 Ip×p + X T X)−1
j,j ≤ (w +
−1
T
σmin (X X)), and then deduce that
s
β k2 + kζζ k2 q 2
w2 + w2 kβ̂
(w Ip×p + X T X)−1
j,j
r−p
s
β k2 ) + σ 2 (n − p)
(1 + η)c0α
w2 (1 + kβ
≤ √
2
2
r−p
w + σmin (X T X)
s
(1 + η)c0α
2σ 2 (n − p)
β k2 ) +
≤ √
2(1 + kβ
≤ |βj |
r−p
σmin (X T X)
(1 + η)c0α
due to our requirement on r − p.
Observe, out of the 3 conditions specified in Claim C.6,
condition (i) merely guarantees that the sample is large
enough to argue that estimations are close to their expect
value; and condition (ii) is there merely to guarantee that
β k ≈ kβ
β k. It is condition (iii) which is non-trivial to
kβ̂
hold, especially together with the conditions of Claim C.5
that pose other constraints in regards to r, n, η and the various other parameters in play. It is interesting to compare
the requirements on r to the lower bound we get in Theorem 3.3 — especially the latter bound. The two bounds
are strikingly similar, with the exception that here we also
β k2
require r − p to be greater than 1+kβ
. This is part of the
βj2
unfortunate effect of altering the matrix A: we cannot give
confidence bounds only for the coordinates j for which βj2
β k2 .
is very small relative to kβ
In summary, we require to have n = p + Ω(ln(1/ν)) and
β k compathat X contains enough sample points to have kβ̂
β k, and then set r and η such that (it is convenient
rable to kβ
to think of η as a small constant, say, η = 0.1)
• r − p = O(η 2 (n − p)) (which implies r = O(n))
2
3
n2
• r = O( η 2 B 2 ln(1/δ)
σmin ( n1 X T X) )
2
βk
• r − p = Ω( 1+kβ
+
β2
β
j
σ2
βj2
−1 1
· σmin
( n X T X))
Differentially Private Ordinary Least Squares
to have that the (1 − α)-confidence interval around βj0
does not intersect the origin. Once again, we comment
that these conditions are sufficient but not necessary, and
furthermore — even with these conditions holding — we
do not make any claims of optimality of our confidence
bound. That is because from Proposition C.4 onwards our
discussion uses upper bounds that do not have corresponding lower bounds, to the best of our knowledge.
D. Confidence Intervals for “Analyze Gauss”
Algorithm
To complete the picture, we now analyze the “Analyze
Gauss” algorithm of Dwork et al (Dwork et al., 2014).
Algorithm 2 works by adding random Gaussian noise to
AT A, where the noise is symmetric with each coordi2
nate abovethe diagonal
sampled i.i.d from N (0, ∆ ) with
∆2 = O B 4 log(1/δ)
.20 Using the same notation for a
2
sub-matrix of A as [X; y ] as before, with X ∈ Rn×p and
y ∈ Rn , we denote the output of Algorithm 2 as
^
TX
X
TX
yg
g
Ty
X
=
Ty
yg
T
X X +N
y TX + nT
Ty
X y +n
y Ty + m
(15)
where N is a symmetric p × p-matrix, n is a p-dimensional
vector and m is a scalar, whose coordinates are sampled
i.i.d from N (0, ∆2 ).
Using the output of Algorithm 2, it is simple to derive anaβ and kζζ k2 (Equations (1) and (2))
logues of β̂
−1
−1
g
^
TX
Ty = X T X + N
X
βe = X
(X T y + n )
(16)
r
+∆
−2
√
^
TX
X
j,j · ln(1/ν) · (B p + 1)
where ρ is such that ρ2 is w.h.p an upper bound on σ 2 ,
defined as
2 def
ρ =
1√
√
n−p−2 ln(4/α)
2
·
√
−1
B2 p p
2
g
^
T
2
kζζ k − C · ∆ 1−η ln(1/ν) + ∆ kX X kF · ln(p/ν)
for some large constant C.
We comment that in practice, instead of using ρ, it might
be better to use the MLE of σ 2 , namely:
−1
def
2 ^
1
g
2
2
T
σ = n−p kζζ k + ∆ kX X kF
instead of ρ2 , the upper bound we derived for σ 2 . (Replacing an unknown variable with its MLE estimator is a common approach in applied statistics.) Note that the assumpβ k ≤ B is fairly benign once we assume each
tion that kβ
β k ≤ B simrow has bounded l2 -norm. The assumption kβ̂
β
ply assumes that β̂ is a reasonable estimation of β , which is
likely to hold if we assume that X T X is well-spread. The
assumption about the magnitude of the least singular value
of X T X is therefore the major one. Nonetheless, in the
case we considered before where each row in X is sampled
i.i.d from N (00, Σ), this assumption
merely means that n is
√
large enough s.t. n = Ω̃(
∆ p ln(1/ν)
η·σmin (Σ) ).
In order to prove Theorem D.1, we require the following
proposition.
(17)
Proposition D.2. Fix any ν ∈ (0, 21 ). Fix any matrix
M ∈ Rp×p . Let v ∈ Rp be a vector with each coordinate
sampled independently
from a Gaussian N (0, ∆2 ).i Then
h
p
we have that Pr kMvv k > ∆ · kM kF 2 ln(2p/ν) < ν.
g
ζ k2 to get a
We now argue that it is possible to use βej and kζ
confidence interval for βj under certain conditions.
Theorem D.1. Fix α, ν ∈ (0, 12 ). Assume that there exists
p
η ∈ (0, 21 ) s.t. σmin (X T X) > ∆ p ln(1/ν)/η. Under the
homoscedastic model, given β and σ 2 , if we assume also
β k ≤ B and kβ̂
β k = k(X T X)−1 X Ty k ≤ B, then
that kβ
w.p. ≥ 1 − α − ν it holds that |βj − βej | it at most
s
−1
−2
p
^
^
TX
TX
O ρ·
X
+
∆
p
ln(1/ν)
·
X
ln(1/α)
j,j
j,j
Proof. Given M , we have that Mvv ∼ N (00, ∆2 · M M T ).
Denoting M ’s singular values as sv1 , . . . , svp , we can rotate Mvv without affecting its l2 -norm and infer that kMvv |2
is distributed like a sum on p independent Gaussians, each
sampled from N (0, ∆2 · svi2 ). Standard union bound gives
that w.p. ≥ 1 − ν non of the pp
Gaussians exceeds its standard deviation by a factor of 2 ln(2p/ν).
Hence, w.p.
P
≥ 1 − ν it holds that kMvv k2 ≤ 2∆2 i svi2 ln(2p/ν) =
2∆2 · trace(M M T ) · ln(2p/ν).
T
g
^
]
e + βe X
e
Ty − 2 y
TXβ
TX β
ζ k2 = yg
kζ
−1
^
]
Ty − y
TX X
TX
= yg
]
Ty
X
20
It is easy to see that the l2 -global sensitivity of the mapping
A 7→ AT A is ∝ B 4 . Fix any A1 , A2 that differ on one row
which is some vector v with kvv k = B in A1 and the all zero
vector in A2 . Then GS22 = kAT1 A1 − AT2 A2 k2F = kvvv T k2F =
trace(vvv T · v v T ) = (vv Tv )2 = B 4 .
Our proof also requires the use of the following equality,
that holds for any invertible A and any matrix B s.t. I +
B · A−1 is invertible:
−1
−1
(A + B) = A−1 − A−1 I + BA−1
BA−1
Differentially Private Ordinary Least Squares
−1
In our case, we have
T
^
TX
we can bound the variance of X
j→ X e by
!
2
−1
−1
^
^
T
TX
.
Appealing to
σ2 X
j,j + kN k · X X j→
−1
^
TX
X
= (X T X + N )−1
T
−1
−1
T
−1 −1
−1
T
− (X X) I + N (X X)
N (X X)
−1
= (X T X)−1 I − I + N (X T X)−1
N (X T X)−1
def
= (X T X)−1 I − Z · (X T X)−1
(18)
= (X X)
T
Proof of Theorem D.1. Fix ν > 0. First, we apply to standard results about Gaussian matrices, such as (Tao, 2012)
(used also by (Dwork et al., 2014) in their analysis),
to see
p
that w.p. ≥ 1 − ν/6 we have kN k = O(∆ p ln(1/ν)).
And so, for the remainder of the proof we fix N subject to
having bounded operator norm. Note that by fixing N we
^
T X.
fix X
β + e with
Recall that in the homoscedastic model, y = Xβ
each coordinate of e sampled i.i.d from N (0, σ 2 ). We
therefore have that
−1
^
TX
βe = X
^
TX
(X Ty + n ) = X
−1
−1
^
TX
=X
−1
^
TX
=β −X
−1
^
TX
β +X
Nβ
!
σ 2 ln(1/α) .
−1
−1 2
−1
2 ^
^
TX
X
X T X j→
j→n ∼ N (0, ∆
). Gaussian concentra−1
^
TX
tion bounds give that w.p ≥ 1−ν/6 we have |X
j→n | =
−1 p
^
TX
ln(1/ν) .
O ∆ X
j→
Plugging this into our above bounds on all terms that appear
in Equation (19) we have that w.p. ≥ 1 − ν/2 − α/2 we
have that βej − βj is at most
^
TX
X e+X
−1
n
^
TX
X
j→ · B∆
O
p
p ln(1/ν)
−1
v
u
−1
−1
u
p
^
^
TX
T
+ O σt X
j,j + ∆ p ln(1/ν) · X X j→
n
2
!
ln(1/α)
−1
^
TX
Denoting the j-th row of X
^
TX
as X
j→ we deduce:
−1
^
TX
+O ∆ X
j→
p
ln(1/ν)
−1
−1
−1
2
^
TX
0, ∆2 I) is samTo bound X
j→n note that n ∼ N (0
^
T X.
pled independently of X
We therefore have that
−1
T
^
TX
X Te + X
−1
u
−1
−1
u
p
^
^
T
TX
O t X
j,j + ∆ p ln(1/ν) · X X j→
β + X Te + n )
(X T Xβ
−1
^
^
T X − N )β
TX
β +X
(X
Gaussian concentration bounds, we have that w.p.
≥ 1v
− α/2 the absolute value of this Gaussian is at most
T
^
^
^
T
TX
TX
β +X
βej = βj − X
j→ X e + X X j→n
j→ Nβ
(19)
^
T X we have
Note that due to the symmetry of X
−1 2
We naı̈vely bound the size of the
−1
−1
^
^
TX
TX
β
βk
X
by
X
j→ Nβ
j→ kN kkβ
−1
p
^
TX
O X
j→ · B∆ p ln(1/ν) .
−1
Te
^
TX
X
j→ X e
term
=
note that e is cho^
TX
sen
independently
of
X
and
since
−1
2
T
^
T
e
∼
N (00, σ I) we have X X j→ X e
∼
−1
−1
^
^
TX
T
N 0, σ2 · eT
· X TX · X
ej .
Since
jX X
To
bound
^
TX
X
j→
−2
^
TX
=X
j,j (the (j, j)-coordinate of the ma−2
^
TX
trix X
), thus |βej − βj | is at most
s
O σ·
r
+∆
−1
^
TX
X
j,j + ∆
−2
−2
p
^
T
p ln(1/ν) · X X j,j ln(1/α)
√
^
TX
X
j,j · ln(1/ν) · (B p + 1)
(20)
we have
−1
^
TX
X
^
TX
· X TX · X
−1
−1
−1
^
^
TX − N ) · X
TX
· (X
^
TX
=X
−1
^
TX
=X
−1
^
TX
−X
−1
^
TX
·N ·X
All of the terms appearing in Equation (20) are known
^
T X, except for σ — which is a parameter of the
given X
model. Next, we derive an upper bound on σ which we can
then plug into Equation (20) to complete the proof of the
theorem and derive a confidence interval for βj .
Differentially Private Ordinary Least Squares
p
most O(∆ · kzz k ln(1/ν)) w.p. ≥ 1 − ν/8. We can
β k kI − Z(X T X)−1 k =
upper bound kzz k ≤ 2kβ̂
B
O( 1−η ), and so this term’s magnitude is upper
√
∆·B ln(1/ν)
bounded by O
.
1−η
Recall Equation (17), according to which we have
−1
g
^
]
Ty − y
TX X
TX
ζ k2 = yg
kζ
]
Ty
X
(18)
= y Ty + m
− (yy T X + n T )(X T X)−1 (I − Z · (X T X)−1 )(X Ty + n )
• Given our assumption about the least singular value
of X T X and with the bound on kN k, we have that
T
^
T X) ≥ σ
σmin (X
min (X X) − kN k > 0 and so
= y Ty + m
− y T X(X T X)−1 X Ty
+ y T X(X T X)−1 Z(X T X)−1 X Ty
T
T
−1
n
T
T
−1
Z(X T X)−1n
− 2yy X(X X)
+ 2yy X(X X)
^
T X is a PSD. Therefore,
the symmetric matrix X
−1
−1/2
^
^
TX
TX
n = kX
n k2 is strictly
the term n T X
positive. Applying Proposition D.2 we have that
−1
T^
TX
w.p.
≥
1
−
ν/8
it
holds
that
n
X
n ≤
−1
2 ^
O ∆ kX T X kF · ln(p/ν) .
n
− nT (X T X)−1 (I − Z · (X T X)−1 )n
β = (X T X)−1 X Ty , and so we have
Recall that β̂
T
β Zβ̂
β
= y T I − X(X T X)−1 X T y + m − β̂
T
^
TX
β (I − Z(X T X)−1 )n
n − nTX
− 2β̂
Plugging all of the above bounds into Equation (21) we get
that w.p. ≥ 1 − ν/2 − α/2 it holds that
−1
n
(21)
and of course, both n and m are chosen independently of
^
T X and y .
X
Before we bound each term in Equation (21), we first give
−1
a bound on kZk. Recall, Z = I + N (X T X)−1
N.
Recall our assumption (given in the
statement
of
Theop
rem D.1) that σmin (X T X) ≥ ∆
p ln(1/ν). This imη
T
−1
plies that kN (X X) k ≤ kN k·σmin (X T X)−1 = O(η).
Hence
√
∆ p ln(1/ν)
kZk ≤ (kI +N (X T X)−1 k)−1 ·kN k = O
1−η
Moreover, this implies that kZ(X T X)−1 k ≤ O
1
and that kI − Z(X T X)−1 k ≤ O 1−η
.
η
1−η
Armed with these bounds on the operator norms of Z and
(I − Z(X T X)−1 ) we bound the magnitude of the different
terms in Equation (21).
• The term y T (I − XX + ) y is the exact term from
the standard OLS, and we know it is distributed like
σ 2 · χ2n−p distribution. Therefore, it is greater than
p
√
σ 2 ( n − p − 2 ln(4/α))2 w.p. ≥ 1 − α/2.
• The scalar m sampled
from m ∼ N (0, ∆2 ) is
p
bounded by O(∆ ln(1/ν)) w.p. ≥ 1 − ν/8.
T
βk ≤
β Zβ̂
β is upper
• Since we assume kβ̂
B, the term β̂
√
2
B
∆
p
ln(1/ν)
.
bounded by B 2 kZk = O
1−η
T
β (I − Z(X T X)−1 )n
n. We thus have
• Denote z Tn = 2β̂
T
that z n ∼ N (0, ∆2 kzz k2 ) and that its magnitude is at
σ2 ≤
1√
√
n−p−2 ln(4/α)
g
ζ k2 + O (1 +
kζ
2
·
√
p
B 2 p+B
)∆ ln(1/ν)
1−η
^
TX
+ ∆ 2 kX
−1
kF · ln(p/ν)
and indeed, the RHS is the definition of ρ2 in the statement
of Theorem D.1.
E. Experiment: Additional Figures
To complete our discussion about the experiments we have
conducted, we attach here additional figures, plotting both
the t-value approximations we get from both algorithms,
and the “high-level decision” of whether correctly reject or
not-reject the null hypothesis (and with what sign). First,
we show the distribution of the t-value approximation for
coordinates that should be rejected, in Figure 2, and then
the decision of whether to reject or not based on this t-value
— and whether it was right, conservative (we didn’t reject
while we needed to) or wrong (we rejected with the wrong
sign, or rejected when we shouldn’t have rejected) in Figure 3. As one can see, Algorithm 1 has far lower t-values
(as expected) and therefore is much more conservative. In
fact, it tends to not-reject coordinate 1 of the real-data even
on the largest value of n (Figure 3c).
However, because Algorithm 1 also has much smaller
variance, it also does not reject when it ought to notreject, whereas Algorithm 2 erroneiously rejects the nullhypotheses. This can be seen in Figures 4 and 5.
Differentially Private Ordinary Least Squares
(a) Synthetic data, coordinate β1 = 0.5
(a) Synthetic data, coordinate β1 = 0.5
(b) Synthetic data, coordinate β2 = −0.25
(b) Synthetic data, coordinate β2 = −0.25
(c) real-life data, coordinate β1 = 14.07
Figure 2. The distribution of the t-value approximations from selected experiments on synthetic and real-life data where the null
hypothesis should be rejected
(c) real-life data, coordinate β1 = 14.07
Figure 3. The correctness of our decision to reject the nullhypothesis based on the approximated t-value where the null hypothesis should be rejected
Differentially Private Ordinary Least Squares
(a) Synthetic data, coordinate β3 = 0
(a) Synthetic data, coordinate β3 = 0
(b) Real-life data, coordinate β2 = 0.57
(b) Real-life data, coordinate β2 = 0.57
Figure 4. The distribution of the t-value approximations from selected experiments on synthetic and real-life data when the null
hypothesis is (essentially) true
Figure 5. The correctness of our decision to reject the nullhypothesis based on the approximated t-value when the null hypothesis is (essentially) true
| 8cs.DS
|
arXiv:1611.05407v1 [math.ST] 16 Nov 2016
A Semidefinite Program for Structured
Blockmodels
David Choi
November 17, 2016
Abstract
Semidefinite programs have recently been developed for the problem
of community detection, which may be viewed as a special case of the
stochastic blockmodel. Here, we develop a semidefinite program that can
be tailored to other instances of the blockmodel, such as non-assortative
networks and overlapping communities. We establish label recovery in
sparse settings, with conditions that are analogous to recent results for
community detection. In settings where the data is not generated by a
blockmodel, we give an oracle inequality that bounds excess risk relative to the best blockmodel approximation. Simulations are presented for
community detection, for overlapping communities, and for latent space
models.
1
Introduction
The stochastic blockmodel [19, 7] is a popular class of models for network data
in which each node is assumed to belong to a latent class. Various sub-families
of the blockmodel now exist, such as community structure [20], hierarchical
community structure [23, 27], and overlapping blockmodels [29, 33, 32], as well
as relatives such as latent space models [18], mixed membership [1], degreecorrected blockmodels [20], and time-varying blockmodels [24].
For all of these models, estimation of the latent nodal classes is an active
area of research. For blockmodels, spectral methods are known to yield asymptotically consistent estimates, provided that the network is sufficiently large and
dense [22, 28]. For the special case of community structure, it is additionally
known that specialized methods can achieve weakly consistent estimates even
when spectral methods fail completely due to sparsity [26, 21, 14, 2, 31]. Examples of such methods include semidefinite programming [3, 17, 25] and message
passing [13, 32]. For other variants of the blockmodel and their relatives, estimation methods also exist but are less understood; in particular, theory analogous
to that of community detection does not yet seem to exist for these cases.
To address this gap, we show in this paper that semidefinite programming
can be applied not only to community detection, but also to other blockmodel
1
sub-families as well. Specifically, we propose a semidefinite program that can be
tailored to any specific instance of the blockmodel. For this program, we prove
estimation bounds that are analogous to those already known for community
detection, including weak consistency in the bounded degree setting. When the
data is not generated from a blockmodel, the semidefinite program can be used
to construct a “de-noised” version of the data matrix, and we provide an oracle
inequality bounding its error relative to the best blockmodel approximation.
The organization of the paper is as follows. Section 2 presents the semidefinite program. Section 3 presents a convergence analysis for sparse data. Section
4 discusses numerical optimization. Section 5 gives simulations results. Proofs
are contained in the appendix.
2
Problem Formulation
In this section, we present the generative models that we will consider in this
paper; derive a semidefinite relaxation for the combinatorial problem of estimating the latent nodal classes; and present estimators for blockmodel and
non-blockmodel settings.
2.1
Preliminary Notation
Given a matrix M ∈ RnK×nK , let M (ij) ∈ RK×K for i, j ∈ [n] denote its (i, j)th
submatrix of size K ×K, and similarly for x ∈ RnK let x(i) ∈ RK for i = 1, . . . , n
denote its ith subvector of length K, so that
(1)
x
M (1,1) · · · M (1,n)
..
.
.
.
.
.
.
and
x = . .
M =
.
.
.
M (n,1)
···
x(n)
M (n,n)
(ij)
We will use Mab to denote the (a, b)th entry of the submatrix M (ij) , and
(i)
likewise use xa to denote the ath entry of the subvector x(i) . This implies that
(ij)
(i)
Mab = Ma+(i−1)K,b+(j−1)K and xa = xa+(i−1)K .
2.2
Generative Models
Stochastic Blockmodel Let A ∈ {0, 1}n×n denote the symmetric adjacency
matrix of a undirected network with n nodes. In a stochastic blockmodel with K
classes, each node has a random latent class zi ∈ [K], and the upper triangular
entries of A are independent Bernoulli when conditioned on z:
zi ∼ Discrete(π)
(1)
Aij ∼ Bernoulli(θzi ,zj ),
Aji = Aij ,
2
i, j ∈ [n], i < j
(2)
where π is a probability distribution over [K] giving the expected class frequencies, and θ ∈ [0, 1]K×K is a symmetric matrix that gives the connection
probabilities between each class type.
General Model Under a more general model, A ∈ {0, 1}n×n is a random
matrix generated by
Aij ∼ Bernoulli(Pij )
i<j
(3)
Aji = Aij ,
where P ∈ [0, 1]n×n is symmetric and satisfies Pii = 0 for i ∈ [n]. It can be seen
that the stochastic blockmodel is a special case of (3), where Pij = θzi ,zj .
2.3
Semidefinite program
We will assume that A ∈ {0, 1}n×n is observed, and the estimation task is to
find z ∈ [K]n maximizing the generic combinatorial problem
maxn
z∈[K]
n
n X
X
fij (zi , zj ),
(4)
i=1 j=1
under some choice of objective functions fij : [K]2 7→ R for i, j ∈ [n]. In this
paper, we will let {fij } equal the likelihood function
fij (zi , zj ) = Aij log θ̂zi zj + (1 − Aij ) log(1 − θ̂zi ,zj ),
(5)
in which case (4) finds the maximum likelihood assignment of z under a specified parameter matrix θ̂ ∈ [0, 1]K×K . Note that θ̂ may differ from the actual
generative model for A.
Optimizing (4) is not computationally tractable, so we will relax it into a
semidefinite program. Let F ∈ RnK×nK denote a matrix with submatrices
F (ij) ∈ RK×K given by
(ij)
a, b ∈ [K],
Fab = fij (a, b)
(6)
so that (4) can be expressed as
maxn
z∈[K]
n X
n
X
eTzi F (ij) ezj ,
(7)
i=1 j=1
where e1 , . . . , eK ∈ {0, 1}K denote the canonical basis in RK . This can be
further rewritten as
ez1
max
hF, xxT i subject to x = ... .
z∈[K]n ,x∈{0,1}nK
ezn
3
This suggests the following semidefinite program relaxation, where xxT is approximated by a positive semidefinite matrix X:
max
X∈RnK×nK
hF, Xi
(8)
s.t. X 0, X ≥ 0
K X
K
X
(ij)
∀ i, j ∈ [n],
Xab = 1
a=1 b=1
where X 0 denotes that X positive semidefinite, and X ≥ 0 denotes that
X is elementwise non-negative. For any feasible X ∈ RnK×nK , each submatrix
X (ij) ∈ RK×K is nonnegative and sums to one, and can be viewed as a relaxed
version of the indicator matrix ezi eTzj encoding the class pair (zi , zj ).
2.4
Matrix Denoising and Blockmodel Estimation
Let X̂ ∈ [0, 1]nK×nK denote the solution to the semidefinite program (8), and let
A be generated from the general model (3), with generative matrix P ∈ [0, 1]n×n .
Let P̂ ∈ [0, 1]n×n denote a “MAP estimate” of P , constructed by treating each
submatrix X̂ (ij) as a probability distribution over [K]2 :
P̂ij = arg max
ϑ∈[0,1]
K X
K
X
(ij)
X̂ab · 1{θ̂ab = ϑ}
(9)
a=1 b=1
Alternatively, let P̃ denote a randomized estimate of P , where each dyad is an
independent random variable with distribution
(ij)
P̃ij = θ̂ab with probability X̂ab
∀i < j
(10)
P̃ji = P̃ij .
Let ẑ ∈ [K]n denote the cluster labels found by spectral clustering of P̂ – i.e.,
applying K-means to the first K eigencoordinates of P̂ . If A is generated by a
blockmodel, then the generative P will be block structured, with blocks induced
by θ and z. In this case, we will use ẑ to estimate z, up to label-switching
permutation.
To estimate θ up to permutation, let θ̂est ∈ [0, 1]K×K denote the matrix of
between-block densities induced by ẑ,
Pn Pn
j=1 Aij 1{ẑi = a, ẑj = b}
i=1
est
a, b ∈ [K].
(11)
θ̂ab = Pn Pn
i=1
j=1 1{ẑi = a, ẑj = b}
2.5
Discussion
Related Work Semidefinite programs have been used for community detection in [3, 17, 25], as well as [9] which allowed for outliers, and [11] which
4
allowed for degree heterogeneity. In each of these works, the network is required
to exhibit assortatitve block structure. For the general model (3) without such
restrictions, estimation has been considered in [10], but only for dense settings.
To the best of our knowledge, networks that are both non-assortative and sparse
(i.e., bounded degree), such as those presented in Sections 5.2 and 5.3, have not
been considered in previous work.
Additionally, the semidefinite program presented here bears resemblance to
one appearing in [34] for lower bounding the optimal objective value of the
quadratic assignment problem (without finding a feasible solution)1 , and also
recent work on estimating pairwise alignments beween objects [6].
3
Convergence Analysis
In this section, we analyze the solution of the semidefinite program (8) for
both matrix denoising and label recovery. Analogous to existing results for
community detection, our results will imply weak consistency (i.e., performance
better than random guessing) in the regime where the average degree of A is
asymptotically bounded above some constant, and consistency of ẑ, as well as
vanishing excess risk of P̂ , when the average degree → ∞.
The organization of this section is the following: Section 3.1 defines basic
notation; Section 3.2 states the required conditions; and Section 3.3 presents
the convergence results, which are proven in Appendix A.
3.1
Preliminaries
The following notation will be used. Given A ∈ {0, 1}n×n that is generated by
(3) under some P ∈ [0, 1]n×n , let ρ denote its expected density
X
1
Pij .
ρ=
n(n − 1) ij
Given θ̂ ∈ [0, 1]K×K , let F ∈ RnK×nK denote the objective function of the
semidefinite program (8), with submatrices F (ij) ∈ RK×K given by
(ij)
Fab = Aij log θ̂ab + (1 − Aij ) log(1 − θ̂ab )
i, j ∈ [n], a, b ∈ [K].
(12)
nK×nK
Let F̄ ∈ R
denote a idealized version of F in which A is replaced by its
expectation P , with submatrices F̄ (ij) ∈ RK×K given by
(ij)
F̄ab = Pij log θ̂ab + (1 − Pij ) log(1 − θ̂ab )
i, j ∈ [n], a, b ∈ [K].
(13)
Let X ⊂ RnK×nK denote the feasible region of the semidefinite program (8):
(
)
K X
K
X
(ij)
nK×nK
X = X∈R
: X 0, X ≥ 0, and
Xab = 1 for all i, j ∈ [n] .
a=1 b=1
(14)
1 which
can be used as an intermediate step in a branch-and-bound algorithm.
5
Let X̂ denote any solution to the semidefinite program (8), which can be written
as
maximize hF, Xi over all X ∈ X .
For any matrix B ∈ [0, 1]K×K , let the function QB be given by
QB (a, b) = {(c, d) ∈ [K]2 : Bcd = Bab }
a, b ∈ [K],
so that QB identifies the subsets of [K]2 that have equal values under B.
3.2
Assumptions
Assumption 1 and 2 will apply when P̃ is used to estimate the general model
(3). Assumption 1 requires P to have density ρ exceeding 1/n. Assumption 2
bounds the entries of θ̂ to differ from ρ by roughly at most a constant factor.
Assumption 1. Let A ∈ {0, 1}n×n be generated by (3), with P ∈ [0, 1]n×n
(which evolves with n) satisfying ρ ≥ 1/n.
Assumption 2. The matrix θ̂ ∈ [0, 1]K×K (which may evolve with n) satisfies
ρ
θ̂ab
≤ cρ for some c > 0 and all n and a, b ∈ [K].
c ≤ 1−θ̂
ab
Assumptions 3 and 4 will apply when A is generated by a stochastic blockmodel, and are sufficient to show that ẑ converges to the true z, up to label
switching. Assumption 3 describes a parametrization that is commonly used
for sparse blockmodels. Assumption 4 places bounds on the misspecification
between θ̂ and θ∗ in the sparse blockmodel setting.
Assumption 3. For all n, let A ∈ {0, 1}n×n and z ∈ [K]n be generated by a
stochastic blockmodel with parameters (π, θ∗ ). Let π be constant, and let θ∗ =
αB ∗ , where α ∈ R satisfies α → 0 and α > 1/n, and B ∗ ∈ RK×K
is constant,
+
rank K, and satisfies
K X
K
X
∗
πa πb Bab
= 1.
(15)
a=1 b=1
P
Aij , and B̂ ∈ RK×K
is a
+
∀a, b ∈ [K]
(16)
Assumption 4. Let θ̂ = α̂B̂, where α̂ =
1
n(n−1)
ij
fixed matrix such that B̂ and B ∗ satisfy
QB̂ (a, b) = QB ∗ (a, b)
∗
Bab
log
B̂ab
B̂cd
− (B̂ab − B̂cd ) > 0
∀a, b ∈ [K] and c, d ∈
/ QB ∗ (a, b). (17)
Assumption 4 states that B̂ and B ∗ need not have identical values, but
should have the same structure (as given by QB̂ and QB ∗ ). Additionally, for all
∗
a, b ∈ [K], the entry B̂ab should be the closest element of B̂ to Bab
, in terms of
the Bregman divergence associated with the Poisson likelihood.
6
3.3
Results
Theorem 1 holds when A is generated from the general model (3), including
non-blockmodels. It gives an oracle inequality on the quality of the randomized
estimate P̃ given by (10), relative to the best blockmodel approximation to the
generative P .
Theorem 1. Let Assumptions 1 and 2 hold. Let P̃ denote the randomized
estimate of P given by (10). Then
n
n X
n
n
X
1 XX
1
1
min
,
KL(P
,
P̃
)
≤
KL(P
,
θ̂
)
+
O
√
ij
ij
ij ẑi ,ẑj
P
n2 ρ i=1 j=1
n2 ρ ẑ∈[K]n i=1 j=1
nρ
with OP (·) having constant terms depending only on K and c (which appears in
Assumption 2).
Theorem 2 assumes the sparse blockmodel setting of Assumptions 3 and 4,
and shows that both P̂ and the randomized estimate P̃ asymptotically recover
P , with vanishing fraction of incorrect values.
Theorem 2. Let Assumptions 3 and 4 hold. Let P̂ denote the estimate of P
given by (9). Then
n
n
1
1 XX
1{P̂ij 6= θ̂zi zj } = OP √
,
n2 i=1 j=1
nα
(18)
with the same result if P̂ is replaced by P̃ given by (10).
Corollary 1 follows from Theorem 2, and states that the eigencoordinates
from P̂ , which are used to compute ẑ, converge (up to a unitary transformation)
to those of P . As P is block structured, this suggests that ẑ will converge to z
up to a permutation of the labels. It is proven in Appendix C, and is a direct
application of the Davis-Kahan theorem in [30, Th. 4].
Corollary 1. Let Assumptions 3 and 4 hold, and let V = (v1 , . . . , vK ) denote
the eigenvectors of P . Let V̂ = (v̂1 , . . . , v̂K ) denote the eigenvectors of the K
largest eigenvalues (in absolute value) of P̂ or P̃ . Let O denote the set of K ×K
orthogonal matrices. It holds that
1
.
min kV̂ O − V k2F ≤ OP √
O∈O
nα
Remark Weak consistency in the bounded degree regime is implied by each
of the results, in that the estimation error is bounded away from the performance of random guessing, provided that limn→∞ ρ ≥ c0 /n (for Theorem 1) or
limn→∞ α ≥ c0 /n (for Theorem 2), for some constant c0 .
7
4
Numerical optimization
The semidefinite program (8) can be solved by the alternating direction method
of multipliers (ADMM) [8]. While this semidefinite program is much larger than
those previously introduced for community detection, speedups can often be
achieved by exploiting problem structure, resulting in competitive computation
times.
To solve (8) using ADMM, we introduce the decision variable X ∈ RnK×nK
and auxiliary variables Y, W, U, V ∈ RnK×nK , which are all initialized to zero
and evolve according to the following rule:
1
1
Wt − Ut + Y t − V t + F
(19)
X t+1 = ΠX
2
ρ
W t+1 = max(0, X t+1 + U t )
Y t+1 = ΠS+ (X t+1 + V t )
U t+1 = U t + X t+1 − W t+1
V t+1 = V t + X t+1 − Y t+1 ,
where F ∈ RnK×nK is given by (6), ρ ∈ R is a positive stepsize parameter, the
operator ΠS+ denotes projection onto the set of positive semidefinite matrices,
and the operator ΠX denotes projection onto the affine subspace of all matrices
X ∈ RnK×nK satisfying the linear constraints
K X
K
X
(ij)
Xab = 1
∀i, j ∈ [n].
a=1 b=1
The slowest step of the ADMM iterations is the computation of Y t+1 =
ΠS+ (X t+1 + V t ), which requires the eigendecomposition of an nK × nK matrix. In comparison, semidefinite programs for community detection require
only the decomposition of an n × n matrix in each ADMM iteration. However,
in many settings of interest, X t+1 + V t will be highly structured and have fast
eigendecomposition methods.
In particular, let the K × K submatrices [X t+1 + V t ](ij) for i, j ∈ [n] be
symmetric and share a common set of orthonormal eigenvectors v1 , . . . , vK . It
then holds for some ` ≤ K and some partition S0 , . . . , S` of [K] that
[X t+1 + V t ](ij) =
`
X
l=0
λijl
X
vj vjT
∀ i, j ∈ [n],
(20)
j∈Sl
where λijl is the eigenvalue corresponding to the eigenvectors {vj : j ∈ Sl }. In
this case, Y t+1 can be computed in the following manner:
1. Find the eigendecomposition for each submatrix [X t+1 + V t ](ij) ∈ RK×K ,
yielding eigenvectors v1 , . . . , vK , sets S0 , . . . , S` partitioning [K], and eigenvalues {λijl } satisfying (20).
8
2. Let Λ0 , . . . , Λ` ∈ Rn×n and E0 , . . . , E` ∈ RK×K be given by
X
Λl (i, j) = λijl
and
El =
vj vjT
l = 0, . . . , `,
j∈Sl
so that
X t+1 + V t =
`
X
Λl ⊗ El .
l=0
3. Return
ΠS+ (X t+1 + V t ) =
`
X
ΠS+ (Λl ) ⊗ El ,
(21)
l=0
which holds because the matrices E0 , . . . , E` are orthogonal and positive
semidefinite.
Since Λ0 , . . . , Λ` ∈ Rn×n , the above steps require ` + 1 eigendecompositions of
n × n matrices. As the computation time of eigendecomposition scales cubically
in the matrix size, the resulting speedup can be quite significant in practice; for
example, if K = 10 and ` = log2 K, the speedup is a factor of roughly 200 (e.g.,
going from 3 hours to ≤ 1 minute to solve the semidefinite program).
When will the decomposition (20) hold? A sufficient condition is that θ̂ is
in the span of an association scheme, which is evidently a fundamental concept
in combinatorics [15, 16], and is defined as follows:
Definition 1. [15, Sec 2] A set of binary-valued and symmetric matrices B0 , . . . , B`
form an association scheme if the following properties hold:
1. B0 = I
P`
T
2.
i=0 Bi = 11
3. For all i, j, it holds that Bi Bj ∈ span(B0 , . . . , B` )
The key property of association schemes is the following result, which is Theorem 2.6 in [5], or follows from E1-E4 in [15]. The result states that B0 , . . . , B`
share a common set of eigenvectors.
Lemma 1 ([15, 5]). Let B0 , . . . , B` ∈ RK×K form an association scheme.
Then B0 , . . . , B` have the same eigenvectors v1 , . . . , vK , and for some partition
(i)
S0 , . . . , S` of [K] and scalars {λj }`i,j=0 it holds for that
Bi =
`
X
j=0
(i)
λj
X
vk vkT ,
i = 0, . . . , `.
k∈Sj
Additionally, the vector 1 will be one of the eigenvectors.
9
(22)
Theorem 3, which is proven in Appendix B, states that if the parameter matrix θ̂ ∈ RK×K is in the span of an association scheme, then the decomposition
(20) will hold, with ` equal to the number of matrices in the association scheme,
and with precomputable {Sl } and v1 , . . . , vK :
Theorem 3. Let F be defined as (12), for any symmetric A ∈ {0, 1}n×n
and θ̂ ∈ span(B0 , . . . , B` ), where {Bi }`i=0 form an association scheme. Let
X, W, Y, U, V ∈ RnK×nK be initialized to zero and evolve by the ADMM equations (19). Then (20) holds for all t, with v1 , . . . , vK and S0 , . . . , S` satisfying
(22).
In Section 5, we will give three examples of semidefinite programs in which
θ̂ can be shown to belong to association schemes, with ` = 1, ` = 2, and ` ≤ K
respectively. We remark that association schemes were originally invented by
statisticians for the study of experiment design [4, 5], and that they have been
used with semidefinite programming in [12] to lower bound the optimal objective
value of the traveling salesperson problem (without finding a feasible solution).
5
Simulation results
To illustrate the usage and performance of the semidefinite program, we show
simulated results for three examples: community structure, overlapping communities, and latent space models.
5.1
Community Structure
In the best-understood blockmodel setting, A is generated by θ ∈ [0, 1]K×K
which has parameters γ0 , γ1 ∈ [0, 1], and equals
θ = γ0 I + γ1 (11T − I).
(23)
In this model, nodes connect with probability γ0 if they are in the same class,
and probability γ1 if they are in different classes.
Under (23), the parameter matrix θ can be written as
θ = γ0 B0 + γ1 B1 ,
for B0 = I and B1 = 11T −I. By manual verification, this can be seen to satisfy
the requirements given in Definition 1 for an association scheme with ` = 1 so
that fast methods can be used to evaluate the ADMM iterations.
Figure 1 shows an adjacency matrix A generated by (23), the estimated
class labels ẑ found by spectral clustering of P̂ (as discussed in Section 2.4),
and the estimates found by direct spectral clustering of A. In this instance,
the semidefinite program yields nearly perfect class estimates, while spectral
clustering of A fails due to its sparsity.
Figure 2 shows the average simulated performance, over a range of values
for the network size n and average degree nρ. We see that for nρ ≥ 10, the
10
0
200
400
600
800
0
200
400
600
800
nz = 14424
(a) Adj. Matrix A
(b) ẑ
(c) Spectral clusters
Missclassification Rate
Figure 1: Community structured blockmodel (23) with n = 800 nodes, K = 4
classes, average degree = 18, and unbalanced class sizes. (a) adjacency matrix A.
(b) similarity matrix for ẑ, which resulted in 4 miss-classified nodes. (c) shows
the similarity matrix for class estimates found by spectral clustering of A, which
split the largest community, failed to find the smallest two communities, and
resulted in 447 miss-classified nodes.
0.6
0.6
0.6
0.6
0.5
0.5
0.5
0.5
0.4
0.4
0.4
0.4
0.3
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.1
0
0
0.2
SDP
spectral
0.1
0
200
600
1000
Number of Nodes
(a) nρ = 5
200
600
1000
0
200
(b) nρ = 10
600
1000
(c) nρ = 20
200
600
1000
(d) nρ = 30
Figure 2: Misclassification rates for ẑ found by semidefinite programming (blue
line, solid) vs. spectral clustering of A (red line, dotted), under the community
structured blockmodel (23) with K = 4, n = {100, 200, 400, 800, 1200}, and
expected average degree nρ = {5, 10, 20, 30}. 100 simulations per trial, standard
errors shown.
semidefinite program has roughly constant error if the average degree is fixed,
which is consistent with Theorem 2 (and other known results for community
detection). In contrast, the misclassification rate for spectral clustering of A
increases with the sparsity of the graph. This is exemplified by subplot (d),
where spectral clustering of A performs well for small networks but degrades
severely as n increases.
5.2
Two Overlapping Groups of Communities
Let parameters γ0 , γ1 ∈ [0, 1] satisfy γ0 > γ1 , and let K = k 2 for some integer
k, so that each a ∈ [K] has a k-ary representation (a1 , a2 ) ∈ [k]2 . Let θ equal
(
γ0 if a1 = b1 or a2 = b2
θab =
a, b ∈ [K].
(24)
γ1 otherwise
11
In this model, there exist two types of community structure, each type comprised of k communities, where node belongs to one community of each type.
Two nodes have a higher probability of connection if they share at least one
community in common.
We observe that θ = γ0 B0 +γ0 B1 +γ1 B2 , where B0 , B1 , and B2 ∈ {0, 1}K×K
are given by
B0 = I
B1 = (I − 11T ) ⊗ I + I ⊗ (I − 11T )
B2 = (I − 11T ) ⊗ (I − 11T ).
By manual verification, {B0 , B1 , B2 } can be seen to satisfy the requirements
given in Definition 1 for an association scheme with ` = 2.
For this model, it may be of interest to not only compute ẑ and θ̂est , which
estimate z and θ up to label switching, but to also estimate the 2k overlapping
communities, which we will denote by C1 , . . . , C2k ⊂ [n]. It can be seen that
C1 , . . . , C2k are given by
Cl+(m−1)k = {i ∈ [n] : zim = l}
l ∈ [k], m ∈ [2],
where each zi ∈ [K] has k-ary representation (zi1 , zi2 ). Equivalently, C1 , . . . , C2k
may also be defined as follows: Let G denote a graph with K vertices, and with
edges induced by thresholding θ between γ0 and γ1 :
(
1 if θab ≥ (γ0 + γ1 )/2
Gab =
a, b ∈ [K].
0 otherwise.
It can be seen that G has 2k maximal cliques C1 , . . . , C2k , for which the communities C1 , . . . , C2k satisfy
C` = {i ∈ [n] : zi ∈ Cπ(`) }
` = 1, . . . , 2k,
(25)
for some permutation π of [2k].
Thus, to estimate C1 , . . . , C2k , we can construct θ̂est ∈ [0, 1]K×K as given by
(11), and estimate G by
γ0 + γ1
est
a, b ∈ [K],
(26)
Ĝab = 1 θ̂ab ≥
2
which has maximal cliques Ĉ1 , . . . , Ĉm0 ⊂ [K] for some m0 . Using the maximal
cliques of Ĝ, we can estimate overlapping communities Cˆ1 , . . . , Cˆm0 ⊂ [n] by
Cˆ` = {i ∈ [n] : ẑi ∈ Ĉ` },
` = 1, . . . , m0 .
(27)
We remark that even in settings where the model (24) is not valid, by (26)
and (27) the subsets Cˆ1 , . . . , Cˆm0 are still interpretable as overlapping subsets of
densely connected nodes.
12
0
200
400
600
800
0
200
400
600
800
nz = 40414
(a) Adjacency (b) Cˆ1 , Cˆ2 , Cˆ3 1 2 (c) Cˆ4 , Cˆ5 , Cˆ6 1 2 (d) Nonoverlap- (e)
Spectral
matrix A
34
34
ping model
clustering of A
Figure 3: Overlapping community model (24), with n = 800, nρ = 50, and
6 overlapping communities C1 , . . . , C6 . (a) adjacency matrix A. (b) similarity
matrix for Cˆ1 , Cˆ2 , Cˆ3 . (c) similarity matrix for Cˆ4 , Cˆ5 , Cˆ6 . (d) similarity matrix for
ẑ, but assuming non-overlapping model (23). (e) similarity matrix using spectral
clustering of A. Cˆ1 , . . . , Cˆ6 had 32 errors, while ẑ using the non-overlapping
model (23) had 397 errors, and spectral clustering of A had 187 errors.
Figure 3a shows an adjacency matrix A generated by (24), with k = 3 and
K = 9. The pattern of θ is clearly visible in A. Figures 3b and 3c show the
estimated communities Cˆ1 , . . . , Cˆ6 using (27). For comparison, Figure 3d shows ẑ
when the semidefinite program assumes (23) instead of (24), and Figure 3e shows
the estimate of z under spectral clustering of A. In this instance, Cˆ1 , . . . , Cˆ6
accurately estimate the true communities C1 , . . . , C6 , with a misclassification
rate of 0.04. This is consistent with Theorem 2, which predicts that P̂ will be
nearly block-structured, implying that the subsequent steps of estimating ẑ, θ̂est ,
and Ĝ will succeed as well. In contrast, the two alternative methods give poor
estimates for z, with misclassification rates of 0.50 and 0.23, respectively.
Figure 4 shows the average misclassification rate for Cˆ1 , . . . , Cˆ6 , in simulations over a range of values for the network size n and average degree nρ. For
comparison, the misclassification rate for spectral clustering of A (which estimates z instead of C1 , . . . , C6 ) is shown as well. The misclassification rate of the
semidefinite program is not quite constant in n for fixed nρ, suggesting that the
asymptotic results of Theorem 2 may require larger n for this model compared
to the results shown in Figure 2. However, as nρ increases the semidefinite program estimates the overlapping communities (and hence z as well) with much
better accuracy than spectral clustering of A, which shows little improvement
with increasing nρ.
5.3
Latent Space Models
We consider a latent space model, reminiscent of [18], in which each node is
assigned a latent coordinate vector y1 , . . . , yn ∈ [0, 1]d . Conditional on {yi }ni=1 ,
each dyad is independent Bernoulli with log odds given by
log odds Aij = −kyi − yj k/σ,
(28)
where σ ≥ 0 is a bandwidth parameter. In general, (28) is not equivalent to a
blockmodel.
13
Missclassification Rate
1
1
1
1
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0
0.4
SDP
spectral
0.2
0
500
1000
1500
Number of Nodes
(a) nρ = 30
500
1000
1500
(b) nρ = 40
500
1000
1500
(c) nρ = 50
0
500
1000
1500
(d) nρ = 60
Figure 4: Misclassification rates for Cˆ1 , . . . , Cˆ6 found by semidefinite programming (blue line, solid), and for estimates of z found by spectral clustering of A
(red line, dotted), under the overlapping community model (24) with K = 9,
k = 3, n = {400, 800, 1200, 1600}, expected average degree nρ = {30, 40, 50, 60},
and unbalanced class sizes. 100 simulations per trial, standard errors shown.
Let D ∈ Rn×n denote the matrix of squared distances between the latent
coordinates, given by
Dij = kyi − yj k2 .
It is known that the first d eigencoordinates of (I − 11T /n)D(I − 11T /n) will
recover y1 , . . . , yn up to a unitary transformation.
To estimate D, we will approximate (28) by a blockmodel with K = (2k)d
classes, where each class represents a coordinate in Rd , so that θ̂ equals
logit θ̂ab = −δ(γa , γb )/σ,
(29)
where γ1 , . . . , γK ∈ Rd , and δ is a distance metric. In order for θ̂ to belong to
an association scheme, we choose γ1 , . . . , γK extending beyond [0, 1]d to form a
grid in [0, 2]d , and choose δ to be a toric distance. That is, given a ∈ [K], let
(a1 , . . . , ad ) denote its (2k)-ary representation, and let γ1 , . . . , γK be given by
a
ad
1
,...,
γa =
,
a ∈ [K],
k
k
and let δ : [0, 2]d × [0, 2]d 7→ R equal the distance on a torus of circumference 2,
1/2
d
X
δ(γa , γb ) =
min(|γaj − γbj |, 2 − |γaj − γbj |)2 ,
j=1
where γaj is the jth element of the vector γa . Since θ̂ab depends on γa and γb
only through their element-wise differences |γaj −γbj |, for j = 1, . . . , d, it follows
that θ̂ can be written as a weighted sum of matrices
θ̂ =
2k−1
X
j1 =0
···
2k−1
X
γj1 ···jd C (j1 ) ⊗ · · · ⊗ C (jd ) ,
jd =0
14
where C (j) ∈ {0, 1}2k×2k for j = 0, . . . , 2k − 1 is the circulant matrix given by
(
1 if min(|a − b|, 2k − |a − b|) = j
(j)
Cab =
0 otherwise.
By manual inspection, it can be seen that C = {C (0) , . . . , C (2k−1) } satisfies the
requirements of an association scheme, and hence that also C ⊗ · · · ⊗ C is an
association scheme as well, with ` ≤ (2k)d = K. (See [5, Chapters 1-3] for a
complete treatment.)
Let X̂ denote the solution to the semidefinite program (8) with θ̂ given by
(29). To estimate D, let D̂ ∈ Rn×n equal the MAP-style estimate under X̂,
D̂ij = arg max
x∈R
K X
K
X
(ij)
X̂ab 1{δ(γa , γb )2 = x},
(30)
a=1 b=1
and let D̃ ∈ Rn×n denote the randomized estimate
(ij)
D̃ij = δ(γa , γb )2 with probability Xab .
(31)
Given D̂, we will take the first d eigencoordinates of (I − 11T /n)D̂(I − 11T /n)
to estimate y1 , . . . , yn (or similarly using D̃).
Corollary 2 bounds the error between the randomized estimate D̃ and the
true distances D, relative to rounding the coordinates {yi }ni=1 to their closest
points on a grid over [0, 1]d :
Corollary 2. Let L : R × R 7→ R denote the loss function
!
0
e−x /σ
e−x/σ
0
,
,
L(x, x ) = KL
1 + e−x/σ 1 + e−x0 /σ
and let η1 , . . . , ηkd ∈ R form a grid on [0, 1]d ,
a
ad
1
ηa =
,...,
,
k
k
a ∈ [k d ],
where (a1 , . . . , ad ) denotes the k-ary representation of a ∈ [k d ]. Then
q
n
n
n
n
p
1 XX
1 X X p
L
D
,
D̃
≤
min
L
Dij , kηzi − ηzj k
ij
ij
2
2
n ρ i=1 j=1
z∈[kd ]n n ρ
i=1 j=1
1
+ OP √
nρ
15
Proof. It holds that
q
n
n
n
n
p
1 XX
1 X X p
1
Dij , D̃ij ≤ minn 2
Dij , δ(γzi , γzj ) + OP √
L
L
n2 ρ i=1 j=1
nρ
z∈[K] n ρ
i=1 j=1
n
n
1 X X p
1
≤ min 2
Dij , δ(ηzi , ηzj ) + OP √
L
nρ
z∈[kd ]n n ρ
i=1 j=1
n
n
1
1 X X p
Dij , kηzi − ηzj k + OP √
≤ min 2
,
L
nρ
z∈[kd ]n n ρ
i=1 j=1
where the first inequality holds by Theorem 1 and definition of L; the second
inequality holds because the minimization is over a strictly smaller set than in
the previous line; and the last inequality holds because δ(x, x0 ) = kx − x0 k for
all (x, x0 ) ∈ [0, 1]d × [0, 1]d .
Figure 5 shows latent coordinates y1 , . . . , yn ∈ R2 arranged in a circle, from
which an adjacency matrix was generated by (28) with n = 500 and nρ = 20.
The figure also shows the estimated coordinates ŷ derived from the randomized
D̃ given by (31), from the MAP-style D̂ given by (30), and by applying the
USVT method of [10] to the adjacency matrix A, in which a spectral estimate
of P is constructed from A, and then inverted to form an estimate of D. In this
instance, both D̃ and D̂ yield estimates ŷ that are similar, and are substantially
more accurate than the USVT approach which failed due to the sparsity of A.
Figure 6 shows a different configuration for y1 , . . . , yn , with similar results.
Figure 7 shows the average simulated estimation accuracy using D̂, for a
range of values for the network size n and average degree nρ. For comparison,
the performance of the spectral USVT method is shown as well. We see that
the estimation error for D̂ is near-constant when the average degree is fixed. In
contrast, the estimation error for the spectral USVT method worsens with the
sparsity of A. This is exemplified by subplot (d), in which the USVT method
performs well for small networks but degrades severely as n increases.
A
A.1
Proof of Theorems 1 and 2
Intermediate Results
We first present intermediate results that will be used in the proof of Theorems
1 and 2. Let X ∗ denote any solution to the idealized problem
maximize hF̄, Xi over all X ∈ X .
Lemma 2 is the main technical result, and states that for the general matrix
model (3), X̂ nearly optimizes the desired objective function F̄ , even though
only the noisy proxy F is available. Its proof closely follows the approach of
[17].
16
(a) y
(b) ŷ from D̃
(c) ŷ from D̂
(d) ŷ, USVT
Figure 5: Latent space model (28), with n = 500, nρ = 20, and latent coordinates y1 , . . . , yn ∈ R2 arranged in a circle. (a) Latent coordinates y. (b)
estimated ŷ using randomized D̃ from (31). (c) estimated ŷ using MAP-style
D̂ from (30). (d) estimated ŷ, using USVT spectral method [10] directly on A.
RMS errors for (b), (c), and (d) were 0.08, 0.07, and 0.25 respectively.
(a) y
(b) ŷ from D̃
(c) ŷ from D̂
(d) ŷ, USVT
Estimation Error
Figure 6: Latent space model (28), with n = 1000, nρ = 140, and latent
coordinates y1 , . . . , yn ∈ R2 arranged in a cross. (a) Latent coordinates y. (b)
estimated ŷ using randomized D̃ from (31). (c) estimated ŷ using MAP-style
D̂ from (30). (d) estimated ŷ, using USVT spectral method [10] directly on A.
RMS errors for (b), (c), and (d) were 0.044, 0.040, and 0.11 respectively.
0.3
0.3
0.3
0.3
0.25
0.25
0.25
0.25
0.2
0.2
0.2
0.2
0.15
0.15
0.15
0.15
0.1
SDP
USVT
0.05
0
200
600
1000
Number of Nodes
(a) nρ = 5
0.1
0.1
0.1
0.05
0.05
0.05
0
200
400
600
800 1000 1200
(b) nρ = 15
0
200
400
600
800 1000 1200
(c) nρ = 25
0
200
400
600
800 1000 1200
(d) nρ = 35
Figure 7: Estimation error for ŷ found by semidefinite programing (blue line,
solid) vs. USVT method on A (red line, dotted), under the latent space model
(28) with n = {200, 300, 400, 600, 800, 1000, 1200}, expected average degree nρ =
{5, 15, 25, 35}, and latent coordinates y1 , . . . , yn ∈ R2 arranged in a circle. 100
simulations per trial, standard errors shown.
17
Lemma 2. Let Assumptions 1 and 2 hold. Then for some C1 , C2 > 0 it holds
for all n that
1
C1
∗
P
hF̄, X − X̂i ≥ √
= e−C2 n ,
n2 ρ
nρ
where C1 and C2 depend only on K and c.
Lemma 3 gives a condition under which P̂ and P̃ will both be approximately
block structured. It roughly states that if there exists z ∈ [K]n such that
(ij)
(zi , zj ) ∈ arg maxab F̄ab for all i, j ∈ [n], then both P̂ and P̃ will asymptotically
have block structure corresponding to z.
Lemma 3. Let Assumptions 1 and 2 hold. If there exists z ∈ [K]n and ∆ > 0
such that P, θ̂ and F̄ satisfy
(ij)
F̄z(ij)
≥ F̄ab + ρ∆
i ,zj
∀i, j ∈ [n] and a, b ∈
/ Qθ̂ (zi , zj ),
(32)
then it holds that
n
n
K
K
X X X X (ij)
1
X̂ab · 1{(a, b) ∈
/ Qθ̂ (zi , zj )} ≤ OP
n(n − 1) i=1 j=1 a=1
b=1
1
√
nρ
,
(33)
Lemma 4 states that the error between P and the randomized estimate P̃
converges to its expectation.
Lemma 4. Let Assumptions 1 and 2 hold. Then
n
n
n
n
i
XX
XX h
1
1
KL Pij , P̃ij =
E KL Pij , P̃ij
n(n − 1)ρ i=1 j=1
n(n − 1)ρ i=1 j=1
1
+ OP √
.
nρ
Bernstein’s inequality states that for independent x1 , . . . , xn satisfying |xi | ≤
b, with variance σi2 and expectation Exi = 0, it holds that
!
n
1X
nt2
P
xi ≥ t ≤ exp − 2 Pn
.
2
n i=1
i=1 σi + bt/3
n
Grothendieck’s inequality [17, Th. 3.1] states that there exists a universal
constant CG such that for any matrix M ∈ Rn×n ,
max |hM, Xi| ≤ CG
X∈M
max
s,t∈{−1,1}n
|sT M t|,
(34)
where M = {U V T : all rows of U, V ∈ B2n }, and B2n = {x ∈ Rn : kxk2 ≤ 1} is
the n-dimensional unit ball.
Lemmas 2, 3, and 4 are proven in Section A.3.
18
A.2
Proof of Theorems 1 and 2
Proof of Theorem 1. Given z ∈ [K]n , let the vector x(z) ∈ {0, 1}nK be given
by
T
x(z) = eTz1 · · · eTzn .
Theorem 1 holds by the the following steps:
n
n
XX
1
KL Pij , P̃ij
n(n − 1)ρ i=1 j=1
n X
n
h
i
X
1
1
=
E KL Pij , P̃ij + OP √
n(n − 1)ρ i=1 j=1
nρ
n
n X
X
1
1
H(Pij ) − hF̄, X̂i + OP √
=
n(n − 1)ρ i=1 j=1
nρ
n
n X
X
1
1
H(Pij ) − hF̄, X ∗ i + OP √
≤
n(n − 1)ρ i=1 j=1
nρ
(35)
(36)
(37)
n
n X
X
1
1
T
H(Pij ) − maxn hF̄, x(z)x(z) i + OP √
≤
(38)
n(n − 1)ρ i=1 j=1
nρ
z∈[K]
n
n X
X
1
1
KL Pij , θ̂zi ,zj + OP √
≤ minn
(39)
nρ
z∈[K] n(n − 1)ρ
i=1 j=1
where (35) holds by Lemma 4; (36) and (39) follow from the identity
(ij)
KL(Pij , θ̂ab ) = H(Pij ) − F̄ab ,
with (36) additionally using (10), the definition of P̃ ; (37) holds by Lemma 2;
and (38) holds because x(z)x(z)T ∈ X , implying that hF̄, X ∗ − x(z)x(z)T i ≥ 0.
Proof of Theorem 2. Define π̂ by
n
π̂a =
1X
1{zi = a}.
n i=1
By Bernstein’s
inequality, it can be seen that α̂ = ρ(1 + oP (1)), and that ρ =
P
1
P
satisfies
(n2 ) i<j ij
ρ=
K X
K
X
∗
π̂a π̂b θab
(1 + o(1))
a=1 b=1
= α(1 + oP (1)),
19
so that α̂ → ρ → α. As a result, Assumptions 1 and 2 can be seen to hold (in
probability). Let ∆0 be defined by
∆0 =
∗
min Bab
log
min
a,b∈[K]2 c,d∈Q
/ B∗
B̂ab
B̂cd
− (B̂ab − B̂cd ).
(ij)
(ij)
It can be seen that ∆0 > 0 by assumption. To bound F̄zi ,zj − F̄cd
apply Lemma 3, observe for all i, j ∈ [n] and c, d ∈
/ Qθ̂ (zi , zj ):
(ij)
F̄z(ij)
− F̄cd = Pij log
i ,zj
θ̂zi ,zj
θ̂cd
= αBz∗i ,zj log
= αBz∗i ,zj log
= ρBz∗i ,zj log
+ (1 − Pij ) log
B̂zi ,zj
B̂cd
B̂zi ,zj
B̂cd
B̂zi ,zj
B̂cd
≥ ρ∆0 + oP (α),
so as to
1 − θ̂zi ,zj
1 − θ̂cd
1 − α̂B̂zi ,zj
1 − α̂B̂zi ,zj
+ log
− αBz∗i ,zj log
1 − α̂B̂cd
1 − α̂B̂cd
+ α̂(B̂zi ,zj − B̂cd ) + O(α̂2 ) + O(αα̂)
− ρ(B̂zi ,zj − B̂cd ) + oP (α)
where the oP (α) terms are bounded uniformly over all (i, j) ∈ [n]2 and (c, d) ∈
[K]4 . This implies that for all i, j ∈ [n] and all c, d ∈
/ QB ∗ (zi , zj ), it holds that
(ij)
F̄z(ij)
− F̄cd ≥ ρ(∆0 + oP (1)),
i ,zj
implying that the conditions of Lemma 3 will hold in probability for any ∆ < ∆0 .
Lemma 3 thus implies
n K K
n
1
1 X X X X (ij)
X̂ab · 1{(a, b) ∈
/ QB ∗ (zi , zj )} = OP √
.
n2 i=1 j=1 a=1
nα
b=1
Let Eij for i, j ∈ [n] be given by
Eij =
K X
K
X
(ij)
X̂ab · 1{(a, b) ∈
/ QB ∗ (zi , zj )}.
a=1 b=1
To show (18), we apply (40) as follows:
n
n
n
n
1 XX
1 XX
1{P̂ij 6= θ̂zi zj } ≤ 2
1{Eij ≥ 1/2}
n2 i=1 j=1
n i=1 j=1
n
n
1 XX
2Eij
n2 i=1 j=1
1
= OP √
,
nα
≤
20
(40)
where the last equality follows from (40).
To show that (18) holds with P̃ in place of P̂ , we observe the following
bounds:
n X
n
n
n
X
1
1 XX
E 2
1{P̃ij 6= θ̂zi zj } = 2
Eij
n i=1 j=1
n i=1 j=1
1
= OP √
nα
and
n X
n
h
i
X
4 X
1
1{P̃ij 6= θ̂zi zj } = 4
Var 1{P̃ij 6= θij }
Var 2
n i=1 j=1
n i<j
4 X
P P̃ij 6= θij
≤ 4
n i<j
4 X
Eij
n4 i<j
1
1
.
= 2 OP √
n
nα
√
Applying Chebychev, which states that P |X − EX| ≥ k Var X ≤ k −2 , with
=
k = n3/4 α−1/4 yields that
n
n
1 XX
1
1{P̃ij 6= θ̂zi zj } = OP √
.
n2 i=1 j=1
nα
A.3
Proof of Lemmas 2, 3, and 4
Proof of Lemma 2. Let F̃ ∈ RnK×nK be a re-centered version of F̄ , with submatrices F̃ (ij) ∈ RK×K given by
F̃ (ij) = F̄ (ij) + (Aij − Pij ) log ρ · 11T .
Alegebraic manipulation yields that
hF̄, X ∗ i − hF̄, X̂i = hF̃, X ∗ i + hF̄ − F̃, X ∗ i − hF̃, X̂i − hF̄ − F̃, X̂i
= hF̃, X ∗ i − hF̃, X̂i
∗
∗
≤ hF̃, X i − hF, X i + hF, X̂i − hF̃, X̂i
(41)
(42)
≤ 2 max hF̄ − F, Xi
X∈X
≤ 2CG
max
s,t∈{−1,1}nK
21
sT (F − F̃ )t,
(43)
where (41) holds because F̃ (ij) − F̄ (ij) ∝ 11T for all i, j ∈ [n], implying that
hF̃ − F̄, Xi = hF̃ − F̄, X 0 i for all X, X 0 ∈ X ; (42) holds because X̂ maximizes
hF, Xi, implying that hF, X ∗ − X̂i ≤ 0; and (43) follows by Grothendieck’s
inequality and because X ⊂ M.
It remains to bound the right hand side of (43). By the definition of F̃ , it
can be seen that
(ij)
(ij)
Fab − F̃ab = (Aij − Pij ) log
θ̂ab /ρ
1 − θ̂ab
,
so that
sT (F − F̃ )t =
n X
n X
K X
K
X
(j)
s(i)
a tb (Aij − Pij ) log
i=1 j=1 a=1 b=1
θ̂ab /ρ
1 − θ̂ab
.
Given s, t ∈ {−1, 1}nK define xij ≡ xij (s, t) by
xij = (Aij − Pij )
K
K X
X
(j)
(j)
(j)
s(i)
a tb + sa tb
a=1 b=1
log
θ̂ab /ρ
1 − θ̂ab
,
P
so that sT (F − F̃ )t = i<j xij . Using ρ/c ≤ θ̂ab /(1 − θ̂ab ) ≤ cρ which holds by
Assumption 2 and letting C = 2| log c|, it can be seen that
Var(xij + xji ) ≤ Pij (1 − Pij )K 4 C 2
By Bernstein’s inequality, it follows that
X
1
xij ≥ ≤ exp −
P n
2
i,j:i<j
|xij | ≤ K 2 C.
and
n
2
2
K 4 C 2 ρ + 2K 2 C
!
,
−1 P
where we have used the fact that n2
i<j Pij (1 − Pij ) ≤ ρ. Letting =
pρ
C1 n for any value of C1 implies
r
1
X
nC 2
1
ρ
xij ≥ C1
≤ exp − 4 22 1 2
.
P n
n
K C + 2K C
2 i,j:i<j
Applying a union bound over all s, t ∈ {−1, 1}nK implies
r
1
2
X
1
ρ
nK
2 nC1
P
max
x
≥
C
≤
2
exp
−
,
ij
1
n
n
K 4 C 2 + 2K 2 C
s,t∈{−1,1}nK
2 i,j:i<j
which implies for all C1 satisfying 21 C12 /(K 4 C 2 + 2K 2 CC1 ) > K log 2 that
r
X
1
ρ
= e−C2 n ,
P
max
xij ≥ C1
(44)
n
n
s,t∈{−1,1}nK
2 i,j:i<j
where C2 = 12 C12 /(K 4 C 2 + 2K 2 CC1 ) − K log 2. Since sT (F − F̃ )t =
combining (44) and (43) proves the lemma.
22
P
i<j
xij ,
Proof of Lemma 3. It holds that
hF̄, X̂i =
n X
n X
K X
K
X
(ij)
(ij)
F̄ab X̂ab
i=1 j=1 a=1 b=1
=
n X
n
X
"
F̄z(ij)
i ,zj
+
≤
(ij)
F̄ab
−
F̄z(ij)
i ,zj
#
(ij)
X̂ab
a=1 b=1
i=1 j=1
n X
n
X
K X
K
X
"
F̄z(ij)
i ,zj
− ρ∆
K X
K
X
#
(ij)
X̂ab
· 1{(a, b) ∈
/ Qθ̂ (zi , zj )} .
a=1 b=1
i=1 j=1
Rearranging and using hF̄, X ∗ i =
n
P
n
ij
(ij)
F̄zi ,zj (as implied by (32)) yields
K
K
hF̄, X ∗ i − hF̄, X̂i X X X X (ij)
X̂ab · 1{(a, b) ∈
/ Qθ̂ (zi , zj )},
≥
ρ∆
i=1 j=1 a=1
b=1
and dividing both sides by n2 and using Lemma 2 yields that with probability
at least 1 − e−C2 n ,
n
n K K
C1 /∆
1 X X X X (ij)
X̂ab · 1{(a, b) ∈
/ Qθ̂ (zi , zj )},
≥
√
nρ
n2 i=1 j=1 a=1
b=1
proving the lemma.
Proof of Lemma 4. Let xij for i, j ∈ [n] be given by
xij = H(Pij ) − Pij log ρ − KL Pij , P̃ij .
It can be seen that xij for i < j are independent random variables with distributions
xij = Pij log
θ̂ab /ρ
1 − θ̂ab
+ log(1 − θ̂ab )
(ij)
with probability X̂ab ,
with xji = xij for i > j and xii = 0. By definition of {xij }, it holds that
n X
n
X
n X
n
h
i X
KL Pij , P̃ij − E KL Pij , P̃ij
=
(Exij − xij ) .
i=1 j=1
i=1 j=1
To bound the right hand side of (45), observe that by Assumption 2
|xij | ≤ Pij | log c| + | log(1 + cρ)|
cρ
≤ Pij | log c| +
1 + cρ
23
(45)
and hence it holds that
max |xij − Exij | ≤ O(1)
ij
X
1
Var(xij + xji ) ≤ O(ρ),
n(n − 1) i,j:i<j
Applying Bernstein’s inequality thus yields
X
1
(xij − Exij ) ≥ ≤ exp −
P n
2
i,j:i<j
1
and letting = (ρ/n)1/2 implies that
n
2
)ρ
P
i,j (xij
(
Combining this bound with (45) proves the lemma.
B
n
2
2
O(ρ) + O(1)
!
,
− Exij ) = OP (nρ)−1/2 .
Proof of Theorem 3
Lemma 1 can be found as E1-E4 in [15], and as Theorem 2.6 in [5].
Proof of Theorem 3. Let B = span(B0 , . . . , B` ), and let A denote the set
A = {X ∈ RnK×nK : X (ij) ∈ B
∀ i, j},
and observe the following properties of A:
1. If M ∈ A, then its submatrices M (ij) must be symmetric and have the
same eigenvectors since they are in the span of an association scheme.
Specifically, if X t+1 + V t ∈ A then (20) holds.
2. A is a linear space, so if M1 , M2 ∈ A, then M1 + M2 ∈ A as well.
3. 0 ∈ A, so the initial values X 0 , W 0 , Y 0 , U 0 , V 0 are in A
P`
4. Let B = i=0 γi Bi for some weights γ0 , . . . , γ` . BecausePB0 , . . . , B` are
binary with disjoint
support, it holds that max(0, B) = i max(0, γi )Bi
P
and log(B) = i log(γi )Bi .
5. If θ̂ ∈ B then log θ̂ and log(1 − θ̂) are in B as well, by property 4. This
implies the matrix F given by (12) is in A, since each submatrix F (ij) is
a linear combination of log θ̂ and log(1 − θ̂).
6. If M ∈ A then max(0, M ) is also in A, by property 4.
7. If M ∈ A, then the eigenvectors of each submatrix are orthogonal and
include the vector 1, by Lemma 1. As a result, the effect of the projection
ΠX (M ) is to change the eigenvalue associated with the vector 1 in each
submatrix, implying that ΠX (M ) ∈ A.
24
8. If X t+1 + V t ∈ A, then (20) and hence (21) hold, which implies that
ΠS+ (X t+1 + V t ) ∈ A.
Using these properties, we show by induction that X t , W t , Y t , U t , V t ∈ A for
all t. As for the base case, it holds that X 0 , W 0 , Y 0 , U 0 , V 0 ∈ A by property
3. Now suppose that X t , W t , Y t , U t , V t ∈ A for any t. Then by (19) it follows
that X t+1 ∈ A by properties 2, 5, and 7; that W t+1 ∈ A by properties 2 and 6;
that Y t+1 ∈ A by properties 2 and 8; and that U t+1 , V t+1 ∈ A by property 2.
This completes the induction argument.
Since X t , W t , Y t , U t , V t ∈ A for all t, it follows by property 2 that X t+1 +
t
V ∈ A for all t. By property 1, this implies that (20) holds for all t, proving
the theorem.
C
Proof of Corollary 1
Here we prove Corollary 1, which states that the eigencoordinates of P̂ or P̃ will
approximate those of P (up to a unitary transform) when Theorem 2 holds. This
suggests that ẑ, which is computed by spectral clustering of P̂ , will converge to
z up to label permutation.
C.1
Intermediate results
Lemma 5 bounds the eigenvalues of P under Assumptions 1 and 2 .
Lemma 5. Let Assumptions 1 and 2 hold. Let D ∈ [0, 1]K×K denote the
matrix D = diag(π). Let λ∗1 , . . . , λ∗K and λ1 , . . . , λK respectively denote the
sorted eigenvalues of D1/2 B ∗ D1/2 and P . It holds that
λk = nα(λ∗k (1 + oP (1)))
k = 1, . . . , K.
We will use the following version of the Davis-Kahan theorem, taken from
[30, Th. 4]:
Theorem 3. Let P, P̂ ∈ Rn×n be symmetric, with singular values σ1 ≥ . . . ≥ σn
2
and σ̂1 ≥ . . . ≥ σ̂n respectively. Fix 1 ≤ r ≤ s ≤ n and assume that min(σr−1
−
2
) > 0, where σ02 = ∞ and σn+1 = −∞. Let d = s−r+1, and let V =
σr2 , σs2 −σs+1
(vr , vr+1 , . . . , vs ) ∈ Rn×d and V̂ = (v̂r , v̂r+1 , . . . , v̂s ) ∈ Rn×d have orthonormal
columns satisfying P vj = σj uj and P̂ v̂j = σ̂j ûj for j = r, r + 1, . . . , s. Then
there exists orthogonal Ô ∈ Rd×d such that
kV̂ Ô − V kF ≤
23/2 (2σ1 + kP̂ − P kop )kP̂ − P kF
2
2 )
min(σr−1
− σr2 , σs2 − σs+1
25
C.2
Proof of Corollary 1 and Lemma 5
2
Proof of Corollary 1. Let r = 1 and s = K = rank(P ), so that σs2 − σs+1
= λ2K .
By Theorem 3, it holds that
kV̂ Ô − V
k2F
≤
23/2 (2σ1 + kP̂ − P kop )kP̂ − P kF
λ2K
!2
.
It follows that
kV̂ Ô − V
k2F
≤
=
!2
23/2 O(nα)kP̂ − P kF
(nαλ∗K )2 (1 + oP (1))
!2
23/2 O(1)kP̂ − P kF
,
nα(λ∗K )2 (1 + oP (1))
(46)
where in the first inequality follows from λK = nαλ∗K (1 + oP (1)) by Lemma 5,
and also from
2σ1 + kP̂ − P kop ≤ 3kP kop + kP̂ kop
≤ 3n max Pij + n max P̂ij
ij
≤
∗
3nα max Bab
ab
ij
+ nα max B̂ab .
ab
By Theorem 2 it holds that
n
n
1
1 XX
2
k
P̂
−
P
k
=
(P̂ij − Pij )2
F
n2
n2 i=1 j=1
n
n
1 XX
1{P̂ij − Pij } · max(P̂ij − Pij )2
i,j
n2 i=1 j=1
1
≤ OP √
· O(α2 ).
nα
≤
Substituting (47) into (46) yields
O(1)
kV̂ Ô − V k2F ≤ √ ,
nα
completing the proof.
Proof of Lemma 5. Let π̂ ∈ [0, 1]K be given by
n
1X
π̂a =
1{zi = a},
n i=1
26
(47)
and let D̂ = diag(π̂). Let ũ ∈ RK denote an eigenvector of D̂1/2 B ∗ D̂1/2 with
eigenvalue λ̂, let ṽ = D̂−1/2 ũ, and let v ∈ Rn be given by vi = ṽzi for i ∈ [n].
It can be seen that
[P v]i = [nαB ∗ D̂ṽ]zi
= [nαB ∗ D̂1/2 ũ]zi
= [nαD̂−1/2 D̂1/2 B ∗ D̂1/2 ũ]zi
= [nαD̂−1/2 λ̂ũ]zi
= [nαλ̂ṽ]zi
= nαλ̂vi ,
showing that v is an eigenvector of P with eigenvalue nαλ̂. Since D̂ → D, it
follows that the eigenvalues of D̂1/2 B ∗ D̂1/2 converge to those of D1/2 B ∗ D1/2 ,
completing the proof.
References
[1] Edoardo M Airoldi, David M Blei, Stephen E Fienberg, and Eric P Xing.
Mixed membership stochastic blockmodels. Journal of Machine Learning
Research, 9(Sep):1981–2014, 2008.
[2] Arash A Amini, Aiyou Chen, Peter J Bickel, Elizaveta Levina, et al.
Pseudo-likelihood methods for community detection in large sparse networks. The Annals of Statistics, 41(4):2097–2122, 2013.
[3] Arash A Amini and Elizaveta Levina. On semidefinite relaxations for the
block model. arXiv preprint arXiv:1406.5647, 2014.
[4] RA Bailey. Designs on association schemes. Lecture Notes-Monograph
Series, pages 79–102, 2003.
[5] Rosemary A Bailey. Association schemes: Designed experiments, algebra
and combinatorics, volume 84. Cambridge University Press, 2004.
[6] Afonso S Bandeira, Yutong Chen, and Amit Singer. Non-unique games
over compact groups and orientation estimation in cryo-em. arXiv preprint
arXiv:1505.03840, 2015.
[7] Peter J Bickel and Aiyou Chen. A nonparametric view of network models
and newman–girvan and other modularities. Proceedings of the National
Academy of Sciences, 106(50):21068–21073, 2009.
[8] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating
direction method of multipliers. Foundations and Trends R in Machine
Learning, 3(1):1–122, 2011.
27
[9] T Tony Cai, Xiaodong Li, et al. Robust and computationally feasible
community detection in the presence of arbitrary outlier nodes. The Annals
of Statistics, 43(3):1027–1059, 2015.
[10] Sourav Chatterjee et al. Matrix estimation by universal singular value
thresholding. The Annals of Statistics, 43(1):177–214, 2015.
[11] Yudong Chen, Xiaodong Li, and Jiaming Xu. Convexified modularity
maximization for degree-corrected stochastic block models. arXiv preprint
arXiv:1512.08425, 2015.
[12] Etienne De Klerk, Dmitrii V Pasechnik, and Renata Sotirov. On semidefinite programming relaxations of the traveling salesman problem. SIAM
Journal on Optimization, 19(4):1559–1573, 2008.
[13] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová. Inference and phase transitions in the detection of modules in
sparse networks. Physical Review Letters, 107(6):065701, 2011.
[14] Chao Gao, Zongming Ma, Anderson Y Zhang, and Harrison H Zhou.
Achieving optimal misclassification proportion in stochastic block model.
arXiv preprint arXiv:1505.03772, 2015.
[15] Michel X Goemans and Franz Rendl. Semidefinite programs and association
schemes. Computing, 63(4):331–340, 1999.
[16] Ronald L Graham. Handbook of combinatorics, volume 1. 1995.
[17] Olivier Guédon and Roman Vershynin. Community detection in sparse networks via grothendiecks inequality. Probability Theory and Related Fields,
pages 1–25, 2015.
[18] Peter D Hoff, Adrian E Raftery, and Mark S Handcock. Latent space
approaches to social network analysis. Journal of the american Statistical
association, 97(460):1090–1098, 2002.
[19] Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt.
Stochastic blockmodels: First steps. Social networks, 5(2):109–137, 1983.
[20] Brian Karrer and Mark EJ Newman. Stochastic blockmodels and community structure in networks. Physical Review E, 83(1):016107, 2011.
[21] Florent Krzakala, Cristopher Moore, Elchanan Mossel, Joe Neeman, Allan
Sly, Lenka Zdeborová, and Pan Zhang. Spectral redemption in clustering sparse networks. Proceedings of the National Academy of Sciences,
110(52):20935–20940, 2013.
[22] Vince Lyzinski, Daniel L Sussman, Minh Tang, Avanti Athreya, Carey E
Priebe, et al. Perfect clustering for stochastic blockmodel graphs via adjacency spectral embedding. Electronic Journal of Statistics, 8(2):2905–2922,
2014.
28
[23] Vince Lyzinski, Minh Tang, Avanti Athreya, Youngser Park, and Carey E
Priebe. Community detection and classification in hierarchical stochastic
blockmodels. arXiv preprint arXiv:1503.02115, 2015.
[24] Catherine Matias and Vincent Miele. Statistical clustering of temporal networks through a dynamic stochastic block model. arXiv preprint
arXiv:1506.07464, 2015.
[25] Andrea Montanari and Subhabrata Sen. Semidefinite programs on sparse
random graphs and their application to community detection. arXiv
preprint arXiv:1504.05910, 2015.
[26] Elchanan Mossel, Joe Neeman, and Allan Sly. Stochastic block models and
reconstruction. arXiv preprint arXiv:1202.1499, 2012.
[27] Tiago P Peixoto. Hierarchical block structures and high-resolution model
selection in large networks. Physical Review X, 4(1):011047, 2014.
[28] Karl Rohe, Sourav Chatterjee, and Bin Yu. Spectral clustering and the
high-dimensional stochastic blockmodel. The Annals of Statistics, pages
1878–1915, 2011.
[29] Jaewon Yang and Jure Leskovec. Overlapping community detection at
scale: a nonnegative matrix factorization approach. In Proceedings of the
sixth ACM international conference on Web search and data mining, pages
587–596. ACM, 2013.
[30] Yi Yu, Tengyao Wang, and Richard J Samworth. A useful variant of the
davis–kahan theorem for statisticians. Biometrika, 102(2):315–323, 2015.
[31] Anderson Y Zhang and Harrison H Zhou. Minimax rates of community
detection in stochastic block models. arXiv preprint arXiv:1507.05313,
2015.
[32] Pan Zhang and Cristopher Moore. Scalable detection of statistically significant communities and hierarchies, using message passing for modularity. Proceedings of the National Academy of Sciences, 111(51):18144–18149,
2014.
[33] Yuan Zhang, Elizaveta Levina, and Ji Zhu. Detecting overlapping communities in networks using spectral methods. arXiv preprint arXiv:1412.3432,
2014.
[34] Qing Zhao, Stefan E Karisch, Franz Rendl, and Henry Wolkowicz. Semidefinite programming relaxations for the quadratic assignment problem. Journal of Combinatorial Optimization, 2(1):71–109, 1998.
29
| 10math.ST
|
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS1
The Exact Solution to Rank-1 L1-norm TUCKER2
Decomposition
Panos P. Markopoulos,∗ Dimitris G. Chachlakis, and Evangelos E. Papalexakis
November 1, 2017
arXiv:1710.11306v1 [cs.DS] 31 Oct 2017
Abstract
We study rank-1 L1-norm-based TUCKER2 (L1-TUCKER2) decomposition of 3-way tensors, treated as a
collection of N D × M matrices that are to be jointly decomposed. Our contributions are as follows. i) We prove
that the problem is equivalent to combinatorial optimization over N antipodal-binary variables. ii) We derive the
first two algorithms in the literature for its exact solution. The first algorithm has cost exponential in N ; the second
one has cost polynomial in N (under a mild assumption). Our algorithms are accompanied by formal complexity
analysis. iii) We conduct numerical studies to compare the performance of exact L1-TUCKER2 (proposed) with
standard HOSVD, HOOI, GLRAM, PCA, L1-PCA, and TPCA-L1. Our studies show that L1-TUCKER2 outperforms
(in tensor approximation) all the above counterparts when the processed data are outlier corrupted.
Index Terms
Data analysis, L1-norm, outliers, robust, TUCKER decomposition, tensors.
I. I NTRODUCTION AND P ROBLEM S TATEMENT
Introduced by L. R. Tucker [1] in the mid-1960s, TUCKER decomposition is a fundamental method n-way
tensor analysis, with applications in a wide range of fields, including machine learning, computer vision [2], [3],
wireless communications [4], biomedical signal processing [5], and social-network data analysis [6], [7] to name
a few. Considering that the n-way tensor under processing is formed by the concatenation (say, across the n-th
mode, with no loss of generality) of a number of coherent (same class, or distribution) (n − 1)-way coherent
tensor measurements, then TUCKER decomposition simplifies to TUCKER2 decomposition. TUCKER2 strives to
jointly decompose the collected (n − 1)-way tensors and unveil the low-rank multi-linear structure of their class, or
distribution. Higher-Order SVD (HOSVD) and Higher-Order Orthogonal Iteration (HOOI) algorithms [8] are wellknown solvers for TUCKER2 (and TUCKER) decompositions. A detailed presentation of TUCKER, TUCKER2,
P. P. Markopoulos and D. G. Chachlakis are with the Department of Electrical and Microelectronic Engineering, Rochester Institute of
Technology, Rochester, NY 14623 USA (e-mail: [email protected], [email protected]).
E. E. Papalexakis is with the Department of Computer Science and Engineering, University of California Riverside, Riverside, CA 92521
USA (e-mail: [email protected]).
∗
Corresponding author.
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS2
and the respective solvers is offered in [9]–[11]. Note that both types of solvers can generally only guarantee a
locally optimal solution.
For n = 2, TUCKER/TUCKER2 take the familiar form of Principal-Component Analysis (PCA). Thus, similar to
PCA, TUCKER/TUCKER2 are sensitive against outliers within the processed tensor [12]–[14]. On the other hand,
L1-Principal-component Analysis (L1-PCA) [15]–[17], substituting the L2-norm in PCA by the outlier-resistant L1norm, has illustrated remarkable outlier-resistance. Extending this formulation to tensor processing, one can similarly
endow robustness to TUCKER and TUCKER2 decompositions by substituting the L2-norm in their formulations
by the L1-norm. Indeed, an approximate algorithm for L1-norm-based TUCKER2 (L1-TUCKER2) was proposed in
[12]. However, L1-TUCKER2 remains to date unsolved. In this work, we offer for the first time the exact solution
to L1-TUCKER2 for the special case of rank-1 approximation, and provide two optimal algorithms. A formal
problem statement follows.
Consider a collection of N real-valued matrices of equal size, X1 , X2 , . . . , XN ∈ RD×M . For any rank d ≤
PN
>
2
min{D, M }, a TUCKER2 decomposition strives to jointly analyze {Xi }N
i=1 , by maximizing
i=1 kU Xi VkF
over U ∈ RD×d and V ∈ RM ×d , such that U> U = V> V = Id ; then, Xi is low-rank approximated as
UU> Xi VV> . The squared Frobenius norm k · k2F returns the summation of the squared entries of its matrix
argument. Among other methods in the tensor-processing literature, TUCKER2 coincides with Multilinear PCA
[18] (for zero-centered matrices) and the Generalized Low-Rank Approximation of Matrices (GLRAM) [19]. Clearly,
for N = 1, TUCKER2 simplifies to the rank-d approximation of matrix X1 ∈ RD×M , solved by means of the
familiar singular-value decomposition (SVD) [20]; i.e., the optimal arguments U and V are built by the d left-hand
and right-hand singular vectors of X1 , respectively.
To counteract against the impact of any outliers in {Xi }N
i=1 , in this work, we consider the L1-norm-based
TUCKER2 reformulation
L1-TUCKER2:
maximize
D×d
>
U∈R
; U U=Id
V∈RM ×d ; V> V=Id
N
X
kU> Xi Vk1 ,
(1)
i=1
where the L1-norm k · k1 returns the summation of the absolute values of its matrix argument. The problem in (1)
was studied in [12] under the title L1-Tensor Principal-Component Analysis (TPCA-L1).1 Authors in [12] presented
an approximate algorithm for its solution which they employed for image reconstruction. To date, (1) has not been
solved exactly in the literature, even for the special case of rank-1 approximation (i.e., d = 1). In this work,
we deliver, for the first time, the exact solution to L1-TUCKER2 for d = 1, by means of two novel algorithms.
In addition, we provide numerical studies that demonstrate the outlier-resistance of exact L1-TUCKER2, and its
1
In this work, we refer to the problem as L1-TUCKER2, so as to highlight its connection with the TUCKER2 formulation (instead of
the general TUCKER formulation).
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS3
superiority (in joint-matrix decomposition and reconstruction) over L2-norm-based (standard) TUCKER2, GLRAM,
TPCA-L1, PCA, and L1-PCA.
II. E XACT S OLUTION
A. Reformulation into combinatorial optimization
For rank d = 1, L1-TUCKER2 in (1) takes the form
N
X
maximize
M ×1
u∈RD×1 ; v∈R
; kuk2 =kvk2 =1
|u> Xi v|
(2)
i=1
First, we focus on the absolute value in (2) and notice that, for any a ∈ RN ,
PN
i=1 |ai |
=
PN
i=1 sgn (ai ) ai
=
sgn (a)> a = maxb∈{±1}N b> a, where sgn (·) returns the {±1}-sign of its (vector) argument. In view of the
above, Lemma 1 follows.
Lemma 1. For any given u ∈ RD and v ∈ RM , it holds that
N
X
>
|u Xi v| =
max
b∈{±1}N
i=1
u
N
X
>
!
bi Xi
v.
(3)
i=1
The maximum in (3) is attained for b = [sgn u> X1 v , sgn u> X2 v , . . . , sgn u> XN v ]> .
In addition, the following well-known Lemma 2 derives by the matrix-approximation optimality of SVD [20].
Lemma 2. For any given b ∈ {±1}N , it holds that
max
>
u
u∈RD×1 ; kuk2 =1
v∈RM ×1 ; kvk2 =1
N
X
!
bi Xi
v = σmax
i=1
N
X
!
bi Xi
(4)
i=1
where σmax (·) returns the highest singular value of its matrix argument. The maximum in (4) is attained if u and
P
v are the left-hand and right-hand dominant singular vectors of N
i=1 bi Xi , respectively.
·
D×M N . Then, for any
To compact our notation, we concatenate {Xi }N
i=1 into X = [X1 , X2 , . . . , XN ] ∈ R
P
b ∈ {±1}N , it holds N
i=1 bi Xi = X(b ⊗ IM ), where ⊗ denotes the Kronecker matrix product [21]. Then, in
view of Lemma 1 and Lemma 2, we can rewrite the L1-TUCKER2 in (2) as
max
D×1
u∈R
; kuk2 =1
v∈RM ×1 ; kvk2 =1
=
max
b∈{±1}N
u∈RD×1 ; kuk2 =1
v∈RM ×1 ; kvk2 =1
N
X
|u> Xi v|
u> (X(b ⊗ IM )) v
= max N σmax (X(b ⊗ IM )) .
b∈{±1}
(5)
i=1
(6)
(7)
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS4
It is clear that (7) is a combinatorial problem over the size-2N feasibility set {±1}N . The following Proposition
1 derives straightforwardly from Lemma 1, Lemma 2, and (5)-(7) and concludes our transformation of (2) into a
combinatorial problem.
Proposition 1. Let bopt be a solution to the combinatorial
maximize
σmax (X(b ⊗ IM ))
N
(8)
b∈{±}
and denote by uopt ∈ RD and vopt ∈ RM the left- and right-hand singular vectors of X(bopt ⊗ IM ) ∈ RD×M ,
>
>
respectively. Then, (uopt , vopt ) is an optimal solution to (2). Also, bopt = [sgn u>
opt X1 vopt , . . . , sgn uopt XN vopt ]
P
>
>
>
and N
i=1 |uopt Xi vopt | = uopt (X(bopt ⊗ IM ))vopt = σmax (X(bopt ⊗ IM )). In the special case that uopt Xi vopt = 0,
for some i ∈ {1, 2, . . . , N }, [bopt ]i can be set to +1, having no effect to the metric of (8).
Given bopt , (uopt , vopt ) are obtained by SVD of X(bopt ⊗ IM ). Thus, by Proposition 1, the solution to L1TUCKER2 for low-rank d = 1 is obtained by the solution of the combinatorial problem (8) and a D-by-M SVD.
B. Connection to L1-PCA and hardness
In the sequel, we show that for M = 1 and d = 1, L1-TUCKER2 in (2) simplifies to L1-PCA [15]–[17].
·
Specifically, for M = 1, matrix Xi is a D × 1 vector, satisfying Xi = xi = vec(Xi ), and (2) can be rewritten as
max
N
X
u∈RD ; v∈R; kuk2 =|v|=1
|u> xi v|.
(9)
i=1
It is clear that for every u, an optimal value for v is trivially v = 1 (or, equivalently, v = −1); thus, for X =
[x1 , x2 , . . . , xN ] ∈ RD×N , (9) becomes
max
N
X
u∈RD ; kuk2 =1
|u> xi | =
i=1
max
u∈RD ; kuk2 =1
kX> uk1 ,
(10)
which is the exact formulation of the well-studied L1-PCA problem [15]–[17]. We notice also that for M = 1 the
combinatorial optimization (8) in Proposition 1 becomes
max
b∈{±1}N
σmax (X(b ⊗ 1)) =
max
b∈{±1}N
kXbk2 ,
(11)
since the maximum singular-value of a vector coincides with its Euclidean norm, which is in accordance to the
L1-PCA analysis in [16], [17]. Based of the equivalence of L1-PCA to (11), [16] has proven that L1-PCA of X
is formally NP-hard in N , for jointly asymptotic N and rank(X). Thus, by its equivalence to L1-PCA for d = 1
and M = 1, L1-TUCKER2 is also NP-hard in N , for jointly asymptotic N and rank(X).
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS5
Fig. 1. For ρ = 3 and N = 4, we draw W ∈ Rρ×N , such that WW> = I3 and Assumption1 holds
true. Then, we plot the nullspaces of
all 4 columns of W (colored planes). We observe that the planes partition R3 into K = 2( 30 + 31 + 32 ) = 2(1 + 3 + 3) = 14 coherent
cells (i.e., 7 visible cells above the cyan hyperplane and 7 cells below.)
C. Exact Algorithm 1: Exhaustive search
Proposition 1 shows how the solution to (2) can be obtained through the solution to the combinatorial problem
in (8). Our first exact algorithm solves (8) straightforwardly by an exhaustive search over its feasibility set. In
fact, noticing that σmax (·) is invariant to negations of its matrix argument, we obtain a solution bopt to (8) by
an exhaustive search in the size-2N −1 set Bex = {b ∈ {±1}N : b1 = 1}. For every value that b takes in Bex ,
we conduct SVD to X(b ⊗ IM ) to calculate σmax (X(b ⊗ IM )), with cost O(min{D, M }DM ) [20]. Since it
entails 2N −1 SVD calculations, the cost of this exhaustive-search algorithm is O(2N −1 min{D, M }DM ); thus, it
is exponential to the number of jointly processed matrices, N , and at most quadratic to the matrix sizes, D and M .
D. Exact Algorithm 2: Search with cost polynomial in N
In the sequel, we focus on the case where N is low-bounded by the constant DM and present an algorithm that
solves (2) with polynomial cost in N . DM < N emerges as a case of interest in signal processing applications
when {Xi }N
i=1 are measurements of a D × M fixed-size sensing system (e.g., D × M images). By Proposition 1,
for the optimal solutions bopt and (uopt , vopt ) of (8) and (2), respectively, it holds
>
>
>
>
X>
u
,
.
.
.
,
sgn
v
X
u
bopt = [sgn vopt
opt N opt ] ,
1 opt
(12)
>
with sgn u>
opt Xi vopt = +1, if uopt Xi vopt = 0. In addition, for every i ∈ {1, 2, . . . , N }, we find that
>
>
>
>
vopt
X>
u
=
Tr
X
u
v
opt
opt
i
i
opt = xi (vopt ⊗ uopt ).
(13)
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS6
Therefore, defining Y = [x1 , x2 , . . . , xN ] ∈ RDM ×N , (12) can be rewritten as
bopt = sgn Y> (vopt ⊗ uopt ) .
(14)
svd
Consider now that Y is of some rank ρ ≤ min{DM, N } and admits SVD Y = QSW, where Q> Q = WW> =
·
Iρ and S is the ρ×ρ diagonal matrix that carries the ρ non-zero singular-values of Y. Defining popt = S> Q> (vopt ⊗
uopt ), (15) can be rewritten as
bopt = sgn W> popt .
(15)
In view of (15) and since sgn (·) is invariant to positive scalings of its vector argument, an optimal solution to (8),
bopt , can be found in the binary set
B = {b ∈ {±1}N : b = sgn W> c , c ∈ Rρ }.
(16)
Certainly, by definition, (16) is a subset of {±1}N and, thus, has finite size upper bounded by 2N . This, in turn,
implies that there exist instances of c ∈ Rρ that yield the same value in sgn W> c . Below, we delve into this
observation to build a tight superset of B that has polynomial size in N , under the following mild “general position”
assumption [22].
Assumption 1. For every I ⊂ {1, 2, . . . , N } with |I| = ρ − 1, it holds that rank([W]:,I ) = ρ − 1; i.e., any
collection of ρ − 1 columns of W are linearly independent.
·
For any i ∈ {1, 2, . . . , N }, define wi = [W]:,i and denote by Ni the nullspace of wi . Then, for every c ∈ Ni , the
(non-negative) angle between c and wi , φ(c, wi ), is equal to
π
2
and, accordingly, wi> c = kck2 kwi k2 cos (φ(c, wi )) =
0. Clearly, the hyperplane Ni partitions Rρ in two non-overlapping halfspaces, Hi+ and Hi− [23], such that
sgn c> wi = +1 for every c ∈ Hi+ and sgn c> wi = −1 for every c ∈ Hi− . In accordance with Proposition 1,
we consider that Hi+ is a closed set that includes its boundary Ni , whereas Hi− is open and does not overlap with
Ni . In view of these definitions, we proceed with the following illustrative example. Consider some ρ > 2 and two
+ , H− }
column indices m < i ∈ {1, 2, . . . , N }. Then, hyperplanes Nm and Ni divide Rρ in the halfspace pairs {Hm
m
and {Hi+ , Hi− }, respectively. By Assumption 1,2 each one of the two halfspaces defined by Nm will intersect with
+ ∩ H+ , C = H+ ∩ H− ,
both halfspaces defined by Ni , forming the four halfspace-intersection “cells” C1 = Hm
2
m
i
i
− ∩ H− , C = H− ∩ H+ . It is now clear that, for any k ∈ {1, 2, 3, 4}, [sgn [W]> c ]
C3 = Hm
4
m,i is the same for
m
i
i
every c ∈ Ck . For example, for every c ∈ C2 , it holds that [sgn [W]> c ]m = +1 and [sgn [W]> c ]i = −1.
Next, we go one step further and consider the arrangement of all N hyperplanes {Ni }N
i=1 . Similar to our discussion
2
If wm and wi are linearly independent, then Nm and Ni intersect but do not coincide.
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS7
above, these hyperplanes partition Rρ in K cells {Ck }K
k=1 , where K depends on ρ and N . Formally, for every k ,
the k -th halfspace-intersection set is defined as
·
Ck =
\
Hi+
\
−
Hm
,
(17)
m∈Ik−1
i∈Ik+
for complementary index sets Ik+ and Ik− that satisfy Ik+ ∩ Ik− = ∅ and Ik+ ∪ Ik− = {1, 2, . . . , N } [24], [25]. By
the definition in (17), and in accordance with our example above, every c ∈ Ck lies in the same intersection of
halfpsaces and, thus, yields the exact same value in sgn W> c . Specifically, for every c ∈ Ck , it holds that
h
sgn W> c
+1, i ∈ I +
k
= sgn wi> c =
.
−1, i ∈ I −
i
i
(18)
k
·
In view of (18), for every k ∈ {1, 2, . . . , K} and any c ∈ Ck , we define the “signature” of the k -th cell bk =
ρ
sgn W> c . Moreover, we observe that Ck ∩ Cl = ∅ for every k 6= l and that ∪K
k=1 Ck = R . By the above
observations and definitions, (16) can be rewritten as
B=
K
[
{sgn W> c : c ∈ Ck } = {b1 , b2 , . . . , bK }.
(19)
k=1
Importantly, in [24], [26], it was shown that the exact number of coherent cells formed by the nullspaces of N
points in Rρ that are in general position (under Assumption 1) is exactly
ρ−1
X
N −1
K=2
≤ 2N ,
j
(20)
j=0
with equality in (20) if and only if ρ = N . Accordingly, per (20), the cardinality of B in (16) is equal to |B| =
P
N −1
2 ρ−1
. For clarity, in Fig. 1, we plot the nullspaces (colored planes) of the columns of arbitrary W ∈ R3×4
j=0
j
that satisfies both WW> = I3 and Assumption 1. It is interesting that exactly K = 14 < 24 = 16 coherent cells
emerge by the intersection of the formed halfspaces. In the sequel, we rely on (19) to develop a conceptually simple
method for computing a tight superset of the cell signatures in B .
·
Under Assumption 1, for any I ⊆ {1, 2, . . . , N } with |I| = ρ − 1, the hyperplane intersection VI = ∩i∈I Ni
is a line (1-dimensional subspace) in Rρ . By its definition, this line is the verge between all cells that are jointly
bounded by the ρ − 1 hyperplanes in {Ni }i∈I . Consider now a vector c ∈ Rρ that crosses over the verge VI (at any
point other than 0ρ ). By this crossing, the value of [sgn W> c ]I will change so that sgn W> c adjusts to the
signature of the new cell to which c just entered. At the same time, a crossing over VI cannot be simultaneously
·
over any of the hyperplanes in {Ni }i∈I c , for I c = {1, 2, . . . , N } \ I ; this is because, under Assumption 1, it is
only at 0ρ that more than ρ − 1 hyperplanes can intersect. Therefore, it is clear that [sgn W> c ]I c will remain
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS8
Algorithm 2: Polynomial in N
Input: {Xi }N
i=1
0: Y ← [vec(X1 ), vec(X2 ), . . . , vec(XN )]
1: (Q, Sd×d , W) ← svd(Y), mt ← 0
2: For every I ⊆ {1, 2, . . . , N }, |I| = d − 1
3:
Build BI in (21)
4:
For every b ∈ BI
5:
(U, Σ, V) ← svd(X(b ⊗ IM ))
6:
m ← max{diag(Σ)}
7:
if m > mt ,
8:
mt ← m, bt ← b, u ← [U]:,1 , v ← [V]:,1
Output: bopt ← bt , uopt ← u, and vopt ← v
Fig. 2. Algorithm for the exact solution of rank-1 L1-TUCKER2 in (2), with cost O(N ρ+1 ).
invariant during this crossing and, in fact, equal to [sgn W> v ]I c , for any v ∈ VI with v> c > 0. In view of the
above, for any v ∈ VI \ 0ρ , the set
·
BI = {b ∈ {±1}N : [b]I c = [sgn W> v ]I c }
(21)
contains the signatures of all sets that are bounded by the verge VI . Moreover, it has been shown (see, e.g., [26])
that, for every cell, there exists at least one such verge that bounds it. Therefore, it derives that the set
[
Bpol =
BI
(22)
I⊂{1,2,...,N }; |I|=ρ−1
includes all cell signatures and, thus, is a superset of B . We notice that, for every I , BI has size 2ρ−1 . Since I
N
N
can take ρ−1
distinct values, we find that Bpol is upper bounded by 2ρ−1 ρ−1
. Thus, both |Bpol | and |B| are
polynomial, in the order of O(N ρ−1 ).
Practically, for every I , v can be calculated by Gram-Schmidt orthogonalization of [W]:,I with cost O(ρ3 ).
N
Keeping the dominant terms, the construction of Bpol costs O(N ρ−1 ) and can be parallelized in ρ−1
processes.
Then, testing every entry of Bpol for optimality in (8) costs an additional O(N ). Thus, the overall cost of our second
algorithm, taking also into account the O(N ) (for constant DM ) SVD cost for the formation of W, is O(N ρ ).
The presented algorithm is summarized in Fig. 2.
III. N UMERICAL S TUDIES
20×20 where A = b uv> and kuk = kvk = 1, b ∼ N (0, 49),
Consider {Xi }14
i
i
2
2
i
i=1 such that Xi = Ai + Ni ∈ R
and each entry of Ni is additive white Gaussian noise (AWGN) from N (0, 1). We consider that Ai is the rank1 useful data in Xi that we want to reconstruct, by joint analysis (TUCKER2-type) of {Xi }14
i=1 . By irregular
corruption, 30 entries in 2 out of the 14 matrices (i.e., 60 entries out of the total 5600 entries in {Xi }14
i=1 ) have
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS9
14000
12000
10000
8000
6000
4000
2000
0
6
8
10
12
14
16
18
20
22
Fig. 3. Reconstruction MSE versus corruption variance σc2 (dB).
14
been further corrupted additively by noise from N (0, σc2 ). To reconstruct {Ai }14
i=1 from {Xi }i=1 , we follow one
of the two approaches below.
In the first approach, we vectorize the matrix samples and perform standard matrix analysis. That is, we obtain the
first (d = 1) principal component (PC) of [vec(X1 ), vec(X2 ), . . . , vec(XN )], q. Then, for every i, we approximate
Ai by Âi = mat(qq> ai ), where mat(·) reshapes its vector argument into a 20 × 20 matrix, in accordance with
vec(·). In the second approach, we process the samples in their natural form, as matrices, analyzing them by
TUCKER2. If (u, v) is the TUCKER2 solution pair, then we approximate Ai by Âi = uu> Xi vv> . For the first
approach, we obtain q by PCA (i.e., SVD) and L1-PCA [16]. For the second approach, we conduct TUCKER2
by HOSVD [3], HOOI [9], GLRAM [19], TPCA-L1 [12], and the proposed exact L1-TUCKER2. Then, for each
P14
2
reconstruction method, we measure the mean of the squared error
i=1 kAi − Âi kF over 1000 independent
realizations for corruption variance σc2 = 6, 8, . . . , 22dB. In Fig. 3, we plot the reconstruction mean squared error
(MSE) for every method, versus σc2 . We observe that PCA and L1-PCA exhibit the highest MSE due to the
vectorization operation (L1-PCA outperforms PCA clearly, across all values of σc2 ). Then, all TUCKER2-type
methods perform similarly well when σc2 is low. As the outlier variance σc2 increases, the performance of L2-normbased TUCKER2 (HOSVD, HOOI) and GLRAM deteriorates severely. On the other hand, the L1-norm-based
TPCA-L1 exhibits some robustness. The proposed exact L1-TUCKER2 maintains the sturdiest resistance against
the corruption, outperforming its counterparts across the board.
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS10
R EFERENCES
[1] L. R. Tucker, “Some mathematical notes on three-mode factor analysis,” Psychometrika, vol. 31, pp. 279-311, 1966.
[2] M. A. O. Vasilescu and D. Terzopoulos, “Multilinear analysis of image ensembles: Tensorfaces, in Proc. 7th European Conf. on Comput.
Vision (ECCV), Copenhagen, Denmark, May 2002, pp. 447-460.
[3] L.D. Lathauwer, B.D. Moor, and J. Vandewalle, “A multilinear singular value decomposition,” SIAM J. Matrix Anal. App., vol. 21, pp.
1253-1278, 2000.
[4] P. R. B. Gomes, A. L. F. de Almeida, and J. P. C. L. de Costal, “Fourth-order tensor method for blind spatial signature estimation,” in
Proc. IEEE Int. Conf. Acoust., Speech Signal Process., Florence, Italy, May 2014, pp. 2992-2996.
[5] M. Mørup, L. Hansen, K. Lars, and S. M. Arnfred, “Algorithms for sparse nonnegative TUCKER decompositions,” Neural computation
(MIT Press), vol. 20, pp. 2112-2131, 2008.
[6] J. Sun, S. Papadimitriou, C. Lin, N. Cao, S. Liu, and W. Qian, “Multivis: Content-based social network exploration through multi-way
visual analysis,” in Proc. SIAM Int. Conf. Data Mining (SDM), Sparks, NV, May 2009, pp. 1064-1075.
[7] T. G. Kolda and J. Sun, “Scalable tensor decompositions for multi-aspect data mining,” in Proc. IEEE Int. Conf. Data mining, Pisa,
Italy, Dec. 2008, pp. 363-372.
[8] L. D. Lathauwer, B. De Moor, and J. Vandewalle, “On the best rank-1 and rank-(R1,R2,.,RN) approximation of higher-order tensors,”
SIAM J. Matrix Anal. Appl., vol. 21, pp. 1324-1342, May 2000.
[9] T. G. Kolda and B. W. Bader, “Tensor decompositions and applications,” SIAM Review, vol. 51, pp. 455-500, 2009.
[10] J. Sun, D. Tao, S. Papadimitriou, P. Yu, and C. Faloutsos, “Incremental tensor analysis: Theory and applications, ACM Trans. Knowl.
Disc. Data, vol. 2, pp. 1-37, Oct. 2008.
[11] N. Sidiropoulos, L. D. Lathauwer, X. Fu, K. Huang, E. Papalexakis, and C. Faloutsos, “Tensor decomposition for signal processing
and machine learning,” IEEE Trans. Signal Process., vol. 63 no.13, pp. 3551-3582, Jul. 2017.
[12] Y. Pang, X. Li, and Y. Yuan, “Robust tensor analysis with L1-norm,” IEEE Trans. Circuits Syst. Video Technol., vol. 20, pp. 172-178,
Feb. 2010.
[13] X. Fu, K. Huang, W. K. Ma, N. D. Sidiropoulos, and R. Bro, “Joint tensor factorization and outlying slab suppression with applications,”
IEEE Trans. Signal Process., vol. 63, no. 23, pp. 6315-6328, Dec. 2015.
[14] X. Cao, X. Wei, Y. Han, and D. Lin, “Robust face clustering via tensor decomposition,” IEEE Trans. Cybern., vol. 45, no. 11, pp.
2546-2557, Nov. 2015.
[15] P. P. Markopoulos, G. N. Karystinos, and D. A. Pados, “Some options for L1-subspace signal processing,” in Proc. 10th Intern. Sym.
Wireless Commun. Sys. (IEEE ISWCS 2013), Ilmenau, Germany, Aug. 2013, pp. 1-5.
[16] P. P. Markopoulos, G. N. Karystinos and D. A. Pados, “Optimal algorithms for L1-subspace signal processing,” IEEE Trans. Signal
Process., vol. 62, pp. 5046-5058, Oct. 2014.
[17] P. P. Markopoulos, S. Kundu, S. Chamadia, and D. A. Pados, “Efficient L1-norm Principal-Component Analysis via bit flipping,” IEEE
Trans. Signal Process., vol. 65, pp. 4252-4264, Aug. 2017.
[18] H. Lu, K. N. Plataniotis and A. N. Venetsanopoulos, “MPCA: Multilinear principal component analysis of tensor objects,” IEEE Trans.
Neural Net., vol. 19, pp. 18-39, Jan. 2008.
[19] J. Ye, “Generalized low rank approximations of matrices,” Mach. Lear., vol. 61, pp. 167-191, Nov. 2005.
[20] G. H. Golub and C. F. Van Loan, Matrix Computations. (3rd ed.) Baltimore, MD: Johns Hopkins University Press, 1996.
[21] C. F. Van Loan, “The ubiquitous Kronecker product, J. Comput. Appl. Math., pp. 85-100, 2000.
[22] P. B. Yale, Geometry and Symmetry. San Francisco, CA: Holden-Day, 1968.
[23] P. Orlik and H. Terao, Arrangements of Hyperplanes. New York, NY: Springer-Verlag, 1992.
THIS IS A PREPRINT; AN EDITED/FINALIZED VERSION OF THIS MANUSCRIPT HAS BEEN SUBMITTED TO IEEE SIGNAL PROCESSING LETTERS11
[24] R. O. Winder, ”Partitions of N-Space by hyperplanes,” SIAM J. Appl. Math., vol. 14, pp. 811-818, 1966.
[25] S. K. Brown and P. Diaconis, “Random walks and hyperplane arrangements,” Ann. of Probability, pp. 1813-1854, 1998.
[26] G. N. Karystinos and A. P. Liavas, “Efficient Computation of the Binary Vector That Maximizes a Rank-Deficient Quadratic Form,”
IEEE Trans. Inf. Theory, vol. 56, pp. 3581-3593, Jul. 2010.
| 8cs.DS
|
How to Round Subspaces: A New Spectral Clustering
Algorithm
Ali Kemal Sinop∗
arXiv:1503.00827v5 [cs.DS] 19 Oct 2015
October 20, 2015
Abstract
A basic problem in spectral clustering is the following. If a solution obtained from the
spectral relaxation is close to an integral solution, is it possible to find this integral solution even
though they might be in completely different basis? In this paper, we propose a new spectral
clustering algorithm. It can recover
√ a k-partition such that the subspace corresponding to the
span of its indicator vectors is O( OPT) close to the original subspace in spectral norm with
OPT being the minimum possible (OPT ≤ 1 always). Moreover our algorithm does not impose
any restriction on the cluster sizes. Previously, no algorithm was known which could find a
k-partition closer than o(k · OPT).
We present two applications for our algorithm. First one finds a disjoint union of bounded
degree expanders which approximate a given graph in spectral norm. The second one is for
approximating the sparsest k-partition in a graph where each cluster have expansion at most
φk provided φk ≤ O(λk+1 ) where λk+1 is the (k + 1)st eigenvalue of Laplacian matrix. This
significantly improves upon the previous algorithms, which required φk ≤ O(λk+1 /k).
1
Introduction
In this paper, we study the following problem. If the solution of spectral relaxation for some
k-way partitioning problem is close to an integral solution, can we still find this integral solution?
The main difficulty is due to the rotational invariance of the spectral relaxation. The basis of an
integral solution might be completely different than the basis of solutions for the spectral relaxation.
Arguably, this is an important problem in spectral clustering, which is a widely used approach for
many data clustering and graph partitioning problems arising in practice.
In spectral clustering, one uses the top (or bottom) k-eigenvectors of some matrix derived from
the input (usually the Laplacian or adjacency matrix of some graph derived from the distances
or nearest neighbors) to find a k-partition. If the clusters are separated in a nice way, then these
k-eigenvectors will be close a k-partition up to an arbitrary rotation. Hence a crucial part of spectral
clustering methods is how to “round” these k-eigenvectors to a close-by k-partition.
Formally, we study the problem of approximating a k-dimensional linear subspace of Rn by
another subspace which is k-piecewise constant: Every vector of this subspace has its coordinates
∗
Email: [email protected]. This work was done while author was visiting Simons Institute for the Theory of
Computing, University of California at Berkeley. This material is based upon work supported by the National Science
Foundation under Grant No. 1540685.
1
comprised of at most k distinct values. Or equivalently, given a k-by-n orthonormal matrix Y of
the form Y = [y1 , . . . , yn ] (think of Y as an embedding of n points in Rk ), our problem is to find a
k-partition Γ = {T1 , . . . , Tk } so as to minimize the total variance under any direction:
XX
min
max
hz, yu − cS i2 .
Γ
z∈Rk :kzk2 =1
S∈Γ u∈S
Here cS is the mean of points in the cluster S. If we use C ∈ Rn×k to denote the matrix of cluster
means with each row of C being one of the cluster means, then our objective can be stated more
concisely as minC kY − Ck2 with k · k2 being the spectral norm. Geometrically speaking, this
corresponds to finding a k-piecewise constant subspace that makes the minimum angle with Y .
This is the problem of clustering with spectral norm [KK10]. In 2-dimensions, where k = 2, optimal
solution corresponds to one of the threshold cuts. From this perspective, our problem can be seen
as a generalization of thresholding to higher dimensions.
Our main contribution is a new spectral clustering
algorithm that can recover a k-partition
√
0
0
whose center matrix C satisfies kY − C k2 ≤ O( OPT), where OPT is √
the minimum possible
(observe that OPT ≤ 1). Furthermore, the recovered k-partition will be O( OPT)-close in Jaccard
index to the optimum partition: Each cluster we found will be close to a unique cluster among
the optimum k-partition. Previously, no algorithm was known to find a k-partition closer than
o(k · OPT).
We also study two closely related problems. In the first one, the goal is to approximate a matrix
in spectral norm by a block diagonal matrix, with every block being the normalized adjacency
matrix of a clique. In our second application, we turn to the problem of k- EXPANSION. Given an
undirected, weighted graph G = (V, C); find a k-partition Γ = {S1 , . . . , Sk } of the nodes so as to
minimize the maximum expansion:
def
Φk (G) = min max
Γ
S∈Γ
C(S, S)
.
min(|S|, |S|)
(1)
Here C(S, S) denotes the total weight of edges crossing S. Our second application is for approximating the optimum k-partition of k- EXPANSION on graphs whose spectrum grows faster than φk
(we will make this precise later).
The choice of spectral norm to measure the closeness of associated subspaces is quite natural
from the perspective of our second application. Given any subspace, we show how to construct
graphs in polynomial time, such that approximating k- EXPANSION on such graphs implies a
solution for the spectral clustering problem. From this perspective, we can see that the subspace
rounding problem is a prerequisite toward obtaining a o(k)-factor approximation algorithm for
the problem of k- EXPANSION, where the best known is O(k 4 ) due to [LGT14].
1.1
Related Work
Spectral methods have been successfully used for clustering tasks [Bol13] arising in many different
areas such as VLSI [AKY99], machine learning, data analysis [NJW01] and computer vision [SM00,
YS03]. They are usually obtained by formulating the clustering task as a combinatorial optimization
problem (such as sparsest/normalized cuts [SM00]), then solving the corresponding basic SDP
relaxation, whose solution is often given by k extremal eigenvectors of an associated matrix.
2
One of the first spectral clustering algorithms with worst case guarantees was given in [KVV04]
for the graph partitioning problem assuming certain conditions on the internal versus external
conductance. The problem of finding a k-partition so as to minimize the spectral norm was
first introduced by [KK10] in the context of learning mixtures of Gaussians. The best known
approximation factor is O(k) due to [AS12].
A problem closely related to spectral clustering is k- EXPANSION, as defined
in (1). When
√
all cluster sizes are constrained to be nearly equal, this problem admits a O( log n log k)-factor
approximation [BFK+ 11]. On the other hand, if a bi-criteria approximation
is sought, then one can
√
find (1 − Ω(1))k clusters each of which has expansion at most O( log n log k) times the optimum
[LM14].
If we look at the basic SDP relaxation of k- EXPANSION, then the optimal fractional solution is
given by the k smallest eigenvectors of the corresponding graph Laplacian matrix. In fact, this is
the main motivation behind the usage of k-eigenvectors for clustering tasks in practice. A natural
question is whether one can “round” these eigenvectors to a k-partition (the so called Cheeger
inequalities).
When k = 2, it was shown in [AM85] that simple thresholding yields a 2-partition
√
with O( φ2 ) expansion, where φk is the optimal value for k- EXPANSION. Later a better bound
was given in [KLL+ 13], assuming there is some gap between eigenvalues. When k > 2, bi-criteria
versions of Cheeger’s inequality are known [ABS10, LRTV12, LGT14]. Here the guarantees on the
√
e φk ), where O
e hides the dependencies on logarithmic factors; but the
expansion are of the form O(
algorithms can only find (1 − Ω(1))k parts.
The problem becomes significantly harder when exactly k clusters are desired. In this case,
it was shown in [LGT14] that a method similar
to the one proposed in [NJW01] will yield a k√
partition with maximum expansion O(k 4 φk ). This is the best known approximation algorithm
for k- EXPANSION problem and, as of yet, there is no algorithm known which achieves a polylogarithmic approximation.
Perhaps the simplest case of k- EXPANSION is when there is a gap between the (k + 1)st smallest
λ
≥ 1ε . One might think of this as a
eigenvalue of Laplacian matrix, λk+1 , and φk of the form φk+1
k
stability criteria: It implies that all k-partitions with maximum expansion ≤ O(φk ) are O(ε)-close
to each other. To put it in another way, approximating the optimum k-partition is at least as easy as
finding a k-partition with minimum possible expansion among all its clusters. For the case of k = 2,
it is trivial to show that thresholding the second smallest eigenvector of Laplacian yields ε-close
partition to the optimal one. On the other hand, when k > 2, the best prior result is due to [AS12],
which can find a k-partition that is O(kε)-close to the optimal one. In other words, when ε k1 ,
there is no algorithm known to find a non-trivial approximation of the optimum k-partition.
1.2
Organization
We first introduce some useful notation and background in Section 2. After this, we state our
main contributions in Section 3. Then we propose a new spectral clustering algorithm in Section 4.
√
In Section 5, we will prove that our algorithm always finds a k-partition that is ε-close to any given
subspace, where ε is the optimum. In Section 6, we discuss some applications of our algorithm.
Our main applications will be:
• (Section 6.1) Approximating a graph using disjoint union of expanders and,
• (Section 6.2) k- EXPANSION when φk ≤ O(λk+1 ).
3
Finally, in Section 7, we present a simple reduction from k- EXPANSION to our problem: This means
any algorithm for k- EXPANSION has to solve our subspace rounding problem as well.
2
Notation and Background
def
Let [m] = {1, 2, . . . , m}. We will associate V = [n] with the set of nodes. For any vector q ∈ RΥ ,
def
1
kqk2 q
def
1
n
and q = kqk
2 q. Given a subset S ⊆ V , we use eS ∈ R to denote the indicator vector for
2
(
1 if i ∈ S,
S, eS (i) =
.
0 else.
Matrices. We use Rr×c to denote the set of r-by-c real matrices. Likewise, we use Sc and Sc+ ⊆ Sc
to denote the set of c-by-c symmetric and positive semidefinite matrices, respectively. Finally let
Sk (Rn ) be the set of all n-by-k orthonormal matrices (Stiefel manifold) for k ≤ n:
n
o
def
Sk (Rn ) = A ∈ Rn×k AT A = Ik .
q =
Given an r-by-c matrix A ∈ Rr×c , we use σi (A), i ∈ {1, 2, . . . , min(r, c)} to refer to the ith largest
singular value of A. We define σmin (A) as the minimum singular value of A, and kAk2 as the
√2-norm
of
A, which is kAk2 = σ1 (A). Likewise kAkF denotes the Frobenius norm of A, kAkF = AT A =
qP
2
j σj (A). Given matrix R ⊆ [r], C ⊆ [c], we will use AR,C to refer to the minor corresponding to
rows R and columns C.
Finally, we will use AΠ , A⊥ ∈ Sr+ to denote the r-by-r projection matrices onto the column space
and co-kernel of A, respectively. Observe that for any A ∈ Sk (Rn ), AΠ = AAT and A⊥ = In − AAT .
One way of measuring the closeness of two subspaces is to look at how much (in degrees) we
need to rotate a vector in one subspace to the closest vector in the other subspace. It is well known
that this quantity is related to the spectral norm. For completeness, we provide a formal version of
this statement along with its proof:
Proposition 2.1 ([SS90]). Given two linear k-dimensional subspaces of Rn with orthonormal basis A, B ∈
Sk (Rn ) respectively; the cosine of the largest angle between these two subspaces is given by the following:
def
cos(∠AB) =
|hx, yi|
.
x∈span(A) y∈span(B) kxk2 kyk2
min
max
We have sin(∠AB) = kA⊥ Bk2 = kB ⊥ Ak2 .
Proof. From the definition of ]AB, it is easy to see how it measures the maximum degrees necessary
to rotate a point in A to any point in B and vice versa. We will now prove the second statement.
Any point x in span(A) can be written as Ap for some p ∈ Rk . Moreover A is orthonormal, thus
kxk = kApk = kpk. This allows us to rewrite cos(]AB) as follows:
|hx, yi|
|hAp, yi|
= min max
.
p y∈span(B) kpk2 kyk2
x∈span(A) y∈span(B) kxk2 kyk2
min
max
For any p, best y is given by B Π Ap. Moreover kB Π Apk22 + kB ⊥ Apk22 = kpk22 , thus:
s
kB Π Apk2
kB ⊥ Apk22
= min
= 1 − max
.
p
p
kpk2
kpk22
4
Consequently, sin(∠AB) = maxp
kB ⊥ Apk22
kpk22
= kB ⊥ Ak22 .
Definition 2.2. Let SetV (k) be the family of sets of k non-empty subsets of V . We will use DisjV (k) ⊆
SetV (k) to denote the set of k disjoint subsets of V : Γ ∈ DisjV (k) if and only if Γ ∈ SetV (k) and S ∩ T = ∅
for all S 6= T ∈ Γ.
In order to compare subspaces with k-partitions, we need to identify a canonical representation
of the subspaces associated with k-partitions. The most natural representation is to use each basis
vector as the normalized indicator of one of the clusters.
Notation 2.3 (Basis Matrices of k-Partitions). Given k-subsets Γ = {A1 , . . . , Ak } of V , let Γ ∈ Rn×k
def
be the corresponding normalized incidence matrix Γ = eA1 . . . eAk . We will use ΓΠ ∈ Sn+ and
Γ⊥ ∈ Sn+ to denote the associated projection matrices so that ΓΠ Γ = Γ and Γ⊥ Γ = 0.
Multiplication with either of the projection matrices ΓΠ and Γ⊥ has a natural correspondence
with means and the differences to means:
Proposition 2.4. If Γ ∈ DisjV , then Γ is an orthonormal matrix, Γ ∈ Sk (Rn ), and Γ⊥ is a Laplacian
matrix. For any Y ∈ Rk×n , ith column of:
• Y ΓΠ is the mean of points in the same cluster with i provided i is in any cluster of Γ, and 0 otherwise.
• Y Γ⊥ is the difference between yi and its associated center as defined above.
For example, kY Γ⊥ k2F measures the sum of squared distances of each point to the center of its cluster or
origin if it is not in any cluster.
We will measure the distance between sets in a way similar to cosine distance.
def
Notation 2.5. Given p, q ∈ Rn , we define ∆(p, q) as ∆(p, q) = 1 − hp, qi2 . Note that ∆(p, q) =
2
1
⊗2 − q ⊗2 k2 . For convenience, we will use ∆(S, q) as ∆(e , q). In particular, ∆(A, B) = 1 − |A∩B| .
S
2 kp
|A||B|
Our measure of set similarity is closely related to the Jaccard index.
Proposition 2.6. For any pair of subsets A, B ⊆ V :
|A∆B|
1 |A∆B|
≤ ∆(A, B) ≤
.
4 |A ∪ B|
|A ∪ B|
Proof. Since |A∪B|2 ≥ |A||B|, we immediately see that 1−∆(A, B) ≥ |A∩B|
|A∪B| . For the other direction,
p
suppose ∆(A, B) ≤ ε and |A| ≥ |B|. Then (1 − ε) |A||B| ≤ |A ∩ B| which implies
s
s
|A ∩ B|
|B|
|A ∩ B|
≥ (1 − ε)
≥ (1 − ε)
.
|A|
|A|
|A|
Therefore |A ∩ B| ≥ (1 − ε)2 |A| and |A∆B| = |A| + |B| − 2|A ∩ B| ≤ |B| − (1 − 4ε + 2ε2 )|A| ≤ 4ε|B|.
In particular,
|A∆B|
≥ 4∆(A, B).
|A ∪ B|
5
We will now generalize our set similarity measure to k-partitions.
b ∈ SetV (k); we define ∆(Γ, Γ)
b as:
Notation 2.7. Given Γ, Γ
def
b = min max ∆(S, π(S)).
∆(Γ, Γ)
b S∈Γ
π:Γ↔Γ
b whenever ∆(A, B) ≤ ε (resp. ∆(Γ, Γ)
b ≤ ε).
We say A (resp. Γ) is ε-close to B (resp. Γ)
Observe that our notion of proximity is a very strong bound. For example if Γ is ε-close to Γ∗ ,
then any subset S ∈ Γ∗ of size |S| < 1ε has to be preserved exactly in Γ.
The next theorem says that the similarity measure we use for k-partitions in Notation 2.7 is
tightly related to the spectral norm distance between the corresponding basis.
b 2 ≤ 2∆(Γ, Γ).
b ∈ DisjV (k); ∆(Γ, Γ)
b ≤ kΓ⊥ Γk
b Moreover, after appropriately
Theorem 2.8. Given Γ, Γ
2
2
b
b
b
ordering the columns of Γ, kΓ − Γk2 ≤ 4∆(Γ, Γ).
Proof of Theorem 2.8 is given in Section 8.1.
Proposition 2.9. Given A, B ∈ Sk (Rn ), σmin (AT B) =
p
1 − kA⊥ Bk22 .
Proof. B T A⊥ B = B T B − B T AAT B = Ik − B T AAT B. Since kAT Bk2 ≤ 1, kA⊥ Bk22 = 1 −
σmin (AT B)2 .
Consider two subspaces with basis A and B. If the angle between these two subspaces is small,
then one might intuitively expect that AAT and BB T are very close also. In the next lemma, we
make this intuition formal. We also include its proof for completeness.
Lemma 2.10 ([SS90]). Given A, B ∈ Sk (Rn ); kAAT − BB T k2 = kA⊥ Bk2 .
Proof. We will prove ≤ by upper bounding the spectral norm of (AAT − BB T )2 . Since AAT = AΠ
and BB T = B Π :
(AAT − BB T )2 =AΠ + B Π − AΠ B Π − B Π AΠ
=AΠ B ⊥ + B Π A⊥ .
(2)
If this matrix is zero, then our claim is trivially true. Suppose not. Consider the largest eigenvalue
σ of eq. (2) and a corresponding eigenvector q. We have 0 6= σq = (AΠ B ⊥ + B Π A⊥ )q, which means
either B ⊥ q 6= 0 or A⊥ q 6= 0 (or both). Without loss of generality, we may assume A⊥ q 6= 0:
σq = (AΠ B ⊥ + B Π A⊥ )q =⇒ σA⊥ q = A⊥ B Π A⊥ q.
def
Consequently, q 0 = A⊥ q is an eigenvector of A⊥ B Π A⊥ with eigenvalue σ:
A⊥ B Π A⊥ q 0 = A⊥ B Π A⊥ q = σA⊥ q = σq 0 .
In particular, kAAT − BB T k22 = k(AAT − BB T )2 k2 = σ ≤ kA⊥ B Π A⊥ k2 = kA⊥ BB T A⊥ k2 =
kA⊥ Bk2 . Now we will prove ≥. If we multiply both sides of eq. (2) with A⊥ , we see that
(AAT − BB T )2 A⊥ (AAT − BB T )2 A⊥ = A⊥ B Π A⊥
which implies
k(AAT − BB T )2 k2 ≥ kA⊥ Bk22 .
6
2.1
Graph Partitioning
Given an undirected graph G = (V, C) with nodes V and non-negative edge weights C, we use
AG ∈ SV and LG ∈ SV+ to denote the adjacency and Laplacian matrices of G. Consider the following
k-way graph partitioning problem where the goal is to minimize the maximum ratio of the total
weight of edges cut and the number of nodes inside among all clusters.
Definition 2.11 (k- EXPANSION). Given an undirected graph G = (V, C) with nodes V and non-negative
edge weights C, we define the k-way expansion of G as the following:
def
φk (G) =
min
max
Γ∈DisjV (k) T ∈Γ
C(T, T )
.
|T |
Here C(A, B) denotes the total weight of unordered edges between A and B. For fixed G, we will use
Γ∗ ∈ DisjV (k) to refer to the k-partition which achieves φk (G).
At the first glance, our notion of expansion might seem different than the usual definition given
in eq. (1). However they are indeed the same:
Proposition 2.12. For any G = (V, C) and a k-partition of V , Γ ∈ DisjV (k),
max
T ∈Γ
C(T, T )
C(S, S)
= max
.
S∈Γ min(|S|, |S|)
|T |
In particular, φk (G) = Φk (G).
Proof. For any S, min(|S|, |S|) ≤ |S|, therefore
C(S,S)
|S|
C(S,S)
.
min(|S|,|S|)
≤
Now we will prove the other
C(T,T )
|T | .
For any T 0 ∈ Γ:
X
X
C(T, T ) ≤ φ
|T | = φ|V \ T 0 |.
C(T 0 , T 0 ) ≤
direction. Let φ = maxT ∈Γ
T ∈Γ\T 0
Consequently,
C(T 0 ,T 0 )
|V \T 0 |
=
C(T 0 ,T 0 )
|T 0 |
T ∈Γ\T 0
≤ φ. Recall that
C(T 0 ,T 0 )
|T 0 |
≤ φ, so
C(T 0 ,T 0 )
min(|T 0 |,|T 0 |
≤ φ.
We can capture the objective function of φk using the spectral norm, within a factor of 2:
Lemma 2.13. Given Γ ∈ DisjV (k), 21 kΓT LΓk2 ≤ maxT ∈Γ
def
Proof. Let φ = maxT ∈Γ
C(T,T )
|T | .
C(T,T )
|T |
≤ kΓT LΓk2 .
We need to prove φ ≤ σmax (ΓT LΓ) ≤ 2φ. The lower bound is
trivial, so we only give the proof of upper bound. Note Γ = JD−1/2 where D is a matrix whose
diagonals are |T | for T ∈ Γ and the columns of J are indicator vectors for every T ∈ Γ. Then
ΓT LΓ = D−1/2 J T LJD−1/2 . Define W as the matrix which is equal to J T LJ along its diagonals
and 0 everywhere else. Since J T LJ is a Laplacian matrix J T LJ 2W . Therefore:
ΓT LΓ 2D−1/2 W D−1/2 .
D−1/2 W D−1/2 is diagonal whose entries are
eT
T LeT
|T |
=
C(T,T )
|T |
≤ φ over all T ∈ Γ. Consequently,
σmax (ΓT LΓ) ≤ 2σmax (D−1/2 ZD−1/2 ) ≤ 2φ.
7
Given Lemma 2.13, a simple relaxation for φk (G) (the basic SDP relaxation) is the following:
min kQT LQk2 st QT Q = Ik .
(3)
Note that for any Γ ∈ DisjV (k), Γ is feasible and eq. (3) is indeed a relaxation. Moreover the CourantFischer-Weyl principle implies that the optimum value of eq. (3) is λk with the corresponding
optimal solution being the smallest k-eigenvectors of L. Therefore:
λk ≤ φk (G).
3
(4)
Our Contributions
We re-state our main problem. Given a k-by-n matrix Y : Y T ∈ Sk (Rn ) of the form Y = [y1 , . . . , yn ]
(think of Y as an embedding of n points in Rk ), find a k-partition Γ ∈ DisjV (k) so as to minimize
the total variance under any direction:
XX
hz, yu − cS i2 .
(5)
min
max
Γ
z∈Rk :kzk2 =1
S∈Γ u∈S
Here cu is the mean of points in the same cluster with u if one exists, and cu = 0 otherwise. This is
the problem of clustering with spectral norm [KK10].
Remark 3.1 (Covering all points). For simplicity, we allow some points to be left uncovered by any set in
Γ. However the same guarantees still hold even if we require Γ to cover all points: We arbitrarily assign
uncovered points to clusters while making sure that the relative cluster sizes do not change. This procedure
changes the approximation ratio by a factor of 2.
We can express eq. (5) more succinctly as the following:
min kY Γ⊥ k22
Proposition 2.4
=
Γ
eq. (5)
(6)
There are two closely related problems, whose optimum is within square root of eq. (5) (Lemma 2.10):
• Finding a k-by-k rotation matrix R : RT R = RRT = Ik and a k-partition Γ ∈ DisjV (k) so as to
minimize the following:
min kRY − Γk2 .
R,Γ
(7)
• Approximate the Gram matrix of Y , Y T Y , using block diagonal matrices with each block
being constant. This is equivalent to:
min kY T Y − ΓΓT k2 .
Γ
(8)
Our main contribution is a new spectral clustering algorithm whose pseudo-code is given through
Algorithms 1 to 5. We prove the following guarantee on its outputs.
8
Theorem 3.2 (Restatement of Theorem 5.12). Let Γ∗ ∈ DisjV (k) with kY Γ∗ ⊥ k22 ≤ O(ε). Then
b ← S PECTRAL C LUSTERING(Y ) is a k-partition so that Γ
b ∈ DisjV (k), and it is O(√ε) close to both Γ∗
Γ
and Y :
√
√
b ⊥ k2 ≤ O( ε).
b ≤ O( ε) and kY Γ
∆(Γ∗ , Γ)
2
any cluster size.
Remark 3.3 (Small Clusters). Our main guarantee as stated in Theorem 5.12 works for
√
For example, consider the case of some optimal cluster T ∈ Γ∗ having size |T | ≤ O 1/ ε . For such cluster,
√
any S with ∆(S, T ) ≤ O( ε) has to be exactly equal to T .
√
In other words, our algorithm will recover any T ∈ Γ∗ with |T | ≤ O(1/ ε) exactly.
As an easy consequence, we show how to approximate a graph as a disjoint union of expanders
(provided one exists) in polynomial time.
Corollary 3.4 (Restatement of Corollary 6.4). Given a graph G, if there exists Γ∗ ∈ DisjV (k) such that
Laplacian of G is ε-close (in spectral norm) to the Laplacian corresponding to the disjoint union of normalized
cliques on each T ∈ Γ∗ :
kL − Γ∗ ⊥ k2 ≤ ε,
√
then in polynomial time, we can find Γ ∈ DisjV (k) which is O ε -close to Γ∗ and G:
L − Γ⊥ k2 ≤ O(ε1/4 ).
Next we significantly improve the known bounds for recovering a k-partition when all clusters
have small expansion as in Definition 2.11. Previous spectral clustering algorithms only guarantee
recovering each T ∈ Γ∗ when the (k + 1)st smallest eigenvalue, λk+1 , of the associated Laplacian
matrix for G satisfies
λk+1 > Ω(k · φk ).
Our new algorithm significantly relaxes this requirement to λk+1 > Ω(φk ).
Theorem 3.5 (Restatement of Theorem 6.1). Given a graph G with Laplacian matrix L, let Γ be the
k-partition obtained by running Algorithm 3 on the smallest k eigenvectors of L. Then:
s
!
φk
.
∆(Γ, Γ∗ ) ≤ O
λk+1
Finally, we show that any approximation algorithm for k- EXPANSION implies the same approximation bound for the spectral clustering problem restricted to orthonormal matrices. In other
words, the spectral clustering problem is a prerequisite for approximating k- EXPANSION even on
graphs whose normalized Laplacian matrix has its (k + 1)st eigenvalue λk+1 larger than a constant.
def
Theorem 3.6 (Restatement of Theorem 7.1). Given Y : Y T ∈ Sk (Rn ), let Γ∗ = argminΓ kY Γ⊥ k22
2
with ε = kY Γ⊥
graph X, whose normalized Laplacian
∗ k2 . Then there exists a weighted, undirected, regular√
st
matrix has its (k + 1) smallest eigenvalue at least λk+1 ≥ 1 − O( ε) such that:
• Each T ∈ Γ∗ has small expansion, φX (T ) ≤ O(ε);
• If Γ ∈ DisjV (k) is a k-partition with maxS∈Γ φX (S) ≤ δ, then ∆(Γ, Γ∗ ) ≤ O(δ +
Moreover such X can be constructed in polynomial time.
9
√
ε).
4
Our Algorithm
The pseudo-code of our clustering algorithm and its sub-procedures are listed in Algorithms 1 to 5.
b = S PECTRAL C LUSTERING(Y ) (Algorithm 3), where Y is a
The main procedure is invoked by Γ
b is a
k-by-n matrix Y such as the smallest k-eigenvectors of some Laplacian matrix. The output Γ
k-partition close to Y . We use Γ∗ = (T1 , . . . , Tk ) to denote the closest k-partition to Y . We will refer
to T ’s as true clusters.
4.1
Intuition
First we start with the discussion of some of the main challenges involved in spectral clustering
and the intuition behind the major components of our algorithm.
Finding a Cluster. Since there are k directions and k clusters, we can think of each direction being
associated with one of the clusters. Moreover, for one of the true clusters T , the total correlation of
its center with remaining directions will be very small. By utilizing this intuition, we can easily
find such a subset, say S (Algorithm 4). However this property need not be true for all T ’s: Even
though each T will be at most ε-close to every non-associated direction, the total correlation might
be (k − 1)ε 1, thus overwhelming the correlation it had with its associated direction. In fact,
this is the reason why k-means type procedures will fail to find every cluster when ε > 1/k. To
remedy this, each time we find S, we can try to “peel” it off. A natural approach is to project
the columns of Y onto the orthogonal complement of the center of S. Similar ideas were used
before in the context of learning mixtures of anisotropic Gaussians [BV08] and column based matrix
reconstruction [DR10, GS12]. After projection, we obtain a new (k − 1)-by-n orthonormal matrix,
Z, corresponding to remaining k − 1 clusters.
Boosting. Unfortunately we can not iterate the above approach: No matter how accurate we are
in S, there will be some error: If Y were ε-close to (T1 , . . . , Tk ), then we can at best guarantee that
Z is 2ε-close to (T2 , . . . , Tk ). After k iterations, our error will be 2O(k) ε, which is much worse than
k-means! In our algorithm, we keep the error from accumulating via a boosting step (Algorithm 1).
Unraveling. Even with boosting, there remains one issue: The clusters we found may overlap with
one another. Unlike other distance based clustering problems such as k-means, the assignment
problem (“which cluster does this node belong to?”) is quite non-trivial in spectral clustering even
if we are given all k-centers. There is no simple local procedure which can figure out the assignment
of node u by only looking at yu and the cluster centers. We deal with this issue by reducing the
ownership problem to finding a matching in a bipartite graph (Algorithm 5). Our approach is
very similar to the one used in [BS06] for a special case of Santa Claus problem. Unfortunately,
this operation has a cascading effect: Adding a new cluster might considerably change the previous
clusters and their centers. Dealing with this challenge is what causes our final algorithm to be
rather involved.
Final Algorithm. The final difficulty we face is that, the boosted cluster might cannibalize other
much smaller clusters. We overcome this issue by maintaining both estimates for every true cluster:
a coarse estimate, which is the core we originally found; and the finer estimate obtained after
boosting. Due to the cascading effect of unraveling whenever we add a new cluster, we have to
re-compute the centers and project onto their orthogonal complement at every round.
10
4.2
Overview
Our algorithm proceeds iteratively. At rth iteration, it finds a core Sr for one of the yet-unseen true
clusters, say Tr . The main invariant we need from the core set is the following (Lemma 5.7):
(i) Sr is noticeably close to Tr , say ∆(Sr , Tr ) ≤
1
100 .
(ii) All the remaining k − r true clusters, {Tr+1 , . . . , Tk }, have small overlap with Sr in the sense
1
that |Tj ∩ Sr | ≤ 100
|Tj | over all j > r.
cr . For boosting to work however
After having found Sr , our algorithm needs to boost Sr to S
(Theorem 5.8), we need the invariant (ii) to be true for all j 6= i. We do this by (U1 , . . . , Ur ) ←
U NRAVEL(S1 , . . . , Sr ). Since Ui ’s are close to Si ’s, each Tj with j < r will still be mostly overlapping with Sj (Theorem 5.6); hence Tj ’s with j < r can not overlap with Ur as Uj ’s are disjoint.
cr . The only invariant we
Consequently we can use Ur instead of Sr for boosting so as to obtain S
cr is that it is much closer to Tr than Sr :
require from S
√
cr , Tr ) ≤ O( ε).
∆(S
By using the centers of boosted sets instead of core sets, we can make sure that the error does not
b 0 )⊥ Y (Lemma 5.5).
accumulate after the projection step Z ← (Y Γ
c1 , . . . , S
ck ) one last time and output
After k iterations, we apply U NRAVEL to all boosted sets (S
the result.
Remark 4.1. In Algorithms 1 and 4, the last step involves computing the top singular vector. In both cases,
we have:
• A good initial guess (the indicator vector),
• Large separation between σ1 and σ2 .
Thus, we can simply use power method for O(log 1/ε) many iterations to compute a sufficiently accurate
approximation of the top right singular vector, from which we can obtain an approximation of the top left
singular vector easily. It was previously shown in [GKB13] that power method is sufficient in the context of
spectral clustering.
5
Analysis of the Algorithm
In this section, we prove the correctness of our algorithm. Our main result is the following. Its
proof is given at the end of Section 5.5.
Theorem 5.1 (Restatement of Theorem 5.12). Let Γ∗ ∈ DisjV (k) with kY Γ∗ ⊥ k22 ≤ O(ε). Then
b ← S PECTRAL C LUSTERING(Y ) is a k-partition so that Γ
b ∈ DisjV (k), and it is O(√ε) close to both Γ∗
Γ
and Y :
√
√
b ⊥ k2 ≤ O( ε).
b ≤ O( ε) and kY Γ
∆(Γ∗ , Γ)
2
In order to keep the analysis simple, we make no effort toward optimizing the constants.
11
Algorithm 1 Sb = B OOST(Y, S).
1. p ← top left singular vector of YS .
2. Return R OUND(Y T p).
Algorithm 2 S = R OUND(q).
1. F ← {u | squ ≥ sqv } | v ∈ V, s ∈ {±1}}.
2. Return argmaxS∈F |hq, eS i|.
b = S PECTRAL C LUSTERING(Y ).
Algorithm 3 Γ
Algorithm 4 S = F IND C LUSTER(Y ).
1. For r ← 1 to rank(Y ) do:
1. Choose δ 0 ∈ [0, 1] as the minimum for which the
following returns some S.
2. For each c ∈ V :
b 0 ← U NRAVEL(S
c1 , S
c2 , . . . , S
[
Γ
r−1 ).
0
⊥
b
Z ← (Y Γ ) Y .
Sr ← F IND C LUSTER(Z).
(U1 , ..., Ur ) ← U NRAVEL(S1 , ..., Sr ).
cr ← B OOST(Y, Ur ).
S
b
2. Return U NRAVEL(Γ).
(a)
(b)
(c)
(d)
(e)
kY
−Yc k
kY
i. S ← {π(1), . . . , π(m)}.
ii. q ← top right singular vector of YS .
iii. Return R OUND(q).
b = U NRAVELδ (Γ) (see Figure 1 for a sample graph construction.)
Algorithm 5 Γ
1. Choose δ 0 ∈ [0, 1] as the minimum for which the following returns a matching.
2. Construct a bipartite graph H = (L, R, E), where left side L is V :
(a) For all S ∈ Γ, there is a block BS of d(1 − δ)|S|e identical nodes in R.
(b) For all S ∈ Γ and u ∈ S, there is an edge between u and all nodes in BS .
3. Find a matching that covers R.
b ... | S ∈ Γ).
4. For all S ∈ Γ, Sb ← nodes matched to the block BS . Return (S,
Input Γ = {S1 , S2 , S3 , S4 }.
Bipartite Graph.
Figure 1 The graph constructed by U NRAVEL on input S1 = {a, b, c, e}, S2 = {d, e, f, g},
S3 = {g, h, i, j, k}, S4 = {k} for δ = 14 .
5.1
−Yc k
π(i)
π(i+1)
(a) π be an ordering st kY
k ≤ kYπ(i+1) k .
π(i)
n P
o
2
0
(b) m ← min j
i≤j kYπ(i) k ≥ 1 − δ .
P
(c) If j≤m kYπ(j) − Yc k2 ≤ δ 0 , then:
Preliminaries
In the following proposition, we show that for any pair of symmetric matrices that are close to each
other in spectral norm, if there is a gap between the largest and second largest eigenvalues; then
the largest eigenvectors of both matrices will be very close to each other. This can also be obtained
using Wedin’s theorem [SS90], but we chose to give a simple and self-contained proof.
12
Proposition 5.2. Given A, B ∈ Sn with maximum eigenvectors p, q ∈ Rn :
hp, qi2 ≥ 1 −
2kA − Bk2
.
σ1 (A) − σ2 (A)
def
def
Proof. Suppose kpk = kqk = 1. Let δ = kA − Bk2 and θ = hp, qi2 . We have
σ1 (A) =pT Ap ≤ δ + σ1 (B)hq, pi2 + σ2 (B)kp⊥ qk2
≤δ + (σ1 (A) + δ)θ + (σ2 (A) + δ)(1 − θ)
=δ + θ(σ1 − σ2 ) + (σ2 + 2δ).
Hence θ ≥
σ1 −σ2 −2δ
σ1 −σ2 .
The main tool we use to identify the clusters will be the eigenvalues of principal minors of the
Gram matrix, Y T Y . Basically, eigenvalues measure how much true clusters, T , overlap with given
principal minor. We make this connection formal in the following claim:
Claim 5.3. Given Y ∈ Rm×n , Γ ∈ DisjV (r) and subset S ⊆ V , let ρ = |S ∩ T |/|T | T ∈ Γ . Then for
any i:
|σi2 (YS ) − (ρ)↓i | ≤ kY T Y − ΓΠ k2 .
Here (ρ)↓i is the ith largest element of ρ.
Proof. We have kY T Y − Γ∗ Π k2 ≥ kYST YS − (Γ∗ Π )S,S k2 . Observe that the eigenvectors of (Γ∗ Π )S,S
0
are eT 0 ∩S with corresponding eigenvalue |T|T∩S|
over all T 0 ∈ Γ∗ . Thus ρ’s are the eigenvalues of
0|
(Γ∗ Π )S,S .
Consider a principal minor corresponding to some S whose largest eigenvalue is large, and
second largest eigenvalue is small. The previous claim implies that there is a unique optimal cluster
T which is almost contained by S. However this is still not sufficient: S might be much larger than
T . In the next lemma, we show that, one can take the top right singular vector of YS and round
(threshold) it to obtain another subset Sb ⊆ S which is now very close to T .
Lemma 5.4 (Initial Guess). Given Y ∈ Rm×n , Γ∗ ∈ DisjV (r) with kY T Y − Γ∗ Π k2 ≤ δ, and subset
def
def
S ⊆ V , let q ∈ RS be the top right singular vector of YS . If we define σ1 = σ12 (YS ) and σ2 = σ22 (YS ), then
the subset Sb ⊆ S obtained by:
Sb ← R OUND(q)
satisfies the following. For T being argmaxT ∈Γ∗
|S∩T |
|T | :
b
p
|T ∩ S|
4δ
q
≥ σ1 − δ 1 −
σ1 − σ2
b
|T ||S|
and
b ≤ (σ2 + δ)|T 0 |.
∀T 0 6= T ∈ Γ∗ : |T 0 ∩ S|
13
def
def
Proof. We will use σ1 = σ12 (YS ), σ2 = σ22 (YS ) and ∆ = σ1 − σ2 . Since q is the top right singular
vector of YS , kYS qk2 = σ1 and kqk = 1.
|S∩T |
|S∩T 0 |
0
By Claim 5.3, |T|T∩S|
| − σ1 ≤ δ, which implies |T | ≥ σ1 − δ. For any T 6= T , |T 0 | ≤ σ2 + δ.
Via Proposition 5.2, we see that:
hq, eT ∩S i2 ≥ 1 −
2δ
2δ
2δ
≥1−
=⇒ hq, eSbi2 ≥ 1 − .
σ1 − σ2
∆
∆
Provided that δ ≤ 14 ∆, both hq, eT ∩S i and hq, eSbi have the same sign. Therefore:
heT ∩S , eSbi ≥ hq, eT ∩S ihq, eSbi − kq ⊥ eT ∩S kkq ⊥ eSbk ≥ 1 −
4δ
.
∆
b ≤ (α + δ)|T 0 | and
Consequently, using the fact Sb ⊆ S, we see that any T 0 6= T has |T 0 ∩ S|
b
T ∩ S ∩ Sb = T ∩ S:
b
b
|T ∩ S|
1
|T ∩ S|
4δ
q
≤q
≤√
.
1−
∆
σ
−
δ
1
b
b
|T ∩ S||S|
|T ||S|
After we found new clusters, we iterate by projecting Y onto the orthogonal complement of
their center. In the following lemma, we prove that, as long as the clusters were close to optimal
ones, the projection preserves remaining clusters.
Lemma 5.5. Given Y : Y T ∈ Sk (Rn ) and Γ ∈ DisjV (r) with the cluster centers in Γ being linearly
independent, suppose there exists Γ∗ ∈ DisjV (k) of the form Γ∗ = Γ0∗ ]Γ00∗ : Γ0∗ ∈ DisjV (r), Γ00∗ ∈ DisjV (k−r)
such that:
• kΓ − Γ∗ 0 k22 ≤ α.
• kY Γ∗ ⊥ k22 ≤ ε.
def
Then Γ00∗ is a good spectral clustering for Z = (Y Γ)⊥ Y , in the sense that kZ(Γ∗ 00 )⊥ k22 ≤ ε + α. In addition,
σ1 (Z) = . . . = σk−r (Z) = 1, σk−r+1 (Z) = 0.
Proof. Note that Z has all singular values either 0 or 1:
ZZ T = (Y Γ)⊥ Y Y T (Y Γ)⊥ = (Y Γ)⊥ .
Moreover (Y Γ)⊥ has rank rank(Y ) − rank(Y Γ) = k − r. Since spectral clustering is invariant under
change of basis, we can assume Z ∈ R(k−r)×n so that all singular values of Z are 1. This means Z
is orthonormal. It is obtained from Y by a linear transformation, therefore Γ∗ is a good spectral
clustering for Z also:
ε ≥kY Γ∗ ⊥ Y T k2 ≥ kZΓ∗ ⊥ Z T k2
=kZZ T − (ZΓ∗ )(ZΓ∗ )T k2
=kIk−r − (ZΓ∗ )(ZΓ∗ )T k2 .
14
In other words,
(1 − ε)Ik−r ZΓ∗ Π Z T = Z (Γ∗ 0 )Π + (Γ∗ 00 )Π Z T
=⇒ εIk−r Z(Γ∗ 00 )⊥ Z T − Z T Γ∗ Π Z T .
Now we will upper bound kZ(Γ∗ 0 )Π Z T k2 = kZΓ∗ 0 k2 = k(Y Γ)⊥ (Y Γ∗ 0 )k2 by a simple CauchySchwarz:
k(Y Γ)⊥ (Y Γ∗ 0 )k22 =k(Y Γ)⊥ (Y Γ∗ 0 ) − (Y Γ)⊥ (Y Γ)k22
=k(Y Γ)⊥ Y (Γ∗ 0 − Γ)k22
≤k(Y Γ)⊥ Y k22 kΓ∗ 0 − Γk22
≤kY k22 kΓ∗ 0 − Γk22 ≤ kΓ∗ 0 − Γk22 ≤ α.
As a consequence, kZ(Γ∗ 00 )⊥ k22 ≤ ε + kZΓ∗ 0 k22 ≤ ε + α.
Now we will start with the proof of correctness for our unraveling procedure.
5.2
Correctness of U NRAVEL (Algorithm 5)
We will now prove that if the input of U NRAVEL(Γ) (Algorithm 5) is a list of (possibly overlapping)
sets which are close to some k-partition (ground truth), then the output will be a list of k disjoint
sets which are also close to the ground truth. Our algorithm is based on formulating this as a simple
maximum bipartite matching problem.
Lemma 5.6. Given Γ ∈ SetV (k), if there exists Γ∗ ∈ DisjV (k) which is δ-close to Γ, then U NRAVEL(Γ) (Alb ∈ DisjV (k) such that:
gorithm 5) will output Γ
b with U ⊆ S and |U | ≥ (1 − δ)|S|.
• For each S ∈ Γ, there exists U ∈ Γ
b is 4δ-close to Γ∗ .
• Γ
Proof. It is easy to see that if all blocks are matched, then the resulting assignment is a collection
of k-disjoint subsets so it has the first property. For the second property, consider any S ∈ Γ with
b and T ∈ Γ∗ .
corresponding subsets U ∈ Γ
|U |
|S \ U |
|S ∩ U |2
=1−
=
≤ δ.
|S||U |
|S|
|S|
p
p
p
√
b is 4δ-close to Γ∗ .
Hence ∆(U, T ) ≤ ∆(U, S) + ∆(S, T ) ≤ 2 δ. This implies Γ
Now we will prove that if δ-close Γ∗ ∈ DisjV (k) exists, then there is always a matching of all
blocks. Suppose π : Γ ↔ Γ∗ is a matching which minimizes maxS ∆(S, π(S)). Then:
∆(U, S) = 1 −
|S ∩ π(S)|2 ≥ (1 − δ)|S||π(S)| =⇒ |S ∩ π(S)| ≥ (1 − δ)|S|.
By Hall’s theorem, we have to show that for any set of right nodes, B ⊆ R, B’s neighbors on the
left, N (B), are more than B: |N (B)| ≥ |B|. Observe that if B contains some nodes of block BS ,
then adding the whole block BS to B does not increase |N (B)|, because all nodes in BS have the
15
same set of neighbors. Hence the only subsets we need to consider are of the form B = ∪S∈A BS
over all A ⊆ Γ:
|N (B)| = ∪S∈A S ≥ ∪S∈A (S ∩ π(S))
X
X
=
|S ∩ π(S)| ≥ (1 − δ)
|S| = |B|.
S∈A
5.3
S
Correctness of F IND C LUSTER (Algorithm 4)
Here we prove that, given an orthonormal basis Y , if it is close to a k-partition up to rotation,
then F IND C LUSTER(Y ) (Algorithm 4) will output a set which is close to one of the clusters in this
k-partition.
Lemma 5.7. Given Y ∈ Rm×n , whose singular values are 0 or 1, if there exists Γ∗ ∈ DisjV (k) with
kY T Y − Γ∗ Π k22 ≤ δ for some small enough constant δ, then F IND C LUSTER(Y ) (Algorithm 4) will output
S 0 ⊆ V such that:
√
• There exists T ∈ Γ∗ with ∆(S 0 , T ) ≤ O( δ),
√
• Any T 0 6= T : T 0 ∈ Γ∗ has |T 0 ∩ S 0 | ≤ O( δ)|T 0 |.
Proof. kY T Y − Γ∗ Π k22 ≤ δ implies σk (Y ) = 1 and σk+1 (Y ) = 0. Therefore kY Γ∗ ⊥ k2F ≤ δk. In other
words,
X
X
(1 − kY eT k22 ) ≤ δk.
(kYT k2F − kY eT k22 ) =
T ∈Γ∗
T ∈Γ∗
So:
X h
i
max(1, kYT k2F ) − kY eT k22 ≤ 2δk.
T ∈Γ∗
As a consequence, there exists T ∈ Γ∗ and some c ∈ T such that:
kYT k2F ≥ kY eT k22 ≥ 1 − 2δ.
4δ ≥
1
|T |
X
kYu − Yv k22 ≥
u∈T,v∈T
X
kYu − Yc k2 .
u∈T
kYu − Yc k2
4δ
≥ Eu∼kYu k2
.
1 − 2δ
kYu k2
def kYu −Yc k2
kYu k2
Let’s define ρu =
and sort the nodes in ascending order so that ρ1 ≤ ρ2 ≤ . . . ≤ ρn :
Eu∼kYu k2 ρu ≤
4δ
≤ 9δ
1 − 2δ
provided δ ≤ 14 .
√
√
By a simple Markov inequality, sum of all kYuP
k2 over u ∈ T with ρu √
≤ 3 δ is atP
least 1 − 3 δ.
Consequently, the smallest integer m for which 1≤u≤m kYu k2 ≥ 1 − 3 δ satisfies 1≤u≤m kYu −
√
Yc k2 ≤ 3 δ.
16
√
From now on, we assume S is a subset and δ 0 : δ 0 ≤ 3 δ with:
X
X
kYS k2F =
σj2 (YS ) ≥ 1 − δ 0 and δ 0 ≥
kYu − Yc k2 .
j
u∈S
Recall that variance is lower bounded by sum of the squares of all but largest singular values:
X
X
σj2 (YS ).
kYu − Yc k2 ≥ kYS k2F − σ12 (YS ) =
j≥2
u∈S
Hence σ12 (YS ) ≥ 1 − 2δ 0 and σ22 (YS ) ≤ δ 0 . Provided that δ ≤
(σ1 − δ) 1 −
1
100 ,
which implies δ 0 ≤
3
10 :
√
4δ
≥ 1 − 8 δ.
σ1 − σ2
For such S, Lemma 5.4 tells us that the new subset Sb ⊆ S obtained by rounding the top right
def
singular vector of YS satisfies, for T = argmaxT ∈Γ∗
√
b T ) ≤ 1 − (σ1 − δ) 1 − 4δ
• ∆(S,
σ1 −σ2 ≤ 8 δ.
|S∩T |
|T | :
√
b ≤ 4 δ|T 0 |.
• For any T 0 6= T ∈ Γ∗ , |T 0 ∩ S|
5.4
Correctness of B OOST (Algorithm 1)
As we mentioned earlier, if we keep finding clusters and removing them iteratively, the error will
quickly accumulate and degrade the quality of remaining clusters. To prevent this, we apply a
boosting procedure as described in Sb ← B OOST(Y, S) (Algorithm 1) every time we find a new
cluster. The main idea is that, if S is close to some cluster in ground truth, say T , and far from
others; then the top left singular vector, say p, of the vectors associated with S will be close to the
ones S ∩ T . Unfortunately, we can not use simple perturbation bounds such as Wedin’s theorem.
We have to make full use of eq. (5) instead: Under projection by p, the vectors of T tend to stay
together; therefore vectors in T \ S will be very close to the vectors in S ∩ T . Hence indeed p will
be close most of T \ S in addition to S ∩ T .
Theorem 5.8 (Boosting). Given Y ∈ Sk (Rn ) and Γ∗ ∈ DisjV (k) with kY Γ∗ ⊥ k22 ≤ ε, consider any
subset S ⊆ V . Suppose there exists T ∈ Γ∗ with |S ∩ T | ≥ (1 − α)|T | such that for any T 0 6= T ∈ Γ∗ ,
|S ∩ T 0 | ≤ α|T 0 | for some α ≤ 41 . Then for Sb ← B OOST(Y, S) (Algorithm 1), Sb satisfies:
√
b T ) ≤ c0 ε
∆(S,
for some constant c0 ≤ 50.
Remark 5.9. Note that Theorem 5.8 allows us to convert a subset with non-negligible overlap into a
subset which is very close. Unfortunately, the new subset we obtain is no longer guaranteed to have small
intersection with other T 0 6= T .
17
def
of Theorem 5.8. By Lemma 2.10, for δ =
√
ε:
kY T Y − Γ∗ Π k2 ≤ δ.
def
We will use σ1 = σ12 (YS ) and σ2 = σ22 (YS ). By Claim 5.3, σ1 − |S ∩ T |/|T | ≤ δ and σ2 ≤ β + δ.
def
Since p is the left top singular vector of YS , kYST pk2 = σ1 . We define q = Y T p so that kqS k2 = σ1 :
kqS k2
δ
≥1 −
|S ∩ T |/|T |
|S ∩ T |/|T |
δ
4δ
≥1 −
≥1− .
1−α
3
(9)
Since σ1 ((Γ∗ Π )S,S ) − σ2 ((Γ∗ Π )S,S ) ≥ 1 − 2α, Proposition 5.2 implies:
heT ∩S , qi2 ≥ kqS k2 1 −
2δ
≥ kqS k2 1 − 4δ .
1 − 2α
def
Let’s use µA to denote the mean of p on subset A, µA = heA , qi. We will assume, without loss of
√
generality, µS∩T ≥ 0. Hence µS∩T ≥ √kqS k
1 − 4δ. On the other hand,
|S∩T |
ε ≥pT Y Γ∗ ⊥ Y T p = q T Γ∗ ⊥ q
X
1
≥
(qi − qj )2 ≥ |S ∩ T |(µS∩T − µT )2 .
|T |
{i,j}∈(T2 )
r
ε
µT ≥µS∩T −
.
|S ∩ T |
s
p
ε|T |
heT , qi ≥ |T |µS∩T −
|S ∩ T |
s
√
√
|T |
≥
kqS k 1 − 4δ − ε .
|S ∩ T |
Using eq. (9), we can lower bound this quantity as:
r
r ε
4δ
≥
1−
1 − 4δ −
3
1−α
r
4ε
26
≥1 − 4δ −
≥ 1 − δ.
3
5
Since kqk = kY T pk ≤ 1, we have ∆(p, T ) ≤ 1 − heT , qi2 ≤ 11δ. Using Proposition 5.10, we see that
Sb ← R OUND(q) satisfies:
√
b T ) ≤ 4∆(q, T ) ≤ 44δ = 44 ε.
∆(S,
18
5.5
Correctness of R OUND (Algorithm 2)
Possibly the simplest case for our problem is when Y is 1-dimensional, i.e. it is a vector. As we
argued in the introduction, our algorithm boils down to simple thresholding in this case.
Proposition 5.10. Given q 6= 0 ∈ Rn , for any T 6= ∅, S ← R OUND(q) (Algorithm 2) satisfies ∆(S, T ) ≤
4∆(q, T ).
def
Proof. Let ε = ∆(q, T ). Without loss of generality, we may assume q1 ≥ . . . ≥ qm > 0 ≥ . . . ≥ qn ,
kqk = 1 and hq, eT i ≥ 0. We have:
P
√
hq, eT i
j≤|T | qj
1−ε≤ p
≤ p
|T |
|T |
P
P
j≤|S 0 | qj
j≤|S| qj
p
p
≤ max
≤
≤ |hq, eS i|.
S0
|S 0 |
|S|
p
p
p
√
So ∆(S, q) ≤ ε and ∆(S, T ) ≤ ∆(S, q) + ∆(q, T ) ≤ 2 ε.
5.6
Correctness of S PECTRAL C LUSTERING (Algorithm 3)
Finally, we put everything together and prove the correctness of S PECTRAL C LUSTERING(Y ) (Algorithm 3). In the following lemma, we will show that the algorithm will iteratively find sets
S1 , S2 , . . . (think of them as coarse approximations of Ti ’s) and Sb1 , Sb2 , . . . (think of them as fine
approximations of Ti ’s) such that at each iteration, each Si and Sbi will correspond to a unique
Ti . Moreover, each Si will have very small overlap with remaining Tj ’s coming after themselves.
Even though Si ’s might still have large overlap with previous Tj ’s for j < i, we can easily use
U NRAVEL(Γ) (Algorithm 5) to rectify this issue.
Lemma 5.11. Let Γ∗ ∈ DisjV (k) with kY Γ∗ ⊥ k22 ≤ ε for some ε ≤ ε0 , where ε0 ∈ (0, 1) is a conb = (S
c1 , . . . , S
cr ) as found by
stant. For any r ∈ [k], consider the sequences Γ = (S1 , . . . , Sr ) and Γ
S PECTRAL C LUSTERING(Y ) at the start of rth iteration. Then there exists an ordering of Γ∗ :
Γ∗ = (T1 , T2 , . . . , Tr , Tr+1 , . . . , Tk ),
|
{z
} |
{z
}
def 0
= Γ∗
with the following properties for some α ≤
1
16
def 00
= Γ∗
and β ≤ 100:
(a) For every i ≤ r:
• ∆(Si , Ti ) ≤ α.
• For all j > i, |Tj ∩ Si | ≤ α|Tj |.
√
(b) For every i ≤ r, ∆(Sbi , Ti ) ≤ β ε.
Proof. By induction on r. For r = 0, (a) and (b) are trivially true.
Given r, suppose (a) and (b) are true with (T1 , . . . , Tk ), Γ0∗ and Γ00∗ being as described.
19
b = (S
c1 , . . . , S
cr ). (b) means
At the beginning of (r + 1)st iteration, we have Γ = (S1 , . . . , Sr ) and Γ
√
0
b
Γ is β ε-close to Γ∗ .
b 0 ← U NRAVEL(Γ),
b by Lemma 5.6, Γ
b 0 is 4β √ε-close to Γ∗ and Γ
b 0 ∈ DisjV (r). Using
After Γ
√
b 0 − Γ∗ 0 k2 ≤ 8β ε. Now we can invoke Lemma 5.5, which implies:
Theorem 2.8, we see that kΓ
2
√
√
kZ(Γ∗ 00 )⊥ k22 ≤ ε + 8β ε ≤ 9β ε.
√
Provided 9β ε ≤ δ0 for some small enough constant δ0 , we can use Lemma 5.7 to see that the
subset Sr+1 ← F IND C LUSTER(Z) satisfies
∆(Sr+1 , T ) ≤ α,
and
∀T 0 ∈ Γ00∗ : T 0 6= T =⇒ |T 0 ∩ Sr+1 | ≤ α|T 0 |.
We reorder (Tr+1 , . . . , Tk ) so that Tr+1 = T . Then (a) holds true when i = r + 1. Since T1 , . . . , Tr
remain the same, (a) is true for all i ≤ r + 1.
Consider (U1 , . . . , Ur+1 ) ← U NRAVEL(Γ). By Lemma 5.6, we know that ∆(Ui , Ti ) ≤ 4α and
Ui ⊆ Si , |Ui | ≥ (1 − α)|Si | for each i ∈ [r + 1]. In particular, for all i ≤ r + 1:
|Ui ∩ Ti | ≥ (1 − 4α)|Ti |.
Ur+1 being a subset of Sr+1 means |Ur+1 ∩ Tj | ≤ |Sr+1 ∩ Tj | ≤ α|Tj | whenever j > r + 1. Now we
will prove the case of j ≤ r. Using the fact that U ’s are disjoint, for any j ≤ r:
|Ur+1 ∩ Tj | ≤ |Tj | − |Tj ∩ Uj | ≤ |Tj | − (1 − 4α)|Tj | = 4α|Tj |.
Consequently, for any j 6= r + 1:
|Ur+1 ∩ Tj | ≤ 4α|Tj |.
[
After executing S
r+1 ← B OOST (Y, Ur+1 ), noting α ≤
1
16 ,
we see via Theorem 5.8:
√
√
[
∆(S
r+1 , Tr+1 ) ≤ c0 ε = β ε.
Combined with the fact that Sbi and Ti remain the same for i ≤ r, (b) also remains true for all
i ≤ r + 1.
By induction, we now see that both (a) and (b) are true for all r ≤ k.
b ← S PECTRAL C LUSTERING(Y ) is a
Theorem 5.12. Let Γ∗ ∈ DisjV (k) with kY Γ∗ ⊥ k22 ≤ O(ε). Then Γ
√
b ∈ DisjV (k), and it is O( ε) close to both Γ∗ and Y :
k-partition so that Γ
√
√
b ⊥ k2 ≤ O( ε).
b ≤ O( ε) and kY Γ
∆(Γ∗ , Γ)
2
b
b = (S
c1 , . . . , S
ck ) is β √ε-close to Γ∗ . Lemma 5.6 implies that U NRAVEL(Γ)
Proof. By Lemma 5.11, Γ
√
outputs a disjoint collection of k-subsets which is 4β ε-close to Γ∗ . For the second bound:
1 b⊥ 2
b ⊥ k2 + kY Γ∗ Π Γ
b ⊥ k2
kY Γ k2 ≤kY Γ∗ ⊥ Γ
2
2
2
b ⊥ k2 + kY Γ∗ k2 · kΓ∗ T Γ
b ⊥ k2
≤kY Γ∗ ⊥ k22 · kΓ
2
2
2
√
T b⊥ 2
b
≤ε + kΓ∗ Γ k ≤ ε + 2∆(Γ∗ , Γ) ≤ O( ε).
2
In the second to last inequality, we used Theorem 2.8.
20
6
Applications
In this section, we will show some applications of our spectral clustering algorithm.
6.1 k- EXPANSION
Our first application is approximating non-expanding k-partitions in graphs. One may also interpret
this as applying our subspace rounding algorithm on the basic SDP relaxation for k- EXPANSION
problem.
Theorem 6.1. Given a graph G with Laplacian matrix L, let Γ be the k-partition obtained by running Algorithm 3 on the smallest k eigenvectors of L. Then:
s
!
φk
∆(Γ, Γ∗ ) ≤ O
.
λk+1
Remark 6.2 (Faster Algorithm). By slightly modifying our algorithm to take advantage of the underlying
graph structure, one can obtain a faster randomized algorithm having the same guarantees with Theorem 6.1
with expected running time O(k 2 (n + m)).
Proof. From Lemma 2.13, we know that φk ≤ σmax (Γ∗ T LΓ∗ ) ≤ 2φk . Now consider the matrix
Z = Y T , whose columns are the smallest k eigenvectors of L. We have L λk+1 Z ⊥ which means:
λk+1 · Γ∗ T Z ⊥ Γ∗ Γ∗ T LΓ∗ 2φIk .
This implies
O(ε) ≥ kΓ∗ T Z ⊥ Γ∗ k2 = kIk − Γ∗ T ZZ T Γ∗ k2 .
p
Thus σk (Z T Γ∗ ) = σmin (Z T Γ∗ ) ≥ 1 − O(ε) and:
kY Γ∗ ⊥ k22 =kZ T Γ∗ ⊥ Zk2 = kIk − Z T Γ∗ Γ∗ T Zk2
=1 − σmin (Γ∗ T Z) ≤ O(ε).
The claim follows from Theorem 5.12.
6.2
Matrix and Graph Approximations
Our next application is for approximating a matrix in terms of k-block diagonal matrices corresponding to the adjacency matrices of normalized cliques, under spectral norm.
def
Theorem 6.3. Given a matrix X ∈ Sn , let ε = minΓ∗ ∈DisjV (k) kX − Γ∗ Π k2 . In polynomial time, we can
√
find Γ ∈ DisjV (k) such that ∆(Γ∗ , Γ) ≤ O( ε) and:
kX − ΓΠ k2 ≤ O ε1/4 .
Proof. Let Y be the matrix whose
Consider Γ ← S PECTRAL C LUSTERING(Y ):
rows
are
the
top
kY T Y − Xk2 ≤ kX − Γ∗ Γ∗ T k2 ≤ ε
21
k
eigenvectors
of
X.
which means
kY T Y − Γ∗ Γ∗ T k2 ≤ 2ε =⇒ kY Γ∗ ⊥ k22 ≤ 2ε.
√
√
By Theorem 5.12, ∆(Γ∗ , Γ) ≤ O( ε) and kY Γ⊥ k22 ≤ O( ε):
kΓΓT − Xk2 ≤ ε + kΓ∗ Γ∗ T − ΓΓT k2 ≤ O(ε1/4 ).
Our final application is for approximating a graph Laplacian via another Laplacian corresponding to the graph formed as a disjoint union of k normalized cliques (expanders), again
under spectral norm. Since we are working with Laplacian matrices, this means the new graph
approximates cuts of the original graph also.
Corollary 6.4. Given a graph G, if there exists Γ∗ ∈ DisjV (k) such that Laplacian of G is ε-close (in
spectral norm) to the Laplacian corresponding to the disjoint union of normalized cliques on each T ∈ Γ∗ :
kL − Γ∗ ⊥ k2 ≤ ε,
√
then we can find Γ ∈ DisjV (k) which is O ε -close to Γ∗ and G in polynomial time:
L − Γ⊥ k2 ≤ O(ε1/4 ).
Proof. Since kL − Γ∗ ⊥ k2 = k(I − L) − Γ∗ Π k2 , we can apply Theorem 6.3 on the matrix I − L. The
rest follows easily.
7 k- EXPANSION Implies Spectral Clustering
In this section, we will show that approximation algorithms for various graph partitioning problems
imply similar approximation guarantees for our clustering problem.
def
Theorem 7.1. Given Y : Y T ∈ Sk (Rn ), let Γ∗ = argminΓ kY Γ⊥ k22 with ε = kY Γ∗ ⊥ k22 . Then there
exists a weighted, undirected, regular graph X, whose normalized Laplacian matrix has its (k + 1)st smallest
√
eigenvalue λk+1 is at least λk+1 ≥ 1 − O( ε) such that:
• Each T ∈ Γ∗ has small expansion, φX (T ) ≤ O(ε),
• If Γ ∈ DisjV (k) is a k-partition with maxS∈Γ φX (S) ≤ δ, then ∆(Γ, Γ∗ ) ≤ O(δ +
√
ε).
Moreover such X can be constructed in polynomial time.
Proof. Consider the following SDP. Here we chose ε ∈ [0, 1] to be the minimum value where this
SDP remains feasible:
√
(i) X Y Π + εY ⊥ .
(ii) Y XY T (1 − ε)Ik ,
(iii) X is doubly stochastic, diagonally dominant, PSD and has trace k.
22
It is easy to see that Γ∗ Π is a feasible solution (Lemma 2.10). Moreover any feasible solution X
corresponds to the adjacency matrix of a graph which is undirected and has all degrees equal to 1.
Now we will show that it has the other properties:
√
√
• λk+1 = 1 − σk+1 (X) ≥ 1 − σk+1 (Y Π + εY ⊥ ) = 1 − ε.
• maxT φX (T ) ≤ 2kΓ∗ T (I − X)Γ∗ k2 ≤ O(kY T (I − X)Y k2 ) + ε ≤ O(ε).
• Recall maxS φX (S) ≥ 12 kΓT (I − X)Γk2 so σk (Y T Γ) ≥ 1 − O(δ + ε). In particular σk (Γ∗ T Γ) ≥
√
√
1 − O( ε + δ). Using Theorem 2.8, we have ∆(Γ, Γ∗ ) ≤ O( ε + δ).
8
8.1
Omitted Proofs
Proof of Theorem 2.8
b 2 ≤ 2∆(Γ, Γ).
b ∈ DisjV (k); ∆(Γ, Γ)
b ≤ kΓ⊥ Γk
b
Theorem 8.1 (Restatement of Theorem 2.8). Given Γ, Γ
2
2
b kΓ − Γk
b ≤ 4∆(Γ, Γ).
b
Moreover, after appropriately ordering the columns of Γ,
2
b T Γ⊥ Γk
b 2 is
Upper Bound. Recall that Γ⊥ is a Laplacian matrix. From Lemma 2.13, we see that kΓ
b let S be the matching set in Γ so
within factor-2 of the maximum diagonal element. Given T ∈ Γ
b T Γ⊥ Γ
b corresponding to T is:
b The diagonal element of Γ
that ∆(S, T ) ≤ ∆(Γ, Γ).
2
eT T Γ⊥ eT ≤ ke⊥
S eT k = 1 −
|S ∩ T |2
= ∆(S, T ) ≤ ε.
|S||T |
b 2 ≤ 2ε. Before proving the lower bound, we will show how this bounds kΓ − Γk
b 2:
Therefore kΓ⊥ Γk
2
1
b T (Γ − Γ)
b = Ik − 1 (ΓT Γ
b +Γ
b T Γ).
(Γ − Γ)
2
2
q
b 2 ≤ 2ε.
b 2 ≤ 1 − σmin (Γ
b T Γ) = 1 − 1 − kΓ⊥ Γk
So 12 kΓ − Γk
2
2
b and π2 : Γ
b → Γ as the following:
Lower Bound. We define π1 : Γ → Γ
def
∀S ∈ Γ : π1 (S) = argmax
b
T ∈Γ
|S ∩ T |
,
|T |
|S ∩ T |
b : π2 (T ) def
∀T ∈ Γ
= argmax
.
|S|
S∈Γ
Consider M = {(S, π1 (S)) | S ∈ Γ}: By Claims 8.3 and 8.4, M is indeed a perfect matching between
b Now consider any matched pair (S, T ) ∈ M . Without loss of generality, say |S| ≥ |T |. By
Γ and Γ.
Claim 8.2, |S ∩ T | ≥ (1 − ε)|S|. Since |S∆T | = |S| + |T | − 2|S ∩ T |:
|S∆T | ≤ |S| + |T | − 2(1 − ε)|S| = 2ε|S| + (|T | − |S|) ≤ 2ε|S|.
We finish our proof with Claims 8.2, 8.3 and 8.4.
23
Claim 8.2. If π1 (S) = T , then |S ∩ T | ≥ (1 − ε)|T |. Similarly, if π2 (T ) = S, then |S ∩ T | ≥ (1 − ε)|S|.
bΓ
b T Γ ∈ Sk so that λmin (P ) = σ 2 (ΓT Γ).
b Thus λmin (P ) =
Proof. Consider the matrix P = ΓT Γ
+
min
b 2 ≥ 1 − ε. In particular, all diagonals of P are at least 1 − ε. Consider any diagonal
σmin (ΓT Γ)
corresponding to S ∈ Γ:
X |S ∩ T |2
bΓ
b T eS
eTS Γ
=
|S|
|S||T |
b
T ∈Γ
!
|S ∩ T 0 | X |S ∩ T |
≤ max
|T 0 |
|S|
b
T 0 ∈Γ
1−ε≤
b
T ∈Γ
= max
b
T 0 ∈Γ
which, by construction, is equal to
T 0|
|S ∩
|T 0 |
,
|S∩π1 (S)|
|π1 (S)| .
This proves the first part of the claim. The second part
b and Γ.
follows immediately by applying the same argument on Γ
Claim 8.3. Both π1 and π2 are bijections.
Proof. Suppose π1 (S) = π1 (S 0 ) = T for some S 6= S 0 . Since S, S 0 are disjoint and ε < 12 :
|T | ≥ |S ∩ T | + |S 0 ∩ T | ≥ 2(1 − ε)|T | > |T |,
a contradiction. A similar argument shows that π2 is a bijection as well.
Claim 8.4. π1 = π2−1 .
b are bijections by Claim 8.3, there exists a cycle of the form
Proof. Suppose not. Since both Γ and Γ
(S0 , T0 , . . . , Sm−1 , Tm−1 , Sm = S0 )
where π1 (Si ) = Ti and π2 (Ti ) = Si+1 for some m ≥ 2. By construction, |Si ∩ Ti | ≥ (1 − ε)|Ti | which
means ε|Ti | ≥ |Ti \ Si |. Since Si and Si+1 are disjoint, |Ti \ Si | ≥ |Ti ∩ Si+1 |. Again, by construction,
|Ti ∩ Si+1 | ≥ (1 − ε)|Si+1 |. Therefore ε|Ti | ≥ (1 − ε)|Si+1 | which implies |Ti | ≥ 1−ε
ε |Si+1 | > |Si+1 |
since ε < 1/2. By a similar argument, we can also show that |Si | > |Ti |. Consequently, |S0 | > |S1 | >
. . . > |Sm | = |S0 | which is a contradiction. So all cycles have length 2, which implies π1 = π2−1 .
Acknowledgments
We thank Moses Charikar and Ravishankar Krishnaswamy for stimulating discussions about the
problem. We also thank Anup Rao for pointing out that Lemma 2.10 holds with equality.
24
References
[ABS10]
Sanjeev Arora, Boaz Barak, and David Steurer. Subexponential algorithms for Unique
Games and related problems. In 51th Annual IEEE Symposium on Foundations of Computer
Science, FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages 563–572, 2010.
[AKY99] Charles J. Alpert, Andrew B. Kahng, and So-Zen Yao. Spectral partitioning with multiple
eigenvectors. Discrete Applied Mathematics, 90(1-3):3–26, 1999.
[AM85]
Noga Alon and V. D. Milman. λ1 , Isoperimetric inequalities for graphs, and superconcentrators. J. Comb. Theory, Ser. B, 38(1):73–88, 1985.
[AS12]
Pranjal Awasthi and Or Sheffet. Improved spectral-norm bounds for clustering. In
Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques 15th International Workshop, APPROX 2012, and 16th International Workshop, RANDOM
2012, Cambridge, MA, USA, August 15-17, 2012. Proceedings, pages 37–49, 2012.
[BFK+ 11] Nikhil Bansal, Uriel Feige, Robert Krauthgamer, Konstantin Makarychev, Viswanath
Nagarajan, Joseph Naor, and Roy Schwartz. Min-max graph partitioning and small set
expansion. In IEEE 52nd Annual Symposium on Foundations of Computer Science, FOCS
2011, Palm Springs, CA, USA, October 22-25, 2011, pages 17–26, 2011.
[Bol13]
Marianna Bolla. Spectral Clustering and Biclustering: Learning Large Graphs and Contingency
Tables. Wiley, 2013.
[BS06]
Nikhil Bansal and Maxim Sviridenko. The Santa Claus problem. In Proceedings of the
38th Annual ACM Symposium on Theory of Computing, Seattle, WA, USA, May 21-23, 2006,
pages 31–40, 2006.
[BV08]
S. Charles Brubaker and Santosh Vempala. Isotropic PCA and affine-invariant clustering.
In 49th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2008, October
25-28, 2008, Philadelphia, PA, USA, pages 551–560, 2008.
[DR10]
Amit Deshpande and Luis Rademacher. Efficient volume sampling for row/column
subset selection. In 51th Annual IEEE Symposium on Foundations of Computer Science,
FOCS 2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages 329–338, 2010.
[GKB13] Alex Gittens, Prabhanjan Kambadur, and Christos Boutsidis. Spectral clustering via the
power method – provably. CoRR, abs/1311.2854, 2013.
[GS12]
Venkatesan Guruswami and Ali Kemal Sinop. Optimal column-based low-rank matrix
reconstruction. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2012, Kyoto, Japan, January 17-19, 2012, pages 1207–1214, 2012.
[KK10]
Amit Kumar and Ravindran Kannan. Clustering with spectral norm and the k-Means
algorithm. In 51th Annual IEEE Symposium on Foundations of Computer Science, FOCS
2010, October 23-26, 2010, Las Vegas, Nevada, USA, pages 299–308, 2010.
25
[KLL+ 13] Tsz Chiu Kwok, Lap Chi Lau, Yin Tat Lee, Shayan Oveis Gharan, and Luca Trevisan.
Improved Cheeger’s inequality: analysis of spectral partitioning algorithms through
higher order spectral gap. In Symposium on Theory of Computing Conference, STOC’13,
Palo Alto, CA, USA, June 1-4, 2013, pages 11–20, 2013.
[KVV04] Ravi Kannan, Santosh Vempala, and Adrian Vetta. On clusterings: Good, bad and
spectral. Journal of the ACM, 51(3):497–515, 2004.
[LGT14]
James R. Lee, Shayan Oveis Gharan, and Luca Trevisan. Multiway spectral partitioning
and higher-order Cheeger inequalities. Journal of the ACM, 61(6):37, 2014.
[LM14]
Anand Louis and Konstantin Makarychev. Approximation algorithm for sparsest kpartitioning. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7, 2014, pages 1244–1255, 2014.
[LRTV12] Anand Louis, Prasad Raghavendra, Prasad Tetali, and Santosh Vempala. Many sparse
cuts via higher eigenvalues. In Proceedings of the 44th Symposium on Theory of Computing
Conference, STOC 2012, New York, NY, USA, May 19 - 22, 2012, pages 1131–1140, 2012.
[NJW01] Andrew Y. Ng, Michael I. Jordan, and Yair Weiss. On spectral clustering: Analysis and
an algorithm. In Advances in Neural Information Processing Systems 14 [Neural Information
Processing Systems: Natural and Synthetic, NIPS 2001, December 3-8, 2001, Vancouver, British
Columbia, Canada], pages 849–856, 2001.
[SM00]
Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Trans.
Pattern Anal. Mach. Intell., 22(8):888–905, 2000.
[SS90]
G. W. Stewart and Ji-guang Sun. Matrix Perturbation Theory. Academic Press, 1990.
[YS03]
Stella X. Yu and Jianbo Shi. Multiclass spectral clustering. In 9th IEEE International
Conference on Computer Vision (ICCV 2003), 14-17 October 2003, Nice, France, pages 313–
319, 2003.
26
| 8cs.DS
|
arXiv:1610.10023v1 [math.AC] 31 Oct 2016
QUASI-PRÜFER EXTENSIONS OF RINGS
GABRIEL PICAVET AND MARTINE PICAVET-L’HERMITTE
Abstract. We introduce quasi-Prüfer ring extensions, in order
to relativize quasi-Prüfer domains and to take also into account
some contexts in recent papers, where such extensions appear in a
hidden form. An extension is quasi-Prüfer if and only if it is an INC
pair. The class of these extensions has nice stability properties.
We also define almost-Prüfer extensions that are quasi-Prüfer, the
converse being not true. Quasi-Prüfer extensions are closely linked
to finiteness properties of fibers. Applications are given for FMC
extensions, because they are quasi-Prüfer.
1. Introduction and Notation
We consider the category of commutative and unital rings. An epimorphism is an epimorphism of this category. Let R ⊆ S be a (ring)
extension. The set of all R-subalgebras of S is denoted by [R, S]. The
extension R ⊆ S is said to have FIP (for the “finitely many intermediate algebras property”) if [R, S] is finite. A chain of R-subalgebras of S
is a set of elements of [R, S] that are pairwise comparable with respect
to inclusion. We say that the extension R ⊆ S has FCP (for the “finite
chain property”) if each chain in [R, S] is finite. Dobbs and the authors
characterized FCP and FIP extensions [10]. Clearly, an extension that
satisfies FIP must also satisfy FCP. An extension R ⊆ S is called FMC
if there is a finite maximal chain of extensions from R to S.
We begin by explaining our motivations and aims. The reader who is
not familiar with the notions used will find some Scholia in the sequel,
as well as necessary definitions that exist in the literature. Knebusch
and Zang introduced Prüfer extensions in their book [26]. Actually,
these extensions are nothing but normal pairs, that are intensively
studied in the literature. We do not intend to give an extensive list of
recent papers, written by Ayache, Ben Nasr, Dobbs, Jaballah, Jarboui
and some others. We are indebted to these authors because their papers are a rich source of suggestions. We observed that some of them
2010 Mathematics Subject Classification. Primary:13B22; Secondary:
Key words and phrases. flat epimorphism, FIP, FCP extension, minimal extension, integral extension, Morita, Prüfer hull, support of a module.
1
2
G. PICAVET AND M. PICAVET
are dealing with FCP (FIP, FMC) extensions, followed by a Prüfer
extension, perhaps under a hidden form. These extensions reminded
us quasi-Prüfer domains (see [18]). Therefore, we introduced in [38]
quasi-Prüfer extensions R ⊆ S as extensions that can be factored
R ⊆ R′ ⊆ S, where the first extension is integral and the second is
Prüfer. Note that FMC extensions are quasi-Prüfer.
We give a systematic study of quasi-Prüfer extensions in Section 2
and Section 3. The class of quasi-Prüfer extensions has a nice behavior
with respect to the classical operations of commutative algebra. An important result is that quasi-Prüfer extensions coincide with INC-pairs.
Another one is that this class is stable under forming subextensions
and composition. A striking result is the stability of the class of quasiPrüfer extensions by absolutely flat base change, like localizations and
Henselizations. Any ring extension R ⊆ S admits a quasi-Prüfer closure, contained in S. Examples are provided by Laskerian pairs, open
pairs and the pseudo-Prüfer pairs of Dobbs-Shapiro [15].
Section 4 deals with almost-Prüfer extensions, a special kind of quasiPrüfer extensions. They are of the form R ⊆ T ⊆ S, where the first
extension is Prüfer and the second is integral. Any ring extension
admits an almost-Prüfer closure, contained in S. The class of almostPrüfer extensions seems to have less properties than the class of quasiPrüfer extensions but has the advantage of the commutation of Prüfer
closures with localizations at prime ideals. We examine the transfer of
the quasi (almost)-Prüfer properties to subextensions.
Section 5 study the transfer of the quasi (almost)-Prüfer properties
to Nagata extensions.
In section 6, we complete and generalize the results of Ayache-Dobbs
in [5], with respect to the finiteness of fibers. These authors have evidently considered particular cases of quasi-Prüfer extensions. A main
result is that if R ⊆ S is quasi-Prüfer with finite fibers, then so is
R ⊆ T for T ∈ [R, S]. In particular, we recover a result of [5] about
FMC extensions.
Now Section 7 gives calculations of |[R, S]| with respect to its Prüfer
closure, quasi-Prüfer (almost-Prüfer) closure in case R ⊆ S has FCP.
1.1. Recalls about some results and definitions. The reader is
warned that we will mostly use the definition of Prüfer extensions by
flat epimorphic subextensions investigated in [26]. The results needed
may be found in Scholium A for flat epimorphic extensions and some
results of [26] are summarized in Scholium B. Their powers give quick
proofs of results that are generalizations of results of the literature.
QUASI-PRÜFER EXTENSIONS
3
As long as FCP or FMC extensions are concerned, we use minimal
(ring) extensions, a concept introduced by Ferrand-Olivier [17]. An
extension R ⊂ S is called minimal if [R, S] = {R, S}. It is known that
a minimal extension is either module-finite or a flat epimorphism [17]
and these conditions are mutually exclusive. There are three types of
integral minimal (module-finite) extensions: ramified, decomposed or
inert [36, Theorem 3.3]. A minimal extension R ⊂ S admits a crucial
ideal C(R, S) =: M which is maximal in R and such that RP = SP
for each P 6= M, P ∈ Spec(R). Moreover, C(R, S) = (R : S) when
R ⊂ S is an integral minimal extension. The key connection between
the above ideas is that if R ⊆ S has FCP or FMC, then any maximal
(necessarily finite) chain of R-subalgebras of S, R = R0 ⊂ R1 ⊂ · · · ⊂
Rn−1 ⊂ Rn = S, with length n < ∞, results from juxtaposing n
minimal extensions Ri ⊂ Ri+1 , 0 ≤ i ≤ n − 1.
Following [24], we define the length ℓ[R, S] of [R, S] as the supremum
of the lengths of chains in [R, S]. In particular, if ℓ[R, S] = r, for some
integer r, there exists a maximal chain in [R, S] with length r.
As usual, Spec(R), Max(R), Min(R), U(R), Tot(R) are respectively
the set of prime ideals, maximal ideals, minimal prime ideals, units,
total ring of fractions of a ring R and κ(P ) = RP /P RP is the residual
field of R at P ∈ Spec(R).
If R ⊆ S is an extension, then (R : S) is its conductor and if P ∈
Spec(R), then SP is the localization SR\P . We denote the integral
S
closure of R in S by R (or R).
A local ring is here what is called elsewhere a quasi-local ring. The
support of an R-module E is SuppR (E) := {P ∈ Spec(R) | EP 6= 0}
and MSuppR (E) := SuppR (E) ∩ Max(R). Finally, ⊂ denotes proper
inclusion and |X| the cardinality of a set X.
Scholium A We give some recalls about flat epimorphisms (see [27,
Chapitre IV], except (2) which is [31, Proposition 2]).
(1) R → S is a flat epimorphism ⇔ for all P ∈ Spec(R), either
RP → SP is an isomorphism or S = P S ⇔ RP ⊆ SP is a flat
epimorphism for all P ∈ Spec(R).
(2) (S) A flat epimorphism, with a zero-dimensional domain, is surjective.
(3) If f : A → B and g : B → C are ring morphisms such that g ◦ f
is injective and f is a flat epimorphism, then g is injective.
(4) Let R ⊆ T ⊆ S be a tower of extensions, such that R ⊆ S is a
flat epimorphism. Then T ⊆ S is a flat epimorphism but R ⊆ T
does not need. A Prüfer extension remedies to this defect.
4
G. PICAVET AND M. PICAVET
(5) (L) A faithfully flat epimorphism is an isomorphism. Hence,
R = S if R ⊆ S is an integral flat epimorphism.
(6) If f : R → S is a flat epimorphism and J an ideal of S, then
J = f −1 (J)S.
(7) If f : R → S is an epimorphism, then f is spectrally injective
and its residual extensions are isomorphisms.
(8) Flat epimorphisms remain flat epimorphisms under base change
(in particular, after a localization with respect to a multiplicatively closed subset).
(9) Flat epimorphisms are descended by faithfully flat morphisms.
1.2. Recalls and results on Prüfer extensions. We recall some
definitions and properties of ring extensions R ⊆ S and rings R. There
are a lot of characterizations of Prüfer extensions. We keep only those
that are useful in this paper. We give the two definitions that are dual
and emphasize some characterizations in the local case.
Scholium B
(1) [26] R ⊆ S is called Prüfer if R ⊆ T is a flat epimorphism for
each T ∈ [R, S].
(2) R ⊆ S is called a normal pair if T ⊆ S is integrally closed for
each T ∈ [R, S].
(3) R ⊆ S is Prüfer if and only if it is a normal pair [26, Theorem
5.2(4)].
(4) R is called Prüfer if its finitely generated regular ideals are
invertible, or equivalently, R ⊆ Tot(R) is Prüfer [20, Theorem
13((5)(9))].
Hence Prüfer extensions are a relativization of Prüfer rings. Clearly,
a minimal extension is a flat epimorphism if and only if it is Prüfer.
We will then use for such extensions the terminology: Prüfer minimal
extensions. The reader may find some properties of Prüfer minimal
extensions in [36, Proposition 3.2, Lemma 3.4 and Proposition 3.5], asserted by L. Dechene in her dissertation, but where in addition R must
be supposed local. The reason why is that this word has surprisingly
disappeared during the printing process of [36].
We will need the two next results. Some of them do not explicitely
appear in [26] but deserve to be emphasized. We refer to [26, Definition
1, p.22] for a definition of Manis extensions.
Proposition 1.1. Let R ⊆ S be a ring extension.
(1) R ⊆ S is Prüfer if and only if RP ⊆ SP is Prüfer for each
P ∈ Spec(R) (respectively, P ∈ Supp(S/R)).
QUASI-PRÜFER EXTENSIONS
5
(2) R ⊆ S is Prüfer if and only if RM ⊆ SM is Manis for each
M ∈ Max(R).
Proof. (1) The class of Prüfer extensions is stable under localization [26,
Proposition 5.1(ii), p.46-47]. To get the converse, use Scholium A(1).
(2) follows from [26, Proposition 2.10, p.28, Definition 1, p.46].
Proposition 1.2. Let R ⊆ S be a ring extension, where R is local.
(1) R ⊆ S is Manis if and only if S \ R ⊆ U(S) and x ∈ S \ R ⇒
x−1 ∈ R. In that case, R ⊆ S is integrally closed.
(2) R ⊆ S is Manis if and only if R ⊆ S is Prüfer.
(3) R ⊆ S is Prüfer if and only if there exists P ∈ Spec(R) such
that S = RP , P = SP and R/P is a valuation domain. Under
these conditions, S/P is the quotient field of R/P .
Proof. (1) is [26, Theorem 2.5, p.24]. (2) is [26, Scholium 10.4, p.147].
Then (3) is [10, Theorem 6.8].
Next result shows that Prüfer FCP extensions can be described in a
special manner.
Proposition 1.3. Let R ⊂ S be a ring extension.
(1) If R ⊂ S has FCP, then R ⊂ S is integrally closed ⇔ R ⊂ S is
Prüfer ⇔ R ⊂ S is a composite of Prüfer minimal extensions.
(2) If R ⊂ S is integrally closed, then R ⊂ S has FCP ⇔ R ⊂ S is
Prüfer and Supp(S/R) is finite.
Proof. (1) Assume that R ⊂ S has FCP. If R ⊂ S is integrally closed,
then, R ⊂ S is composed of Prüfer minimal extensions by [10, Lemma
3.10]. Conversely, if R ⊂ S is composed of Prüfer minimal extensions,
R ⊂ S is integrally closed, since so is each Prüfer minimal extension. A
Prüfer extension is obviously integrally closed, and an FCP integrally
closed extension is Prüfer by [10, Theorem 6.3].
(2) The logical equivalence is [10, Theorem 6.3].
Definition 1.4. [26] A ring extension R ⊆ S has:
bS , called the
(1) a greatest flat epimorphic subextension R ⊆ R
Morita hull of R in S.
eS , called the Prüfer hull
(2) a greatest Prüfer subextension R ⊆ R
of R in S.
b := R
bS and R
e := R
eS , if no confusion can occur. R ⊆ S is
We set R
e
called Prüfer-closed if R = R.
eS is denoted by P(R, S) in [26] and R
bS is the weakly
Note that R
surjective hull M(R, S) of [26]. Our terminology is justified because
6
G. PICAVET AND M. PICAVET
Morita’s work is earlier [30, Corollary 3.4]. The Morita hull can be
computed by using a (transfinite) induction [30]. Let S ′ be the set of
all s ∈ S such that there is some ideal I of R, such that IS = S and
Is ⊆ R. Then R ⊆ S ′ is a subextension of R ⊆ S. We set S1 := S ′
and Si+1 := (Si )′ ⊆ Si . By [30, p.36], if R ⊂ S is an FCP extension,
b = Sn for some integer n.
then R
At this stage it is interesting to point out a result; showing again
that integral closedness and Prüfer extensions are closely related.
Proposition 1.5. Olivier [33, Corollary, p.56] An extension R ⊆ S is
integrally closed if and only if there is a pullback square:
R −−−→
y
S
y
V −−−→ K
where V is a semi-hereditary ring and K its total quotient ring.
In that case V ⊆ K is a Prüfer extension, since V is a Prüfer ring,
whose localizations at prime ideals are valuation domains and K is an
absolutely flat ring. As there exist integrally closed extensions that
are not Prüfer, we see in passing that the pullback construction may
not descend Prüfer extensions. The above result has a companion for
minimal extensions that are Prüfer [21, Proposition 3.2].
Proposition 1.6. Let R ⊆ S be an extension and T ∈ [R, S], then
eT = R
e ∩ T . Therefore, for T, U ∈ [R, S] with T ⊆ U, then R
eT ⊆ R
eU .
R
eT is the greatest Prüfer extenProof. Obvious, since the Prüfer hull R
sion R ⊆ V contained in T .
e if R ⊆ S has FCP.
We will show later that in some cases Te ⊆ U
2. Quasi-Prüfer extensions
We introduced the following definition in [38, p.10].
Definition 2.1. An extension of rings R ⊆ S is called quasi-Prüfer if
one of the following equivalent statements holds:
(1) R ⊆ S is a Prüfer extension;
(2) R ⊆ S can be factored R ⊆ T ⊆ S, where R ⊆ T is integral
and T ⊆ S is Prüfer. In that case R = T
To see that (2) ⇒ (1) observe that if (2) holds, then T ⊆ R is integral
and a flat injective epimorphism, so that R = T by (L) (Scholium A(5)).
QUASI-PRÜFER EXTENSIONS
7
We observe that quasi-Prüfer extensions are akin to quasi-finite extensions if we refer to Zariski Main Theorem. This will be explored in
Section 6, see for example Theorem 6.2.
Hence integral or Prüfer extensions are quasi-Prüfer. An extension
is clearly Prüfer if and only if it is quasi-Prüfer and integrally closed.
Quasi-Prüfer extensions allow us to avoid FCP hypotheses.
We give some other definitions involved in ring extensions R ⊆ S.
The fiber at P ∈ Spec(R) of R ⊆ S is FibR,S (P ) := {Q ∈ Spec(S) |
Q ∩ R = P }. The subspace FibR,S (P ) of Spec(S) is homeomorphic
to the spectrum of the fiber ring FR,S (P ) := κ(P ) ⊗R S at P . The
homeomorphism is given by the spectral map of S → κ(P ) ⊗R S and
κ(P ) → κ(P ) ⊗R S is the fiber morphism at P .
Definition 2.2. A ring extension R ⊆ S is called:
(1) incomparable if for each pair Q ⊆ Q′ of prime ideals of S, then
Q ∩ R = Q′ ∩ R ⇒ Q = Q′ , or equivalently, κ(P ) ⊗R T is a
zero-dimensional ring for each T ∈ [R, S] and P ∈ Spec(R),
such that κ(P ) ⊗R T 6= 0.
(2) an INC-pair if R ⊆ T is incomparable for each T ∈ [R, S].
(3) residually algebraic if R/(Q ∩ R) ⊆ S/Q is algebraic for each
Q ∈ Spec(S).
(4) a residually algebraic pair if the extension R ⊆ T is residually
algebraic for each T ∈ [R, S].
The following characterization was announced in [38]. We were unaware that this result is also proved in [7, Corollary 1], when we present
it in ArXiv. However, our proof is largely shorter because we use the
powerful results of [26].
Theorem 2.3. An extension R ⊆ S is quasi-Prüfer if and only if
R ⊆ S is an INC-pair and, if and only if, R ⊆ S is a residually
algebraic pair.
Proof. Suppose that R ⊆ S is quasi-Prüfer and let T ∈ [R, S]. We
set U := RT . Then R ⊆ U is a flat epimorphism by definition of a
Prüfer extension and hence is incomparable as is R ⊆ R . It follows
that R ⊆ U is incomparable. Since T ⊆ U is integral, it has going-up.
It follows that R ⊆ T is incomparable. Conversely, if R ⊆ S is an
INC-pair, then so is R ⊆ S. Since R ⊆ S is integrally closed, R ⊆ S
is Prüfer [26, Theorem 5.2,(9’), p.48]. The second equivalence is [14,
Proposition 2.1] or [18, Theorem 6.5.6].
Corollary 2.4. An extension R ⊆ S is quasi-Prüfer if and only if
R ⊆ T is Prüfer for each T ∈ [R, S].
8
G. PICAVET AND M. PICAVET
It follows that most of the properties described in [6] for integrally
closed INC-pairs of domains are valid for arbitrary ring extensions.
Moreover, a result of Dobbs is easily gotten: an INC-pair R ⊆ S is
an integral extension if and only if R ⊆ S is spectrally surjective [14,
Theorem 2.2]. This follows from Scholium A, Property (L).
Example 2.5. Quasi-Prüfer domains R with quotient fields K can be
characterized by R ⊆ K is quasi-Prüfer. The reader may consult [9,
Theorem 1.1] or [18]. In view of [2, Theorem 2.7], R is a quasi-Prüfer
domain if and only if Spec(R(X)) → Spec(R) is bijective.
We give here another example of quasi-Prüfer extension. An extension R ⊂ S is called a going-down pair if each of its subextensions has
the going-down property. For such a pair, R ⊆ T has incomparability
for each T ∈ [R, S], at each non-maximal prime ideal of R [3, Lemma
5.8](ii). Now let M be a maximal ideal of R, whose fiber is not void in
T . Then R ⊆ T is a going-down pair, and so is R/M ⊆ T /MT because
MT ∩ R = M. By [3, Corollary 5.6], the dimension of T /MT is ≤ 1.
Therefore, if R ⊂ S is a going-down pair, then R ⊂ S is quasi-Prüfer
if and only if dim(T /MT ) 6= 1 for each T ∈ [R, S] and M ∈ Max(R).
Also open-ring pairs R ⊂ S are quasi-Prüfer by [8, Proposition 2.13].
An i-pair is an extension R ⊆ S such that Spec(T ) → Spec(R) is
injective for each T ∈ [R, S], or equivalently if and only if R ⊆ S is
quasi-Prüfer and R ⊆ R is spectrally injective [38, Proposition 5.8].
These extensions appear frequently in the integral domains context.
Another examples are given by some extensions R ⊆ S, such that
Spec(S) = Spec(R) as sets, as we will see later.
3. Properties of quasi-Prüfer extensions
We now develop the machinery of quasi-Prüfer extensions.
Proposition 3.1. An extension R ⊂ S is (quasi-)Prüfer if and only if
RP ⊆ SP is (quasi-)Prüfer for any P ∈ Spec(R) (P ∈ MSupp(S/R)).
Proof. The proof is easy if we use the INC-pair property definition of
quasi-Prüfer extension (see also [6, Proposition 2.4]).
Proposition 3.2. Let R ⊆ S be a quasi-Prüfer extension and ϕ : S →
S ′ an integral ring morphism. Then ϕ(R) ⊆ S ′ is quasi-Prüfer and
S ′ = ϕ(S)ϕ(R), where ϕ(R) is the integral closure of ϕ(R) in S ′ .
Proof. It is enough to apply [26, Theorem 5.9] to the Prüfer extension
R ⊆ S and to use Definition 2.1.
This result applies with S ′ := S ⊗R R′ , where R → R′ is an integral
morphism. Therefore integrality ascends the quasi-Prüfer property.
QUASI-PRÜFER EXTENSIONS
9
We know that a composite of Prüfer extensions is Prüfer [26, Theorem 5.6, p.51]. The following Corollary 3.3 contains [7, Theorem 3].
Corollary 3.3. Let R ⊆ T ⊆ S be a tower of extensions. Then R ⊆ S
is quasi-Prüfer if and only if R ⊆ T and T ⊆ S are quasi-Prüfer. It
follows that R ⊆ T is quasi-Prüfer if and only if R ⊆ RT is quasiPrüfer.
Proof. Consider a tower (T ) of extensions R ⊆ R ⊆ S := R′ ⊆ R′ ⊆ S ′
(a composite of two quasi-Prüfer extensions). By using Proposition 3.2
we see that R ⊆ S = R′ ⊆ R′ is quasi-Prüfer. Then (T ) is obtained
by writing on the left an integral extension and on the right a Prüfer
extension. Therefore, (T ) is quasi-Prüfer. We prove the converse.
If R ⊆ T ⊆ S is a tower of extensions, then R ⊆ T and T ⊆ S
are INC-pairs whenever R ⊆ S is an INC-pair. The converse is then a
consequence of Theorem 2.3.
The last statement is [7, Corollary 4].
Using the above corollary, we can exhibit new examples of quasiPrüfer extensions. We recall that a ring R is called Laskerian if each of
its ideals is a finite intersection of primary ideals and a ring extension
R ⊂ S a Laskerian pair if each T ∈ [R, S] is a Laskerian ring. Then [42,
Proposition 2.1] shows that if R is an integral domain with quotient
field F 6= R and F ⊂ K is a field extension, then R ⊂ K is a Laskerian
pair if and only if K is algebraic over R and R (in K) is a Laskerian
Prüfer domain. It follows easily that R ⊂ K is quasi-Prüfer.
Next result generalizes [25, Proposition 1].
Corollary 3.4. An FMC extension R ⊂ S is quasi-Prüfer.
Proof. Because R ⊂ S is a composite of finitely many minimal extensions, by Corollary 3.3, it is enough to observe that a minimal extension
is either Prüfer or integral.
Corollary 3.5. Let R ⊆ S be a quasi-Prüfer extension and a tower
R ⊆ T ⊆ S, where R ⊆ T is integrally closed. Then R ⊆ T is Prüfer.
T
Proof. Observe that R ⊆ T is quasi-Prüfer and then that R = R .
Next result deals with the Dobbs-Shapiro pseudo-Prüfer extensions
of integral domains [15], that they called pseudo-normal pairs. Suppose
that R is local, we call here pseudo-Prüfer an extension R ⊆ S such that
there exists T ∈ [R, S] with Spec(R) = Spec(T ) and T ⊆ S is Prüfer
[15, Corollary 2.5]. If R is arbitrary, the extension R ⊆ S is called
pseudo-Prüfer if RM ⊆ SM is pseudo-Prüfer for each M ∈ Max(R).
In view of the Corollary 3.3, it is enough to characterize quasi-Prüfer
extensions of the type R ⊆ T with Spec(R) = Spec(T ).
10
G. PICAVET AND M. PICAVET
Corollary 3.6. Let R ⊆ T be an extension with Spec(R) = Spec(T )
and (R, M) local. Then R ⊆ T is quasi-Prüfer if and only if Spec(R) =
Spec(U) for each U ∈ [R, T ] and, if and only if R/M ⊆ T /M is an
algebraic field extension. In such a case, R ⊆ T is Prüfer-closed.
Proof. It follows from [1] that M ∈ Max(T ). Part of the proof is gotten
by observing that R ⊆ U is an INC extension if Spec(R) = Spec(U).
e is a spectrally
Another one is proved in [1, Corollary 3.26]. Now R ⊆ R
e
surjective flat epimorphism and then, by Scholium A, R = R.
Let R ⊆ S be an extension and I an ideal shared with R and S. It
is easy to show that R ⊆ S is quasi-Prüfer if and only if R/I ⊆ S/I is
quasi-Prüfer by using [26, Proposition 5.8] in the Prüfer case. We are
able to give a more general statement.
Lemma 3.7. Let R ⊆ S be a (quasi-)Prüfer extension and J an ideal
of S with I = J ∩ R. Then R/I ⊆ S/J is a (quasi-)Prüfer extension.
If R ⊆ S is Prüfer and N is a maximal ideal of S, then R/(N ∩ R) is
a valuation domain with quotient field S/N.
Proof. Assume first that R ⊆ S is Prüfer. We have J = IS by
Scholium A(6), because R ⊆ S is a flat epimorpism. Therefore, any
D ∈ [R/I, S/J] is of the form C/J where C ∈ [R, S]. We can write
C/IS = (C + I)/IS ∼
= C/C ∩ IS. As R ⊆ C is a flat epimorphism,
C ∩ IS = IC. It then follows that D = C ⊗ R/I and we get easily
that R/I ⊆ S/J is Prüfer, since R/I ⊆ D is a flat epimorphism. The
quasi-Prüfer case is an easy consequence.
With this lemma we generalize and complete [23, Proposition 1.1].
Proposition 3.8. Let R ⊆ S be an extension of rings. The following
statements are equivalent:
(1) R ⊆ S is quasi-Prüfer;
(2) R/(Q ∩ R) ⊆ S/Q is quasi-Prüfer for each Q ∈ Spec(S) ;
(3) (X − s)S[X] ∩ R[X] 6⊆ M[X] for each s ∈ S and M ∈ Max(R);
(4) For each T ∈ [R, S], the fiber morphisms of R ⊆ T are integral.
Proof. (1) ⇒ (2) is entailed by Lemma 3.7. Assume that (2) holds and
let M ∈ Max(R) that contains a minimal prime ideal P , lain over by a
minimal prime ideal Q of S. Then (2) ⇒ (3) follows from [23, Proposition 1.1(1)], applied to R/(Q ∩ R) ⊆ S/Q. If (3) holds, argue as in
the paragraph before [23, Proposition 1.1] to get that R ⊆ S is a Pextension, whence an INC-extension by [14, Proposition 2.1]. Because
integral extensions have incomparability, we see that (4) ⇒ (1). Corollary 3.3 shows that the reverse implication holds, if any quasi-Prüfer
QUASI-PRÜFER EXTENSIONS
11
extension R ⊆ S has integral fiber morphisms. For P ∈ Spec(R), the
extension RP /P RP ⊆ SP /P SP is quasi-Prüfer by Lemma 3.7. The
ring RP /P RP is zero-dimensional and RP /P RP → SP /P SP , being a
flat epimorphism, is therefore surjective by Scholium A (S). It follows
that the fiber morphism at P is integral.
Remark 3.9. The logical equivalence (1) ⇔ (2) is still valid if we
replace quasi-Prüfer with integral in the above proposition. It is enough
to show that an extension R ⊆ S is integral when R/P ⊆ S/Q is
integral for each Q ∈ Spec(S) and P := Q ∩ R. We can suppose that
S = R[s] ∼
= R[X]/I, where X is an indeterminate, I an ideal of R[X]
and Q varies in Min(S), because for an extension A ⊆ B, any element
of Min(A) is lain over by some element of Min(B). If Σ is the set of
unitary polynomials of R[X], the assumptions show that any element
of Spec(R[X]), containing I, meets Σ. As Σ is a multiplicatively closed
subset, I ∩ Σ 6= ∅, whence s is integral over R.
But a similar result does not hold if we replace quasi-Prüfer with
Prüfer, except if we suppose that R ⊆ S is integrally closed. To see
this, apply the above proposition to get a quasi-Prüfer extension R ⊆ S
if each R/P ⊆ S/Q is Prüfer. Actually, this situation already occurs
for Prüfer rings and their factor domains, as Lucas’s paper [29] shows.
More precisely, [29, Proposition 2.7] and the third paragraph of [29, p.
336] shows that if R is a ring with Tot(R) absolutely flat, then R is
a quasi-Prüfer ring if R/P is a Prüfer domain for each P ∈ Spec(R).
Now example [29, Example 2.4] shows that R is not necessarily Prüfer.
We observe that if R ⊆ S is quasi-Prüfer, then R/M is a quasiPrüfer domain for each N ∈ Max(S) and M := N ∩ R (in case R ⊆ S
is integral, R/M is a field). To prove this, observe that R/M ⊆ S/N
can be factored R/M ⊆ κ(M) ⊆ S/N. As we will see, R/M ⊆ κ(M)
is quasi-Prüfer because R/M ⊆ S/N is quasi-Prüfer.
The class of Prüfer extensions is not stable by (flat) base change.
For example, let V be a valuation domain with quotient field K. Then
V [X] ⊆ K[X] is not Prüfer [26, Example 5.12, p.53]. Thus if we
consider an ideal I of R and J := IS, R ⊆ S Prüfer may not imply
R/I ⊆ S/IS Prüfer except if IS ∩ R = I. This happens for instance
for a prime ideal I of R that is lain over by a prime ideal of S.
Proposition 3.10. Let R ⊆ S be a (quasi)-Prüfer extension and R →
T a flat epimorphism, then T ⊆ S ⊗R T is (quasi)-Prüfer. If in addition
S and T are both subrings of some ring and R ⊆ T is an extension,
then T ⊆ T S is (quasi)-Prüfer.
12
G. PICAVET AND M. PICAVET
Proof. For the first part, it is enough to consider the Prüfer case. It is
well known that the following diagram is a pushout if Q ∈ Spec(T ) is
lying over P in R:
RP −−−→
y
SP
y
TQ −−−→ (T ⊗R S)Q
As RP → TQ is an isomorphism since R → T is a flat epimorphism
by Scholium A, it follows that RP ⊆ SP identifies to TQ → (T ⊗R S)Q .
The result follows because Prüfer extensions localize and globalize.
In case R → T is a flat epimorphic extension, the surjective maps
T ⊗R S → T S and R ⊗R T → RT are isomorphisms because R → T R
(resp. S → ST ) is injective and R → T ⊗T R (resp. S → S ⊗R T ) is a
flat epimorphism. Then it is enough to use Scholium A.
The reader may find in [26, Corollary 5.11, p.53] that if R ⊆ A ⊆ S
and R ⊆ B ⊆ S are extensions and R ⊆ A and R ⊆ B are both Prüfer,
then R ⊆ AB is Prüfer.
Proposition 3.11. Let R ⊆ A and R ⊆ B be two extensions, where
A and B are subrings of a ring S. If they are both quasi-Prüfer, then
R ⊆ AB is quasi-Prüfer.
Proof. Let U and V be the integral closures of R in A and B. Then R ⊆
A ⊆ AV is quasi-Prüfer because A ⊆ AV is integral and Corollary 3.3
applies. Using again Corollary 3.3 with R ⊆ V ⊆ AV , we find that
V ⊆ AV is quasi-Prüfer. Now Proposition 3.10 entails that B ⊆ AB
is quasi-Prüfer because V ⊆ B is a flat epimorphism. Finally R ⊆ AB
is quasi-Prüfer, since a composite of quasi-Prüfer extensions.
It is known that an arbitrary product of extensions is Prüfer if and
only if each of its components is Prüfer [26, Proposition 5.20, p.56].
The following result is an easy consequence.
Proposition 3.12. Let {Ri ⊆ Si |i = 1, . . . , n} be a finite family of
quasi-Prüfer extensions, then R1 × · · · × Rn ⊆ S1 × · · · × Sn is quasiPrüfer. In particular, if {R ⊆ Si |i = 1, . . . , n} is a finite family of
quasi-Prüfer extensions, then R ⊆ S1 × · · · × Sn is quasi-Prüfer.
In the same way we have the following result deduced from [26,
Remark 5.14, p.54].
Proposition 3.13. Let R ⊆ S be an extension of rings and an upward
directed family {Rα |α ∈ I} of elements of [R, S] such that R ⊆ Rα is
quasi-Prüfer for each α ∈ I. Then R ⊆ ∪[Rα |α ∈ I] is quasi-Prüfer.
QUASI-PRÜFER EXTENSIONS
13
Proof. It is enough to use [26, Proposition 5.13, p.54] where Aα is the
integral closure of R in Rα .
A ring morphism R → T preserves the integral closure of ring morT ⊗R S
∼
phisms R → S if T
= T ⊗R R for every ring morphism R → S. An
absolutely flat morphism R → T (R → T and T ⊗R T → T are both
flat) preserves integral closure [33, Theorem 5.1]. Flat epimorphisms,
Henselizations and étale morphisms are absolutely flat. Another examples are morphisms R → T that are essentially of finite type and
(absolutely) reduced [37, Proposition 5.19](2). Such morphisms are flat
if R is reduced [28, Proposition 3.2].
We will prove an ascent result for absolutely flat ring morphisms.
This will be proved by using base changes. For this we need to introduce
some concepts. A ring A is called an AIC ring if each monic polynomial
of A[X] has a zero in A. We recalled in [35, p.4662] that any ring A
has a faithfully flat integral extension A → A∗ , where A∗ is an AIC
ring. Moreover, if A is an AIC ring, each localization AP at a prime
ideal P of A is a strict Henselian ring [35, Lemma II.2].
Theorem 3.14. Let R ⊆ S be a (quasi-) Prüfer extension and R → T
an absolutely flat ring morphism. Then T → T ⊗R S is a (quasi-)
Prüfer extension.
Proof. We can suppose that R is an AIC ring. To see this, it is enough
to use the base change R → R∗ . We set T ∗ := T ⊗R R∗ , S ∗ := S ⊗R R∗ .
We first observe that R∗ ⊆ S ∗ is quasi-Prüfer for the following reason:
the composite extension R ⊆ S ⊆ S ∗ is quasi-Prüfer because the last
extension is integral. Moreover, R∗ → T ∗ is absolutely flat. In case
T ∗ ⊆ T ∗ ⊗R∗ S ∗ is quasi-Prüfer, so is T ⊆ T ⊗R S, because T → T ∗ =
T ⊗R R∗ is faithfully flat and T ∗ ⊆ T ∗ ⊗R∗ S ∗ is deduced from T ⊆R S
by the faithfully flat base change T → T ⊗R S. It is then enough to
apply Proposition 3.17.
We thus assume from now on that R is an AIC ring.
Let N ∈ Spec(T ) be lying over M in R. Then RM → TN is absolutely
flat [32, Proposition f] and RM ⊆ SM is quasi-Prüfer. Now observe
that (T ⊗R S)N ∼
= TN ⊗RM SM . Therefore, we can suppose that R and
T are local and R → T is local and injective. We deduce from [33,
Theorem 5.2], that RM → TN is an isomorphism. Therefore the proof
is complete in the quasi-Prüfer case. For the Prüfer case, we need only
to observe that absolutely flat morphisms preserve integral closure and
a quasi-Prüfer extension is Prüfer if it is integrally closed.
14
G. PICAVET AND M. PICAVET
Proposition 3.15. Let R ⊆ S be an extension of rings and R → T a
base change which preserves integral closure. If T ⊆ T ⊗R S has FCP
and R ⊆ S is Prüfer, then T ⊆ T ⊗R S is Prüfer.
Proof. The result holds because an FCP extension is Prüfer if and only
if it is integrally closed.
e ⊆ Te needs not to be an isomorphism, since
We observe that T ⊗R R
this property may fail even for a localization R → RP , where P is a
prime ideal of R.
Proposition 3.16. Let R ⊆ S be an extension of rings, R → R′ a
faithfully flat ring morphism and set S ′ := R′ ⊗R S. If R′ ⊆ S ′ is
(quasi-) Prüfer (respectively, FCP), then so is R ⊆ S.
Proof. The Prüfer case is clear, because faithfully flat morphisms descend flat epimorphisms (Scholium A (9)). For the quasi-Prüfer case,
we use the INC-pair characterization and the fact that FR,S (P ) →
FR′ ,S ′ (P ′ ) is faithfully flat for P ′ ∈ Spec(R′ ) lying over P in R [22,
Corollaire 3.4.9]. The FCP case is proved in [11, Theorem 2.2].
Proposition 3.17. Let R ⊆ S be a ring extension and R → R′ a
spectrally surjective ring morphism (for example, either faithfully flat
or injective and integral). Then R ⊆ S is quasi-Prüfer if R′ → R′ ⊗R S
is injective (for example, if R → R′ is faithfully flat) and quasi-Prüfer.
Proof. Let T ∈ [R, S] and P ∈ Spec(R) and set T ′ := T ⊗R R′ . There
is some P ′ ∈ Spec(R′ ) lying over P , because R → R′ is spectrally
surjective. There is a faithfully flat morphism FR,T (P ) → FR′ ,T ′ (P ′) ∼
=
′
FR,T (P ) ⊗k(P ) κ(P ) [22, Corollaire 3.4.9]. By Theorem 2.3, the result
follows from the faithful flatness of FR,T (P ) → FR′ ,T ⊗R R′ (P ′ ).
Theorem 3.18. Let R ⊆ S be a ring extension.
=⇒
e
(1) R ⊆ S has a greatest quasi-Prüfer subextension R ⊆ R = R.
=⇒
e =: R
~ is quasi-Prüfer and then R
~ ⊆ R.
(2) R ⊆ RR
=⇒
(3) R
R
=⇒
e R = R.
e
= R and R
Proof. To see (1), use Proposition 3.13 which tells us that the set of all
quasi-Prüfer subextensions is upward directed and then use Proposi=⇒
=⇒
tion 3.12 to prove the existence of R . Then let R ⊆ T ⊆ R be a tower
=⇒
e ⊆ =⇒
R , we
with R ⊆ T integral and T ⊆ R Prüfer. From T ⊆ R ⊆ R
=⇒
e
deduce that T = R and then R = R.
e can be factored R ⊆ R
e ⊆ RR
e and is a tower of
(2) Now R ⊆ RR
e → RR
e is integral.
quasi-Prüfer extensions, because R
QUASI-PRÜFER EXTENSIONS
15
=⇒
(3) Clearly, the integral closure and the Prüfer closure of R in R are
=⇒
=⇒
e with R , and R, R
e ⊆ R.
the respective intersections of R and R
This last result means that, as long integral closures and Prüfer
=⇒
closures of subsets of R are concerned, we can suppose that R ⊆ S is
quasi-Prüfer.
4. Almost-Prüfer extensions
We next give a definition “dual” of the definition of a quasi-Prüfer
extension.
4.1. Arbitrary extensions.
Definition 4.1. A ring extension R ⊆ S is called an almost-Prüfer
extension if it can be factored R ⊆ T ⊆ S, where R ⊆ T is Prüfer and
T ⊆ S is integral.
Proposition 4.2. An extension R ⊆ S is almost-Prüfer if and only if
e ⊆ S is integral. It follows that the subring T of the above definition
R
e=R
b when R ⊆ S is almost-Prüfer.
is R
Proof. If R ⊆ S is almost-Prüfer, there is a factorization R ⊆ T ⊆
e⊆R
b ⊆ S, where T ⊆ R
b is both integral and a flat epimorphism by
R
e=R
b by Scholium A (5) (L).
Scholium A (4). Therefore, T = R
Corollary 4.3. Let R ⊆ S be a quasi-Prüfer extension, and let T ∈
^
∩T
[R, S]. Then, T ∩ R ⊆ T R is almost-Prüfer. Moreover, T = R
TR
.
Proof. T ∩ R ⊆ T is quasi-Prüfer by Corollary 3.3. Being integrally
closed, it is Prüfer by Corollary 3.5. Moreover, T ⊆ T R is an integral
^
extension. Then, T ∩ R ⊆ T R is almost-Prüfer and T = R
∩T
TR
.
We note that integral extensions and Prüfer extensions are almostPrüfer and hence minimal extensions are almost-Prüfer. There are
quasi-Prüfer extensions that are not almost-Prüfer. It is enough to
consider [39, Example 3.5(1)]. Let R ⊆ T ⊆ S be two minimal extensions, where R is local, R ⊆ T integral and T ⊆ S is Prüfer. Then
b and
R ⊆ S is quasi-Prüfer but not almost-Prüfer, because S = R
e The same example shows that a composite of almost-Prüfer
R = R.
extensions may not be almost-Prüfer.
But the reverse implication holds.
16
G. PICAVET AND M. PICAVET
Theorem 4.4. Let R ⊆ S be an almost-Prüfer extension. Then R ⊆ S
e = R,
b (R)
e P =R
fP for each P ∈ Spec(R).
is quasi-Prüfer. Moreover, R
In this case, any flat epimorphic subextension R ⊆ T is Prüfer.
e ⊆ S, be an almost-Prüfer extension, that is R
e ⊆ S is
Proof. Let R ⊆ R
e is Prüfer. Now the Morita
integral. The result follows because R ⊆ R
hull and the Prüfer hull coincide by Proposition 4.2. In the same way,
e P →R
fP is a flat epimorphism and (R)
e P → SP is integral.
(R)
We could define almost-Prüfer rings as the rings R such that R ⊆
e = Tot(R) (by TheoTot(R) is almost-Prüfer. But in that case R
rem 4.4), so that R is a Prüfer ring. The converse evidently holds.
Therefore, this concept does not define something new.
We observed in [10, Remark 2.9(c)] that there is an almost-Prüfer
FMC extension R ⊆ S ⊆ T , where R ⊆ S is a Prüfer minimal extension and S ⊆ T is minimal and integral. But R ⊆ T is not an FCP
extension.
Proposition 4.5. Let R ⊆ S be an extension verifying the hypotheses:
(i) R ⊆ S is quasi-Prüfer.
(ii) R ⊆ S can be factored R ⊆ T ⊆ S, where R ⊆ T is a flat
epimorphism.
(1) Then the following commutative diagram (D) is a pushout,
R −−−→
y
R
y
T −−−→ T R
T R ⊆ S is Prüfer and R ⊆ T R is quasi-Prüfer. Moreover,
FR,R (P ) ∼
= FT,T R (Q) for each Q ∈ Spec(T ) and P := Q ∩ R.
(2) If in addition R ⊆ T is integrally closed, (D) is a pullback,
T ∩ R = R, (R : R) = (T : T R) ∩ R and (T : T R) = (R : R)T .
Proof. (1) Consider the injective composite map R → R ⊗R T → T R.
As R → R ⊗R T is a flat epimorphism, because deduced by a base
change of R → T , we get that the surjective map R ⊗R T → T R
is an isomorphism by Scholium A (3). By fibers transitivity, we have
FT,RT (Q) ∼
= κ(Q)⊗k(P ) FR,R (P ) [22, Corollaire 3.4.9]. As κ(P ) → κ(Q)
is an isomorphism by Scholium A, we get that FR,R (P ) ∼
= FT,RT (Q).
(2) As in [5, Lemma 3.5], R = T ∩ R. The first statement on the
conductors has the same proof as in [5, Lemma 3.5]. The second holds
because R ⊆ T is a flat epimorphism (see Scholium A (6)).
QUASI-PRÜFER EXTENSIONS
17
Theorem 4.6. Let R ⊂ S be a quasi-Prüfer extension and the diagram
(D’):
R −−−→ R
y
y
(1)
(2)
(3)
(4)
(5)
e −−−→ RR
e
R
e = R and
(D’) is a pushout and a pullback, such that R ∩ R
e : RR)
e ∩ R so that (R
e : RR)
e = (R : R)R.
e
(R : R) = (R
=⇒
e = S, where
e=R
~ ⊆ R =R
e =R
R ⊂ S can be factored R ⊆ RR
the first extension is almost-Prüfer and the second is Prüfer.
e = R.
e
e⇔R
R ⊂ S is almost-Prüfer if and only if S = RR
e =R
e=R
~ is the greatest almost-Prüfer subextension of
R ⊆ RR
e=R
eR~ .
R ⊆ S and R
e
Supp(S/R) = Supp(R/R)
∪ Supp(R/R) if R ⊆ S is almostPrüfer. (Supp can be replaced with MSupp).
Proof. To show (1), (2), in view of Theorem 3.18, it is enough to apply
=⇒
e and S = R , because R ⊆ RR
e is almostProposition 4.5 with T = R
Prüfer whence quasi-Prüfer, keeping in mind that a Prüfer extension is
integrally closed, whereas an integral Prüfer extension is trivial. Moree = RR
e because RR
e⊆R
e is both integral and integrally closed.
over, R
(3) is obvious.
(4) Now consider an almost-Prüfer subextension R ⊆ T ⊆ U, where
R ⊆ T is Prüfer and T ⊆ U is integral. Applying (3), we see that
U U
e ⊆ RR
e in view of Proposition 1.6.
U =R R
e
(5) Obviously, Supp(R/R)
∪ Supp(R/R) ⊆ Supp(S/R). Conversely,
e M = RM .
let M ∈ Spec(R) be such that RM 6= SM , and RM = (R)
e M = RM , which is absurd.
Then (3) entails that SM = (R)M (R)
Corollary 4.7. Let R ⊆ S be an almost-Prüfer extension. The following conditions are equivalent:
(1) Supp(S/R) ∩ Supp(R/R) = ∅.
e ∩ Supp(R/R)
e
(2) Supp(S/R)
= ∅.
e
(3) Supp(R/R) ∩ Supp(R/R) = ∅.
eP = R
fP for each
Proof. Since R ⊆ S is almost-Prüfer, we get (R)
e
P ∈ Spec(R). Moreover, Supp(S/R) = Supp(R/R) ∪ Supp(R/R) =
e ∪ Supp(R/R).
e
Supp(S/R) ∪ Supp(R/R) = Supp(S/R)
e ∩ Supp(R/R).
e
(1) ⇒ (2): Assume that there exists P ∈ Supp(S/R)
e P 6= SP , RP , so that RP ⊂ SP is neither Prüfer, nor integral.
Then, (R)
18
G. PICAVET AND M. PICAVET
But, P ∈ Supp(S/R) = Supp(S/R) ∪ Supp(R/R). If P ∈ Supp(S/R),
then P 6∈ Supp(R/R), so that (R)P = RP and RP ⊂ SP is Prüfer,
a contradiction. If P ∈ Supp(R/R), then P 6∈ Supp(S/R), so that
(R)P = SP and RP ⊂ SP is integral, a contradiction.
e
(2) ⇒ (3): Assume that there exists P ∈ Supp(R/R)
∩ Supp(R/R).
e
Then, RP 6= (R)P , (R)P , so that RP ⊂ SP is neither Prüfer, nor ine ∪ Supp(R/R).
e
tegral. But, P ∈ Supp(S/R) = Supp(S/R)
If P ∈
e
e
e
Supp(S/R), then P 6∈ Supp(R/R), so that (R)P = RP and RP ⊂ SP
e
e
is integral, a contradiction. If P ∈ Supp(R/R),
then P 6∈ Supp(S/R),
e P = SP and RP ⊂ SP is Prüfer, a contradiction.
so that (R)
(3) ⇒ (1): Assume that there exists P ∈ Supp(S/R) ∩ Supp(R/R).
Then, (R)P 6= RP , SP , so that RP ⊂ SP is neither Prüfer, nor integral.
e
e
If P ∈ Supp(R/R),
But, P ∈ Supp(S/R) = Supp(R/R) ∪ Supp(R/R).
then P 6∈ Supp(R/R), so that (R)P = RP and RP ⊂ SP is Prüfer,
e
a contradiction. If P ∈ Supp(R/R), then P 6∈ Supp(R/R),
so that
e
(R)P = RP and RP ⊂ SP is integral, a contradiction.
Proposition 4.5 has the following similar statement proved by Ayache
and Dobbs. It reduces to Theorem 4.6 in case R ⊆ S has FCP because
of Proposition 1.3.
Proposition 4.8. Let R ⊆ T ⊆ S be a quasi-Prüfer extension, where
T ⊆ S is an integral minimal extension and R ⊆ T is integrally closed .
Then the diagram (D) is a pullback, S = T R and (T : S) = (R : R)T .
Proof. [5, Lemma 3.5].
Proposition 4.9. Let R ⊆ U ⊆ S and R ⊆ V ⊆ S be two towers
of extensions, such that R ⊆ U and R ⊆ V are almost-Prüfer. Then
g=U
e Ve .
R ⊆ UV is almost-Prüfer and UV
Proof. Denote by U ′ , V ′ and W ′ the Prüfer hulls of R in U, V and
W = UV . We deduce from [26, Corollary 5.11, p.53], that R ⊆ U ′ V ′
is Prüfer. Moreover, U ′ V ′ ⊆ UV is clearly integral and U ′ V ′ ⊆ W ′
because the Prüfer hull is the greatest Prüfer subextension. We deduce
g=U
e Ve .
that R ⊆ UV is almost-Prüfer and that UV
Proposition 4.10. Let R ⊆ U ⊆ S and R ⊆ V ⊆ S be two towers
of extensions, such that R ⊆ U is almost-Prüfer and R ⊆ V is a flat
epimorphism. Then U ⊆ UV is almost-Prüfer.
Proof. Mimic the proof of Proposition 4.9 and use [26, Theorem 5.10,
p.53].
Proposition 4.11. Let R ⊆ S be an almost-Prüfer extension and
R → T a flat epimorphism. Then T ⊆ T ⊗R S is almost-Prüfer.
QUASI-PRÜFER EXTENSIONS
Proof. It is enough to use Proposition 3.10 and Definition 4.1.
19
Proposition 4.12. An extension R ⊆ S is almost-Prüfer if and only
fP = (R)
e P for each P ∈ Spec(R).
if RP ⊆ SP is almost-Prüfer and R
e P ⊆R
fP . Suppose
Proof. For an arbitrary extension R ⊆ S we have (R)
eP =R
fP by
that R ⊆ S is almost-Prüfer, then so is RP ⊆ SP and (R)
Theorem 4.4. Conversely, if R ⊆ S is locally almost-Prüfer, whence
fP = (R)
e P holds
locally quasi-Prüfer, then R ⊆ S is quasi-Prüfer. If R
e P so that S = RR
e and
for each P ∈ Spec(R), we have SP = (RR)
R ⊆ S is almost-Prüfer by Theorem 4.6.
Corollary 4.13. An FCP extension R ⊆ S is almost-Prüfer if and
only if RP ⊆ SP is almost-Prüfer for each P ∈ Spec(R).
Proof. It is enough to show that R ⊆ S is almost-Prüfer if RP ⊆ SP
is almost-Prüfer for each P ∈ Spec(R) using Proposition 4.12. Any
e ⊂ R1 is integral by definition of R.
e Assume that
minimal extension R
′
′
]
e P ⊂ (R
e
e
(R)
P ), so that there exists R2 ∈ [R, S] such that (R)P ⊂ (R2 )P
e P , for
is a Prüfer minimal extension with crucial maximal ideal Q(R)
e with Q ∩ R ⊆ P . In particular, R
e ⊂ R′ is not
some Q ∈ Max(R)
2
e R′ ] such that
integral. We may assume that there exists R1′ ∈ [R,
2
e Using
R1′ ⊂ R2′ is a Prüfer minimal extension with P 6∈ Supp(R1′ /R).
e R′ ] such that R
e ⊂ R2 is a
[39, Lemma 1.10], there exists R2 ∈ [R,
2
Prüfer minimal extension with crucial maximal ideal Q, a contradic]
e P ⊂ SP is integral for each P , whence (R)
e P = (R
tion. Then, (R)
P ).
We now intend to demonstrate that our methods allow us to prove
easily some results. For instance, next statement generalizes [5, Corollary 4.5] and can be fruitful in algebraic number theory.
Proposition 4.14. Let (R, M) be a one-dimensional local ring and
R ⊆ S a quasi-Prüfer extension. Suppose that there is a tower R ⊂
T ⊆ S, where R ⊂ T is integrally closed. Then R ⊆ S is almost-Prüfer,
e and S is zero-dimensional.
T =R
Proof. Because R ⊂ T is quasi-Prüfer and integrally closed, it is Prüfer.
If some prime ideal of T is lying over M, R ⊂ T is a faithfully flat
epimorphism, whence an isomorphism by Scholium A, which is absurd.
Now let N be a prime ideal of T and P := N ∩ R. Then RP is zerodimensional and isomorphic to TN . Therefore, T is zero-dimensional.
It follows that T R is zero-dimensional. Since RT ⊆ S is Prüfer, we
deduce from Scholium A, that RT = S. The proof is now complete.
20
G. PICAVET AND M. PICAVET
We also generalize [5, Proposition 5.2] as follows.
Proposition 4.15. Let R ⊂ S be a quasi-Prüfer
extension, such that
q
R is local with maximal ideal N := (R : R). Then R is local and
[R, S] = [R, R] ∪ [R, S]. If in addition R is one-dimensional, then
either R ⊂ S is integral or there is some minimal prime ideal P of R,
such that S = (R)P , P = SP and R/P is a one-dimensional valuation
domain with quotient field S/P .
Proof. R is obviously local. Let T ∈ [R, S] \ [R, R] and s ∈ T \ R.
Then s ∈ U(S) and s−1 ∈ R by Proposition 1.2 (1). But s−1 6∈ U(R),
so that s−1 ∈ N. It follows that there exists some integer n such that
s−n ∈ (R : R), giving s−n R ⊆ R, or, equivalently, R ⊆ Rsn ⊆ T .
Then, T ∈ [R, S] and we obtain [R, S] = [R, R] ∪ [R, S].
Assume that R is one-dimensional. If R ⊂ S is not integral then
R ⊂ S is Prüfer and R is one-dimensional. To complete the proof, use
Proposition 1.2 (3).
4.2. FCP extensions. In case we consider only FCP extensions, we
obtain more results.
Proposition 4.16. Let R ⊆ S be an FCP extension. The following
statements are equivalent:
(1) R ⊆ S is almost-Prüfer.
(2) RP ⊆ SP is either integral or Prüfer for each P ∈ Spec(R).
e ∩ Supp(R/R)
e
(3) RP ⊆ SP is almost-Prüfer and Supp(S/R)
= ∅.
(4) Supp(R/R) ∩ Supp(S/R) = ∅.
Proof. The equivalence of Proposition 4.12 shows that (2) ⇔ (1) holds
because Tb = Te and over a local ring T , an almost-Prüfer FCP extension
T ⊆ U is either integral or Prüfer [39, Proposition 2.4] . Moreover when
e P =R
fP
RP ⊆ SP is either integral or Prüfer, it is easy to show that (R)
Next we show that (3) is equivalent to (2) of Proposition 4.12.
e ∩ Supp(R/R)
e
Let P ∈ Supp(S/R)
be such that RP ⊆ SP is almoste
e P ⊂ SP . Since R ⊂ R
e
Prüfer. Then, (R)P 6= RP , SP , so that RP ⊂ (R)
e P , giving (R)
e P ⊆R
fP and RP 6= R
fP . It follows
is Prüfer, so is RP ⊂ (R)
fP = SP in view of the dichotomy principle [39, Proposition 3.3]
that R
fP 6= (R)
e P.
since RP is a local ring, and then R
fP 6= (R)
e P , i.e. P ∈ Supp(S/R). Then,
Conversely, assume that R
fP , so that R
fP = SP , as we have just seen. Hence RP ⊂ SP
RP 6= R
is integrally closed. It follows that RP = RP = RP , so that P 6∈
e
eP 6=
by Theorem 4.6(5). Moreover, R
Supp(R/R) and P ∈ Supp(R/R)
QUASI-PRÜFER EXTENSIONS
21
e
e ∩
SP implies that P ∈ Supp(S/R).
To conclude, P ∈ Supp(S/R)
e
Supp(R/R).
(1) ⇔ (4) An FCP extension is quasi-Prüfer by Corollary 3.4. Supe we
pose that R ⊆ S is almost-Prüfer. By Theorem 4.6, letting U := R,
get that U ∩R = R and S = RU. We deduce from [39, Proposition 3.6]
that Supp(R/R) ∩ Supp(S/R) = ∅. Suppose that this last condition
holds. Then by [39, Proposition 3.6] R ⊆ S can be factored R ⊆ U ⊆ S,
where R ⊆ U is integrally closed, whence Prüfer by Proposition 1.3,
and U ⊆ S is integral. Therefore, R ⊆ S is almost-Prüfer.
Lemma 4.17. Let B ⊂ D and C ⊂ D be two integral minimal extensions and A := B ∩ C. If A ⊂ D has FCP, then, A ⊂ D is integral.
Proof. Set M := (B : D) and N := (C : D).
If M 6= N, then, A ⊂ D is integral by [13, Proposition 6.6].
Assume that M = N. Then, M ∈ Max(A) by [13, Proposition 5.7].
Let B ′ be the integral closure of A in B. Then M is also an ideal of
B ′ , which is prime in B ′ , and then maximal in B ′ . If A ⊂ D is an
FCP extension, so is B ′ ⊆ B, which is a flat epimorphism, and so is
B ′ /M ⊆ B/M. Then, B ′ = B since B ′ /M is a field. It follows that
A ⊆ B is an integral extension, and so is A ⊂ D.
Proposition 4.18. Let R ⊂ S be an FCP almost-Prüfer extension.
e=R
b is the least T ∈ [R, S] such that T ⊆ S is integral.
Then, R
Proof. We may assume that R ⊂ S is not integral. If there is some
e such that U ⊆ R
e is integral, then U = R.
e Set X := {T ∈
U ∈ [R, R]
e is a minimal element of X.
[R, S] | T ⊆ S integral}. It follows that R
e is the least element of X.
We are going to show that R
e S] ≥ 1 and let R
e = R0 ⊂ R1 ⊂ · · · ⊂ Rn−1 ⊂ Rn = S
Set n := ℓ[R,
e S], with length n. There does not exist a
be a maximal chain of [R,
e
maximal chain of R-subalgebras
of S with length > n. Let T ∈ X. We
e
intend to show that T ∈ [R, S]. It is enough to choose T such that T
is a minimal element of X. Consider the induction hypothesis: (Hn ):
e S] when n := ℓ[R,
e S].
X ⊆ [R,
e ⊂ S is minimal. Let T ∈ X and
We first show (H1 ). If n = 1, R
e
T1 ∈ [T, S] be such that T1 ⊂ S is minimal. Assume that T1 6= R.
e⊂R
e is integral, which contradicts the
Lemma 4.17 shows that T1 ∩ R
e so that T = R
e for the same
beginning of the proof. Then, T1 = R,
contradiction and (H1 ) is proved.
Assume that n > 1 and that (Hk ) holds for any k < n. Let T ∈ X
e S], then
and T1 ∈ [T, S] be such that T1 ⊂ S is minimal. If T1 ∈ [R,
22
G. PICAVET AND M. PICAVET
e T1 ] ≤ n−1. But we get that T ∈ [R, T1 ], with T ⊆ T1 integral.
k := ℓ[R,
e is also the Prüfer hull of R ⊆ T1 , with k := ℓ[R,
e T1 ] ≤ n−1.
Moreover, R
e T1 ] ⊂ [R,
e S].
Since (Hk ) holds, we get that T ∈ [R,
e
If T1 6∈ [R, S], set U := T1 ∩Rn−1 . We get that T1 ⊂ S and Rn−1 ⊂ S
are minimal and integral. Using again Lemma 4.17, we get that U ⊂ S
e Rn−1 ] = n − 1 and U ∈ [R, Rn−1 ]. As before, R
e is
is integral, with ℓ[R,
e Rn−1 ],
also the Prüfer hull of R ⊆ Rn−1 . Since (Hn−1 ) holds, U ∈ [R,
e
so that T1 ∈ [R, S], a contradiction. Therefore, (Hn ) is proved.
We will need a relative version of the support. Let f : R → T be a
ring morphism and E a T -module. The relative support of E over R is
SR (E) := a f (SuppT (E)) and MSR (E) := SR (E)∩Max(R). In particular, for a ring extension R ⊂ S, we have SR (S/R) := SuppR (S/R)).
Proposition 4.19. Let R ⊆ S be an FCP extension. The following
statements hold:
e
(1) Supp(R/R)
∩ Supp(R/R) = ∅.
e
e R)
e ∩ Supp(R/R)
e
(2) Supp(R/R)
∩ Supp(R/R) = Supp(R/
= ∅.
e
(3) MSupp(S/R) = MSupp(R/R) ∪ MSupp(R/R).
e is
Proof. (1) is a consequence of Proposition 4.16(4) because R ⊆ R
almost-Prüfer.
e
We prove the first part of (2). If some M ∈ Supp(R/R)∩Supp(R/R),
′
e M , T := (R)M
it can be supposed in Max(R). Set R := RM , U := (R)
and M ′ := MRM . Then, R′ 6= U, T , with R′ ⊂ U FCP Prüfer and
R′ ⊂ T FCP integral, an absurdity [39, Proposition 3.3].
e R)
e ∩
To show the second part, assume that some P ∈ Supp(R/
e
Supp(R/R).
Then, P 6∈ Supp(R/R) by the first part of (2), so that
e P = RP R
eP = R
eP , a contradiction.
RP = RP , giving (R)
S
S
(3) Obviously, MSupp(S/R) = MS (S/R) = MS (S/T )∪MS (T /T )
T
T
∪MS (T /U )∪MS (U /U)∪MS (U/R). By [39, Propositions 2.3 and
S
S
T
3.2], we have MS (S/T ) ⊆ S (T /T ) = S (R/R ) = MS (R/R) =
T
T
MSupp(R/R), MS (T /U ) = S (R /R) ⊆ S (R/R) = Supp(R/R)
T
T
and MS (U /U) = S (R /R) = Supp(R/R). To conclude, MSupp(S/R) =
e
MSupp(R/R)
∪ MSupp(R/R).
Proposition 4.20. Let R ⊂ S be an FCP extension and M ∈ MSupp(S/R),
g
e
e
e
then R
M = (R)M if and only if M 6∈ MSupp(S/R) ∩ MSupp(R/R).
g
e
Proof. In fact, we are going to show that R
M 6= (R)M if and only if
e ∩ MSupp(R/R).
e
M ∈ MSupp(S/R)
QUASI-PRÜFER EXTENSIONS
23
e ∩ MSupp(R/R).
e
g
Let M ∈ MSupp(S/R)
Then, R
M 6= RM , SM and
g
e
g
then RM ⊂ R
⊂
S
.
Since
R
⊂
R
is
Prüfer,
so is RM ⊂ R
M
M
M
e M ⊆ R
g
g
by Proposition 1.2, giving (R)
and
R
=
6
R
.
Therefore,
M
M
M
g
g
e
R
=
S
[39,
Proposition
3.3]
since
R
is
local,
and
then
R
M
M
M
M 6= (R)M .
g
e
g
g
Conversely, if R
M 6= (R)M , then, RM 6= RM , so that RM = SM , as
we have just seen and then RM ⊂ SM is integrally closed. It follows
that RM = RM = RM , so that M 6∈ MSupp(R/R). Hence, M ∈
e
eM 6= SM ⇒ M ∈
MSupp(R/R)
by Proposition 4.19(3). Moreover, R
e To conclude, M ∈ MSupp(S/R)
e ∩ MSupp(R/R).
e
MSupp(S/R).
g
e
If R ⊆ S is any ring extension, with dim(R) = 0, then R
M = (R)M
for any M ∈ Max(R). Indeed by Scholium A (2), the flat epimorphism
e is bijective as well as RM → (R)
e M . This conclusion is still
R → R
valid in another context.
Corollary 4.21. Let R ⊂ S be an FCP extension. Assume that one
of the following conditions is satisfied:
e ∩ MSupp(R/R)
e
(1) MSupp(S/R)
= ∅.
e
(2) S = RR, or equivalently, R ⊆ S is almost-Prüfer.
g
e
Then, R
M = (R)M for any M ∈ Max(R).
Proof. (1) is Proposition 4.20. (2) is Proposition 4.12.
Proposition 4.22. Let R ⊂ S be an almost-Prüfer FCP extension.
e in T R.
e
Then, any T ∈ [R, S] is the integral closure of T ∩ R
e and V := T R.
e Since R ⊂ S is almost-Prüfer,
Proof. Set U := T ∩ R
e
e
e is also the Prüfer hull of
U ⊆ R is Prüfer and R ⊆ V is integral and R
U ⊆ V . Because R ⊂ S is almost-Prüfer, for each M ∈ MSuppR (S/R),
RM ⊆ SM is either integral, or Prüfer by Proposition 4.16, and so is
g
e
UM ⊆ VM . But R
M = (R)M by Corollary 4.21 is also the Prüfer hull
′
′
of UM ⊆ VM . Let T be the integral closure of U in V . Then, TM
is
the integral closure of UM in VM .
′
e M,
Assume that UM ⊆ VM is integral. Then VM = TM
and UM = (R)
e M = TM , giving TM = T ′ .
so that VM = TM (R)
M
′
e M,
Assume that UM ⊆ VM is Prüfer. Then UM = TM
and VM = (R)
′
e M = TM , giving TM = T .
so that UM = TM ∩ (R)
M
′
To conclude, we get that TM = TM
for each M ∈ MSuppR (S/R).
′
Since RM = SM , with TM = TM
for each M ∈ Max(R)\MSuppR (S/R),
′
we get T = T , whence T is the integral closure of U ⊆ V .
24
G. PICAVET AND M. PICAVET
We build an example of an FCP extension R ⊂ S where we have
g
e M for some M ∈ Max(R). In particular, R ⊂ S is not
RM 6= (R)
almost-Prüfer.
Example 4.23. Let R be an integral domain with quotient field S
and Spec(R) := {M1 , M2 , P, 0}, where M1 6= M2 are two maximal
ideals and P a prime ideal satisfying P ⊂ M1 ∩ M2 . Assume that
there are R1 , R2 and R3 such that R ⊂ R1 is Prüfer minimal, with
C (R, R1 ) = M1 , R ⊂ R2 is integral minimal, with C (R, R2 ) = M2
and R2 ⊂ R3 is Prüfer minimal, with C (R2 , R3 ) = M3 ∈ Max(R2 )
such that M3 ∩ R = M2 and M2 R3 = R3 . This last condition is
satisfied when R ⊂ R2 is either ramified or inert. Indeed, in both
cases, M3 R3 = R3 ; moreover, in the ramified case, we have M32 ⊆ M2
and in the inert case, M3 = M2 [36, Theorem 3.3]. We apply [13,
Proposition 7.10] and [10, Lemma 2.4] several times. Set R2′ := R1 R2 .
Then, R1 ⊂ R2′ is integral minimal, with C (R1 , R2′ ) =: M2′ = M2 R1
and R2 ⊂ R2′ is Prüfer minimal, with C (R2 , R2′ ) =: M1′ = M1 R2 ∈
Max(R2 ). Moreover, M1′ 6= M3 , Spec(R1 ) = {M2′ , P1 , 0}, where P1
is the only prime ideal of R1 lying over P . But, P = (R : R1 ) by
[17, Proposition 3.3], so that P = P1 . Set R3′ := R3 R2′ . Then, R2′ ⊂
R3′ is Prüfer minimal, with C (R2′ , R3′ ) =: M3′ = M3 R2′ ∈ Max(R2′ )
and R3 ⊂ R3′ is Prüfer minimal, with C (R3 , R3′ ) = M1′′ = M1 R3 ∈
Max(R3 ). It follows that we have Spec(R3′ ) = {P ′ , 0} where P ′ is the
only prime ideal of R3′ lying over P . To end, assume that R3′ ⊂ S is
Prüfer minimal, with C (R3′ , S) = P ′. Hence, R2 is the integral closure
of R in S. In particular, R ⊂ S has FCP [10, Theorems 6.3 and
3.13] and is quasi-Prüfer. Since R ⊂ R1 is integrally closed, we have
e Assume that R1 6= R.
e Then, there exists T ∈ [R1 , S] such
R1 ⊆ R.
that R1 ⊂ T is Prüfer minimal and C (R1 , T ) = M2′ , a contradiction
by Proposition 4.16 since M2′ = C (R1 , R2′ ), with R1 ⊂ R2′ integral
e It follows that M1 ∈ MSupp(R/R).
e
minimal. Then, R1 = R.
But,
′
e
e
P = C (R3 , S) ∩ R ∈ Supp(S/R) and P ⊂ M1 give M1 ∈ MSupp(S/R),
g
e
so that R
M1 6= (R)M1 by Proposition 4.20 giving that R ⊂ S is not
almost-Prüfer.
We now intend to refine Theorem 4.6, following the scheme used in
[4, Proposition 4] for extensions of integral domains.
Proposition 4.24. Let R ⊆ S and U, T ∈ [R, S] be such that R ⊆ U is
integral and R ⊆ T is Prüfer. Then U ⊆ UT is Prüfer in the following
cases and R ⊆ UT is almost-Prüfer.
e
(1) Supp(R/R)∩Supp(R/R)
= ∅ (for example, if R ⊆ S has FCP).
(2) R ⊆ U preserves integral closure.
QUASI-PRÜFER EXTENSIONS
25
Proof. (1) We have ∅ = MSupp(U/R)∩MSupp(T /R), since U ⊆ R and
e Let M ∈ MSupp((UT )/R). For M ∈ MSupp(U/R), we have
T ⊆ R.
RM = TM and (UT )M = UM . If M ∈
/ MSupp(U/R), then UM = RM
and (UT )M = TM , so that UM ⊆ (UT )M identifies to RM ⊆ TM .
Let N ∈ Max(U) and set M := N ∩ R ∈ Max(R) since R ⊆ U is
integral. If M 6∈ Supp(R/R), then RM = RM = UM and N is the
only maximal ideal of U lying over M. It follows that UM = UN and
(UT )M = (UT )N by [10, Lemma 2.4]. Then, UN ⊆ (UT )N identifies
e
to RM ⊆ TM which is Prüfer. If M 6∈ Supp(R/R),
then RM = TM
gives UM = (UT )M , so that UN = (UT )N by localizing the precedent
equality and UN ⊆ (UT )N is still Prüfer. Therefore, U ⊆ UT is locally
Prüfer, whence Prüfer by Proposition 1.1.
(2) The usual reasoning shows that U ⊗R T ∼
= UT , so that U ⊆ UT
UT
UT
is integrally closed. Since U is contained in R , we get that U = R .
Now observe that R ⊆ UT is almost-Prüfer, whence quasi-Prüfer. It
follows that U ⊆ UT is Prüfer.
Next propositions generalize Ayache’s results of [4, Proposition 11].
Proposition 4.25. Let R ⊆ S be a quasi-Prüfer extension, T, T ′ ∈
[R, S] and U := T ∩ T ′ . The following statements hold:
Te = (T^
∩ R) for each T ∈ [R, S].
′
e
e
T ∩ T ⊆ T^
∩ T ′.
Let Supp(T /T ) ∩ Supp(Te/T ) = ∅ (this assumption holds if R ⊆
S has FCP). Then, T ⊆ T ′ ⇒ Te ⊆ Te′ .
e
(4) If Supp(U /U) ∩ Supp(U/U)
= ∅, then Te ∩ Te′ = T^
∩ T ′.
(1)
(2)
(3)
Proof. (1) We observe that R ⊆ T is quasi-Prüfer by Corollary 3.3.
Since T ∩ R is the integral closure of R in T , we get that T ∩ R ⊆ T is
∩ R.
Prüfer. It follows that T ∩ R ⊆ Te is Prüfer. We thus have Te ⊆ T^
e
To prove the reverse inclusion, we set V := T ∩ R and W := V ∩ T .
We have W ∩ R = Ve ∩ R = V , because V ⊆ Ve ∩ R is integral and
Prüfer since we have a tower V ⊆ Ve ∩ R ⊆ Ve . Therefore, V ⊆ W
is Prüfer because W ∈ [V, Ve ]. Moreover, T ⊆ Te ⊆ Ve , since V ⊆ Te
is Prüfer. Then, T ⊆ W is integral because W ∈ [T, T ], and we have
V ⊆ T ⊆ W . This entails that T = W = Ve ∩ T , so that T ⊆ Ve is
Prüfer. It follows that Ve ⊆ Te since T ∈ [V, Ve ].
(2) A quasi-Prüfer extension is Prüfer if and only if it is integrally
closed. We observe that T ∩ T ′ ⊆ Te ∩ Te′ is integrally closed, whence
Prüfer. It follows that Te ∩ Te′ ⊆ T^
∩ T ′.
26
G. PICAVET AND M. PICAVET
(3) Set U = T ∩ R and U ′ = T ′ ∩ R, so that U, U ′ ∈ [R, R] with
U ⊆ U ′ . In view of (1), we thus can suppose that T, T ′ ∈ [R, R]. It
follows that T ⊆ T ′ is integral and T ⊆ Te is Prüfer. We deduce from
Proposition 4.24(1) that T ′ ⊆ T ′ Te is Prüfer, so that TeT ′ ⊆ Te′ , because
Supp(T /T ) ∩ Supp(Te/T ) = ∅ and T = R. Therefore, we have Te ⊆ Te′ .
e /U) = ∅. Then, T ∩T ′ ⊂ T, T ′
(4) Assume that Supp(U /U)∩Supp(U
gives T^
∩ T ′ ⊆ Te ∩ Te′ in view of (3), so that T^
∩ T ′ = Te ∩ Te′ by (2).
Proposition 4.26. Let R ⊆ S be a quasi-Prüfer extension and T ⊆ T ′
a subextension of R ⊆ S. Set U := T ∩ R, U ′ := T ′ ∩ R, V := T R and
V ′ := T ′ R. The following statements hold:
(1) T ⊆ T ′ is integral if and only if V = V ′ .
(2) T ⊆ T ′ is Prüfer if and only if U = U ′ .
(3) Assume that U ⊂ U ′ is integral minimal and V = V ′ . Then,
T ⊂ T ′ is integral minimal, of the same type as U ⊂ U ′ .
(4) Assume that V ⊂ V ′ is Prüfer minimal and U = U ′ . Then,
T ⊂ T ′ is Prüfer minimal.
(5) Assume that T ⊂ T ′ is minimal and set P := C(T, T ′ ).
(a) If T ⊂ T ′ is integral, then U ⊂ U ′ is integral minimal if and
only if P ∩ U ∈ Max(U).
(b) If T ⊂ T ′ is Prüfer, then V ⊂ V ′ is Prüfer minimal if and
only if there is exactly one prime ideal in V lying over P .
Proof. In [R, S] we have the integral extensions U ⊆ U ′ , T ⊆ V, T ′ ⊆
V ′ and the Prüfer extensions V ⊆ V ′ , U ⊆ T, U ′ ⊆ T ′ . Moreover, R
is also the integral closure of U ⊆ V ′ .
(1) is gotten by considering the extension T ⊆ V ′ , which is both
T ⊆ V ⊆ V ′ and T ⊆ T ′ ⊆ V ′ .
(2) is gotten by considering the extension U ⊆ T ′ , which is both
U ⊆ T ⊆ T ′ and U ⊆ U ′ ⊆ T ′ .
(3) Assume that U ⊂ U ′ is integral minimal and V = V ′ . Then,
T ⊂ T ′ is integral by (1) and T 6= T ′ because of (2). Set M := (U :
U ′ ) ∈ SuppU (U ′ /U). For any M ′ ∈ Max(U) such that M ′ 6= M, we
′
′
′
have UM ′ = UM
′ , so that TM ′ = T M ′ because UM ′ ⊆ TM ′ is Prüfer.
′
′
′
But, U ⊆ T is almost-Prüfer, giving T = T U . By Theorem 4.6,
(T : T ′ ) = (U : U ′ )T = MT 6= T because T 6= T ′ . We get that U ⊆ T
Prüfer implies that M 6∈ SuppU (T /U) and UM = TM . It follows that
T ′ M = TM U ′ M = U ′ M . Therefore, TM ⊆ T ′ M identifies to UM ⊆ U ′ M ,
which is minimal of the same type as U ⊂ U ′ by [13, Proposition 4.6].
Then, T ⊂ T ′ is integral minimal, of the same type as U ⊂ U ′ .
(4) Assume that V ⊂ V ′ is Prüfer minimal and U = U ′ . Then,
T ⊂ T ′ is Prüfer by (2) and T 6= T ′ because of (1). Set Q := C(V, V ′ )
QUASI-PRÜFER EXTENSIONS
27
and P := Q ∩ T ∈ Max(T ) since Q ∈ Max(V ). For any P ′ ∈ Max(T )
such that P ′ 6= P , and Q′ ∈ Max(V ) lying above P ′, we have VQ′ = VQ′ ′ ,
so that VP ′ = V ′ P ′ . It follows that T ′ P ′ ⊆ V ′ P ′ is integral, so that
TP ′ = T ′ P ′ and P ′ 6∈ SuppT (T ′/T ). We get that T ⊂ T ′ is Prüfer
minimal in view of [10, Proposition 6.12].
(5) Assume that T ⊂ T ′ is a minimal extension and set P := C(T, T ′ ).
(a) Assume that T ⊂ T ′ is integral. Then, V = V ′ and U 6= U ′
by (1) and (2). We can use Proposition 4.5 getting that P = (U :
U ′ )T ∈ Max(T ) and Q := (U : U ′ ) = P ∩ U ∈ Spec(U). It follows that
Q 6∈ SuppU (T /U), so that UQ = TQ and UQ′ = TQ′ . Then, UQ ⊂ UQ′ is
integral minimal, with Q ∈ SuppU (U ′ /U).
If Q 6∈ Max(U), then U ⊂ U ′ is not minimal by the properties of the
crucial maximal ideal.
Assume that Q ∈ Max(U) and let M ∈ Max(U), with M 6= Q.
′
Then, UM = UM
because M + Q = U, so that U ⊂ U ′ is a minimal
extension and (a) is gotten.
(b) Assume that T ⊂ T ′ is Prüfer. Then, V 6= V ′ and U = U ′ by (1)
and (2). Moreover, P T ′ = T ′ gives P V ′ = V ′ . Let Q ∈ Max(V ) lying
over P . Then, QV ′ = V ′ gives that Q ∈ SuppV (V ′ /V ). Moreover, we
have V ′ = V T ′ . Let P ′ ∈ Max(T ), P ′ 6= P . Then, TP ′ = TP′ ′ gives
VP ′ = VP′ ′ . It follows that SuppT (V ′ /V ) = {P } and SuppV (V ′ /V ) =
{Q ∈ Max(V ) | Q ∩ T = P }. But, by [10, Proposition 6.12], V ⊂ V ′ is
Prüfer minimal if and only if |SuppV (V ′ /V )| = 1, and then if and only
if there is exactly one prime ideal in V lying over P .
Lemma 4.27. Let R ⊆ S be an FCP almost-Prüfer extension and
U ∈ [R, R], V ∈ [R, S]. Then U ⊆ V has FCP and is almost-Prüfer.
Proof. Obviously, U ⊆ V has FCP and R is the integral closure of U
in V . Proposition 4.16 entails that SuppR (R/R) ∩ SuppR (S/R) = ∅.
We claim that SuppU (R/U) ∩ SuppU (V /R) = ∅. Deny and let Q ∈
SuppU (R/U) ∩ SuppU (V /R). Then, RQ 6= UQ , VQ . If P := Q ∩ R. we
get that RP 6= UP , VP , giving RP 6= RP , SP , a contradiction. Another
use of Proposition 4.16 shows that U ⊆ V is almost-Prüfer.
Proposition 4.28. Let R ⊆ S be an FCP almost-Prüfer extension and
T ⊆ T ′ a subextension of R ⊆ S. Set U := T ∩ R and V ′ := T ′ R. Let
W be the Prüfer hull of U ⊆ V ′ . Then, W is also the Prüfer hull of
T ⊆ T ′ and T ⊆ T ′ is an FCP almost-Prüfer extension.
Proof. By Lemma 4.27, we get that U ⊆ V ′ is an FCP almost-Prüfer
extension. Let Te be the Prüfer hull of T ⊆ T ′ . Since U ⊆ T and T ⊆ Te
are Prüfer, so is U ⊆ Te and Te ⊆ V ′ gives that Te ⊆ W . Then, T ⊆ W
is Prüfer as a subextension of U ⊆ W .
28
G. PICAVET AND M. PICAVET
Moreover, in view of Proposition 4.18, W is the least U-subalgebra
of V ′ over which V ′ is integral. Since T ′ ⊆ V ′ is integral, we get that
W ⊆ T ′ , so that W ∈ [T, T ′ ], with W ⊆ T ′ integral as a subextension
of W ⊆ V ′ . It follows that W is also the Prüfer hull of T ⊆ T ′ and
T ⊆ T ′ is an FCP almost-Prüfer extension.
5. The case of Nagata extensions
In this section we transfer the quasi-Prüfer (and almost-Prüfer) properties to Nagata extensions.
Proposition 5.1. Let R ⊆ S be a Prüfer (and FCP) extension, then
R(X) ⊆ S(X) is a Prüfer (and FCP) extension.
Proof. We can suppose that (R, M) is local, in order to use Proposition 1.2(3). Then it is enough to know the following facts: V (X) is
a valuation domain if so is V ; R[X]P [X] ∼
= RP (X) where
= R(X)P (X) ∼
∼
P (X) = P R(X) and R(X)/P (X) = (R/P )(X) for P ∈ Spec(R). If in
addition R ⊆ S is FCP, it is enough to use [11, Theorem 3.9]: R ⊂ S
has FCP if and only if R(X) ⊂ S(X) has FCP.
Proposition 5.2. If R ⊆ S is quasi-Prüfer, then so is R(X) ⊆ S(X),
R(X) = R(X) ∼
= R ⊗R R(X) and S(X) ∼
= S ⊗R R(X).
Proof. It is enough to use proposition 5.1, because R(X) = R(X). The
third assertion results from [34, Proposition 4 and Proposition 7].
Proposition 5.3. If R ⊆ S is almost-Prüfer, then so is R(X) ⊆ S(X).
^ = R(X)
e
It follows that R(X)
for an almost-Prüfer extension R ⊆ S.
e is Prüfer and R
e ⊆ S is
Proof. If R ⊆ S is almost-Prüfer, then R ⊆ R
e
e
integral and then R(X) ⊆ R(X)
is Prüfer and R(X)
⊆ S(X) is integral,
^ = R(X).
e
whence R(X) ⊆ S(X) is almost-Prüfer with R(X)
e = R.
Lemma 5.4. Let R ⊂ S be an FCP ring extension such that R
^ = R(X).
Then, R(X)
^ there is some T ′ ∈ [R(X), R(X)]
^ such that
Proof. If R(X) 6= R(X),
′
′
R(X) ⊂ T is Prüfer minimal. Set C (R(X), T ) ∈ MSupp(S(X)/R(X))
=: M ′ . There is M ∈ MSupp(S/R) such that M ′ = MR(X) [11,
Lemma 3.3]. But, M ′ 6∈ MSupp(R(X)/R(X)) = MSupp(R(X)/R(X))
by Proposition 4.19(2), giving that M 6∈ MSupp(R/R) = S (R/R).
Then [39, Proposition 1.7(3)] entails that M ∈ S (S/R). By [39,
Proposition 1.7(4)], there are some T1 , T2 ∈ [R, S] with T1 ⊂ T2 Prüfer
minimal (an FCP extension is quasi-Prüfer), with M = C (T1 , T2 ) ∩ R.
QUASI-PRÜFER EXTENSIONS
29
We can choose for T1 ⊂ T2 the first minimal extension verifying the preceding property. Therefore, M 6∈ S (T1 /R), so that M 6∈ S (T1 /R) =
Supp(T1 /R). By [39, Lemma 1.10], we get that there exists T ∈ [R, T2 ]
such that R ⊂ T is Prüfer minimal, a contradiction.
^
e
Proposition 5.5. If R ⊂ S is an FCP extension, then, R(X)
= R(X).
e is Prüfer, R(X) ⊆ R(X)
e
Proof. Because R ⊆ R
is Prüfer by Corol^ Assume that R(X)
^ and set
e
e
lary 5.1. Then, R(X)
⊆ R(X).
6= R(X)
^
e so that T = Te, giving T
e
T := R,
(X) = T (X) = R(X)
by Lemma 5.4.
^
^ is a Prüfer extension, contradicting the definition
Hence T
(X) ⊂ R(X)
^
^
e
of T
(X). So, R(X)
= R(X).
Proposition 5.6. Let R ⊆ S be an almost-Prüfer FCP extension, then
\ = R(X).
^
b
R(X)
= R(X)
\ = R(X)
^ = R(X)
e
b
Proof. We have a tower R(X) ⊆ R(X)
= R(X),
where the first and the third equalities come from Theorem 4.4 and
the second from Proposition 5.5.
We end this section with a special result.
Proposition 5.7. Let R ⊆ S be an extension such that R(X) ⊆ S(X)
\
b
has FIP, then R(X)
= R(X).
Proof. The map [R, S] → [R(X), S(X)] defined by T 7→ T (X) =
\ = T (X) for
R(X) ⊗R T is bijective [12, Theorem 32], whence R(X)
\ is a flat epimorphism.
b
some T ∈ [R, S]. Moreover, R(X)
→ R(X)
b = T and the result follows.
Since R → R(X) is faithfully flat, R
6. Fibers of quasi-Prüfer extensions
We intend to complete some results of Ayache-Dobbs [5]. We begin
by recalling some features about quasi-finite ring morphisms. A ring
morphism R → S is called quasi-finite by [40] if it is of finite type
and κ(P ) → κ(P ) ⊗R S is finite (as a κ(P )-vector space), for each
P ∈ Spec(R) [40, Proposition 3, p.40].
Proposition 6.1. A ring morphism of finite type is incomparable if
and only if it is quasi-finite and, if and only if its fibers are finite.
Proof. Use [41, Corollary 1.8] and the above definition.
30
G. PICAVET AND M. PICAVET
Theorem 6.2. An extension R ⊆ S is quasi-Prüfer if and only if
R ⊆ T is quasi-finite (respectively, has finite fibers) for each T ∈ [R, S]
such that T is of finite type over R, if and only if R ⊆ T has integral
fiber morphisms for each T ∈ [R, S].
Proof. It is clear that R ⊆ S is an INC-pair implies the condition
because of Proposition 6.1. To prove the converse, let T ∈ [R, S] and
write T as the union of its finite type R-subalgebras Tα . Now let
Q ⊆ Q′ be prime ideals of T , lying over a prime ideal P of R and
set Qα := Q ∩ Tα and Q′α := Q′ ∩ Tα . If R ⊆ Tα is quasi-finite, then
Qα = Q′α , so that Q = Q′ and then R ⊆ T is incomparable. The last
statement is Proposition 3.8.
Corollary 6.3. An integrally closed extension is Prüfer if and only if
each of its subextensions R ⊆ T of finite type has finite fibers.
Proof. It is enough to observe that the fibers of a (flat) epimorphism
have a cardinal ≤ 1, because an epimorphism is spectrally injective.
A ring extension R ⊆ S is called strongly affine if each of its subextensions R ⊆ T is of finite type. The above considerations show that
in this case R ⊆ S is quasi-Prüfer if and only if each of its subextensions R ⊆ T has finite fibers. For example, an FCP extension is
strongly affine and quasi-Prüfer. We also are interested in extensions
R ⊆ S that are not necessarily strongly affine and such that each of its
subextensions R ⊆ T have finite fibers.
Next lemma will be useful, its proof is obvious.
Lemma 6.4. Let R ⊆ S be an extension and T ∈ [R, S]
(1) If T ⊆ S is spectrally injective and R ⊆ T has finite fibers, then
R ⊆ S has finite fibers.
(2) If R ⊆ T is spectrally injective, then T ⊆ S has finite fibers if
and only if R ⊆ S has finite fibers.
Remark 6.5. Let R ⊆ S be an almost-Prüfer extension, such that the
e ⊆ S has finite fibers and let P ∈ Spec(R).
integral extension T := R
The study of the finiteness of FibR,S (P ) can be reduced as follows. As
R ⊆ S is an epimorphism, because it is Prüfer, it is spectrally injective
(see Scholium A). The hypotheses of Proposition 4.5 hold. We examine
three cases. In case (R : R) 6⊆ P , it is well known that RP = (R)P so
that |FibR,S (P )| = 1, because R → S is spectrally injective. Suppose
now that (R : R) = P . From (R : R) = (T : S)∩R, we deduce that P is
lain over by some Q ∈ Spec(T ) and then FibR,R (P ) ∼
= FibT,S (Q). The
conclusion follows as above. Thus the remaining case is (R : R) ⊂ P
QUASI-PRÜFER EXTENSIONS
31
and we can assume that P T = T for if not FibR,R (P ) ∼
= FibT,S (Q) for
some Q ∈ Spec(T ) by Scholium A (1).
e⊆S
Proposition 6.6. Let R ⊆ S be an almost-Pr̈ufer extension. If R
eP : SP ) is a maximal ideal of R
eP for
has finite fiber morphisms and (R
e then R ⊆ R and R ⊆ S have finite fibers.
each P ∈ SuppR (S/R),
Proof. The Prüfer closure commutes with the localization at prime
e Let P be a prime ideal
ideals by Proposition 4.12. We set T := R.
of R and ϕ : R → RP the canonical morphism. We clearly have
FibR,. (P ) = a ϕ(FibRP ,.P (P RP )). Therefore, we can localize the data at
P and we can assume that R is local.
In case (T : S) = T , we get a factorization R → R → T . Since
R → T is Prüfer so is R → R and it follows that R = R because a
Prüfer extension is integrally closed.
From Proposition 1.2 applied to R ⊆ T , we get that there is some
P ∈ Spec(R) such that T = RP , R/P is a valuation ring with quotient
field T /P and P = PT . It follows that (T : S) = PT = P ⊆ R, and
hence (T : S) = (T : S) ∩ R = (R : R). We have therefore a pushout
diagram by Theorem 4.6:
R′ := R/P −−−→ R/P := R′
y
y
T ′ := T /P −−−→ S/P := S ′
where R/P is a valuation domain, T /P is its quotient field and R/P →
S/P is Prüfer by [26, Proposition 5.8, p. 52].
Because R′ → S ′ is injective and a flat epimorphism, there is a bijective map Min(S ′ ) → Min(R′ ). But T ′ → S ′ is the fiber at P of T → S
and is therefore finite. Therefore, Min(S ′ ) is a finite set {N1 , . . . , Nn }
of maximal ideals lying over the minimal prime ideals {M1 , . . . , Mn } of
R′ lying over 0 in R′ . We infer from Lemma 3.7 that R′ /Mi → S ′ /Ni
is Prüfer, whence integrally closed. Therefore, R′ /Mi is an integral domain and the integral closure of R′ in S ′ /Ni . Any maximal ideal M of
R′ contains some Mi . To conclude it is enough to use a result of Gilmer
[19, Corollary 20.3] because the number of maximal ideals in R′ /Mi is
less than the separable degree of the extension of fields T ′ ⊆ S ′ /Ni .
e : S) is a maximal ideal of R.
e We
Remark 6.7. (1) Suppose that (R
e : S)P ⊆ (R
eP : SP ) and the hypotheses on (R
e : S) of
clearly have (R
the above proposition hold.
32
G. PICAVET AND M. PICAVET
e ⊆ S is a tower of finitely many integral minimal
(2) In case R
e =
extensions Ri−1 ⊆ Ri with Mi = (Ri−1 : Ri ), then SuppRe (S/R)
e where Ni = Mi ∩ R. If the ideals Ni are dif{N1 , . . . , Nn } ⊆ Max(R)
e ⊆ S is integral minimal and the
ferent, each localization at Ni of R
above result may apply. This generalizes the Ayache-Dobbs result [5,
e ⊆ S is supposed to be integral minimal.
Lemma 3.6], where R
Proposition 6.8. Let R ⊆ S be a quasi-Prüfer ring extension.
(1) R ⊆ S has finite fibers if and only if R ⊆ R has finite fibers.
(2) R ⊆ R has finite fibers if and only if each extension R ⊆ T ,
where T ∈ [R, S] has finite fibers.
Proof. (1) Let P ∈ Spec(R) and the morphisms κ(P ) → κ(P ) ⊗R R →
κ(P )⊗R S. The first (second) morphism is integral (a flat epimorphism)
because deduced by base change from the integral morphism R → R
(the flat epimorphism R → S). Therefore, the ring κ(P ) ⊗R R is zero
dimensional, so that the second morphism is surjective by Scholium
A (2). Set A := κ(P ) ⊗R R and B := κ(P ) ⊗R S, we thus have a
module finite flat ring morphism A → B. Hence, AQ → BQ is free for
each Q ∈ Spec(A) [16, Proposition 9] and BQ 6= 0 because it contains
κ(P ) 6= 0. Therefore, AQ → BQ is injective and it follows that A ∼
= B.
(2) Suppose that R ⊆ R has finite fibers and let T ∈ [R, S], then
R ⊆ RT is a flat epimorphism by Proposition 4.5(1) and so is κ(P ) ⊗R
R → κ(P ) ⊗R RT . Since Spec(κ(P ) ⊗R RT ) → Spec(κ(P ) ⊗R R)
is injective, R ⊆ RT has finite fibers. Now R ⊆ T has finite fibers
because T ⊆ RT is integral and is therefore spectrally surjective.
Remark 6.9. Actually, the statement (1) is valid if we only suppose
that R ⊆ S is a flat epimorphism.
Next result contains [5, Lemma 3.6], gotten after a long proof.
Corollary 6.10. Let R ⊆ S be an almost-Prüfer extension. Then
R ⊆ S has finite fibers if and only if R ⊆ R has finite fibers, and if
e ⊆ S has finite fibers.
and only if R
Proof. By Proposition 6.8(1) the first equivalence is clear. The second
is a consequence of Lemma 6.4(2).
The following result is then clear.
Theorem 6.11. Let R ⊆ S be a quasi-Prüfer extension with finite
fibers, then R ⊆ T has finite fibers for each T ∈ [R, S].
Corollary 6.12. If R ⊆ S is quasi-finite and quasi-Prüfer, then R ⊆ T
e ⊆ S is module finite.
has finite fibers for each T ∈ [R, S] and R
QUASI-PRÜFER EXTENSIONS
33
Proof. By the Zariski Main Theorem, there is a factorization R ⊆ F ⊆
S where R ⊆ F is module finite and F ⊆ S is a flat epimorphism
[40, Corollaire 2, p.42]. To conclude, we use Scholium A in the rest of
e ⊗R F → S is injective because F → R
e ⊗R F
the proof. The map R
is a flat epimorphism and is surjective, since it is integral and a flat
e ⊗R F → S is a flat epimorphism .
epimorphism because R
Corollary 6.13. An FMC extension R ⊆ S is such that R ⊆ T has
finite fibers for each T ∈ [R, S].
Proof. Such an extension is quasi-finite and quasi-Prüfer. Then use
Corollary 6.12.
[5, Example 4.7] exhibits some FMC extension R ⊆ S, such that
R ⊆ R has not FCP. Actually, [R, R] is an infinite (maximal) chain.
Proposition 6.14. Let R ⊆ S be a quasi-Prüfer extension such that
R ⊆ R has finite fibers and R is semi-local. Then T is semi-local for
each T ∈ [R, S].
Proof. Obviously R is semi-local. From the tower R ⊆ T R ⊆ S we
deduce that R ⊆ T R is Prüfer. It follows that T R is semi-local [5,
Lemma 2.5 (f)]. As T ⊆ T R is integral, we get that T is semi-local.
The following proposition gives a kind of converse.
Proposition 6.15. Let R ⊆ S be an extension with R semi-local. Then
R ⊆ S is quasi-Prüfer if and only if T is semi-local for each T ∈ [R, S].
Proof. If R ⊆ S is quasi-Prüfer, R ⊆ S is Prüfer. Let T ∈ [R, S]
and set T ′ := T R, so that T ⊆ T ′ is integral, and R ⊆ T ′ is Prüfer
(and then a normal pair). It follows from [5, Lemma 2.5 (f)] that T ′ is
semi-local, and so is T .
If T is semi-local for each T ∈ [R, S], so is any T ∈ [R, S]. Then,
(R, S) is a residually algebraic pair [6, Theorem 3.10] (generalized to
arbitrary extensions) and so is RM ⊆ SM for each M ∈ Max(R),
whence is Prüfer [6, Theorem 2.5] (same remark) and Proposition 1.2.
Then, R ⊆ S is Prüfer by Proposition 1.1 and R ⊆ S is quasi-Prüfer.
7. Numerical properties of FCP extensions
Lemma 7.1. Let R ⊂ S be an FCP extension. The map ϕ : [R, S] →
{(T ′, T ′′ ) ∈ [R, R]×[R, S] | SuppT ′ (R/T ′)∩SuppT ′ (T ′′ /R) = ∅}, defined
by ϕ(T ) := (T ∩ R, RT ) for each T ∈ [R, S], is bijective. In particular,
if R ⊂ S has FIP, then |[R, S]| ≤ |[R, R]||[R, S]|.
34
G. PICAVET AND M. PICAVET
Proof. Let (T ′ , T ′′) ∈ [R, R]×[R, S]. Then, R is also the integral closure
of T ′ in T ′′ (and in S).
Let T ∈ [R, S]. Set T ′ := T ∩ R and T ′′ := RT . Then (T ′ , T ′′ ) ∈
[R, R] × [R, S]. Assume that T ′ = T ′′ , so that T ′ = T ′′ = R, giving T =
R and SuppT ′ (R/T ′ ) = SuppT ′ (T ′′ /R) = ∅. Assume that T ′ 6= T ′′ . In
view of [39, Proposition 3.6], we get SuppT ′ (R/T ′ )∩SuppT ′ (T ′′ /R) = ∅.
Hence ϕ is well defined.
Now, let T1 , T2 ∈ [R, S] be such that ϕ(T1 ) = ϕ(T2 ) = (T ′ , T ′′).
Assume T ′ 6= T ′′ . Another use of [39, Proposition 3.6] gives that T1 =
T2 . If T ′ = T ′′ , then, T ′ = T ′′ = R, so that T1 = T2 = R. It follows
that ϕ is injective. The same reference gives that ϕ is bijective.
Proposition 7.2. Let R ⊂ S be a FCP extension. We define two
order-isomorphisms ϕ′ and ψ as follows:
~ → [R, R] × [R, R]
~ defined by ϕ′ (T ) := (T ∩ R, T R)
ϕ′ : [R, R]
~ → [R, R]
e × [R,
e R]
~ defined by ψ(T ) := (T ∩ R,
e T R).
e
ψ : [R, R]
Proof. This follows from [39, Lemma 3.7] and Proposition 4.19. (We
~ = R.)
e
recall that R
~ R)
e = Supp(R/R),
Corollary 7.3. If R ⊆ S has FCP, then Supp(R/
e
~
e
~
= Supp(R/R)
and Supp(R/R)
= Supp(R/R)∪Supp(R/R).
Supp(R/R)
e R),
e B := Supp(R/R),
e
e
Proof. Set A := Supp(R/
C := Supp(R/R)
e
with
and D := Supp(R/R). Then, A ∪ B = C ∪ D = Supp(R/R),
A ∩ B = C ∩ D = B ∩ D = ∅ by Proposition 4.19.
Assume that A ∪ B 6= B ∪ D and let P ∈ (A ∪ B) \ (B ∪ D).
e P = (R)P (R)
e P = RP , a contradiction. It follows that
Then, RP 6= (R)
A ∪ B = B ∪ D. Intersecting the two members of this equality with
A and D, we get A = A ∩ D = D. In the same way, intersecting the
equality A ∪ B = C ∪ D = C ∪ A by B and C, we get B = C.
Corollary 7.4. Let R ⊂ S be an FCP extension. We define two orderisomorphisms
~ by ϕ1 (T ) := T R
e → [R, R]
ϕ1 : [R, R]
e R]
~ by ψ1 (T ) := T R.
e
ψ1 : [R, R] → [R,
Proof. We use notation of Proposition 4.19. We begin to remark that
e play symmetric roles.
R and R
e be such that ϕ1 (T ) = ϕ1 (T ′ ). Since T ∩ R =
Let T, T ′ ∈ [R, R]
′
T ∩ R = R by Proposition 4.19, we get ϕ(T ) = ϕ(T ′ ), so that T = T ′
and ϕ1 is injective. A similar argument shows that ψ1 is injective.
QUASI-PRÜFER EXTENSIONS
35
e There exists T ∈ [R, R]
e such that ϕ(T ) = (R, U), so
Let U ∈ [R, R].
e
e
that R = T ∩ R and U = T R. Let M ∈ Supp(R/R)
= Supp(R/R)
∪
e
Supp(R/R) by Corollary 7.3. If M ∈ Supp(R/R),
then M 6∈ Supp(R/R)
e M = RM R
eM = R
eM . If M ∈
by Proposition 4.19, giving TM ⊆ R
Supp(R/R), the same reasoning gives TM ⊆ RM , so that RM = TM ∩
eM . Then, TM = R
eM . It follows that T ⊆ R,
e
RM = TM , but RM = R
e
giving T ∈ [R, R] and ϕ1 is surjective, hence bijective. A similar argument shows that ψ1 is surjective, hence bijective.
~
e × [R, R] → [R, R]
Corollary 7.5. If R ⊂ S has FCP, then θ : [R, R]
′
′
defined by θ(T, T ) := T T , is an order-isomorphism. In particular, if
e
~ ≤ |[R, S]|.
R ⊂ S has FIP, then |[R, R]||[R,
R] = |[R, R]k
Proof. Using notation of Proposition 7.2 and Corollary 7.4, we may
remark that ψ◦θ = Id×ψ1 . Since ψ and Id×ψ1 are order-isomorphisms,
so is θ. The FIP case is obvious.
Gathering the previous results, we get the following theorem.
Theorem 7.6. If R ⊂ S has FCP, the next statements are equivalent:
(1) Supp(R/R) ∩ Supp(S/R) = ∅.
(2) The map ϕ : [R, S] → [R, R] × [R, S] defined by ϕ(T ) := (T ∩
R, T R) is an order-isomorphism.
(3) R ⊆ S is almost-Prüfer.
e
(4) Supp(S/R) = Supp(R/R).
e → [R, S] defined by ϕ1 (T ) := T R is an
(5) The map ϕ1 : [R, R]
order-isomorphism.
e S] defined by ψ1 (T ) := T R
e is an
(6) The map ψ1 : [R, R] → [R,
order-isomorphism.
e × [R, R] → [R, S] defined by θ(T, T ′ ) := T T ′
(7) The map θ : [R, R]
is an order-isomorphism.
e = Supp(R/R).
If one of these conditions holds, then Supp(S/R)
If R ⊂ S has FIP, the former conditions are equivalent to each of
the following conditions:
e
R]|.
(8) |[R, S]| = |[R, R]||[R,
(9) k[R, S]| = |[R, R]||[R, S]|.
e = |[R, S]|.
(10) |[R, R]|
e S]|.
(11) |[R, R]| = |[R,
Proof. (1) ⇒ (2) by [39, Lemma 3.7].
36
G. PICAVET AND M. PICAVET
(2) ⇒ (1). If the statement (2) holds, there exists T ∈ [R, S] such
that T ∩ R = R and T R = S. Then, [39, Proposition 3.6] gives that
Supp(R/R) ∩ Supp(S/R) = ∅.
(1) ⇒ (3) by [39, Proposition 3.6].
(3) ⇒ (4), (5), (6) and (7): Use Corollary 7.3 to get (4), Corollary 7.4
to get (5) and (6), and Corollary 7.5 to get (7). Moreover, (3) and
e = Supp(R/R).
Corollary 7.3 give Supp(S/R)
(4) ⇒ (1) by Proposition 4.19(2).
e
(5), (6) or (7) ⇒ (3) because, in each case, we have S = RR.
Assume now that R ⊂ S has FIP.
Then, obviously, (7) ⇒ (8), (2) ⇒ (9), (5) ⇒ (10) and (6) ⇒ (11).
e so
e
R]| = |[R, R]|,
(9) ⇒ (3) by Corollary 7.5, which gives |[R, R]||[R,
e and then S = R.
e
that |[R, S]| = |[R, R]|,
(8) ⇒ (1): Using the map ϕ of Lemma 7.1, we get that {(T ′ , T ′′) ∈
[R, R] × [R, S] | SuppT ′ (R/T ′ ) ∩ SuppT ′ (T ′′ /R) = ∅} = [R, R] × [R, S],
so that SuppR (R/R) ∩ SuppR (S/R) = ∅.
(10) ⇒ (3) and (11) ⇒ (3) by Corollary 7.4.
Example 7.7. We give an example where the results of Theorem 7.6
do not hold if R ⊆ S has not FCP. Set R := ZP and S := Q[X]/(X 2 ),
e = Q because R ⊂ R
e is Prüfer (minimal)
where P ∈ Max(Z). Then, R
e ⊂ S is integral minimal. Set M := P RP ∈ Max(R) with (R, M)
and R
a local ring. It follows that M ∈ Supp(R/R) ∩ Supp(S/R) because
R ⊂ S is neither integral, nor Prüfer. Similarly, M ∈ Supp(R/R) ∩
e
Supp(R/R).
Indeed, R ⊂ R has not FCP.
We end the paper by some length computations in the FCP case.
Proposition 7.8. Let R ⊆ S be an FCP extension. The following
statements hold:
e = ℓ[R, R]
~ and ℓ[R, R] = ℓ[R,
e R]
~
(1) ℓ[R, R]
~ = ℓ[R, R]
e + ℓ[R,
e R]
~ = ℓ[R, R] + ℓ[R, R]
~
(2) ℓ[R, R]
~ = |Supp (R/R)|
e = |SuppR (R/R)|.
e
~
(3) ℓ[R, R]
= ℓ[R, R]
R
Proof. To prove (1), use the maps ϕ1 and ψ1 of Corollary 7.4. Then (2)
follows from [11, Theorem 4.11] and (3) from [10, Proposition 6.12].
References
[1] D.F. Anderson and D.E. Dobbs, Pairs of rings with the same prime ideals,
Can. J. Math. XXXII, (1980), 362–384.
[2] D. F. Anderson, D. E. Dobbs and M. Fontana, On treed Nagata rings, J.
Pure Appl. Algebra, 61, (1989), 107–122.
QUASI-PRÜFER EXTENSIONS
37
[3] A. Ayache, M. Ben Nasr, O. Echi and N. Jarboui, Universally catenarian
and going-down pairs of rings, Math. Z, 238, (2001), 695–731.
[4] A. Ayache, A constructive study about the set of intermediate rings, Comm.
Algebra, 41 (2013), 4637–4661.
[5] A. Ayache and D. E. Dobbs, Finite maximal chains of commutative rings,
JAA, 14, (2015), 14500751–1450075-27.
[6] A. Ayache and A. Jaballah, Residually algebraic pairs of rings, Math. Z,
225, (1997), 49–65.
[7] M. Ben Nasr and N. Jarboui, New results about normal pairs of rings with
zero-divisors, Ricerche mat. 63 (2014), 149–155.
[8] R.D. Chatham, Going-down pairs of commutative rings, Rendiconti del
Circolo Matematico di Palermo, Serie II, Tomo L, (2001), 509–542.
[9] G. W. Chang and M. Fontana, Uppers to 0 in polynomial rings and Prüfer
-like domains, Comm. Algebra, 37 (2009), 164–192.
[10] D. E. Dobbs, G. Picavet and M. Picavet-L’Hermitte, Characterizing the
ring extensions that satisfy FIP or FCP, J. Algebra, 371 (2012), 391–429.
[11] D. E. Dobbs, G. Picavet and M. Picavet-L’Hermitte, Transfer results for
the FIP and FCP properties of ring extensions, Comm. Algebra, 43 (2015),
1279–1316.
[12] D. E. Dobbs, G. Picavet and M. Picavet-L’Hermitte, When an extension
of Nagata rings has only finitely many intermediate rings, each of those is
a Nagata ring?, Int. J. Math. Math. Sci., 2014 (2014), Article ID315919,
13 pp.
[13] D. E. Dobbs, G. Picavet, M. Picavet-L’Hermitte and J. Shapiro, On intersections and composites of minimal ring extensions, J P J. Algebra, Number
Theory Appl., 26 (2012), 103–158.
[14] D. E. Dobbs, On characterizations of integrality involving the lying-over
and incomparability properties, J. Comm. Algebra, 1 (2009), 227–235.
[15] D.E. Dobbs and J. Shapiro, Pseudo-normal pairs of integral domains, Houston J. Math., 40 (2014), 1–9.
[16] S. Endo, On semi-hereditary rings, J. Math. Soc. Japan, 13 (1961), 109–
119.
[17] D. Ferrand and J.-P. Olivier, Homomorphismes minimaux d’anneaux, J.
Algebra, 16 (1970), 461–471.
[18] M. Fontana, J. A. Huckaba and I. J. Papick, Prüfer domains, Dekker, New
York, 1997.
[19] R. Gilmer, Multiplicative Ideal Theory, Dekker, New York, 1972.
[20] M. Griffin, Prüfer rings with zero divisors, Journal fur die reine and angewande Mathematik, 239, (1969), 55–67.
[21] M. Grandet, Une caratérisation des morphismes minimaux non entiers,
C.R. Acad. Sc. Paris, 271, (1970), Série A 581–583.
[22] A. Grothendieck and J. Dieudonné, Eléments de Géométrie Algébrique,
Springer Verlag, Berlin, (1971).
[23] E. Houston, Uppers to zero in polynomial rings, pp. 243–261, in: Multiplicative Ideal Theory in Commutative Algebra, Springer-Verlag, New York,
2006.
[24] A. Jaballah, Finiteness of the set of intermediary rings in normal pairs,
Saitama Math. J., 17 (1999), 59–61.
38
G. PICAVET AND M. PICAVET
[25] N. Jarboui and E. Massaoud, On finite saturated chains of overrings,
Comm. Algebra, 40, (2012), 1563–1569.
[26] M. Knebusch and D. Zhang, Manis Valuations and Prüfer Extensions I,
Springer, Berlin, 2002.
[27] D. Lazard, Autour de la platitude, Bull. Soc. Math. France, 97, (1969),
81–128.
[28] M. Lazarus, Fermeture intégrale et changement de base, Ann. Fac. Sci.
Toulouse, 6, (1984), 103–120.
[29] T. G. Lucas, Some results on Prüfer rings, Pacific J. Math., 124, (1986),
333–343.
[30] K. Morita, Flat modules, Injective modules and quotient rings, Math. Z.,
120 (1971), 25–40.
[31] J.P. Olivier, Anneaux absolument plats universels et épimorphismes à buts
réduits, Séminaire Samuel. Algèbre Commutative, Tome 2 (1967-1968),
exp. no 6, p. 1–12.
[32] J. P. Olivier, Montée des propriété par morphismes absolument plats, J.
Alg. Pure Appl., Université des Sciences et Technique du Languedoc, Montpellier, France (1971).
[33] J.P. Olivier, Going up along absolutely flat morphisms, J. Pure Appl. Algebra, 30 (1983), 47–59.
[34] G. Picavet, Propriétés et applications de la notion de contenu, Comm.
Algebra,13, (1985), 2231–2265.
[35] G. Picavet, Universally going-down rings, 1-split rings, and absolute integral closure, Comm. Algebra, 31, (2003), 4655–4685.
[36] G. Picavet and M. Picavet-L’Hermitte, About minimal morphisms, pp.
369–386, in: Multiplicative Ideal Theory in Commutative Algebra, Springer,
New York, 2006.
[37] G. Picavet, Seminormal or t-closed schemes and Rees rings, Algebra Repr.
Theory, 1, (1998), 255–309.
[38] G. Picavet and M. Picavet-L’Hermitte, Some more combinatorics results
on Nagata extensions, Palestine J. Math., 1, (Spec.1), (2016), 49–62.
[39] G. Picavet and M. Picavet-L’Hermitte, Prüfer and Morita hulls of FCP
extensions, Comm. Algebra, 43, (2015), 102-119.
[40] M. Raynaud, Anneaux locaux Henséliens, Lect. Notes in Math., Springer,
Vol. 169, (1970).
[41] H. Uda, Incomparability in ring extensions, Hiroshima Math. J., 9, (1979),
451–463.
[42] S. Visweswaran, Laskerian pairs, J. Pure Appl. Algebra, 59, (1989), 87–110.
Université Blaise Pascal, Laboratoire de Mathématiques, UMR6620
CNRS, 24, avenue des Landais, BP 80026, 63177 Aubière CEDEX, France
E-mail address: [email protected]
E-mail address: picavet.gm(at)wanadoo.fr
| 0math.AC
|
Stateful Behavioral Types for ABS
Eduard Kamburjan and Tzu-Chun Chen
arXiv:1802.08492v1 [cs.PL] 23 Feb 2018
Department of Computer Science, Technische Universität Darmstadt, Germany
[email protected], [email protected]
Abstract. It is notoriously hard to correctly implement a multiparty
protocol which involves asynchronous/concurrent interactions and the
constraints on states of multiple participants. To assist developers in implementing such protocols, we propose a novel specification language to
specify interactions within multiple object-oriented actors and the sideeffects on heap memory of those actors; a behavioral-type-based analysis
is presented for type checking. Our specification language formalizes a
protocol as a global type, which describes the procedure of asynchronous
method calls, the usage of futures, and the heap side-effects with a firstorder logic. To characterize runs of instances of types, we give a modeltheoretic semantics for types and translate them into logical constraints
over traces. We prove protocol adherence: If a program is well-typed
w.r.t. a protocol, then every trace of the program adheres to the protocol, i.e., every trace is a model for the formula of its type.
1
Introduction
The combination of actors [26] in object-oriented languages (e.g., Scala [33] and
ABS [28]), sometimes called Active Objects [7,32], is an active research area for
system models and is frequently used in practice [36]. In general, Active Objects
communicate internally within an object via the object’s heap memory, and communicate externally by asynchronous method calls through futures [3,22], which
are constructs for synchronizing executions invoked by those calls. Encapsulated
heap memory and asynchronous calls with explicit synchronization points (specified by futures) make it hard to specify and verify protocols for Active Objects.
The main obstacle, as Din and Owe [15] pointed out, is to bridge the gap between
local perspectives of single objects and global perspectives of the whole system,
e.g., it is non-trivial to precisely specify the communication within an object’s
heap memory from a global perspective [16]. Multiparty session types (short as
MPST) [13,27], one important member of behavioral types [2,19], are established
theories for typing globally stateless concurrent interactions (i.e., method calls)
among multiple participants (i.e., objects) to ensure communication safety. Current works in MPST [5,37] have attempted to specify states in communication
by using global values and assuming that channels are the only communication
concept; however, global values are not sufficient to specify the non-trivial interplay of processes with the heap memory inside an object, and channels are not
able to fully represent the usage of futures.
1
translation
execution
We integrate a stateful analysis and the specification of Din et al. [15] into
MPST since MPST is able to formally specify the scenario of global interactions.
By specifying communication from a global perspective and functional properties of local states in communication, we aim to ensure (1) the correctness of
heap accesses of methods of actors and (2) the deadlock freedom of well-typed
processes which interact with each other. The verification of local specifications
compositionally guarantees the global specification.
We specify passed data and modifications of heap memory with first-order
logic (FOL) formulas and transform our behavioral types into logical constraints
on traces. Moreover, from the model-theoretic perspective, we define protocol adherence as that every generated trace of a well-typed(`) program is a model(|=)
for the translation of the type. The following scenario illustrates our approach
and the challenges for specifying and verifying protocols for Active Objects.
Assume we have a GUI U, a computation server S,
Program ` Protocol
and an interface server I such that U, without knowing
S, wants to compute some data by sending them to I via
a method call. After executing this call, U gets ready for
the next action by setting field intern to value expect and
Trace |= Constraint terminating its process to stay responsive. I delegates U
’s task to S and remains responsive to other requests without waiting for S’s
computation through invoking a new method call on U with future x, which will
carry the computation result, bigger than 0, back to U. The code and figure below
implement this scenario:
1 object U{
2 TState intern = init;
3 Int resume(FuthInti x){
4
if( this.intern!=expect } return −1;
5
Int r = x.get; return r; }
6 Unit start(Int j){
7
FuthUniti f = I!cmp(this, j);
8
this.intern = expect; }}
9
10
11
12
13
14
15
16
object I{
Unit cmp(Int dat){
FuthInti f = S!cmp(dat);
FuthInti f’ = U!resume(f);}}
object S{ Int cmp(Int i){ ... }}
main { U!start(20); }
In the code, ! denotes a non-blocking call, I!cmp calls U
I
S
method cmp of I, U!start calls the initial method of U,
cmp
U!resume calls method resume for continuation, and S!cmp
cmp
starts the actual computation at S. The challenge for
resume
formal specifications is to express that (1) I is transparent
(get)
to U and S such that I must pass the same data to S that
it received from U and I does not read the return value from S, and (2) U changes
its heap to expect and reads the correct future.
Contributions. We propose (1) a specification language for actors’ behaviors,
which integrates FOL to specify heap memory, (2) a type system integrating a
validity calculus for it and (3) model-theoretic semantics for protocol adherence.
Roadmap. Section 2 gives an overview of our approach and illustrates the workflow. We first present the core ideas without loops and branching: Section 3
introduces a core language for Active Objects, Sync, and its dynamic logic, Section 4 gives the types and operations on them and Section 5 gives the type
2
system. Section 6 gives language and type constructs for loops and Section 7
gives language and type constructs for branching. Section 8 gives a conclusion
and related work.
2
Challenges and Overview of the Workflow
We consider object-oriented actors, which use method calls, futures, and heap
memory for communication. Every method call is asynchronous and starts a new
process at the callee object. At each such call, the active process at the caller
gets a fresh future identity on which one may synchronize on the termination
of the started process. An object delete may only switches its active process to
another process if the currently active process terminates. Using futures gives
programmers the control of when to make a synchronization – however, it leads
to the following complications:
Unexposed State In the Active Object concurrency model, each process has
exactly one future, thus reading from a future is synchronizing with an unknown process and depends on the state of the process’s object. Futures cannot be analyzed in isolation — reading from a future must take the unexposed
state of the object into account.
Mixed Communication Paradigms Communication through the heap memory is hard to describe with data types, as it requires fine-grained specification of computation and has no explicit caller or callee. Thus, it is difficult to
isolate the parts of the program realizing the communication protocol. Additionally, method calls are asynchronous, while future reads are synchronous.
Two-Fold Endpoints In the Active Object model, the callee endpoints of
methods calls are objects, but the caller endpoints and the endpoints for future synchronization are processes. The interplay of multiple objects, which
contain multiple processes, must be captured in the analysis by a two-fold
notion of endpoints: Objects and processes are both endpoints.
Example 1: Specifying global types. Our specification language for side-effects is
a FOL for specifying local memory instead of global values since (1) global values
are not natural in an Active Object setting, and (2) a logic over memory locations
(variables and fields) allows us to use a well-established theory of first order
(FO) dynamic logic [23] to capture the semantics of methods. We formalize the
scenario in Section 1 by the following global type in our specification language:
.
G =main →
− U : starthU.state = expecti . U →
− I : comp .
f
.
.
I−
→ S : comphi = dat, result > 0i . I →
− U : resumehx = f, >i . U ↑ x . End
where > denotes true. Following the syntax of Honda et al. [27], U →
− I : comp
denotes a message comp from U to I, i.e., the call to a method comp. Formula
.
U.state = expect is the postcondition for the process started by this call. If two
formulas are provided, the first is the precondition describing the state of the
caller and the second is the postcondition describing the state of the callee and
3
GX
prp (G X)
G X0
prp∗ (G X0 )
prp∗ (G X) m1
∗
prp∗ (G X) m2
G
|
{z
}|
Step 1: Generating
local types for objects
{z
Step 2: Propagating
guarantees in objects
}|
prp∗ (G X0 ) m01
prp∗ (G X0 ) m02
{z
}
Step 3: Generating
local types for methods
Fig. 1. Workflow for Phase 1: G is a global type and denotes projection on object X
resp. method m. Function prp∗ is the function propagating guarantees.
the return value, which is denoted by keyword result. The annotation f denotes
the memory location where the future of the denoted call is stored. Formula
.
i = dat states that dat, the parameter of S.cmp, carries the same value as re.
ceived by I.cmp on parameter i, while formula x = f requires that parameter
x of the call at method resume carries the future of the previous call to comp.
Finally, U ↑ x describes a read of U on the future stored in the location x. Note
that we specify locations in formulas and avoid a situation where that an endpoint must guarantee an obligation containing values that it cannot access. For
example Bocchi et al [5] allow this situation and thus require additional analyses
of history-sensitivity and temporal-satisfiability.
For analysis, we adopt an approach similar to MPST: We project a global
type on endpoints defined inside it, to automatically derive local specifications
for all objects and methods. Additionally, formulas, which are used to specify
conditions on the heap memory, are projected on the logical substructure of the
callee, because the callee cannot access the caller’s fields.
Two-phase Analysis. The analysis requires that the protocol is encoded as a
global type, which defines the order of method calls and future read operations
between objects, annotated with FO specifications of heap memory and passed
data. Our analysis mechanism has two phases. In Phase 1, a global type is used
to generate local types for all endpoints. In Phase 2, the endpoints are type
checked against their local types via generating a causality graph. The workflow
of Phase 1 is based on MPST’s approach, but is adjusted to the Active Object
concurrency model:
Phase 1. The workflow of Phase 1 is shown in Fig. 1.
– Step 1 : The global type is projected onto the participating objects and generates object types. Such a type specifies the obligation of an object for
running methods in a certain order, and for guaranteeing the FOL specifications of the object’s state. During projection, the FO-specifications are
projected onto the substructure of this object, i.e., the object only has to
prove specifications of memory that it can access.
4
– Step 2 : FO-specifications are propagated within an object type: as the order
of method executions is specified by the specification, the postcondition of
a method can be assumed as a precondition for the next method.
– Step 3 : An object type is projected on its methods, producing method types.
A global type encodes the following obligations (short as Obl. ) for the implementation: (Obl. a) the whole system does not deadlock, (Obl. b) for each object,
the observable events (calls and reads) are ordered as specified in the global type
and (Obl. c) for each method, the observable events are ordered as specified in
the local type derived from the global type, and adhere to the FO-specifications.
In the following, we demonstrate the Phase 1 workflow of the global type in
Example 1. We do not formally introduce the syntax at this point.
Step 1: Object Types. Continue with Example 1, projecting G on object U we
get
.
.
?starth>i.I!comph>i.Put state = expect.?resumeh∃f. x = f i.Read x.Put result > 0
Type ?starth>i denotes a starting point for runtime execution and type I! comph>i
denotes an invocation of method comp. Type Put ϕ specifies the termination of
the currently active process in a state where ϕ holds. The position of Put is
derived automatically and we do not specify in a global type when a process
terminates. The precondition of resume is weakened: The location f is not visible
to U, so callee U can use not all information from caller I. To check what information U is able to access, it is syntactically checked that all locations in ϕ are
visible to U. Type Read x specifies a synchronization on the future stored in x.
Step 2: Propagation. In the next step we propagate the postcondition of the
last process to the precondition of the next process. No process is specified
as active between start and resume, so the heap is not modified — thus, the
.
postcondition of start still holds when resume starts. Adding state = expect to
the precondition of resume strengthens the assumption for the type checking of
resume. The propagation of conditions results in:
prp∗ (G U) =?starth>i . I!comph>i . Put state = expect .
.
.
?resumeh∃f. x = f ∧ state = expecti . Read x . Put result > 0
Step 3: Method Types. We generate a method type to specify a method in isolation. Projecting the object type in Step 2 on method resume generates:
.
.
prp∗ (G U) resume =?resumeh∃f. x = f ∧ state = expecti.Read x.Put result > 0
Method types share the syntax with object types, and the projection from
object types is splitting at positions where one method ends and another starts.
Phase 2. After generating method types, Phase 2 of the analysis type checks the
implementation of methods against their method types, where the formulas are
checked for validity. The type checking of method types guarantees the correct
local order of events (Obl. b and c). State specifications are checked by integrating a validity calculus into the type system. To guarantee (Obl. a), we require
the following analyses:
5
Causality Graph. We generate a causality graph to ensure deadlock freedom
(Obl. a): A deadlock free causality graph for Active Objects is cycle-free [17,25].
A causality graph is also used to ensure that methods of one object are executed
in the order specified in the global type that the object obeys to (Obl. b).
U
I
?
!
?
↓
?
!
↑
!
?
S
↓
↓
↓
The nodes are the local types from the projected object types. An edge connecting two nodes models that the statement for the first type causes the statement
for the second type; for example, there are edges from a call to the corresponding
receiving type. The graph is partially generated from G, and partially generated
from the code: The edge connecting the gray nodes is added by a Points-To analysis, which maps a location of future to the methods resolving this future. The
dotted line is considered for cycle-freedom check— the termination of a method
causes the next to start— but not for method order check.
Model-theoretic Semantics. One of our contribution is defining and verifying
protocol adherence from a model-theoretic view: The property that a program
follows a specified scenario in the sense that every generated trace is a model for
the translation of the global type. We thus define protocol adherence through a
logical characterization of global types and translate types into constraints over
traces, which are sequences of configurations generated by the program. This
declarative approach for defining protocol adherence allows us to connect the
FO properties embedded in the type to the execution of methods by using a
dynamic logic to capture side-effects. If s is the body of some method m, the
dynamic logic formula ϕ → [s]ψ expresses that the first-order formula ψ holds
after executing m if ϕ holds in the beginning. If the dynamic logic formula is
valid, then for every generated trace, at every position i where m terminates ψ
holds if at the corresponding position j < i where it starts ϕ holds.
3
The Core Actor-Based Language using Futures
This section introduces Sync, a simple Active Object language based on ABS [28].
Due to space limitation, we present the basic constructs of Sync below, but
Projection
L
G
Points-To
Future locations
Type System
SDL-formulas
Validity Calculus
guarantees
guarantees
guarantees
guarantees through well-formedness
Deadlock Freedom
Local Action Order
Protocol Adherence
State Specification
Composition Guarantees
Fig. 2. Workflow for Phase 2 in our analysis.
6
leave constructs of branching and repetition to later Sections. Sync-program
consists of a main statement and a set of actors, which are objects that do
have fields and method but do not share states. Inside an object, processes do
not interleave such that an active process must terminate before another can be
scheduled. Therefore, methods can be considered sequential. We assume standard
operations and literals on types of booleans, integers, lists and Object.
Definition 1 (Syntax of Sync). Let e denote expressions, T denote data types,
x denote variable and field names, X denote object names, and Fut<T> denote
future types. · represents possibly empty lists and [·] represents optional elements.
Prgm ::= O main{X!m(e)} O ::= object X {M T x = e} M ::= T m(T x){s; return e}
s ::= [Fut<T>] x = X!m(e) | [T] x = e | T x = e.get | skip | s; s
Objects communicate only by asynchronous method calls using futures. Upon
a method call, a fresh future is generated on callee side and is passed to the
caller. The callee writes the return value into the future upon termination of
the corresponding process; anyone with the access to the future can read, but
not write, into it. Sync is a standard imperative language with two additional
statements: (1) x = X!m(e) calls method m with parameters e on object X. The
generated future is stored in x. The caller continues execution, while the callee is
computing the call on m or scheduling m for later if another process is currently
active. (2) e.get reads a value from the future stored in e. If the process computing
this future has not terminated, the reading process blocks.
To define a small-step reduction relation over events for the semantics of
Sync, we first define an event as a process action in communication:
Definition 2 (Events). Let f, f 0 range over futures. An event, denoted by ev,
is defined by the following grammar:
ev ::=iEv(X, X0 , f, m, e) | iREv(X, f, m) | fEv(X, f, m, e) | fREv(X, f, e) | noEv
An invocation iEv(X, X0 , f, m, e) models that X calls method X0 .m using f by passing e as parameters. An invocation reaction iREv(X, f, m) models that X starts
the execution of m to resolve f . A resolving fEv(X, f, m, e) models that X resolves
f , which contains contains e at the moment, by finishing the execution of m. A
fetch fREv(X, f, e) models that X reads value e from f .
Fig. 3 defines the reduction relation, denoted by →ev , for the semantics.
JeKσ,ρ denotes the evaluation of an expression under stores σ and ρ. Rule (call)
executes a method call on the object stores in e by generating a fresh future f 0
and an invocation event. The new process is not set as active. By rule (start), the
new process becomes active upon creation: the object X must be inactive. An
invocation reaction event is generated. Rule (sync) synchronizes on a future f 0
stored in e, by checking whether the configuration contains prc(X0 , f 0 , val(e0 ), σ 0 ),
i.e. f 0 is resolved, and reads the return value e0 . Rule (end) terminates a process.
In all other rules, the ev parameter is noEv.
A configuration is composed of processes and objects. A process has a unique
future f , a store σ which maps fields to literals, and the id of its object X.
An object has a unique name X, an active future f , and a store ρ which maps
variables to literals.
7
C does not contain f 0
(call)
JeKσ,ρ = X0
C = ob(X, f, ρ) C0
ev = iEv(X, X0 , f, m, JeKσ,ρ )
c(m, e0 )) C
prc(X, f, m(e!m0 (e0 ); s), σ) C →ev prc(X, f, m(s), σ) prc(X0 , f 0 , m0 (M (m0 )), M
(start)
ev = iREv(X, f, m)
prc(X, f, m(s), σ) ob(X, ⊥, ρ) C →ev prc(X, f, m(s), σ) ob(X, f, ρ) C
C = prc(X0 , f 0 , val(e0 ), σ 0 ) C0
(sync)
(end)
JeKσ,ρ = f 0
ev = fREv(X, f, e)
prc(X, f, m(e.get; s), σ) ob(X, f, ρ) C →ev prc(X, f, m(s), σ) ob(X, f, ρ) C
ev = fEv(X, f, m, e)
prc(X, f, m(return e), σ) ob(X, f, ρ) C →ev prc(X, f, val(JeKσ,ρ ), σ) ob(X, ⊥, ρ) C
Fig. 3. The semantics rules.
Definition 3 (Runtime Syntax of Processes and Objects). The following
grammar defines runtime processes and objects as configurations C:
C ::= prc(X, f, m(s), σ) | prc(X, f, val(e), σ) | ob(X, f, ρ) | C C
A process either is executing a method m for a request carried by f at some
object X, represented by prc(X, f, m(s), σ), or has returned e, represented by
prc(X, f, val(e), σ). An object ob(X, f, ρ) has its name X, the future of the active
process f and the heap ρ. We write ob(X, ⊥, ρ) to indicate that X is inactive. The
composition of configurations is commutative and associative, i.e. C C0 = C0 C
and C (C0 C00 ) = (C C0 ) C00 .
We denote the initial configuration of a program Prgm with I(Prgm). If all
processes of a configuration C have terminated, the configuration also terminates.
c(m, e) for the
The method body of method m is denoted M (m) and we write M
initial local store of a task executing m with parameters e.
We use traces, sequences of pairs of events and configurations, to capture the
behavior of a program. We only consider terminating runs and define big-step
semantics Prgm ⇓ tr for finite traces:
Definition 4 (Run and Big-Step Semantics). A run from C1 to Cn is a
sequence of configurations C1 , . . . , Cn with events ev1 , . . . , evn−1 such that:
C1 →ev1 C2 →ev2 . . . →evn−1 Cn
A trace tr of a run is a sequence (ev1 , C1 ), . . . , (evm , Cm ) where for every 1 ≤
j < m there is a C such that Cj →evj C is in the run and evj 6= noEv holds. Let
Prgm be a Sync program. Prgm generates tr, denoted by Prgm ⇓ tr, if and only
if there is a run from its initial configuration to a terminated configuration such
that tr is the trace of this run.
Dynamic Logic. A dynamic logic combines FO-formulas over the heap with
symbolic executions [1,8] of statements. An symbolic execution uses symbolic
values to describe a possible set of actual values; and does not reason about one
execution of the statement, but describes a set of executions.
8
Example 2. Formula ∃Int a. a > 0 ∧ i > a → [j = i*2;]j > 0 describes that if
there is a number a bigger than 0 and smaller than the value stored in i, then
after executing j = i*2;, variable j contains a positive value.
Based on ABSDL [14], we present Sync Dynamic Logic (short as SDL), which
extends first-order logic over program variables and heap memory with modalities that model the effect of statements. In this logic, method parameters are
special variables and a modality is a formula [s]ϕ which holds in a configuration, say C, if ϕ holds in every configuration reached from C after executing s.
Configurations are FO models where we evaluate modality-free SDL formulas.
Definition 5 (Formulas ϕ). We define the set of formulas ϕ and terms t by
the following grammar, where p ranges over predicate symbols, f ranges over
function symbols, x ranges over logical variables, and v ranges over logical and
program variables. The set of all formulas is denoted by SDL.
.
ϕ ::= tt | ¬ϕ | ϕ ∨ ϕ | p(t . . . t) | t ≥ t | t = t | ∃T x; ϕ | [s]ϕ
t ::= v | f(t . . . t)
Local program variables (i.e., v) are modeled as special function symbols. To
model heap accesses, following Schmitt et al. [35], we use two function symbols
store and select with the connecting axiom select(store(heap, f, o, value), f, o) =
value where heap is a special local program variable modeling the heap explicitly.
A special function symbol result is interpreted as the return value of a method,
and a logical variable is free if it is not bound by any quantifier.
Definition 6. A formula ϕ is valid if it evaluates to true in every configuration.
We differentiate between global formulas which refer to the heap of multiple
objects and X-formula which refer only to X. A X-formula, containing only the
function symbols for elements from X and special function symbol self that
models the reference to X, reasons about X – it can be proven that it holds for
a given state, only by checking locally the code of X. A validity calculus for a
superset of SDL is presented in [15].
Definition 7. Let ϕ be a formula. The weakened X-formula ϕ@X is obtained by
replacing all function symbols in ϕ which are not exclusive to X (i.e., refer to the
fields of other objects) by free variables and existentially quantifying over them.
Example 3. Let fl be a field, X an object and i the parameter of some method
in class X. Consider ϕ = X.fl > 0 ∧ i > X.fl. The formula ϕ is an X-formula, as
ϕ = ϕ@X. The weakening for some object X0 is ϕ@X0 = ∃Int a.a > 0 ∧ i > a.
ϕ@X0 does not reason about X.fl, but still has the information that i > 1.
4
Behavioral-Type-Based Stateful Specifications of
Futures
Definition 8 defines a specification language for writing global types to specify
the behavior of the system. Due to space limitation, we only represent the key
constructs here but leave other constructs to later Sections.
9
Definition 8 (Syntax of Global Types). Let ϕ, ψ range over modality-free
SDL formulas and Xi range over object names. [·] denotes optional elements.
[x]
G ::= main →
− X : mhϕi.G
G ::= X1 −−→ X2 : mhϕ, ψi.G | X ↑ e.G | End
[x]
The calling type X1 −−→ X2 : mhϕ, ψi specifies a method call from X1 to m at
X2 . If x is written above the arrow, the future of this call must be stored in
location x. The SDL-formula ϕ specifies (1) the call parameters passed to the
callee and (2) the memory of X1 at the moment of the call. Formula ψ is the
postcondition of the callee process and specifies the state of X2 and the return
value. The initial method call main →
− X : mhψi only specifies the postcondition
of the process running X.m. Type X ↑ e specifies a synchronization on the future
stored in e. Its parameter e is an expression – it is possible to specify a reading
action from a future stored in a list. Every synchronization must be specified.
End specifies the end of communication. G denotes a complete protocol with an
initializing method call, while G denotes partial types. Even without fields in the
formula, the implementation is referenced in the specifications, as the endpoints
are object names. Object and method types use share the syntax, together we
call them local types. The grammar of local types is defined as follows:
Definition 9 (Syntax of Local Types). Let ϕ range over modality-free SDL
formulas and Xi range over object names. [·] denotes optional elements.
L ::=?mhϕi.L
L ::=?mhϕi.L | X![x] mhϕi.L | Put ϕ.L | Read e.L | skip.L | End
Receiving type ?mhϕi denotes the start of a process computing m in a state
where formula ϕ holds. Here ϕ is a precondition of m and describes the local
state and method parameters of m. Type Put ϕ denotes the end of a process in
a state where ϕ holds. Here ϕ is a postcondition and describes the return value
and the local store. Contrary to global types, a postcondition of a process is
not annotated at the call, but at the point of termination, because the point
of termination is now explicit. Type X![x] mhϕi corresponds to the caller side of
[x]
X1 −−→ X2 : mhϕ, ψi. Type Read e models a read from e and skip denotes no
communication. We use L to denote a complete local type and L to denote
partial types.
Projection has three steps: (1) projection of global types on objects, (2)
condition propagation, and (3) projection of object types on methods.
Projection on Objects. The projection on objects needs to (1) ensure that every
object has access to all locations occurring in its specification; and (2) add Put ϕ
at the correct position, which requires an additional parameter in the projection
to keep track what process is specified to be active and what is its postcondition.
To track the postcondition of the last active method of an object, we use a
partial function ac : O * SDL, to map objects to formulas. If no method was
active on X yet, ac is undefined, written ac(X) = ⊥. The projection of G on an
object X is denoted by G ac X. The selected projection rules for methods calls
and termination are given in Fig. 4. We write ac⊥ for
the function defined by
ψ
if X = X0
0
∀X. ac(X) = ⊥. For updates, we write ac[X 7→ ψ](X ) =
.
0
ac(X ) otherwise
10
main −
→ X2 : mhϕi.G ac X1 =?mhϕ@X2 i.(G ac[X2 7→ψ] X1 ) if ac = ac⊥ ∧ X2 = X1
main −
→ X2 : mhϕi.G ac X1 = skip.(G ac[X2 7→ψ] X1 ) if ac = ac⊥ ∧ X2 6= X1
x
→ X2 : mhϕ, ψi.G ac X1 = X2 !x mhϕi.(G ac[X2 7→ψ] X1 ) if ac(X1 ) 6= ⊥ ∧ ϕ = ϕ@X1
X1 −
?mhϕ@X2 i.(G ac[X2 7→ψ] X1 )
if ac(X2 ) = ⊥
x
X1 −
→ X2 : mhϕ, ψi.G ac X2 =
Put ac(X2 ).?mhϕ@X2 i.(G ac[X2 7→ψ] X1 ) if ac(X2 ) 6= ⊥
x
→ X2 : mhϕ, ψi.G ac X = skip.(G ac[X2 7→ψ] X) if X2 6= X 6= X1
X1 −
Read e.(G ac[X2 7→ψ] X) if ac(X1 ) 6= ⊥ ∧ X1 = X
(X1 ↑ e.G 0 ) ac X =
skip.(G ac[X2 7→ψ] X)
if X1 6= X
Put ac(X).End if ac(X) 6= ⊥
End ac X =
End
if ac(X) = ⊥
Fig. 4. The selected rules for projection on objects.
When projecting on caller X1 , a sending local type is generated if X1 if has
an active process (ac(X1 ) 6= ⊥) and the precondition can be proven by the caller
(ϕ = ϕ@X1 ). If the callee has an active process (ac(X2 ) 6= ⊥), then for this
process the termination type is added before the receiving type; if the callee is
specified as being inactive (ac(X2 ) = ⊥), then only the receiving type is added.
When projecting on any other object, skip is added. In any case, ac is updated
and maps the callee to a new postcondition. In the rest of paper, we omit ac⊥
and write just G X.
Propagation. In our concurrency model the heap does not change if no process is
active — all guarantees from the last active process still hold for the next process.
By propagation, formulas are added from the postcondition of one method to
the precondition of the next. Propagation moves formulas from where they must
hold to all points where they still are assumed to hold. Propagation replaces a
partial local type, if the partial type matches the given pattern.
Definition 10 (Propagation). The propagation function prp is defined via
term rewriting (denoted ) as follows. prp∗ denotes the fixpoint of rewriting.
Put ϕ.?mhψi
Put ϕ.?mhψ ∧ ϕ@Xi where X is the target object
Projection on Methods. The projection on a method, say L m0 m, results in a set
of method types. A method may have multiple methods types, as long as the
method types can are distinguishable. We call the method types in the result set
are distinguishable if they have non-overlapping preconditions, e.g., for any two
preconditions ϕ1 and ϕ2 , formula ¬(ϕ1 ∧ ϕ2 ) is valid.
The rules for projection on a method are straightforward and we refer to
Section 2 for an example. Full rules are given in Figure 5.
Definition 11 (Well-Formedness). A global type G is well-formed, if the projections on all methods are defined and all types of a method are distinguishable.
11
{(?mhϕi, m)} if m0 = ⊥
Put ϕ m0 m =
{(skip, m0 )} otherwise
{(Read e, m0 )} if m = m0
Read e m0 m =
{(skip, m0 )}
otherwise
00
0
{(X1 !v m hϕi, m )} if m = m0
X1 !v m00 hϕi m0 m =
{(skip, m0 )}
otherwise
?mhϕi m0 m =
skip m0 m = {(skip, m0 )}
(L1 .L2 ) m0 m =
{(Put ϕ, ⊥)} if m = m0
{(skip, ⊥)}
otherwise
End m0 m = {(skip, m0 )}if m0 = ⊥
{(L01 .L02 , m000 ) | (L01 , m00 ) ∈ L1 m0 ∧m00 6= ⊥ ∧ (L02 , m000 ) ∈ L2 m00 }
∪ {(L02 , m000 ) | (L01 , ⊥) ∈ L1 m0 ∧(L02 , m000 ) ∈ L2 ⊥ }
Fig. 5. Projection on Methods.
Semantics of Types as Constraints on Traces. To formalize the behavioral types
of the last section, we transform them into first-order constraints over traces.
We define C as a function transforming global types to constraints on traces.
Recall that we have defined C for configurations and ev for events. The primitive
C(i) references the ith configuration and ev(i) references the ith event in a trace.
We use events and formulas as colors and thus include futures, method names,
literals and object names in the domain. Constraints refer to SDL formulas ϕ
with C(i) |= ϕ, meaning that in the ith configuration, ϕ holds. The translation
of G1 .G2 models that there is a position i such that, the C(G1 ) holds on the
subtrace before i and the C(G2 ) holds on the subtrace after i. To construct such
a constraint, we use relativization [24], a syntactic restriction of a constraint γ
to a substructure described by another constraint χ.
Definition 12. Let χ be a constraint with a free variable x of data type T and
γ a constraint. We denote the relativization of γ with χ by γ[x ∈ T/χ]. For all
quantifiers of type T in constraint γ, relativization adds χ(x) as restrictions. The
main rule for relativization is (∃y ∈ T.γ)[x ∈ T/χ] = ∃y ∈ T.χ(y) ∧ γ[x ∈ T/χ].
The following example relativizes a constraint over natural numbers:
γ = ∀i ∈ N. isEven(i) ⇒ ∃j ∈ N. j < i
χ=k≥1
γ[k ∈ N/χ] = ∀i ∈ N. i ≥ 1 ⇒ isEven(i) ⇒ ∃j ∈ N. j ≥ 1 ∧ j < i
The rules for translating G into a constraint C(G) are defined as follows.
Definition 13 (Semantics of Global Types). Predicate res(i) holds if ev(i)
is a resolving event and A(i, X) holds if X is active in C(i).
.
C(main −
→ X2 : mhψi.G) = ∃j, k. ∃f. ∃e0 . ev(j) = iREv(X2 , f, m) ∧ C(j) |= ϕ@X2 ∧
.
0
ev(k) = fEv(X2 , f, e ) ∧ C(k) |= ψ ∧ ∀l.l 6= j ∧ l 6= k ⇒ res(l) ∧ C(G)
x
C(X1 −
→ X2 : mhϕ, ψi) = ∃i, j, k. ∃f. ∃e, e0 .
.
.
ev(i) = iEv(X1 , X2 , f, m, e) ∧ C(i) |= ϕ ∧ ev(j) = iREv(X2 , f, m) ∧ C(j) |= ϕ@X2 ∧
.
.
ev(k) = fEv(X2 , f, e0 ) ∧ C(k) |= ψ ∧ C(i) |= X1 .x = f ∧ ∀l. l 6= i ∧ l 6= j ∧ l 6= k ⇒ res(l)
C(X1 −
→ X2 : mhϕ, ψi) = ∃i, j, k. ∃f. ∃e, e0 .
12
.
.
ev(i) = iEv(X1 , X2 , f, m, e) ∧ C(i) |= ϕ ∧ ev(j) = iREv(X2 , f, m) ∧ C(j) |= ϕ@X2 ∧
.
0
ev(k) = fEv(X2 , f, e ) ∧ C(k) |= ψ ∧ ∧∀l. l 6= i ∧ l 6= j ∧ l 6= k ⇒ res(l)
.
.
C(X ↑ e) = ∃i. ∃f. ∃e0 . ∃X0 . ev(i) = fREv(X, X0 , f, e0 ) ∧ C(i) |= e = f ∧ ∀l. l 6= i ⇒ res(l)
^
C(G1 .G2 ) =
∃i ∈ N. C(G1 )[j ∈ N/A(j, X) ⇒ j < i] ∧ C(G2 )[j ∈ N/A(j, X) ⇒ j ≥ i]
X
C(End) = true
The constraint describing the call type is that the three events modeling (1)
call, (2) start of the process and (3) termination of the process exist and that
at the configurations for these events, the projected formulas hold. Every other
event is a fEv: The exact position of termination (i.e., fEv events) is not specified
in global types, so we do not constrain them. Reading from a location is defined
analogously, and sequential composition is described above.
Expressing a global type as a constraint on traces also enables us to connect
the types with SDL, because the semantics of SDL is defined in terms of traces.
Consider again Example 2: If the statement s in the modality is the body of a
method m with precondition ϕ and postcondition ψ, then the formula ϕ ⇒ [s]ψ
expresses that if ev(i) is the invocation reaction event of m and C(i) |= ϕ, then
at position j where the corresponding resolving event is, C(j) |= ψ holds.
We also give a translation of local types:
C(L1 . L2 ) = ∃i ∈ N. C(L1 )[n ∈ N/n < i] ∧ C(L2 )[n ∈ N/n ≥ i]
C(?mhϕi) = ∃i. ∀j. i = j ∧ ∃f. iREv(X, f, m) ∧ C(i) |= ϕ@X
.
C(X0 !x mhϕi) = ∃i. ∀j. i = j ∧ ∃f, e. iEv(X, X0 , f, m, e) ∧ C(i) |= ϕ@X ∧ C(i) |= (x = f )
C(X0 !mhϕi) = ∃i. ∀j. i = j ∧ ∃f, e. iEv(X, X0 , f, m, e) ∧ C(i) |= ϕ@X
C(Put ϕ) = ∃i. ∀j. i = j ∧ ∃f, e. fEv(X, f, e) ∧ C(i) |= ϕ@X
C(Read e) = ∃i. ∀j. i = j ∧ ∃f. fREv(X, X0 , f, e0 ) ∧ e = f
C(End) = true
5
Analysis
While our aim is to verify protocol adherence, we also verify deadlock freedom,
which additionally requires the Points-To analysis. Deadlock freedom is equivalent to cycle-freedom of causality graphs [17] in Active Objects. The causality
graph of a global type G is G(G) = (V, E). Each node L ∈ V is a singular partial
local type, and each edge (L1 , L2 ) ∈ E models that L2 must happen after L1 .
Definition 14 (Causality Graph). Let G be a well-formed global type. The
nodes of its causality graph G(G) are all local types derived from projecting G
on all endpoints of G. An edge (L1 , L2 ) is added if either (1) L1 .L.L2 is a partial
type for some L in some projection on some object in G or (2) L1 is a sending
type and L2 is a receiving type derived from the projection of a calling type.
13
Note that global types do not contain sufficient information to deduce all
causality, e.g., the causality of get statements can not be deduced from a global
type because synchronizations on futures are specified over locations. We use a
Points-To analysis for futures [17] instead. For generating a causality graph, we
first derive a partial causality graph from the global type, and then we apply
the Points-To analysis during type checking for the graph completion by deducing the missing edges. The Points-To analysis, defined below, determines what
methods are responsible to resolve the futures in a given expression.
Definition 15 (Points-To). Given an expression e in a program, the PointsTo analysis determines the set p2(e) of methods such that there is a reachable
configuration C = ob(X, f, ρ) prc(X, f, m(s), σ) C0 with JeKσ,ρ = f and m ∈ p2(e).
Whenever a e.get-statement is checked against a type Read e, edges between
the node of termination type of the methods which e can point to, and the node
of the current type Read e, are added. The graph on page 6 is admissible:
Definition 16 (Admissibility). A causality graph is admissible if (1) every
path is cycle-free and (2) for every object X, for any pair of receiving types of X,
there exists a connecting path without an edge of the form (Put ϕ, ?mhψi).
Type System and Analysis. Auxiliary SDL-formula post(X.m, ϕ) models that the
value in every future resolved by X.m satisfies ϕ, while formula Post(G) represents
the conjunction of all postconditions specified in G. Figure 6 gives the selected
typing rules that invokes validity calculus [15] and Points-To analysis.
Before introducing the typing rules, we define Roles(G) as the set of in G,
G(G)+E as the set of edges of G(G) and E (i.e., E are added into G(G)), term(m)
as the set of ↓ nodes of method m, and node(s) as the set of nodes referring to
the types that have typed s. We define three kinds of type judgments:
The Type Judgment for Programs. ` Prgm : G checks Prgm against global
type G. The well-formedness of G is ensured during type checking. Rule (TMain) checks that every endpoint in G is implemented in Prgm, the main block
makes the correct initializing call and checks each object against its object type.
The edges collected from the typing rules for objects are added to the partial
causality graph G(G) and the resulting graph is checked for admissibility.
The Type Judgment for Objects. Φ ` O : L B E checks whether O is well-typed
by L under a given E with Φ. E is a set of causality edges and Φ is a set of SDL
formulas. Rule (T-Object) projects L on all methods, checks each method mi by
L mi and collects all resulting edges.
The Type Judgment for Statements. Φ, s ` s : L B E checks whether s is welltyped by L under a given E with Φ, s. The environment s are the statements
type-checked so far. Whenever an SDL formula is checked, a validity check is
performed and s is added in the modality to consider the side-effects on the heap
memory so far. These are not recorded in E: The causality edges only record
14
(T-Main)
S
Oi = object Xi {. . . } iRoles(G) = {X1 , . . . , Xn }
G(G) + i≤n Ei admissible
∃j ≤ n. G = main −
→ Xj : mhϕi.G ∀i ≤ n. Post(G) ` Oi : prp∗ (G Xi ) B Ei
(T-Object)
` O1 . . . On main{Xj !m()} : G
∀i ≤ n. L ac mi =?mi hϕi i.Li S
∀i ≤ n. Φ, ϕi , skip ` si : Li B Ei
E = i≤n Ei
Φ ` object X{T1 m1 (T x){s1 } . . . Tn mn (T x){sn }
Φ ⇒ [s; return e]ϕ
(T-Return)
Φ, s ` return e : Put ϕ B E
(T-Call)
Φ, s; X!m(e) ` s0 : L B E
Φ ⇒ [s; X!m(e)]ϕ
(T-Call-2)
Φ, s; T x = X!m(e) ` s0 : L B E
Φ ⇒ [s; T x = X!m(e)]ϕ
Φ, s ` T x = X!m(e); s0 : X!x mhϕi.L B E
(T-Assign)
Φ, s ` X!m(e); s0 : X!mhϕi.L B E
T x = e} : L B E
Φ, s; T x = e ` s0 : L B E 0
Φ, s ` T x = e; s0 : L B E
Φ, s; T x = e.get ` s0 : L B E 0
E = E ∪ {(n, n )|∃m ∈ p2(e). n ∈ term(m) ∧ n0 ∈ node(s; e.get)}
0
(T-Get)
0
Φ, s ` T x = e.get; s0 : Read e.L B E
Fig. 6. The selected typing rules.
which method a get statement synchronizes on. Rule (T-Return) checks that after
executing all the type-checked statements, the return statement results in a state
where ϕ holds. Rule (T-Call) also checks the formula ϕ which describes the state
when the call has to be executed. Rule (T-Get) additionally executes the Points-To
analysis and adds all the edges as described in the previous section.
Theorem 1 (Deadlock Freedom and Protocol Adherence). Let Prgm be
a program and G be a global type. If Prgm is well-typed against G then (1) Prgm
does not deadlock and (2) every generated trace from Prgm satisfies C(G):
` Prgm : G → ∀tr. Prgm ⇓ tr → tr |= C(G)
Decidability and Types for Validation. The judgment ` Prgm : G is undecidable
if the validity of the FO logic used for specifying side-effects is undecidable.
A developer can choose an FOL fragment with decidable validity to trade off
between expressiveness and analyzability, e.g., if the developer of the analysis
chooses a more restricted fragment, which may limit the expressiveness of the
specification, then the validity of the FO logic used for specifying side-effects
may become decidable.
When using an undecidable FOL fragment, our approach can be used as a
validation tool to check whether the implemented (sub-)system will be behaving
as expected. Our approach is able to be integrated into the development process similarly as invariant-based approaches, and applies techniques proposed by
MPST to connecting global and local views of concurrent programs, a notoriously difficult problem when using contracts and invariants [15].
15
Protocol Adherence. Current work on MPST defines protocol adherence as a
fidelity theorem, which defines that every sequence of interactions in a session
follows the scenario declared in MPST [27] as follows: An operational semantics
for types is defined and it is shown that the semantics of the language is a refinement of the semantics of the types. Similarly, behavioral contracts [11] define
protocol adherence by compliance, which compares the interaction of contracts.
These are operational approaches to specification. We define protocol adherence
from a declarative perspective: We define a property that we require to hold for
all traces of a well-typed program. A declarative specification can be analyzed
with tools for logical specification, and can enable easier integration of other
static analysis tools (e.g., to consider state), since they are only required to have
a logical characterization or to be specified over traces.
6
Loops
In this section, we extend the syntax with loop and use a Kleene-star types (G)∗ϕ
(resp. (L)∗ϕ ), which repeats the type G (resp. L) finitely often. The formula ϕ is
a loop invariant and has to be satisfied whenever a loop iteration starts or ends.
Definition 17 (Syntax with Repetition).
s ::= . . . | while(e){s}
G ::= . . . | (G)∗ϕ .G
L ::= . . . | (L)∗ϕ . L
By syntactic restrictions, the local type of an object cannot have the form of
(L)∗ϕ .L0 , i.e., it is not allowed to start with a loop. The reason for this restriction is
that every loop has an invariant that an object must guarantee before executing
the next iteration. If an object is not active before the loop, it can not guarantee
the invariant in the very beginning.
In the following, we give the projection rules for loop.
Definition 18 (Projection Rules for Loops). The auxiliary predicate rcv (X, G)
holds if X is specified as being called in G.
Put ac(X).(L00 )∗ϕ@X .L0 if L 6= skip ∧ rcv (X, G) ∧ ac(X) 6= ⊥
∗
0
if L 6= skip ∧ ¬rcv (X, G) ∧ ac(X) 6= ⊥
(G)ϕ .G ac X = (L)∗ϕ@X .L0
0
L
if L = skip
Where G.End ac[X7→⊥] X = L00 ,G ac X = L and G 0 ac X = L0
if m00 = m0 ∧ L (m00 ) m0 6= {(skip, m000 )}
(L m00 m0 )∗ϕ , m00
∗
0
0
(L)ϕ m00 m = L (m00 ) m
if m00 6= m0 ∧ L (m00 ) m0 6= {(skip, m000 )}
00
{(skip, m )}
L (m00 ) m0 = {(skip, m000 )}
The first rule projects global types to object types. The first case is applied if
the object participates in the repetition of the inner type G by being repeatedly
called. The last active process must terminate first and the repeatedly called
method must terminate within the repetition. The termination inside the loop
16
is ensured by projecting the inner type with an appended End. The second case
is applied if the object participates with only one active process in the whole
repeated communication (L 6= skip ∧ ¬rcv (X, G) ∧ ac(X) 6= ⊥). Finally, the last
case removes the repetition if the object does not participate in it. Note that
there is no case where the object is repeatedly called, but not active before,
following the syntactic restriction discussed above.
The second rule projects object types to methods. The rule distinguishes
its cases by checking if the whole process is inside the repetition or not. If the
process is completely inside, the repetition is removed, as it is not visible to the
method.
Definition 19 (Rules for Propagation for Loops).
(1)
Put ϕ . (L)∗ψ
(3)
(L)∗ϕ
0
(4)
.
(L)∗ψ
(?m hϕi . L .
Put ϕ ∧ ψ . (L)∗ψ
(L)∗ϕ∧ψ
Put ϕ0 )∗ψ
.
(2)
(L)∗ψ . ?m0 hϕi
(L)∗ψ . ?m0 hϕ ∧ ψi
(L)∗ψ
0
(?m hϕ ∧ ψi . L . Put ϕ0 ∧ ψ)∗ψ
A fixpoint operator for propagation is used in presence of loops. Loop invariants have to hold before the first repetition, so the first rule ensures that
the last process before a repetition satisfies the invariant when terminating. The
second rule adds an invariant to the next process (for the same reason as the first
rule). The third rule is another case of the first one, in case two repetitions are
succeeding each other. Finally, the last rule adds the invariant to the processes
inside the repetition. This rule enables the type system to use the invariant in
the first method of the repetition and ensures that the last method reestablishes
the invariant.
For loops, first-order constraints, however, are not expressive enough. The
constraint with the Kleene star resemble regular languages and we thus use
monadic second order logic (MSO) to capture repetition. MSO extends firstorder logic with a quantifier ∃Y ⊆ Z which quantities over subsets of Z and a ∈
primitive to express membership of those sets.
Definition 20 (Translation into MSO).
C(Gϕ∗ ) = ∃X ⊆ N. ∃i, j ∈ X. ∀k ∈ N. i < k ≤ j ∧ ∀i ∈ X. C(i) |= ϕ∧
∀i, j ∈ X. ∀k ∈ X. k ≥ j ∨ k ≤ i ⇒ C(G) [n ∈ N/i < n ≤ j]
C(L∗ϕ ) = ∃X ⊆ N. ∃i, j ∈ X. ∀k ∈ N. i < k ≤ j ∧ ∀i ∈ X. C(i) |= ϕ∧
∀i, j ∈ X. ∀k ∈ X. k ≥ j ∨ k ≤ i → C(L) [x ∈ N/i < x ≤ j]
For both global and local types, repetition uses a set of boundary indices X,
between which the inner translation must hold plus constraints for the invariant.
We give the following typing rule for loops, which resembles loop invariant
rules from the Hoare or sequent calculi.
17
Definition 21 (Typing Rule for Loops).
(T-While)
ϕ ∧ Post(G), skip ` s : L B E
Φ ⇒ [s00 ]ϕ
ϕ ∧ Post(G) ⇒ [s]ϕ
ϕ ∧ Post(G), skip ` s : L B E 0
E = E 0 ∪ E 00
0
0
00
Φ, s00 ` while e {s}; s0 : (L)∗ϕ . L0 B E
The first premise continues the type checking of the program, in an environment where only the information in the invariant (and the global information
in Post, as defined in Section 5 is available. The second premise checks that the
invariant holds initially and the third premise checks that the invariant is preserved by the loop body. The forth premise checks the loop body and the last
premise combines the derived causality edges. The extension of the causality
graph is described in [29].
Example 4. Consider a big data analysis webtool like amazon cloud with a clientside GUI U and a server-side computational server S. U is a state-machine and
we model the following scenario: U sends data to the computational server. To
stay responsive, U ends its initial process, but is called repeatedly by the server
to update the progress. Whenever U is updated, the server also gets information
whether the GUI is hidden on the users screen or not. The sequence diagram in
Figure 7 illustrates the protocol.
S
comp
update
run
U
Fig. 7. Sequence Diagram for Example 4
Example 5. It is important to specify the behavior of the state of U. One needs
to ensure that during updating, U stays in a state expecting to receive updates
from the server: The field expect that saves the data is not Nil; A specification
is
∗
x
main −
→ G : runh>i . G −
→ S : comp . S −
→ G : update . S ↑ x
U.expect6=Nil
. end
The formula U.expect 6= Nil is the invariant, that the field expect is a nonempty list. The type S ↑ x describes that the server reads the return value from
18
the variable x and the previous action stores the future that identifies the call
to the update in x. Every read from futures must be annotated, as these are
critical for deadlock: If the process running start would attempt to read from
the future for comp, then the system would deadlock: U has an active process
waiting for S and S has an active process waiting for a call to update, which
can not be processed as U has already an active process. The local type for the
method update of U is
?updatehself .expect 6= Nili.Put self .expect 6= Nil
There is no repetition, because the repetition is only visible for the object, not a
single process. Thus the pre- and postcondition are derived from the invariant.
Similarly, there is no End, because the End type models the termination of the
protocol which again in visible for the object (no more incoming messages arrive)
but not for a single process.
Example 6. The code in Figure 8 implements the behavior of the GUI and the
Server from Example 4 and can be checked with the types from Example 5.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
object U{
List<Int> expect = Nil;
Unit run(List<Int> l){ S!comp(l); expect = l;}
Bool update(Int val, Int res){
return updateScreen(expect,val,res); //crashes if called in wrong state
}
}
object S{
Unit comp(List<Int> l){
Int i = 0;
while(i < length(l)){
Int res = compute(l[i]); //dummy
Fut<Bool> b = U!update(l[i],res);
b.get; i++;
}
}
}
Fig. 8. The code for Example 4, with loop.
7
Branching
Active Object have multiple ways to communicate the choice how to continue
the protocol and how an object reacts on it:
19
– (1) The choice is communicated via method selection, i.e., each branch corresponds to a different method call.
– (2) The choice is communicated via futures, i.e., other objects must react to
the choice of an object by reading its future.
– (3) The choice is communicated via the heap memory, i.e., processes must
behave according to some condition for the memory.
We aim to stick with standard imperative statements and must regard the
restriction that an if statement can only choose between two branches, while a
protocol may describe more than two. In our analysis of branching, choice is
communicated:
– (1) method calls and condition on the passed data for new process running
on other objects
– (2) the condition on future for already running processes running on (possibly) other objects.
– (3) via post-conditions to processes running later on the same object.
Definition 22 (Syntax with Branching).
s ::= . . . | if(e)s else s fi
G ::= . . . | X hϕi i, Xij hϕij i j∈J ; Gi
i∈I
L ::= . . . | ⊕ {Li }i∈I | &{X.mhϕi i; Li }i∈I
The global type X hϕi i, Xij hϕij i j∈J ; Gi i∈I describes that X chooses a branch
Gi . The formulas ϕi are additional postconditions for the choosing
process. Other
process can read the choice by reading this future. In Xij hϕij i j∈J , we describe
that the currently active process of Xij has the additional postcondition ϕij . The
local type ⊕{Li }i∈I is an active choice and &{X.mhϕi i; Li }i∈I is a passive choice.
The branch must be taken by reading the future from X.m and evaluating ϕi .
Definition 23 (Projection Rules for Branching). Given the ith branch
hϕi i, Xij hϕij i j∈J ; Gi , we denote the updated ac function with
aci = ac[X 7→ ac(X) ∧ ϕi @X][Xij 7→ ac(Xij ) ∧ ϕij @Xij ]j∈J
The auxiliary predicate allAct states that all mentioned objects and occur in all
branches are active and dist states that a set of formulas does not overlap.
^
allAct = ac(X) 6= ⊥ ∧ i∈I ac(Xij ) 6= ⊥ ∧ ∀i, i0 ∈ I. ∀j. Xij = Xi0 j
j∈J
dist({ϕ1 , . . . , ϕn }) = ∀i, j < n. i 6= j → (ϕi ∧ ϕj is unsatisfiable)
Figure 9 shows the projection rules for branching.
The projection rule from global to object-local types has four cases: the first
two are straightforward for the choosing process and the currently active reacting
processes. The third case handles objects which behave the same in all branches
and the forth handles objects which are active in only one. The projection on the
20
passive choice moves the Read type from its position after the choice in front
of it: The global type has no explicit point where a process terminates, thus
the read must be after the choice which adds the postcondition to the choosing
process. However the get statement must be before the if statement, which relies
on the read value in the guard.
⊕{Li }i∈I
0
&{X.mhϕ
i @X i; Li }
0
X hϕi i, Xij hϕij i j∈J ; Gi i∈I ac X = L
L
if
if
if
if
X = X0 ∧ allAct ∧ Gi aci X0 = Li
Xij = X0 ∧ allAct ∧ Gi aci X0 = Li
ac(X0 ) = ⊥ ∧ ∀i. Gi aci X0 = L
ac(X0 ) = ⊥ ∧ ∃i. Gi aci X0 = L
∧∀j 6= i. Gj acj X0 = skip
S
Li m0 m if m 6= m0
⊕{Li }i∈I m0 m = i∈I
⊕ {Li }i∈I if m = m0 ∧ ∀i ∈ I. Li m0 m = L0i
S
if m 6= m0
i∈I Li m0 m
Read e.&{X.mhϕi i; Li }i∈I if m = m0 ∧ dist((ϕ0i )i∈ I )∧
&{X.mhϕi i; Li }i∈I m0 m =
∀i ∈ I. Li m0 m = L0i = Read e.L00i
Fig. 9. Projection Rules for Branching
Definition 24 (Translation into MSO for Branching). For the translation
into MSO constraints, we use the auxiliary predicate firstTerm(i, X) that states
that the ith position in the trace refers to the first resolving event from X and the
auxiliary predicate lastTerm(i, X.m) that states that the ith position in the trace
refers to the last resolving event of X.m.
.
firstTerm(i, X) =∀j. ∃f. ∃m. ∃e. ev (j) = fEv(X, f, m, e) → i ≤ j
.
lastTerm(i, X.m) =∀j. ∃f. ∃e. ev (j) = fEv(X, f, m, e) → i ≥ j
Additionally to the translation of the branches, it encodes that the choosing process terminates before any process that relies on the communication of its choice
via the return value. The rules are as follows:
_
i∈I
C X hϕi i, Xij hϕij i j∈J ; Gi i∈I =
^
C(Gi ) ∧ ∃k. firstTerm(k, X)∧
∃kj . firstTerm(kj , Xij ) ∧ k ≥ kj ∧ σ(h)[kj ] |= ϕij
j∈J
C(⊕{Li }) =
_
C(&{X.Mhϕi i; Li }) =
_
C(Li )
i
i
∃j ∈ N. lastTerm(j, p.m) ∧ σ(j) |= ϕi ∧ C(Li )
In the following we present the rules for branching. The typing rules split
the branches into two disjoint sets and shows that the guard of the if statement together with the added choice-conditions of the branch selects the correct
21
continuation of the type. Once the sets of branches are singletons, the choice
operators can be removed.
Definition 25 (Typing Rules).
(T-Offer)
I = I1 ∪ I2
I1 ∩ I2 = ∅
E = E1 ∪ E2
∀i ∈ I1 . Φ ∧ post(X.m, ϕi ) ⇒ e)
∀i ∈ I2 . Φ ∧ post(X.m, ϕi ) ⇒ ¬e)
W
Φ; e; Wi∈I1 post(X.m, ϕi ), s ` s0 ; s000 : &{X.mhϕi i; Li }i∈I1 B E1
Φ; ¬e; i∈I2 post(X.m, ϕi ), s ` s0 ; s000 : &{X.mhϕi i; Li }i∈I2 B E2
Φ, s ` if e then s0 else s00 fi ; s000 : &{X.mhϕi i; Li }i∈I B E
(T-Offer-Single)
(T-Select-Single)
Φ, s ` s : L B E
Φ, s0 ` s : L B E
0
Φ, s0 ` s : &{X.m : ϕ; L} B E Φ, s0 ` s : ⊕{L} B E
I = I1 ∪ I2
I1 ∩ I2 = ∅
E = E1 ∪ E2
Φ; e, s ` s ; s : ⊕{Li }i∈I1 B E1
Φ; ¬e, s ` s00 ; s000 : ⊕{Li }i∈I2 B E2
0
(T-Select)
000
Φ, s ` if e then s0 else s00 fi ; s000 : ⊕{Li }i∈I B E
The extension of the causality graph is described in [29].
We use the following example to illustrate how we handle branching.
Example 7. Consider the scenario: Client X1 wants to access data on server X2
and sends its login data by calling method acc. Then X2 decides. If the login
data is invalid, X2 logs the denied access by calling logging server S and returns
−1 to X1 ; if the access succeeds, it returns the data, a value > 0, to X2 . X1
reacts on the return value and returns a boolean indicating whether the access
was successful. This is formalized by the following type:
x
main −
→ X1 : start . X1 −
→ X2 : acc.X2
.
→ S : log.End
hresult = −1i X1 h¬resulti ; X1 ↑ x.X2 −
hresult > 0i X1 hresulti ; X1 ↑ x.End
The local type for X1 .start is the following. Note that the Read type is now
before the branching.
?start.X2 !x acc.Read x.&
8
.
X2 .acchresult = −1i ; Put ¬result
X2 .acchresult > 0i ; Put result
Conclusion and Related Work
In this paper, we propose a two-phase analysis, which gives developers a means
to specify and verify protocols where information is not only transmitted via
asynchronous method calls but also through the heap memory of Active Objects;
and the model-theoretic semantics, which give a declarative definition of protocol
adherence. The following works are related:
22
Actors and Objects. Our previous type system for Active Objects [30] cannot
specify and verify state and future passing. Furthermore, we do not need to check
the scheduler in this work. Crafa and Padovani [12,34] investigate behavioral
types for the object-oriented join calculus with typestate, a concurrency model
similar to actors, but with an encoding for futures. Gay et al. [18] model channels as objects, integrating MPST with classes; Dezani-Ciancaglini et al. [13] use
MPST in the object-oriented language MOOSE, where types describe communication through shared channels. We ensure deadlock freedom with the approach
similar to Giachino et al. [20,21], who ensure deadlock freedom by inferring behavioral contracts and applying a cycle detection algorithm; however, they do
not consider protocol adherence.
State and Contracts. In an non-object-oriented setting, Bocchi et al. [4,5,6] develop a MPST discipline with assertions for endpoint state. The specifications
use global values in global types and require complex checks for history-sensitivity
and temporal-sensitivity to ensure that an endpoint proves its obligations. They
explicitly track values over several endpoints, while we implicitly do so by equations over locations. In a stateless setting, Laneve and Padovani [31] combine contracts and behavioral types, and Toninho and Yoshida use dependent MPST [37]
to reason about passed data.
Logics. Session types as formulas have been examined by Caires et al. [9] and
Carbone et al. [10] for intuitionistic and linear logics as types-as-proposition for
the π-calculus. In our work, we examine a dynamic logic combines FO-formulas,
called SDL, over the heap with symbolic executions of statements, and examine
a declarative formulation of protocol adherence.
23
References
1. Anand, S., and Harrold, M. J. Heap cloning: Enabling dynamic symbolic
execution of java programs. In ASE’11 (2011), IEEE Computer Society, pp. 33–
42.
2. Ancona, D., Bono, V., and Bravetti, M. Behavioral Types in Programming
Languages. Now Publishers Inc., Hanover, MA, USA, 2016.
3. Baker, H. G., and Hewitt, C. The incremental garbage collection of processes.
SIGART Newsletter 64 (1977), 55–59.
4. Bocchi, L., Demangeon, R., and Yoshida, N. A multiparty multi-session
logic. In TGC’12 (2012), vol. 8191 of LNCS, Springer, pp. 97–111.
5. Bocchi, L., Honda, K., Tuosto, E., and Yoshida, N. A theory of design-bycontract for distributed multiparty interactions. In CONCUR’10 (2010), vol. 6269
of LNCS, Springer, pp. 162–176.
6. Bocchi, L., Lange, J., and Tuosto, E. Three algorithms and a methodology
for amending contracts for choreographies. Sci. Ann. Comp. Sci. 22, 1 (2012),
61–104.
7. Boer, F. D., Serbanescu, V., Hähnle, R., Henrio, L., Rochas, J., Din,
C. C., Johnsen, E. B., Sirjani, M., Khamespanah, E., Fernandez-Reyes,
K., and Yang, A. M. A Survey of Active Object Languages. ACM Comput.
Surv. 50, 5 (2017), 76:1–76:39.
8. Cadar, C., and Sen, K. Symbolic execution. In RV’12 (2012), vol. 7687 of
LNCS, Springer, p. 2.
9. Caires, L., and Pfenning, F. Session types as intuitionistic linear propositions.
In CONCUR’10 (2010), vol. 6269 of LNCS, Springer, pp. 222–236.
10. Carbone, M., Lindley, S., Montesi, F., Schürmann, C., and Wadler, P.
Coherence generalises duality: A logical explanation of multiparty session types.
In CONCUR’16 (2016), vol. 59 of LIPIcs, Schloss Dagstuhl - Leibniz-Zentrum fuer
Informatik, pp. 33:1–33:15.
11. Castagna, G., Gesbert, N., and Padovani, L. A theory of contracts for web
services. ACM Trans. Program. Lang. Syst. 31, 5 (2009), 19:1–19:61.
12. Crafa, S., and Padovani, L. The chemical approach to typestate-oriented programming. ACM Trans. Program. Lang. Syst. 39, 3 (2017), 13:1–13:45.
13. Dezani-Ciancaglini, M., Drossopoulou, S., Mostrous, D., and Yoshida,
N. Objects and session types. Information and Computation 207, 5 (2009), 595 –
641.
14. Din, C. C., Bubel, R., and Hähnle, R. Key-abs: A deductive verification tool
for the concurrent modelling language ABS. In CADE’15 (2015), vol. 9195 of
LNCS, Springer, pp. 517–526.
15. Din, C. C., and Owe, O. A sound and complete reasoning system for asynchronous communication with shared futures. J. Log. Algebr. Meth. Program. 83,
5-6 (2014), 360–383.
16. Din, C. C., Tarifa, S. L. T., Hähnle, R., and Johnsen, E. B. History-based
specification and verification of scalable concurrent and distributed systems. In
ICFEM’15 (2015), vol. 9407 of LNCS, Springer, pp. 217–233.
17. Flores-Montoya, A., Albert, E., and Genaim, S. May-happen-in-parallel
based deadlock analysis for concurrent objects. In FMOODS/FORTE’13 (2013),
vol. 7892 of LNCS, Springer, pp. 273–288.
18. Gay, S. J., Gesbert, N., Ravara, A., and Vasconcelos, V. T. Modular
session types for objects. Logical Methods in Computer Science 11, 4 (2015).
24
19. Gay, S. J., Vasconcelos, V. T., Wadler, P., and Yoshida, N. Theory and
applications of behavioural types (dagstuhl seminar 17051). Dagstuhl Reports 7, 1
(2017), 158–189.
20. Giachino, E., Henrio, L., Laneve, C., and Mastandrea, V. Actors may
synchronize, safely! In PPDP’16 (2016), ACM, pp. 118–131.
21. Giachino, E., Laneve, C., and Lienhardt, M. A framework for deadlock
detection in coreabs. Software & Systems Modeling 15, 4 (Oct 2016), 1013–1048.
22. Halstead Jr., R. H. Multilisp: a language for concurrent symbolic computation.
ACM TOPLAS 7, 4 (1985), 501–538.
23. Harel, D. First-Order Dynamic Logic. Springer-Verlag New York, Inc., Secaucus,
NJ, USA, 1979.
24. Henkin, L. Relativization with respect to formulas and its use in proofs of independence. Compositio Mathematica 20 (1968), 88–106.
25. Henrio, L., Laneve, C., and Mastandrea, V. Analysis of synchronisations in
stateful active objects. In iFM’17 (2017), vol. 10510 of LNCS, Springer, pp. 195–
210.
26. Hewitt, C., Bishop, P., and Steiger, R. A universal modular ACTOR formalism for artificial intelligence. In Proceedings of the 3rd International Joint Conference on Artificial Intelligence (1973), IJCAI’73, Morgan Kaufmann Publishers
Inc., pp. 235–245.
27. Honda, K., Yoshida, N., and Carbone, M. Multiparty asynchronous session
types. J. ACM 63, 1 (Mar. 2016), 9:1–9:67.
28. Johnsen, E. B., Hähnle, R., Schäfer, J., Schlatte, R., and Steffen, M.
ABS: A core language for abstract behavioral specification. In FMCO’10 (2010),
pp. 142–164.
29. Kamburjan, E. Session Types for ABS. Tech. rep., TU Darmstadt, 2016.
http://formbar.raillab.de/en/techreportsessiontypesabs/.
30. Kamburjan, E., Din, C. C., and Chen, T. Session-based compositional analysis
for actor-based languages using futures. In ICFEM’16 (2016), vol. 10009 of LNCS,
Springer, pp. 296–312.
31. Laneve, C., and Padovani, L. The Pairing of Contracts and Session Types. In
Concurrency, Graphs and Models (Ugo65’08) (2008), vol. 5065 of LNCS, Springer,
pp. 681–700.
32. Lavender, R. G., and Schmidt, D. C. Pattern languages of program design 2.
Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1996, ch. Active
Object: An Object Behavioral Pattern for Concurrent Programming, pp. 483–499.
33. Odersky, M et al. Scala programming language. http://www.scala-lang.org.
34. Padovani, L. Deadlock-Free Typestate-Oriented Programming. Submitted to
The Art, Science, and Engineering of Programming, preprint available under
https://hal.archives-ouvertes.fr/hal-01628801/file/main.pdf, Nov. 2017.
35. Schmitt, P. H., Ulbrich, M., and Weiß, B. Dynamic frames in Java dynamic
logic. In FoVeOOS’10 (2011), vol. 6528 of LNCS, Springer, pp. 138–152.
36. Tasharofi, S., Dinges, P., and Johnson, R. E. Why do scala developers mix
the actor model with other concurrency models? In ECOOP’13 (2013), vol. 7920
of LNCS, Springer, pp. 302–326.
37. Toninho, B., and Yoshida, N. Certifying data in multiparty session types. J.
Log. Algebr. Meth. Program. 90 (2017), 61–83.
25
A
Soundness
The proof for Theorem 1 is similar to the proof for Theorem 2 in [29], we thus
only give a sketch and point out where the proofs differ.
Propagation
First, we state the correctness of propagation. Let tr X be the projection of
trace tr on X, i.e., tr X results from tr by replacing all events not issued by X.
Lemma 1. Let Prgm be a program and G a type for Prgm. If in all traces
produced by Prgm, the order of invocation events is the same, then every trace
that satisfies the translation of the propagated type iff it satisfies the translation
of the original type:
` Prgm : G → ∀tr. ∀X. Prgm ⇓ tr → tr X |= C(prp∗ (G X)) ↔ tr X |= C(G X)
Proof. We fix X and denote G X with L. We show this by induction on the
number n of applications of prp for the fixpoint.
Induction Base, n = 0 Then prp∗ (L) = L and the lemma holds trivially.
0
Induction Step, n = n0 + 1 By induction hypothesis there is a type L0 = prpn (L)
such that the desired property holds. We make a case distinction on the applied case in the definition of prp in its last application:
- Case 1 Put ϕ.?mhψi
Put ϕ.?mhψ ∧ ϕ@Xi
In this case we have to show that the start of execution of method m the
formula ψ as holds. Let m0 be the method whose termination action Put ϕ
is responsible for ϕ. By assumption, the order of invocation events is fixed
and the program can be typed. Thus, there is no trace such that between
the invocation action of m0 and m, there is another invocation event. Thus,
each trace tr that contains pairs of the form (iREv(X, f, m), C), for some
f, C s.t. C |= ψ contains this pair as part of a subtrace of the following
form:
h
i
fEv(X, f 0 , m0 , e0 ), C0 , iREv(X, f, m), C
for some f 0 , C s.t. C |= ϕ. Every state change on X must be executed by
some process on X, but as there is no such such process (as there would
be an invocation event for it) between the two events in the subtrace, the
state-part of ϕ, i.e. ϕ@X, still holds at the invocation event: C |= ϕ@X.
This is exactly the condition captured by this propagation case.
- Case 2 Put ϕ . (L)∗ψ
Put ϕ ∧ ψ . (L)∗ψ
By the definition of C((L)∗ ) there is a set of indices X in every trace,
such that every such position i ∈ X, the invariant holds and for every
pair of consecutive positions i, j ∈ X, the subtrace tr[i..j] satisfies C(L).
Now, Put ϕ is the last event before the repetition, thus before the very
first position i0 ∈ X there is a pair
tr[i0 − 1] = fEv(X, f, m, e), C
26
In C, no process is active at X. We only regard traces produced Prgm,
thus tr is well-formed1 and i0 must be a invocation reaction event
tr[i0 ] = iREv(X, f 0 , m0 ), C
Such that C |= ψ. With the same argument as above, the condition at ψ
must hold at i0 − 1, as there was no process who could have changed it.
Note, that i0 6= 0 as every local type starts with a receiving action, not
a repetition.
- Case 3 (L)∗ψ . ?m0 hϕi
(L)∗ψ . ?m0 hϕ ∧ ψi
This case is analogous to case 2.
(L)∗ϕ∧ψ . (L)∗ψ
- Case 4 (L)∗ϕ . (L)∗ψ
This case is analogous to case 2.
- Case 5 (?m0 hϕi . L . Put ϕ0 )∗ψ
(?m0 hϕ ∧ ψi . L . Put ϕ0 ∧ ψ)∗ψ
∗
By the definition of C((L) ), there is a set of indices X in every trace,
such that every such position i ∈ X, the invariant holds and for every
pair of consecutive positions i, j ∈ X, If the repetition start with a
receiving action and ends with a termination, the chosen positions are
those of the termination actions and the first action before (Again, the
syntactic form guarantees such a position). This reduces this case to
show that the same propagation as in case 1 holds and thus technical
details are analogous to case 1.
Main Theorem
Given a well-formed global type G we can say that another (possibly not wellformed) global G0 is a prefix of G if we can extend G0 to G by concatenating
another global type:
G0 v G ⇐⇒ G = G0 .G
And similarly for local types L and traces tr.
The main lemma is similar to subject reduction in non-model-theoretic semantics for types, as it connects types and operational semantics of the language.
It states that each step in the execution preserves the property that the trace so
far is a prefix of a a trace which is a model for the type.
Lemma 2. Let Prgm be a program and G a well-formed type with ` Prgm : G.
Every prefix of every trace of Prgm satisfies the translation of a prefix of G:
` Prgm : G → ∀tr. Prgm ⇓ tr → ∀tr0 . tr0 v tr → ∃G0 . G0 v G∧tr0 |= C(G0 )
The property that the whole execution of the program satisfies the translation
of the whole type and not some prefix follows from well-formedness of global
types (Theorem 1 in [29]) and deadlock freedom. The main differences to the
proof in [29] are the following:
1
The well-formedness of traces is defined in [15,29]
27
All assumed conditions hold at the point they are used We distinguish
between the following kinds of assumed conditions:
– The precondition at method start. The precondition is a conjunction
ϕ1 ∧ ϕ2 where ϕ1 is resulting from the projection and ϕ2 from the propagation. That ϕ2 holds follows from Lemma 1. That ϕ1 holds follows
from the fact that the precondition is projected from a formula ψ in the
global call, which is fully proven by the caller, checked in rule (T-Call)
with the condition that ψ is equal to its projection on the caller.
– The selection condition of the passive choice. It must connect that the
additional condition executes the statement branch which is typed with
the corresponding type branch. This follows directly from the two additional promises of (T-Offer) with respect to (T-Select).
Methods are executed in the right order Assume there is a method m1 that
is executed on some object X before m2 in the type, but executed the other
way around in the generated trace. Then m2 does not depend on m1 . But by
assumption the program has been typed. Rule (T-Main) checks, however, that
the start of m1 causes m2 in its admissibility check and this means that m2
cannot be executed before m1 (Lemma 36 in [29]).
Deadlock Freedom A deadlocked configuration is a configuration which is not
terminated, yet cannot continue execution. First we observe that every deadlock is caused by processes blocking at get statements. It cannot be a single
process, because a process has no access on its own future. It can also not be
stored in the heap between call and execution start, as this would mean that
another method was active to store it and this would violate the condition
that all methods are executed in the right order shown above.
Assume there would be a deadlock. W.l.o.g. we assume that only two processes are involved, p1 executing m1 and p2 executing m2 . If p1 blocks while
attempting to read a future belonging to m2 then the Points-To analysis will
include m2 in the set of possible method in rule (T-Get). As m1 has been type
checked, this mean that an edge from the corresponding termination to the
corresponding read would be added to the causality graph. The same holds
for m2 . The termination of mi is in the same object type as the reading type,
thus there is a path from the read in mi to the termination in mi . The resulting
graph is pictured below and contains a cycle. The absence of cycles is however checked in rule (T-Main). For a full formalization of deadlocks through
causality graphs, we refer to [17].
m2
...
↑
...
↓
m1
...
↑
...
↓
28
| 6cs.PL
|
arXiv:1712.08626v1 [cs.AI] 22 Dec 2017
Obtaining Accurate Probabilistic Causal Inference by
Post-Processing Calibration
Fattaneh Jabbari1 , Mahdi Pakdaman Naeini2 , Gregory F. Cooper1,3
1
Intelligent Systems Program, University of Pittsburgh, Pittsburgh, PA
2
Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA
3
Department of Biomedical Informatics, University of Pittsburgh, Pittsburgh, PA
[email protected], [email protected], [email protected]
Abstract
Discovery of an accurate causal Bayesian network structure from observational
data can be useful in many areas of science. Often the discoveries are made under
uncertainty, which can be expressed as probabilities. To guide the use of such
discoveries, including directing further investigation, it is important that those
probabilities be well-calibrated. In this paper, we introduce a novel framework to
derive calibrated probabilities of causal relationships from observational data. The
framework consists of three components: (1) an approximate method for generating
initial probability estimates of the edge types for each pair of variables, (2) the
availability of a relatively small number of the causal relationships in the network
for which the truth status is known, which we call a calibration training set, and
(3) a calibration method for using the approximate probability estimates and the
calibration training set to generate calibrated probabilities for the many remaining
pairs of variables. We also introduce a new calibration method based on a shallow
neural network. Our experiments on simulated data support that the proposed
approach improves the calibration of causal edge predictions. The results also
support that the approach often improves the precision and recall of predictions.
1
Introduction
Much of science consists of discovering and modeling causal relationships in nature. Increasingly,
scientists have available multiple complex data and a large number of samples, each of which has
an enormous number of measurements recorded, thanks to rapid advancements in sophisticated
measurement technology, where this data are often purely observational. In past 25 years, there has
been tremendous progress in developing computational methods for discovering causal knowledge
from observational data [11, 24, 19, 23, 12]. A primary use of such methods is to analyze observational
data to generate novel causal hypotheses that are likely to be correct when subjected to experimental
validation; such an approach can significantly increase the efficiency of causal discovery in science.
To make informed decisions about which novel causal hypotheses to investigate experimentally,
scientists need to know how likely the hypotheses are to be true. In probabilistic terms, this means
they need to have the probabilities of the hypotheses (as output by a causal discovery algorithm) be
well-calibrated. Informally, we say that probabilities are well-calibrated if events predicted to occur
with probability p do occur about p fraction of the time, for all p [4]. In general, it is important to use
calibrated probabilities when making decisions using decision theory.
In this paper, we focus on the discovery of causal Bayesian network (CBN) structure from observational data. In particular, we focus on the discovery of the causal relationships (edge types)
between pairs of measured variables. If a causal arc is novel and important, it may be worthwhile to
experimentally investigate it. The extent to which it is worth doing so depends in part on how high is
A NIPS 2017 Workshop: From ’What If?’ To ’What Next?’, Long Beach, CA, USA.
the calibrated probability that the causal arc is present. We introduce a method to calibrate edge type
probabilities in CBNs with thousands of measured variables and arbitrarily many latent variables.
The method requires the following components: (1) a method for generating initial probability
estimates of the edge types for each pair of variables; in general those estimates need not be wellcalibrated, (2) the truth status of a small unbiased sample of the causal relationships in the network,
which we call a calibration training set, and (3) a calibration method for using the uncalibrated
probability estimates and the calibration training set to generate calibrated probabilities for the large
number of remaining pairs of variables.
We use a bootstrapping method [8] for generating probability estimates of edge types. This method
resamples a dataset n times with replacement and learns a model for each dataset. In particular, for
each dataset we use the Really Fast Causal Inference (RFCI) algorithm [3] to estimate the underlying
generative network when allowing for the possibility of latent confounders. For any given pair of
nodes (A, B), the probability they have a given edge type (e.g., A → B) is estimated as the fraction
of that edge type for (A, B) in the n networks. Previously, researchers have successfully applied this
approach for estimating the probabilities of edge types in Bayesian networks [9]. These bootstrap
estimates are not guaranteed to represent calibrated posterior probabilities, however, even in the
large sample limit of the number of bootstrap samples. A key reason is that heuristic search, while
practically necessary, may get stuck in local maxima. Thus, there is a need to map those estimates to
calibrated probabilities, which is the focus of the current paper.
The bootstrapping approach described above provides empirical estimates of edge-type posterior
probabilities for both constraint-based (e.g., PC, FCI [23], and RFCI [3]) and Bayesian structure
learning algorithms (e.g., GES [2]). Bayesian model averaging [15, 10, 14, 7, 13] provides an
alternative approach for estimating edge probabilities. However, such Bayesian methods are typically
applicable when using datasets in which the number of random variables is in the double digits (for
exact search methods) to triple digits (for heuristic search methods). In contrast, we are interested
in providing calibrated estimates of edge probabilities for datasets that may contain thousands of
variables, as typically encountered with modern biological data. We also note that Bayesian model
averaging methods are sensitive to the method applied for heuristic search [9] and to the structure
and parameter priors that are used, even if they are non-informative. Consequently, their generated
probabilities are still subject to possibly being uncalibrated. Finally, there are no computationally
tractable Bayesian methods for discovering CBNs that contain more than a few latent confounders;
in contrast, constraint-based methods exist that can perform discovery of CBNs with hundreds (or
more) latent variables on datasets with thousands of variables in a feasible amount of time [3].
We assume the availability of a calibration training set that allows us to induce a mapping from
bootstrap probability estimates to calibrated posterior probabilities. The training set should contain
the truth status for the subset of edge types. In the domain of biomedical applications, the truth status
might come, for example, from results published in the literature. We emphasize that the calibration
training set can be very small, relative to the number of total node pairs. In the experiments we
performed, it consists of less than 0.02% of all the node pairs. Using it, our goal is to generate better
calibrated probabilities for the remaining 99.98% of node pairs. In an application using biomedical
data, for example, a biomedical scientist who chose to experimentally test causal relationships that
have high probabilities (i.e., close to 1) that are well-calibrated could be confident that the experiments
would usually corroborate those relationships. We introduce a new neural-network-based calibration
method that uses the calibration training set to construct a mapping from bootstrap probability
estimates to calibrated posterior probabilities of edge types for all node pairs in a CBN (except those
few that are used for training). We apply that mapping to all of those node pairs.
In this paper, we use simulated data to investigate two main questions1 . First, how calibrated are
the bootstrap-derived probabilities of edge types? Second, how calibrated are the probabilities
produced by our neural-network-based calibration method? Given a finite calibration training set,
the latter method is not guaranteed to always output perfectly calibrated probabilities either. Our
main hypothesis in this paper is that this calibration method will output probabilities that are better
calibrated than are the bootstrap probabilities, while being at least as discriminative in terms of
measures such as precision, recall, and F1 score.
1
Note that it is difficult to obtain gold standards for the causal relationships among the variables in large observational datasets. As a result,
the use of simulated data is important and commonly done to evaluate causal discovery methods.
2
2
Method
In this section we briefly describe the RFCI search, bootstrap RFCI, and calibration model.
2.1
Overview of RFCI
Colombo et al.[3] developed an algorithm called Really Fast Causal Inference (RFCI), which identifies
the causal structure of the data-generating process in the presence of latent variables using Partial
Ancestral Graphs (PAGs) as a representation. A PAG encodes a Markov equivalence class of Bayesian
networks (possibly with latent variables) that exhibit the same conditional independence relationships.
RFCI has two stages: (1) adjacency search: this involves a selective search for the (in)dependencies
among the measured variables, (2) orientation phase: this orients the endpoints among pairs of nodes
that are connected according to the first stage. As is typical of constraint-based causal discovery
algorithms, RFCI outputs a single graph structure (PAG) and does not provide any information about
the uncertainty of the edges between the nodes in the structure.
2.2
Bootstrap RFCI
Considering the PAG generated by RFCI, it is possible to partition all pairs of nodes (A, B) into the
following seven classes: (1) A · · · B: there is no edge between A and B; (2) A → B: a directed edge
from A to B means that A is a direct or indirect cause of B; (3) B → A: this is similar to (2); (4) A
◦→ B: this edge type indicates either A is a cause of B, there is an unmeasured confounder of A
and B, or both; (5)B ◦→ A: this is similar to (4); (6) A◦–◦ B: this edge type expresses that A is a
cause of B, B is a cause of A, there is an unmeasured confounder of A and B, or that there is an
unmeasured confounder and one of those two causal relationships holds; (7) A ↔ B: a bi-directed
edge between A and B represents the presence of an unmeasured confounder of A and B.
The bootstrap RFCI (BRFCI) method that we apply has three main steps. First, it performs bootstrap
sampling over the training data n times (n = 200 in our experiments) to create n different bootstrap
training datasets. In the second step, it runs RFCI on each of n datasets to obtain n PAGs. Finally, for
every pair of nodes, it uses the frequency counts of each edge class for that pair over the generated
PAGs to determine a probability distribution for the seven possible edge classes. As mentioned,
these bootstrap estimates are not guaranteed to be calibrated. In the following section, we describe a
post-processing method to map the bootstrap probabilities to calibrated probabilities.
2.3
Calibration Model
For a pair of nodes (A, B), the resulting output of the BRFCI method will be seven jointly exhaustive
and mutually exclusive class probabilities that correspond to the seven classes described above.
Therefore, we need to apply a calibration method that post-processes a multi-class classification score
(in our case seven classes). One simple approach to devise such a multi-class calibration model is to
use a well-performing non-parametric binary classifier calibration method such as isotonic regression
[25], averaging over Bayesian binning (ABB) [16], or Bayesian binning into quantiles (BBQ) [17] to
post-process the corresponding output probabilities of each class separately. This is performed in
a one-versus-remainder fashion as described in [25]. The major drawback of this approach is that
such binary calibration methods are histogram-based non-parametric and they require a considerable
amount of data to produce well-calibrated probabilities. However, it is often too expensive or not
feasible to obtain the truth status for a large number of node pairs in real applications of causal
discovery. Consequently, the availability of only a small calibration training set is a critical constraint
in the design of the calibration approach.
To resolve this problem, we make a simple extension to Platt’s method [20], which is a parametric
binary classifier calibration approach. Platt’s method uses a sigmoid transformation to map the output
of a binary classifier into a calibrated probability. It then uses a logistic loss function to learn the two
parameters of the model. The method has two advantages: (1) it has only two parameters that make it
a viable choice for low sample size calibration datasets, and (2) the method runs in O(1) at test time,
and thus, it is fast. A natural extension to Platt’s method for the multi-class calibration task is to use a
combination of a softmax transfer function and a cross-entropy loss function instead of a sigmoid
function and a logistic loss function, respectively. Minimizing the cross entropy is equivalent to
minimizing the empirical Kullback-Leibler divergence of the estimated probabilities and the observed
3
Σ
P(cal
A→B)
...
P(B→A)
P( A!→B)
Σ
Σ
P(B!→A)
P( A!−!B)
P( A↔B)
Σ
Σ
Σ
cal
P(B→A)
P(cal
A!→B)
cal
P(B!→A)
P(cal
A!−!B)
P(cal
A↔B)
Figure 1: The structure of post-processing calibration method.
The inputs on the left are the bootstrap probabilities for seven
edge types. The outputs on the right are the corresponding postprocessed probabilities that are intended to be better calibrated.
V = 2000, E = 2000
V = 2000, E = 4000
800
P(cal
A... B)
Σ
P( A→B)
Softmax
P( A... B)
Output Layer
Number of nodes
Hidden Layer
600
400
200
0
0
2
4
6 8 10 12 14
Degree
Figure 2: Parent size distribution for simulated CBNs with V = 2000 nodes and
E = 2000 or 4000 edges.
ones. The minimum will be achieved by the true probability distribution and minimizing the cross
entropy function will result in finding the closest distribution parameterized by the model to the
observed distribution of data [18].
The model that uses the softmax transfer function and optimizes the cross entropy loss function is
called softmax regression [18]. The softmax regression-based calibration model inherits the desirable
properties of Platt’s method. However, similar to Platt’s method, the mapping that the softmax
regression-based calibration method can learn is restrictive since the final separating boundaries
between each pair of classes are always linear. A simple relaxation of this restriction is to use a
shallow neural network with one hidden layer. Figure 1 shows the architecture of the shallow neural
network model that we use to post-process the bootstrap-generated probabilities.
In our experiments, we train 10 different such shallow neural networks by setting the number of
neurons in the hidden layer to be 4, 5, 6, or 7 randomly. At test time, we use the average of the 10
different outputs generated by these models as the final calibrated probability estimates. The averaging
is helpful since it reduces the variance error of the predictions and improves the final performance
of the post-processed probabilities [5]. We will use the notation fcal (p1, p2, p3, p4, p5, p6, p7) to
denote the mapping from a seven-element vector of uncalibrated probabilities that are input to a
seven-element vector of calibrated probabilities that are output. We implemented our model using the
scikit flow Python package2 , which uses the tensorflow machine learning package [1]. We used the
cross-entropy loss function and the adagrad optimization method [6] to learn the parameters; we set
the learning rate and the batch size to be 0.1 and 10, respectively.
3
Experimental Methods
This section describes the experimental methods that we used to evaluate the performance of the
calibrated network discovery method introduced above. The evaluation involves the following steps:
1. Create a random causal Bayesian network, BN , with V real-valued nodes and E edges. We set V
to be 1000 and 2000. We also set the average number of edges per node to be 1 and 2 (i.e., E = V
and E = 2V ). To construct the BN , we first ordered the nodes. Then, we randomly added edges in a
forward direction until obtaining the specified mean graph density. This process generates a graph
with a power-law-type distribution over the number of parents, with some nodes having many more
than the average number of parents (Figure 2). In each BN , the nodes correspond to continuous
random variables where for every pair of nodes, (A, B), we parametrize a relation A → B in the BN
as a structural equation model (SEM): A = A and B = Aβ + B , where A and B are zero-mean
Gaussian noise terms and β is a linear coefficient. In our experiments, similar to Ramsey[21, 22],
variances of A and B are uniformly randomly chosen from the interval [1.0, 3.0] and β is drawn
uniformly randomly from the interval [−1.5, 0.2] ∪ [0.2, 1.5]. This choice of parameter values for the
simulations implies that, on average, around half of the variance of the variables is due to the error
term, which makes structure learning more difficult [21, 22].
2. Simulate a dataset D of size 1000 from BN , subject to constraints that are described below.
2
https://github.com/tensorflow/skflow
4
3. Set a percentage of variables, h, to be unobserved (i.e., latent). These latent variables are randomly
chosen from confounder variables (i.e., common causes) in a given data-generating BN . We set h to
be either 10% or 20%.
4. Generate 200 bootstrap datasets, DB[1..200] from D.
5. For each bootstrap dataset, learn a PAG using the RFCI method; let P AG[1..200] designate these
PAGs. RFCI uses Fisher’s Z test to check conditional independence of variables in the dataset. We
set the significance level at which independence judgments were made to be α = 0.001 and 0.005.
6. For each node pair (A, B), calculate the probability distribution Pe (A, B) of the edge types of
(A, B) using maximum likelihood estimates on the counts in P AG[1..200].
7. Perform a stratified random sampling on the node pairs to obtain N training samples for calibration
and use the rest of the data for testing. We set N = 70, 140, or 210. In obtaining N samples, we used
stratified random sampling to select N /7 samples for each of the seven edge classes3 . In particular, we
first sorted the probability scores of edges in each edge class according to the bootstrap probabilities.
We then partitioned the instances into 5 bins of {[0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), [0.8, 1]}
based on their bootstrap probabilities. Finally, we sampled separately from each bin with equal
frequency.
8. Learn the calibration function fcal using the calibration training data.
9. For each node pair (A, B) in the test set, derive Pecal (A, B) = fcal (Pe (A, B)).
10. Compare the performance of Pecal versus Pe in correctly predicting the data-generating structure
of BN for the test set pairs, and doing so in a manner that is well calibrated.
In running the above evaluation procedure, step (5) is the most time consuming part that involves
running RFCI on 200 bootstrap datasets. However, this is still feasible due to the run-time efficiency
of the RFCI method and our use of parallel computing4 . For all simulations, we used Tetrad5 , which
is an open-source, freely available software application that is coded in Java.
Steps (1) through (10) above were repeated for 10 randomly generated BNs and the performance
results were averaged. For a given node pair, we take the predicted edge type for that pair to be the
one that has the highest probability. Note that although there are seven different edge classes, we
consider only five edge types for performance evaluation, because A → B and B → A are both
directed edge types, and A◦→ B and B ◦→ A are both partially directed edge types.
The first two edge-type-based evaluation measures are precision (P) and recall (R). To compute
these measures for each edge type, we calculated the four basic statistics of true positives (TP), false
positives (FP), true negatives (TN), and false negatives (FN) for each of the types separately. Precision
is then derived as the ratio T P/(T P + F P ). Recall is derived as the ratio T P/(T P + F N ). We
also report the F1-score (i.e., the harmonic mean of the precision and recall), which is a summary
measure that shows the overall performance of the predictions in terms of both precision and recall.
We also evaluated the edge-type-based predictions in terms of maximum calibration error (MCE)
[17]. We calculated the MCE for each edge type by partitioning the output space of the estimated
edge-type probabilities, which is on the interval [0, 1], into equal-frequency bins with 100 randomly
chosen instances. The estimated probability for each instance is located in one of the bins. For each
bin, we define the associated calibration error as the absolute difference between the mean value
of the predictions and the actual observed frequency of positive instances. The MCE calculates the
maximum calibration error over all the bins. The lower the value of MCE, the better the calibration
of the probability scores. The lowest possible value of MCE is 0 and the highest possible value is 1.
We also report the overall (micro averaged) MCE as a summary measure which shows the performance
of the predictions in terms of calibration. To compute this measure, we augmented the seven-element
probability distribution vectors, Pe (A, B) for all test instances to form an aggregated vector Pall . We
also augmented their corresponding 1-of-k binary labels [18, 1] to form an aggregated binary vector
Zall . The overall MCE is defined as the maximum calibration error calculated based on Pall and
Zall .
3
Using stratified random sampling is crucial due to severe class imbalance of the data (i.e., more than 99% of the pairs are no-edge type).
The running times of the experiments varied from 11 to 176 minutes on a 16-core compute node, which is computationally feasible,
because step (5) needs to be done one time only.
5
https://github.com/cmu-phil/tetrad
4
5
4
Experimental Results
This section presents the results of our experiments in evaluating the performance of the generated
probabilities for the five edge types before and after calibration. We use the shallow neural network
calibration method to learn the calibration function fcal from calibration training data. Since the
purpose of this paper is not to compare calibration methods, we do not report the results of experiments
on using other calibration methods (e.g., IsoReg or Platt’s method). Rather, we only report the results
of calibration using the neural network method which we found performs well with relatively small
calibration training sample sizes compared to the other calibration methods that we tried.
For each set of configurations (e.g., N = 210, V = 2000, E = 2000), we report the average results
of using 10 randomly simulated CBNs. Tables 1a, 1b, 1c, and 1d show the results for CBNs with 2000
nodes (due to the page limit, the results of experiments for V = 1000 are not included but similar
results are achieved). In these tables, boldface indicates the results that are statistically significantly
superior, based on a two-sided Wilcoxon signed rank test at 5% significance level. Tables 1a-1d
indicate that by post-processing the bootstrap probabilities, we can improve the overall edge-type
performance both in terms of discrimination and calibration. The only exception is the A ↔ B edge
type for which we lose discrimination. This is happening because the original precision and recall
of the bootstrap probabilities are very low for this edge type. Consequently, we often obtain very
few positive instances from this edge type in the calibration training set, which negatively affects
the performance of predictions after calibration. Note that for the no-edge type we do not report
precision, recall, and F1, because they were always very close to 1.
Figure 3 shows the calibration diagram [4] of the estimated probabilities before and after calibration
when we use 210 calibration training instances. We emphasize that observing 210 calibration
instances is equivalent to observing less than 0.02% of all node pairs in the CBN (i.e., there are
1999 × 103 node pairs in a CBN with 2000 nodes). To draw the calibration diagrams, we partitioned
the output space of the estimated probabilities into five equal-size bins. In each bin, we draw the
average frequency of positive class versus the mean of the predictions that are located in that bin. In
these diagrams, the straight dashed line connecting (0, 0) to (1, 1) represents a perfectly calibrated
model. The closer a calibration curve is to this line, the better calibrated is the prediction model.
Figure 3 shows the proposed shallow neural network post-processing method often improves the
calibration performance of the predictions for the A → B edge-type, which is the most important
edge type since it is the one that is most likely to drive experimentation. In particular, directed edges
that a scientist considers to be high probability, as well as novel and important, would be prime
candidates for experimental validation. Furthermore, for this edge-type, the high probability region is
arguably the most critical one for making decisions about which directed arcs to investigate, such that
false positive experimental investigations are minimized.
Also, the associated diagrams of the no-edge type (i.e., A · · · B) in Figure 3 show that the estimated
probabilities are pretty well-calibrated after calibration. This is an interesting observation considering
the fact that the precision and recall are also always very close to 1 for this edge type after using
the post-processing calibration method. These results indicate that when the calibrated probabilities
indicate with high probability that there is no edge between a pair of variables, then those nodes rarely
are directly causally related. This result provides confidence in not prioritizing the experimental
investigation of such node pairs for direct causal relationships.
Another interesting observation in Figure 3 is that the bootstrap probabilities of the A ↔ B edge-type
are highly overestimated. This results in high false positive rate for that edge-type, and consequently,
increases the false negative rate for other edge types. Note that post-processing the bootstrap
probabilities does not generate high probabilities for the A ↔ B edge-type and consequently there
is no red circle in the high-probability bins but some in low probability bins, which is appropriate,
because those edges, which are output by RFCI, are seldom correct.
Overall, the calibration diagrams in Figure 3 show that post-processing the bootstrap probabilities
using the proposed shallow-neural network model generally improves the calibration performance
of the predictions. A key advantage of the shallow neural network approach for post-processing
the estimated probabilities is that we can readily condition on other types of features for learning a
calibration mapping (e.g., features extracted from the structure of the predicted PAGs by the RFCI
method, such as the indegree of B when we are generating a calibrated probability for the edge type
A → B). Such conditioning on local or global features of the learned graph could potentially yield
improvements in the post-processed calibrated probabilities. This is an area for future research.
6
Table 1: Simulation results. V , E, and h represent the number of variables, edges, and percentage of hidden
variables in the data-generating CBN, respectively. α is the significance level used in the RFCI method and N is
the calibration training set size. Boldface indicates the results that are significantly better, based on the Wilcoxon
signed rank test at 5% significance level. For MCE, lower is better.
(a) V = 2000, h = 0.1, and α = 0.001
N
E
2K
70
4K
2K
140
4K
2K
210
4K
method
before
after
before
after
before
after
before
after
before
after
before
after
P
0.61
0.69
0.66
0.67
0.57
0.66
0.66
0.67
0.53
0.65
0.65
0.67
A→B
R F1
0.33 0.42
0.37 0.47
0.30 0.41
0.38 0.48
0.28 0.37
0.31 0.41
0.29 0.40
0.37 0.48
0.24 0.32
0.28 0.37
0.29 0.40
0.36 0.47
P
0.35
0.43
0.65
0.63
0.32
0.39
0.65
0.65
0.26
0.35
0.65
0.65
A→B
R F1
0.15 0.21
0.15 0.21
0.25 0.36
0.33 0.42
0.11 0.15
0.10 0.14
0.24 0.35
0.31 0.42
0.09 0.13
0.08 0.13
0.23 0.34
0.30 0.41
P
0.60
0.65
0.74
0.70
0.49
0.62
0.74
0.70
0.33
0.43
0.75
0.72
A→B
R F1
0.19 0.28
0.32 0.41
0.29 0.41
0.39 0.50
0.13 0.20
0.23 0.32
0.28 0.41
0.40 0.50
0.07 0.11
0.14 0.19
0.27 0.40
0.37 0.48
P
0.25
0.39
0.72
0.68
0.15
0.24
0.73
0.71
0.07
0.13
0.73
0.71
A→B
R F1
0.05 0.07
0.10 0.14
0.22 0.34
0.30 0.42
0.02 0.03
0.06 0.09
0.21 0.32
0.29 0.41
0.00 0.01
0.02 0.03
0.20 0.31
0.28 0.40
MCE
0.12
0.10
0.30
0.20
0.10
0.09
0.30
0.17
0.10
0.09
0.31
0.19
P
0.45
0.57
0.51
0.46
0.42
0.58
0.51
0.47
0.40
0.58
0.51
0.47
A◦→ B
R F1
0.07 0.12
0.62 0.59
0.03 0.06
0.43 0.44
0.06 0.10
0.63 0.60
0.03 0.06
0.44 0.46
0.05 0.09
0.64 0.61
0.03 0.05
0.43 0.45
MCE
0.44
0.25
0.45
0.23
0.46
0.27
0.45
0.24
0.46
0.26
0.46
0.23
P
0.79
0.79
0.62
0.68
0.64
0.79
0.44
0.68
0.62
0.80
0.27
0.66
A◦–◦ B
R F1
0.06 0.12
0.44 0.56
0.03 0.05
0.27 0.38
0.05 0.09
0.41 0.54
0.02 0.03
0.26 0.37
0.04 0.07
0.40 0.53
0.01 0.02
0.23 0.34
MCE
0.65
0.27
0.44
0.26
0.65
0.29
0.44
0.26
0.66
0.25
0.44
0.29
P
0.05
0.03
0.04
0.00
0.04
0.02
0.04
0.00
0.04
0.02
0.04
0.01
A↔B
R F1
0.25 0.08
0.02 0.02
0.09 0.06
0.00 0.00
0.23 0.07
0.01 0.01
0.09 0.06
0.00 0.00
0.23 0.07
0.01 0.01
0.09 0.06
0.00 0.00
MCE
0.94
0.18
0.95
0.09
0.93
0.17
0.95
0.09
0.93
0.16
0.95
0.08
A ··· B
MCE
0.22
0.14
0.51
0.21
0.21
0.12
0.51
0.18
0.21
0.12
0.51
0.17
Overall
MCE
0.33
0.23
0.55
0.22
0.34
0.26
0.55
0.21
0.35
0.26
0.56
0.20
MCE
0.66
0.27
0.40
0.22
0.67
0.30
0.40
0.20
0.67
0.31
0.40
0.20
P
0.11
0.21
0.09
0.08
0.10
0.20
0.09
0.06
0.10
0.19
0.09
0.14
A↔B
R F1
0.18 0.13
0.10 0.12
0.06 0.07
0.01 0.01
0.17 0.13
0.10 0.11
0.06 0.07
0.01 0.01
0.17 0.13
0.10 0.11
0.06 0.07
0.01 0.01
MCE
0.90
0.29
0.91
0.14
0.90
0.27
0.91
0.14
0.90
0.28
0.91
0.12
A ··· B
MCE
0.24
0.13
0.53
0.20
0.23
0.12
0.53
0.19
0.23
0.12
0.52
0.18
Overall
MCE
0.34
0.24
0.56
0.23
0.33
0.23
0.59
0.21
0.33
0.23
0.58
0.21
MCE
0.70
0.25
0.46
0.28
0.69
0.24
0.46
0.27
0.62
0.21
0.47
0.25
P
0.05
0.00
0.05
0.01
0.05
0.00
0.05
0.00
0.05
0.00
0.05
0.00
A↔B
R F1
0.30 0.09
0.00 0.00
0.12 0.07
0.00 0.00
0.29 0.09
0.00 0.00
0.11 0.06
0.00 0.00
0.29 0.09
0.00 0.00
0.11 0.06
0.00 0.00
MCE
0.96
0.09
0.97
0.10
0.96
0.09
0.97
0.09
0.96
0.09
0.97
0.08
A ··· B
MCE
0.24
0.14
0.38
0.17
0.24
0.13
0.38
0.17
0.24
0.11
0.38
0.15
Overall
MCE
0.42
0.25
0.55
0.24
0.38
0.24
0.55
0.25
0.34
0.25
0.54
0.24
MCE
0.49
0.20
0.43
0.25
0.47
0.20
0.39
0.21
0.39
0.20
0.37
0.21
P
0.11
0.06
0.10
0.05
0.11
0.09
0.10
0.04
0.12
0.09
0.10
0.03
A↔B
R F1
0.21 0.15
0.01 0.02
0.08 0.09
0.00 0.01
0.21 0.14
0.01 0.02
0.08 0.08
0.00 0.01
0.20 0.15
0.01 0.02
0.08 0.08
0.00 0.00
MCE
0.92
0.14
0.94
0.13
0.92
0.14
0.94
0.12
0.92
0.13
0.94
0.12
A ··· B
MCE
0.22
0.13
0.40
0.18
0.22
0.12
0.39
0.15
0.22
0.11
0.40
0.16
Overall
MCE
0.52
0.23
0.49
0.25
0.50
0.23
0.48
0.23
0.48
0.23
0.48
0.25
(b) V = 2000, h = 0.2, and α = 0.001
N
E
2K
70
4K
2K
140
4K
2K
210
4K
method
before
after
before
after
before
after
before
after
before
after
before
after
MCE
0.10
0.08
0.28
0.18
0.09
0.08
0.28
0.15
0.10
0.08
0.29
0.15
P
0.37
0.53
0.49
0.44
0.32
0.55
0.48
0.43
0.32
0.55
0.49
0.43
A◦→ B
R F1
0.03 0.05
0.34 0.41
0.02 0.04
0.31 0.36
0.02 0.04
0.34 0.41
0.02 0.03
0.31 0.36
0.02 0.03
0.32 0.40
0.01 0.03
0.31 0.36
MCE
0.48
0.26
0.44
0.25
0.50
0.26
0.45
0.24
0.50
0.27
0.45
0.23
P
0.56
0.77
0.36
0.65
0.44
0.79
0.20
0.65
0.40
0.80
0.13
0.63
A◦–◦ B
R F1
0.03 0.05
0.30 0.42
0.02 0.04
0.21 0.31
0.02 0.03
0.28 0.40
0.01 0.03
0.22 0.32
0.01 0.02
0.27 0.39
0.01 0.01
0.20 0.30
(c) V = 2000, h = 0.1, and α = 0.005
N
E
2K
70
4K
2K
140
4K
2K
210
4K
method
before
after
before
after
before
after
before
after
before
after
before
after
MCE
0.16
0.12
0.29
0.20
0.16
0.11
0.29
0.19
0.16
0.11
0.30
0.18
P
0.36
0.43
0.39
0.42
0.27
0.44
0.39
0.44
0.15
0.46
0.35
0.44
A◦→ B
R F1
0.01 0.02
0.62 0.51
0.01 0.01
0.44 0.43
0.01 0.01
0.62 0.51
0.01 0.01
0.44 0.44
0.00 0.00
0.63 0.53
0.00 0.01
0.45 0.44
MCE
0.43
0.27
0.46
0.27
0.43
0.26
0.46
0.27
0.44
0.26
0.47
0.26
P
0.17
0.76
0.18
0.66
0.00
0.77
0.12
0.73
0.00
0.76
0.10
0.73
A◦–◦ B
R F1
0.00 0.00
0.18 0.28
0.00 0.00
0.14 0.22
0.00 0.00
0.18 0.28
0.00 0.00
0.16 0.26
0.00 0.00
0.15 0.24
0.00 0.00
0.17 0.27
(d) V = 2000, h = 0.2, and α = 0.005
2K
210
4K
0.2
1.0
0.8
0.6
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
1.0
0.8
0.2
Estimated Probability
0.2
Estimated Probability
1.0
0.8
0.6
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
1.0
0.8
0.4
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
1.0
0.8
0.6
0.2
Estimated Probability
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
1.0
0.8
0.6
0.4
0.00.0 0.2 0.4 0.6 0.8 1.0
A···B
1.0
0.8
0.6
0.4
0.4
0.00.0 0.2 0.4 0.6 0.8 1.0
A↔B
A◦–◦B
1.0
0.8
0.6
0.4
0.00.0 0.2 0.4 0.6 0.8 1.0
A◦–◦ B
R F1
0.00 0.00
0.12 0.19
0.00 0.00
0.12 0.20
0.00 0.00
0.11 0.17
0.00 0.00
0.12 0.20
0.00 0.00
0.06 0.10
0.00 0.00
0.12 0.20
0.4
0.6
0.4
P
0.06
0.67
0.12
0.70
0.00
0.72
0.00
0.67
0.00
0.66
0.00
0.71
0.6
0.4
Estimated Probability
Objective Probability
1.0
0.8
0.6
0.4
0.00.0 0.2 0.4 0.6 0.8 1.0
V = 2k,
E = 4k
Objective Probability
0.6
V = 2k,
E = 2k
A◦→B
Objective Probability
Objective Probability
A→B
1.0
0.8
MCE
0.42
0.24
0.45
0.27
0.44
0.25
0.46
0.25
0.47
0.24
0.46
0.26
Objective Probability
4K
A◦→ B
R F1
0.00 0.01
0.45 0.41
0.00 0.01
0.32 0.35
0.00 0.00
0.45 0.42
0.00 0.00
0.33 0.35
0.00 0.00
0.45 0.42
0.00 0.00
0.33 0.35
Objective Probability
2K
140
P
0.17
0.39
0.32
0.37
0.14
0.40
0.26
0.38
0.06
0.40
0.26
0.38
Objective Probability
4K
before
after
before
after
before
after
before
after
before
after
before
after
MCE
0.12
0.09
0.30
0.22
0.11
0.07
0.31
0.20
0.11
0.07
0.32
0.21
Objective Probability
2K
70
method
Objective Probability
E
Objective Probability
N
0.4
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
0.2
0.00.0 0.2 0.4 0.6 0.8 1.0
Estimated Probability
Figure 3: The calibration curves of the edge-type probabilities before (blue crosses) and after (red circles)
calibration. The closer the predictions to the diagonal, the more calibrated are the probabilities. In these results,
the calibration training set size is 210, the percentage of hidden variables is 0.1, and the significance level of the
test of independence in RFCI is 0.005
7
5
Conclusion
In this paper, we introduced a new approach for improving the calibration of CBN structure discovery.
We used a bootstrapping method to obtain estimated probabilities of the causal relationships between
each pair of random variables. Although we applied the bootstrapping method to the RFCI algorithm,
it can be applied with any other type of the network discovery method, as long as the method is
sufficiently fast to run hundreds of times on a dataset to obtain bootstrap probability estimates. To
calibrate the bootstrap probabilities, we devised a natural extension of Platt’s calibration method that
supports multi-class calibration using a shallow neural network. Our experiments on a wide range
of large simulated datasets show that by using only a small set of instances as gold standards for
training the calibration model, we can obtain substantial improvements in terms of precision, recall,
and calibration, relative to the bootstrap probabilities. In future work, we plan to expand the range of
simulated experiments we perform, as well as evaluate the method using real biomedical data for
which the truth status is known from the literature for a relatively small subset of variables.
Acknowledgement We thank the members of the Center for Causal Discovery for their helpful
feedback. The research reported in this publication was supported by grant U54HG008540 awarded
by the National Human Genome Research Institute through funds provided by the trans-NIH Big
Data to Knowledge (BD2K) initiative. The content is solely the responsibility of the authors and does
not necessarily represent the official views of the National Institutes of Health. This material is also
based upon work supported by the National Science Foundation under Grant No. IIS-1636786. Any
opinions, findings, and conclusions or recommendations expressed in this material are those of the
author(s) and do not necessarily reflect the views of the National Science Foundation.
References
[1] Martín Abadi, Ashish Agarwal, et al. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv:1603.04467,
2016.
[2] David Maxwell Chickering. Optimal structure identification with greedy search. Journal of Machine Learning Research, 3:507–554,
2003.
[3] Diego Colombo, Marloes Maathuis, Markus Kalisch, and Thomas Richardson. Learning high-dimensional directed acyclic graphs with
latent and selection variables. The Annals of Statistics, 40(1):294–321, 2012.
[4] Morris DeGroot and Stephen Fienberg. The comparison and evaluation of forecasters. The Statistician, 1983.
[5] Pedro Domingos. A few useful things to know about machine learning. Communications of the ACM, 55(10):78–87, 2012.
[6] John Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization. Journal of
Machine Learning Research, 12:2121–2159, 2011.
[7] Daniel Eaton and Kevin Murphy. Exact Bayesian structure learning from uncertain interventions. In International Conference on
Artificial Intelligence and Statistics, pages 107–114, 2007.
[8] Bradley Efron and Robert Tibshirani. An Introduction to the Bootstrap. CRC press, 1994.
[9] Nir Friedman, Moises Goldszmidt, and Abraham Wyner. Data analysis with Bayesian networks: A bootstrap approach. In Proceedings
of the Fifteenth Conference on Uncertainty in artificial intelligence, pages 196–205, 1999.
[10] Nir Friedman and Daphne Koller. Being Bayesian about network structure. a Bayesian approach to structure discovery in Bayesian
networks. Machine learning, 50(1-2):95–125, 2003.
[11] Clark Glymour and Gregory Cooper. Computation, Causation, and Discovery. MIT Press, 1999.
[12] Phyllis Illari, Federica Russo, and Jon Williamson. Causality in the Sciences. OUP Oxford, 2011.
[13] Mikko Koivisto. Advances in exact Bayesian structure discovery in Bayesian networks. arXiv:1206.6828, 2012.
[14] Mikko Koivisto and Kismat Sood. Exact Bayesian structure discovery in Bayesian networks. Journal of Machine Learning Research,
5:549–573, 2004.
[15] David Madigan, Jeremy York, and Denis Allard. Bayesian graphical models for discrete data. International Statistical Review, pages
215–232, 1995.
[16] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Binary classifier calibration using a Bayesian non-parametric approach.
In Proceedings of the SIAM International Conference on Data Mining, pages 208–216, 2015.
[17] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using Bayesian binning. In
AAAI, pages 2901–2907, 2015.
[18] Michael A Nielsen. Neural networks and deep learning. Determination Press, 2015.
[19] Judea Pearl. Causality: Models, reasoning and inference. Econometric Theory, 19:675–685, 2003.
[20] John Platt. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large
Margin Classifiers, 10(3):61–74, 1999.
[21] Joseph Ramsey. Scaling up greedy equivalence search for continuous variables. arXiv:1507.07749, 2015.
[22] Ricardo Silva, Richard Scheines, Clark Glymour, and Peter Spirtes. Learning the structure of linear latent variable models. Journal of
Machine Learning Research, 7:191–246, 2006.
[23] Peter Spirtes. Introduction to causal inference. Journal of Machine Learning Research, 11:1643–1662, 2010.
[24] Peter Spirtes, Clark Glymour, and Richard Scheines. Causation, Prediction, and Search. MIT press, 2000.
[25] Bianca Zadrozny and Charles Elkan. Transforming classifier scores into accurate multiclass probability estimates. In Proceedings of the
ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 694–699, 2002.
8
| 2cs.AI
|
arXiv:1802.01227v1 [math.ST] 5 Feb 2018
Copula-based Partial Correlation Screening: a Joint and
Robust Approach
Jialiang Li1,2,3 and Xiaochao Xia1,4∗
1
Department of Statistics and Applied Probability, National University of Singapore
2
Duke-NUS Graduate Medical School
4
3
Singapore Eye Research Institute
College of Science, Huazhong Agricultural University, Wuhan, China
Abstract
Screening for ultrahigh dimensional features may encounter complicated issues
such as outlying observations, heteroscedasticity or heavy-tailed distribution, multicollinearity and confounding effects. Standard correlation-based marginal screening
methods may be a weak solution to these issues. We contribute a novel robust joint
screener to safeguard against outliers and distribution mis-specification for both the
response variable and the covariates, and to account for external variables at the screening step. Specifically, we introduce a copula-based partial correlation (CPC) screener.
We show that the empirical process of the estimated CPC converges weakly to a Gaussian process and establish the sure screening property for CPC screener under very
mild technical conditions, where we need not require any moment condition, weaker
than existing alternatives in the literature. Moreover, our approach allows for a diverging number of conditional variables from the theoretical point of view. Extensive
simulation studies and two data applications are included to illustrate our proposal.
1
Keywords: Copula; Partial correlation; Outlier; Joint screening; Sure independent
screening.
1
Introduction
With the arrival of a big data era, ultrahigh dimensional data become readily available from
many business and scientific research fields, including medicine, genetics, finance and economics. Such massive data usually carry two common features: (i) the number of predictors
or features can be tremendous and diverge to infinity with the sample size and (ii) the data
distribution is very likely to be heteroscedastic and heavy-tailed for both the response and
the covariates. These two features are observed in the two real data sets investigated in
this paper. It is usually hoped that screening identifies important predictors among numerous candidates. We note that ultimately data scientists for such large scale data still
need to construct a comprehensive joint model to accurately predict the future outcome.
Thus a purely marginal screening approach as is usually adopted in the literature may not
adequately serve the model building purpose. We contribute a new screening method that
addresses the above issues and complements the existing methodology.
Variable screening serves as a fast and efficient computing device. Abundant feature
screening methods are proposed in recent decades, including the sure independence screening
(SIS) by Fan and Lv (2008) who first established the sure screening property under Gaussian
linear model, the sure independent ranking screening (SIRS, Zhu et al. (2011)), the Kendall’s
τ based screening (Kendall-SIS, Li et al. (2012a)), the distance correlation based screening
(DC-SIS, Li et al. (2012b)), the quantile-adaptive screening (QaSIS, He et al. (2013)), empirical likelihood screening (Chang et al. (2013, 2016)), the censored rank independence screening for lifetime data (CRIS, Song et al. (2014)), the screening method based on quantile correlation (QC-SIS, Li et al. (2015)), the conditional quantile screening (CQ-SIS, Wu and Yin
2
(2015)), the survival impaction index screening (SII, Li et al. (2016)), the nonparametric
independence screening (NIS, Fan et al. (2011)), among many others. These screening tools
might suffer the following two drawbacks: First, almost all methods evaluate a marginal
association between the response and the predictors without adjusting external variables.
Therefore some jointly important markers may be incorrectly screened out if their marginal
signal is not as strong as the spurious markers in the ranked list. On the other hand,
marginally important variables may be jointly ineffective and hence including them in a
multivariate model may lead to less convincing prediction. To take into account the joint
effects, a marginal feature screening is usually followed by an iterative calculation, such as
the iterative SIS (ISIS) in Fan and Lv (2008), which is computationally expensive and does
not come with any theoretical guarantee. Secondly, distribution of the response and the
covariates may be rather different from the light-tailed symmetric normal distribution and
very often there are outliers affecting the computed screening indices. Some of the aforementioned procedures address the robustness of the response but to our knowledge none of
the existing work addresses the robustness of the covariates yet, which is a harder problem
with higher dimension.
We aim to tackle the two problems with a new screener. Specifically, to address the first
issue, we develop a joint feature screening method by incorporating additional information.
Recently a few conditional feature screening methods have been proposed. For instance,
Liu et al. (2014) considered a sure independence screening procedure via conditional Pearson correlation coefficient through a kernel smoothing. Their method can be employed to
handle ultrahigh dimensional varying-coefficient feature variables, which were investigated in
Fan et al. (2014) and Cheng et al. (2014) as well. In addition, Xia et al. (2018) considered a
robust screening method based on conditional quantile correlation, a generalized conception
of Li et al. (2015). However, these authors only considered a single conditional variable. To
extend to multivariate conditional variables, Chu et al. (2016) studied several confounding
3
variables. Barut et al. (2016) extended Fan and Song (2010))’s approach to allow for a portion of predictors as conditional variables. Our work provides a more general framework
where all the ultrahigh dimensional markers and other low dimensional confounders can be
jointly considered during the screening process. For the second issue, we incorporate robust
copula-based correlation and partial correlation in our screening methods. The nonparametric copula is a well-known distribution-free summary measure and naturally leads to a
screener against outliers and distribution mis-specification. To the best of our knowledge,
there are very few works applying this classical dependence concept in high-dimensional setting. Xia et al. (2018) proposed a robust conditional feature screening approach, however,
their method is only robust for the response but not for the covariates. Another relevant
recent work is Ma et al. (2017).
The contribution of this paper can be summarized as follows. Firstly, we proposed a double robust copula-based correlation (CC). Copula is very popular bivariate function to model
the nonlinear dependence between paired variates. See Nelsen (2007) for an introduction for
copula. The CC characterizes the empirical dependence between two random variables evaluated at a level pair and is invariant under monotone transformation for both variables. We
study the asymptotic process properties of the CC. A marginal variable screening approach
via CC (CC-SIS) can be performed and achieves the desired sure screening consistency
(Fan and Lv (2008)). Secondly, extending copula-based correlation to copula-based partial
correlation (CPC), we then construct a more general framework for joint screening. The
importance of each marker is evaluated in the presence of conditional variables. This provides a fast way for conditional feature screening as it avoids the iterated computation. CPC
is also robust by its construction from a nonparametric estimation and thus may be more
reliable than a similar approach in Ma et al. (2017) with a broader range of application. We
provide both theoretical and numerical support for the proposed screening method. Our
data analysis indicates that the final multivariate regression models built after our screening
4
approach indeed predict the outcome with improved accuracy.
The rest of the paper is organized as follows. Section 2 presents CC, its empirical estimate
and CC-SIS. Asymptotic properties for the estimated CC functions are established in Section
2 as well. Methodologies and large sample properties for the CPC and the CPC-SIS are
presented in Section 3. Further implementation details on CPC-SIS for different cases are
given in Section 4. Simulation studies and two applications are carried out in Section 5.
Section 6 concludes the paper. All the technical proofs are relegated to the Appendix.
2
Copula-based Correlation and Variable Screening
Consider two random variables X and Y . Let FX be the cumulative distribution function
(CDF) of X, which is assumed to be right continuous. FX−1 (τ ) = inf{x : FX (x) ≥ τ } is the τ
quantile of FX . FY,X is the joint CDF of Y and X, and FY |X is the conditional distribution
−1
function of Y given X with a density fY |X . Use Fn,X , Fn,X
and Fn,Y,X to denote empirical
versions of FX , FX−1 and FY,X , respectively, based on a sample of size n. Let D[a, b] be the
Banach space of all càdlàg functions z : [a, b] 7→ R on an interval [a, b] ⊂ R̄ equipped with the
uniform norm, and ℓ∞ ([a, b]2 ) denotes the collection of all bounded functions z : [a, b]2 7→ R.
d
We use → to denote convergence in distribution.
2.1
Copula-based Correlation
We propose the following copula-based correlation (CC)
̺Y,X (τ, ι) =
FY,X (FY−1 (τ ), FX−1 (ι)) − τ ι
p
,
τ (1 − τ )ι(1 − ι)
0 ≤ τ, ι ≤ 1,
(1)
where the first term in the numerator is a copula function of (X, Y ) defined by C(u, v) =
FU,V (u, v) with U = FY (Y ) and V = FX (X), evaluated at (u, v) = (τ, ι) (see Corollary 2.3.7
5
of Nelsen (2007), p.22). By a simple algebra, it follows that ̺Y,X (τ, ι) =
−1
cov(ψτ (Y −FY−1 (τ )),ψι (X−FX
(ι)))
√
τ (1−τ )ι(1−ι)
−1
E[ψτ (Y −FY−1 (τ ))ψι (X−FX
(ι))]
√
τ (1−τ )ι(1−ι)
, where ψτ (u) = τ − I(u ≤ 0) and I(·) is the indicator function.
Since var(ψτ (Y − FY−1 (τ ))) = τ (1 − τ ), ̺Y,X (τ, ι) in (1) is indeed a legitimate correlation
coefficient that lives between −1 and 1. Like other correlation measures, CC equals 0 if X
and Y are independent.
CC can measure the nonlinear dependence between X and Y , thus incorporates all kinds
of bivariate joint distribution of X and Y . In addition, because the indicator function is
unaffected by outliers and extreme values, CC is robust for certain heavy-tailed distribution
for both Y and X. We note that monotone transformation of X and Y does not alter the
value of CC.
Given a sample of i.i.d. observations {(Xi , Yi ), i = 1, . . . , n}, we can construct an empirical estimate of ̺Y,X (τ, ι) as
−1
−1
n−1
Fn,Y,X (Fn,Y
(τ ), Fn,X
(ι)) − τ ι
p
=
̺bY,X (τ, ι) =
τ (1 − τ )ι(1 − ι)
Pn
i=1 ψτ (Yi
p
−1
−1
− Fn,Y
(τ ))ψι (Xi − Fn,X
(ι))
τ (1 − τ )ι(1 − ι)
. (2)
Let σY,X (τ, ι) = FY,X (FY−1 (τ ), FX−1 (ι)), σX|Y (τ, ι) = FX|Y =F −1 (τ ) (FX−1 (ι)) and σY |X (τ, ι) =
Y
FY |X=F −1(ι) (FY−1 (τ )). In the following, we fix the level at (τ, ι) and write σY,X , σX|Y and
X
σY |X for simplicity. Furthermore, define
h
1
I(Y ≤ FY−1 (τ ), X ≤ FX−1 (ι))
ξ(Y, X; τ, ι) = p
τ (1 − τ )ι(1 − ι)
−σX|Y (τ, ι)I(Y ≤
FY−1 (τ ))
− σY |X (τ, ι)I(X ≤
FX−1 (ι))
i
.
We have the weak convergence result for ̺bY,X (τ, ι) established in the next theorem.
Theorem 2.1. Let 0 < a < b < 1 and suppose that marginal distributions FX and FY are
continuously differentiable on the intervals [FX−1 (a)−ε, FX−1 (b)+ε] and [FY−1 (a)−ε, FY−1 (b)+ε]
with positive derivatives fX and fY , respectively, for some ε > 0. Furthermore, assume that
6
=
conditional density functions fY |X and fX|Y are continuous on the product of these intervals.
Then we have
√
in ℓ∞ ([a, b]2 ), where
w
n{b
̺Y,X (τ, ι) − ̺Y,X (τ, ι)}
w
GY,X (τ, ι)
denotes ”converge weakly”, and GY,X (τ, ι) is Gaussian process with
mean zero and covariance function Ω1 (τ1 , ι1 ; τ2 , ι2 ) ≡ E{[ξ(Y, X; τ1 , ι1 ) − Eξ(Y, X; τ1 , ι1 )] ×
[ξ(Y, X; τ2, ι2 ) − Eξ(Y, X; τ2 , ι2 )]}.
We may write the covariance function Ω1 (τ1 , ι1 ; τ2 , ι2 ) =
FY,X (FY−1(τ1 ∧ τ2 ), FX−1(ι1 ∧ ι2 )) − FY,X (FY−1 (τ1 ), FX−1 (ι1 ))FY,X (FY−1 (τ2 ), FX−1 (ι2 ))
−σX|Y (τ2 , ι2 )[FY,X (FY−1 (τ1 ∧ τ2 ), FX−1 (ι1 )) − FY,X (FY−1 (τ1 ), FX−1 (ι1 ))τ2 ]
−σY |X (τ2 , ι2 )[FY,X (FY−1 (τ1 ), FX−1 (ι1 ∧ ι2 )) − FY,X (FY−1 (τ1 ), FX−1 (ι1 ))ι2 ]
−σX|Y (τ1 , ι1 )[FY,X (FY−1 (τ1 ), FX−1 (ι2 )) − FY,X (FY−1 (τ2 ), FX−1(ι2 ))τ1 ]
+σX|Y (τ1 , ι1 )σX|Y (τ2 , ι2 )(τ1 ∧ τ2 − τ1 τ2 )
+σX|Y (τ1 , ι1 )σY |X (τ2 , ι2 )[FY,X (FY−1 (τ1 ), FX−1(ι2 )) − τ1 ι2 ]
−σY |X (τ1 , ι1 )[FY,X (FY−1 (τ2 ), FX−1 (ι1 ∧ ι2 )) − FY,X (FY−1 (τ2 ), FX−1 (ι2 ))ι1 ]
+σY |X (τ1 , ι1 )σX|Y (τ2 , ι2 )[FY,X (FY−1 (τ2 ), FX−1(ι1 )) − τ2 ι1 ]
+σY |X (τ1 , ι1 )σY |X (τ2 , ι2 )(ι1 ∧ ι2 − ι1 ι2 ) /[τ1 (1 − τ1 )ι1 (1 − ι1 )τ2 (1 − τ2 )ι2 (1 − ι2 )]1/2 .
√
d
̺Y,X (τ, ι) → N(0, Ω1 ), where Ω1 ≡
In particular, at fixed (τ, ι), if ̺Y,X (τ, ι) = 0, then nb
2
2
Ω1 (τ, ι; τ, ι) = σY,X − σY,X
+ (τ − τ 2 )σX|Y
+ (ι − ι2 )σY2 |X − 2(1 − τ )σY,X σX|Y − 2(1 −
ι)σY,X σY |X + 2[σY,X − τ ι]σX|Y σY |X /[τ (1 − τ )ι(1 − ι)]. If Y and X are independent, then
Ω1 = 1, producing the same null distribution used in classical correlation and auto-correlation
studies. Compared with Li et al. (2015), our result is free of the moment conditions on X,
7
while Li et al. (2015) requires the existence of a fourth order moment on X to achieve the
convergence in law.
In order to make statistical inference, such as constructing a confidence interval for
̺Y,X (τ, ι) and testing a hypothesis like H0 : ̺Y,X (τ, ι) = 0, we need to estimate the covariance function Ω1 (τ1 , ι1 ; τ2 , ι2 ). To this end, denote m1 (y) = E{I(X ≤ FX−1 (ι))|Y = y}
and m2 (x) = E{I(Y ≤ FY−1 (τ ))|X = x}. We can use the nonparametric approach like the
Nadaraya-Watson (NW) method to obtain estimates m
b 1 (y) and m
b 2 (x) for m1 (y) and m2 (x),
−1
−1
respectively, where the unknown FX−1 (ι) and FY−1 (τ ) are replaced by Fn,X
(ι) and Fn,Y
(τ ),
−1
respectively. Therefore, we obtain the estimates σ
bX|Y (τ, ι) = m
b 1 (Fn,Y
(τ )) and σ
bY |X (τ, ι) =
−1
m
b 2 (Fn,X
(ι)). Next, we give an estimate of Ω1 (τ1 , ι1 ; τ2 , ι2 ). Denote ξbn (Yi , Xi ; τ, ι) = I(Yi ≤
p
−1
−1
−1
−1
Fn,Y
(τ ), Xi ≤ Fn,Y
(ι))−b
σX|Y (τ, ι)I(Yi ≤ Fn,Y
(τ ))−b
σY |X (τ, ι)I(Xi ≤ Fn,X
(τ )) / τ (1 − τ )ι(1 − ι)
P
and ξ n (Y, X; τ, ι) = n−1 ni=1 ξbn (Yi , Xi ; τ, ι). Then, we obtain a consistent estimate of
b 1 (τ1 , ι1 ; τ2 , ι2 ) = n−1 Pn [ξbn (Yi , Xi ; τ1 , ι1 )−ξ n (Y, X; τ1 , ι1 )]×[ξbn (Yi , Xi ; τ2 , ι2 )−
Ω1 (τ1 , ι1 ; τ2 , ι2 ) as Ω
i=1
ξ n (Y, X; τ2 , ι2 )].
As is relevant to this paper, we consider a single Y but many X candidates. In practice,
to compare the dependence strength of two random variables X1 and X2 on Y , we may check
the difference ̺Y,X1 (τ, ι) − ̺Y,X2 (τ, ι). In particular, we may test a hypothesis regarding this
difference. Given a sample {(Yi , Xi1 , Xi2 ), i = 1, . . . , n}, similarly to (2), we can define
̺bY,X1 (τ, ι) and ̺bY,X2 (τ, ι). The following theorem can be applied to answer this question.
Theorem 2.2. Let 0 < a < b < 1 and suppose that marginal distributions FXk and FY are
continuously differentiable on the intervals [FX−1k (a)−ε, FX−1k (b)+ε] and [FY−1 (a)−ε, FY−1 (b)+ε]
with positive derivatives fXk and fY , respectively, for some ε > 0 and k = 1, 2. Furthermore,
assume that conditional density functions fY |Xk and fXk |Y , k = 1, 2, are continuous on the
product of these intervals. Then we have
√
n{[b
̺Y,X1 (τ, ι) − ̺bY,X2 (τ, ι)] − [̺Y,X1 (τ, ι) − ̺Y,X2 (τ, ι)]}
8
w
GY,X1 ,X2 (τ, ι),
in ℓ∞ ([a, b]2 ), where GY,X1 ,X2 (τ, ι) is Gaussian process with mean zero and covariance function Ξ1 (τ1 , ι1 ; τ2 , ι2 ) ≡ E{[η(Y, X1 , X2 ; τ1 , ι1 ) − Eη(Y, X1 , X2 ; τ1 , ι1 )] × [η(Y, X1 , X2 ; τ2 , ι2 ) −
Eη(Y, X1 , X2 ; τ2 , ι2 )]}, η(Y, X1 , X2 ; τ, ι) = ξ(Y, X1 ; τ, ι) − ξ(Y, X2 ; τ, ι) and ξ(Y, X; τ, ι) is defined in Theorem 2.1.
It follows from Theorem 2.2 that for a fixed pair (τ, ι), if ̺Y,X1 (τ, ι) = ̺Y,X2 (τ, ι), then
√
d
(1)
(2)
n{[b
̺Y,X1 (τ, ι) − ̺bY,X2 (τ, ι)] → N(0, Ξ1 ), where Ξ1 ≡ Ξ1 (τ, ι; τ, ι) = Ω1 + Ω1 − 2A12 , where
(k)
Ω1 is the same as Ω1 only except that X involved in Ω1 is substituted by Xk for k = 1, 2,
and A12 ≡ A12 (τ, ι) = [σY,X1 ,X2 (τ, ι)−σY,X1 σY,X2 ]−(1−τ )σX2 |Y σY,X1 −σY |X2 [σY,X1 ,X2 (τ, ι)−
ισY,X1 ]−(1−τ )σX1 |Y σY,X2 +τ (1−τ )σX1 |Y σX2 |Y +σX1 |Y σY |X2 (σY,X2 −τ ι)−σY |X1 [σY,X1 ,X2 (τ, ι)−
p
ισY,X2 ] + σY |X1 σX2 |Y (σY,X1 − τ ι) + σY |X1 σY |X2 [σX1 ,X2 (ι, ι) − ι2 ] / τ (1 − τ )ι(1 − ι), where
σY,X1 ,X2 (τ, ι) = FY,X1 ,X2 (FY−1 (τ ), FX−11 (ι), FX−12 (ι)) and σX1 ,X2 (ι, ι) = FX1 ,X2 (FX−11 (ι), FX−12 (ι)).
If Y, X1 and X2 are mutually independent, then Ξ1 = 2. Next, we estimate the covariance function Ξ1 (τ1 , ι1 ; τ2 , ι2 ). Let ηbn (Yi , Xi1 , Xi2 ; τ, ι) = ξbn (Yi , Xi1 ; τ, ι) − ξbn (Yi , Xi2 ; τ, ι),
P
where ξbn (Yi , Xi ; τ, ι) is given before, and η n (Y, X1 , X2 ; τ, ι) = n−1 ni=1 ηbn (Yi , Xi1 , Xi2 ; τ, ι).
P
b 1 (τ1 , ι1 ; τ2 , ι2 ) = n−1 n [b
Then, Ξ1 (τ1 , ι1 ; τ2 , ι2 ) can be estimated as Ξ
i=1 ηn (Yi , Xi1 , Xi2 ; τ1 , ι1 )−
η n (Y, X1 , X2 ; τ1 , ι1 )] × [b
ηn (Yi , Xi1 , Xi2 ; τ2 , ι2 ) − ηn (Y, X1 , X2 ; τ2 , ι2 )].
2.2
CC-based Variable Screening
Suppose that we collect a sample {(Yi , Xi ), i = 1, · · · , n} consisting of n independent copies
of (Y, X), where Y is the response variable and X = (X1 , · · · , Xp )T is a vector of p predictors.
When the number, p, of predictors is of an exponential order of sample size n, i.e., the socalled ultrahigh dimension, and most of p predictors are irrelevant, we can use CC as a
screener to identify the sparse set of informative predictors. We write pn instead of p to
emphasize the dependence on sample size. An empirical estimate for CC between Y and Xj
9
is given by
̺bY,Xj (τ, ι) =
n−1
Pn
i=1
−1
−1
ψτ (Yi − Fn,Y
(τ ))ψι (Xij − Fn,X
(ι))
j
p
.
τ (1 − τ )ι(1 − ι)
(3)
Then, we may select an empirical active set to be
ca = j : |b
M
̺Y,Xj (τ, ι)| ≥ νn , 1 ≤ j ≤ pn ,
(4)
where νn is a user-specified threshold parameter that controls the size of finally screened
model. Using CC may lead to sure independence screening (SIS) property and this procedure
will be abbreviated as CC-SIS.
Denote the true active set by M∗a = {j : |̺Y,Xj (τ, ι)| > 0, j = 1, . . . , pn }. Write FY−1|X (τ ) =
inf{y : P (Y ≤ y|X) ≥ τ }, uj = |̺Y,Xj (τ, ι)| and u
bj = |b
̺Y,Xj (τ, ι)|. To establish the screening
consistency, we need the following conditions.
(C1) In a neighbourhood of FY−1 (τ ), the density fY (y) of Y is uniformly bounded away from
zero and infinity and has bounded derivative. For every 1 ≤ j ≤ pn , in a neighbourhood
of FX−1j (ι), the density fXj (x) of Xj is uniformly bounded away from zero and infinity
and has bounded derivative.
(C2) minj∈M∗a uj ≥ C0 n−κ for some κ > 0 and C0 > 0.
Theorem 2.3. (Screening Property for CC-SIS) Suppose that the condition (C1) holds,
(i) for any constant C > 0, then there exists some positive constant c̃1 such that for sufficiently large n,
P
bj − uj ≥ Cn−κ ≤ 6pn exp(−c̃1 n1−2κ ).
max u
1≤j≤pn
(ii) In addition, if condition (C2) is further satisfied and by choosing νn = C1 n−κ with
10
C1 ≤ C0 /2, we have
ca ≥ 1 − 6sn exp(−c̃1 n1−2κ )
P M∗a ⊂ M
for sufficiently large n, where sn = |M∗a | is the cardinality of set M∗a .
This result implies that the CC-SIS can select all the truly active predictors with an
overwhelming probability. The dimensionality can be as high as pn = o(exp(n1−2κ )), similar
to many model-free feature screening methods (see Li et al. (2012a) and Wu and Yin (2015)
for example). Moreover, our result requires less condition on both the predictors and the
response due to the nonparametric nature. In fact, the condition does not involve any
moment assumption of the predictors or the response.
In practice, the threshold parameter νn plays an important role in producing a satisfied model. Small νn value will result in a large number of predictors after screening,
which in turn leads to many incorrect positives. Here we consider a data-driven procedure
to determine the threshold for the CC-SIS by controlling the false discovery rates (FDR).
By Theorem 2.1, for covariate j such that ̺Y,Xj (τ, ι) = 0, it follows that asymptotically,
√ b
n[Ω1 (τ, ι; τ, ι)]−1/2 ̺bY,Xj (τ, ι) ∼ N(0, 1). We can use high-criticism t-tests to select varib 1 (τ, ι; τ, ι)]−1/2 |b
ca,δ = {j : √n[Ω
̺Y,Xj (τ, ι)| ≥ δ} for a small δ > 0. This controls the
ables M
ca,δ ∩ (M∗ )c |/|(M∗ )c |} defined by Zhao and Li (2012). The following proposition
FDR E{|M
a
a
justifies this FDR procedure under the condition below.
(C3) For j ∈ (M∗a )c , ̺Y,Xj (τ, ι) = 0.
Proposition 2.4. (FDR Property) Under conditions (C1)-(C3) and the condition of Theorem 2.1, if we choose δ = Φ−1 (1 − d¯n /(2pn )) and Φ(·) is CDF of standard normal variable
and d¯n is the number of false positives that can be tolerated, then for some constant ca > 0,
11
we have
n |M
ca,δ ∩ (M∗ )c | o d¯n
√
a
n.
E
≤
+
c
/
a
|(M∗a )c |
pn
3
Copula-based Partial Correlation and Variable Screening
3.1
Copula-based Partial Correlation, CPC
To facilitate a joint screening procedure (Ma et al. (2017)), we define a copula-based partial
correlation (CPC) for Y and X conditional on a q-dimensional random vector Z as
E{ψτ (Y − ZT α0 )ψι (X − ZT θ 0 )}
p
,
̺Y,X|Z (τ, ι) =
τ (1 − τ )ι(1 − ι)
(5)
where α0 = argminα E{ρτ (Y − ZT α)} and θ 0 = argminθ E{ρι (X − ZT θ)} and ρw (u) = u[w −
−1
I(u ≤ 0)] for w = τ or ι. Note that this implies that ZT α0 = FY−1|Z (τ ) and ZT θ 0 = FX|Z
(ι).
Parameters α and θ can be interpreted as the marginal increment on conditional quantiles
of Y and X given Z, respectively, when increasing by a unit of Z. CPC is actually the CC
between Y and Xj after removing the confounding effects of Z. Linear partial correlation has
been widely used in regression diagnostics and describes the association of the response and
predictor conditional on specifical values of other predictors. The unconditional ̺Y,X (τ, ι)
value may be spurious due to lurking variables and does not necessarily imply the same
̺Y,X|Z (τ, ι) value conditional on Z. Our copula based version is relatively more robust for
real data analysis. The CC is a special case of CPC when Z is a constant.
With sample observations {(Yi , Xi , Zi ), i = 1, . . . , n}, we can obtain the following estimate
P
b = argmin 1 Pn ρι (Xi − ZT θ).
b = argminα n1 ni=1 ρτ (Yi − ZTi α) and θ
of ̺Y,X|Z (τ, ι). Let α
θn
i
i=1
12
Both can be obtained from a quantile regression straightforwardly. An empirical estimator
for ̺Y,X|Z (τ, ι) is
̺bY,X|Z (τ, ι) =
n−1
Pn
i=1
b
b ι (Xi − ZTi θ)
ψτ (Yi − ZTi α)ψ
p
.
τ (1 − τ )ι(1 − ι)
(6)
To study the asymptotic property of ̺bY,X|Z (τ, ι), we denote
∆11 = E{fY |Z (ZT α0 )ZZT }, ∆12 = E FX|Z,Y =ZT α0 (ZT θ 0 )fY |Z (ZT α0 )Z ,
∆21 = E FY |Z,X=ZT θ0 (ZT α0 )fX|Z (ZT θ 0 )Z , ∆22 = E{fX|Z (ZT θ 0 )ZZT },
Σ11 = E{FY,X|Z (ZT α0 , ZT θ 0 )}[1 − E{FY,X|Z (ZT α0 , ZT θ 0 )}],
Σ22 = E{ψτ2 (Y − ZT α0 )ZZT },
Σ12 = E{FY,X|Z (ZT α0 , ZT θ 0 )Z},
Σ33 = E{ψι2 (X − ZT θ 0 )ZZT },
Σ23 = E{ψτ (Y − ZT α0 )ψι (X − ZT θ 0 )ZZT },
where α0 and θ 0 are defined in (5).
We have the following asymptotic result for the CPC.
Theorem 3.1. Let 0 < a < b < 1. Suppose that ∆11 and ∆22 are uniformly positive
definite matrices in τ and ι, and there exists a constant π > 0 such that fY |Z (ZT α0 + ·),
fY |Z,X (ZT α0 + ·), fX|Z,Y (ZT θ 0 + ·) and fX|Z (ZT θ 0 + ·) are uniformly integrable on [−π, π]
and uniformly bounded away from zero and infinity in τ and ι. Then we have
√
n{b
̺Y,X|Z (τ, ι) − ̺Y,X|Z (τ, ι)}
w
GY,X|Z (τ, ι)
in ℓ∞ ([a, b]2 ), where GY,X|Z (τ, ι) is Gaussian process with mean zero and covariance function
Ω2 (τ1 , ι1 ; τ2 , ι2 ) ≡ E{[ζ(Y, X, Z; τ1 , ι1 )−Eζ(Y, X, Z; τ1 , ι1 )]×[ζ(Y, X, Z; τ2 , ι2 )−Eζ(Y, X, Z; τ2 , ι2 )]}
−1
T 0
T
and ζ(Y, X, Z; τ, ι) = [I(Y ≤ ZT α0 , X ≤ ZT θ 0 ) − ∆T12 ∆−1
11 I(Y ≤ Z α )Z − ∆21 ∆22 I(X ≤
p
ZT θ 0 )Z]/ τ (1 − τ )ι(1 − ι).
13
If Z ≡ 1, in another word, there is no conditional variable available, the asymptotic
distribution in Theorem 3.1 reduces to that in Theorem 2.1. The above result implies
√
d
̺Y,X|Z (τ, ι) → N(0, Ω2 ), where Ω2 ≡
that for a fixed pair (τ, ι), if ̺Y,X|Z (τ, ι) = 0, then nb
Ω2 (τ, ι; τ, ι) = E{[ζ(Y, X, Z; τ, ι)−Eζ(Y, X, Z; τ, ι)]2 } =
−1
1
[Σ11 +∆T12 ∆−1
11 Σ22 ∆11 ∆12 +
τ (1−τ )ι(1−ι)
−1
−1
−1
−1
−1
T
T
T
∆T21 ∆−1
22 Σ33 ∆22 ∆21 − 2(1 − τ )∆12 ∆11 Σ12 − 2(1 − ι)∆21 ∆22 Σ12 + 2∆12 ∆11 Σ23 ∆22 ∆21 ].
This theorem can be used for statistical inference if we can find a consistent estimate
of Ω2 . To this end, let e∗1 = Y − ZT α0 and e∗2 = X − ZT θ 0 and assume that the random
vectors (e∗1 , Z, X) and (e∗2 , Z, Y ) have joint densities fe∗1 ,Z,X and fe∗2 ,Z,Y , respectively. Denote by fe∗1 , fe∗2 , fe∗1 |Z , fe∗1 |Z,X , fe∗2 |Z and fe∗2 |Z,Y the marginal densities of e∗1 and of e∗2 , the
conditional densities of e∗1 given Z and (Z, X) and of e∗2 given Z and (Z, Y ), respectively.
Then, it can be verified that ∆11 = E{fe∗1 |Z (0)ZZT } = fe∗1 (0)E{ZZT |e∗1 = 0} and, similarly, ∆12 = fe∗1 (0)E{I(X ≤ ZT θ 0 )Z|e∗1 = 0}, ∆21 = fe∗2 (0)E{I(Y ≤ ZT α0 )Z|e∗2 = 0}
and ∆22 = fe∗2 (0)E{ZZT |e∗2 = 0}. To estimate these quantities, we first calculate the
b and then obtain the corresponding quantile residb and θ
quantile regression estimates α
b for i = 1, . . . , n. Next, we provide an estib and ê∗2i = Xi − ZTi θ
uals ê∗1i = Yi − ZTi α
mate for ∆12 and the estimators for ∆11 , ∆21 and ∆22 can be obtained similarly. We can
use nonparametric NW estimates used in estimating σX|Y (τ, ι) and σY |X (τ, ι) in Section
2.1 to obtain estimates for each component of m(s) = E{I(X ≤ ZT θ 0 )Z|e∗1 = s} using
b i ), i = 1, . . . , n}, and denote it by m(s).
b
Then we obtain
the data {(ê∗1i , I(Xi ≤ ZTi θ)Z
b 12 = fbe∗ (0)m(0),
b
∆
where fbe∗1 (0) is a nonparametric kernel density estimate for fe∗1 (0) in ∆12
1
b 12 is consistent under some regubased on {ê∗1i , i = 1, . . . , n}. It can be shown that such ∆
b 11 = n−1 Pn I(Yi ≤
larity conditions. For other unknown terms involved in Ω2 , we have Σ
i=1
P
P
n
Tb 2 b
−1
2
T
b − [n−1 n I(Yi ≤ ZT α,
b Xi ≤ ZTi θ)
b i ZTi ,
ZTi α,
i b Xi ≤ Zi θ)] , Σ22 = n
i=1
i=1 ψτ (Yi − Zi α)Z
P
P
Tb
b i ZT , Σ
b 12 = n−1 n I(Yi ≤ ZT α,
b 33 = n−1 n ψ 2 (Xi − ZT θ)Z
Σ
i b Xi ≤ Zi θ)Zi , and
i
i
i=1
i=1 ι
P
Tb
T
b 23 = n−1 n ψτ (Yi − ZT α)ψ
Σ
ι (Xi − Zi θ)Zi Zi . Using the plug-in approach, a consisi b
i=1
b 2.
tent estimate of Ω2 is thus obtained and denoted by Ω
14
The next theorem can be used to test whether ̺τ,ι (Y, X1 |Z) = ̺τ,ι (Y, X2 |Z) for two different random variables X1 and X2 . Write θ 0k = argminθ E{ρι (Xk − ZT θ)} for k = 1, 2 and
(k)
let ∆12 be ∆12 , where the involved X and θ 0 are replaced by Xk and θ 0k , respectively, for
(k)
(k)
(k)
(k)
(k)
(k)
k = 1, 2. In the same manner, we can define ∆21 , ∆22 , Σ11 , Σ33 , Σ12 , Σ23 and, accord(k)
ingly, Ω2
with k = 1, 2. In addition, we write ∆31 = E{FY,X1 ,X2 |Z (ZT α0 , ZT θ 01 , ZT θ 02 )} −
E{FY,X1 |Z (ZT α0 , ZT θ 01 )}E{FY,X2 |Z (ZT α0 , ZT θ 02 )}, ∆32 = E{FY,X1 ,X2 |Z (ZT α0 , ZT θ 01 , ZT θ 02 )Z}
and ∆33 = E{ψι (X1 − ZT θ 01 )ψι (X2 − ZT θ 02 )ZZT }.
(k)
Theorem 3.2. Let 0 < a < b < 1. Suppose that matrices ∆11 and ∆22 , k = 1, 2, are
uniformly positive definite in τ and ι, and there exists a constant π > 0 such that fY |Z (ZT α0 +
·), fY |Z,Xk (ZT α0 +·), fXk |Z,Y (ZT θ 0 +·) and fXk |Z (ZT θ 0 +·) are uniformly integrable on [−π, π]
for k = 1, 2 and uniformly bounded away from zero and infinity in τ and ι. Then
√
n{[b
̺Y,X1 |Z (τ, ι) − ̺bY,X2 |Z (τ, ι)] − [̺Y,X1 |Z (τ, ι) − ̺Y,X2 |Z (τ, ι)]}
w
GY,X1 ,X2 |Z (τ, ι)
in ℓ∞ ([a, b]2 ), where GY,X1 ,X2 |Z (τ, ι) is Gaussian process with mean zero and covariance function Ξ2 (τ1 , ι1 ; τ2 , ι2 ) ≡ E{[β(Y, X1 , X2 , Z; τ1 , ι1 )−Eβ(Y, X1 , X2 , Z; τ1 , ι1 )]×[β(Y, X1 , X2 , Z; τ2 , ι2 )−
Eβ(Y, X1 , X2 , Z; τ2 , ι2 )]} and β(Y, X1 , X2 , Z; τ, ι) = ζ(Y, X1, Z; τ, ι) − ζ(Y, X2 , Z; τ, ι), where
ζ(Y, X1, Z; τ, ι) is given in Theorem 3.1.
For fixed (τ, ι), if ̺τ,ι (Y, X1 |Z) = ̺τ,ι (Y, X2 |Z), then
(1)
(2)
√
d
n[b
̺Y,X1 |Z (τ, ι) − ̺bY,X2 |Z (τ, ι)] →
N(0, Ξ2 ), where Ξ2 ≡ Ξ2 (τ, ι; τ, ι) = Ω2 +Ω2 −2B12 and B12 ≡ B12 (τ, ι) =
(2)
(1)
(1)
(2)
(1)
(2)
1
[∆31 −
τ (1−τ )ι(1−ι)
(1)
(2)
(2)
(2)
−1
T −1
T −1
T −1
−1
(1−τ )(∆12 )T ∆−1
11 Σ12 −(1−τ )(∆12 ) ∆11 Σ12 +(∆12 ) ∆11 Σ22 ∆11 ∆12 +(∆12 ) ∆11 Σ23 (∆22 ) ∆21 +
(1)
(1)
(1)
(1)
(1)
(2)
(2)
(2)
(1)
(1)
(1)
(2)
T
−1
T
−1
−1
(∆21 )T (∆22 )−1 ∆32 +(∆21 )T (∆22 )−1 Σ23 ∆−1
11 ∆12 +(∆21 ) (∆22 ) (ιΣ12 −∆32 )+(∆21 ) (∆22 ) ∆33 (∆22 )
Given a sample of observations {(Yi , Xi1 , Xi2 , Zi ), i = 1, . . . , n}, the asymptotic variance Ξ2
b2 = Ω
b (1) + Ω
b (2) − 2B
b12 , where Ω
b (1) and Ω
b (2) are defined as Ω
b 2 given
can be estimated as Ξ
2
2
2
2
b12 , we only need to estimate ∆31 , Σ32 and Σ32 since the rest
above. To obtain the estimate B
of unknown quantities involved in B12 can be estimated using the previous methods. To this
15
Tb
b 31 = n−1 Pn I(Yi ≤ ZT α,
end, we can use the following estimates: ∆
i b Xi1 ≤ Zi θ 1 , Xi2 ≤
i=1
−1 Pn
Tb
T
b2 ) − n−1 Pn I(Yi ≤ ZT α,
b2 ) , ∆
b 32 =
b Xi2 ≤ ZTi θ
ZTi θ
i b Xi1 ≤ Zi θ 1 ) n
i=1
i=1 I(Yi ≤ Zi α,
Pn
P
Tb
−1
b1 , Xi2 ≤ ZT θ
b
b
b Xi1 ≤ ZTi θ
n−1 ni=1 I(Yi ≤ ZTi α,
i 2 )Zi , ∆33 = n
i=1 ψι (Xi1 − Zi θ 1 )ψι (Xi2 −
b2 )Zi ZT , where θ
bk = argmin n−1 Pn ρι (Xik − ZT θ), k = 1, 2.
ZTi θ
θ
i
i
i=1
3.2
CPC-based Variable Screening
We may now propose a joint robust screening using the CPC. There are two practical scenarios to favor joint screening over marginal screening. First, very often we may acquire
low-dimensional variables W ∈ Rr in addition to ultrahigh dimensional covariates X. For
example, when studying the relationship between a disease phenotype Y and genetic variables X, we may also have patient demographical information or environmental variables and
include them in W. Consequently, we have a data set {(Yi , Xi , Wi ), i = 1, . . . , n}. Second,
even if there is no external variables W, it may still be necessary to consider a joint screening
by removing the effects from correlated components in X. For instance, some covariates,
XSj , may be closely correlated to Xj and influence the observed correlation between Y and
Xj indirectly, where Sj is a subset of {1, · · · , pn } \ {j}. Ma et al. (2017) also considered a
set Sj which is referred to as a conditional set with relatively small size (< n). To account
for both scenarios, we may consider the conditional variables Z = (WT , XTSj )T in this paper.
We allow that the conditional variables Z to differ with j. However, for simplicity of presentation, we still use Z instead of Zj , and we denote by qn the dimension of Z. In principle, we
only need qn = max1≤j≤pn (r + |Sj |) for sure screening. In practice, we may select a proper
Sj as follows: Treat Xj as the response and X−j = {Xk , k 6= j, 1 ≤ k ≤ pn } as the predictors
and then apply any sensible marginal screening method such as the CC-SIS to pick out the
top ℓ most important predictors and set them as the conditional variables.
For ultrahigh dimensional covariates X = (X1 , · · · , Xpn )T , we can define CPC between
16
Y and the jth marker Xj given Z in the same way as in (5), namely,
̺Y,Xj |Z (τ, ι) =
E{ψτ (Y − ZT α0 )ψι (Xj − ZT θ 0j )}
p
,
τ (1 − τ )ι(1 − ι)
(7)
where α0 = argminα E{ρτ (Y − ZT α)} and θ 0j = argminθj E{ρι (Xj − ZT θ j )}. As in (6), a
sample estimate for ̺Y,Xj |Z (τ, ι) can be given as
̺bY,Xj |Z (τ, ι) =
b = argminα n1
where α
Pn
i=1
n−1
Pn
i=1
bj )
b ι (Xij − ZTi θ
ψτ (Yi − ZTi α)ψ
p
,
τ (1 − τ )ι(1 − ι)
(8)
bj = argmin 1 Pn ρι (Xij − ZT θ j ). The
ρτ (Yi − ZTi α) and θ
θj n
i
i=1
CPC screening yields the following empirical active set:
cb = {j : |b
M
̺Y,Xj |Z (τ, ι)| ≥ vn , 1 ≤ j ≤ pn },
(9)
where vn is a user-specified threshold parameter. We refer to this sure independence screening procedure as CPC-SIS. Clearly, CPC-SIS extends earlier conditional sure independence
screening such as Barut et al. (2016).
Let M∗b = {j : |̺Y,Xj |Z (τ, ι)| > 0, j = 1, . . . , p} be the true active set. We write
FY−1|X,W (τ ) = inf{y : P (Y ≤ y|X, W) ≥ τ }. For simplicity, we still use uj = |̺Y,Xj |Z (τ, ι)|
and u
bj = |b
̺Y,Xj |Z (τ, ι)| to denote the underlying and empirical CPC utilities, respectively.
To establish the sure screening property, we need the following conditions, which are very
mild and similarly imposed in Ma et al. (2017).
(D1) (i) The conditional density fY |Z=z (y) of Y given Z = z satisfies the Lipschitz condition
of order 1 and fY |Z=z (y) > 0 for any y in a neighborhood of ZT α0 = zT α0 .
(ii) For every 1 ≤ j ≤ pn , the conditional density fXj |Z=z (x) of Xj given Z = z satisfies
the Lipschitz condition of order 1 and fXj |Z=z (x) > 0 for any x in a neighborhood of
ZT θ 0 = zT θ 0 .
17
(D2) (i) There exist some finite constants m1 , m2 and m3 such that
max |Zij | ≤ m1 ,
i,j
max |ZTi α0 | ≤ m2 ,
i
max |ZTi θ 0j | ≤ m3 .
i,j
(ii) There exist two positive finite constants cmin and cmax such that
cmin ≤ λmin (E(ZZT )) ≤ λmax (E(ZZT )) ≤ cmax ,
where λmin (E(ZZT )) and λmax (E(ZZT )) stand for the minimum and maximum eigenvalues of E(ZZT ), respectively.
(D3) minj∈M∗b uj ≥ C0∗ n−κ for some κ > 0 and C0∗ > 0.
Theorem 3.3. (Screening Property for CPC-SIS) Suppose that the conditions (D1) and
(D2) hold,
(i) for any constant C > 0, then there exists some positive constant c̃∗1 such that for sufficiently large n,
P
bj − uj ≥ Cn−κ ≤ 12pn exp(−c̃∗1 qn−1 n1−2κ ).
max u
1≤j≤pn
(ii) In addition, if condition (D3) is further satisfied and by choosing vn = C2 n−κ with
C2 ≤ C0∗ /2, we have
∗
c
P Mb ⊂ Mb ≥ 1 − 12sn exp(−c̃∗1 qn−1 n1−2κ )
for sufficiently large n, where sn = |M∗b |.
When conditional variables are available, our proposed CPC-SIS method can handle the
dimensionality of order pn = o(exp(qn−1 n1−2κ )). If qn = O(1), then the dimension can be as
18
high as o(n1−2κ ), the same order as that of the CC-SIS. Moreover, the proposed CPC-SIS
can be readily used for the ultrahigh dimensional data as long as qn = o(n1−2κ ).
As in Section 2.2, we can determine a proper vn by controlling FDR. By Theorem 3.1, for
√ b −1/2
covariate j such that ̺Y,Xj |Z (τ, ι) = 0, we have nΩ
̺bY,Xj |Z (τ, ι) ∼ N(0, 1) asymptotically.
2
cb,δ = {j : √nΩ
b −1/2 |b
Then, we select variables M
̺Y,Xj |Z (τ, ι)| ≥ δ} for a small δ > 0, which
2
cb,δ ∩ (M∗ )c |/|(M∗)c |} under the following condition.
controls the FDR E{|M
b
b
(D4) For j ∈ (M∗a )c , we have that ̺Y,Xj |Z (τ, ι) = 0.
Proposition 3.4. (FDR Property) Under conditions (D1)-(D4) and the condition of Theorem 3.1, if we choose δ = Φ−1 (1 − d¯n /(2pn )) and Φ(·) and d¯n are the same as those in
Proposition 2.4, then for some constant cb > 0, we have
n |M
cb,δ ∩ (M∗ )c | o d¯n
√
b
E
≤
+ cb / n.
∗ c
|(Mb ) |
pn
4
Implementation on CPC-SIS
We provide more details on the implementation of CPC-SIS. We consider three practical
types of conditional variables in the following.
Case 1.
If W is not available, we consider the conditional variables from X itself for
each Xj , namely, Z = XSj for j = 1, . . . , pn . We start with an empty active set A(0) = ∅.
• Step 1. For j = 1, . . . , pn , select confounding sets Sjν ’s via the partial correlation based
consequential test (Ma et al. (2017)).
• Step 2. In the kth iteration, where k = 1, . . . , d∗ and d∗ = ⌊2(n/ log n)1/2 ⌋, for given
A(k−1) , we update Sj = A(k−1) ∪ Sjν and then find the variable index j ∗ such that
j ∗ = argmaxj6∈A(k−1) |b
̺Y,Xj |Z (τ, ι)|. Update A(k) = A(k−1) ∪ {j ∗ }.
19
• Step 3. In the kth iteration, where k = d∗ + 1, . . . , dn , we set Sj = A(d ) ∪ Sjν and then
∗
cb
find j ∗ = argmaxj6∈A(k−1) |b
̺Y,Xj |Z (τ, ι)|. Update A(k) = A(k−1) ∪ {j ∗ }. Use A(dn ) ≡ M
as the final set of selected covariates.
Case 2.
If W is available, we consider the same conditional variables for each target
Xj , namely, Z = W for j = 1, . . . , pn .
• Step 1. For j = 1, . . . , pn , compute the CPC utility statistics u
bj = |b
̺Y,Xj |Z (τ, ι)|.
• Step 2. Rank the covariates in terms of their u
bj ’s in a decreasing order and then select
the top dn covariates as the final set of selected covariates.
Case 3.
If W is available, we slightly modify the algorithm in Case 1. The steps
are the same as those in Case 1 only except that we consider the conditional variables
Z = (WT , XTSj )T in each iteration for 1 ≤ k ≤ dn for each step.
Case 1 only utilizes the confounding information from covariates X itself while Case 2
incorporates the exogenous conditional information but ignores the confounding effect from
X itself. Case 3 is a the most flexible version incorporating all types of covariate information.
We will implement Case 3 for the real data analysis in this paper.
5
5.1
Numerical Studies
Simulation Studies
In this section, we conduct simulations to examine the finite sample performances of the
two proposed copula-based correlations: CC and CPC, as well as the two so-constructed
screening procedures: CC-SIS and CPC-SIS.
20
5.1.1
Inference Performance
We consider two simulation examples with fixed dimension p in this subsection, and illustrate
the practical performance of estimated CC and CPC, respectively. We consider sample size
n = 200 and 400 and set the number of repetitions to be N = 5000 in Examples 1 and 2.
Example 1. We generate the response from two models
• (a1) Y = exp(2X1 ) + exp((2 + c0 )X2 ), where (X1 , X2 ) is from a standard bivariate
normal distribution with corr(X1 , X2 ) = ρ;
• (a2) Y = 2X01 + (2 + c0 )X02 + ε,
where in model (a2), covariates X1 and X2 are both generated from a mixture distribution
of a normal distribution with probability 0.9 and a cauchy distribution with probability
0.1, that is, X1 = 0.9X01 + 0.1ǫ1 and X2 = 0.9X02 + 0.1ǫ2 with (X01 , X02 ) following a
standard bivariate normal distribution with corr(X01 , X02 ) = ρ, ǫ1 ∼
1
Cauchy(0, 1)
5
and
ǫ2 ∼ 15 Cauchy(0, 1) are independent and the model error ε is generated from N(0, 1). Our
interest of this example is to test H0 : ̺Y,X1 (τ, ι) = ̺Y,X2 (τ, ι) at the significance level
α0 = 0.05 under various values of c0 . We consider different ρ’s and set c0 = 0, 1, 2, 4, where
c0 = 0 implies that H0 holds true, whereas H0 should be rejected with large probability for
other values of c0 . We report the empirical size and power for each setup over 5000 runs in
Table 1. Observing Table 1, we can see that our proposed CC testing procedure based on
Theorem 2.2 performs satisfactorily across different quantile levels (τ, ι) since the empirical
size are close to nominal level α0 = 0.05 and enlarging the sample size from 200 to 400
generally tends to improve the performance. Also we can see that when c0 runs away from
0, the empirical power increases to 1, and when the correlation between X1 and X2 is low,
the performance will be better.
Example 2. In this example, we generate conditional variables Z = (Z1 , Z2 , Z3 , Z4 )T from
the 4-dimensional multivariate normal distribution N(04 , Σ) with Σ = (ρ|j−k|)1≤j,k≤4. We
21
ρ
0
Method(τ, ι)
CC(0.25,0.25)
CC(0.5,0.5)
CC(0.75,0.75)
0.5
CC(0.25,0.25)
CC(0.5,0.5)
CC(0.75,0.75)
0.9
CC(0.25,0.25)
CC(0.5,0.5)
CC(0.75,0.75)
n
200
400
200
400
200
400
200
400
200
400
200
400
200
400
200
400
200
400
c0 = 0
0.056
0.051
0.059
0.054
0.057
0.055
0.055
0.056
0.057
0.053
0.054
0.052
0.055
0.052
0.061
0.050
0.045
0.047
Model (a1)
1
2
0.07
0.07
0.07
0.09
0.33
0.69
0.57
0.93
0.77
0.99
0.96
1.00
0.06
0.08
0.06
0.08
0.26
0.56
0.44
0.83
0.71
0.99
0.95
1.00
0.08
0.17
0.09
0.29
0.17
0.36
0.27
0.59
0.52
0.91
0.85
1.00
4
0.08
0.08
0.94
1.00
1.00
1.00
0.14
0.18
0.85
0.98
1.00
1.00
0.47
0.75
0.63
0.87
0.99
1.00
c0 = 0
0.051
0.055
0.057
0.048
0.053
0.050
0.048
0.052
0.050
0.049
0.052
0.043
0.048
0.052
0.048
0.048
0.053
0.049
Model (a2)
1
2
0.48
0.90
0.78
1.00
0.56
0.95
0.85
1.00
0.49
0.90
0.78
1.00
0.27
0.62
0.45
0.90
0.29
0.68
0.51
0.93
0.26
0.63
0.47
0.89
0.10
0.19
0.14
0.35
0.09
0.20
0.15
0.36
0.09
0.20
0.14
0.36
4
1.00
1.00
1.00
1.00
1.00
1.00
0.94
1.00
0.96
1.00
0.93
1.00
0.43
0.73
0.45
0.75
0.43
0.72
Table 1: Empirical size and power for testing H0 : ̺Y,X1 (τ, ι) = ̺Y,X2 (τ, ι) using CC across
(τ, ι) = (0.25, 0.25), (0.5, 0.5) and (0.75, 0.75) for Example 1.
generate Y from the model
• Y = 2X1 + (2 + c0 )X2 + ZT b + ε,
where b = (3, 4, 3, 4)T , X1 = ZT b + ǫ1 and X2 = ZT b + ǫ2 , where ǫ1 ∼ 13 t(3) and ǫ2 ∼ 31 t(3)
are independent. The other setups are the same as Example 1. In this example, our interest
is to test H0 : ̺Y,X1 |Z (τ, ι) = ̺Y,X2 |Z (τ, ι) at the significance level α0 = 0.05. Here, we consider
the performance of CPC with (τ, ι) = (0.5, 0.5) and the corresponding empirical size and
power are reported in Table 2. A similar conclusion can be drawn as in Example 1. These
numerical results empirically demonstrate that our theoretical result in Theorem 3.2 is valid.
5.1.2
Screening Performance for CC-SIS and CPC-SIS
Throughout this subsection, we adopt the following simulation setup: the sample size
n = 200, the covariate dimension p = 1000, and the number of simulations N = 200
22
ρ
0
0.5
0.9
n
200
400
200
400
200
400
c0 = 0
0.032
0.035
0.035
0.036
0.032
0.044
ε ∼ N (0, 1)
1
2
0.22
0.57
0.51
0.94
0.22
0.57
0.50
0.94
0.22
0.59
0.50
0.94
4
0.90
1.00
0.90
1.00
0.91
1.00
c0 = 0
0.033
0.041
0.037
0.048
0.042
0.050
ε ∼ 13 Cauchy(0, 1)
1
2
0.25
0.61
0.58
0.95
0.26
0.61
0.59
0.95
0.26
0.61
0.58
0.96
4
0.87
1.00
0.88
1.00
0.87
1.00
Table 2: Empirical size and power for testing H0 : ̺Y,X1 |Z (τ, ι) = ̺Y,X2 |Z (τ, ι) using CPC with
(τ, ι) = (0.5, 0.5) in Example 2.
for each simulation setup. Moreover, for the purpose of comparison, we use three criteria
for evaluation: the first criterion is the minimum model size (MMS), namely, the smallest
number of the selected covariates that contain all the active covariates, and its robust standard deviation (RSD); the second is the rank for each active covariates (Rj ); and the third
is the proportion of all the active covariates being selected (P) with the screening threshold
specified as ⌊n/ log n⌋ over N simulations. We report the median of MMS and Rj .
In Example 3, we compare our CC-SIS methods with a few existing methods: SIS
(Fan and Lv (2008)), SIRS (Zhu et al. (2011)), DC-SIS (Li et al. (2012b)), Kendall-SIS
(Li et al. (2012a)), QC-SIS (Li et al. (2015)) and CQC-SIS (Ma and Zhang (2016)). In Example 4, we compare our CPC-SIS procedure with the aforementioned marginal screening
methods as well as the QPC-SIS (Ma et al. (2017)), where confounding effects arise from
covariates X. In this example, we employ the algorithm in Case 1 given in Section 4 in order
to compare with the QPC-SIS by Ma et al. (2017). In Example 5, we apply the algorithm
in Case 2 and compare with the QPC-SIS by Ma et al. (2017).
Example 3. This example is used to assess the performance of the proposed CC-SIS.
Let X∗ = (X1∗ , . . . , Xp∗n )T be a latent random vector having the pn -dimensional normal
distribution N(0pn , Σ) with Σ = (ρ|k−l| )1≤k,l≤pn , where we set the correlation ρ = 0.4 and 0.8.
We write ǫ = (ǫ1 , . . . , ǫpn )T with each component ǫj being independent of other components
and having the standard Cauchy distribution, i.e., ǫj ∼ Cauchy(0, 1). We generate covariates
X from a mixture distribution: X = 0.8X∗ + 0.2ǫ, and simulate the response data from the
23
following three models:
• (b1) Y = 3X1∗ + 3X2∗ + 2X3∗ + 2X4∗ + 2X5∗ + ε,
∗
• (b2) Y = 5X1∗ I(X1∗ < 0) + 5X2∗ I(X2∗ > 0) + 5 sin(X10
) + ε,
• (b3) Y = exp{3β1 sin(X1∗ ) + 2β2 exp(X2∗ ) + 1.5β3 I(X3∗ > 0) + 2 log(|X4∗ |)} + ε,
where ε is simulated from two scenarios: ε ∼ N(0, 1) and ε ∼ Cauchy(0, 1) and, in model
√
(b3), we set βj = cj (−1)U (4 log n/ n + Z0 ) for j = 1, 2 and 3, where U ∼ Bernoulli(0.4),
Z0 ∼ N(0, 0.52 ) and (c1 , c2 , c3 ) = (1, 0.5, 1). The resulting screening results in terms of MMS
and P are presented in Table 3.
Eyeballing Table 3, we can make some key observations. Under models (b1) and (b2),
our CC-SIS outperforms SIS, SIRS, DC-SIS and QC-SIS.In this case, both response and
covariates are heavy-tailed and thus traditional linear correlation screening methods all fail
to work. Our methods are also comparable to the nonparametric Kendall’s τ which achieves
high accuracy but is slower due to the numerical integration in its implementation. (b3) is a
difficult case and very hard to screen accurately. Under this case, our CC-SIS still performs
much better than other methods.
Example 4. This example is designed to evaluate the performance of the proposed CPCSIS. We generate the response from the following two models
√
• (d1) Y = βX1 + βX2 + βX3 − 3β ρX4 + ε,
√
• (d2) Y = βX1∗ + βX2∗ + βX3∗ − 3β ρX4∗ + ε,
where β = 4 for both models. In model (d1), we consider the observed covariates X ∼
N(0pn , Σ) with Σ = (σij )1≤i,j≤pn , implying that covariates X have no outliers and the response Y is fully dependent on observed X up to a random noise. Under this setting, it is
desired to expect that the QPC-SIS performs better than our CPC-SIS since X is normal.
24
Method(τ, ι)
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
CQC-SIS
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
CQC-SIS
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
CQC-SIS
ρ = 0.4
ε ∼ N (0, 1)
ε ∼ Cauchy(0, 1)
MMS(RSD)
P
MMS(RSD)
P
Model (b1)
655 (396)
0.03
790 (227)
0.00
615 (176)
0.00
669 (158)
0.00
634 (441)
0.09
725 (350)
0.04
5 (0)
0.99
5 (1)
0.96
9 (18)
0.80
20 (41)
0.70
6 (7)
0.93
9 (14)
0.84
9 (15)
0.85
14 (40)
0.70
710 (315)
0.01
782 (271)
0.00
716 (334)
0.01
790 (250)
0.02
711 (297)
0.01
751 (314)
0.01
573 (156)
0.00
614 (149)
0.00
Model (b2)
482 (408)
0.13
638 (376)
0.01
551 (161)
0.00
589 (195)
0.00
284 (585)
0.30
607 (493)
0.12
3 (0)
1.00
3 (0)
1.00
6 (20)
0.78
13 (45)
0.68
3 (1)
0.98
4 (3)
0.95
7 (15)
0.85
14 (43)
0.66
615 (356)
0.06
588 (391)
0.05
584 (460)
0.07
640 (373)
0.05
588 (414)
0.03
579 (353)
0.02
512 (153)
0.00
562 (155)
0.00
Model (b3)
737 (296)
0.00
768 (204)
0.00
767 (215)
0.00
745 (182)
0.00
790 (245)
0.00
789 (213)
0.01
504 (467)
0.12
475 (442)
0.10
475 (234)
0.02
679 (317)
0.02
483 (357)
0.09
478 (337)
0.07
426 (419)
0.17
281 (460)
0.19
791 (199)
0.00
795 (191)
0.00
785 (210)
0.00
810 (209)
0.01
762 (257)
0.00
793 (240)
0.00
743 (216)
0.00
749 (177)
0.00
ρ = 0.8
ε ∼ N (0, 1)
ε ∼ Cauchy(0, 1)
MMS(RSD)
P
MMS(RSD)
P
522
513
460
5
5
5
5
649
665
599
442
(431)
(149)
(527)
(0)
(0)
(0)
(0)
(444)
(438)
(379)
(118)
0.12
0.00
0.20
1.00
1.00
1.00
1.00
0.06
0.07
0.04
0.00
608
547
615
5
5
5
5
678
693
670
474
(389)
(138)
(489)
(0)
(0)
(0)
(0)
(406)
(398)
(408)
(104)
0.03
0.00
0.13
1.00
1.00
1.00
1.00
0.03
0.06
0.02
0.00
351
523
230
4
5
4
6
411
539
486
441
(414)
(173)
(569)
(1)
(2)
(2)
(3)
(418)
(474)
(374)
(122)
0.15
0.00
0.35
1.00
0.99
1.00
0.99
0.07
0.13
0.09
0.00
686
543
506
4
7
5
7
598
592
604
481
(331)
(158)
(514)
(1)
(4)
(1)
(3)
(411)
(425)
(401)
(150)
0.03
0.00
0.17
1.00
0.94
1.00
0.96
0.05
0.10
0.06
0.00
524 (399)
664 (240)
722 (276)
5 (192)
252 (365)
18 (341)
12 (309)
794 (222)
750 (266)
698 (408)
623 (210)
0.01
0.00
0.02
0.65
0.21
0.56
0.59
0.01
0.02
0.01
0.00
619 (368)
703 (203)
719 (309)
16 (377)
279 (455)
53 (357)
113 (349)
781 (225)
796 (219)
714 (311)
628 (260)
0.01
0.00
0.01
0.60
0.14
0.47
0.47
0.01
0.04
0.02
0.00
Table 3: Simulation results for Example 3, where MMS stands for the median of the minimum
model size and its robust standard deviations (RSD) are given in parenthesis, P is the proportion
of screened sets that cover all active predictors with screening parameter dn = ⌊n/ log n⌋.
25
In model (d2), we generate covariates X from a mixture distribution 0.9X∗ + 0.1ǫ, where
X∗ ∼ N(0pn , Σ) with Σ = (σij )1≤i,j≤pn , and each element of ǫ is independent and distributed
as 15 Cauchy(0, 1). In this model, the covariates are contaminated with outliers, while the
heterogenicity of response Y stems merely from the random error ε. In addition, in model
√
(d1), we let σii = 1 and σij = ρ, j 6= i except that σ4j = σj4 = ρ. In model (d1), we set
σij = 0 if i > 4 or j > 4 and the rest are the same as model (d1). For both models, at the
population level, the covariate X4 is marginally uncorrelated with Y . We consider two cases
of ρ = 0.95 and 0.5 for simulation comparison.
Tables 4 and 5 report the screening results regarding the rank Rj and MMS. From the
tables, we can see that all the marginal screening approaches fail to pick out the covariate
X4 with very large values of the rank R4 . QPC-SIS and our CPC-SIS work much better
than those marginal methods. Particularly, under model (d1), our CPC-SIS has a very
competitive performance to QPC-SIS. Under model (d2), when the covariates are highly
correlated (ρ = 0.95), our proposal CPC-SIS(0.5,0.5) has the best performance.
Example 5. In this example, we examine the case of Z = W. Since conditional variables
selected for each Xj are the same, so we can simply employ our proposed CPC-SIS and
the QPC-SIS of Ma et al. (2017) for variable screening. We generate the response from the
model
• Y = 2X1 + 2X2 − 4X3 + 3X4 + ε,
where Xj = WT b+Uj , W has the same distribution as Z in Example 2, b = (2, 4/3, 2, 4/3)T
and Uj ∼ 13 Cauchy(0, 1) for j = 1, · · · , pn . The model error ε is the same as that in Example
4. The simulation results are given in Table 6. As expected, we can see that all the marginal
screening procedures fail to work since they are unable to identify the covariate X3 . Our
proposed CPC-SIS still performs better than QPC-SIS in terms of MMS.
26
ρ
Method(τ, ι)
R1
R2
0.5
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
2
3
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
28
3
3
91
56
114
3
3
3
2
2
2
2
2
2
2
3
2
2
2
2
2
2
2
2
2
2
2
2
2
2
3
52
3
3
95
49
92
3
3
3
2
2
2
2
2
2
0.95
ε ∼ N (0, 1)
R3
R4
MMS (RSD)
Model (d1)
2
450
450 (235)
3
469
501 (243)
2
497
497 (280)
2
454
454 (228)
2
367
372 (331)
2
415
425 (272)
2
371
387 (324)
2
478
493 (270)
2
437
437 (236)
2
437
438 (227)
2
4
4 (0)
2
4
4 (0)
2
4
4 (0)
2
4
4 (0)
2
4
4 (0)
2
4
4 (0)
3
470
498 (306)
45
474
541 (331)
3
506
556 (220)
3
482
529 (296)
104 303
481 (276)
67
365
499 (383)
94
297
470 (301)
3
470
512 (387)
3
465
499 (346)
3
468
520 (379)
2
4
4 (0)
2
4
4 (0)
2
4
4 (0)
2
4
4 (1)
2
4
4 (0)
2
4
4 (1)
R1
ε ∼ 13 Cauchy(0, 1)
R2
R3
R4
MMS (RSD)
3
3
2
2
2
2
2
2
2
2
2
2
2
2
2
2
214
24
10
3
88
53
125
3
3
3
2
2
2
2
2
2
3
3
2
2
3
2
3
2
2
2
2
2
2
2
2
2
269
26
12
3
99
68
63
3
3
3
2
2
2
2
2
2
3
4
2
2
2
2
2
2
2
2
2
2
2
2
2
2
262
43
15
3
88
47
138
3
3
3
2
2
2
2
2
2
411
432
464
467
382
396
378
473
461
408
4
4
4
4
4
4
466
473
495
486
318
354
318
467
467
469
4
4
4
5
4
4
476
466
464
468
390
408
388
477
463
421
4
4
4
4
4
4
735
584
610
557
485
610
492
522
506
508
4
4
4
5
4
4
(270)
(258)
(220)
(280)
(311)
(275)
(311)
(245)
(331)
(247)
(0)
(0)
(0)
(0)
(0)
(0)
(353)
(376)
(240)
(325)
(292)
(377)
(289)
(375)
(382)
(329)
(0)
(0)
(0)
(1)
(0)
(1)
Table 4: Simulation results for Example 4 (d1), where Rj indicates the median of the rank of
the relevant predictors and MMS stands for the median of the minimum model size and its robust
standard deviations (RSD) are given in parenthesis.
27
ρ
Method(τ, ι)
R1
R2
0.5
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
2
99
2
2
3
3
3
3
2
3
3
2
3
2
3
2
182
355
238
192
266
292
262
313
274
259
112
106
148
14
5
11
2
93
2
2
3
3
3
3
2
3
3
3
3
2
3
2
169
388
191
154
281
224
265
295
304
341
150
147
164
11
5
13
0.95
ε ∼ N (0, 1)
R3
R4
MMS (RSD)
Model (d2)
2
530
566 (359)
100 453
485 (304)
2
558
629 (393)
2
514
514 (362)
3
437
442 (362)
3
339
348 (358)
3
440
446 (315)
3
499
594 (352)
2
514
583 (377)
3
532
588 (345)
3
2
4 (1)
3
2
4 (0)
3
2
4 (1)
2
4
4 (2)
3
1
4 (0)
2
4
4 (1)
200 557
712 (285)
375 507
729 (290)
236 540
762 (335)
181 534
675 (274)
268 443
695 (197)
303 473
672 (305)
265 441
694 (199)
338 595
791 (259)
308 529
795 (283)
279 529
746 (276)
158
4
530 (465)
103
3
518 (524)
214
3
574 (437)
11
3
270 (478)
6
1
110 (404)
15
3
225 (453)
R1
ε ∼ 13 Cauchy(0, 1)
R2
R3
R4
MMS (RSD)
10
107
2
2
4
2
3
3
2
3
3
3
3
2
3
2
518
450
347
224
254
318
268
393
291
362
163
148
146
13
4
13
12
110
2
2
4
3
4
3
3
3
3
3
3
2
3
2
434
401
329
222
260
305
423
394
282
374
186
102
196
16
4
39
14
112
2
2
4
3
5
2
2
3
3
3
3
2
3
3
435
403
365
196
255
213
284
357
290
376
162
96
126
34
4
14
535
478
554
578
453
459
449
531
501
460
2
2
2
4
2
4
496
507
507
482
438
462
431
480
480
459
5
3
6
6
1
5
667 (351)
507 (356)
638 (322)
578 (381)
456 (333)
463 (343)
451 (333)
618 (363)
533 (322)
537 (337)
4 (1)
4 (0)
4 (3)
5 (3)
4 (0)
5 (3)
770 (282)
720 (305)
825 (262)
734 (309)
688 (205)
666 (300)
693 (195)
755 (280)
753 (258)
806 (268)
583 (475)
514 (535)
591 (483)
311 (591)
24 (371)
299 (537)
Table 5: Simulation results for Example 4 (d2), where Rj indicates the median of the rank of
the relevant predictors and MMS stands for the median of the minimum model size and its robust
standard deviations (RSD) are given in parenthesis.
28
ρ
0.5
0.95
Method(τ, ι)
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
SIS
SIRS
DC-SIS
Kendall-SIS
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CPC-SIS(0.25,0.25)
CPC-SIS(0.5,0.5)
CPC-SIS(0.75,0.75)
R1
12
192
336
2
3
3
3
183
269
223
6
4
5
5
5
5
5
190
441
2
3
3
3
326
400
312
5
4
5
6
5
6
R2
12
215
305
2
3
4
3
249
241
174
7
5
7
4
6
8
9
231
303
3
4
4
4
225
244
168
6
5
5
5
6
6
ε ∼ N (0, 1)
R3
R4
MMS (RSD)
377
4
488 (518)
910 188
910 (96)
511 320
744 (162)
997
1
997 (15)
841
2
841 (235)
886
2
886 (175)
817
2
817 (240)
713 208
796 (172)
672 307
823 (160)
701 259
829 (179)
3
3
107 (167)
2
3
53 (95)
3
3
75 (148)
5
1
28 (58)
1
2
14 (27)
6
1
41 (88)
538
4
581 (378)
894 199
894 (95)
508 239
771 (171)
991
1
991 (32)
806
2
806 (247)
831
2
831 (250)
804
2
804 (314)
624 157
795 (188)
597 226
804 (189)
650 143
785 (185)
3
3
73 (165)
3
3
47 (112)
2
3
52 (146)
6
1
31 (47)
1
2
20 (40)
5
1
41 (80)
R1
14
236
305
2
5
3
4
178
232
154
7
5
8
7
6
7
17
246
189
2
4
3
4
95
135
73
12
4
8
6
6
6
ε ∼ 13 Cauchy(0, 1)
R2
R3
R4
MMS (RSD)
14
396
5
532 (494)
222 938 193
938 (98)
272 546 324
746 (154)
2
998
1
998 (10)
3
830
2
830 (263)
3
895
2
895 (207)
5
865
2
865 (184)
148 748 212
806 (170)
231 687 296
824 (172)
191 731 209
822 (174)
9
3
4
109 (187)
5
3
5
77 (118)
6
3
4
94 (165)
8
8
2
62 (110)
6
1
2
19 (34)
9
7
2
57 (75)
11
525
4
586 (445)
220 895 186
895 (103)
273 558 348
776 (174)
2
990
1
990 (46)
3
825
2
825 (254)
4
836
2
836 (257)
5
758
2
758 (315)
202 682 252
805 (192)
235 661 318
812 (193)
154 695 272
815 (178)
6
3
4
128 (214)
8
3
3
78 (115)
6
3
4
105 (182)
5
7
2
55 (90)
5
1
2
18 (42)
7
7
2
46 (89)
Table 6: Simulation results for Example 5, where Rj indicates the median of the rank of the
relevant predictors and MMS stands for the median of the minimum model size and its robust
standard deviations (RSD) are given in parenthesis.
29
5.2
5.2.1
Real Data Applications
Rats Data
We illustrate the CC-SIS and CPC-SIS with the gene expression data on 120 male rats
of 12 weeks old, including expression measurements of 31,099 gene probes. It has been
analyzed in Scheetz et al. (2006) for investigation of the gene regulation in the mammalian
and is available at ftp://ftp.ncbi.nlm.nih.gov/geo/series/GSE5nnn/GSE5680/matrix.
We follow Ma et al. (2017) and consider the expression of gene TRIM32 (probe 1389163 at)
as the response variable Y since it was identified to cause Bardet-Biedl syndrome, closely
associated with the human hereditary disease of the retina Chiang et al. (2006). The other
gene probes are treated as the covariates X. We first apply Iglewicz and Hoaglin (1993)’s
(IH) approach to check outliers. IH constructs a Z-score Zi = 0.6745(xi − x̄)/MAD, where
MAD denotes the median absolute deviation, and recommends that any i such that Zi > 3.5
be labeled as outliers. IH method is quite popular in real applications such as engineering.
Consequently we find that there are over 60% gene probes having one or more outliers. Figure
1 displays the box-plots for two selected genes and the the response. If one only employs the
conventional screening method ignoring the outliers, it would lead to inappropriate results.
The copula-based methods may thus be more robust in this situation. In this data analysis,
we have the sample {(Yi , Xi ∈ Rpn ), 1 ≤ i ≤ n} with n = 120 and pn = 31098.
We report the overlaps of the top ⌊n/ log n⌋ = 25 selected genes by various methods in
Table 7. We can see that different methods select quite different genes and such low level
of agreement should not be overlooked in practice. Robust and joint screening methods like
what we propose in this paper lead to entirely different set of genes which are otherwise
screened out by the conventional non-robust and marginal screening approaches. We notice
that the CPC-SIS and QPC-SIS have a couple of overlaps, partly because both are conditional
screening procedures and able to adjust the confounder effects.
30
rats data
breast cancer data
4
name
gene ID 104
gene ID 586
Y
6
Gene Expressions
Gene Expressions
8
name
2
gene ID 40
gene ID 54
Y
0
4
−2
gene ID 104 gene ID 586
Y
gene ID 40 gene ID 54
Y
Figure 1: Box-plots for the response and two randomly selected genes for the two datasets.
The left panel is for the rats data and the right panel is for the breast cancer data.
Table 8 gives a summary of top 10 gene probes by different methods along with the pvalue resulted from a marginal Wald-test. We then use these 10 genes as regressors and build
a joint statistical models to predict Y . Linear regression and quantile regression are both
considered for this purpose and we displays the mean of their prediction errors (PE1 and
PE2) over 500 random partitions, where the partition ratio of training sample to test sample
is 4 : 1 for each partition. The PE is computed as the average of {(Yi − Ybi )2 , i ∈ testing set}
and Ybi is the predicted value at the ith test data point using the model constructed by the
training sample with the 10 genes in Table 8. We can see that our proposed copula-based
partial correlation screening performs the best with the smallest prediction error. Such a
nice prediction result may be attributed to the fact that CPC selects appropriate markers for
joint modelling after addressing the distribution heterogeneity and the conditional effects.
The heterogeneity problem typically inflates the variance while a purely marginal screener
could introduce bias. The prediction error, consists of the variance and the bias components,
is thus much smaller after employing the CPC screening method.
31
QC-SIS(τ )
QPC-SIS(τ )
CC-SIS(τ, ι)
CPC-SIS(τ, ι)
SIS SIRS DC-SIS Kendall-SIS 0.25 0.5 0.75 0.25 0.5 0.75 (0.25,0.25) (0.5,0.5) (0.75,0.75) (0.25,0.25) (0.5,0.5) (0.75,0.75)
SIS
25
0
1
3
1
1
0
0
0
0
1
3
3
0
0
0
SIRS
0
25
0
0
0
0
0
0
0
0
0
0
0
0
0
0
DC-SIS
1
0
25
1
2
2
1
0
0
0
0
0
1
0
0
0
Kendall
3
0
1
25
5
12
3
0
0
0
2
5
3
0
0
0
QC-SIS(0.25)
1
0
2
5
25
5
0
0
0
0
3
2
1
0
0
0
QC-SIS(0.5)
1
0
2
12
5
25
3
0
0
0
1
7
1
0
0
0
QC-SIS(0.75)
0
0
1
3
0
3
25
0
0
0
0
1
1
0
0
0
QPC-SIS(0.25)
0
0
0
0
0
0
0
25
3
2
0
0
1
2
1
0
QPC-SIS(0.5)
0
0
0
0
0
0
0
3
25
0
0
0
0
1
1
0
QPC-SIS(0.75)
0
0
0
0
0
0
0
2
0
25
0
0
0
0
1
0
CC-SIS(0.25,0.25)
1
0
0
2
3
1
0
0
0
0
25
1
0
0
0
0
CC-SIS(0.5,0.5)
3
0
0
5
2
7
1
0
0
0
1
25
1
0
0
0
CC-SIS(0.75,0.75)
3
0
1
3
1
1
1
1
0
0
0
1
25
0
0
0
CPC-SIS(0.25,0.25) 0
0
0
0
0
0
0
2
1
0
0
0
0
25
0
0
CPC-SIS(0.5,0.5)
0
0
0
0
0
0
0
1
1
1
0
0
0
0
25
1
CPC-SIS(0.75,0.75) 0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
25
Table 7: The overlaps of selected genes using various approaches for the rats data, where the
screening threshold parameter is set as ⌊n/ log n⌋ = 25 for each method and the CPC-SIS applies
the algorithm in Case 1.
Rank
1
2
3
4
5
6
7
8
9
10
PE1
PE2
SIS
SIRS
DC-SIS
Kendall-SIS
QC-SIS(0.5)
ID
p-value
ID
p-value
ID
p-value
ID
p-value
ID
p-value
14770 2.1E-05 2828
1.000
146
6.4E-04 6083 8.3E-10 22641 5.2E-11
21977 1.8E-05 20503 0.480
260
7.2E-04 5002 4.5E-10 14810 6.4E-12
6436 2.0E-08
233
0.152 30768 2.3E-06 14726 4.4E-16 22339 8.0E-13
4797 1.7E-08 3962
0.716 30745 1.6E-05 14810 6.4E-12 5002 4.5E-10
21150 2.3E-07 7656
0.063
285
1.2E-04 25297 1.5E-11 20898 7.2E-14
25573 4.5E-10 20453 0.468 30791 1.6E-07 5259 6.5E-12 31008 1.5E-07
12127 9.4E-09 22023 0.047
3849 1.1E-04 5223 2.9E-10 26828 1.8E-08
9235 1.6E-07
157
0.208
4626 1.1E-04 31008 1.5E-07 24529 4.2E-10
3682 2.5E-06 2575
0.153
4490 2.8E-10 22339 8.0E-13 14414 1.5E-07
8670 4.5E-11 2841
0.284
3967 2.1E-07 6021 1.7E-07 20724 2.8E-10
0.0394
0.0252
0.0290
0.0310
0.0283
0.0377
0.0269
0.0349
0.0342
0.0344
CC-SIS(0.5,0.5)
ID
p-value
14726
4.4E-16
6889
6.6E-14
14701
6.2E-14
20898
7.2E-14
22339
8.0E-13
23278
6.2E-13
25117
8.8E-14
30548
6.9E-14
4512
8.0E-12
4712
8.4E-12
0.0330
0.0360
QPC-SIS(0.5)
ID
p-value
18602
0.473
4101
0.003
12365
0.141
8399
1.000
5063
0.026
9223
0.467
21746
0.716
14019
0.717
30361
0.277
24759
0.010
0.0269
0.0307
CPC-SIS(0.5, 0.5)
ID
p-value
1621
0.469
11288
0.001
12480
1.000
4398
0.271
29604
0.467
22679
1.000
22267
0.065
17039
0.148
11796
0.720
20967
0.026
0.0247
0.0257
Table 8: Summary of top 10 gene probes selected by different screening methods for the rats data.
√ b −1/2
̺bY,X (0.5, 0.5)|)),
ID means the selected gene ID and p-values are computed as 2(1 − Φ(| nΩ
1
where Φ is the cumulative distribution function of standard normal random variable. PE is the mean
of prediction errors over 500 random partitions with the partition ratio of training sample to testing
sample being 4 : 1, where prediction error is defined as the average of {(Yi − Ybi )2 , i ∈ testing set}.
PE1 and PE2 indicate that Ybi is the predicted value via fitting a median regression model and
linear model, respectively, using the top 10 genes selected.
32
5.2.2
Breast Cancer Data
Breast cancer was the second most common cancer in the world and the most leading cause
in women. There were nearly 1.7 million new cases diagnosed in 2012, according to the
worldwide statistics1 . Meanwhile, approximately 252,710 new cases of invasive breast cancer
and 40,610 breast cancer deaths are expected to occur among US women in 2017 as reported
in DeSantis et al. (2017). Although major progresses in breast cancer treatment were made,
there is also limited ability to predict the metastatic behavior of tumor. Van’t Veer et al.
(2002) was the first to study the breast cancer study involving 97 lymph node-negative breast
cancer patients 55 years old or younger, of which 46 developed distant metastases within 5
years (metastatic outcome coded as 1) and 51 remained metastases free for at least 5 years
(metastatic outcome coded as 0). This expression data set with clinical variables has been
well analyzed in many papers for classification (Boulesteix et al. (2008), Yu et al. (2012),
among others).
In this study, after removing the genes with missing values, there are expression levels of
24,188 gene probes entering into the next analysis. In addition to gene expression measurements, the data for several clinical factors are available as well. Our interest is to identify
which gene probes affect the tumor size given other clinical factors (W) including age, histological grade, angioinvasion, lymphocytic infiltration, estrogen receptor and progesterone
receiptor status. Therefore, we have the data {(Yi , Xi ∈ Rpn , Wi ∈ Rr ), 1 ≤ i ≤ n} with
n = 97, pn = 24, 188 and r = 6 for further analysis.
Using the IH method on outlier detection, we find that 18,098 gene probes have at least
1 and at most 29 outliers, suggesting that approximately three quarters of overall gene
probes contain extremely large values. The right panel of Figure 1 displays the empirical
distribution of the response and two typical covariates. Thus, it is more suitable to apply
robust joint screening approach such as the proposed CPC-SIS. We consider the three cases
1
http://www.wcrf.org/int/cancer-facts-figures/worldwide-data
33
SIS
SIRS
DC-SIS
Kendall
QC-SIS(0.25)
QC-SIS(0.5)
QC-SIS(0.75)
QPC-SIS(0.25)
QPC-SIS(0.5)
QPC-SIS(0.75)
CC-SIS(0.25,0.25)
CC-SIS(0.5,0.5)
CC-SIS(0.75,0.75)
CPC-SISa1(0.5,0.5)
CPC-SISa2(0.5,0.5)
CPC-SISa3(0.5,0.5)
QC-SIS(τ ) QPC-SIS(τ )
CC-SIS(τ, ι)
CPC-SISa1 CPC-SISa2 CPC-SISa3
SIS SIRS DC-SIS Kendall-SIS 0.25 0.5 0.75 0.25 0.5 0.75 (0.25,0.25) (0.5,0.5) (0.75,0.75) (0.5,0.5)
(0.5,0.5)
(0.5,0.5)
21
8
7
4
1
7
5
1
1
0
2
1
1
0
1
0
8
21
13
9
1 15 2
1
0
0
3
5
0
0
1
0
7
13
21
11
0
8
3
0
0
0
1
6
0
0
1
0
4
9
11
21
0
3
4
0
0
0
0
3
0
0
0
0
1
1
0
0
21 1
0
0
0
0
1
0
0
0
0
0
7
15
8
3
1 21 0
1
0
0
3
7
0
0
1
0
5
2
3
4
0
0 21
0
1
0
0
0
5
0
0
0
1
1
0
0
0
1
0
21 0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0 21
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
21
0
0
0
0
0
0
2
3
1
0
1
3
0
0
0
0
21
1
0
0
0
0
1
5
6
3
0
7
0
0
0
0
1
21
0
0
1
0
1
0
0
0
0
0
5
0
0
0
0
0
21
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
21
0
0
1
1
1
0
0
1
0
0
0
0
0
1
1
0
21
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
21
Table 9: The overlaps of selected genes probes using various approaches for the breast cancer
data, where the screening threshold parameter is set as dn = ⌊n/ log n⌋ = 21 for each method. The
CPC-SISa1 means the CPC-SIS in Case 1, the CPC-SISa2 indicates the CPC-SIS in Case 2 and
the CPC-SISa3 stands for the CPC-SIS in Case 3.
discussed in Section 4 and denote the methods as CPC-SISa1 , CPC-SISa2 and CPC-SISa3 ,
respectively). Table 9 gives the overlaps of the selected genes by various methods. A similar
conclusion to that in the rats data analysis can be made. Furthermore, Table 10 presents a
summary of top 10 gene probes selected by various methods. The results on PE1 and PE2 in
Table 10 empirically verifies that our proposed CPC-SIS in Case 3 has the most satisfactory
performance in out-of-sample prediction.
6
Conclusion and Discussion
We propose a copula-based correlation and partial correlation to facilitate robust marginal
and joint screening for ultrahigh dimensional data sets. Large sample properties for the
estimated correlation and sure screening properties for CC and CPC screeners were provided.
Empirical studies including simulations and two data applications show that our proposed
CC-SIS and CPC-SIS outperform the existing variable screening approaches, when outliers
are present in both covariates and response. Therefore, our current proposals are more
applicable to the ultrahigh dimensional heterogeneous data. We provide a guideline to carry
out variable screening as follows. If the response and predictors are all normal without
34
SIS
Rank
1
2
3
4
5
6
7
8
9
10
PE1
PE2
Rank
1
2
3
4
5
6
7
8
9
10
PE1
PE2
ID
24032
11913
11870
17439
6841
20938
10692
19897
9164
17050
p-value
3.0E-06
1.2E-07
2.9E-06
6.9E-06
6.8E-08
2.3E-02
2.0E-01
1.5E-03
1.5E-03
2.2E-02
1.566
1.550
CC-SIS(0.5,0.5)
ID
p-value
12801
1.5E-06
13742
1.5E-05
402
6.2E-05
4862
3.4E-04
8349
3.1E-05
9158
1.9E-03
12074
6.8E-06
14466
1.8E-03
18903
2.5E-02
19774
8.1E-06
1.404
1.290
ID
24032
6841
9164
13025
2172
17439
20121
11870
22705
10408
SIRS
p-value
0.000
0.000
0.001
0.001
0.013
0.000
0.000
0.000
0.028
0.002
1.483
1.398
QPC-SIS(0.5)
ID
p-value
11696
0.001
672
0.005
21944
0.021
6466
0.024
518
0.758
12635
0.022
12567
0.609
7160
0.483
21188
0.007
11916
0.495
1.466
1.437
ID
8349
24032
13025
23670
20121
6841
15674
1644
5644
20238
DC-SIS
p-value
3.1E-05
3.0E-06
6.7E-04
5.9E-03
1.2E-07
6.8E-08
1.2E-06
6.9E-06
2.2E-04
2.3E-02
1.419
1.378
CPC-SISa1 (0.5, 0.5)
ID
p-value
301
0.005
18678
0.000
3524
0.603
5422
0.021
14782
0.023
21431
0.922
5239
0.295
777
0.352
20958
0.132
4849
0.460
1.399
1.345
ID
17679
20238
10408
1644
8339
14028
23670
12305
3929
14466
Kendall-SIS
p-value
6.2E-02
2.3E-02
1.9E-03
6.9E-06
2.6E-03
8.9E-05
5.9E-03
7.3E-04
1.8E-03
1.8E-03
1.399
1.366
CPC-SISa2 (0.5, 0.5)
ID
p-value
20121
0.000
4356
0.001
13084
0.008
13191
0.035
6436
0.299
10179
0.192
20102
0.185
1299
0.179
1830
0.381
6025
0.467
1.436
1.304
ID
24032
22705
6841
14466
4767
5644
20121
23670
13742
17439
QC-SIS(0.5)
p-value
3.0E-06
2.8E-02
6.8E-08
1.8E-03
1.2E-05
2.2E-04
1.2E-07
5.9E-03
1.5E-05
6.9E-06
1.409
1.367
CPC-SISa3 (0.5, 0.5)
ID
p-value
4132
0.136
17568
0.620
5459
0.482
1079
0.352
23942
0.002
14
0.922
1847
0.169
3392
0.505
20369
0.367
390
0.920
1.372
1.289
Table 10: Summary of top 10 gene probes selected by different screening methods for the
breast cancer data. ID means the selected gene ID and p-values are computed as 2(1 −
√ b −1/2
Φ(| nΩ
̺bY,X (0.5, 0.5)|)), where Φ is the cumulative distribution function of standard normal
1
random variable. PE is the mean of prediction errors over 500 random partitions with the partition
ratio of training sample to testing sample being 4 : 1, where prediction error is defined as the
average of {(Yi − Ybi )2 , i ∈ testing set}. PE1 and PE2 indicate that Ybi is the predicted value via
fitting a median regression model and linear model, respectively, using the top 10 genes selected.
heteroscedastic variance and the predictors have low correlation, any marginal screening
methods (SIS, SIRS, DC-SIS) can be applied. If the response contains outliers or follows
a heavy tail distribution and the covariates are normal, robust screening methods (Kendall
SIS, QC-SIS, CQC-SIS, CC-SIS) can be employed. If the covariates are highly correlated and
conditional variables are available, conditional screening procedures (CSIS, QPC-SIS) can
work. If the data is heteroscedastic for both the response and covariates and the covariates
may be highly correlated, then only CPC-SIS can be applied.
The copula formulation may suggest many possible extensions of our methodology. In
particular, we may consider censored survival time outcome in this framework. The estimation of copula-based correlation and partial correlation needs to incorporate the random
censoring for such data and we need to invoke more complicated empirical process theories to argue the weak convergence results. Furthermore, we may even allow the predictors
to be censored. See Cheng and Fine (2008) and Cheng and Li (2015) for some earlier dis35
cussion. Such extension to ultrahigh-dimensional data is non-trivial and requires a further
development.
Acknowledgement We thank Professors Runze Li and Shujie Ma for sharing their R code
for the implementation of QPC approach.
Appendix: Proofs
The following Lemma is a well-known result for the bivariate empirical process theory.
Lemma A.1. (Lemma 3.9.28, van der Vaart and Wellner (1996)) Fix 0 < a < b < 1,
and suppose that H is a distribution function on R2 with marginal distribution functions
F and G that are continuously differentiable on the intervals [F −1 (a) − ε, F −1 (b) + ε] and
[G−1 (a) − ε, G−1 (b) + ε] with positive derivatives f and g, respectively, for some ε > 0.
Furthermore, assume that ∂H/∂x and ∂H/∂y exist and are continuous on the product of
these intervals. Then the map φ : D(R̄) 7→ ℓ∞ ([a, b]2 ) is Hadamard-differentiable at H
tangentially to C(R̄). The derivative is given by
h(F −1 (u), ∞)
∂H −1
(F (u), G−1(v))
∂x
f (F −1 (u))
∂H −1
h(∞, G−1(v)}
−
(F (u), G−1 (v))
.
∂y
g(G−1(v))
φ̇H (h)(u, v) = h(F −1 (u), G−1(v)) −
Proof of Theorem 2.1 By the definition, we have
√
n{b
̺Y,X (τ, ι) − ̺Y,X (τ, ι)} =
√
−1
−1
n[Fn,Y,X (Fn,Y
(τ ), Fn,X
(ι)) − FY,X (FY−1 (τ ), FX−1 (ι))]
p
.
τ (1 − τ )ι(1 − ι)
Employing Lemma A.1 and the empirical process theory (Example 3.9.29 in van der Vaart and Wellner
36
(1996)), we are able to show that
√
n
1 X
a
−1
−1
n[Fn,Y,X (Fn,Y
[ϕ(Xi , Yi ; τ, ι) − E{ϕ(Xi , Yi ; τ, ι)}],
(τ ), Fn,X
(ι)) − FY,X (FY−1 (τ ), FX−1 (ι))] = √
n i=1
a
where = means ”asymptotically equivalent” and ϕ(Xi , Yi ; τ, ι) is given by
ϕ(Yi , Xi ; τ, ι) = I{Yi ≤ FY−1 (τ ), Xi ≤ FX−1 (ι)}
I{Yi ≤ FY−1 (τ )}
∂
FY,X (FY−1 (τ ), FX−1 (ι))
∂y
fY (FY−1 (τ ))
∂
I{Xi ≤ FX−1 (ι)}
− FY,X (FY−1 (τ ), FX−1 (ι))
.
∂x
fX (FX−1 (ι))
−
Furthermore, it is easily verified that
∂
FY,X (FY−1 (τ ), FX−1 (ι)) = fY (FY−1 (τ ))FX|Y =F −1 (τ ) (FX−1 (ι)),
Y
∂y
∂
FY,X (FY−1 (τ ), FX−1 (ι)) = fX (FX−1 (ι))FY |X=F −1 (ι) (FY−1 (τ )).
X
∂x
(A.1)
(A.2)
Hence, by (A.1) and (A.2), it follows that
√
n
1 X
n{b
̺Y,X (τ, ι) − ̺Y,X (τ, ι)} = √
[ξ(Yi , Xi ; τ, ι) − E{ξ(Yi , Xi ; τ, ι)}],
n i=1
a
where ξ(Y, X; τ, ι) is defined in the line before Theorem 2.1.
Because the class of indicator functions is Donsker, thus {ξ(Yi , Xi ; τ, ι), (τ, ι) ∈ [a, b]2 } is
Donsker. By Donsker theorem,
√
in ℓ∞ ([a, b]2 ), where
w
n{b
̺Y,X (τ, ι) − ̺Y,X (τ, ι)}
w
GY,X (τ, ι)
denotes ”converge weakly”, and the process GY,X (τ, ι) has mean zero
and covariance function cov(GY,X (τ1 , ι1 ), GY,X (τ2 , ι2 )) = Ω1 (τ1 , ι1 ; τ2 , ι2 ) given in Theorem
37
2.1.
Proof of Theorem 2.2 Following the arguments in the proof of Theorem 2.1, we can show
√
n{[b
̺Y,X1 (τ, ι) − ̺Y,X1 (τ, ι)] − [b
̺Y,X2 (τ, ι) − ̺Y,X2 (τ, ι)]}
n
1 X
a
=√
[η(Yi , Xi1 , Xi2 ; τ, ι) − E{η(Yi , Xi1 , Xi2 ; τ, ι)}]
n i=1
where η(Yi , Xi1 , Xi2 ; τ, ι) = ξ(Yi , Xi1 ; τ, ι)−ξ(Yi , Xi2 ; τ, ι) and ξ(Y, X; τ, ι) is given in Theorem
2.1. Since {ξ(Xi , Yi ; τ, ι), (τ, ι) ∈ [a, b]2 } is Donsker, {η(Yi , Xi1 , Xi2 ; τ, ι), (τ, ι) ∈ [a, b]2 } is also
Donsker. By Donsker Theorem, we have
√
n{[b
̺Y,X1 (τ, ι) − ̺bY,X2 (τ, ι)] − [̺Y,X1 (τ, ι) − ̺Y,X2 (τ, ι)]}
w
GY,X1 ,X2 (τ, ι)
in ℓ∞ ([a, b]2 ), where the process GY,X1 ,X2 (τ, ι) has mean zero and covariance function
cov(GY,X1 ,X2 (τ1 , ι1 ), GY,X1 ,X2 (τ2 , ι2 )) = Ξ(τ1 , ι1 ; τ2 , ι2 ) given in Theorem 2.2.
Proof of Theorem 2.3 (i) We first prove the first statement. Observe that
−1
−1
Fn,Y,Xj (Fn,Y
(τ ), Fn,X
(ι)) − FY,Xj (FY−1 (τ ), FX−1j (ι)) = IInj1 + IInj2
j
(A.3)
P
−1
−1
−1
where IInj1 = Fn,Y,Xj (Fn,Y
(τ ), Fn,X
(ι))−Fn,Y,Xj (FY−1 (τ ), FX−1j (ι)) = n−1 ni=1 I(Yi ≤ Fn,Y
(τ ), Xij ≤
j
−1
Fn,X
(ι))−I(Yi ≤ FY−1 (τ ), Xij ≤ FX−1j (ι)) and IInj2 = Fn,Y,Xj (FY−1 (τ ), FX−1j (ι))−FY,Xj (FY−1 (τ ), FX−1j (ι)) =
j
P
n−1 ni=1 I(Yi ≤ FY−1 (τ ), Xij ≤ FX−1j (ι)) − FY,Xj (FY−1 (τ ), FX−1j (ι)) . In what follows, we are
going to bound the tail probabilities for IInj1 and IInj2 , respectively.
Since |I(Yi ≤ FY−1 (τ ), Xij ≤ FX−1j (ι)) − FY,Xj (FY−1 (τ ), FX−1j (ι))| ≤ 2 and var(I(Yi ≤
FY−1 (τ ), Xij ≤ FX−1j (ι))) ≤ 1, applying Bernstein’s inequality (Lemma 2.2.9, van der Vaart and Wellner
(1996)) gives that for any δ > 0,
max P (|IInj2| > δ/n) ≤ 2 exp(−
1≤j≤pn
38
δ2
).
2(n + 2δ/3)
(A.4)
Next consider IInj1 . Note that
IInj1 = n−1
n
X
i=1
−1
+n
=
ˆ
(1)
IInj1
+
−1
−1
−1
I(Yi ≤ Fn,Y
(τ )) I(Xij ≤ Fn,X
(ι))
−
I(X
≤
F
(ι))
ij
X
j
j
n
X
−1
I(Yi ≤ Fn,Y
(τ )) − I(Yi ≤ FY−1 (τ )) I(Xij ≤ FX−1j (ι))
i=1
(2)
IInj1
(say).
(A.5)
−1
(ι) −
On the other hand, it follows from Lemma B of Serfling (1980) (p.96) that |Fn,X
j
FX−1j (ι)| ≤ 2n−1/2 (log n)1/2 /fXj (FX−1j (ι)) almost surely. Let M1−1 = inf 1≤j≤pn fXj (FX−1j (ι))/2,
−1
which is positive under condition (C1). On the event {|Fn,X
(ι) − FX−1j (ι)| ≤ M1 n−ς1 }
j
P
(1)
−1
with any ς2 < 1/2. we have |IInj1| ≤ n−1 ni=1 |I(Xij ≤ Fn,X
(ι)) − I(Xij ≤ FX−1j (ι))| ≤
j
Pn
−1
−1
−1
1
−ς1
≤ Xij ≤ FX−1j (ι) +
i=1 qι (Xij , FXj (ι)), where qι (Xij , FXj (ι)) = I(FXj (ι) − M1 n
n
R FX−1j (ι)+M1 n−ς1
−1
−ς1
M1 n ). Because E{qι (Xij , FXj (ι))} = F −1 (ι)−M n−ς1 fXj (s)ds ≤ c1 M1 n−ς1 for some pos1
Xj
itive constant c1 by condition (C1). Since qι (Xij , FX−1j (ι)) ≤ 1 and var(qι (Xij , FX−1j (ι))) ≤
E{qι (Xij , FX−1j (ι))}, so applying Bernstein’s inequality (Lemma 2.2.9, van der Vaart and Wellner
(1996)) yields that
n
1X
δ2
δ
−1
−1
[qι (Xij , FXj (ι)) − E{qι (Xij , FXj (ι))}] ≥
≤ 2 exp(−
).
P
n i=1
2n
8(c1 M1 n1−ς1 + δ/3)
This further implies that
(1)
n
1X
qι (Xij , FX−1j (ι)) ≥ δ/n
n i=1
n
1X
[qι (Xij , FX−1j (ι)) − E{qι (Xij , FX−1j (ι))}] ≥ δ/n − c1 M1 n−ς1
≤ P
n i=1
n
1X
−1
−1
≤ P
[qι (Xij , FXj (ι)) − E{qι (Xij , FXj (ι))}] ≥ δ/(2n)
n i=1
max P (|IInj1| ≥ δ/n) ≤ P
1≤j≤pn
≤ 2 exp(−
δ2
8(c1 M1 n1−ς1 + δ/3)
39
),
(A.6)
where the third inequality is due to δ/(2n) ≥ c1 M1 n−ς1 . Similarly, by condition (C1), we
can show that
δ2
(2)
max P (|IInj1| ≥ δ/n) ≤ 2 exp(−
1≤j≤pn
8(c2 M2 n1−ς2 + δ/3)
),
(A.7)
for any ς2 ∈ (0, 1/2) and for some positive constants c2 and M2 . Therefore, combining
(A.4)-(A.7), we obtain that
max P (|IInj1 + IInj2| ≥ 3δ/n) ≤
1≤j≤pn
max P (|IInj2| > δ/n)
1≤j≤pn
(2)
(1)
+ max P (|IInj1| ≥ δ/n) + max P (|IInj1| ≥ δ/n)
1≤j≤pn
1≤j≤pn
δ2
δ
) + 2 exp(−
)
2(n + 2δ/3)
8(c1 M1 n1−ς1 + δ/3)
δ2
)
+2 exp(−
8(c2 M2 n1−ς2 + δ/3)
δ2
≤ 6 exp(−
)
(A.8)
c3 n + c4 δ
2
≤ 2 exp(−
for some positive constants c3 ≥ max(2, 8c1M1 , 8c2 M2 ) and c4 ≥ 8/3, provided that δ/(2n) ≥
max(c1 M1 n−ς1 , c2 M2 n−ς2 ).
Using the basic inequality that ||x| − |y|| ≤ |x − y| and the union bound of probability,
it follows that
uj − uj | ≥ Cn−κ )
uj − uj | ≥ Cn−κ ) ≤ pn max P (|b
P ( max |b
1≤j≤pn
1≤j≤pn
p
≤ pn max P (|IInj1 + IInj2 | ≥ τ (1 − τ )ι(1 − ι)Cn−κ ).
1≤j≤pn
Hence, this in conjunction with the result (A.8) proves the desired result by letting δ =
p
τ (1 − τ )ι(1 − ι)Cn1−κ /3.
40
(ii) Prove the second assertion. By the choice of νn and using condition (C2), we have
da ≥ P min u
P M∗ ⊂ M
uj − uj ) > νn − min uj
bj > νn ≥ P min (b
j∈M∗
j∈M∗
j∈M∗
bj − uj ≥ νn
bj − uj > νn ≥ 1 − P max u
≥ P min uj − max u
j∈M∗
j∈M∗
j∈M∗
≥ 1 − 6sn exp(−c̃1 n1−2κ ).
Thus, this completes the proof.
Proof of Proposition 2.4 Observe that the expected false discovery rate can be rewritten
as
n |M
X
ca,δ ∩ (M∗a )c | o
√
1
b 1 (τ, ι; τ, ι)]−1/2 |b
E
=
P
n[
Ω
̺
(τ,
ι)|
≥
δ
.
Y,X
j
|(M∗a )c |
pn − |M∗a |
∗ c
j∈(Ma )
By Theorem 2.1 and the Berry-Esséen result, it follows that there exists some constant ca > 0
such that
sup P
z
√
b 1 (τ, ι; τ, ι)]−1/2 ̺bY,X (τ, ι) ≤ z − Φ(z) ≤ ca n−1/2 .
n[Ω
j
Combining the above results, we obtain
n |M
X
ca,δ ∩ (M∗ )c | o
√
1
a
E
≤
(2(1 − Φ(δ)) + ca / n).
∗
c
∗
|(Ma ) |
pn − |Ma |
∗ c
j∈(Ma )
Plugging δ = Φ−1 (1 − d¯n /(2pn )) into the above inequality yields the result.
Proof of Theorem 3.1 We first observe that E{ψτ (Y − ZT α0 )Z} = 0 and E{ψτ (X −
ZT θ 0 )Z} = 0, where ZT α0 = ZT α0 and ZT θ 0 = ZT θ 0 . Using the arguments in Li et al.
41
(2015) and Koenker (2005), we can obtain
√
b − α0 ) = E[fYi |Zi (ZTi α0 )Zi ZTi ]
n(α
n
−1
1 X
√
ψτ (Yi − ZTi α0 )Zi + op (1),
n i=1
−1
1 X
√
ψι (Xi − ZTi θ 0 )Zi + op (1).
n i=1
(A.9)
and
√
b − θ 0 ) = E[fX |Z (ZT θ 0 )Zi ZT ]
n(θ
i
i
i
i
n
(A.10)
Let
n
X
b − E{FY,X|Z (ZT α0 , ZT θ 0 )}].
b = 1
b Xi ≤ ZTi θ)
b θ)
[I(Yi ≤ ZTi α,
Un (α,
n i=1
Then, it follows that
√
√
b
b θ)
nUn (α,
,
n{b
̺Y,X|Z (τ, ι) − ̺Y,X|Z (τ, ι)} = p
τ (1 − τ )ι(1 − ι)
where
√
n
X
b − E{FY,X|Z (ZT α,
b
b = √1
b Xi ≤ ZTi θ)
b ZT θ)}
b θ)
I(Yi ≤ ZTi α,
nUn (α,
n i=1
b − E{FY,X|Z (ZT α0 , ZT θ 0 )}
b ZT θ)}
+ E{FY,X|Z (ZT α,
b + Kn2 (α,
b (say).
b θ)
b θ)
=
ˆ Kn1 (α,
(A.11)
Employing the result of empirical processes (Corollary 2.3.12, van der Vaart and Wellner
(1996), p115), we can obtain
n
X
b = √1
b θ)
Kn1 (α,
I(Yi ≤ ZTi α0 , Xi ≤ ZTi θ 0 )
n i=1
−E{FY,X|Z (ZT α0 , ZT θ 0 )} + op (1).
42
(A.12)
By Taylor expansion and using (A.9) and (A.10), we have
b = E FX|Z,Y =ZT α0 (ZT θ 0 )fY |Z (ZT α0 )Z T (α
b θ)
b − α0 )
Kn2 (α,
T b
+E FY |Z,X=ZT θ0 (ZT α0 )fX|Z (ZT θ 0 )Z (θ
− θ 0 ) + op (1)
n
1 X n T −1
= √
∆12 ∆11 ψτ (Yi − ZTi α0 )Zi
n i=1
o
T 0
+∆T21 ∆−1
ψ
(X
−
Z
θ
)Z
i
i + op (1).
22 ι
i
(A.13)
Finally, combining (A.11)-(A.13) and recalling the definitions of ̺bY,X|Z (τ, ι) and ̺Y,X|Z (τ, ι),
we have
p
√
τ (1 − τ )ι(1 − ι) n{b
̺Y,X|Z (τ, ι) − ̺Y,X|Z (τ, ι)}
n
1 X n
I(Yi ≤ ZTi α0 , Xi ≤ ZTi θ 0 ) − E{FY,X|Z (ZT α0 , ZT θ 0 )}
=√
n i=1
o
T 0
T
−1
T 0
+∆T12 ∆−1
ψ
(Y
−
Z
α
)Z
+
∆
∆
ψ
(X
−
Z
θ
)Z
i
i
i
i + op (1), (A.14)
11 τ
i
21 22 ι
i
which implies that
√
a
n{b
̺Y,X|Z (τ, ι)−̺Y,X|Z (τ, ι)} =
√1
n
Pn
i=1 [ζ(Xi , Yi , Zi ; τ, ι)−Eζ(X, Y, Z; τ, ι)].
It can be checked that {ζ(Xi , Yi , Zi ; τ, ι) : (τ, ι) ∈ [a, b]2 } is Donsker. Therefore, by Donsker
Theorem, we can conclude the result.
Proof of Theorem 3.2 This result can be proved using the same arguments in the proof
of Theorem 3.1. We omit the details.
To prove Theorem 3.3, we need the following two lemmas which can be verified using the
same arguments in Ma et al. (2017) and the proofs are omitted.
Lemma A.2. Under conditions (D1)(i) and (D2), for any given constant c∗1 > 0 there exist
some constant c∗2 such that
b − α0 k > c∗1 n−κ ) ≤ 3 exp(−c∗2 qn−1 n1−2κ ).
P (kα
43
Lemma A.3. Under conditions (D1)(ii) and (D2), for every 1 ≤ j ≤ pn and for any given
constant c∗3 > 0 there exist some constant c∗4 such that
bj − θ 0 k > c∗ n−κ ) ≤ 3 exp(−c∗ q −1 n1−2κ ).
P (kθ
j
3
4 n
Proof of Theorem 3.3 To prove the first part of the result, we need to only prove the
following fact. Under conditions (D1) and (D2), for every 1 ≤ j ≤ pn and for any given
constant c∗5 > 0, there exist some positive constants c∗6 such that
n
1X
bj ) − E{ψτ (Yi − ZT α0 )ψι (Xij − ZT θ 0 )} ≥ c∗ n−κ
b ι (Xij − ZTi θ
P
ψτ (Yi − ZTi α)ψ
5
i
i j
n i=1
≤ 12 exp(−c∗6 qn−1 n1−2κ ).
(A.15)
To this end, we first make a decompose as
n
1 X
bj ) − E{ψτ (Yi − ZT α0 )ψι (Xij − ZT θ 0 )}
b ι (Xij − ZTi θ
ψτ (Yi − ZTi α)ψ
i
i j
n i=1
=N
ˆ nj1 + Nnj2 + Nnj3 ,
where
n
Nnj1
1 X
=
ψτ (Yi − ZTi α0 )ψι (Xij − ZTi θ 0j ) − E{ψτ (Yi − ZTi α0 )ψι (Xij − ZTi θ 0j )} ,
n i=1
n
Nnj2 =
1 X
bj ),
b − ψτ (Yi − ZTi α0 ) ψι (Xij − ZTi θ
ψτ (Yi − ZTi α)
n i=1
n
Nnj3
1 X
bj ) − ψι (Xij − ZT θ 0 ) ψτ (Yi − ZT α0 ).
ψι (Xij − ZTi θ
=
i
i j
n i=1
Next, we are going to find the probability bounds for Nnj1 , Nnj2 and Nnj3. For Nnj1, let
∆nij = ψτ (Yi − ZTi α0 )ψι (Xij − ZTi θ 0j ). Since |∆nij | ≤ 1 and thence var(∆nij ) ≤ 1, we can
44
obtain that for any δ > 0,
P (|Nnj1| ≥ δ/n) ≤ 2 exp(−
δ2
),
2(n + δ/3)
(A.16)
by applying Bernstein’s inequality (Lemma 2.2.9, van der Vaart and Wellner (1996)). For
P
b − ψτ (Yi − ZTi α0 ) . It follows that for any
Nnj2 , observe that |Nnj2| ≤ n1 ni=1 ψτ (Yi − ZTi α)
δ > 0,
P (|Nnj2| ≥ δ/n) ≤ P
n
n1 X
n
i=1
b − α0 k ≤ c∗1 n−κ
b − ψτ (Yi − ZTi α0 ) ≥ δ/n, kα
ψτ (Yi − ZTi α)
b − α0 k > c∗1 n−κ ).
+P (kα
o
(A.17)
Let Πni = supkuk≤1 ψτ (Yi − ZTi (α0 + c∗1 n−κ u)) − ψτ (Yi − ZTi α0 ) . Under condition (D1)(i)
and (D2)(ii), there exist a finite positive constant c5 and a u∗ with ku∗ k ≤ 1 such that
R ZT α0 +c∗ n−κ ZTi u∗
1/2
E(Πni ) ≤ E{| ZTi α0 1
fY |Z=z (y)dy|} ≤ c∗1 c5 n−κ E{|ZTi u∗ |} ≤ c∗1 c5 cmax n−κ = c6 n−κ ,
i
where c6 =
1/2
c∗1 c5 cmax
the last inequality uses Cauchy-Schwarz inequality. Analogously, we can
obtain E(Π2ni ) ≤ c7 n−κ for some positive constant c7 . Notice that |Πni | ≤ 1. Hence, applying
Bernstein’s inequality (Lemma 2.2.9, van der Vaart and Wellner (1996)) yields that for any
ǫ > 0,
n
1X
n2 ǫ2
P
Πni − E(Πni ) ≥ ǫ ≤ 2 exp(−
).
n i=1
2(c7 n1−κ + nǫ/3)
Letting δ ≥ 2c6 n1−κ , that is δ/n − c6 n−κ ≥ δ/(2n), and using the above inequality, we can
45
obtain
P
n
1 X
n
i=1
n
1 X
Πni ≥ δ/n ≤ P
Πni − E(Πni ) ≥ δ/n − c6 n−κ
n i=1
n
1X
Πni − E(Πni ) ≥ δ/(2n)
≤ P
n i=1
≤ 2 exp(−
δ2
8(c7 n1−κ + δ/6)
).
(A.18)
Because the first probability on the right hand side of (A.17) is bounded by (A.18). So, by
Lemma A.2 and the results (A.17) and (A.18), we have
P (|Nnj2| ≥ δ/n) ≤ 2 exp(−
δ2
8(c7 n1−κ + δ/6)
) + 3 exp(−c∗2 qn−1 n1−2κ )
(A.19)
provided that δ ≥ 2c6 n1−κ . Similarly, using Lemma A.3, and under conditions (D1)(ii) and
D2(ii), we have
P (|Nnj3| ≥ δ/n) ≤ 2 exp(−
δ2
8(c8 n1−κ + δ/6)
) + 3 exp(−c∗4 qn−1 n1−2κ )
(A.20)
for some positive constant c8 , provided that δ ≥ 2c9 n1−κ for some constant c9 > 0. Let
c10 = max(c7 , c8 ) and c11 = max(c∗2 , c∗4 ). Combining the results (A.16), (A.19) and (A.20)
gives
max P (|Nnj1 + Nnj2 + Nnj3 | ≥ 3δ/n)
1≤j≤pn
δ2
δ2
) + 4 exp(−
) + 6 exp(−c11 qn−1 n1−2κ ), (A.21)
≤ 2 exp(−
1−κ
2(n + δ/3)
8(c10 n
+ δ/6)
provided that δ ≥ 2 max(c6 , c9 )n1−κ . The result in (A.15) follows by letting δ = 3 max(c6 , c9 )n1−κ
in (A.21). In addition, applying (A.15) and using similar arguments in the proof of Theorem
2.3, we can conclude the first part of Theorem 3.3.
46
For the second part of the result, we can verify it by following the proof of Theorem
2.3.
Proof of Proposition 3.4 The proof directly follows the proof of Proposition 2.4. The
details are omitted.
References
Bahadur R.R. (1966). A note on quantiles in large samples. Annals of Mathematical Statistics,
37, 577-580.
Barut E., Fan J. and Verhasselt A. (2016). Conditional sure independence screening. Journal
of American Statistical Association, 111, 1266-1277.
Boulesteix A.L., Porzelius C. and Daumer M. (2008). Microarray-based classification and
clinical predictors: on combined classifiers and additional predictive value. Bioinformatics,
24, 1698-1706.
Chang J., Tang C.Y. and Wu Y. (2013). Marginal empirical likelihood and sure independence
feature screening. Annals of Statistics, 41, 2123-2148.
Chang J., Tang C.Y. and Wu Y. (2016). Local independence feature screening for nonparametric and semiparametric models by marginal empirical likelihood. Annals of Statistics,
44, 515-539.
Cheng Y. and Fine J.P. (2008). Nonparametric estimation of cause-specific cross hazard ratio
with bivariate competing risks data. Biometrika, 95, 233-240.
Cheng Y. and Li J. (2015). Time-dependent diagnostic accuracy analysis with censored
outcome and censored predictor. Journal of Statistical Planning and Inference, 156, 90102.
47
Cheng M., Honda T., Li J. and Peng H. (2014) Nonparametric independence screening and
structure identification for ultra-high dimensional longitudinal data. Annals of Statistics,
42, 1819-1849.
Chiang A.P., Beck J.S., Yan H.J., et al. (2006). Homozygosity mapping with SNP arrays
identifies TRIM32, an E3 ubiquitin ligase, as a BardetBiedl syndrome gene (BBS11).
Proceedings of the National Academy of Sciences, 103, 6287-6292.
Chu W., Li R. and Reimherr M. (2016). Feature screening for time-varying coefficient models
with ultrahigh dimensional longitudinal data. Annals of Applied Statistics, 10, 596-617.
DeSantis C.E., Ma J., Sauer A.G., Newman L. A. and Jemal A. (2017). Breast cancer
statistics, 2017, racial disparity in mortality by state. CA: A Cancer Journal for Clinicians,
67, 439-448. Harvard
Fan J. and Lv J. (2008). Sure independence screening for ultrahigh dimensional feature space.
Journal of the Royal Statistical Society, Series B, 70, 849-911.
Fan J., Feng Y. and Song R. (2011). Nonparametric independence screening in sparse ultrahigh-dimensional additive models. Journal of the American Statistical Association, 106,
544-557.
Fan J., Ma Y., Dai W. (2014) Nonparametric independent screening in sparse ultra-high
dimensional varying coefficient models. Journal of the American Statistical Association,
109, 1270-1284.
Fan J. and Song R. (2010). Sure independence screening in generalized linear models with
NP-dimensionality. Annals of Statistics, 38, 3567-3604.
He X., Wang, L. and Hong H. (2013). Quantile-adaptive model-free variable screening for
high-dimensional heterogeneous data. Annals of Statistics, 41, 342-369.
48
Iglewicz B. and Hoaglin D. (1993), Volume 16: How to detect and handle outliers, The
ASQC Basic References in Quality Control: Statistical Techniques, Edward F. Mykytka,
Ph.D., Editor.
Koenker R. (2005). Quantile Regression, New York: Cambridge University Press.
Li G., Li Y. and Tsai C.L. (2015). Quantile correlations and quantile autoregressive modeling.
Journal of the American Statistical Association, 110, 246-261.
Li G., Peng H., Zhang J. and Zhu L. (2012). Robust rank correlation based screening. Annals
of Statistics, 40, 1846-1877.
Li J., Zheng Q., Peng L. and Huang Z. (2016). Survival impact index and ultrahighdimensional model-free screening with survival outcomes. Biometrics, 72, 1145-1154.
Li R., Zhong W. and Zhu L. (2012). Feature screening via distance correlation learning.
Journal of the American Statistical Association, 107, 1129-1139.
Liu J., Li R. and Wu R. (2014). Feature selection for varying coefficient models with
ultrahigh-dimensional covariates. Journal of the American Statistical Association, 109,
266-274.
Ma X. and Zhang J. (2016). Robust model-free feature screening via quantile correlation.
Journal of Multivariate Analysis, 143, 472-480.
Ma S., Li R. and Tsai C. (2017). Variable screening via quantile partial correlation. Journal
of the American Statistical Association, 112, 650-663.
Nelsen R. B. (2007). An Introduction to Copulas. Springer Science & Business Media.
Serfling R.J. (1980). Approximation Theorems of Mathematical Statistics. Wiley, New York.
49
Scheetz T. E., Kim K.Y. A., Swiderski R. E., et al.(2006). Regulation of gene expression in
the mammalian eye and its relevance to eye disease. Proceedings of the National Academy
of Sciences of the United States of America, 103, 14429-14434.
Song R., Lu W., Ma S. and Jeng X.J. (2014). Censored rank independence screening for
high-dimensional survival data. Biometrika, 101, 799-814.
van der Vaart A. W. and Wellner J. A. (1996). Weak Convergence and Empirical Processes
with Application to Statistics. Springer, New York.
Vant Veer L.J., Dai H., van de Vijver M.J., et al. (2002). Gene expression profiling predicts
clinical outcome of breast cancer. Nature, 415, 530-536.
Wu Y. and Yin G. (2015). Conditional qunatile screening in ultrahigh-dimensional heterogeneous data. Biometrika, 102, 65-76.
Xia X., Li J. and Fu B. (2018). Conditional quantile correlation learning for ultrahigh dimensional varying coefficient models and its application in survival analysis. Statistica Sinica,
Accepted.
Yu T., Li J. and Ma S. (2012). Adjusting confounders in ranking biomarkers: a model-based
ROC approach. Briefings in Bioinformatics, 13, 513-523.
Zhao D.S. and Li Y. (2012). Principled sure independence screening for Cox models with
ultra-high-dimensional covariates. Journal of Multivariate Analysis, 105, 397-411.
Zhu L., Li L., Li R. and Zhu L.X. (2011). Model-Free feature screening for ultrahigh dimensional data. Journal of the American Statistical Association, 106, 1464-1475.
50
| 10math.ST
|
arXiv:1509.01847v5 [math.GR] 27 Feb 2018
THE BASS–JIANG GROUP FOR
AUTOMORPHISM-INDUCED HNN-EXTENSIONS
ALAN D. LOGAN
Abstract. We use the Bass–Jiang group for automorphism-induced
HNN-extensions to build a framework for the construction of tractable
groups with pathological outer automorphism groups. We apply
this framework to a strong form of a question of Bumagin–Wise
on the outer automorphism groups of finitely presented, residually
finite groups.
1. Introduction
Matumoto proved that every group Q can be realised as the outer
automorphism group of some group GQ [Mat89]. Many authors have refined this result by placing restrictions on the groups Q and GQ [Koj88]
[GP00] [DGG01] [BG03] [BW05] [FM06] [Min09] [Log15] [Log16] [Log17a].
Bumagin–Wise asked the following question [BW05, Problem 1]:
Question 1 (Bumagin–Wise). Can every countable group Q be realised
as the outer automorphism group of a finitely generated, residually finite group GQ ?
Our main result is Theorem 1.1, which relates to a strong form of
Question 1 where the group GQ is taken to be finitely presented rather
than finitely generated. Almost nothing is known in general about
the outer automorphism groups of finitely presented, residually finite
groups. A group H has Serre’s property FA if every action of H on
any tree has a global fixed point; this is related to the property of not
splitting non-trivially as a free product with amalgamation or HNNextension [Ser80, Theorem 15].
Theorem 1.1. Suppose Q is a finitely presented group with Serre’s
property FA. Then there exists a finitely presented group GQ such that
Q embeds with finite index into Out(GQ ). Moreover, if Q is residually
finite then GQ can be chosen to be residually finite.
2010 Mathematics Subject Classification. 20E06, 20E26, 20E36, 20F28.
Key words and phrases. HNN-extensions, outer automorphism groups, residually
finite groups.
1
2
ALAN D. LOGAN
Examples of finitely presented groups Q with Serre’s property FA
are SL3 (Z) [Ser74], triangle groups [Ser80], finitely presented groups
with Kazhdan’s property T [Wat82], random groups in the sense of
Gromov [DGP11], R. Thompson’s groups T and V [Far11], and the
Brin–Thompson groups nV [Kat15].
The proof of Theorem 1.1 has two main ingredients. Firstly, Theorem
1.1 applies a certain framework built in this paper and explained below.
Secondly, Theorem 1.1 applies a version of Rips’ construction [BO08].
Rips’ construction is usually applied to obtain finitely generated, nonfinitely presentable groups with pathological properties; our application
to finitely presented groups is unusual.
The framework. This paper principally consists of an in-depth analysis of the Bass–Jiang group for automorphism-induced HNN-extensions.
The Bass–Jiang group of an HNN-extension G, defined in Section 3, is
a subgroup of Out(G) which arises naturally in the context of Bass–
Serre theory. Automorphism-induced HNN-extensions, defined in Section 2, are a class of tractable HNN-extensions; by “tractable” we mean
that they are an easy class of groups to work with and possess nicer
properties than general HNN-extensions. In Section 7 we package the
technical results of this analysis into a framework for the construction
of tractable groups with pathological outer automorphism groups. This
framework is applied to prove Theorem 1.1 (see also [Log16] [Log17a]).
Describing Out(G). Theorem 1.2 is a striking, non-technical special case of the framework described in Section 7. This theorem gives
a convenient description of Out(G) for certain residually finite HNNextensions G. As our motivation comes from Question 1, the assumption that G is residually finite is completely natural. For functions ρ
and σ write ρσ(x) := σ(ρ(x)), and write γg (h) := g −1 hg for all h ∈ G.
For φ ∈ Aut(H) define Fix(φ) := {h ∈ H : φ(h) = h}. Define AK as:
AK := {δ ∈ Aut(H) : δ(K) = K, ∃a ∈ H s.t. δφ(k) = φδγa (k)∀k ∈ K}
and this is a subgroup of Aut(H).
H and
Theorem 1.2. Let G ∼
= hH, t; tkt−1 = φ(k), k ∈ Ki, K
φ ∈ Aut(H). Assume that Z(H) ≤ Fix(φ), that H is finitely generated,
and that G is residually finite. Then we have a short exact sequence:
1 → CH (K) ⋊
NH (K)
AK Inn(H)
→ Out0 (G) →
→1
Z(H)K
Inn(H)
where Out0 (G) is an index-one or -two subgroup of Out(G).
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
3
Theorem 1.2 follows immediately from Theorems 4.7 and 6.2. Indeed,
Theorem 6.2 classifies when Out0 (G) = Out(G); these conditions are
applicable here but are omitted for brevity. Note that the proof of
Theorem 1.1 requires a more general result than Theorem 1.2, as the
groups G = GQ in Theorem 1.1 are not necessarily residually finite.
A note on labels of results. In a previous paper [Log16] we cited
certain results from this current paper. The labels of the results in
this current paper have subsequently changed. We record the relevant
changes here: Theorem A is now Theorem 4.7, Lemma 2.1 is now
Lemma 4.1, Lemma 5.2 is now Lemma 6.3, and Proposition 5.3 is now
Proposition 2.1.
Outline of the paper. In Section 2 we define automorphism-induced
HNN-extensions and explain their “tractability”. In Section 3 we define the Bass–Jiang group and the Pettet group for HNN-extensions.
In Section 4 we prove conditions implying that the Bass–Jiang group
is the whole outer automorphism group; our main result is Theorem
4.7. In Section 5 we give a description of the Bass–Jiang group for
automorphism-induced HNN-extensions; our main result is Theorem
5.6. In Section 6 we prove results relating the structure of the Bass–
Jiang group of an automorphism-induced HNN-extension G to commutativity in the base group of G; our main result is Theorem 6.2. In
Section 7 we package the technical results of this paper into a framework for the construction of tractable groups with pathological outer
automorphism groups, and we prove Theorem 1.1.
Acknowledgements. The author would like to thank Stephen J.
Pride and Tara Brendle for many helpful discussions about this paper.
2. Automorphism-induced HNN-extensions
Let H be a group and let η : K → K ′ be an isomorphism of
subgroups of H. An HNN-extension of H over η is a group with
relative presentation G = hH, t; tkt−1 = η(k), k ∈ Ki. We write
G = H∗(K,K ′,η) . An automorphism-induced HNN-extension is an HNNextension with relative presentation
G = hH, t; tkt−1 = φ(k), k ∈ Ki
4
ALAN D. LOGAN
where φ ∈ Aut(H) and K
H;1 the isomorphism of associated subgroups is induced by the automorphism φ. We write G = H∗(K,φ).
Throughout this paper we use the letters G, H, K and φ as above.
In a general HNN-extension the form of the embeddings of the two
associated subgroups K and K ′ may be completely different. For example, K may be normal while K ′ is malnormal. In contrast, in an
automorphism-induced HNN-extension the form of the embeddings of
K and K ′ = φ(K) are necessarily the same (both normal, both malnormal, and so on), and in practice the image group φ(K) may be
disregarded. For example, supposing H is finitely generated and residually finite, then G = H∗(K,φ) is residually finite if and only if K is
residually separable in H [Log17b]. In this current paper, Theorems
1.2, 5.6 and 6.2 do not mention the image group φ(K).
The Baumslag–Solitar groups BS(m, n) = ha, t; tam t−1 = an i are
the HNN-extensions of the infinite cyclic group with non-trivial associated subgroups. These provide standard examples of HNN-extensions
with pathological properties. For example, Baumslag–Solitar groups
can be non-Hopfian (they are Hopfian if any only if n = 1, or m = 1,
or gcd(m, n) 6= 1 [CL83]); they can be Hopfian but non-residually finite (they are residually finite if and only if n = 1, or m = 1, or
|m| = |n| [Mes72], and note that a finitely generated, residually finite group is Hopfian); and they can have non-finitely generated outer
automorphism group (for example, Out(BS(2, 4)) is not finitely generated [CL83]). However, automorphism-induced Baumslag–Solitar
groups have |m| = |n|, so are residually finite, by the above classification, and have virtually cyclic outer automorphism group [GHMR00].
The above two paragraphs demonstrate that automorphism-induced
HNN-extensions are more tractable and in certain cases have nicer
properties than general HNN-extensions. Therefore, increasing the
complexity of the base group H maintains the tractability of these
HNN-extensions but can allow for pathological properties; see, for example, [Log16], where H is a cubulated hyperbolic group with Serre’s
property FA, and [Log17a], where H is a triangle group.
Notation. We write hg := g −1 hg and we write γg for the inner automorphism defined by γg (h) = g −1hg for all h ∈ G. These contrast
with the notation tkt−1 = φ(k) used here for HNN-extensions because
it makes certain proofs more readable.
1If H
= K then G = H⋊φ Z is the mapping torus of φ. Here the Bass–Jiang group
(see Section 3) is very different because φ extends to an inner automorphism [Log15].
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
5
The normaliser-quotient NH (K)/K. The following proposition
underlies this whole paper. It implies that NH (K)/K embeds into
Aut(G) for G = H∗(K,φ), and this paper represents an effort to obtain
conditions implying that NH (K)/K is “close to” Out(G) (for example,
the are isomorphic, commensurable, and so on).
Proposition 2.1 (Ateş–Logan–Pride [ALP17]). Let G = H∗(K,φ). For
d ∈ NH (K) define ϕd : h 7→ h, t 7→ φ(d)td−1 . Then the map
τ : NH (K) 7→ Aut(G)
d 7→ ϕd
is a homomorphism with kernel K.
Proof. It is routine to prove that ϕd ∈ Aut(G). Then τ is a homomorphism as ϕd1 ϕd2 = ϕd1 d2 , and it has kernel K as φ(d)td−1 = t if and
only if d−1 ∈ K by Britton’s Lemma [LS77, Section IV.2].
In the framework built by this paper and explained in Section 7, the
pathological properties of Out(G), where G = H∗(K,φ), are inherited
from the normaliser-quotient NH (K)/K. In particular, by Proposition
2.1, if NH (K)/K is not residually finite and H is finitely generated
then G is not residually finite [Bau63] (and this is an “if and only if”
statement if NH (K) has finite index in H [Log17b]). We see later, in
Theorem 6.2, that if Z(H) ≤ K ∩ Fix(φ) then NH (K)/K also embeds
into Out(G), while Lemma 6.3 gives a finite-index version of this embedding when Z(H) 6≤ K. Thus the properties of NH (K)/K are in a
certain sense bestowed upon G.
Note that if H is infinite cyclic (and hence G is a Baumslag–Solitar
group) then NH (K)/K is cyclic, and so bestows no pathological properties upon G.
3. The Bass–Jiang group
This paper analyses the Bass–Jiang group for automorphism induced
HNN-extensions. The proof of Theorem 1.1 applies this analysis. In
this section we define the Bass–Jiang group OutH (G) of an HNNextension G = H∗(K,K ′,η) . We also define the Pettet group OutH (G),
which is applied in one of our main technical results, Theorem 4.7.
The Pettet group lies between the Bass–Jiang group and the full outer
automorphism group:
OutH (G) ≤ OutH (G) ≤ Out(G).
6
ALAN D. LOGAN
Note that these groups may be defined for graphs of groups in general
[BJ96] [Pet99], but for brevity we only define them for HNN-extensions.
Bass–Serre theory. The definitions in this section and certain proofs
in Section 4 apply Bass–Serre theory. All relevant definitions and notation are taken from Serre’s book [Ser80]. If G = H∗(K,K ′,η) then G
acts in a standard way on a tree Γ, called the Bass–Serre tree of G.
Vertex-stabilisers Gv of Γ are precisely the conjugates of the base group
H, so Gv = H g for some g ∈ G, while edge-stabilisers Ge are precisely
the conjugates of the associated group K, so Ge = K g for some g ∈ G.
For vertices v, w ∈ V Γ we write [v, w] for the geodesic connecting v
and w, and we define the length of [v, w], denoted |[v, w]|, to be the
number of (positive) edges in [v, w].
The Bass–Jiang group. If ψ ∈ Aut(G) then ψb denotes the coset
of Out(G) containing ψ. The Bass–Jiang group OutH (G) of G =
H∗(K,K ′,η) is the subgroup of Out(G) consisting of those ψb ∈ Out(G)
which have a representative ψ such that ψ(H) = H and ψ(t) = gtǫ
for some g ∈ H, ǫ = ±1. It is helpful to have a geometric view of
the Bass–Jiang group: If AutH (G) is the full pre-image of OutH (G)
in Aut(G) then AutH (G) acts on the Bass–Serre Γ of G such that the
diagram in Figure 1 commutes.
The Bass–Jiang group OutΓ (GΓ ) for the fundamental group GΓ of
a graph of groups Γ was described by Bass–Jiang [BJ96]. Levitt described a group related to OutΓ (GΓ ) in his investigation of Out(G) for
G a one-ended hyperbolic group [Lev05]. Gilbert–Howie–Metaftsis–
Raptis gave conditions implying OutΓ (GΓ ) = Out(GΓ ) for GΓ a generalised Baumslag–Solitar group [GHMR00]. Theorem 4.7 mirrors this
result of Gilbert–Howie–Metaftsis–Raptis.
The Pettet group. The Pettet group OutH (G) of G = H∗(K,K ′,η) is
the subgroup of Out(G) consisting of those ψb ∈ Out(G) which have a
representative ψ such that ψ(H) = H. It is helpful to have a geometric
view of the Pettet group: If AutH (G) is the full pre-image of OutH (G)
in Aut(G) then AutH (G) acts on the Bass–Serre tree Γ of G such that
the diagram in Figure 2 commutes.
M. Pettet studied AutH (G), and implicitly the Pettet group [Pet99].
He focused on conditions for the Pettet group to equal the Bass–Jiang
group; these are related to Conditions (1) and (2) from Theorem 4.7.
Two examples. We now give two examples which demonstrate that,
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
G
θ
AutH (G)
G
θ
7
AutH (G)
Aut(V Γ)
Aut(Γ)
Figure 1. This diagram
commutes, where AutH (G)
is the full pre-image of
the
Bass–Jiang
group,
G = H∗(K,K ′ ,η) , Γ is the
Bass–Serre tree, and θ is the
canonical homomorphism
with θ(G) = Inn(G).
Figure 2. This diagram
commutes, where AutH (G)
is the full pre-image of
the Pettet group, G =
H∗(K,K ′ ,η) , V Γ is the vertices of the Bass–Serre tree,
and θ is the canonical homomorphism with θ(G) =
Inn(G).
in general, OutH (G)
OutH (G) and OutH (G)
Out(G). Our first
example is due to Collins–Levin [CL83] and Levitt [Lev07]. Consider
the HNN-extension G1 = ha, t; t−1 a2 t = a4 i, with base group H1 =
hai. The map ψ1 : a 7→ a, t 7→ t−1 a2 ta−2 t is an automorphism of
G1 , and clearly ψb1 ∈ OutH1 (G1 ). However, ψb1 6∈ OutH1 (G1 ) and so
OutH1 (G1 ) OutH1 (G1 ).
We shall write F (x, y) for the free group with free basis {x, y}. Consider the automorphism-induced HNN-extension G2 = ha, b, t; at = bi,
with base group H2 = F (a, b). Clearly G2 = F (a, t), and hence the
map ψ2 : a 7→ at, t 7→ at2 is an automorphism of G2 . However,
ψb2 6∈ OutH2 (G2 ) as at is not conjugate to any power of a in F (a, t), so
OutH2 (G2 ) Out(G2 ).
4. The Bass–Jiang group versus Out(G)
The Bass–Jiang group OutH (G) is most of interest when it is the
full outer automorphism group Out(G), so OutH (G) = Out(G). In
this section we prove Theorem 4.7, which gives conditions implying
that OutH (G) = Out(G). To prove these conditions we use the Pettet
group OutH (G), as defined in Section 3. Throughout the remainder of
this paper we assume all HNN-extensions are automorphism-induced,
so G = H∗(K,φ) , unless we explicitly state otherwise. We emphasize
that this assumption is necessary for our results.
Lemma 4.1 gives a condition implying OutH (G) = Out(G). Lemma
4.6 gives conditions implying OutH (G) = OutH (G). These lemmas
combine to prove Theorem 4.7.
8
ALAN D. LOGAN
4.1. Conditions implying OutH (G) = Out(G). Lemma 4.1 now
proves that for G = H∗(K,φ), if H has Serre’s property FA then the
Pettet group is equal to the full outer automorphism group. We then
have OutH (G) ≤ OutH (G) = Out(G).
We note the similarity between Lemma 4.1 and a comment of Pettet [Pet99, Introduction]. The principal difference is that Pettet is
additionally assuming conditions implying OutH (G) = OutH (G); we
require no such preliminary assumptions. The proof of Lemma 4.1
shows that under the conditions of the lemma the base group H is
always conjugacy maximal in G = H∗(K,φ) . The proof of Lemma 4.1
may be written purely algebraically; we have used Bass–Serre theory
so as to be similar to the other proofs on Section 4. Recall that K 6= H
in an automorphism-induced HNN-extension.
Lemma 4.1. Let G = H∗(K,φ). If H has Serre’s property FA then
OutH (G) = Out(G).
Proof. We prove that if ψ ∈ Aut(G) then ψ(H) = H g , where g ∈
G; this is sufficient as then ψγg−1 is a representative for ψb such that
ψγg−1 (H) = H. So, let ψ ∈ Aut(G) and consider the action of ψ(H)
on the Bass–Serre tree Γ of G. As H has Serre’s property FA, ψ(H)
stabilises some vertex of Γ. Hence, ψ(H) ≤ H g1 with g1 ∈ G. Similarly,
ψ −1 (H) ≤ H g2 with g2 ∈ G, and so H g3 ≤ ψ(H) with g3 := ψ(g2 )−1 ∈
G.
Suppose that ψ(H)
H g1 and we shall find a contradiction. By
g3
g1
the above, H
H . Now, there exist vertices u, w ∈ V Γ such that
g1
Gu = H and Gw = H g3 . If u = w then H g1 = H g3 , a contradiction.
So u 6= w and H g3 stabilises the geodesic [u, w]. In particular, H g3
stabilises the initial edge of [u, w], and therefore there exists g4 ∈ G
such that H g3 = H g4 ≤ K g4 or H g3 = H g4 ≤ φ(K)g4 . Hence H = K, a
contradiction.
4.2. Conditions implying OutH (G) = OutH (G). We now work
towards proving Lemma 4.6, which gives sufficient conditions for the
Bass–Jiang group to be equal to the Pettet group of G = H∗(K,φ). We
then have OutH (G) = OutH (G) ≤ Out(G).
Conjugacy. We begin our proof of Lemma 4.6 with Lemma 4.2,
which describes conjugacy in automorphism-induced HNN-extensions.
A word U in an HNN-extension G = H∗(K,K ′,η) is a product of the form
h0 tǫ1 h1 tǫ2 h2 · · · hn−1 tǫn hn where hi ∈ H, ǫi = ±1. For words U and V
we write U ≡ V if U and V are exactly the same word, and we say U is
t-reduced if it contains no subwords of the form tkt−1 , k ∈ K, or t−1 k ′ t,
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
9
k ′ ∈ K ′ . Every element of an HNN-extension has a representative word
which is t-reduced, by Britton’s Lemma.
Lemma 4.2. Let G = H∗(K,φ). For all g ∈ G there exists i ∈ Z and
a ∈ H such that if b, c ∈ H with bg = c then c = φi γa (b).
Note that i and a are dependent on g, not on b and c.
Proof. Let g ∈ G be arbitrary. Let u, w ∈ V Γ be the vertices of the
Bass–Serre tree Γ of G such that Gu = H and Gw = H g . Let W be a
t-reduced word in H∗(K,φ) representing g. We induct on n = |[u, w]|,
the length of the geodesic [u, w]. If n = 0 then u = w and the result
trivially holds. For the induction step, note that if n > 0 then [u, w]
decomposes as [u, v] and [v, w] where |[u, v]| = n − 1 and |[v, w]| = 1.
We also have a decomposition W ≡ V dn−1 tǫn dn , di ∈ H and where V
is such that Gv = H V and the result holds for V . Then:
ǫn d
n
c = bW = (bV )dn−1 t
ǫn d
n
= (φj γan−1 (b))dn−1 t
= φi γa (b)
as required, where i = j − ǫn and a = φ−ǫn (an−1 dn−1)dn .
We now prove Lemma 4.3, which corresponds to Condition (3) from
Lemma 4.6 (and Condition (3) from Theorem 4.7).
Lemma 4.3. Let G = H∗(K,φ). Suppose that OutH (G)
OutH (G).
Additionally, suppose that there does not exist any b ∈ H such that
φ(K) γb (K) and there does not exist any c ∈ H such that γc (K)
φ(K). Then there exists a ∈ H such that φ(K) = γa (K).
Proof. Suppose ψb ∈ OutH (G) \ OutH (G), and let ψ ∈ ψb be such that
ψ(H) = H. Recall that ψ acts on the vertices of the Bass–Serre tree
Γ of G. Let u, v ∈ V Γ be connected by a single edge e′ , and suppose
−1
that Gu = H and that Gv = H t . Then Ge′ = K. Write w := ψ(v),
−1
so Gw = H ψ(t) , and consider the geodesic [u, w]. As ψ 6∈ AutH (G),
|[u, w]| > 1. Note that ψ(K) stabilises [u, w] because K stabilises e′
−1
(or, algebraically, because ψ(K) ≤ ψ(H), ψ(H t )), so ψ(K) ≤ G[u,w].
Note also that the first edge e1 in this geodesic has stabiliser an Hconjugate either of K or of φ(K), and without loss of generality we can
assume Ge1 is either K or φ(K).
There exists a edge ej in the geodesic [u, w] such that Gej = ψ(K)
[Pet99, Lemma 2.2], so ψ(K) = K g for some g ∈ G. Therefore, Gej =
K g = ψ(K) ≤ G[u,w] ≤ Ge1 . As Ge1 = K or φ(K), we have that
10
ALAN D. LOGAN
Gej = K g ≤ K or φ(K), so Gej = φi γa0 (K) for some i ∈ Z, a0 ∈ H by
Lemma 4.2.
Recall that [u, w] contains more than one edge. Let f ∈ [u, w] be
ǫ
adjacent to ej . Then Gf = Geaj1 t a2 for some a1 , a2 ∈ H, ǫ = ±1, so
ǫa
2
φi γa0 (K) = Gej ≤ G[u,w] ≤ Gf = Geaj1 t
ǫa
2
= φi γa0 (K)a1 t
.
Noting that φi γa0 (K) ≤ H, we therefore have that every element of
φi γa0 (K) is equal to an element of the form (a1 tǫ a2 )−1 φi γa0 (k)a1 tǫ a2 ,
k ∈ K, and hence of the form φi−ǫ γa3 (k), with i, ǫ, a3 fixed. Hence,
φi γa0 (K) ≤ φi−ǫ γa3 (K), and so γa4 (K) ≤ φ−ǫ (K). By the assumptions
of the lemma γa (K) = φ(K), as required, where a = a4 if ǫ = −1 or
a = φ(a4 )−1 if ǫ = 1.
Residual finiteness. We now prove Lemma 4.4, which provides conditions for OutH (G) = OutH (G) to hold in a general HNN-extension G =
H∗(K,K ′,η) . Lemma 4.5 is essentially Lemma 4.4 for G an automorphisminduced HNN-extension. If the conditions of either of these lemmas
hold then the HNN-extension G is residually finite.
Baumslag–Tretkoff gave conditions for general HNN-extensions G =
H∗(K,K ′,η) to be residually finite [BT78], and the conditions of Lemma
4.4 are a strong form of these residual finiteness conditions. A subgroup
N ≤ H is characteristic if ψ(N) = N for all ψ ∈ Aut(H).
Lemma 4.4. Let G = hH, t; tkt−1 = η(k), k ∈ Ki where η : K → K ′ is
an isomorphism of subgroups of H, and let H be finitely generated and
residually finite. Suppose that for any positive integer n and elements
xi , yi , zi (i = 1, 2, . . . , n) with xi ∈ H \ K, yi ∈ H \ K ′ , zi ∈ K ∩ K ′ \ 1,
there exists a characteristic subgroup N of finite index in H such that
(1) xi K ∩ N = yi K ′ ∩ N = ∅, the empty set, and zi 6∈ N,
(2) η maps K ∩ N into N and η −1 maps K ′ ∩ N into N.
Then OutH (G) = OutH (G).
Proof. Note that η induces an isomorphism η ∗ of KN/N onto K ′ N/N.
We write G = hH/N, t; t(kN)t−1 = η ∗ (kN), k ∈ Ki. Then the natural
map h 7→ hN induces a homomorphism of G onto G [BT78, Proof of
Theorem 4.2]. Note that H/N is finite, so OutH/N (G) = OutH/N (G)
[Pet99, Theorem 1].
Suppose OutH (G) OutH (G). Then there exists µ ∈ AutH (G) with
µ|H ∈ Aut(H) such that µ(t) = h1 tǫ1 h2 tǫ2 . . . hm tǫm is t-reduced with
m > 1. Note that for each j = 1, . . . , m, either hj ∈ H \ K or hj ∈
H \ K ′ . Therefore, pick n and the elements xi , yi , zi , i = 1, . . . , n, such
that that each hj is an xi or yj as appropriate, pick N characteristic in
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
11
H appropriately and form the corresponding group G. Now, µ induces
a map µ∗ : G → G with µ∗ (t) = h1 Ntǫ1 h2 Ntǫ2 . . . hm Ntǫm and µ∗ (h) =
µ(h)N.
We now prove that µ∗ ∈ Aut(G). Firstly, note that µ∗ |H/N ∈
Aut(H/N) as N is characteristic in H. Then, µ∗ (t)(kN)µ∗ (t)−1 =
η ∗ (kN) for all k ∈ K because µ and µ∗ |H/N are homomorphisms.
Hence, µ∗ : G → G is a homomorphism. It is surjective because µ(t)
and H generate G, and it is injective because G is finitely generated
and residually finite [BT78, Theorem 3.1] and hence Hopfian.
Then µ∗ ∈ AutH (G) \ AutH (G) as µ∗ |H/N ∈ Aut(H/N) and the
word h1 Ntǫ1 h2 Ntǫ2 . . . hm Ntǫm is t-reduced (as each hj is an xi or yi as
appropriate). Hence, OutH/N (G) OutH/N (G), a contradiction.
The only difference between the conditions of Lemma 4.4 and those
of Baumslag–Tretkoff is that we require the finite-index subgroup N
to be characteristic rather than normal. Suppose, as with Baumslag–
Tretkoff, we instead assume that N is normal in H. Then, as H is
b ≤f N which
finitely generated, there exists a characteristic subgroup N
b is such that
satisfies the first condition. However, it is not clear that N
b
b
η(K ∩ N ) ≤ N .
Now, if G = H∗(K,φ) is automorphism-induced then the characteristic
b also satisfies the second condition, as for example φ(K ∩
subgroup N
b = φ(K) ∩ φ(N)
b ≤ φ(N)
b =N
b . This observation gives the following
N)
lemma, which corresponds to Condition (3) of Lemma 4.6 (also compare
with [BT78, Lemma 4.4]). A subgroup K of H is residually separable
in H if for all x ∈ H \ K there exists a finite index, normal subgroup
N of H, written N ⊳f H, such that x 6∈ KN.
Lemma 4.5. Let G = H∗(K,φ), and let H be finitely generated and
residually finite. If K is residually separable in H then OutH (G) =
OutH (G).
Proof. As K is residually separable in H the conditions of Lemma 4.4
hold. The result follows.
An alternative proof of Lemma 4.5 is as follows: By combining
Lemma 4.3 with a result of Shirvani [Shi85, Lemma 3], we see that the
conditions of Lemma 4.5 imply that φ(K) = γa (K) for some a ∈ H.
The result follows from Condition (2) of Lemma 4.6, below (the proof
of this condition is independent of the proof of Condition (4)).
Proof of Lemma 4.6. Lemma 4.6 combines Lemma 4.3 and Lemma
12
ALAN D. LOGAN
4.5 with results of Pettet. A subgroup K of a group H is conjugacy
maximal if there does not exist any a ∈ H such that K K a .
Lemma 4.6. Let G = H∗(K,φ). Suppose that one of the following holds:
(1) K is conjugacy maximal in H;
(2) φ(K) = γa (K) for some a ∈ H; or
(3) There does not exist any b ∈ H such that φ(K)
γb (K) and
there does not exist any c ∈ H such that γc (K) φ(K); or
(4) H is finitely generated and residually finite, and K is residually
separable in H; or
(5) H is finitely generated and G is residually finite.
Then OutH (G) = OutH (G).
Proof. If (1) or (2) holds then the result is known to hold [Pet99,
Lemma 2.6 and Theorem 1]. Now, note that it is impossible for
(3) to hold simultaneously to the inequality OutH (G)
Out(G): if
both held then (2) would hold, by Lemma 4.3, and so we obtain
OutH (G) = Out(G), a contradiction. Hence, if (3) holds then the
result holds.
If (4) holds then the result holds by Lemma 4.5. If (5) holds then
(4) holds [Log17b]. The result follows.
4.3. The Bass–Jiang group versus Out(G). Theorem 4.7 now gives
conditions implying that the Bass–Jiang group is the full outer automorphism group.
Theorem 4.7. Let G ∼
H and
= hH, t; tkt−1 = φ(k), k ∈ Ki, K
φ ∈ Aut(H), be an automorphism-induced HNN-extension. If H has
Serre’s property FA then OutH (G) = Out(G). If any one of the following statements holds then OutH (G) = OutH (G):
(1) K is conjugacy maximal in H;
(2) φ(K) = γa (K) for some a ∈ H; or
(3) There does not exist any b ∈ H such that φ(K)
γb (K) and
there does not exist any c ∈ H such that γc (K) φ(K); or
(4) K is residually separable in H; or
(5) G is residually finite.
Proof. Theorem 4.7 follows immediately from Lemmas 4.1 and 4.6.
Note that Theorem 4.7 obtains OutH (G) = Out(G) in two disjoint steps; it proves that OutH (G) = Out(G) and separately that
OutH (G) = OutH (G). M. Pettet proved a similar result to Theorem
4.7 involving Conditions (1) and (2) [Pet99, Theorem 1], but he does
not obtain these disjoint steps due to his more general setting.
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
13
Theorem 4.7 gives a concrete method of proving that an automorphisminduced HNN-extension G = H∗(K,φ) is not residually finite: if there
exists some ψ ∈ AutH (G) \ AutH (G) (that is, OutH (G) OutH (G))
then G is not residually finite. Lemma 4.8, below, provides a way of
proving this inequality.
4.4. A condition implying OutH (G)
OutH (G). We do not
know if the conditions of Lemma 4.6 are both necessary and sufficient for OutH (G) = OutH (G) to hold, and reading closely the proof of
Lemma 4.3 one might suspect that they are. Lemma 4.8 now demonstrates how close the conditions of Lemma 4.6 are to being necessary
and sufficient.
Lemma 4.8 demonstrates this by proving that if Conditions (1) and (3)
of Lemma 4.6 fail “compatibly” with one another then OutH (G)
OutH (G). Note that these two conditions fail for G = H∗(K,φ) if
and only if, after replacing φ with φ−1 if necessary, there exists some
a, b ∈ H such that γa (K)
K
φγb(K). The “compatibility” we
require is that we can take a ∈ φγb(K).
Lemma 4.8. Let G = H∗(K,φ). Suppose also that K a
where a ∈ φγb(K), b ∈ H. Then OutH (G) OutH (G).
K
φγb(K)
Proof. By replacing φ with φγb, we may assume b = 1. Note that a 6∈ K
as γa (K) K, and so the word t2 at−1 φ(a)−1 is t-reduced. We prove
that the map
ψ : h 7→ h
∀h ∈ H
t 7→ t2 at−1 φ(a)−1
is an automorphism of G; then ψ ∈ AutH (G) \ AutH (G) and the result
follows.
Now, φ(a)−1 kφ(a) ∈ φ(K) so ψ : G → G is a homomorphism. To
see that ψ is a surjection we prove that t ∈ im(ψ). Begin with the
following, where a1 = φ−1 (a) ∈ K. Note that a−1 a−1
1 a ∈ K.
−1
(t2 at−1 φ(a)−1 )a−1 = t2 at−1 φ(a−1 a−1
1 aa )
−1
−1
= t2 a−1
1 at φ(a )
= ta−1 tat−1 φ(a−1 )
14
ALAN D. LOGAN
We then have that:
(t2 at−1 φ(a)−1 )φ(a) (t2 at−1 φ(a)−1 )a−1 (t2 at−1 φ(a)−1 )−1
= (t2 at−1 φ(a)−1 )φ(a) ta−1 tat−1 φ(a)−1 (t2 at−1 φ(a)−1 )−1
= (t3 at−1 φ(a)−1 )(t2 at−1 φ(a)−1 )−1
=t
Therefore, ψ is surjective. It is injective because it is invertible, with
inverse ψ −1 : h 7→ h for all h ∈ H, t 7→ tφ(a)ta−1 t−1 .
5. Describing the Bass–Jiang group
In this section we prove Theorem 5.6, which gives a short-exact sequence description of the Bass–Jiang group of G = H∗(K,φ) . Combined
with Theorem 4.7, it allow for a description of Out(G).
Bass–Jiang gave a decomposition of the group OutH (G) for G a
general HNN-extension [BJ96]. Their description took the form of a
filtration. However, their description contains both of the associated
subgroups K and φ(K), while our description in Theorem 5.6 only
(V )
contains K. The difference is in the description of the kernel OutH (G)
(V )
from the short exact sequence in Theorem 5.6 (OutH (G) is formally
(V )
defined below). Specifically, Bass–Jiang find some N E OutH (G) such
that
(V )
OutH (G) ∼ (NH (φ(K)) ∩ NH (K)) /Z(H)
.
=
N
Z(H)K
In general, NH (φ(K)) ∩ NH (K) 6= NH (K). Therefore Bass–Jiang’s de(V )
composition of OutH (G) is different from the decomposition in The(V )
orem 5.6. Our simpler description of OutH (G) is fundamental to the
framework described in Section 7 and is essential in proving Theorem
1.1.
Results of Bass–Jiang. Our proof of Theorem 5.6 is based on two
results of Bass–Jiang, which we state now. We have translated their
results and definitions into our setting of automorphism-induced HNNextensions. The first result is a classification of the elements of the
Bass–Jiang group OutH (G) [BJ96, Theorem 5.6].
Theorem 5.1 (Bass–Jiang). Let G = H∗(K,φ) . If δ ∈ Aut(H), a ∈ H
are such that δ(K) = K and δφ(k) = φδγa (k) for all k ∈ K then the
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
15
map
α(δ,a) : h 7→ δ(h)
t 7→ at
∀h ∈ H
is an automorphism of G.
If ζ ∈ Aut(H), b ∈ H are such that ζ(K) = φ(K), ζ 2γb (K) = K,
and ζφ−1(k) = φζγb(k) for all k ∈ K then the map
β(ζ,b) : h 7→ ζ(h)
∀h ∈ H
−1
t 7→ bt
is an automorphism of G.
Moreover, every element ψb of OutH (G) has a representative in AutH (G)
of the form α(δ,a) or of the form β(ζ,b) .
The second result of Bass–Jiang which we apply requires certain definitions. The restrictions on a and b in the second, equivalent definition
(V )
of OutH (G) follow from Theorem 5.1:
α(δ,a) : a ∈ H, δ(K) = K, δφ(k) = φδγa (k) for all k ∈ K},
Out0H (G) := {b
(V )
OutH (G) := {b
α(γb ,a) : a ∈ H, γb ∈ Inn(H)}
= {b
α(γb ,a) : a ∈ H, b ∈ NH (K), baφ(b)−1 ∈ CH (φ(K))}.
The subgroup Out0H (G) clearly has index two in OutH (G) if there exists
some automorphism β(ζ,b) of G, and index one otherwise.
The second result of Bass–Jiang which we apply is the initial two
steps of their filtration decomposition of OutH (G) [BJ96, Theorem
8.12]. Recall that the subgroup AK ≤ Aut(H) was defined in the
introduction.
Theorem 5.2 (Bass–Jiang). Let G = H∗(K,φ). The surjection
AK Inn(H)
Inn(H)
7→ δb
χ1 : Out0H (G) →
(V )
α
b(δ,a)
has kernel OutH (G), and either Out0H (G) = OutH (G) or there exists
some ζ ∈ Aut(H), b ∈ H such that ζ(K) = φ(K), ζ 2γb (K) = K, and
ζφ−1(k) = φζγb(k) for all k ∈ K, whence Out0H (G) has index two in
OutH (G).
2Section
8.5 of Bass–Jiang’s paper is a restriction of their Theorem 8.1 to HNNextensions. We cite their more general Theorem 8.1 because certain errors have
crept in to the re-statement in Section 8.5.
16
ALAN D. LOGAN
(V )
The subgroup OutH (G) of Out0H (G). We wish to describe the
(V )
kernel OutH (G) from Theorem 5.2; combining the description we obtain with Theorem 5.2 proves the main result of this section, Theorem
5.6. Our description is given by Lemmas 5.4 and 5.5. These lemmas
(V )
both mention the subgroup CK of OutH (G):
CK : = {b
α(1,a) : a ∈ H}
= {b
α(1,a) : a ∈ CH (φ(K))}
Lemmas 5.4 and 5.5 also both apply the following technical result.
Lemma 5.3. Let G = H∗(K,φ), and let g ∈ G and x, y ∈ H. Suppose
that both of the following hold.
(1)
(2)
g −1 ytg = x−1 φ(x)t
g −1 hg = x−1 hx
∀h∈H
Then g ∈ K.
Proof. Begin by writing g as a t-reduced word zW , where z ∈ H and W
is either empty or begins with t±1 . Assume that W is non-empty. Now,
(2) has the form W −1 z −1 hzW = x−1 hx. As K H, h can be chosen
such that the word W −1 z −1 hzW is t-reduced (z −1 hz not in K or φ(K)
as appropriate), a contradiction by Britton’s Lemma [LS77, Section
IV.2]. Hence, W is empty. So g = z ∈ H. It then follows from (1) that
g ∈ K, as required.
Our proofs of Lemmas 5.4 and 5.5, and of certain proofs in Section 6,
require calculations with automorphisms of the form α(δ,a) . We note
the following identities:
α(δ1 ,a1 ) α(δ2 ,a2 ) = α(δ1 δ2 ,δ2 (a1 )a2 )
−1
α(δ,a)
= α(δ−1 ,δ−1 (a−1 ))
Certain calculations are written in terms of cosets α
b(δ,a) , but these
follow immediately from the above two identities.
Lemma 5.4. The map
(V )
χ2 : OutH (G) → NH (K)/Z(H)K
α
b(γb ,a) 7→ bZ(H)K
is a well-defined surjective homomorphism with kernel CK .
Proof. Suppose the map χ2 is well-defined. Note that for all b ∈
NH (K), the pair (γb , b−1 φ(b)) defines the automorphism α(γb ,b−1 φ(b)) by
Theorem 5.1, and so χ2 is surjective. As α
b(γb1 ,a1 ) α
b(γb2 ,a2 ) = α
b(γb1 b2 ,a3 )
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
17
for some a3 ∈ H, we have that χ2 is also a homomorphism. Now,
CK ≤ ker χ2 , but as α
b(γk ,a) = α
b(1,kaφ(k)−1 ) for all k ∈ K we have that
ker χ2 = CK .
To see that the map χ2 is well-defined, suppose α
b(γb1 ,a1 ) = α
b(γb2 ,a2 )
−1
and it is sufficient to prove that b1 b2 ∈ Z(H)K. Note that if α(γb ,a) ∈
Aut(G) then α(γb ,b−1 φ(b)) ∈ Aut(G), and then α(γb ,a) = α(γb ,b−1 φ(b)) α(1,φ(b)−1 ba) .
Therefore, the identity α
b(γb1 ,a1 ) = α
b(γb2 ,a2 ) becomes:
α
b(γb
1
α
b(1,φ(b1 )−1 b1 a1 )
,b−1
1 φ(b1 ))
=α
b(γb
2
α
b(1,φ(b2 )−1 b2 a2 )
,b−1
2 φ(b2 ))
⇒α
b(1,a3 ) = α
b(γ −1
b1
b2
−1 φ(b−1 b ))
,(b−1
2
1 b2 )
1
for some a3 ∈ H. Hence, α(1,a3 ) γg = α(γ −1
b
1
b2
−1 φ(b−1 b ))
,(b−1
2
1 b2 )
1
for some
g ∈ G and a3 ∈ H. Therefore, the conditions of Lemma 5.3 hold for the
two underlying identities (where x := b−1
1 b2 and y := a3 ), hence g ∈ K
−1
−1
with g −1 hg = (b−1
b
)
hb
b
for
all
h
∈
H, and so b−1
2
2
1
1
1 b2 ∈ Z(H)K as
required.
Note that NK := {b
α(γb ,b−1 φ(b)) : b ∈ NH (K)} is a subgroup of
(V )
(V )
OutH (G), and OutH (G) = CK NK . So by Lemma 5.4 we just need
CK ∩ NK = 1 to obtain a semidirect product in the statement of Theorem 5.6. This possible splitting of the map χ2 is discussed in Section 6.
Lemma 5.5. Define L = {φ−1 (k)k −1 : k ∈ K ∩ Z(H)}. Then L E
CH (K)
CH (K) and CK ∼
.
=
L
C (φ(K))
; the result then follows. To begin,
Proof. We prove that CK ∼
= H
φ(L)
note that α
b(1,a1 ) α
b(1,a2 ) = α
b(1,a1 a2 ) , and so CK is a homomorphic image
of CH (φ(K)). We therefore prove that α(1,a) ∈ Inn(G) if and only if
a ∈ φ(L).
Clearly if a = kφ(k)−1 ∈ φ(L) then α(1,a) = γk−1 ∈ Inn(G). On
the other hand, suppose α(1,a) ∈ Inn(G). To see that a ∈ φ(L), note
that the conditions of Lemma 5.3 hold (where x = 1, y = a, and in
the following we take k := g −1 ). Therefore, we have that k −1 hk = h
for all h ∈ H and that k −1 tk = at, where k ∈ K. The first identity
implies k ∈ Z(H) (and so k ∈ Z(H) ∩ K), while the second implies
a = k −1 φ(k). Hence, a ∈ φ(L).
Describing OutH (G). We can now give our description of the Bass–
Jiang group OutH (G) for G = H∗(K,φ).
Theorem 5.6. Let G ∼
H and
= hH, t; tkt−1 = φ(k), k ∈ Ki, K
φ ∈ Aut(H), be an automorphism-induced HNN-extension. Let L =
18
ALAN D. LOGAN
{φ−1(k)k −1 : k ∈ K ∩ Z(H)}. Then L is a normal subgroup of CH (K)
and we have a diagram of exact sequences:
1
CH (K)/L
1
(V )
OutH (G)
Out0H (G)
AK Inn(H)
Inn(H)
1
NH (K)/Z(H)K
1
where either Out0H (G) = OutH (G) or there exists some ζ ∈ Aut(H)
and some a ∈ H such that ζ(K) = φ(K), ζ 2γa (K) = K and ζφ−1(k) =
φζγa(k) for all k ∈ K, whence Out0H (G) has index two in OutH (G).
(V )
Proof. By Theorem 5.2, it is sufficient to prove that OutH (G) decomposes as in the statement of Theorem 5.6. This follows from Lemmas 5.4 and 5.5.
An alternative view. Recall the definition of AK :
AK := {δ ∈ Aut(H) : δ(K) = K, ∃a ∈ H s.t. δφ(k) = φδγa (k)∀k ∈ K}
The following isomorphism can be substituted in to Theorem 5.6. Here,
Inn(NH (K)) = {γb : b ∈ NH (K)}.
AK Inn(H) ∼
AK
=
Inn(H)
Inn(NH (K))
This isomorphism holds because Inn(H) ∩ AK = Inn(NH (K)); that
Inn(H)∩AK is a subgroup of Inn(NH (K)) is clear from the definition of
AK , while if γb (K) = K then γb ∈ AK because γb φ(k) = φγb γb−1 φ(b) (k).
6. The impact of commutativity in H
(V )
The reader might feel cheated that the decomposition of OutH (G) in
Theorem 5.6 did not split over the subgroups CK and NK , as defined in
(V )
(V )
Section 5. Indeed, OutH (G) = CK NK and CK E OutH (G). However,
elements of the form α
b(γc ,c−1 φ(c)) where c ∈ Z(H) are contained in
(V )
both subgroups, so the decomposition of OutH (G) in Theorem 5.6
does not split in general. In this section we prove Theorem 6.2, which
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
19
(V )
assumes that Z(H) ≤ Fix(φ); here the decomposition of OutH (G)
does split over CK and NK . We also prove related results which assume
commutativity conditions on H.
Splitting the description of Theorem 5.6. We begin by “reversing” the map χ2 from Lemma 5.4, which allows us to determine if χ2
splits.
Lemma 6.1. The map
(V )
χ2 : NH (K) → OutH (G)
b 7→ α
b(γb ,b−1 φ(b))
is a homomorphism with kernel JK, where J = Z(H) ∩ Fix(φ).
Proof. As α
b(γb ,b−1
b(γb ,b2 −1 φ(b2 )) = α
b(γb b ,b2 −1 b1 −1 φ(b1 b2 )) , the map χ2
φ(b1 )) α
2
1 2
1 1
is a homomorphism. Note that ker χ2 ≥ JK as for b ∈ J, k ∈ K we
have α(γbk ,(bk)−1 φ(bk)) = γk . To see that the ker χ2 ≤ JK, suppose that
α(γb ,b−1 φ(b)) = γg . Then the conditions of Lemma 5.3 hold for the two
underlying identities (where x = b, y = 1), so k := g ∈ K. Hence,
bk −1 ∈ Z(H) and b−1 φ(b) = k −1 φ(k). The second identity implies that
bk −1 ∈ Fix(φ). Therefore, bk −1 ∈ Z(H) ∩ Fix(φ) = J so b ∈ JK as
required.
The hypotheses of Theorem 6.2 are those of Theorem 5.6 with the
additional assumption that Z(H) ≤ Fix(φ). Here the vertical exact
sequence from Theorem 5.6 splits. Theorem 6.2 does not immediately
follow from Theorem 5.6. The subgroup NK corresponds to the factor NH (K)/Z(H)K, while CK corresponds to the factor CH (K). The
α(γ ,b−1 φ(b))
action of NK on CK is α(1,a)b
= α(1,γφ(b) (a)) .
Theorem 6.2. Let G be as in Theorem 5.6, and additionally assume
Z(H) ≤ Fix(φ). We have a short exact sequence:
1 → CH (K) ⋊
NH (K)
AK Inn(H)
→ Out0H (G) →
→1
Z(H)K
Inn(H)
where the conditions for the index | OutH (G) : Out0H (G)| are as in
Theorem 5.6.
Proof. As Z(H) ≤ Fix(φ), Lemma 6.1 implies that the map χ2 from
Lemma 5.4 splits. Theorem 6.2 then follows from Theorem 5.6.
Virtually embedding NH (K)/K into Out(G). Lemma 6.3 applies Lemma 6.1 to obtain conditions implying a finite-index subgroup
20
ALAN D. LOGAN
of NH (K)/K embeds into Out(G). This normaliser-quotient is the basis of the framework described in Section 7 (see also the discussion after
Proposition 2.1). Lemma 6.3 is fundamental to the paper [Log16]. The
hypotheses of Lemma 6.3 occur when, for example, H is non-elementary
hyperbolic, K EV and V is torsion-free (that is, V contains no elements
of finite order).
Lemma 6.3. If K E V ≤ NH (K) with V ∩ Z(H) ∩ Fix(φ) = 1 then
(V )
V /K embeds into OutH (G) with index dividing |CH (K)|·|NH (K) : V |.
Proof. Recall that J := Z(H) ∩ Fix(φ). To prove that V /K em(V )
beds into OutH (G), note that the map χ2 induces maps NH (K) →
NH (K)/K → NH (K)/JK (the maps are the canonical maps). It is
sufficient to prove that if v ∈ V \ K then vK 67→ JK under the second
map. Suppose vK 7→ JK and vK 6= K, then there exists some k ∈ K
such that vk ∈ J. However, vk ∈ V and V ∩ J = 1 so v = k −1 ∈ K a
(V )
contradiction. Hence, V /K embeds into OutN (G).
(V )
To prove that the index of V /K in OutH (G) divides |CH (K)| ·
|NH (K) : V |, first note that, by considering the map χ2 , the index of
(V )
V /K in OutH (G) clearly divides
(V )
| OutH (G) : im(χ2 )| · |NH (K) : V |.
(V )
We therefore prove that | OutH (G) : im(χ2 )| divides |CH (K)|. It is
routine to construct an isomorphism between P/(P ∩ Q) and P Q/Q,
Q not necessarily normal in P Q. Take P := CK and Q := NK . Not(V )
ing that im (χ2 ) = NK and that CK NK = OutH (G), we have that
(V )
| OutH (G) : im (χ2 ) | divides |CK |. Then |CK | divides |CH (K)| by
Lemma 5.5, as required.
Describing Aut0H (G). Our description of Out0H (G) from Theorems 5.6
and 6.2 can be used to describe the full pre-image Aut0H (G) of Out0H (G)
in Aut(G). However, this description is not necessarily useful. Lemma
6.4 now gives conditions which allow for a complete description of
Aut0H (G). Lemma 6.4 is applied in the paper [Log17a]. Recall that
NK = {b
α(γb ,b−1 φ(b)) : b ∈ NH (K)}.
Lemma 6.4. If AK ≤ Inn(H) and CH (K) = 1 then Out0H (G) ∼
=
NH (K)/K and Aut0H (G) can be described as follows.
Aut0H (G) = Inn(G) ⋊ Out0H (G)
NH (K)
∼
= G⋊
K
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
21
Proof. Note that α(γb ,a) = α(γb ,b−1 φ(b)) α(1,φ(b)−1 ba) , with φ(b)−1 ba ∈ CH (φ(K)).
(V )
Then α(1,φ(b)−1 ba) = 1 as CH (K) = 1, so OutH (G) = NK . As AK ≤
Inn(H), Theorem 5.2 gives us that Out0H (G) = NK .
The representatives α(γb ,b−1 φ(b)) of NK generate a subgroup of Aut(G)
which contains non-trivial inner automorphisms. We therefore consider
the set Φ := {ϕd ; d ∈ NH (K)}. As ϕd = α(γd ,d−1 φ(d)) γd−1 , the set Φ
contains a representative for each coset of Aut0H (G)/ Inn(G). Hence,
Φ Inn(G) = Aut0H (G). Then ϕd ∈ Inn(G) if and only if d ∈ K, if and
only if ϕd = 1, so Φ ∩ Inn(G) = 1. Therefore, Aut0H (G) = Inn(G) ⋊ Φ.
The result then follows as Φ ∼
= NH (K)/K by Proposition 2.1 while
Inn(G) ∼
= G because CH (K) = 1.
7. Tractable groups with pathological outer
automorphism groups
In this section we package the technical results of this paper into a
framework for the construction of tractable groups with pathological
outer automorphism groups. We then apply this framework to prove
Theorem 1.1 from the introduction. Other applications of this framework can be found in [Log16] and [Log17a].
The framework. In the aforementioned framework the pathological properties of Out(G), where G = H∗(K,φ), are inherited from the
normaliser-quotient NH (K)/K. The framework is as follows:
(1) Obtain K < H such that NH (K)/K has the required properties, e.g. is a stipulated group, is non-residually finite, etc.
(2) Apply Theorems 5.6 or 6.2, or Lemma 6.3 to obtain a description of OutH (G) in terms of NH (K)/K,
(3) Apply Theorem 4.7 to obtain OutH (G) = Out(G), and so obtain a description of Out(G) in terms of NH (K)/K,
(4) Apply Lemma 6.4 to obtain AutH (G) = Inn(G) ⋊ Out(G), and
so obtain a description of Aut(G) in terms of NH (K)/K.
Proof of Theorem 1.1. We now prove Theorem 1.1.
Proof of Theorem 1.1. Suppose Q is finitely presented with Serre’s property FA. Now, there exist torsion-free hyperbolic groups with Serre’s
property FA [Pri83], and hence there exists a torsion-free hyperbolic
group H with Serre’s property FA, and with a finitely generated noncyclic normal subgroup K such that H/K ∼
= Q [BO08]. Form GQ =
hH, t; k t = k, k ∈ Ki. Note that GQ is residually finite if (and only if)
Q is residually finite [Log16, Proposition 2.2].
22
ALAN D. LOGAN
Firstly note that OutH (GQ ) = Out(GQ ), which follows from Theorem 4.7 because K is normal so conjugacy maximal, and because
H has Serre’s property FA. Secondly, note that NH (K)/K embeds
with finite index into OutH (GQ ), which follows from Theorem 5.6 because CH (K) = 1 as H is torsion-free hyperbolic and K is non-cyclic,
and because Out(H) is finite as H is hyperbolic with Serre’s property
FA [Lev05].
Hence, Q ∼
= H/K = NH (K)/K ≤f OutH (GQ ) = Out(GQ ) as required.
References
[ALP17]
F. Ateş, A. D. Logan, and S.J. Pride, Automata and Zappa-Szèp products of groups, in preparation (2017).
[Bau63]
G. Baumslag, Automorphism groups of residually finite groups, J. London Math. Soc. 38 (1963), 117–118. MR 0146271
[BG03]
G. Braun and R. Göbel, Outer automorphisms of locally finite p-groups,
J. Algebra 264 (2003), no. 1, 55–67. MR 1980685
[BJ96]
H. Bass and R. Jiang, Automorphism groups of tree actions and of
graphs of groups, J. Pure Appl. Algebra 112 (1996), no. 2, 109–155.
MR 1402782
[BO08]
I. Belegradek and D. Osin, Rips construction and Kazhdan property
(T), Groups Geom. Dyn. 2 (2008), no. 1, 1–12. MR 2367206
[BT78]
B. Baumslag and M. Tretkoff, Residually finite HNN extensions,
Comm. Algebra 6 (1978), no. 2, 179–194. MR 484178
[BW05]
I. Bumagin and D. T. Wise, Every group is an outer automorphism
group of a finitely generated group, J. Pure Appl. Algebra 200 (2005),
no. 1-2, 137–147. MR 2142354
[CL83]
D. J. Collins and F. Levin, Automorphisms and Hopficity of certain
Baumslag-Solitar groups, Arch. Math. (Basel) 40 (1983), no. 5, 385–
400. MR 707725
[DGG01] M. Droste, M. Giraudet, and R. Göbel, All groups are outer automorphism groups of simple groups, J. London Math. Soc. (2) 64 (2001),
no. 3, 565–575. MR 1865550
[DGP11] F. Dahmani, V. Guirardel, and P. Przytycki, Random groups do not
split, Math. Ann. 349 (2011), no. 3, 657–673. MR 2755002
[Far11]
D. Farley, A proof that Thompson’s groups have infinitely many relative
ends, J. Group Theory 14 (2011), no. 5, 649–656. MR 2831963
[FM06]
R. Frigerio and B. Martelli, Countable groups are mapping class groups
of hyperbolic 3-manifolds, Math. Res. Lett. 13 (2006), no. 5-6, 897–910.
MR 2280783
[GHMR00] N. D. Gilbert, J. Howie, V. Metaftsis, and E. Raptis, Tree actions
of automorphism groups, J. Group Theory 3 (2000), no. 2, 213–223.
MR 1753479
[GP00]
R. Göbel and A. T. Paras, Outer automorphism groups of metabelian
groups, J. Pure Appl. Algebra 149 (2000), no. 3, 251–266. MR 1762767
BASS–JIANG GROUP FOR AUT.-INDUCED HNN-EXTENSIONS
[Kat15]
[Koj88]
[Lev05]
[Lev07]
[Log15]
[Log16]
[Log17a]
[Log17b]
[LS77]
[Mat89]
[Mes72]
[Min09]
[Pet99]
[Pri83]
[Ser74]
[Ser80]
[Shi85]
[Wat82]
23
M. Kato, Higher dimensional Thompson groups have Serre’s property
FA, arXiv: 1504.06680 (2015).
S. Kojima, Isometry transformations of hyperbolic 3-manifolds, Topology Appl. 29 (1988), no. 3, 297–307. MR 953960
G. Levitt, Automorphisms of hyperbolic groups and graphs of groups,
Geom. Dedicata 114 (2005), 49–70. MR 2174093
, On the automorphism group of generalized Baumslag–Solitar
groups, Geometry & Topology 11 (2007), no. 1, 473–515.
A. D. Logan, On the outer automorphism groups of finitely generated,
residually finite groups, J. Algebra 423 (2015), 890–901. MR 3283743
, On a question of Bumagin and Wise, New York J. Math. 22
(2016), 865–873. MR 3548127
, Every group is the outer automorphism group of an HNNextension of a fixed triangle group, arXiv:1709.06441 (2017).
, The residual finiteness of (hyperbolic) automorphism-induced
HNN-extensions, arXiv:1711.05797 (2017).
R. C. Lyndon and P. E. Schupp, Combinatorial group theory, SpringerVerlag, Berlin-New York, 1977, Ergebnisse der Mathematik und ihrer
Grenzgebiete, Band 89. MR 0577064
T. Matumoto, Any group is represented by an outerautomorphism
group, Hiroshima Math. J. 19 (1989), no. 1, 209–219. MR 1009671
S. Meskin, Nonresidually finite one-relator groups, Trans. Amer. Math.
Soc. 164 (1972), 105–114. MR 0285589
A. Minasyan, Groups with finitely many conjugacy classes and their
automorphisms, Comment. Math. Helv. 84 (2009), no. 2, 259–296.
MR 2495795
M. R. Pettet, The automorphism group of a graph product of groups,
Comm. Algebra 27 (1999), no. 10, 4691–4708. MR 1709254
Stephen J. Pride, Some finitely presented groups of cohomological dimension two with property (FA), J. Pure Appl. Algebra 29 (1983),
no. 2, 167–168. MR 707619
J.P. Serre, Amalgames et points fixes, Proceedings of the Second International Conference on the Theory of Groups (Australian Nat. Univ.,
Canberra, 1973), Springer, Berlin, 1974, pp. 633–640. Lecture Notes in
Math., Vol. 372. MR 0376882
, Trees, Springer, 1980.
M. Shirvani, On residually finite HNN-extensions, Arch. Math. (Basel)
44 (1985), no. 2, 110–115. MR 780256
Y. Watatani, Property T of Kazhdan implies property FA of Serre,
Math. Japon. 27 (1982), no. 1, 97–103. MR 649023
School of Mathematics and Statistics, University of Glasgow, Glasgow, G12 8QW, UK.
E-mail address: [email protected]
| 4math.GR
|
Partial Consistency with Sparse Incidental Parameters
Jianqing Fan1 , Runlong Tang2 , and Xiaofeng Shi1
arXiv:1210.6950v3 [math.ST] 2 Aug 2017
1
2
Princeton University
Johns Hopkins University
March 14, 2018
Abstract
Penalized estimation principle is fundamental to high-dimensional problems. In the literature, it has been extensively and successfully applied to various models with only structural
parameters. As a contrast, in this paper, we apply this penalization principle to a linear regression model with a finite-dimensional vector of structural parameters and a high-dimensional
vector of sparse incidental parameters. For the estimators of the structural parameters, we derive their consistency and asymptotic normality, which reveals an oracle property. However, the
penalized estimators for the incidental parameters possess only partial selection consistency but
not consistency. This is an interesting partial consistency phenomenon: the structural parameters are consistently estimated while the incidental ones cannot. For the structural parameters,
also considered is an alternative two-step penalized estimator, which has fewer possible asymptotic distributions and thus is more suitable for statistical inferences. We further extend the
methods and results to the case where the dimension of the structural parameter vector diverges
with but slower than the sample size. A data-driven approach for selecting a penalty regularization parameter is provided. The finite-sample performance of the penalized estimators for the
structural parameters is evaluated by simulations and a real data set is analyzed.
Keywords: Structural Parameters, Sparse Incidental Parameters, Penalized Estimation, Partial Consistency,
Oracle Property, Two-Step Estimation, Confidence Intervals
1
Introduction
Since the pioneering papers by Tibshirani (1996) and Fan and Li (2001), the penalized estimation
methodology for exploiting sparsity has been studied extensively. For example, Zhao and Yu (2006)
1
provides an almost necessary and sufficient condition, namely Irrepresentable Condition, for the
LASSO estimator to be strong sign consistent. Fan and Lv (2011) shows that an oracle property
holds for the folded concave penalized estimator with ultrahigh dimensionality. For an overview on
this topic, see Fan and Lv (2010).
All the aforementioned papers consider models only with the so-called structural parameters,
which are related to every data point. By contrast, we consider in this paper another type of models
where there are not only the structural parameters but also the so-called incidental parameters,
each of which is related to only one data point. Specifically, suppose data {X i , Yi }ni=1 are from the
following linear model:
Yi = µ?i + X Ti β ? + i ,
(1.1)
where the vector of incidental parameters µ? = (µ?1 , · · · , µ?n )T is sparse, the vector of structural
parameters β ? = (β1? , · · · , βd? )T is of main interest, and for each i, X i is a d-dimensional covariate
vector, and i is a random error. Let ν = (µ?T , β ?T )T . Then, in model (1.1), a different data point
(X i , Yi ) depends on a different subset of ν, that is, µ?i and β ? .
Model (1.1) arises as a working model for estimation from Fan et al. (2012b), which considers a
large-scale hypothesis testing problem under arbitrary dependence of test statistics. By Principal
Factor Approximation, a method proposed by Fan et al. (2012b), the dependent test statistics
Z = (Z1 , · · · , Zp )T ∼ N (µ, Σ) can be decomposed as Zi = µi +bTi W +Ki , where µ = (µ1 , · · · , µp )T
and bi is the ith row of the first k unstandardized principal components, denoted by B, of Σ
and K = (K1 , · · · , Kp )T ∼ N (0, A) with A = Σ − BB T . The common factor W drives the
dependence among the test statistics. This realized but unobserved factor is critical for False
Discovery Proportion (FDP) estimation and power improvements by removing the common factor
{bTi W } from the test statistics. Hence, an important goal is to estimate W with given {bi }ni=1 .
In many applications on large-scale hypothesis testing, the parameters {µi }pi=1 are sparse. For
example, genome-wide association studies show that the expression level of gene CCT8 is highly
related to the phenotype of Down Syndrome. It is of interest to test the association between each
of millions of SNP’s and the CCT8 gene expression level. In the framework of Fan et al. (2012b),
each µi stands for such an association. That is, if µi = 0, the ith SNP is not associated with the
CCT8 gene expression level; otherwise, it is associated. Since most of the SNP’s are not associated
the CCT8 gene expression level, it is reasonable to assume {µi }pi=1 are sparse. Replacing Zi , µi , bi ,
W , k, p, and Ki with Yi , µ?i , X i , β ? , d, n, and i respectively, we obtain model (1.1) formally. It
is of interest to study this model independently with simplifications.
2
Although model (1.1) emerges from a critical component of estimating FDP in Fan et al.
(2012b), it stands with its own interest. For example, in some applications, there are only few
signals (nonzero µ?i ’s) and what is interesting is to learn about β ? , which reflects the relationship
between the covariates and response. For another example, those few nonzero µ?i ’s might be some
measurement or recording errors of the responses {Yi }. In these cases, model (1.1) is suitable for
modeling data with contaminated responses and a method producing a reliable estimator for β ? is
essentially a robust replacement for ordinary least squares estimate, which is sensitive to outliers.
Several models with structural and incidental parameters have first been studied in a seminal
paper by Neyman and Scott (1948), which points out the inconsistency of the maximum likelihood
estimators (MLE) of structural parameters in the presence of a large number of incidental parameters and provides a modified MLE. However, their method does not work for model (1.1) due to
no exploration of the sparsity of incidental parameters. Kiefer and Wolfowitz (1956) shows the
consistency of the MLE of the structural parameters when the incidental parameters are assumed
to be from a common distribution. That is, they eliminate the essential high-dimensional issue
of the incidental parameters by randomizing them. In contrast, this paper considers deterministic incidental parameters and handles the high-dimensional issue by penalization with a sparsity
assumption. Basu (1977) considers the elimination of nuisance parameters via marginalizing and
conditioning methods and Moreira (2009) solves the incidental parameter problem with an invariance principle. For a review of the incidental parameter problems in statistics and economics, see
Lancaster (2000).
Without loss of generality, suppose the first s incidental parameters {µ?i }si=1 are nonvanishing
and the remaining are zero. Then, model (1.1) can be written in a matrix form as Y = Xν + ,
where
X=
Is
X T1,s
0
0
X Ts+1,n
I n−s
,
X Ti,j = (X i , X i+1 , · · · , X j )T , I k is a k × k identity matrix, 0 is a generic block of zeros and
ν = (µ?1 , · · · , µ?s , β T , µ?s+1 , · · · , µ?n )T . Although this is a sparse high-dimensional problem, the
matrix X does not satisfy the sufficient conditions of the theoretical results in Zhao and Yu (2006)
and Fan and Lv (2011) due to the inconsistency of the estimation of the incidental parameters in
ν and the penalty should not be simply placed on all parameters. For details, see Supplement C.
3
In this paper, we investigate a penalized estimator of (µ? , β ? ) defined by
(µ̂, β̂) = argmin
n
X
(Yi − µi −
(µ,β)∈Rn+d i=1
X Ti β)2
+
n
X
pλ (|µi |),
(1.2)
i=1
where pλ is a penalty function with a regularization parameter λ. Since only the incidental parameters are sparse, the penalty is imposed on them. An iterative algorithm is proposed to numerically
compute the estimators. The estimator β̂ possesses consistency, asymptotic normality and an oracle property. On the other hand, the nonvanishing elements of µ? cannot be consistently estimated
even if β ? were known. So, there is a partial consistency phenomenon.
Penalized estimation (1.2) is a one-step method. For the estimation of β ? , We also propose a twostep method whose first step is designed to eliminate the influence of the data with large incidental
parameters. The estimator β̃ from the two-step method has fewer possible asymptotic distributions
than β̂ and thus is more suitable for constructing confidence regions for β ? . It is asymptotically
equivalent to the one-step estimator β̂ when the sizes of the nonzero incidental parameters are
small enough, that is, when the incidental parameters are really sparse. Also, the two-step method
improves the convergence rate and efficiency over the one-step method for challenging situations
where large nonzero incidental parameters increase the asymptotic covariance or even reduce the
convergence rate for the one-step method.
The rest of the paper is organized as follows. In Section 2, the model and penalized estimation
method are formally introduced and the corresponding penalized estimators are characterized. In
Section 3, asymptotic properties of the penalized estimators are derived; a penalized two-step
estimator is proposed and its theoretical properties are obtained; we also provide a data-driven
approach for selecting the regularization parameter. In Section 4, we consider the case where
the number of covariates grows with but slower than the sample size. In Section 5, we present
simulation results and analyze a read data set. Section 6 concludes this paper with a discussion
and all the proofs and some theoretical results are relegated to the appendix and supplements.
2
Model and Method
The matrix form of model (1.1) is given by
Y = µ? + Xβ ? + ,
(2.1)
where Y = (Y1 , Y2 , · · · , Yn )T , X = (X 1 , X 2 , · · · , X n )T , and = (1 , 2 , · · · , n )T . The covariates
{X i }ni=1 are independent and identically distributed (i.i.d.) copies of X 0 ∈ Rd , which is a random
4
vector with mean zero and a covariance matrix ΣX > 0. They are independent of the random
errors {i }, which are i.i.d. copies of 0 , which is a random variable with mean zero and variance
σ 2 > 0. Denote an bn and an bn if an = o(bn ) and bn = o(an ), respectively. There is an
assumption on the covariates and random errors.
Assumption (A): There exist positive sequences κn
√
n, γn
√
n such that
P ( max kX i k2 > κn ) → 0 and P ( max |i | > γn ) → 0, as n → ∞,
1≤i≤n
1≤i≤n
(2.2)
where k · k2 stands for the l2 norm of Rd .
Suppose there are three types of incidental parameters in model (1.1) or (2.1): for simplicity on
1
the indexes, the first s1 incidental parameters {µ?i }si=1
are large in the sense that |µ?i | max{κn , γn }
for 1 ≤ i ≤ s1 ; the next s2 ones {µ?i }si=s1 +1 are nonzero and bounded by γn with s = s1 + s2 ; the
last n − s ones {µ?i }ni=s+1 are zero. Note that it is unknown to us which µ?i ’s are large, bounded or
zero. The sparsity of µ? is understood by s1 + s2 n, i.e. s1 + s2 = o(n). Denote the vectors of
the three types of incidental parameters µ?1 , µ?2 , and µ?3 , respectively.
The penalized estimation (1.2) can be written as
(µ̂, β̂) = argmin L(µ, β),
L(µ, β) = kY − µ −
Xβk22
+
(µ,β)
n
X
pλ (|µi |).
(2.3)
i=1
The penalty function pλ can be the soft (i.e. L1 or LASSO), hard, SCAD or a general folded
concave penalty function (Fan and Li, 2001). For simplicity, we next consider only the soft penalty
function, that is, pλ (|µi |) = 2λ|µi |. The cases with the hard and SCAD penalties can be considered
in a similar way.
By Lemma D.1 in Supplement D, A necessary and sufficient condition for (µ̂, β̂) to be a minimizer of L(µ, β) is that β̂ = (X T X)−1 X T (Y − µ̂), Yi − µ̂i − X Ti β̂ = λSign(µ̂i ) for i ∈ Iˆ0c and
|Yi − X Ti β̂| ≤ λ for i ∈ Iˆ0 , where Sign(·) is a sign function and Iˆ0 = {1 ≤ i ≤ n : µ̂i = 0}.
Numerically, the special structure of L(µ, β) suggests a marginal decent algorithm for the minimization problem in (2.3), which iteratively computes µ(k) = argmin L(µ, β (k−1) ) and β (k) =
µ∈Rn
argmin
L(µ(k) , β)
until convergence. The advantage of this algorithm is that there exist analytic
β∈Rd
solutions to the above two minimization problems. They are respectively the soft-threshold estimators with residuals {Yi − X Ti β (k−1) } and ordinary least-squares estimator with responses Y − µ(k) .
In this section and the next, the number of the covariates d is assumed to be a fixed integer. A
case where d diverging to infinity will be considered in Section 4. For the case where d is finite, we
make the following assumption on λ.
5
Assumption (B): The regularization parameter λ satisfies
κn λ, αγn ≤ λ, and λ min{µ? ,
√
n},
(2.4)
where κn and γn are defined in (2.2), α is a constant greater than 2, and µ? = min1≤i≤s1 |µ?i |.
For simplicity, abbreviate “with probability going to one” to “wpg1”. A stopping rule for the above
algorithm is based on the successive difference kβ (k+1) − β (k) k2 . By Proposition D.2 in Supplement
D, wpg1, the iterative algorithm stops at the the second iteration, given the initial estimator is
bounded wpg1.
Suppose {β (k) } has a theoretical limit β (∞) , corresponding to which, there is a limit estimator
µ(∞) . Then, (µ(∞) , β (∞) ) is a solution of the following system of nonlinear equations
β = (X T X)−1 X T (Y − µ),
(2.5)
and, with soft-threshold applied to each component, it follows
µ = (|Y − X T β| − λ)+ Sign(Y − X T β),
(2.6)
where (·)+ returns the maximum value of the input and zero. By Lemma D.3 in Supplement D, a
necessary and sufficient condition for (µ̂, β̂) to be a minimizer of L(µ, β) is that it is a solution to
equations (2.5) and (2.6). Hence, (µ(∞) , β (∞) ) is a minimizer of L(µ, β) and can also be denoted
as (µ̂, β̂).
Note that β̂ is also the minimizer of the profiled loss function L̃(β) = L(µ(β), β), where
µ(β) as a function of β is given by (2.6) . Interestingly, this profiled loss function is a criterion
function equipped with the famous Huber loss function (see Huber (1964) and Huber (1973)).
Pn
T
Specifically, the profiled loss function can be expressed as L̃(β) =
i=1 ρ(Yi − X i β), where
ρ(x) = x2 I(|x| ≤ λ) + (2λx − λ2 )I(|x| > λ) is exactly the Huber loss function, which is optimal in a
minimax sense. This equivalence between the penalized estimation and Huber’s robust estimation
indicates that the penalization principle is versatile and can naturally produce an important loss
function in robust statistics. This equivalence also provides a formal endorsement of the least
absolute deviation robust regression (LAD) in Fan et al. (2012b) and indicates that it is better
to use all data with LAD regression rather than 90% of them. It is worthwhile to note that the
penalized estimation is only formally equal to the Huber’s. Our model (2.1) considers deterministic
sparse incidental parameters µ?i ’s, while the model in Huber’s works assumes random contamination
as in Kiefer and Wolfowitz (1956). Recently, there appear a few papers on robust regression in
6
high-dimensional settings, see, for example, Chen et al. (2010), Lambert-Lacroix and Zwald (2011),
? and Bean et al. (2012). Portnoy and He (2000) provide a high level review of literature on robust
statistics.
From the equations (2.5) and (2.6), β̂ is a solution to
ϕn (β) = 0, where ϕn (β) = β − (X T X)−1 X T (Y − µ(β)).
(2.7)
In general, this is a Z-estimation problem. The following theoretical analysis is based on this
characterization of β̂.
At the end of this section, we provide for further analysis some notations and an expansion of
P
P
P
P
P
ϕn (β). Let S = ni=1 X i X Ti , SS = i∈S X i X Ti , SµS = i∈S X i µ?i , SS = i∈S X i i , S = ni=1 X i
P
and SS = i∈S X i , where S is a subset of {1, 2, · · · , n}. It is straightforward to show
ϕn (β) = (SS10 + SS11 + SS12 )(β − β ? ) − (SµS11 + SµS12 )
− (SS10 + SS11 + SS12 ) − λ(SS20 + SS21 + SS22 − SS30 − SS31 − SS32 ),
(2.8)
where the index sets S10 = {s + 1 ≤ i ≤ n : |X Ti (β ? − β) + i | ≤ λ}, S11 = {1 ≤ i ≤ s1 :
|µ?i + X Ti (β ? − β) + i | ≤ λ} and S12 = {s1 + 1 ≤ i ≤ s : |µ?i + X Ti (β ? − β) + i | ≤ λ}; S20 , S21
and S22 are defined similarly except that the absolute operation is omitted and “≤” is replaced by
“>”; S30 , S31 and S32 , are defined similarly with S20 , S21 and S22 except that “> λ” is replaced
by “< −λ”. Note that all these index sets depend on β.
3
Asymptotic Properties
In this section, we consider the asymptotic properties of the penalized estimators β̂ and µ̂. Assumption (A), together with Assumption (B), enables the penalized estimation method to distinguish
the large incidental parameters from others, and thus simplifies the asymptotic properties of the
index sets Sij ’s in (2.8) in the sense that they become independent of β wpg1. Denote a hypercube
of β ? by BC (β ? ) = {β ∈ Rd : |βj − βj? | ≤ C, 1 ≤ j ≤ d} with a constant C > 0.
Lemma 3.1 (On Index Sets Sij ’s). Under Assumptions (A) and (B), for every C > 0 and every
? , S
?
?
?
β ∈ BC (β ? ), wpg1, S10 = S10
11 = ∅, S12 = S12 , S20 = ∅, S21 = S21 , S22 = ∅, S30 = ∅, S31 = S31
? = {s + 1, s + 2, · · · , n}, S ? = {s + 1, s + 2, · · · , s},
and S32 = ∅, where the limit index sets S10
1
12
? = {1 ≤ i ≤ s : µ? > 0} and S ? = {1 ≤ i ≤ s : µ? < 0}.
S21
1
1
31
i
i
7
By Lemma 3.1, wpg1, the solution β̂ to (2.7) has an analytic expression:
−1 µ
? + SS ? )
β̂ = β ? + (SS10
[SS ? + (SS10
? + SS ? ) + λ(SS ? − SS ? )],
12
21
31
12
12
(3.1)
from which, we derive asymptotic properties of β̂. Some analysis needs the following assumption.
Assumption (C): There exists some constant δ > 0 such that EkX 0 k2+δ
< ∞ and kµ?2 k2 /kµ?2 k2+δ
2
Ps
? 2+δ 1/(2+δ) .
diverges to infinity, where kµ?2 k2+δ =
i=s1 +1 |µi |
The following result shows the existence of a unique consistent estimator of β ? .
Theorem 3.2 (Existence and Consistency of β̂). Under Assumptions (A) and (B), if either s2 =
o(n/(κn γn )) or Assumption (C) holds, then, for every fixed C > 0, wpg1, there exists a unique
P
estimator β̂ n ∈ BC (β ? ) such that ψn (β̂ n ) = 0 and β̂ n −→ β ? .
In Theorem 3.2, there are two different kinds of sufficient conditions: on is on s2 , which is the
size of bounded incidental parameters µ?2 , and the other is Assumption (C), which is about the
norms of µ?2 . They come from different analysis approaches on the term SµS ? in (3.1). One does
12
not imply the other. For details, see Supplement E. Specially, if s2 =
O(nα2 )
for some α2 ∈ (0, 1)
and κn γn n(1−α2 ) , then β̂ is consistent by Theorem 3.2.
Next, we consider the asymptotic distributions of the consistent estimator β̂ n obtained in The? = {1 ≤ i ≤ s : µ? > 0}
orem 3.2. Without loss of generality, we assume the sizes of index sets S21
1
i
? = {1 ≤ i ≤ s : µ? < 0} are asymptotically equivalent to as and (1 − a)s with a constant
and S31
1
1
1
i
a ∈ (0, 1). Similar to Theorem 3.2, there are two different sets of conditions on µ?2 corresponding
to two different analysis approaches. Denote ∼ as the asymptotic equivalence and Dn = kµ?2 k2 .
Theorem 3.3 (Asymptotic Distributions on β̂ n ). Under Assumptions (A) and (B), suppose s2
√
n/(κn γn ) holds or Assumption (C) and Dn2 /n = o(1) hold.
√
d
n(β̂ n − β ? ) −→ N (0, σ 2 Σ−1
X ); [main case]
√
d
+
(2) If s1 ∼ bn/λ2 , then n(β̂ n − β ? ) −→ N (0, (b + σ 2 )Σ−1
X ), for every constant b ∈ R ;
√
d
(3) If s1 n/λ2 , then rn (β̂ n − β ? ) −→ N (0, Σ−1
X ), where rn ∼ n/(λ s1 ).
(1) If s1 n/λ2 , then
When the incidental parameters are really sparse, the size s1 of large incidental parameters is
small and the size s2 or the magnitude Dn of bounded incidental parameters is also small so that the
conditions of case (1) tends to hold. This case is of most interest and we denote it as the main case.
The other cases are presented to provide a relatively complete picture of the asymptotic distributions
8
of β̂. In fact, Theorem E.1 in Supplement E shows more possible asymptotic distributions. Note
that the constant a does not appear in the limit distributions of Theorem 3.3 due to cancelation and
√
that the sub- n consistency emerges in case (3) when s1 is large, because for this case the impact
of the large incidental parameters is too big to be handled efficiently by the penalized estimation.
For case (2), in one direction, as b → 0, its condition and limit distribution become those of case
(1); in the other direction, as b increases, it approaches case (3). This boundary phenomenon was
in spirit similar to that in Tang et al. (2012). Specially, if λ nα1 , κn γn nα2 , s1 n1−α1 and
√
d
s2 n1/2−α2 , for some α1 ∈ (0, 1) and α2 ∈ (0, 1/2), then n(β̂ n − β ? ) −→ N (0, σ 2 Σ−1
X ) by the
main case of Theorem 3.3.
Remark 1 (An Oracle Property). Suppose an oracle tells the true µ? . Then, with the adjusted
(O)
responses Y −µ? , the oracle estimator of β ? is given by β̂
= (XX T )−1 X T (Y −µ? ). The limiting
√
(O)
distribution of n(β̂ n − β ? ) is N (0, σ 2 Σ−1
X ). Comparing this with the main case of Theorems 3.3,
it follows that the penalized estimator β̂ n enjoys an oracle property.
Although mainly interested in the estimation of β ? , we also obtain the soft-threshold estimator
µ̂ of µ? : for each i,
µ̂i = µi (β̂) = (|Yi − X Ti β̂| − λ)+ sgn(Yi − X Ti β̂).
(3.2)
Denote E = {µ̂i 6= 0, for i = 1, 2, · · · , s1 ; and µ̂i = 0, for i = s1 + 1, s1 + 2, · · · , n}.
P
Theorem 3.4 (Partial Selection Consistency on µ̂). Under Assumptions (A) and (B), if β̂ −→ β ? ,
then P (E) → 1.
Theorem 3.4 shows that, wpg1, the indexes of µ?1 and µ?3 are estimated correctly, but those of
µ?2 wrongly. We call this a partial selection consistency phenomenon.
3.1
Two-Step Estimation
Theorems 3.3 shows that the penalized estimator β̂ n has multiple different limit distributions,
which complicates the application of these theorems in practice. In addition, the convergence rate
√
of β̂ n is less than the optimal rate n in the challenging cases where the impact of large incidental
parameters is substantial. To address these issues, we propose the following two-step estimation
method: firstly, we apply the penalized estimation (2.3) and let Iˆ0 = {1 ≤ i ≤ n : µ̂i = 0}; secondly,
we define the two-step estimator as
β̃ = (X TIˆ X Iˆ0 )−1 X TIˆ Y Iˆ0 ,
0
0
9
(3.3)
where X Iˆ0 consists of X i ’s whose indexes are in Iˆ0 and Y Iˆ0 consists of the corresponding Yi ’s. The
following theorem shows that β̃ is consistent and Asymptotic Gaussian.
Theorem 3.5 (Consistency and Asymptotic Normality on β̃). Suppose Assumptions (A) and
P
(B) hold. If either s2 = o(n/(κn γn )) or Assumption (C) holds, then β̃ −→ β ? . If either s2 =
√
√
d
o( n/(κn γn )) holds or Assumptions (C) and Dn2 /n = o(1) hold, then n(β̃−β ? ) −→ N (0, σ 2 Σ−1
X ).
Comparing Theorem 3.5 with Theorem 3.3, we see that β̂ has three possible asymptotic distributions but β̃ has only one since for β̃ the conditions on s1 disappear. It is because the two-step
method identifies and removes large incidental parameters by exploiting the partial selection consistency property of µ̂ in Theorem 3.4. Further, the two-step estimator improves the convergence rate
to the optimal one over the one-step estimator for the challenging case with s1 n/λ2 . Because
of these advantages, we suggest to use the two-step method to make statistical inferences.
When the incidental parameters are sparse in the sense that the size or the magnitude of the
√
bounded incidental parameters are small, i.e. s2 = o( n/(κn γn )) or Dn2 /n = o(1), it follows,
√
d
by Theorem 3.5, n(β̃ − β ? ) −→ N (0, σ 2 Σ−1
X ), from which a confidence region with asymptotic
confidence level 1 − α is given by
√
1/2
{β ∈ Rd : σ −1 nkΣX (β̃ − β)k2 ≤ qα (χd )},
(3.4)
where qα (χd ) is the upper α-quantile of χd , the square root of the chi-squared distribution with
degrees of freedom d. For each component βj? of β ? , an asymptotic 1 − α confidence interval is
given by
−1/2
[β̃j ± n−1/2 σΣX
−1/2
where ΣX
(j, j)zα/2 ],
(3.5)
(j, j) is the square root of the (j, j) entry of Σ−1
X and zα/2 is the upper α/2-quantile
of N (0, 1). The confidence region (3.4) and interval (3.5) involve unknown parameters ΣX and σ.
They can be estimated by Σ̂X = (1/n)X T X and
σ̂ = #(Iˆ0 )−1/2 kY Iˆ0 − X TIˆ β̃k2 .
0
(3.6)
By Law of Large Numbers, Σ̂X is consistent. By lemma E.3 in Supplement, σ̂ is also consistent.
Hence, after replacing ΣX and σ in the confidence region (3.4) and interval (3.5) with Σ̂X and σ̂,
the resulting confidence region and interval have the asymptotic confidence level 1 − α.
10
3.2
Theoretical and Data-Driven Regularization Parameters
By Assumption (B), the theoretical regularization parameter λ depends on κn and γn , which are also
crucial to the boundary conditions of the asymptotic properties of the penalized estimators β̂ and
β̃. By Assumption (A), κn and γn is determined by the distributions of X 0 and 0 , respectively. It
is of interest to explicitly derive κn and γn for some typical cases on the covariates and errors. When
√
the covariates are bounded with CX > 0 and the random errors follow N (0, σ 2 ), let κn = dCX
p
and γn = 2σ 2 log(n). They satisfy (2.2) in Assumption (A) and the specification of λ (2.4) in
p
√
Assumption (B) becomes α 2σ 2 log(n) ≤ λ min{µ? , n}. When X 0 and 0 follow N (0, ΣX )
2 the maximum of diagonal elements of Σ . We can take
and N (0, σ 2 ), respectively. Denote by σX
X
q
p
p
√
2 log(n) and γ =
2σ 2 log(n). Then, (2.4) becomes log(n) λ min{µ? , n}. A
κn = 2dσX
n
case on exponentially tailed random variables has all been considered in Supplement E.2.
Although the theoretical specification of λ guaranties desired asymptotic properties, a datadriven specification is of interest in practice. A popular way to specify λ is to use multi-fold
cross-validation. The validation set, however, needs to be made as little contaminated as possible.
We propose the following procedure to identify a data-driven regularization parameter:
Step 1: On the training and testing data sets.
(OLS)
1. Apply ordinary least squares (OLS) to all the data and obtain residuals ˆi
X Ti β̂
(OLS)
= Yi −
for each i.
(OLS)
2. Identify the set of “pure” data corresponding to the npure smallest values in {|ˆ
i
3. Compute the updated OLS estimator β̂
residuals
(OLS,2)
{ˆ
i
}
(OLS,2)
|}.
with the “pure” data and obtain updated
for each i.
(OLS,2)
4. Identify the updated “pure” data set corresponding to the npure smallest {|ˆ
i
|} and
label the remaining as the “contaminated” data set.
5. Randomly select a subset from the updated “pure” data set as a testing set and merge
the remaining “pure” data set and the “contaminated” one into a training set.
Step 2: On the range [λL , λU ] of the regularization parameter.
1. Compute the standard deviation σ̂pure of the residuals of the “pure” data set.
2. Set λL = αl σ̂pure and λU = αu σ̂pure , where αl < αu are positive constants.
Step 3: On the data-driven regularization parameter.
11
1. For each grid point of λ in the interval [λL , λU ], apply a penalized method to the train2
ing set and obtain the estimator β̂ λ,train and the corresponding test error σ̂λ,test
=
P
T
2
testing set (Yi − X i β̂ λ,train ) .
2
2. Identify the data-driven regularization parameter λopt , which minimizes σ̂λ,test
among
the grid points.
This simple data-driven procedure can certainly be further improved. For example, In Step 1, the
sub-steps 3 and 4 can be repeated more times to obtain a better “pure” data set. In Step two, the
(OLS,2)
range for λ can also be obtained from quantiles of {|ˆ
i
(OLS,2)
on σ̂pure and quantiles of {|ˆ
i
|}. We can also hybrid quantities based
|} to determine [λL , λU ].
The good performance of this data-driven regularization parameter will be demonstrated in
Subsection 5.2.
4
Diverging number of structural parameters
In Sections 2 and 3, we have considered model (2.1) under the assumption that the number of
covariates d is a fixed integer. However, when there are a moderate or large number of covariates,
it is appropriate to assume that d diverges to infinity with the sample size. In this section, we
consider model (2.1) with the assumption that d → ∞ and d n.
Since the number of covariates grows orderly slower than the sample size, we chose to continue
use the penalized estimation (2.3) for (µ? , β ? ) and the penalized two-step estimation (3.3) for β ? .
The corresponding estimators are still denoted as (µ̂, β̂) and β̃, but we should keep it in mind that
their dimensions diverge to infinity with n. The characterizations of β̂ in Lemmas D.1 and D.3
are still valid since they are finite-sample results. The iteration algorithm also wpg1 stops at the
second iteration, which is shown by Proposition F.3 in Supplement F.
As before, it is critical to properly specify the regularization parameter λ for the case with a
diverging number of covariates.
Assumption (B’) The regularization parameter λ satisfies
√
dκn λ, αγn ≤ λ, and λ µ? ,
where κn and γn are defined in (2.2) and α > 2.
12
(4.1)
Comparing Assumption (B’) with Assumption (B), the main difference in formation is that κn is
√
changed to dκn . In fact, κn in (4.1) also depends on d, which will be shown in Supplement E.2.
This difference is caused by the assumption that d diverges to ∞.
Lemma 4.1 (On Index Sets Sij ’s). Under Assumptions (A) and (B’), the conclusion of Lemma
3.1 holds.
Thus, wpg1, still valid is the crucial analytic expression of β̂ (3.1), from which we derive its
theoretical properties. They are similar to those in the previous section, with additional technical
complexity caused by the diverging dimension d.
Denote k·kF,d = d−1/2 k·kF , where k·kF is the Frobenius norm, and κX = d−1
Pd
4 1/2 ,
j=1 (E[X0j ])
the average of the square root of the fourth marginal moments of X 0 . We make the following
assumptions on ΣX and κX .
Assumption (D): kΣ−1
X kF,d is bounded.
Assumption (E): κX is bounded.
Theorem 4.2 (Existence and Consistency on β̂). Suppose Assumptions (A), (B’), (D) and (E)
hold. If there exists rd , a sequence of positive numbers depending on d, such that d3 /n → 0,
√
√
(rd d)2 /n → 0, s1 = o(n/(rd dκn λ)) and s2 = o(n/(rd dκn γn )), then, for every fixed C > 0, wpg1,
P
there exists a unique estimator β̂ ∈ BC (β ? ) such that ψn (β̂) = 0 and rd kβ̂ − β ? k2 −→ 0.
Next, we consider the asymptotic distribution on β̂. Since the dimension of β̂ diverges to infinity,
following Fan and Lv (2011), it is more appropriate to study its linear maps. Let An be a q × d
matrix, where q is a fixed integer, Gn = An ATn with the largest eigenvalue λmax (Gn ), and GX,n =
T
2
An Σ−1
X An . Denote by λmin (ΣX ) the smallest eigenvalue of ΣX , σX,max = max1≤j≤d Var[X0j ],
2
σX,min
= min1≤j≤d Var[X0j ] and γX,max = max1≤j≤d E|X0j |3 . Abbreviate “with respect to” by
“wrt”. We assume further
Assumption (D’): λmin (ΣX ) is bounded away from zero, which implies Assumption (D).
Assumption (D”): kΣX kF,d is bounded.
Assumption (F): kAn kF and λmax (Gn ) are bounded and GX,n converges to a q × q symmetric
matrix GX wrt k·kF .
Assumption (G): σX,max > 0; σX,max and γX,min are bounded from above and σX,min is bounded
away from zero.
13
Similar to the main case of Theorem 3.3, a properly scaled β̂ n is asymptotically Gaussian.
Theorem 4.3 (Asymptotic Distribution on β̂). Suppose Assumptions (A), (B’), (D’), (D”), (E),
√
√
√ √
(F) and (G) hold. If d5 log d = o(n), s1 = o( n/(λ dκn )) and s2 = o( n/( dκn γn )), then
√
d
nAn (β̂ − β ? ) −→ N (0, σ 2 GX ).
The penalized estimator µ̂ obtained by (3.2) is partially consistent.
Theorem 4.4 (Partial Selection Consistency on µ̂). Suppose Assumptions (A) and (B’)hold and
√
β̂ is a consistent estimator of β ? wrt rd k·k2 . If rd ≥ 1/ d, then P (E) → 1.
We can construct the penalized two-step estimator β̃ through (3.3) with µ̂. This two-step
estimator is consistent by Theorem F.4 in Supplement F and its asymptotic distribution, as an
extension of the main case in Theorem 3.5, is given by the following theorem.
Theorem 4.5 (Asymptotic Distribution on β̃). Suppose all the assumptions and conditions of The√
d
orem 4.3 hold except that the condition on s1 is not required. Then nAn (β̃−β ? ) −→ N (0, σ 2 GX ).
From Theorems 4.3 and 4.5, Wald-type asymptotic confidence regions of β ? are availabe. For
example, a confidence region based on β̃ with asymptotic confidence level 1 − α is given by
√
−1/2
{β ∈ Rd : σ −1 nkGX,n An (β̃ − β)k2 ≤ qα (χq )}.
(4.2)
−1
Since GX,n involves the unknown ΣX , we estimate it by ĜX,n = An Σ̂X ATn . On the other hand,
σ is estimated by σ̂ in (3.6) as before. After plugging ĜX,n and σ̂ into (4.2), we obtain
√
−1/2
{β ∈ Rd : σ̂ −1 nkĜX,n An (β̃ − β)k2 ≤ qα (χq )}.
(4.3)
By Lemma A.5 in the appendix, the consistency of σ̂ is assured. Then, Theorem A.6 in the appendix
guarantees the asymptotic validity of the confidence region (4.3).
5
Numerical Evaluations and Real Data Analysis
In this section, we evaluate the finite-sample performance of the penalized estimation through
simulations and use it to analyze a real data set. The model for simulations is given by, for
? + . For simplicity, the deterministic sparse
i = 1, · · · , n, Yi = µ?i + Xi,1 β1? + · · · + Xi,50 β50
i
incidental parameters {µ?i } are generated as i.i.d. copies of µ: µ is 0, U [−c, c] and W (c + Exp(τ ))
with probabilities p0 , p1 , and p2 respectively, where U [−c, c] is a uniform random variable in [−c, c],
14
W takes values 1 and −1 with probabilities pw and 1 − pw , respectively, and Exp(τ ) follows an
exponential distribution with mean 1/τ > 0. Note that c can be viewed as a contamination
parameter. The larger c is, the more contaminated the data. On the other hand, pw determines
? = 1;
the asymmetry of the incidental parameters. The regression coefficients β1? = · · · = β50
i.i.d.
{(Xi,1 , · · · , Xi,50 )} ∼ N (0, ΣX ), where ΣX (i, j) = 2 exp(−|i − j|)), which is a Toeplitz matrix
and the constant 2 is used to inflate the covariance a little; the covariates are independent of
i.i.d.
{i } ∼ N (0, 1); and n = 500; p0 = 0.8, p1 = 0.1, p2 = 0.1, c is 0.5, 1, 3 or 5, pw is 0.5 or 0.75 and
τ = 1.
5.1
Performance of Penalized Methods
The following methods for estimating β ? are evaluated. (i) Oracle method (O): an oracle knows the
index set S of zero µ?i ’s. Its performance is used as a benchmark. (ii) Ordinary least squares method
(OLS): all µ?i ’s are thought as zeros. (iii) Four penalized least squares (PLS) methods, namely, PLS
with soft penalty (PLS.Soft or S), PLS with hard penalty (PLS.Hard or H), two-step PLS with soft
penalty (PLS.Soft.TwoStep or S.TS) and two-step PLS with hard penalty (PLS.Hard.TwoStep or
P
P
(O)
H.TS). More specifically, the oracle estimator of β ? is given by β̂
= ( i∈S X i X Ti )−1 i∈S X i Yi .
The hard penalty function is pλ (|t|) = λ2 −(|t|−λ)2 {|t| < λ} (see Fan and Li (2001)). Each method is
evaluated by the square root of the empirical mean squared error (RMSE). Every penalized method
is evaluated with a grid of values for the regularization parameter λ, ranging from .5 to 5 by .25.
The sequence plot of Figure 1 shows 500 realized incidental parameters µ?i ’s with c = 3 and
pw = 0.75. They are used in data generation of simulations. The scatter plot of Figure 1 shows
the responses Yi ’s against the first covariate Xi1 ’s of a generated data set and the red stars stand
for the contaminated sample points, the ones with nonzero µ?i ’s. With fifty covariates, usually it is
difficult to graphically identify the contaminated data points.
With the above incidental parameters, those six methods are evaluated by simulations with
iteration number 1000. Since ΣX is a Toeplitz matrix with equal diagonal elements, the asymptotic
variances of the estimators of β1? and β2? are different and representative for estimators of other
βi? ’s. So, we only report simulation results on the estimation of β1? and β2? .
Figure 2 shows RMSE’s of six estimators for β1? . RMSE’s for β2? are similar. As expected,
the oracle method has the smallest RMSE and OLS the largest. RMSE of each PLS method as a
function of λ forms a convex curve, which achieves a minimal RMSE significantly below the green
line of OLS and close to the cyan line of O. More specifically, RMSE of PLS.Hard achieves the
15
Marginal Data Plot
40
Sequence Plot of Incidental Parameters
●
●
●
●
●
●
●
● ●
●
●
●
● ● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
0
●
●
−40
−5
●
●
●
●
●
●
●
●
0
●
−20
0
mu
●
●
●
100
200
300
●
●
● ● ●
●
●
●
●
●
●
●
●●
●
●
●●●●● ● ●
●
●
●
●
●
●
●●
●●
●
●
●
●
● ●●
●
●
●●
●
●
●
●●●
● ●
●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
● ● ●
●
●
● ●
●
●
●
● ●●
● ●●
● ●
● ●●● ●●
●●
●
●
●
● ●
●
●
●
●
●
●
●
● ●
●
● ●
●●
● ●●●
●
● ●● ● ● ● ●● ●●
●● ●●
●● ●
●
●
●● ● ● ● ●●●
●
●
● ● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●● ●
●●
●
● ●
●
●
● ●
●
●
●
●●
● ●
● ●● ● ●
●
●
● ●
●
● ●
●
● ●
● ●● ●●
● ●
● ● ●●●
●●
●
●
●● ●●●● ●● ●● ●
● ● ●● ● ● ● ●
● ●●
●●
●● ●
●
●●
●
●
●●
●● ●
● ●●●●●
●●
●
● ●
●
●●
●
●
●●●
●
●
●
● ● ●●
●
● ● ● ● ●●●
●
●
●●
●
● ● ●
●
●
●
●
● ●● ●
● ●
●
●
●
●
●● ● ●
●
●
●
●
●
● ●
●
●
●
● ●
●
● ● ●●●
●
●●
●
●
●
● ●●
●
●
● ●
●
●
●●● ●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
Y
5
●
●
●
20
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
400
500
−4
−2
Index
0
2
4
X1
Figure 1: The sequence plot on the left shows 500 incidental parameters µ?i ’s with c = 3 and pw = 0.75. The
non-zero µ?i ’s are in red. The scatter plot on the right shows the responses Yi ’s against the first covariate
Xi1 ’s of a data set generated with those 500 incidental parameters. The red stars stand for the contaminated
sample points, the ones with nonzero µ?i ’s.
minimal RMSE when λ is around 2.75. On the other hand, RMSE of PLS.Soft decreases a little
till λ is around .75, then increases and stays above RMSE of PLS.Hard. This reflects the fact
that a large λ in a soft-threshold method usually causes bias. PLS.Hard.TwoStep has very similar
performance with PLS.Hard for all λ. PLS.Soft.TwoStep has similar performance with PLS.Soft
when λ is small. However, as λ becomes large, PLS.Soft.TwoStep moves closer to PLS.Hard than
PLS.Soft. It is because PLS.Soft.TwoStep and PLS.Hard.TwoStep have similar estimation when λ
is large. The minimal RMSE of PLS.Soft is slightly larger than those of other PLS Methods.
Table 1 depicts the minimal RMSE’s of the estimators for β1? and β2? with the corresponding optimal λ’s and biases. The biases are ignorable comparing withe the RMSE’s. The optimal λ for PLS.Soft and other PLS methods are around .75 and 2.5, respectively. This indicates
the simple soft threshold method tends to work best with a small λ due to the bias issue. Denote the empirical relative efficiency (ERE) of an estimator A with respect to another estimator
B as RMSE(B)/RMSE(A). Then, for the estimation of β1? , the ERE’s of PLS.Soft, PLS.Hard,
PLS.Soft.TwoStep and PLS.Hard.TwoStep with respect to O are around 78%, 87%, 87% and 87%,
respectively; the ERE’s of the PLS methods with respect to OLS are around 176%, 196%, 196%
and 196%, respectively. The ERE’s for β2? are similar. Thus, in terms of ERE (and RMSE), the
PLS methods perform closely to O and significantly better than OLS.
16
RMSE's for beta1 with Deterministic Sparse Inicidental Parameters
OLS
●
S
H
S.TS
H.TS
0.08
0.09
O
●
●
●
0.07
●
●
●
●
●
●
●
●
●
●
●
●
●
0.04
0.05
●
●
0.06
RMSE
●
1
2
3
4
5
lambda
Figure 2:
RMSE’s of the O(Oracle), OLS, S(PLS.Soft), H(PLS.Hard), S.TH(PLS.Soft.TwoStep) and
H.TH(PLS.Hard.TwoStep) estimators of β1? with the incidental parameters shown in Figure 1. The top and
bottom solid horizontal lines show the RMSE’s for OLS and O, respectively. Other four horizontal lines
indicate the minimal RMSE’s for those four PLS methods and the corresponding four best λ’s are shown by
the vertical lines.
From Table 1, we can also see that the RMSE’s of the estimators for β1? are always smaller than
those for β1? . This is because the first covariate is less correlated with others covariates than the
second one.
In order to examine the performance of the methods with general incidental parameters but
not just those in Figure 1, we generate µ? randomly for each iteration. The iteration number for
each simulation is also 1000.
Figure 3 shows the RMSE’s of six estimators of β1? under two settings: pw = 0.5 and c = 1 or
5. Each plot in Figure 3 presents a similar pattern with Figure 2. When pw is fixed at 0.5, the
RMSE’s of each non-oracle estimator of β1? increases as the contamination parameter c increases
from 1 to 5. This indicates that each non-oracle estimator performs worse as the data becomes more
contaminated. However, the PLS estimators are more robust than OLS, which is very sensitive to
the change of c. We have also done simulations with pw = 0.75 and the RMSE’s of the estimators
of β1? are similar to those with pw = 0.75 so that the corresponding plots are similar to those in
Figure 3. In other words, the RMSE’s of all estimators are stable with respect to pw , which means
17
O
OLS
S
H
S.TS
H.TS
S.P
H.P
LAD
Bias(β̂1 )×104
.8
17.8
1.3
.6
5.9
20.4
9.5
11.5
10.4
RMSE(β̂1 )×102
4.0
8.4
5.1
4.6
4.6
4.6
5.4
5.0
5.4
.75
2.75
2.25
2.5
2.45
2.47
λ
Bias(β̂2 )×104
-2.3
-46.7
24.6
43.5
-21.8
10.0
32.3
13.9
-7.3
RMSE(β̂2 )×102
4.4
9.0
5.3
4.9
5.0
4.9
5.8
5.5
5.9
.75
2.75
2.75
2.5
2.45
2.47
λ
Table 1: RMSE’s of the Oracle, OLS and LAD estimators and The minimal RMSE’s of the penalized
estimators of β1? and β2? with the corresponding optimal or data-driven λ’s and biases when the incidental
parameters shown in Figure 1 are used. For a data-driven method, the data-driven λ is different in each
iteration so that the reported λ’s are averages. The lines over the numbers emphasize the numbers are
averages. The standard deviations for the data-driven λ’s of S.P and H.P are 0.37 and 0.34, respectively.
RMSE's for beta1 with Random Sparse Incidental Parameters
c = 1 , p.w = 0.5
●
1
●
●
●
2
●
●
●
●
●
3
●
●
●
4
●
●
●
0.08
0.06
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.04
0.06
RMSE
0.04
●
●
●
●
●
RMSE
0.08
O
OLS
S
H
S.TS
H.TS
●
●
c = 5 , p.w = 0.5
5
1
lambda
2
3
4
5
lambda
Figure 3: Similar to Figure 2, these plots show the RMSE’s of the O(Oracle), OLS, S(PLS.Soft),
H(PLS.Hard), S.TH(PLS.Soft.TwoStep) and H.TH(PLS.Hard.TwoStep) estimators of β1? with randomly
generated µ? under two settings with pw = 0.5 and c = 1 or 5.
18
RMSE(βˆ1 )×102
O
OLS
S
H
S.TS
H.TS
S.P
H.P
LAD
(pw , c) = (.5, .5)
4.06
4.06
3.92
3.95
3.90
3.96
4.01
4.38
4.85
1.25
3.5
2.75
2.5
2.08
2.15
4.11
4.27
4.21
4.27
4.01
4.59
2
3.75
4.25
3.5
2.20
2.24
4.81
4.99
4.81
4.80
5.64
5.19
1
2.5
2
2.25
2.42
2.45
4.96
4.63
4.66
4.64
6.36
4.87
1
2.25
2.75
3
2.52
2.58
3.91
3.88
3.98
3.91
4.08
4.41
1.5
5
3.25
4.5
2.08
2.14
4.19
4.15
4.15
4.20
4.16
4.59
3.25
3
4
3.5
2.15
2.23
4.91
4.93
4.80
5.02
5.35
5.06
1
2.25
2
2.5
2.44
2.47
5.01
4.66
4.75
4.66
6.22
4.90
1.5
2.25
2.5
3
2.55
2.60
λ
(pw , c) = (.5, 1)
4.06
4.58
λ
(pw , c) = (.5, 3)
4.12
6.47
λ
(pw , c) = (.5, 5)
4.07
8.50
λ
(pw , c) = (.75, .5)
4.13
4.02
λ
(pw , c) = (.75, 1)
4.17
4.41
λ
(pw , c) = (.75, 3)
4.15
5.99
λ
(pw , c) = (.75, 5)
λ
3.97
8.41
4.75
5.49
5.52
4.79
4.98
5.52
5.76
Table 2: Similar to Table 1, this one shows the RMSE’s and minimal RMSE’s of nine estimators of β1? under
eight settings on randomly generated µ? with pw = 0.5, 0.75 and c = 0.5, 1, 3, 5. The standard deviations of
the data-driven λ’s of S.P and H.P for different settings are between 0.2 and 0.45.
the magnitudes of the nonzero incidental parameters matter most but not their signs. We also note
that some penalized methods perform closely to or even outperform the oracle one when c is small
as showed in the plot with c = 1. This happens because that O ignores all the contaminated data
points, even those with very light contamination, but the penalized methods exploit information
in such points.
Table 2 contains the RMSE’s of the estimators of β1? under eight settings with pw = 0.5 or
0.75 and c = 0.5, 1, 3 or 5. For each pw , as c increases from 0.5 to 5, the RMSE’s of O with the
multiplication factor 102 is almost constantly around 4, those of OLS increases from about 4 to 8.5
and those of PLS ones grow from about 4 to 5, which confirms the robustness of the PLS estimators.
When c ≤ 1 is small with respect to the variance of random error σ = 1, the data points are only
slightly contaminated. OLS and PLS methods perform similar to O. However, when c ≥ 3 is large,
19
which means the data are more contaminated, the RMSE’s of OLS become significantly larger, but
PLS methods perform still closely to O.
5.2
Performance of Data-Driven Penalized Methods
Previous simulations have shown the PLS methods with optimal λ’s have good RMSE’s comparing
with those of the Oracle and OLS ones. In practice, however, these optimal λ’s are unknown. One
approach to obtain a data driven λ has been introduced in Subsection 3.2. Since, as shown in the
previous simulation results, the two-step PLS methods perform similarly with the one-step PLS
methods, i.e. PLS.Soft and PLS.Hard, only the latter are studied by simulations with data-driven
λ’s and denoted as PLS.Soft.Prac (S.P) and PLS.Hard.Prac (H.P), respectively. For estimating
data-driven λ’s, αl = 2 and αu = 7. The size of the pure data set npure is n/2 and that of the
testing data set is npure /2.
Simulations are first run with the deterministic sparse incidental parameters as showed in
Figure 1. We can see in Table 1 that the RMSE’s of estimators of β1? from PLS.Soft.Prac and
PLS.Hard.Prac are around 5.4 and 5.0, slightly larger than the optimal values 5.1 and 4.6.,respectively. However, they are still significantly smaller than RMSE of OLS, which is 8.4. The
observations of the estimators of β2? are similar. As before, we also evaluate the performance of
the data-driven PLS methods with random sparse incidental parameters. Table 2 shows that, for
a given pw , when c is small such as 0.5 and 1, the RMSE’s of PLS.Soft.Prac and PLS.Hard.Prac
are close to those of PLS.Soft and PLS.Hard with the optimal λ’s. In these cases, PLS.Soft.Prac
performs slightly better than PLS.Hard.Prac, and even better than PLS.Soft with the optimal λ
and the Oracle method. On the other hand, for a given pw , when c is large such as 3 and 5, the
RMSE’s of PLS.Soft.Prac and PLS.Hard.Prac are greater than those of PLS.Soft and PLS.Hard,
respectively, but still less than those of OLS. In these cases, the RMSE’s of PLS.Soft.Prac are larger
than those of PLS.Hard.Prac, which indicates the bias issue of the soft threshold method. Thus,
the data-driven regularization parameter works well with penalized estimation. When the data is
slightly contaminated, the soft penalty is preferred; otherwise, the hard penalty is recommended.
Tables 1 and 2 also contain RMSE of the least absolute deviation regression method (LAD)
used in Fan et al. (2012b) with all but not part of the sample points with small residuals. Generally speaking, in both deterministic and random incidental parameter cases, LAD performs
similarly with the PLS methods with data-driven λ’s. More specifically, when c ≤ 1 is small,
PLS.Soft.Prac outperform LAD; otherwise, LAD performs better. For all the cases, LAD is domi20
nated by PLS.Hard.Prac. These observations confirm that LAD is an effective robustness method
and the penalized methods make improvement.
5.3
Data-Driven Confidence Intervals
We next turn to investigate the finite-sample performance of the asymptotic confidence interval (CI)
(3.5) for βj? with j = 1, 2 based on PLS two-step methods. Since CI (3.5) is based on the properties of
the penalized two-step estimator with the soft penalty, we focus on PLS.TS.Soft with a data-driven
regularization parameter λ. The choice of λ in Subsection 3.2 for minimizing RMSE is usually no
longer suitable for constructing confidence intervals, since it is designed to achieve minimal RMSE.
We propose to first obtain σ̂pure as in the data-driven procedure in Subsection 3.2 and then simply
set the data-driven λ be five times of σ̂pure . Since σ̂pure tends to underestimate σ, this data-driven
λ is usually not large with respect to σ. Denote this method as PLS.TwoStage.Soft.Prac or S.TS.P.
−1
After plugging in σ̂ and σ̂j−1 , the square root of the (j, j)th element of Σ̂X , and replace n by
m = #(Iˆ0 ) in the theoretical CI (3.5), we obtain a data-driven CI [β̃j ± m−1/2 σ̂σ̂j−1 zα/2 ], where β̃j
is the PLS.TwoStage.Soft.Prac estimator of βj? for each j.
This data-driven CI is compared with CI’s based on Oracle and OLS methods. More specif(O)
ically, denote the Oracle and OLS estimators of βj? as β̂j
(O)
corresponding CI’s are given by [β̂j
−1/2 (O) −1
σ̂ σ̂j zα/2 ]
± mo
(OLS)
and β̂j
, respectively. Then, the
(OLS)
and [β̂j
± n−1/2 σ̂ (OLS) σ̂j−1 zα/2 ],
where mo is the number of zero incidental parameters and σ̂ (O) and σ̂ (OLS) are the estimators of σ
from O and OLS methods, respectively.
The simulation settings are the same to the previous ones with deterministic sparse incidental
parameters except the following changes. (a) The number of covariates d is reduced to 5 from 50.
This is because when d = 50 and the nominal level is 95%, even the empirical coverage rate (CR)
of the oracle confidence interval for β1? becomes 93.5%, not very close to 95%. (b) The iteration
number is increased from 1000 to 10000 to improve the accuracy of CR’s. (c) The probabilities of
nonzero incidental parameters (p1 , p2 ) are set to be (0.01, 0.01), (0.03, 0.03) and (0.05, 0.05); the
contamination parameter c is increased to 10. In order to achieve good second order asymptotic
approximation, we can either increase the sample size or enlarge the signal noise ratio. Here we
adopt the latter.
Table 3 reports the empirical coverage rates (CR) and average lengths (AL) of the CI’s of β1? and
β2? from O, OLS and PLS.TS.Soft.Prac methods under three different settings on the incidental
parameters. For the oracle method, these three settings are the same and thus only one set of
21
(p1 , p2 )
β1?
O
OLS
S.TS.P
.950
.944
.951
.948
(.05,.05)
.946
(.01,.01)
(.01,.01)
(.03,,03)
(.03,,03)
CR
AL
(.05,.05)
.955
.133
β2?
OLS
S.TS.P
.948
.945
.946
.947
.945
.949
.946
.200
.135
.213
.144
.279
.137
.297
.146
.382
.139
.407
.149
CR
AL
O
.948
.142
Table 3: Coverage rates (CR) and average length (AL) of 95% confidence intervals for β1? and β2? from O,
OLS, PLS.TwoStage.Soft.Prac methods under three settings on deterministic sparse incidental parameters.
simulation results are presented. Table 3 shows that the CR’s of all methods under all settings are
close to the nominal level .95. The OLS treats the deterministic incidental parameters as random
ones and achieves excellent CR’s. However, the AL’s of OLS are significantly larger than those
of O and PLS.TS.Soft.Prac, especially when there are more non-zero incidental parameters. On
the other hand, the AL’s of PLS.TS.Soft.Prac are only slightly larger than those of O. This means
PLS.TS.Soft.Prac has excellent efficiency in terms of AL’s given excellent CR’s. Also note that
the AL’s for β1? are less than those for β2? . This is because the asymptotic variance of β̂1 is less
than that of β̂2 when the covariance matrix ΣX is a Toeplitz matrix. Simulations with random
incidental parameters under the same settings have also been done and the results are similar to
those in Table 3 with slightly inflated AL’s for OLS and PLS.TS.Soft.Prac due to the randomness
of the incidental parameters.
5.4
Real Data Analysis
We implement the penalized estimation with the soft penalty in the method of estimating false
discovery proportion of a multiple testing procedure proposed by Fan et al. (2012b) for investigating the association between the expression level of gene CCT8, which is closely related to Down
Syndrome phenotypes, and thousands of SNPs. The data set consists of three populations: 60 Utah
residents (CEU), 45 Japanese and 45 Chinese (JPTCHB) and 60 Yoruba (YRI). More details on
the data set can be found in Fan et al. (2012b).
In the testing procedure by Fan et al. (2012b), a filtered least absolute deviation regression
(LAD) is used to estimate the loading factors with 90% of the cases (SNPs) whose test statistics
are small and thus the resulting estimator is statistically biased. We upgrade this step with S.P
22
10
●
●
●
●
●
●
●
●
●
3.0
3.5
●
●
2.0
2.5
3.0
●
0.8
50
●
0
0
2.5
●
●
●
2.0
●
●
● ●
●
●
●
●
●
●
●
●
●
0.4
● ●
●
●● ●
●
Estimated FDP(t)
●
●
●● ●
●●●
●
●●
●●
LAD
S.P
●
●
0.0
●
● ●
30
R(t)
●
●
●
●
●
●
●
●
●
30
Estimated V(t)
●
●
●
10
50
CEU
●
●
●●
●
●●
●●
●
●●
3.5
2.0
2.5
3.0
3.5
50
●
●● ●
●
●
●
●
●
20
●
●
●
●
10
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ●●
●
●
●
0.8
●
●
●
●
●
●
●
●
0.4
●
Estimated FDP(t)
●
Estimated V(t)
30
●
●
●
30
40
●
●
●
40
●
●
●
●
0.0
R(t)
●●
●●
20
●
10
50
JPTCHB
●
●
●
●
4
5
6
7
8
●
0
●
9
4
5
6
7
8
9
4
5
6
7
8
9
●●
●
●
●
● ●
●
●
●●
●
0
5
6
−log(t)
7
●
●●
●●
●●
●●
●
● ●
●
●
8
●●
●● ●
●
●●
4
5
6
−log(t)
7
8
●
●●
0.8
●
●●
● ●
●
●● ●
●
●
●
●
● ● ●
●
●
●●
●
●
●
●
●
0.4
80 100
60
●
●
●
●●
●
●
4
Estimated FDP(t)
●
● ●
●
●●
●●
●
●
●
●
●
●
●
●
●
●●●
●
●●
●
●
●
●● ●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
0.0
●
●
●
●
●
●
●
●
40
Estimated V(t)
●
●
●
●●
●
●
●
●
●
●
0
40
60
●
●●
●
●
●●
●
●
●
●●
●
20
●
●
20
R(t)
80 100
YRI
●●
●
●
●
●●
●●
●
●●
4
5
6
7
8
−log(t)
Figure 4: Discovery number R(t), estimated false discovery number V (t) and estimated false discovery
proportion FDP(t) as functions of a threshold on t for populations CEU, JTPCHB and YRI. The x-axis is
− log10 (t).
23
Population
t
R(t)
[
FDP(t)
with LAD
[
FDP(t)
with S.P
CEU
6.12 × 10−4
4
.810
.845
JPTCHB
1.51 × 10−9
5
.153
.373
YRI
2.54 × 10−9
2
.227
.308
ˆ
Table 4: Discovery numbers R(t) and estimated false discover proportions FDP(t)s
from methods with LAD
and S.P for specific values of threshold t.
described in Subsections 3.2 and 5.2 and re-estimate the number of false discoveries V (t) and the
false discovery proportion FDP(t) as functions of − log10 (t), where t is a thresholding value. Figure
[
4 shows the number of total discoveries R(t), V̂ (t) and FDP(t)
from procedures using filtered LAD
[
and S.P. It is clear that V̂ (t) and FDP(t)
with S.P are uniformly larger than but reasonably close
[
to those with filtered LAD. Table 4 contains R(t) and FDP(t)
with filtered LAD and S.P for several
specific thresholds. The estimated FDPs with S.P for CEU and YRI are slightly larger than those
[ for JPTCHB with S.P is more than double of that with filtered LAD. This
with LAD and FDP
suggests that the estimation of FDP with filtered LAD might tend to be optimistic.
6
Conclusion and Discussion
This paper considers the estimation of structural parameters with a finite or diverging number
of covariates in a linear regression model with the presence of high-dimensional sparse incidental
parameters. By exploiting the sparsity, we propose an estimation method penalizing the incidental
parameters. The penalized estimator of the structural parameters is consistent and asymptotically
Gaussian and achieves an oracle property. On the contrary, the penalized estimator of the incidental
parameters possesses only partial selection consistency but not consistency. Thus, the structural
parameters are consistently estimated while the incidental parameters not, which presents a partial
consistency phenomenon. Further, in order to construct better confidence regions for the structural
parameters, we propose a two-step estimator, which has fewer possible asymptotic distributions and
can be asymptotically even more efficient than the one-step penalized estimator when the size and
magnitude of nonzero incidental parameters are substantially large.
Simulation results show that the penalized methods with best regularization parameters achieve
significantly smaller mean square errors than the ordinary least squares method which ignores the
incidental parameters. Also provided is a data-driven regularization parameter, with which the
24
penalized estimators continue to significantly outperform ordinary least squares when the incidental
parameters are too large to be neglected. In terms of average length together with excellent
coverage rates, the advantage of the confidence intervals based on the two-step estimator with an
alternative data-driven regularization parameter is verified by simulations. A data set on genomewide association is analyzed with a multiple testing procedure equipped with a data-driven penalized
method and false discovery proportions are estimated.
In econometrics, a fixed effect panel data model is given by, for 1 ≤ i ≤ n and 1 ≤ t ≤ T ,
Yit = µ?i + X Tit β ? + it ,
(6.1)
where µ?i ’s are unknown fixed effects. When T diverges, the fixed effects can be consistently
estimated. When T is finite and greater than or equal to 2, although the fixed effects can no
longer be consistently estimated, they can be removed by a within-group transformation: for each
i, Yit − Ȳi = (X it − X̄ i )T β ? + it − ¯i , where Ȳi , X̄ i and ¯i are the averages of Yit ’s, X it ’s and it ’s,
respectively. When T is equal to 1, however, the within-group transformation fails. Note that,
with T = 1, Model (6.1) becomes Model (1.1) so that the proposed penalized estimations provide
a solution under the sparsity assumption on the fixed effects.
Although this paper only illustrates the partial consistency phenomenon of a penalized estimation method for a linear regression model, such a phenomenon shall universally exist for a general
parametric model, which contains both a structural parameter and a high-dimensional sparse incidental parameter. For example, consider a panel data logistic regression model: P (Yit = 1|X it ) =
(1 + exp{−(µ?i + X Tit β ? )})−1 . When T is finite, the fixed effects µ?i ’s cannot be removed by the
within-group transformation as in the panel data linear model (6.1). However, the proposed penalized estimations can still provide a solution.
Further, if the structural parameter has a dimension diverging faster than the sample size and
is sparse, it is expected that the partial consistency phenomenon will continue to appear when
sparsity penalty is imposed on both the structural and incidental parameters.
Acknowledgement
The authors thank the Editor, an associate editor, and three referees for their many helpful comments that have resulted in significant improvements in the article. The research was partially
supported by NSF grants DMS-1206464 and DMS-0704337 and NIH Grants R01-GM100474-01
and R01-GM072611-08.
25
A
Appendix
In this appendix, we provide the proofs of the theoretical results in Section 4. The proofs of the
results in Sections 2 and 3 are in Supplements D and E.
Denote Sk,l = S{k,k+1,··· ,l} and Sk,l = S{k,k+1,··· ,l} . Let B = {maxs+1≤i≤n kX i k2 ≤ κn } and
T
D = ni=1 {−γn ≤ i ≤ γn }. Then P (B) → 1 and P (D) → 1 by (2.2).
Proof of Lemma 4.1. We first consider Si0 ’s, then Si1 ’s, and finally Si2 ’s with i = 1, 2, 3. Consider
? }. Note that P (A) ≥ P (A|B)P (B) and P (B) → 1. It
S10 , S20 and S30 . Let A = {S10 = S10
√
suffices to show that P (A|B) → 1. By λ dκn , it follows P (A|B) ≥ P ({s + 1 ≤ i ≤ n :
√
√
? |B) ≥ P ({s + 1 ≤ i ≤
−λ + maxs+1≤i≤n kX i k2 dC ≤ i ≤ λ − maxs+1≤i≤n kX i k2 dC} ⊃ S10
√
√
?
? ) ≥ P (D) → 1. Thus, wpg1, S
n : −λ + κn dC ≤ i ≤ λ − κn dC} ⊃ S10
10 = S10 . From
? , it follows that, wpg1, S
S10 ∪ S20 ∪ S30 = S10
20 = S30 = ∅. Consider S21 , S31 and S11 . Recall
√
that µ? = min{|µ?i | : 1 ≤ i ≤ s1 } and note that λ − µ? + dCκn < −γn when n is large. Let
? and S
?c
?
S211 = S21 S21
212 = S21 S21 . We will show P (S211 = S21 ) → 1 and P (S212 = ∅) → 1.
? ) → 1. Denote A = {S
?
Then P (S21 = S21
1
211 ⊃ S21 }. On the event B, S211 ⊃ {1 ≤ i ≤ s1 :
√
i > λ − µ? + dCκn and µ?i > 0} ⊃ {1 ≤ i ≤ s1 : i > −γn and µ?i > 0}. Then, P (A1 ) ≥
? )P (B) → 1 · 1 = 1. It follows that,
P (A1 |B)P (B) ≥ P ({1 ≤ i ≤ s1 : i > −γn and µ?i > 0} ⊃ S21
? . Note that S
?
?
wpg1, S211 ⊃ S21
211 ⊂ S21 . Then, wpg1, S211 = S21 . Denote A2 = {S212 = ∅}.
√
On the event B, S212 ⊂ {1 ≤ i ≤ s1 : i > λ + µ? − dCκn and µ?i < 0}, which contains
{1 ≤ i ≤ s1 : i > γn }. Then, P (A2 ) ≥ P (A2 |B)P (B) ≥ P ({1 ≤ i ≤ s1 : i > γn } = ∅)P (B) →= 1.
? ) → 1. Similarly, we can show, wpg1, S
?
Then, wpg1, S212 = ∅. Thus, P (S21 = S21
31 = S31 .
? ∪ S ? . Then, wpg1, S
Note that S11 , S21 and S31 are disjoint and their union is S21
11 = ∅.
31
√
? }. Note that −λ − µ? + dCκ < −γ and
Consider S12 , S22 and S32 . Denote A = {S12 = S12
n
n
i
√
λ − µ?i − dCκn > γn when n is large for s1 + 1 ≤ i ≤ s. On the event B, S12 ⊃ {s1 + 1 ≤ i ≤
√
√
s : −λ − µ?i + dCκn ≤ i ≤ λ − µ?i − dCκn }, which contains {s1 + 1 ≤ i ≤ s : −γn ≤ i ≤ γn }.
? )P (B) → 1. Thus,
Then, P (A) ≥ P (A|B)P (B) ≥ P ({s1 + 1 ≤ i ≤ s : −γn ≤ i ≤ γn } = S12
? . Note that S , S
?
wpg1, S12 = S12
12
22 and S32 are disjoint and their union is S12 . Then, wpg1,
S22 = S32 = ∅.
2 = (1/d)
Before proceeding to the proofs of Theorems 4.2 to A.6, we denote σ̄X
P
P
2
and σ̄XX
= (1/d2 ) dk=1 dl=1 Var[X0k X0l ] and make the following assumptions.
2 is bounded.
Assumption (E1): σ̄X
26
Pd
j=1 Var[X0j ]
2
Assumption (E2): σ̄XX
is bounded.
Assumption (E) in Section 4 implies Assumptions (E1) and (E2) by Cauchy-Schwarz inequality.
For simplicity, we adopt the notation ., which means the left hand side is bounded by a constant
times the right, where the constant does not affect related analysis. Below are three lemmas needed
for proving Theorems 4.2 to A.6. Their proofs are in Supplement F. Suppose that M and E are
matrices and k·k is a matrix norm and that {An } is a sequence of random d × d matrices and A a
deterministic d × d matrix, and denote Σ̂n = (1/n)Sn , the sample covariance matrix.
Lemma A.1 (Stewart (1969)). If kIk = 1 and kM −1 kkEk < 1, then
Lemma A.2. If kA−1 kF,d
k(M + E)−1 − M −1 k
kM −1 kkEk
≤
.
kM −1 k
1 − kM −1 kkEk
√
P
P
−1
is bounded, An −→ A, and rd ≥ 1/ d, then A−1
n −→ A , where the
convergence in probability is wrt rd k·kF .
P
Lemma A.3. If Assumption (E2) holds and rd2 d4 /n → 0, then Σ̂n −→ ΣX wrt rd k·kF .
Proof of Theorem 4.2. By the proof of Lemma 4.1, wpg1, the solution β̂ n to ϕn (β) = 0 on
BC (β ? ) is explicitly given by β̂ n = β ? + T0−1 (T1 + T2 + T3 − T4 ), where T0 = (1/n)Ss1 +1,n ,
?
? and T4 = (λ/n)SS ? . Then, rd kβ̂ − β k2 ≤
T1 = (1/n)SµS ? , T2 = (1/n)Ss1 +1,n , T3 = (λ/n)SS21
n
31
12
√
P
kT0−1 kF,d 4i=1 rd dkTi k2 . We will show that kT0−1 kF,d is bounded by a positive constant wpg1
√
P
and rd dkTi k2 −→ 0 for i = 1, 2, 3, 4. Then, rd kβ̂ n − β ? k2 = oP (1). Consider T0 . By Lemma
P
A.3, kT0 − ΣX kF,d −→ 0 under Assumption (E2) and the condition d3 /n → 0. Then, by Lemma
P
−1
A.2, together with Assumption (D), kT0−1 − Σ−1
X kF,d −→ 0. This implies that, wpg1, kT0 kF,d is
√
√
bounded by a positive constant. Consider T1 . Wpg1, rd dkT1 k2 ≤ rd ds2 κn γn /n = o(1) for s2 =
√
P
o(n/(rd dκn γn )). Consider T2 . For any δ > 0, P (kT2 k2 > δ) ≤ (1/δ 2 )P k(1/n) ni=s1 +1 X i i k22 ≤
√
Pd
2 /(nδ 2 ), where σ̄ 2 = (1/d)
2
2 2 2 2
2
dσ 2 σ̄X
j=1 σj . Thus, P (rd dkT2 k2 > δ) ≤ rd d σ σ̄X /(nδ ) → 0 by AsX
√
√
sumption (E1) and (rd d)2 /n → 0. Consider T3 and T4 . Wpg1, rd dkT3 k2 ≤ rd dλs1 κn /n = o(1)
√
√
for s1 = o(n/(rd dλκn )). Similarly, rd dkT4 k2 = oP (1).
The next lemma is needed for proving Theorem 4.3 and its proof is in Supplement F. Suppose
2
{ξ i } are i.i.d. copies of ξ 0 , a d-dimensional random vector with mean zero. Denote σξ,max
=
2
max1≤j≤d Var[ξ0j ], σξ,min
= min1≤j≤d Var[ξ0j ] and γξ,max = max1≤j≤d E|ξ0j |3 .
Lemma A.4. Suppose σξ,max and γξ,max are bounded from above and σξ,max is bounded from zero.
√
√
√ P
If d = o( n), then (1/ n) ni=1 ξ i = OP ( d log d) wrt k·k2 .
27
Proof of Theorem 4.3. We reuse the notations Ti ’s in the proof of Theorems 4.2, from which,
√
√
nAn (β̂ n − β ? ) = V1 + V2 + V3 − V4 , where Vi = B n Ti for i = 1, 2, 3, 4 and B n = nAn T0−1 .
d
It is sufficient to show that V2 −→ N (0, σ 2 GX ) and other Vi ’s are oP (1). Consider V1 . We
√
have kV1 k2 ≤ ndkAn kF kT0−1 kF,d kT1 k2 . By Assumption (F), kAn kF is bounded. By Lemmas
A.2 and A.3 and Assumption (D), for d = o(n1/3 ), wpg1, kT0−1 kF,d is bounded. We have, wpg1,
p
√ √
kT1 k2 ≤ s2 κn γn /n. Then, kV1 k2 . d/ns2 κn γn , Thus, kV1 k2 = oP (1) for s2 = o( n/( dκn γn )).
√
√
nAn (T0−1 − Σ−1
Consider V2 . We have V2 = V21 + V22 , where V21 = nAn Σ−1
X T2 and V22 =
X )T2 .
p
Pn
√
First, note that V21 = (n − s1 )/n i=s1 +1 Z n,i , where Z n,i = (1/ n − s1 )An Σ−1
X X i i . On one
Pn
hand, for every δ > 0, i=s1 +1 EkZ n,i k22 {kZ n,i k2 > δ} ≤ (n − s1 )EkZ n,0 k42 /δ 2 , and EkZ n,0 k42 =
2
−1
T
1
2
E40 E(X T0 Σ−1
X An An ΣX X 0 ) ,
(n−s1 )2
−2
d
4
2
which is ≤ (n−s
2 E0 λmax (Gn )λmin (ΣX )κX . Then, by As1)
P
√
sumptions (D’), (E) and (F) and for d = o( n), ni=s1 +1 EkZ n,i k22 {kZ n,i k2 > δ} → 0. On the other
P
T
2
hand, ni=s1 +1 Cov(Z n,i ) = σ 2 An Σ−1
X An → σ GX by Assumption (F). Thus, by central limit thed
orem (see Proposition 2.27 in van der Vaart (1998)), V21 −→ N (0, σ 2 GX ). Next, consider V22 . Note
−1/2 k√nT k . By Assumption (F), kA k
that kV22 k2 ≤ kAn kF (d log(d))1/2 kT0−1 −Σ−1
2 2
n F
X kF (d log(d))
5
is O(1); by Lemmas A.2 and A.3, (d log(d))1/2 kT0−1 −Σ−1
X kF is oP (1) for d log(d) = o(n); by Lemma
√
√
P
A.4, (d log(d))−1/2 k nT2 k2 = (d log(d))−1/2 k √1n Ss1 +1,n k2 is OP (1) for d = o( n). Then, V22 −→ 0.
d
Thus, by slutsky’s lemma, V2 −→ N (0, σ 2 GX ). Consider V3 and V4 . First consider V3 . By noting
√
√
√
√
√
that s1 = o( n/(λ dκn )), wpg1, kV3 k2 ≤ ndkAn kF kT0−1 kF,d kT3 k2 . dλs1 κn / n → 0. Thus,
kV3 k2 = oP (1). In the same way, kV4 k2 = oP (1).
T1
Proof of Theorem 4.4. By the definition of E, we have P (E) = T1 T2 T3 , where T1 = P ( si=1
{|µ?i +
T
T
X Ti (β ? −β̂)+i | > λ}), T2 = P ( si=s1 +1 {|µ?i +X Ti (β ? −β̂)+i | ≤ λ}) and T3 = P ( ni=s+1 {|X Ti (β ? −
β̂) + i | ≤ λ}). We will show that each Ti converges to one. Then, P (E) → 1. Denote C =
{rd kβ̂ − β ? k2 ≤ 1}. Then P (C) → 1 since β̂ is a consistent estimator of β ? wrt rd k·k2 . Consider
S
T1 . We have 1 − T1 ≤ T11 + T12 , where T11 = P ( i∈S ? {|µ?i + X Ti (β ? − β̂) + i | ≤ λ}) and
21
S
T12 = P ( i∈S ? {|µ?i +X Ti (β ? − β̂)+i | ≤ λ}). It is sufficient to show that both T11 and T12 converge
√31
S
to zero. By dκn λ µ? , T11 ≤ P ( i∈S ? {i ≤ λ − µ? + kX i k2 · kβ̂ − β ? k2 }, C) + P (C c ), which
21
√
S
?
c
is ≤ P ( i∈S ? {i ≤ λ − µ + dκn }) + P (C ) ≤ s1 P {0 ≤ −γn } + P (C c ) −→ 0. Similarly, T12 → 0.
21
√
T
Thus T1 → 1. Consider T2 and T3 . By αγn ≤ λ and dκn λ, T2 ≥ P ( si=s1 {−λ−µ?i +(1/rd )κn ≤
√
√
T
i ≤ λ − µ?i − (1/rd )κn }, C), which is ≥ P ( si=s1 {−λ − µ?i + dκn ≤ i ≤ λ − µ?i − dκn }, C) ≥
T
P ( si=s1 {−γn ≤ i ≤ γn }, C) → 1. Then T2 → 1. Similarly, T3 → 1.
Proof of Theorem 4.5. Note that
√
1/2
nAn ΣX (β̃ − β ? ) = R̃1 + R̃2 + V1 + V2 , where R̃1 =
28
√
nAn R1 ,
R̃2 =
√
nAn R2 , R1 = (X TIˆ X Iˆ0 )−1 X TIˆ Y Iˆ0 {Iˆ0 6= I0 }, R2 = −(X TI0 X I0 )−1 X TI0 Y I0 {Iˆ0 6= I0 },
0
0
and Vi ’s are defined in the proof of Theorem 4.3. Since P (kR̃1 k2 = 0) ≥ P {Iˆ0 = I0 } → 1,
we have R̃1 = oP (1). Similarly, R̃2 = oP (1). By the proof of Theorem 4.3, V1 = oP (1) and
d
V2 −→ N (0, σ 2 GX ). Therefore, the desired result follows by Slutsky’s lemma.
Lemma A.5 (Consistency on σ̂). Suppose the assumptions and conditions of Theorem 4.2 hold
√
P
with rd ≥ d. If s2 = o(n/γn2 ), then σ̂ −→ σ.
√
Proof of Lemma A.5. Since the assumptions and conditions of Theorem 4.2 hold with rd ≥ d,
√
the penalized estimators β̂ and β̃ are consistent estimators of β ? wrt dk·k2 by Theorems 4.2 and
F.4 in Supplement F. Let A = {Iˆ0 = I0 }. Then A occurs wpg1 by Theorem 4.4.
Note that σ̂ 2 = T A + σ̂ 2 Ac , where T = (n − s1 )−1 kY I0 − X TI0 β̃k22 . It suffices to show that
Pn
P6
P
?
T
2
−1
T −→ σ 2 . Note that T =
i=s1 +1 [X i (β − β̃)] , T2 = (n −
i=1 Ti , where T1 = (n − s1 )
P
P
P
s1 )−1 ni=s1 +1 2i , T3 = 2(n − s1 )−1 ni=s1 +1 X Ti (β ? − β̃)i , T4 = (n − s1 )−1 si=s1 +1 µ?2
i , T5 = 2(n −
P
P
P
s1 )−1 si=s1 +1 µi X Ti (β ? − β̃) and T6 = 2(n−s1 )−1 si=s1 +1 µ?i i . It is clear that T2 −→ σ 2 . Thus, it
√
is sufficient to show other Ti ’s are oP (1). For every η > 0, wpg1, dkβ ? − β̃k2 ≤ η. By Assumption
√
T 2
1 Pn
2 . η 2 . For every
dkβ ? − β̃k2 )2 ≤ 2η 2 d1 EkX T0 k22 = 2η 2 σ̄X
(E1), wpg1, |T1 | ≤ d1 n−s
kX
k
(
i
2
i=s
+1
1
1
√
T
1 Pn
kX
k
η > 0, wpg1, |T3 | ≤ 2 √1d n−s
dkβ ? − β̃k2 ≤ 4η √1d EkX T0 0 k2 = 4σησ̄X . η. For
i
2
i
i=s
+1
1
1
√
√
1
s
γ
κ
dkβ ? −
s2 = o(n/γn2 ), |T4 | ≤ (n − s1 )−1 s2 γn2 → 0. For s2 = o( dn/(γn κn )), |T5 | ≤ 2 √1d n−s
2
n
n
1
P
1
1
β̃k2 ≤ 2η √1d n−s
s2 γn κn −→ 0. For s2 = o(n/γn ), wpg1, |T6 | ≤ 4 n−s
γn s2 E|0 | → 0.
1
1
Theorem A.6 (Asymptotic Distributions on β̂ and β̃ with ĜX,n ). Under the assumptions and
√ −1/2
d
conditions of Theorem 4.3, if d8 (log(d))2 = o(n), then nĜX,n An (β̂ − β ? ) −→ N (0, σ 2 I q ).
Similarly, under the assumptions and conditions of Theorem 4.5, If d8 (log(d))2 = o(n), then
√ −1/2
d
nĜX,n An (β̃ − β ? ) −→ N (0, σ 2 I q ).
−1/2
Note that a stronger requirement on d is required to handle ĜX,n in Theorem A.6.. Below is
a lemma needed for proving Theorem A.6.
Lemma A.7 (Wihler (2009)). Suppose A and B are m × m symmetric positive-semidefinite
matrices. Then, for p > 1, kA1/p − B 1/p kpF ≤ m(p−1)/2 kA − BkF . Specifically, for p = 2,
kA1/2 − B 1/2 kF ≤ (m1/2 kA − BkF )1/2 .
Proof of Theorem A.6. We only show the result on β̂. since the result on β̃ can be obtained in a sim√ −1/2
ilar way. We reuse the notations Ti ’s in the proof of Theorems 4.2, from which, nĜX,n An (β̂ n −
√ −1/2
√
−1/2
−1/2
β ? ) = M + R, where M = nGX,n An (β̂ n − β ? ) and R = n(ĜX,n − GX,n )An (β̂ n − β ? ). By
29
d
P
Theorem 4.3, M −→ N (0, σ 2 GX ). Then, it is sufficient to show that R −→ 0 wrt k·k2 . We have
√
−1/2
−1/2
R = R1 + R2 + R3 − R4 , where Ri = B n Ti for i = 1, 2, 3, 4 and B n = n(ĜX,n − GX,n )An T0−1 .
We will show each Ri converges to zero in probability, which finishes the proof. Before that,
−1/2
−1/2
−1/2
−1/2
we first establish an inequality for kĜX,n − GX,n kF . By Lemma A.7, kĜX,n − GX,n kF ≤
−1
√
P
1/2 . Note that, by Lemma A.3, kΣ̂ − Σ k −→
( qkĜX,n − G−1
0 for d4 = o(n). Then, by
n
X F
X,n kF )
−1
P
2
Lemma A.2, kĜX,n − GX,n kF ≤ kAn k2F kΣ̂n − Σ−1
X kF . kAn kF kΣ̂n − ΣX kF −→ 0. Thus, by
−1
2
Lemma A.2, kĜX,n − G−1
X,n kF . kĜX,n − GX,n kF . kAn kF kΣ̂n − ΣX kF . Since q is a fixed in−1/2
√
−1/2
teger, it follows kĜX,n − GX,n kF . kAn kF ( qkΣ̂n − ΣX kF )1/2 . kAn kF (kΣ̂n − ΣX kF )1/2 .
√ √
−1/2
−1/2
Consider R1 . Note that kR1 k2 ≤ n dkĜX,n − GX,n kF kAn kF kT0−1 kF,d kT1 k2 , which is .
√
n(dkΣ̂n − ΣX kF )1/2 kAn k2F kT0−1 kF,d kT1 k2 . By Lemmas A.2 and A.3, dkΣ̂n − ΣX kF = oP (1)
for d6 = o(n). By Assumption (F), kAn kF is bounded. By Lemmas A.2 and A.3 and Assumption (D), for d = o(n1/3 ), wpg1, kT0−1 kF,d is bounded. Also note that, wpg1, kT1 k2 ≤
√
√
s2 κn γn /n. Then, kR1 k2 . s2 κn γn / n. Thus, kR1 k2 = oP (1) for s2 = o( n/(κn γn )). Consider
√
−1/2
−1/2
R2 . Note that kR2 k2 ≤ kĜX,n − GX,n kF kAn kF kT0−1 kF k nT2 k2 , which is . (d2 log(d)kΣ̂n −
√
ΣX kF )1/2 kAn k2F kT0−1 kF,d (d log(d))−1/2 k nT2 k2 . By Lemmas A.2 and A.3, d2 log(d)kΣ̂n −ΣX kF is
oP (1) for d8 (log(d))2 = o(n). By Assumption (F), kAn kF is O(1). By Lemmas A.2 and A.3 and As√
sumption (D), for d = o(n1/3 ), wpg1, kT0−1 kF,d is bounded. By Lemma A.4, (d log(d))−1/2 k nT2 k2 =
√
P
(d log(d))−1/2 k √1n Ss1 +1,n k2 is OP (1) for d = o( n). Thus, R2 −→ 0. Consider R3 and R4 .
√
√
−1/2
−1/2
By s1 = o( n/(λκn )), wpg1, kR3 k2 ≤ nkĜX,n − GX,n kF kAn kF kT0−1 kF kT3 k2 , which is .
√
√
n(dkΣ̂n − ΣX kF )1/2 kAn k2F kT0−1 kF,d kT3 k2 . λs1 κn / n → 0. Thus, kR3 k2 = oP (1). In the same
way, kR4 k2 = oP (1).
B
Supplementary Materials
Additional materials for Sections 1 to 4 can be found in the file of supplementary materials.
References
Debabrata Basu. On the elimination of nuisance parameters. Journal of the American Statistical
Association, 72:355–366, 1977.
Derek Bean, Peter Bickel, Noureddine El Karoui, Chinghway Lim, and Bin Yu. Penalized robust
regression in high-dimension. 2012.
30
Louis H.Y. Chen, Larry Goldstein, and Qi-Man Shao. Normal Approximation by Stein’s Method.
Springer-Verlag, 2010.
Jianqing Fan and Runze Li. Variable selection via nonconcave penalized likelihood and its oracle
properties. Journal of the American Statistical Association, 96(456):1348–1360, Dec 2001. ISSN
0162-1459.
Jianqing Fan and Jinchi Lv. A selective overview of variable selection in high dimensional feature
space. Statistica Sinica, 20:101–148, 2010.
Jianqing Fan and Heng Peng. On non-concave penalized likelihood with diverging number of
parameters. The Annals of Statistics, 32:928–961, 2004.
Jianqing Fan, Yuan Liao, and Martina Mincheva. High dimensional covariance matrix estimation
in approximate factor models. 2011.
Jianqing Fan, Yingying Fan, and Emre Barut. Adaptive robust variable selection. 2012a.
Jianqing Fan, Yang Feng, and Xin Tong. A road to classification in high dimensional space: the
regularized optimal affine discriminant. Journal of the Royal Statistical Society Series B, 2012b.
Peter J. Huber. Robust estimation of a location parameter. The Annals of Mathematical Statistics,
35:73–101, 1964.
Peter J. Huber. Robust regression: Asymptotics, conjectures and monte carlo. The Annals of
Statistics, 1:799–821, 1973.
Johannes Jahn. Introduction to the Theory of Nonlinear Optimization. Springer Berlin Heidelberg,
2007.
J. Kiefer and J. Wolfowitz. Consistency of the maximum likelihood estimator in the presence of
infinitely many incidental parameters. The Annals of Mathematical Statistics, 27:887–906, 1956.
Michael R. Kosorok. Bootstrapping the grenander estimator. In Beyond Parametrics in Interdisciplinary Research: Festschrift in Honor of Professor Pranab K. Sen, pages 282–292. Institute
of Mathematical Statistics: Hayward, CA., 2008.
S. Lambert-Lacroix and L. Zwald. Robust regression through the huber’s criterion and adaptive
lasso penalty. Electronic Journal of Statistics, 5:1015–1053, 2011. ISSN 1935-7524.
31
Tony Lancaster. The incidental parameter problem since 1948. Journal of Econometrics, 95:
391–413, 2000.
Marcelo J. Moreira. A maximum likelihood method for the incidental parameter problem. The
Annals of Statistics, 37(6A):3660–3696, 2009. ISSN 0090-5364.
Jerzy Neyman and Elizabeth L. Scott. Consistent estimates based on partially consistent observations. Econometrica, 16:1–32, 1948.
Stephen Portnoy and Xuming He. A robust journey in the new millennium. Journal of the American
Statistical Association, 95:1331–1335, 2000.
Albert N. Shiryaev. Probability. Springer-Verlag, second edition, 1995.
G. W. Stewart. On the continuity of the generalized inverse. SIAM Journal on Applied Mathematics,
17:33–45, 1969.
Runlong Tang, Moulinath Banerjee, and Michael R. Kosorok. Likelihood based inference for current
status data on a grid: A boundary phenomenon and an adaptive inference procedure. The Annals
of Statistics, 40(1):45–72, 2012.
Robert Tibshirani. Regression shrinkage and selection via the lasso. J. Royal. Statist. Soc B, 58:
267–288, 1996.
Aad W. van der Vaart. Asymptotic Statistics. Cambridge University Press, 1998.
Aad W. van der Vaart and Jon A. Wellner. Weak Convergence and Empirical Processes. Springer,
1996.
Thomas P. Wihler. On the holder continuity of matrix functions for normal matrices. Journal of
inequalities in pure and applied mathematics, 10, 2009.
Peng Zhao and Bin Yu. On model selection consistency of lasso. The Journal of Machine Learning
Research, 7(Nov):2541–2563, 2006.
32
Supplementary Materials for Paper:
“Partial Consistency with Sparse Incidental Parameters”
by Jianqing Fan, Runlong Tang and Xiaofeng Shi
C
Supplement for Section 1
In this supplement, we first show the method proposed by Neyman and Scott (1948) does not work
for model (1.1) and then explain which assumptions or conditions for the consistent results of the
penalized methods in Zhao and Yu (2006), Fan and Peng (2004) and Fan and Lv (2011) are not
satisfied for model (1.1).
Although the modified equations of maximum likelihood method proposed by Neyman and Scott
(1948) could handle “a number of important cases” with incidental parameters, unfortunately, it
does not work for model (1.1). More specifically, consider the simplest case of model (1.1) with
d = 1:
Yi = µ?i + Xi β ? + i , for i = 1, 2, · · · , n,
where {i } are i.i.d. copies of N (0, σ 2 ). Using the notations of Neyman and Scott (1948), the
√
likelihood function for (Xi , Yi ) is pi = pi (β, σ, µi |Xi , Yi ) = ( 2πσ)−1 exp{−(2σ 2 )−1 (Yi −µ?i −Xi β)2 },
√
and the log-likelihood function is log pi = − log( 2πσ) − (2σ 2 )−1 (Yi − µ?i − Xi β)2 . Then, the score
functions are
1
∂ log pi
= 2 (Yi − µ?i − Xi β)Xi ,
∂β
σ
∂ log pi
1
1
φi2 =
= + 3 (Yi − µ?i − Xi β)2 ,
∂σ
σ σ
∂ log pi
1
ωi =
= 2 (Yi − µ?i − Xi β).
∂µi
σ
φi1 =
From the equation ωi = 0, we have µ̂i = Yi − Xi β. Plugging this µ̂i into φi1 and φi2 (replacing
µi with µ̂i ), we obtain φi1 = 0 and φi2 = 1/σ. Then, Ei1 = Eφi1 = 0 and Ei2 = Eφi1 = 1/σ.
Thus, Ei1 and Ei2 do only depend on the structural parameters (β ? and σ). However, we then
have Φi1 = φi1 − Ei1 = 0 and Φi2 = φi2 − Ei2 = 0. This means Fn1 = Fn2 = 0, independent of
structural parameters! Consequently, the estimation equations degenerate to two 0 = 0 equations,
which means the modified equation of maximum likelihood method does not work for model (1.1).
i
Next, we explicitly explain which assumptions or conditions for the consistent results of the
penalized methods in Zhao and Yu (2006), Fan and Peng (2004) and Fan and Lv (2011) are not
valid for model (1.1).
Zhao and Yu (2006) derive strong sign consistency for lasso estimator. However, their consistency results Theorems 3 and 4 do not apply to model (1.1), since the above specific design matrix
X does not satisfy their regularity condition (6) on page 2546. More specifically, with model (1.1),
0
0
I
X
1
s
1,s
a.s.
n
−→
,
C11
=
n X T Pn X X T
0 Σ
1,s
i
i=1
X
i
where ΣX is the covariance matrix of the covariates. This means that some of the eigenvalues of
n goes to 0 as n → ∞. Then the regularity condition (6), which is
C11
n
αT C11
α ≥ a positive constant , for all α ∈ Rs+d such that kαk22 = 1,
does not hold any more. Thus the consistency results Theorems 3 and 4 in Zhao and Yu (2006) is
not applicable for model (1.1).
Fan and Peng (2004) show the consistency with Euclidean metric of a penalized likelihood
estimator when the dimension of the sparse parameter increases with the sample size in Theorem
1 on Page 935. Under their framework, the log-likelihood function of the data point Vi = (X i , Yi )
for each i from model (1.1) with random errors being i.i.d. copies of N (0, σ 2 ) is given by
log fn (Vi , µi , β) ∝ −
1
(Yi − µi − X Ti β)2 ,
2σ 2
where ∝ means “proportional to”. As we can see that log-likelihood functions with different i’s
might different since µi ’s might be different for different i’s. This violates a condition that all the
data points are i.i.d. from a structural density in Assumption (G) on Page 934.
This violation might not be essential, however, since we could consider the log-likelihood function for all the data directly. That is, we consider
Ln (µ, β) =
n
X
log fn (Vi , µi , β) ∝ −
n
1 X
(Yi − µi − X Ti β)2 .
2σ 2
i=1
i=1
Then, the Fisher information matrix for (µ, β) is given by
σ −2 I n
0
,
In+d (µ, β) =
0
nσ −2 Σ2X
ii
where In is the n × n identity matrix. Then, the Fisher information for one data point is
−1 σ −2 I
n
0
1
n
.
In+d (µ, β) =
n
0
σ −2 Σ2
X
It is clear that the minimal eigenvalue λmin (In+d (µ, β)/n) = n−1 σ 2 → 0 as n → ∞. This violates
the condition that the minimal eigenvalue should be lower bounded from 0 in Assumption (F) on
Page 934. Thus, the consistency result Theorem 1 in Fan and Peng (2004) can not be applied to
model (1.1).
Fan and Lv (2011) “consider the variable selection problem of nonpolynomial dimensionality
in the context of generalized linear models” by taking the penalized likelihood approach with
folded-concave penalties. Theorem 3 on page 5472 of Fan and Lv (2011) shows that there exists a
consistent estimator of the unknown parameters with the Euclidean metric under certain conditions.
In Condition 4 on page 5472, there is a condition on a minimal eigenvalue
min λmin [X TI Σ(XI δ)X I ] ≥ cn,
δ∈N0
where X I consists of the first s + d columns of the design matrix X. With model (1.1), this
condition becomes
λmin [X TI X I ] ≥ cn,
which is
n
λmin [(1/n)X TI X I ] = λmin [C11
] ≥ c,
n is the matrix defined in Zhao and Yu (2006) and c is a positive constant. Since the minwhere C11
n ] converges to 0, the above condition does not hold. Thus, the consistency
imal eigenvalue λmin [C11
result Theorem 3 of Fan and Lv (2011) is not applicable for model (1.1).
D
Supplement for Section 2
In this supplement, we provide Lemmas D.1 and D.3, Proposition D.2 and their proofs. Before
that, there are two graphs, Figures 5 and 6, illustrating the incidental parameters and the step of
updating the responses in the iteration algorithm with d = 1.
iii
n−s
γn
λ
µ∗
{µ i∗ : s1 + 1 ≤ i ≤ s}
s2
0
∞
{µi∗ : 1 ≤ i ≤ s1}
s1
−∞
− λ − µ∗
−γn
Figure 5: An illustration of three types of µ?i ’s, that is, large µ?1 , bounded µ?2 and zero µ?3 . The
negative half of the real line is folded at 0 under the positive half for convenience. For the penalized least square method with a soft penalty function and under the assumption of fixed d, the
√
specification of the regularization parameter λ is that κn λ, αγn ≤ λ, and λ min{µ? , n}.
Y
0
X
f ( x ) = xβ ( k −1)
Figure 6: An illustration for the updating of responses with d = 1. The solid black line is a fitted
regression line. The dashed black lines are the corresponding shifted regression lines. The circle
and diamond points are the original data points. The circle and triangle points are the updated
data points. That is, the diamond points are drawn onto the shifted regression lines.
iv
Lemma D.1. A necessary and sufficient condition for (µ̂, β̂) to be a minimizer of L(µ, β) is that
β̂ = (X T X)−1 X T (Y − µ̂),
for i ∈ Iˆ0c ,
Yi − µ̂i − X Ti β̂ = λSign(µ̂i ),
|Yi − X Ti β̂| ≤ λ,
for i ∈ Iˆ0 ,
where Sign(·) is a sign function and Iˆ0 = {1 ≤ i ≤ n : µ̂i = 0}.
Proof of Lemma D.1. By subdifferential calculus (see, for example, Theorem 3.27 in Jahn (2007)),
a necessary and sufficient condition for (µ̂, β̂) to be a minimizer of L(µ, β) is that zero is in the
subdifferential of L at (µ̂, β̂), which means that, for each i,
β̂ = (X T X)−1 X T (Y − µ̂),
Yi − µ̂i − X Ti β̂ = λSign(µ̂i ),
|Yi − X Ti β̂| ≤ λ,
if µ̂i 6= 0,
if µ̂i = 0.
Thus, the conclusion of Lemma D.1 follows.
Proposition D.2. Suppose Assumptions (A) and (B) hold and there exist positive constants C1
and C2 such that kβ ? k2 < C1 and kβ (0) k2 < C2 wpg1. If s1 λ/n = O(1) and s2 γn /n = o(1), then,
for every K ≥ 1 and k ≤ K, wpg1 as n → ∞,
kβ (K+1) − β (K) k2 ≤ O((s1 /n)K ),
and
√
kβ (k) k2 ≤ 2 dC1 + C2 .
Remark 2. For any prespecified critical value in the stopping rule, Proposition D.2 implies that
the algorithm stops at the second iteration wpg1. In practice, the sample size n might not be
large enough for the two-iteration estimator to have a decent performance so that more iterations
are usually needed to activate the stopping rule. By Proposition D.2, K iterations will make the
distance kβ (K+1) −β (K) k2 of the small order (s1 /n)K . When s1 /n is small, the algorithm converges
quickly, which has been verified by our simulations.
√
Proof of Proposition D.2. First, we show that, wpg1, kβ (1) k2 is bounded by 2 dC1 + C2 . For each
k ≥ 1, we have
Sβ (k) = SµS11 + SµS12 + SS1 β ? + SS1 + SS2 ∪S3 β (k−1) + λ(SS2 − SS3 ),
v
where Si = ∪3j=1 Sij (β (k−1) ) for i = 1, 2, 3 and Sij ’s are defined at the end of Section 2. Denote
Ak−1 as the event
?
?
?
?
?
{S11 (β (k−1) ) = ∅, S12 (β (k−1) ) = S12
, S1 (β (k−1) ) = S10
∪ S12
; S2 (β (k−1) ) = S21
; S3 (β (k−1) ) = S31
},
? ’s are defined at the beginning of Section 3.
where Sij
By Lemma 3.1, P (A0 ) → 1. Thus, wpg1,
β (1) = T0−1 T1 + T0−1 T2 + T0−1 T3 + T0−1 T4 (β (0) ) + T0−1 T5 ,
where T0 = S/n, T1 = SµS ? /n, T2 = Ss1 +1,n β ? /n, T3 = Ss1 +1,n /n, T4 (β (0) ) = S1,s1 β (0) /n and T5 =
12
√
−1
−1
−1
? − SS ? )λ/n. We will show that, wpg1, kT
(SS21
0 T1 k2 ≤ C2 /4, kT0 T2 k2 ≤ 2 dC1 , kT0 T3 k2 ≤
31
C2 /4, kT0−1 T4 (β (0) )k2 ≤ C2 /4 and kT0−1 T5 k2 ≤ C2 /4. Then, wpg1,
kβ (1) k2 ≤
5
X
√
kT0−1 Ti k2 ≤ 2 dC1 + C2 .
i=1
On T0−1 T1 . For s2 γn /n = o(1), wpg1,
1
s2
1
kT0−1 T1 k2 ≤ k( S)−1 kF k SµS ? k2 ≤ 4kΣ−1
γn → 0.
X kF EkX 0 k2
12
n
n
n
Thus, wpg1, kT0−1 T1 k2 ≤ C2 /4.
On T0−1 T2 . Wpg1,
√
1
1
kT0−1 T2 k2 ≤ k( S)−1 Ss1 +1,n kF kβ ? k2 ≤ 2kI d kF C1 = 2 dC1 .
n
n
On T0−1 T3 . Wpg1,
1
P
kT0−1 T3 k2 ≤ 2kΣ−1
X kF k Ss1 +1,n k2 −→ 0.
n
Thus, wpg1, kT0−1 T3 k2 ≤ C2 /4.
On T0−1 T4 (β (0) ). For s1 /n = o(1),
kT0−1 T4 (β (0) )k2 ≤
s1 √
s1 1 −1 1
P
k( S)
S1,s1 kF kβ (0) k2 ≤ 2 dC2 −→ 0.
n n
s1
n
Thus, wpg1, kT0−1 T4 (β (0) )k2 ≤ C2 /4.
On T0−1 T5 . For s1 λ/n = O(1), wpg1,
kT0−1 T5 k2 ≤ 2kΣ−1
X kF
s1 λ 1
1
P
? k2 + k
? k2 ) −→ 0.
(k SS21
SS31
n
s1
s1
Thus, wpg1, kT0−1 T5 k2 ≤ C2 /4.
vi
Next, consider kβ 2 − β 1 k2 . Since β (1) is bounded wpg1, by Lemma 3.1, A1 occurs wpg1. Then,
β (2) = T0−1 T1 + T0−1 T2 + T0−1 T3 + T0−1 T4 (β (1) ) + T0−1 T5 ,
where T4 (β (1) ) = (1/n)S1,s1 β (1) . Thus, wpg1,
β (2) − β (1) = S−1 S1,s1 (β (1) − β (0) ).
It follows that, for s1 = o(n), wpg1,
√
√
kβ (2) − β (1) k2 ≤ kS−1 S1,s1 kF kβ (1) − β (0) k2 ≤ (2 ds1 /n)(4 dC1 + 2C2 ) → 0.
Then, wpg1, β (2) = β (1) , which means that, wpg1, the iteration algorithm stops at the second
iteration.
Finally, for any K ≥ 1, repeat the above arguments. Then, with at least probability pn,K =
TK
P ( k=0 Ak ), which increases to one by Lemma 3.1, we have
√
√
kβ (K+1) − β (K) k2 ≤ (2 ds1 /n)K (4 dC1 + 2C2 ) = O((s1 /n)K ) → 0,
√
and kβ (k) k2 ≤ 2 dC1 + C2 for all k ≤ K.
Lemma D.3. A necessary and sufficient condition for (µ̂, β̂) to be a minimizer of L(µ, β) is that
it is a solution to equations (2.5) and (2.6).
Proof of Lemma D.3. First, we show a solution of (2.5) and (2.6) satisfies the necessary and sufficient condition in Lemma D.1. Denote a solution of (2.5) and (2.6) as (µ̂, β̂). Then β̂ =
(X T X)−1 X T (Y − µ̂), which is exactly the first condition in Lemma D.1, and, for each i =
1, 2, · · · , n, (µ̂, β̂) satisfies one of three cases: |Yi − X Ti β̂| ≤ λ and µ̂i = 0; Yi − X Ti β̂ > λ and
µ̂i = Yi − X Ti β̂ − λ; Yi − X Ti β̂ < −λ and µ̂i = Yi − X Ti β̂ + λ. If (µ̂, β̂) satisfies the first case,
it satisfies the third condition in Lemma D.1. If (µ̂, β̂) satisfies the second case, then µ̂i > 0 and
Yi − µ̂i − X Ti β̂ = λ = λSign(µ̂i ), which means that the second case satisfies the second condition
in Lemma D.1. Similarly, the third case also satisfies the second condition in Lemma D.1. Thus
(µ̂, β̂) satisfies the necessary and sufficient condition in Lemma D.1.
In the other direction, suppose (µ̂, β̂) satisfies the necessary and sufficient condition in Lemma
D.1. Then, the first condition in Lemma D.1 exactly (2.5). For each i, (µ̂, β̂) satisfies one of three
cases: µ̂i = 0 and |Yi − X Ti β̂| ≤ λ; µ̂i > 0 and Yi − µ̂i − X Ti β̂ = λ; µ̂i < 0 and Yi − µ̂i − X Ti β̂ = −λ.
If (µ̂, β̂) satisfies the first case, it satisfies the first case in (2.6). If (µ̂, β̂) satisfies the second case,
vii
then µ̂i = Yi − X Ti β̂ − λ and Yi − X Ti β̂ > λ, which means that (µ̂, β̂) satisfies the second case
of (2.6). Similarly, If (µ̂, β̂) satisfies the third case, then it satisfies the third case of (2.6). Thus,
(µ̂, β̂) satisfies (2.5) and (2.6).
E
Supplement for Section 3
In this supplement, we provide the proofs of the theoretical results in Section 3. Before that, we
point out that those two different sufficient conditions in Theorem 3.2 come from the different
analysis on the term SµS ? . Each of the two different sufficient conditions does not imply the other.
12
Specifically, on one hand, suppose the absolute values of µ?i ’s are all equal for i = s1 +1, s2 +2, · · · , s.
P
(2+δ)/2 ? 2+δ
Then, kµ?2 k2+δ
= s2
|µs |
and si=s1 +1 |µ?i |2+δ = s2 |µ?s |2+δ . Thus Assumption (C) holds
2
automatically since s2 → ∞. This means that Assumption (C) holds at least when the absolute
magnitudes of µ?i ’s are similar to each other. For this case, there still exists a consistent estimator
even if n/(κn γn ) s2 n. On the other hand, suppose µ?s = γn and the other µ?i ’s are all equal to
P
= [γn2 +(s2 −1)c2 ](2+δ)/2 and si=s1 +1 |µ?i |2+δ = γn2+δ +(s2 −1)c2+δ .
a constant c > 0. Then, kµ?2 k2+δ
2
If s2 γn2 n/(κn γn ), the previous two terms are both asymptotically equivalent to γn2+δ . Thus
Assumption (C) fails but the other sufficient condition holds.
Proof of Lemma 3.1. The proof is the similar to that of Lemma 4.1 and omitted.
Proof of Theorems 3.2. By Lemma 3.1, wpg1, the solution β̂ n to ϕn (β) = 0 on BC (β ? ) is explicitly
given by
β̂ n = β ? + T0−1 (T1 + T2 + T3 − T4 ),
? and T4 = (λ/n)SS ? .
where T0 = (1/n)Ss1 +1,n , T1 = (1/n)SµS ? , T2 = (1/n)Ss1 +1,n , T3 = (λ/n)SS21
31
12
P
P
We will show that T0 −→ Σ−1
X > 0 with the Frobenius norm and Ti −→ 0 with the Euclidean norm
for i = 1, 2, 3, 4. Thus, by Slutsky’s lemma (see, for example, Lemma 2.8 on page 11 of van der
Vaart (1998)), β̂ n is a consistent estimator of β ? .
P
On T0−1 . By law of large number, T0 −→ ΣX > 0. Then, by continuous mapping theorem,
P
T0−1 −→ Σ−1
X > 0.
On T1 : Approach One. Suppose s2 = o(n/(κn γn )). Then,
kT1 k2 ≤
s
s
1 X
1 X
kX i µ?i k2 =
kX i k2 · |µ?i | ≤ s2 κn γn /n = o(1).
n
n
i=s1 +1
i=s1 +1
viii
On T1 : Approach Two. Under Assumption (C), it follows (ΣX
?2 −1/2 Sµ
?
i=s1 +1 µi )
S12
Ps
d
−→
N (0, Id ). In fact, Assumption (C) implies the Lyapunov condition for sequence of random vectors
(see, e.g. Proposition 2.27 on page 332 of (van der Vaart, 1998)). More specifically, recall the
Lyapunov condition is that there exists some constant δ > 0 such that
s
X
s
X
Ek(ΣX
i=s1 +1
−1/2
µ?2
X i µ?i k2+δ
→ 0.
j )
2
j=s1 +1
Then, by Assumption (C),
s
X
Ek(ΣX
i=s1 +1
s
X
1
−2
µ?2
X i µ?i k2+δ
≤(
j )
2
j=s1 +1
s
X
−
µ?2
j )
2+δ
2
j=s1 +1
s
X
− 2+δ
−→ 0,
|µ?i |2+δ λmin2 EkX0 k2+δ
2
i=s1 +1
where λmin > 0 is the minimum eigenvalue of ΣX . Then,
s
s
X
X
1 µ
1
?2 1/2
−1/2 µ
kT1 k2 = k SS ? k2 ≤ k(ΣX
µi ) kF k(ΣX
µ?2
SS ? k2
i )
12
n 12
n
i=s1 +1
=
1
(
n
s
X
i=s1 +1
1/2
1/2
µ?2
kΣX kF OP (1) ≤
i )
i=s1 +1
1
1
(s2 γn2 )1/2 OP (1) ≤ √ γn OP (1) = oP (1).
n
n
On T2 . By law of large number, T2 = oP (1).
√
On T3 and T4 . By noting λ n,
√
s1 1
1
λ
?
? k2 ≤ √ OP (1) = oP (1).
kT3 k2 = kλ SS21 k2 = λ
k √ SS21
n
n
s1
n
Thus T3 = oP (1). In the same way, we can show that T4 = oP (1) holds.
In Theorem, one condition is Dn /n = o(1). It turns out we can consider other conditions on
Dn and derive more possible asymptotic distributions for β̂ n .
Theorem E.1 (Asymptotic Distributions on β̂ n : more cases). Under Assumptions (A), (B) and
(C), for all constants b, c ∈ R+ ,
√
d
(1) when s1 n/λ2 and Dn2 /n = o(1), n(β̂ n − β ? ) −→ N (0, σ 2 Σ−1
X ); [main case]
√
d
(2) when s1 n/λ2 and Dn2 /n ∼ c, n(β̂ n − β ? ) −→ N (0, (c + σ 2 )Σ−1
X );
d
√
d
√
(3) when s1 n/λ2 and Dn2 /n → ∞, rn (β̂ n − β ? ) −→ N (0, Σ−1
X ), where rn ∼ n/Dn
√
d
(4) when s1 ∼ bn/λ2 and Dn2 /n = o(1), n(β̂ n − β ? ) −→ N (0, (b + σ 2 )Σ−1
X );
√
d
(5) when s1 ∼ bn/λ2 and Dn2 /n ∼ c, n(β̂ n − β ? ) −→ N (0, (b + c + σ 2 )Σ−1
X );
(6) when s1 ∼ bn/λ2 and Dn2 /n → ∞, rn (β̂ n − β ? ) −→ N (0, Σ−1
X ), where rn ∼ n/Dn
ix
n;
n;
d
(7) when s1 n/λ2 and Dn2 /n = o(1) or Dn2 /n ∼ c, rn (β̂ n − β ? ) −→ N (0, Σ−1
X ), where rn ∼
√
√
n/(λ s1 ) n;
√
√
√
(8) when s1 n/λ2 and Dn2 /n → ∞, letting rn ∼ min{ bn/(λ s1 ), n/Dn } n,
√
√
d
bn/(λ s1 ) n/Dn , then rn (β̂ n − β ? ) −→ N (0, Σ−1
X );
√
√
d
(8b) if bn/(λ s1 ) ∼ n/Dn , then rn (β̂ n − β ? ) −→ N (0, (1 + b)Σ−1
X );
√
√
d
(8c) if bn/(λ s1 ) n/Dn , then rn (β̂ n − β ? ) −→ N (0, bΣ−1
X ).
(8a) if
Theorem 3.3 groups the results according to the asymptotic magnitude of s1 given an upper
bound of the diverging speed of s2 . Alternatively, Theorem E.1 groups the results according to
the asymptotic magnitudes of s1 and Dn2 . Since both s1 and Dn2 have three cases, Theorem E.1
basically contains nine cases. For the last case, there are further three cases on the relationship
√
√
between bn/(λ s1 ) and n/Dn . As in Theorem 3.3, the first case of Theorem E.1 is denoted as
the main case since for this case the incidental parameters are sparse in the sense that the size
and magnitude of the nonzero incidental parameters µ?1 and µ?2 are well controlled. Note that
√
s2 = o( n/(κn γn )) implies Dn2 /n = o(1). which means that, under Assumption (C), the cases (1),
(4) and (7) of Theorem E.1 actually imply the three results of Theorem 3.3. As in Theorem 3.3,
√
the convergence rate of β̂ n becomes less than n when s1 n/λ2 or Dn2 /n → ∞, that is, when
the size and magnitude of the nonzero incidental parameters are large; the boundary phenomenon
also appears.
Proof of Theorems 3.3 and E.1. It is sufficient to provide the proof for the case where the sizes of
? = {1 ≤ i ≤ s : µ? > 0} and S ? = {1 ≤ i ≤ s : µ? < 0} are both asymptotically
index sets S21
1
1
31
i
i
s1 /2 and b = 2.
From the proof of Theorems 3.2, wpg1,
µ
?
?
β̂ = β ? + S−1
s1 +1,n [SS ? + Ss1 +1,n + λ(SS21 − SS31 )].
12
Let rn be a sequence going to infinity. Then, rn (β̂ n −β ? ) = T0−1 (V1 +V2 +V3 −V4 ), where V1 = rn T1 ,
V2 = rn T2 , V3 = rn T3 , V4 = rn T4 and Ti ’s are defined in the proof of Theorem 3.2. Next we derive
the asymptotic properties of T0 and Vi ’s, from which the desired results follow by Slutsky’s lemma.
P
On T0 . By the proof of Theorem 3.2, T0−1 −→ Σ−1
X
√
√
On V1 : Approach One. If rn = n and s2 = o( n/(κn γn )), then
s
1 µ
1 X
1
1
kT1 k2 = krn SS ? k2 ≤ rn
kX i k2 · |µ?i | ≤ rn s2 κn γn = √ s2 κn γn = o(1).
12
n
n
n
n
i=s1 +1
x
√
n and s2 = o( n/(κn γn )), then T1 = oP (1).
√
On V1 : Approach Two. If rn = n, then
Thus, if rn =
√
n or rn
√
1
Dn 1 µ
Dn 1 µ
T1 = rn SµS ? = rn
SS ? = √
S ?,
n 12
n Dn 12
n Dn S12
P
1/2 . There are three cases on D /√n or D 2 /n. If D 2 /n → 0,
where Dn = kµ?2 k2 = ( si=s1 +1 µ?2
n
n
n
i )
√
d
P
2
2
then T1 −→ 0. If Dn /n → 1, then T1 −→ N (0, ΣX ). If Dn /n → ∞, it means that rn = n is too
√ p
√
d
fast. Let rn ∼ n/Dn = n n/Dn2 n. Then T1 −→ N (0, ΣX );
√
√
√
d
P
On V2 . If rn = n, then T2 −→ N (0, σ 2 ΣX ). Thus, if rn n, T2 −→ 0; if rn n;
P
T2 −→ ∞.
On V3 and V4 . First consider T3 . Denote #(·) as the size function. If rn =
r
s1 /2
1
1
? = λ
p
S ?.
T3 = λrn SS21
? ) S21
n
n
#(S21
√
n, then
p
p
P
? ) = s /2. There are three cases on λ s /(2n). If λ s /(2n) → 0, then T −→
Note that #(S21
0.
1
1
1
3
p
p
d
Note that λ s1 /(2n) → 0 is equivalent to s1 = o(2n/λ2 ). If λ s1 /(2n) → 1, then T3 −→
p
p
N (0, ΣX ). Note that λ s1 /(2n) → 1 is equivalent to s1 ∼ 2n/λ2 . If λ s1 /(2n) → ∞, it means
p
√
√ √
√
√
rn = n is too large. Let rn ∼ n/(λ (s1 /2)) = n 2n/(λ s1 ) n. With this rate rn ,
p
d
T3 −→ N (0, ΣX ). Note that λ s1 /2n → ∞ is equivalent to s1 O(2n/λ2 ). In the same way, T4
can be analyzed and parallel results can be obtained.
Proof of Theorem 3.4. The proof is similar to that of Theorem 4.4 and omitted.
E.1
Supplement for Subsection 3.1
The following Theorem implies Theorem 3.5 since it contains more details.
Theorem E.2 (Consistency and Asymptotic Normality on β̃). Suppose Assumptions (A) and (B)
√
P
hold. If either s2 = o(n/(κn γn )) or Assumption (C) holds, then β̃ −→ β ? . If s2 = o( n/(κn γn )),
√
d
then n(β̃ − β ? ) −→ N (0, σ 2 Σ−1
X ). On the other hand, under Assumption (C),
(1) if Dn2 /n = o(1), then
(2) if Dn2 /n ∼ c, then
√
d
n(β̃ − β ? ) −→ N (0, σ 2 Σ−1
X ); [main case]
√
d
+
n(β̃ − β ? ) −→ N (0, (c + σ 2 )Σ−1
X ), for every constant c ∈ R ;
d
(3) if Dn2 /n → ∞, then rn (β̃ − β ? ) −→ N (0, Σ−1
X ) where rn ∼ n/Dn
xi
√
n.
Proof of Theorem E.2. Denote I0 = {s1 + 1, s1 + 2, · · · , s = s1 + s2 , s + 1, · · · , n}. Note that
√
s2 = o( n/(κn γn )) ensures that β̂ is consistent by Theorem 3.2. By Theorem 3.4, P {Iˆ0 = I0 } goes
to 1. Then,
β̃ = R1 + R2 + T0−1 (T1 + T2 ),
where R1 = (X TIˆ X Iˆ0 )−1 X TIˆ Y Iˆ0 {Iˆ0 6= I0 } and R2 = −(X TI0 X I0 )−1 X TI0 Y I0 {Iˆ0 6= I0 } and Ti ’s are
0
0
defined in the proof of Theorem 3.2. The proof for the consistency is similar to that of Theorem
3.2 and is omitted. Next we show the asymptotic normality. We have,
rn (β̃ − β ? ) = rn R1 + rn R2 + T0−1 (V1 + V2 ),
√
where Vi ’s are defined in the proof of Theorem E.1. Since P ( nR1 = 0) ≥ P {Iˆ0 = I0 } → 1, we
√
√
have nR1 = oP (1). Similarly, nR2 = oP (1). From the analysis on Vi ’s in the proof of Theorem
E.1, the asymptotic distributions follows by Slutsky’s lemma.
Lemma E.3 (Consistency on σ̂). Suppose Assumptions (A) and (B) hold and either s2 = o(n/(κn γn ))
P
or Assumption (C) holds. If s2 = o(n/γn2 ), then σ̂ −→ σ.
Proof of Lemma E.3. When Assumption (C) or s2 = o(n/(κn γn )) holds, the penalized estimators β̂
and β̃ are consistent estimators of β ? by Theorems 3.2 and 3.5. Denote C = {Iˆ0 = I0 }. By Theorem
3.4, C occurs wpg1. Then, σ̂ 2 = T C + σ̂ 2 C c , where T = an kY I0 − X TI0 β̃k22 and an = 1/(n − s1 ). It
P
P
P
is sufficient to show T −→ σ 2 . We have T = 6i=1 Ti , where T1 = an ni=s1 +1 [X Ti (β ? − β̃)]2 , T2 =
Ps
P
P
P
?
T
an ni=s1 +1 2i , T3 = 2an ni=s1 +1 X Ti (β ? − β̃)i , T4 = an si=s1 +1 µ?2
i , T5 = 2an
i=s1 +1 µi X i (β −
P
P
P
β̃) and T6 = 2an si=s1 +1 µ?i i . It is straightforward to show that T2 −→ σ 2 and each other Ti −→ 0
P
under the condition s2 = o(n/γn2 ) and by noting that β̃ −→ β ? . Then σ̂ is a consistent estimator
of σ.
E.2
Supplement for Subsection 3.2
In this supplement, we consider a special case with exponentially tailed covariates and errors. For
convenience, we first introduce the definition of Orlicz norm and related inequalities. For a strictly
increasing and convex function ψ with ψ(0) = 0, the Orlicz norm of a random variable Z with
respect to ψ is defined as
kZkψ = inf{C > 0 : Eψ(|Z|/C) ≤ 1}.
Then, for each x > 0,
P (|Z| > x) ≤ 1/ψ(x/kZkψ ).
xii
(E.1)
(See Page 96 of van der Vaart and Wellner (1996)). Next, we introduce a lemma on Orlicz norm with
ψ1 . Suppose {Zi }ni=1 is a sequence of random variables and {Z i }ni=1 is a sequence of d-dimensional
random vectors with Z i = (Zi1 , Zi2 , · · · , Zid )T . From Lemma 8.3 on Page 131 of Kosorok (2008),
we have the following extension.
Lemma E.4. If for each 1 ≤ i ≤ n and 1 ≤ j ≤ d,
x2
1
x2
1
} and P (|Zij | > x) ≤ c exp{− ·
},
P (|Zi | > x) ≤ c exp{− ·
2 ax + b
2 ax + b
with a, b ≥ 0 and c > 0, then
k max |Zi k|ψ1 ≤ K{a(1 + c) log(1 + n) +
1≤i≤n
p
p
b(1 + c) log(1 + n)},
p
p
√
k max kZ i k2 kψ1 ≤ K{a d(1 + cd) log(1 + n) + bd(1 + cd) log(1 + n)}.
1≤i≤n
where K is a universal constant which is independent of a, b, c, {Zi } and {Z i }.
Proof of Lemma E.4. The proof for random variables {Zi } is the same to the proof of Lemma 8.3
on Page 131 of Kosorok (2008). For random vectors {Z i },
d
X
√
√
1 x2
P (kZ i k2 ≥ x) ≤ P ( max |Zij | > x/ d) ≤
P (|Zij | > x/ d) ≤ c0 exp{− 0
},
1≤j≤d
2 a x + b0
j=1
√
where a0 = a d, b0 = bd and c0 = cd. Then, by the result on random variables, the desired result
on random vectors follows.
Now, suppose, for every x > 0,
1
x2
1
x2
P (|i | > x) ≤ c1 exp{− ·
} and P (|Xij | > x) ≤ c2 exp{− ·
},
2 a1 x + b1
2 a2 x + b2
(E.2)
with ai , bi ≥ 0 and ci > 0 for i = 1, 2. By Lemma E.4, it follows
k max |i k|ψ1 ≤ K{a1 (1 + c2 ) log(1 + n) +
1≤i≤n
p
b1 (1 + c1 )
p
log(1 + n)},
p
p
√
k max kX i k2 kψ1 ≤ K{a2 d(1 + c2 d) log(1 + n) + b2 d(1 + c2 d) log(1 + n)}.
1≤i≤n
p
Thus, from the inequality (E.1), if a1 > 0, let γn log(n); otherwise, let γn log(n).
p
Similarly, if a2 > 0, let κn log(n); otherwise, let κn log(n). Then, such γn and κn satisfy
the condition (2.2). Suppose both a1 and a2 are positive, which means both i and Xij ’s have
exponential tails. As before, set κn = γn = log(n)τn . For this case, the regularization parameter
√
specification (2.4) becomes log(n)τn λ min{µ? , n}.
xiii
At the end of this supplement, we simply list explicit expressions of κn under different assumptions on the covariates for the case with a diverging number of covariates, which are the extension
of the results in Section 3.2. The magnitude of κn becomes larger than that for the case with d
√
fixed while γn keeps the same. Specifically, if X 0 is bounded with CX > 0, then κn = dCX .
q
2 [(3/2) log(d) + log(n)]. If the
If X 0 follows a Gaussian distribution N (0, ΣX ), then κn = 2dσX
P
Orlicz norm kX0j kψ exists for 1 ≤ j ≤ d and their average (1/d) dj=1 kX0j kψ is bounded, then
κn dψ −1 (n); for instance, if ψ = ψp with p ≥ 1, then κn d(log(n))1/p . Finally, if the data {X i }
satisfies the right inequality of (E.2) with a2 > 0, that is, each component of X i is sub-exponentially
tailed, then κn d3/2 log(n). It is worthwhile to note that these expressions of κn depend on a
factor involving the diverging number of covariates d, which will influence the specification of the
regularization parameter and the sufficient conditions of all the theoretical results in Section 4.
F
Supplement for Section 4
In this supplement, we provide Proposition D.2 and its proof, the proofs of the lemmas in the
appendix and some additional results.
We first extend Proposition D.2 to the case with d → ∞ and d n. Before that, we list two
simple lemmas for a diverging d. Suppose {ξ i } is a sequence of i.i.d. copies of ξ 0 , a d-dimensional
P
random vector with mean zero. Denote σ̄ξ2 = (1/d) dj=1 Var[ξ0j ].
Lemma F.1. Suppose σ̄ξ2 is bounded. If d/n = o(1), then
n
k
1X
P
ξ i k2 −→ 0.
n
i=1
Lemma F.2. Suppose σ̄ξ2 is bounded. If d/n = o(1), then
n
1X
P
kξ i k2 − P kξ 0 k2 −→ 0.
n
i=1
Suppose the specification of the regularization parameter is given by
dκn λ, αγn ≤ λ, and λ µ? ,
(F.1)
where α is a constant greater than 2.
Proposition F.3. Suppose assumptions (D) and (G) hold and the regularization parameter satisfies
√
√
(F.1). Suppose there exist constants C1 and C2 such that kβ ? k2 < C1 d and kβ (0) k2 < C2 d
xiv
√
√
wpg1. If the regularization parameter satisfies (2.4), s1 λκn /(n d) = o(1) and s2 κn γn /(n d) =
o(1), then, for every K ≥ 1, with at least probability pn,K which increases to one as n → ∞,
√
kβ (K+1) − β (K) k2 ≤ O(( ds1 κ2n /n)K d) and kβ (k) k2 ≤ (2C1 + C2 )d for all k ≤ K. Specifically,
wpg1, the iterative algorithm stops at the second iteration.
Proof of Proposition F.3. Reuse the notations in the proof of Lemma D.2. First, we show that,
wpg1, kβ (1) k2 ≤ (2C1 + C2 )d. For each k ≥ 1,
Sβ (k) = SµS11 + SµS12 + SS1 β ? + SS1 + SS2 ∪S3 β (k−1) + λ(SS2 − SS3 ),
Since the regularization parameter satisfies (F.1), it is easy to check that the conclusion of Lemma
4.1 continues to hold, which implies P (A0 ) → 1.
Thus, wpg1,
β (1) = T0−1 T1 + T0−1 T2 + T0−1 T3 + T0−1 T4 (β (0) ) + T0−1 T5 .
We will show that, wpg1,
kT0−1 T1 k2 ≤ (C2 /4)d,
kT0−1 T2 k2 ≤ 2C1 d,
kT0−1 T3 k2 ≤ (C2 /4)d,
kT0−1 T4 (β (0) )k2 ≤ (C2 /4)d,
kT0−1 T5 k2 ≤ (C2 /4)d.
Thus, wpg1,
kβ (1) k2 ≤
5
X
kT0−1 Ti k2 ≤ (2C1 + C2 )d.
i=1
√
On T0−1 T1 . Under Assumption (D), for s2 κn γn /(n d) = o(1), wpg1,
1
1
s2
kT0−1 T1 k2 ≤ k( S)−1 kF k SµS ? k2 ≤ 2kΣ−1
X kF,d √ κn γn d → 0.
n
n 12
n d
Thus, wpg1, kT0−1 T1 k2 ≤ C2 d/4.
On T0−1 T2 . Wpg1,
1
1
kT0−1 T2 k2 ≤ k( S)−1 Ss1 +1,n kF kβ ? k2
n
n
√
√
1
1
≤ kI d kF C1 d + k( S)−1 S1,s1 kF C1 d
n
n
√
1
1 −1
≤ C1 d + k( S) kF k S1,s1 kF C1 d,
n
n
xv
and
s
1
1X
s1
1
kX i k22 ≤ κ2n .
k S1,s1 kF =
n
n
n
i=1
Thus, Under Assumption (D), for s1 κ2n /n = o(1), wpg1,
√
ds1 2 √
−1
−1
kT0 T2 k2 ≤ C1 d + 2kΣX kF,d
κn C1 d ≤ 2C1 d.
n
On T0−1 T3 . Under assumptions (D) and (G), for log(d)/n = o(1), wpg1,
√
1 p
1
1
d√
d log(d)k( S)−1 kF,d (d log(d))−1/2 k √ Ss1 +1,n k2
n
n
n
p
d log(d)
P
√
≤
2kΣ−1
X kF,d OP (1) −→ 0.
n
kT0−1 T3 k2 =
Thus, wpg1, kT0−1 T3 k2 ≤ C2 d/4.
On T0−1 T4 (β (0) ). Under Assumption (D), for s1 κ2n /n, wpg1,
√
1
1
dk( S)−1 kF,d k S1,s1 kF kβ (0) k2
n
n
√
s1 2 √ P
−1
≤ d2kΣX kF,d κn C2 d −→ 0.
n
kT0−1 T4 (β (0) )k2 ≤
Thus, wpg1, kT0−1 T4 (β (0) )k2 ≤ C2 d/4.
√
On T0−1 T5 . Under Assumption (D), for s1 κn λ/(n d) = o(1), wpg1,
√
λ
1
? k2 + kSS ? k2 )
dk( S)−1 kF,d (kSS21
31
n
n
√
λ
≤ d2kΣ−1
X kF,d s1 κn ≤ C2 d/4.
n
kT0−1 T5 k2 ≤
Next, consider kβ 2 − β 1 k2 . Since β (1) ≤ (2C1 + C2 )d wpg1, the conclusion of Lemma 4.1 holds,
which implies A1 occurs wpg1.
Then,
β (2) = T0−1 T1 + T0−1 T2 + T0−1 T3 + T0−1 T4 (β (1) ) + T0−1 T5 ,
where
T4 (β (1) ) =
1
S1,s1 β (1) .
n
Thus, wpg1,
β (2) − β (1) = S−1 S1,s1 (β (1) − β (0) ).
xvi
Thus, for d3/2 s1 κ2n /n = o(1), wpg1,
√
1
1
dk S−1 kF,d k S1,s1 kF kβ (1) − β (0) k2
n
n
√ s1 2
−1
≤ 2kΣX kF,d d κn (2C1 + C2 )d . d3/2 s1 κ2n /n → 0.
n
kβ (2) − β (1) k2 ≤
Thus, wpg1, β (2) = β (1) , which means that, wpg1, the iteration algorithm stops at the second
iteration.
T
For any K ≥ 1, repeating the above arguments, with at least probability pn,K = P ( K
k=0 Ak ),
which increases to one, we have β (k) ≤ (2C1 + C2 )d for k ≤ K and
√ s1 2 K
√
kβ (K+1) − β (K) k2 ≤ (2kΣ−1
κn ) (2C1 + C2 )d . ( ds1 κ2n /n)K d → 0.
X kF,d d
n
This completes the proof.
Next, we provide the proofs of Lemmas A.2, A.3 and A.4 in the appendix.
√
P
Proof of Lemma A.2. Let E = An − A. Note that rd ≥ 1/ d. Then, rd kEkF −→ 0 implies
P
kEkF,d −→ 0. Thus, wpg1, kEkF,d is bounded by a constant C > 0. By Lemma A.1,
−1
−1
kA−1
n − A kF,d ≤ kA kF,d
kEkF,d
kA−1 kF,d kEkF,d
≤ C2
.
−1
1 − CkEkF,d
1 − kA kF,d kEkF,d
Therefore,
rd kEkF
P
−→ 0.
1 − CkEkF,d
−1
2
rd kA−1
n − A kF ≤ C
This completes the proof.
Proof of Lemma A.3. For any δ > 0, we have
P (kΣ̂n − ΣX kF > δ) ≤
d X
d
X
d2
k=1 l=1
δ
n
P(
2
1X
d4 1 2
Xik Xil − σkl )2 ≤
σ̄ .
n
n δ 2 XX
i=1
2 r 2 d4 /(nδ 2 ) = o(1) by Assumption (E2) and for r 2 d4 /n → 0.
Thus, P (rd kΣ̂n − ΣX kF > δ) ≤ σ̄XX
d
d
Thus, Σ̂n is a consistent estimator of ΣX wrt rd k·kF .
Proof of Lemma A.4. Let αd =
n
√
d log d and C1 ≥
√
d
2σξ,max . Then
n
X
1 X
1 X ξij
αd C1
P (k √
ξ i k2 > αd C1 ) ≤
P (| √
| > √ ),
σ
n
n
σj d
i=1
j=1
i=1 j
xvii
where σj is the standard deviation of ξ0j . By Berry and Esseen Theorem (see, for example, P375
√ P
in Shiryaev (1995)), there exists a constant C2 > 0 such that P (k(1/ n) ni=1 ξ i k2 > αd C1 ) ≤
T1 + 2T2 , where
T1 =
d
X
P (|N (0, 1)| >
j=1
αd C1
√ ),
σj d
T2 =
d
X
C2 E|ξ0j |3
√ .
σj3 n
j=1
By noting d2 = o(n),
√
σξ,max d
αd C1
αd C1
√ ) < 2d
√ ) → 0,
φ(
T1 ≤
P (|N (0, 1)| >
α d C1
σξ,max d
σξ,max d
j=1
d
X
T2 ≤
d
X
C2 γξ,max
C2 γξ,max
√ = d 3 √ → 0.
3
σ
n
σmin n
j=1 ξ,min
√ P
Therefore, k(1/ n) ni=1 ξ i k2 = OP (αd ).
Next result is on the consistency of the penalized two-step estimator β̃.
Theorem F.4 (Consistency on β̃). Suppose the assumptions and conditions of Theorem 4.2 hold.
√
P
If rd ≥ 1/ d, then β̃ −→ β ? wrt rd k·k2 .
P
Proof of Theorems F.4. By Theorem 4.2, β̂ −→ β ? wrt rd k·k2 . By Theorem 4.4, P {Iˆ0 = I0 } → 1
√
for rd ≥ 1/ d, where I0 = {s1 + 1, s1 + 2, · · · , s = s1 + s2 , s + 1, · · · , n}. Then, wpg1,
β̃ − β ? = R1 + R2 + T0−1 T1 + T0−1 T2 ,
where R1 = (X TIˆ X Iˆ0 )−1 X TIˆ Y Iˆ0 {Iˆ0 6= I0 }, R2 = −(X TI0 X I0 )−1 X TI0 Y I0 {Iˆ0 6= I0 } and Ti ’s are
0
0
defined in the proof of Theorem 4.2. Then,
√
√
rd kβ̃ − β ? k2 ≤ rd kR1 k2 + rd kR2 k2 + kT0−1 kF,d rd dkT1 k2 + kT0−1 kF,d rd dkT2 k2 .
Since P (kR1 k2,d = 0) ≥ P {Iˆ0 = I0 } → 1, we have R1 = oP (1). Similarly, R2 = oP (1). By the proof
√
P
P
of Theorem 4.2, kT0−1 kF,d is bounded and rd dkTi k2 −→ 0 for i = 1, 2. Thus, β̃ −→ β ? wrt rd k·k2
√
and rd ≥ 1/ d.
Finally, we provide some additional results on the asymptotic distributions of β̂ and β̃ with a
√
different scaling. Specifically, the scaling in Section 4 is nAn . Next, we consider another natural
√
1/2
scaling nAn ΣX .
xviii
Theorem F.5 (Asymptotic Distribution on β̂). Suppose assumptions (D’), (D”), (E), (F) and
√
√
(G) hold. If d6 log d = o(n), s1 = o( n/(λdκn )) and s2 = o( n/(dκn γn )), then
√
d
1/2
nAn ΣX (β̂ n − β ? ) −→ N (0, σ 2 G).
Theorem F.6 (Asymptotic Distribution on β̃). Suppose the assumptions and conditions of Theo√
rem F.5 hold except the condition s1 = o( n/(λdκn )). Then
√
d
1/2
nAn ΣX (β̃ − β ? ) −→ N (0, σ 2 G).
By Theorems F.5 and F.6, Wald-type confidence regions can be constructed. In order to validate
these confidence regions with estimated σ and ΣX , we need Lemma A.5 and the following result.
Theorem F.7 (Asymptotic Distributions on β̂ and β̃ with Σ̂n ). Suppose the assumptions and
conditions of Theorem F.5 hold. If d9 (log(d))2 = o(n), then
√
1/2
d
nAn Σ̂n (β̂ − β ? ) −→ N (0, σ 2 G).
Similarly, suppose the assumptions and conditions of Theorem F.6 hold. If d9 (log(d))2 = o(n), then
√
1/2
d
nAn Σ̂n (β̃ − β ? ) −→ N (0, σ 2 G).
Remark 3. A comparison of the assumptions and conditions of Theorem F.7 with those of Theorems F.5 and F.6 reveals that a much stronger requirement on d is needed to ensure Σ̂n is a good estimator of ΣX . Precisely, the former require that d9 (log(d))2 = o(n) and the latter d6 log(d) = o(n).
This stronger requirement on d is a price paid for estimating ΣX .
Remark 4. The condition on the dimension d in Theorems 4.3 and 4.5 is d5 log(d) = o(n), slightly
weaker than the condition d6 log(d) = o(n) in Theorems F.5 and F.6. Accordingly, The condition
on the dimension d in Theorem A.6 is d8 (log(d))2 = o(n), slightly weaker than the condition
√
d9 (log(d))2 = o(n) in Theorem F.7. This means that the scaling nAn is slightly better than the
√
1/2
scaling nAn ΣX in terms of the condition on d. Further, the former scaling is more suitable for
constructing confidence regions for some entries of β ? .
At the end of this supplement, we provide the proofs of the above theorems.
Proof of Theorems F.5. Reuse the notations Ti ’s in the proof of Theorems 4.2, from which,
√
1/2
nAn ΣX (β̂ n − β ? ) = V1 + V2 + V3 − V4 ,
xix
where Vi = B n Ti for i = 1, 2, 3, 4 and B n =
√
d
1/2
nAn ΣX T0−1 . We will show V2 −→ N (0, σ 2 G) and
other Vi ’s are oP (1), from which the desired result follows by applying Slutsky’s lemma.
√
1/2
On V1 . We have kV1 k2 ≤ ndkAn kF kΣX kF,d kT0−1 kF,d kT1 k2 . By Assumption (F), kAn kF is
1/2
bounded. By Assumption (D”), kΣX kF,d is bounded. By Lemmas A.2 and A.3 and Assumption
(D), for d = o(n1/3 ), wpg1, kT0−1 kF,d is bounded. Further, wpg1, kT1 k2 ≤
kV1 k2 .
1
n s2 κn γn .
Then,
√1 s2 dκn γn ,
n
where . means that the left side is bounded by a constant times the right
√
side, as noted at the beginning of the appendix. Thus, kV1 k2 = oP (1) for s2 = o( n/(dκn γn )).
On V2 . We have V2 = V21 + V22 , where
V21 =
√
−1/2
nAn ΣX
First, consider V21 . We have V21 =
T2 ,
V22 =
1/2
nAn ΣX (T0−1 − Σ−1
X )T2 .
p
P
(n − s1 )/n ni=s1 +1 Z n,i , where
Z n,i = √
On one hand, for every δ > 0,
√
1
−1/2
An ΣX X i i .
n − s1
Pn
2
i=s1 +1 EkZ n,i k2 {kZ n,i k2
> δ} ≤ (n − s1 )EkZ n,0 k42 /δ 2 and
1
−1/2
−1/2
E4 E(X T0 ΣX ATn An ΣX X 0 )2
(n − s1 )2 0
1
≤
E4 λmax (Gn )λmin (ΣX )−1 E(X T0 X 0 )2
(n − s1 )2 0
EkZ n,0 k42 =
d
≤
1X
d2
4 1/2 2
E40 λmax (Gn )λmin (ΣX )−1 (
(EX0j
) ) .
2
(n − s1 )
d
j=1
P
√
Thus, by assumptions (D’), (E) and (F), ni=s1 +1 EkZ n,i k22 {kZ n,i k2 > δ} → 0 for d = o( n). On
P
the other hand, ni=s1 +1 Cov(Z n,i ) = σ 2 An ATn → σ 2 G. Thus, by central limit theorem (see, for
d
example, Proposition 2.27 in van der Vaart (1998)), V21 −→ N (0, σ 2 G). Next, consider V22 . We
have
1/2
−1/2 √
kV22 k2 ≤ kAn kF kΣX kF,d d(log(d))1/2 kT0−1 − Σ−1
k nT2 k2 .
X kF (d log(d))
1/2
By Assumption (F), kAn kF is O(1); By Assumption (D”), kΣX kF,d is O(1); by Lemmas A.2
6
and A.3, d(log(d))1/2 kT0−1 − Σ−1
X kF is oP (1) for d log(d) = o(n); By Lemma A.4, together with
√
√
Assumption (G), (d log(d))−1/2 k nT2 k2 = (d log(d))−1/2 k √1n Ss1 +1,n k2 is OP (1) for d = o( n).
P
d
Thus, V22 −→ 0. By slutsky’s lemma, V2 −→ N (0, σ 2 G).
√
On V3 and V4 . First consider V3 . By noting that s1 = o( n/(λdκn )), wpg1, kV3 k2 ≤
√
√
1/2
d nkAn kF kΣX kF,d kT0−1 kF,d kT3 k2 . dλs1 κn / n → 0. Thus, kV3 k2 = oP (1). In the same way,
kV4 k2 = oP (1). This completes the proof.
xx
√
1/2
Proof of Theorem F.6. From the proof of Theorem F.4, we have nAn ΣX (β̃ − β ? ) = R̃1 + R̃2 +
√
√
1/2
1/2
V1 + V2 , where R̃1 = nAn ΣX R1 , R̃2 = nAn ΣX R2 , and Ri ’s and Vi ’s are defined in the
proofs of Theorems F.4 and F.5. Since P (kR̃1 k2 = 0) ≥ P {Iˆ0 = I0 } → 1, we have R̃1 = oP (1).
d
Similarly, R̃2 = oP (1). By the proof of Theorem F.5, V1 = oP (1) and V2 −→ N (0, σ 2 G). Thus, the
asymptotic distribution of β̃ is Gaussian by Slutsky’s lemma.
Proof of Theorem F.7. We only show the result on β̂. since the result on β̃ can be obtained in a
similar way. We reuse the definitions of Ti ’s in the proof of Theorems 4.2, from which,
√
where M =
√
1/2
nAn Σ̂n (β̂ n − β ? ) = M + R,
1/2
nAn ΣX (β̂ n − β ? ) and R =
√
1/2
nAn (Σ̂n
d
1/2
− ΣX )(β̂ n − β ? ). By Theorem F.5,
P
M −→ N (0, σ 2 G). Then, it is sufficient to show that R −→ 0 wrt k·k2 . We have
R = R1 + R2 + R3 − R4 ,
where Ri = B n Ti for i = 1, 2, 3, 4 and B n =
√
1/2
nAn (Σ̂n
1/2
− ΣX )T0−1 . We will show each Ri
converges to zero in probability, which finishes the proof.
1/2
1/2
On R1 . By Lemma A.7, kΣ̂n − ΣX kF ≤ (d1/2 kΣ̂n − ΣX kF )1/2 . Then,
kR1 k2 ≤
≤
√
√
1/2
1/2
nkAn kF kΣ̂n − ΣX kF kT0−1 kF kT1 k2
ndkAn kF (kΣ̂n − ΣX kF,d )1/2 kT0−1 kF,d kT1 k2 .
By Assumption (F), kAn kF is bounded. By Lemma A.3, kΣ̂n − ΣX kF,d = oP (1) for d = o(n1/3 ).
By Lemmas A.2 and A.3 and Assumption (D), for d = o(n1/3 ), wpg1, kT0−1 kF,d is bounded.
We have, wpg1, kT1 k2 ≤
√
o( n/(dκn γn )).
1
n s2 κn γn .
Then, kR1 k2 .
√1 s2 dκn γn .
n
Thus, kR1 k2 = oP (1) for s2 =
On R2 . We have
√
1/2
1/2
kR2 k2 ≤ kAn kF d(log(d))1/2 kΣ̂n − ΣX kF kT0−1 kF,d (d log(d))−1/2 k nT2 k2 ,
and
1/2
1/2
d(log(d))1/2 kΣ̂n − ΣX kF ≤ (d5/2 log(d)kΣ̂n − ΣX kF )1/2 .
By Assumption (F), kAn kF is O(1); by Lemma A.3, d5/2 log(d)kΣ̂n −ΣX kF = oP (1) for d9 (log(d))2 =
6
o(n); by Lemmas A.2 and A.3, d(log(d))1/2 kT0−1 − Σ−1
X kF is oP (1) for d log(d) = o(n); by Lemma
√
√
P
A.4, (d log(d))−1/2 k nT2 k2 = (d log(d))−1/2 k √1n Ss1 +1,n k2 is OP (1) for d = o( n). Thus, R2 −→ 0.
xxi
√
On R3 and R4 . First consider R3 . By noting that s1 = o( n/(λdκn )), wpg1,
√
√
1/2
1/2
kR3 k2 ≤ d nkAn kF (kΣ̂n − ΣX kF,d )1/2 kT0−1 kF,d kT3 k2 . dλs1 κn / n → 0.
Thus, kR3 k2 = oP (1). In the same way, kR4 k2 = oP (1).
References
Jianqing Fan and Jinchi Lv. Non-concave penalized likelihood with np-dimensionality. IEEE
Transactions On Information Theory, 57:5467–5484, 2011.
Jianqing Fan and Heng Peng. On non-concave penalized likelihood with diverging number of
parameters. The Annals of Statistics, 32:928–961, 2004.
Michael R. Kosorok. Introduction to Empirical Processes and Semiparametric Inference. Springer
New York, 2008.
Jerzy Neyman and Elizabeth L. Scott. Consistent estimates based on partially consistent observations. Econometrica, 16:1–32, 1948.
Albert N. Shiryaev. Probability. Springer-Verlag, second edition, 1995.
Aad W. van der Vaart. Asymptotic Statistics. Cambridge University Press, 1998.
Peng Zhao and Bin Yu. On model selection consistency of lasso. The Journal of Machine Learning
Research, 7(Nov):2541–2563, 2006.
xxii
| 10math.ST
|
REMIX: Automated Exploration for Interactive Outlier Detection
Yanjie Fu∗1 , Charu Aggarwal†2 , Srinivasan Parthasarathy‡2 , Deepak S. Turaga§2 and Hui
Xiong¶3
arXiv:1705.05986v1 [cs.AI] 17 May 2017
1
Missouri U. of Science & Technology, Missouri, USA
2
IBM T. J. Watson Research Center, NY, USA
3
Rutgers University, NJ, USA
May 18, 2017
Abstract
Outlier detection is the identification of points in a dataset that do not conform to the norm. Outlier
detection is highly sensitive to the choice of the detection algorithm and the feature subspace used by
the algorithm. Extracting domain-relevant insights from outliers needs systematic exploration of these
choices since diverse outlier sets could lead to complementary insights. This challenge is especially acute
in an interactive setting, where the choices must be explored in a time-constrained manner.
In this work, we present REMIX, the first system to address the problem of outlier detection in an
interactive setting. REMIX uses a novel mixed integer programming (MIP) formulation for automatically
selecting and executing a diverse set of outlier detectors within a time limit. This formulation incorporates
multiple aspects such as (i) an upper limit on the total execution time of detectors (ii) diversity in the
space of algorithms and features, and (iii) meta-learning for evaluating the cost and utility of detectors.
REMIX provides two distinct ways for the analyst to consume its results: (i) a partitioning of the
detectors explored by REMIX into perspectives through low-rank non-negative matrix factorization; each
perspective can be easily visualized as an intuitive heatmap of experiments versus outliers, and (ii) an
ensembled set of outliers which combines outlier scores from all detectors. We demonstrate the benefits
of REMIX through extensive empirical validation on real-world data.
1
Introduction
Outlier detection is the identication of points in a dataset that do not conform to the norm. This is a critical
task in data analysis and is widely used in many applications such as financial fraud detection, Internet
traffic monitoring, and cyber security [4]. Outlier detection is highly sensitive to the choice of the detection
algorithm and the feature subspace used by the algorithm [4, 5, 35]. Further, outlier detection is often
performed on high dimensional data in an unsupervised manner without data labels; distinct sets of outliers
discovered through different algorithmic choices could reveal complementary insights about the application
domain. Thus, unsupervised outlier detection is a data analysis task which inherently requires a principled
exploration of the diverse algorithmic choices that are available.
Recent advances in interactive data exploration [10, 17, 33, 34] show much promise towards automated
discovery of advanced statistical insights from complex datasets while minimizing the burden of exploration
for the analyst. We study unsupervised outlier detection in such an interactive setting and consider three
practical design requirements: (1) Automated exploration: the system should automatically enumerate,
assess, select and execute a diverse set of outlier detectors; the exploration strategy should guarantee coverage
∗ [email protected]
† [email protected]
‡ [email protected]
§ [email protected]
¶ [email protected]
1
in the space of features and algorithm parameters. (2) Predictable response time: the system should
conduct its exploration within a specified time limit. This implies an exploration strategy that is sensitive
to the execution time (cost) of the detectors. (3) Visual interpretability: the system should enable the
user to easily navigate the results of automated exploration by grouping together detectors that are similar
in terms of the data points they identify as outliers.
Exploratory Action
100
75
50
25
0
0
50
100
150
200
Data Point
(a) Perspective 1
Exploratory Action
100
75
50
25
0
0
50
100
150
200
Data Point
(b) Perspective 2
Figure 1: Factorization of an outlier matrix into two perspectives. Each perspective is a heatmapped
matrix whose columns are data points and rows are detectors. The intensity of a cell (s, p) in a perspective
corresponds to the extent to which detector s identifies point p as an outlier. Each perspective clearly
identifies a distinct set of outliers.
1.1
Key Contributions
We present REMIX, a modular framework for automated outlier exploration and the first to address the
outlier detection problem in an interactive setting. To the best of our knowledge, none of the existing outlier
detection systems support interactivity in the manner outlined above – in particular, we are not aware of
any system which automatically explores the diverse algorithmic choices in outlier detection within a given
time limit. The following are the key contributions of our work.
1.1.1
MIP based Automated Exploration
REMIX systematically enumerates candidate outlier detectors and formulates the exploration problem as a
mixed integer program (MIP). The solution to this MIP yields a subset of candidates which are executed
by REMIX. The MIP maximizes a novel aggregate utility measure which trades off between the total utility
of the top-k vs all selected candidates; the MIP also enforces (i) an upper limit on the total cost (budget)
of the selected candidates, and (ii) a set of diversity constraints which ensure that each detection algorithm
and certain prioritized feature subspaces get at least a certain minimum share of the exploration budget.
2
1.1.2
Meta-Learning for Cost and Utility
The REMIX MIP requires an estimate of cost and utility for each candidate detector. In order to estimate
cost, REMIX trains a meta-learning model for each algorithm which uses the number of data points, size
of the feature subspace, and various product terms derived from them as meta-features. It is significantly
harder to estimate or even define utility. REMIX handles this by defining the utility of a detector as a
proxy for its accuracy on the given data. REMIX estimates this by training a meta-learning model for
each algorithm that uses various statistical descriptors of a detector’s feature subspace as meta-features; this
model is trained on a corpus of outlier detection datasets that are labeled by domain experts [2].
1.1.3
Perspective Factorization
The diversity of feature subspaces and algorithms explored by REMIX will result in different detectors
marking a distinct set of data points as outliers. REMIX provides a succinct way for the data analyst to
visualize these results as heatmaps. Consider the outlier matrix ∆ where ∆s,p is the normalized ([0, 1]ranging) outlier score assigned by detector s to data point p. REMIX uses a low rank non-negative matrix
factorization scheme to bi-cluster ∆ into a small user-specified number of perspectives such that there is
consensus among the detectors within a perspective. The idea of outlier perspectives is a generalization of
the idea of outlier ensembles. A notable special case occurs when the number of perspectives equals one:
here, the results from all the detectors are ensembled into a single set of outliers.
REMIX can be used in two modes. (i) Simple: An analyst can gain rapid visual understanding of the
outlier space by providing two simple inputs: an exploration budget and the number of perspectives. This
yields bi-clustered heatmaps of outliers that are easily interpretable as in Figure 1. (ii) Advanced: As
discussed in Section 4.3.1, an advanced user can also re-configure specific modules within REMIX such as
utility estimation or the enumeration of prioritized feature subspaces. Such re-configuration would steer the
REMIX MIP towards alternate optimization goals while still guaranteeing exploration within a given budget
and providing factorized heatmap visualizations of the diverse outlier sets.
The rest of the paper is organized as follows. We survey related work in Section 2, and provide background
definitions and an overview in Section 3. In Sections 4.1 – 4.5, we present the details of feature subspace and
candidate detector enumeration, cost and utility estimation, MIP for automated exploration, and perspective
factorization. We present an evaluation of REMIX on real-world datasets in Section 5 and conclude in Section
6.
2
Related Work
Outlier ensembles [4, 23, 31, 5] combine multiple outlier results to obtain a more robust set of outliers.
Model centered ensembles combine results from different base detectors while data-centered ensembles explore
different horizontal or vertical samples of the dataset and combine their results [4]. Research in the area
of subspace mining [23, 18, 19, 8, 30] focuses on exploring a diverse family of feature subspaces which are
interesting in terms of their ability to reveal outliers. The work in [28] introduces a new ensemble model
that uses detector explanations to choose base detectors selectively and remain robust to errors in those
detectors.
REMIX is related to ensemble outlier detection since setting the number of perspectives to one in REMIX
leads to ensembling of results from the base detectors. However, REMIX has some notable distinctions: the
idea of perspectives in REMIX generalizes the notion of ensembles; setting the number of perspectives to
a number greater than one is possible in REMIX and results in complementary views of the outlier space
which can be visualized as heatmaps; further, REMIX seeks to guarantee coverage not just in the space
of features or feature subspaces, but also available algorithms – subject to a budget constraint on the total
time available for exploration. None of the existing approaches in literature provide this guarantee on the
exploration time.
Multiview outlier detection [9, 29, 27, 13] deals with a setting where the input consists of multiple
datasets (views), and each view provides a distinct set of features for characterizing the objects in the domain.
Algorithms in this setting aim to detect objects that are outliers in each view or objects that are normal but
show inconsistent class or clustering characteristics across different views. REMIX is related to multiview
3
outlier detection through non-negative matrix factorization (NMF) which is often used here for creating a
combined outlier score of objects. In REMIX, NMF is used not just for ensembling detector scores but also
for factorizing detector results into multiple heatmap visualizations.
Automated exploration is a growing trend in the world of commercial data science systems [1, 3]
as well as machine learning and statistical research [6, 32, 20, 25, 12, 14]. While [14] focuses on model
selection, [25, 12] focus on exploration for non-parametric regression models, [1, 3, 6, 32] deal with algorithm
exploration for classification models, while [20] deals with automated feature generation for classification
and regression. Budgeted allocations and recommendations have also been studied in the context of
problems other than outlier analysis [24, 26].
3
3.1
Preliminaries
Baseline Algorithms
Our implementation of REMIX uses the following set A of five baseline outlier detection algorithms that are
well-known. 1) Local outlier factor: LOF [7] finds outliers by measuring the local deviation of a point from
its neighbors. 2) Mahalanobis distance: MD [16] detects outliers by computing the Mahalanobis distance
from a point and the center of the entire dataset as outlier score. 3) Angle-based outlier detection:
ABOD [22] identifies outliers by considering the variances of the angles between the difference vectors of data
points, which is more robust that distance in high-dimensional space. ABOD is robust yet time-consuming.
4) Feature-bagging outlier detection: FBOD [23] is an ensemble method, which is based on the results
of local outlier factor (LOF). During each iteration, a random feature subspace is selected. LOF then is
applied to calculate the LOF scores based on the selected data subset. The final score of FBOD is the
cumulative sum of each iteration. 5) Subspace outlier detection: SOD [21]: SOD aims to detect outliers
in varying subspaces of a high dimensional feature space. Specifically, for each point in the dataset, SOD
explores the axis-parallel subspace spanned by its neighbors and determines how much the point deviates
from the neighbors in this subspace.
3.2
Interactive Outlier Exploration
A dataset in REMIX is a real-valued matrix A with m columns and n rows. A feature subspace is a subset
of columns in A. An outlier detector Da,f is simply a combination of an algorithm a ∈ A and a feature
subspace f of A. The cost and utility values ca,f and ua,f are positive values associated with Da,f which
are intended to be estimates of the execution cost and accuracy of Da,f respectively. REMIX enumerates
candidate outlier detectors based on a family of prioritized feature subspaces Fp and a family of randomly
constructed feature subspaces Fr .
The interactive outlier exploration problem is a budgeted optimization problem which selects a subset
of detectors with a maximization objective that is a linear combination of two quantities: (i) the total
utility of all the selected detectors, and (ii) the total utility of the top-k selected detectors, subject to
the following budget and diversity constraints: (i) the total cost the selected detectors does not exceed the
budget Ttotal , (ii) each algorithm gets a guaranteed share of the exploration budget, and (iii) each prioritized
feature gets a guaranteed share of the exploration budget.
4
The REMIX Framework
We now describe the five components of REMIX shown in Figure 2 starting with feature subspace enumeration.
4.1
Feature Subspace and Candidate Detector Enumeration
Algorithm 1 describes feature subspace enumeration and has three parts: (i) creating a non-redundant feature
bag Fnr (lines 1 – 13), (ii) creating a prioritized family of subspaces Fp (lines 14-18) using a feature ranking
4
Algorithm 1 Feature Subspace Enumeration
Input: Data Matrix A
Output: Feature subspace families Fp and Fr
m
S
1: Fnr =
A∗,j
. Initialize non-redundant feature bag
j=1
2:
while |Fnr | ≥ 2 do
. Bag has at least 2 features
P
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
(i6=j)∧(A∗,i ∈Fnr )
σi,j
. Mean correlation
∀A∗,j ∈ Fnr , σ̂j ←
|Fnr |−1
(p, q) ← arg max(i6=j)∧(A∗,i ∈Fnr )∧(A∗,j ∈Fnr ) σi,j
. Max
if σp,q ≥ α then
. High max correlation?
if σ̂p ≥ σ̂q then
. Greater average correlation?
Fnr ← Fnr \ {A∗,p }
. Drop A∗,p
else
Fnr ← Fnr \ {A∗,q }
. Drop A∗,q
end if
else break
. Break out of while loop
end if
end while
Fp = Φ
. Initialize to null set
for ` ← 1, |Fnr | do
S` ← Top-` features in Fnr ranked by their Laplacian scores
Fp ← Fp ∪ {S` }
. Add prioritized subspace
end for
Fr = Φ
. Initialize to null set
for i ← 1, Γ do
Ti ← Random subspace with each feature sampled independently at random without replacement
from Fnr with probability 21
Fr ← Fr ∪ {Ti }
. Add random subspace
end for
5
Feature Subspace
Enumeration
1
Perspective Factorization
and Ensembling
Candidate Detector
Enumeration
5
2
Mixed Integer Program
based Exploration
Cost and Utility
Estimation
3
4
Figure 2: REMIX Components: REMIX starts by enumerating multiple feature subspaces from the given
dataset, which in turn is used to enumerate candidate detectors. Next, REMIX evaluates the cost and utility
of all the enumerated detectors (the meta-learning components for training the cost and utility models is
not shown in this figure). The cost and utility values are used as part of a mixed integer program (MIP)
which selects a subset of candidates for execution based on a utility maximization objective and budget
and diversity constraints. The outlier results from the detectors executed by REMIX are factorized into
perspectives, and also ensembled into a single set of results.
approach and (iii) creating a randomized family of subspaces Fr (lines 19 – 23) used for maximizing coverage
and diversity during exploration. Redundant features 6∈ Fnr are not part of subspaces in Fp or Fr .
Algorithm 1 begins by initializing Fnr to all features in A (line 1). It then iteratively looks for a member
in Fnr which can be considered redundant and hence dropped from Fnr . In order for a feature A∗,p ∈ Fnr
to be considered redundant, it needs to (i) be in a maximally correlated feature pair {A∗,p , A∗,q } ⊆ Fnr (line
4), (ii) the correlation coefficient σp,q must be above the REMIX’s redundancy threshold α (line 5), and (iii)
of the two features in the pair, A∗,p must have a mean correlation with other features in Fnr that ≥ the
mean correlation of A∗,q (line 7). We set the default value of α in REMIX to 0.9 based on our experimental
evaluation. It is easy to see that at the end of line 13, Algorithm 1 yields a non-redundant feature bag Fnr
with the following property.
Observation 1. For any pair of features {A∗,p , A∗,q } ⊆ Fnr , σp,q < α. For any feature A∗,r 6∈ Fnr ,
∃A∗,s ∈ Fnr , σr,s ≥ α.
The family of randomized feature subspaces Fr is created for the purpose of guaranteeing feature coverage
and diversity during exploration. In particular, we select Γ subspaces, where each subspace consists of
features selected independently at random without replacement with probability 12 from Fnr . We set the
default value of Γ in REMIX to |Fp |/2 to balance the size of the two families.
The family of prioritized feature subspaces Fp is created as follows. The features in Fnr are first sorted
according to their Laplacian scores [15]. We add |Fnr | subspaces to Fp , where the `th subspace is the set
of top-` features in Fnr ranked by their Laplacian scores. We now provide a brief justification for our use
of the Laplacian score. Due to lack of space, refer the reader to [15] for the exact details of the Laplacian
computation. The Laplacian score is computed as a way to reflect a feature’s ability to preserve locality. In
particular, consider the projection A0 of all points in A onto the subspace Fnr ; now, consider the r-nearest
neighborhood of points in A0 . Features that respect this r-nearest neighborhood provide a better separation
of the inlier class from outlier class within the data. We note that our experimental evaluation in Section
5 consistently demonstrates improved outlier detection accuracy with Fp and Fr as opposed to purely Fr .
We also discuss potential alternatives for prioritized feature subspace construction in Section 4.3.1.
We enumerate candidate detectors by a cartesian product of Fp ∪ Fr and the set of baseline detection
algorithms A.
4.2
Cost Estimation
The cost of a candidate detector can be modeled as a function of its algorithm as well as the size of its feature
subspace f (n and |f |). The runtime complexity of algorithms in the ‘big-O’ notation provides an asymptotic
relationship between cost and input size; however, in REMIX, we seek a more refined model which accounts
6
for lower order terms and the constants hidden by ‘big-O’. We do this by training a multivariate linear
regression model for each
algorithm. Specifically, consider the following polynomial: (1 + |f | + n + log |f | +
log n)3 − 1. There are 73 − 1 distinct terms in the expansion of this polynomial which can be derived exactly
given a feature subspace. The terms1 form the explanatory variables while cost is the dependent variable in
the linear regression model.
4.3
Utility Estimation
Given a detector Da,f , the utility estimation algorithm (Algorithm 2) first normalizes the features in f , and
computes a variety of feature-level statistics for each feature ψ ∈ f . These statistics include the Laplacian
score which is a useful measure for unsupervised feature selection (Section 4.1), the standard deviation,
skewness which measures the assymetry of the feature distribution, kurtosis which measures the extent to
which the feature distribution is heavy-tailed, and entropy which is a measure of the information content in
the feature. We note that these computations are done once for each feature in Fnr and is reused for any
given any detector. Next, the algorithm computes the meta-feature vector M F V (f ) of feature-subspace-level
statistics by combining feature-level statistics. For instance, consider the Laplacian scores of all the features
in f ; the mean, median, median absolute deviation, min, max, and standard deviation of all the Laplacian
scores provide 6 of the 30 distinct components of M F V (f ) in this step. Finally, it uses an algorithm specific
utility model Ua (M F V (f )) to estimate the utility ua,f of Da,f . REMIX trains five distinct linear regression
models ULOF , UM D , UABOD , ULOF , and USOD corresponding to each algorithm. These models are trained
based on distinct expert labeled datasets from the outlier dataset repository [2]. The explanatory variables
for this linear regression are the feature-subspace-level statistics described in Algorithm 2. The dependent
variable is the detection accuracy, measured as the fraction of the data points on which both the detector
and the expert labeled ground truth agree on the outlier characterization.
Algorithm 2 Utility Estimation
Input: A candidate detector Da,f
Output: The utility ua,f of the detector
1: ∀ψ ∈ f , normalize ψ
. One time procedure applied to Fnr
2: ∀ψ ∈ f , extract the following 5 feature-level statistics: Laplacian score, standard deviation, skewness,
kurtosis, and entropy
. One time procedure applied to Fnr
3: Extract the following 30 feature-subspace-level statistics: mean, median, median absolute deviation, min,
max, and standard deviation for each of the 5 feature-level statistics extracted in Step 2. Let M F V (f )
contain these feature subspace-level statistics.
4: Return ua,f = Ua (M F V (f )), where Ua is the utility estimation model learnt for algorithm a
4.3.1
Discussion
Estimating or even defining the utility of a detector is significantly harder than estimating its cost. The
goal of utility estimation in REMIX is not to learn a perfect model; rather, the goal is merely to learn a
utility model which can effectively steer the solution of the mixed integer program (MIP) used by REMIX
for exploration (Section 4.4). Our experiments in Section 5 demonstrate that this is indeed the case with
REMIX. Further, REMIX is intended to be a flexible framework where alternative mechanisms for utility
estimation can be plugged in. For instance, consider Cumulative Mutual Information (CMI) metric and the
Apriori-style feature subspace search algorithm presented in [8] for subspace outlier detection. REMIX can
use CMI as the detector utility, and this Apriori-style algorithm as an alternative for enumerating prioritized
feature subspaces. We chose the meta-learning approach in our implementation since this approach is
algorithm agnostic and hence can be generalized easily.
1 The
exponent 3 suffices to model the cost of most known outlier detection algorithms
7
4.4
MIP based Exploration
REMIX executes only a subset of detectors enumerated by its detector enumeration component (Section
4.1). This subset is determined by solving the following mixed integer program (MIP).
max
X
X
za,f ua,f +λ
{z
}
utility of top-k detectors
X
X
X
ya,f ua,f
(1)
a∈A f ∈Fp ∪Fr
a∈A f ∈Fp ∪Fr
|
X
|
{z
utility of all detectors
ya,f ca,f ≤ Ttotal
}
(2)
a∈A f ∈Fp ∪Fr
X
∀a ∈ A :
ya,f ca,f ≥
f ∈Fp ∪Fr
∀f ∈ Fp :
X
ya,f ca,f ≥
a∈A
Ttotal
2 · |A|
Ttotal
2 · |Fp |
∀a ∈ A, ∀f ∈ Fp ∪ Fr : za,f ≤ ya,f
X X
za,f ≤ k
(3)
(4)
(5)
(6)
a∈A f ∈Fp ∪Fr
∀a ∈ A, ∀f ∈ Fp ∪ Fr : yi,j ∈ {0, 1}
(7)
∀a ∈ A, ∀f ∈ Fp ∪ Fr : zi,j ∈ {0, 1}
(8)
Given an algorithm a and a feature subspace f , ya,f is the binary indicator variable in the MIP which
determines if the detector Da,f is chosen for execution in the MIP solution. Recall that ca,f denotes the
estimated cost of Da,f . We observe the following.
Observation 2. Constraint 2 guarantees that in any feasible solution to the MIP, the total cost of the
selected detectors is ≤ Ttotal .
The value of the total exploration budget Ttotal is provided by the analyst as part of her interaction with
REMIX.
Observation 3. Constraint 3 guarantees that in any feasible solution to the MIP, each algorithm is explored
total
for an estimated duration of time which is ≥ T2·|A|
.
Observation 4. Constraint 4 guarantees that in any feasible solution to the MIP, each prioritized feature is
Ttotal
explored for an estimated duration of time which is ≥ 2·|F
. This also implies that the exploration focuses
p|
at least half its total time on prioritized features.
Consider the binary indicator variable za,f corresponding to the detector Da,f . Constraint 5 ensures
that in any feasible solution to the MIP, za,f can be 1 only if it is chosen in the solution (i.e., ya,f = 1).
Constraint 6 ensures at most k of the detectors chosen by the solution have their z-values set to 1. Now
consider an optimal solution to
Pthe MIP.
P Since utility values are non-negative for all detectors, the first
part of the objective function a∈A f ∈Fp ∪Fr za,f ua,f is maximized when exactly k detectors have their
z-values set to 1 (and not fewer) and when the detectors whose z-values are set to 1 are the ones with the
highest utilities amongst the selected detectors. This leads us to the following guarantee.
Theorem 5. The optimal solution to the MIP maximizes the sum of utilities of the top-k detectors with
highest utilities and the total utility of all the selected detectors scaled by a factor λ.
We set k = 10 and λ = 1 in our implementation of REMIX which balances the utility of the top-10
detectors vs the total utility of all the selected detectors.
8
4.5
Perspective Factorization
Each detector executed by REMIX provides an outlier score for each data point and the results from different
detectors could be potentially divergent. We now present a factorization technique called NMFE (nonnegative matrix factorization and ensembling) for bi-clustering the detection results into a few succinct
perspectives. All detectors within a perspective agree on how they characterize outliers although there could
be disagreement across perspectives.
Let ∆s,p be the outlier score assigned by detector s for data point p normalized across data points
to be in the range [0, 1]. Consider the matrix of outlier scores ∆ ∈ [0, 1]t×n , where t is the number of
detectors executed by REMIX and n is the number of data points. We perform a rank-g non-negative
matrix factorization of ∆ ≈ ΛΩ> , where Λ ∈ Rt×g and Ω ∈ Rn×g , by minimizing the Kullback-Leibler
divergence between ∆ and ΛΩ> [11]:
min
Λ,Ω≥0
X
∆s,p log
s,p
∆s,p
− ∆s,p + Λs,∗ Ωp,∗
Λs,∗ Ωp,∗
(9)
The matrix ΛΩT by definition can be expressed as a sum of g rank-1 matrices whose rows and columns
correspond to detectors and data points respectively. In any of these rank-1 matrices, every row (column) is
a scaled multiple of any other non-zero row (column), and every entry is between 0 and 1. These properties
make it possible for the rank-1 matrices to be visualized as perspectives, or heatmaps where the intensity of
a heatmap cell is the value of the corresponding entry in the perspective, as shown in Figure 1. The number
of perspectives g is specified by the user as an input to REMIX. Setting g = 1 simply results in a direct
averaging (ensembling) of all the detector results into a single perspective.
5
Experiments
In this section, we present the evaluation of REMIX on real-world data.
5.1
Data Collection
We chose 98 datasets for our study from the outlier dataset repository [2] of the Delft University of Technology. This corpus contains outlier detection datasets labeled by domain experts. These datasets span varied
domains such as flowers, breast cancers, heart diseases, sonar signals, genetic diseases, arrhythmia abnormals, hepatitis, diabetes, leukemia, vehicles, housing, and satellite images. Each dataset has benchmark
outlier labels, with an entry 1 representing outliers and 0 representing normal points.
Figure 3 illustrate some statistics of the 98 data sets. Specifically, Figure 3(a) shows the numbers of
features for each dataset sorted in a descending order. In this figure, we can observe that most of the
datasets contains less than 100 features while only a small portion of these datasets have more than 1000
features. Figure 3(b) shows the outlier ratio to the total number of data points for each dataset sorted in
a descending order. In this figure, we can find that the outlier ratios of more than 50% of the datasets are
less than 23.2%.
4970
0.5
0.45
1000
0.4
Outlier ratio
Feature number
0.35
100
0.3
0.25
0.2
0.15
10
0.1
0.05
1
0
10
20
30
40
50
60
70
80
90
0
98
0
10
20
30
40
50
60
70
80
90
98
The rankings of data sets with respect to outlier ratio
The rankings of data sets with respect to number of features
(a) Distribution of feature sizes (b) Distribution of outlier ratios
Figure 3: Statistics of the experimental datasets
9
5.2
Cost and Utility Estimation
600
400
Prediction
8
6
0
0
2
200
4
Prediction
10
12
800
14
Among the 98 datasets, we used 96 datasets to train and test the cost and utility estimation models for each
of the 5 baseline algorithms in our REMIX implementation (with a 70%-30% split for train vs test). Figures
4(a) and 4(b) present the performance of the cost and utility models for the LOF algorithm from a specific
run.
Recall that the utility ua,f estimates the fraction of the data points on which the detector da,f and the
expert labels agree on the outlier characterization. Consider the Hamming distance between the outlier bit
vector (1 if outlier, 0 otherwise) created by the detector and the outlier bit vector created by the expert
labels. Clearly, this Hamming distance equals n · (1 − ua,f ) where n is the number of data points. Figure
4(b) plots the estimated Hamming distance on the y-axis and the actual distance as observed by running
the detector on the x-axis. The red lines in Figures 4(a) and 4(b) represent ideal predictors. As expected,
the cost estimator clearly performs better than the utility estimator. However, as mentioned in Section 4.3.1
our goal in utility estimation is not a perfect predictive model but merely to steer the MIP towards better
solutions. We demonstrate this to be the case in our next set of experiments.
0
2
4
6
8
10
12
14
0
200
Groundtruth
400
600
800
Groundtruth
(a) Cost Estimation
(b) Utility Estimation
Figure 4: Cost and Utility Estimation
5.3
Detection Accuracy and Cost
1.2
1
0.08
0.04
LOF
MD
ABOD
FBOD
SOD
Our
0.035
0.03
LOF
MD
ABOD
FBOD
SOD
Our
0.07
0.06
0.8
0.025
0.05
0.6
0.02
0.04
0.015
0.03
0.01
0.02
0.005
0.01
LOF
MD
ABOD
FBOD
SOD
Our
0.4
0.2
0
@10
@13
@15
@17
(a) Precision@N
@20
0
@10
@13
@15
@17
(b) Recall@N
@20
0
@10
@13
@15
@17
@20
(c) Fmeasure@N
Figure 5: Comparison of REMIX with baseline algorithms on the cardiac arrhythmia dataset
We now present a detailed study of the detection accuracy vs cost of REMIX on two datasets from
the corpus. The first dataset is cardiac arrhythmia. Irregularity in heart beat may be harmless or life
threatening. The dataset contains medical records like age, weight and patient’s electrocardiograph related
data of arrhythmia patients and outlier healthy people. The task is to spot outlier healthy people from
arrhythmia patients. The second dataset is about sonar signals bounced off of a metal cylinder. The dataset
contains outlier sonar signals bounced off a roughly cylindrical rock. The task is to separate outlier rock
related sonar signals from cylinder-related sonar signals.
5.3.1
Evaluation Metrics
Recall that the REMIX perspective factorization scheme can be used as an outlier ensembling technique
simply by setting the number of perspectives to 1. We use REMIX in this ensembling mode for the rest of
10
1.2
1
0.35
0.25
LOF
MD
ABOD
FBOD
SOD
Our
0.2
0.8
LOF
MD
ABOD
FBOD
SOD
Our
0.3
0.25
0.15
LOF
MD
ABOD
FBOD
SOD
Our
0.2
0.6
0.15
0.1
0.4
0.1
0.05
0.2
0
0.05
@10
@13
@15
@17
0
@20
@10
(a) Precision@N
@13
@15
@17
@20
0
(b) Recall@N
@10
@13
@15
@17
@20
(c) Fmeasure@N
Figure 6: Comparison of REMIX with baseline algorithms on the sonar signals dataset
0.8
1.1
EE
RSR
RS1
RS1R
NMFE
1
0.9
0.7
0.8
EE
RSR
RS1
RS1R
NMFE
0.75
0.6
0.7
0.5
0.65
EE
RSR
RS1
RS1R
NMFE
0.8
0.4
0.6
0.3
0.55
0.7
0.6
0.5
0.4
@10
@13
@15
@17
(a) Precision@N
@20
0.2
0.5
0.1
0.45
0
@10
@13
@15
@17
(b) Recall@N
@20
0.4
@10
@13
@15
@17
@20
(c) Fmeasure@N
Figure 7: Comparison of REMIX with other exploration and ensemble strategies on the cardiac arrhythmia
dataset
our experiments. We now define Precision@N, Recall@N, and F-measure@N which we use to compare the
detection accuracy of REMIX with various other approaches. Let 1 denote outlier label and 0 denote the
label for normal points in the expert labeled data.
Precision@N Given the top-N list of data points LN sorted
in a descending order of the predicted
T
|LN L=1 |
outlier scores, the precision is defined as: Precision@N =
where L=1 are the data points with
N
expert outlier label = 1.
Recall@N Given the top-N list of data points
T LN sorted in a descending order of the predicted outlier
scores, the recall is defined as: Recall@N = |LN|L=1L|=1 | where L=1 are the outlier data points with label = 1.
F-measure@N F-measure@N incorporates both precision and recall in a single metric by taking their
×Recall@N
harmonic mean: F@N = 2×Precision@N
Precision@N +Recall@N
5.3.2
Baseline Algorithms
We report the performance comparison of REMIX vs baseline algorithms on the sonar signals dataset and the
cardiac arrhythmia dataset in terms of Precision, Recall, and F-measure. In this experiment, we provide sufficient time for the baseline algorithms in {LOF, M D, ABOD, F BOD, SOD} to complete their executions.
Meanwhile, we set a limited exploration time budget of 0.5 second for REMIX.
Results on Effectiveness Comparison. Figure 5 shows that on the cardiac arrhythmia dataset,
REMIX outperforms the five baseline algorithms in terms Precision@N, Recall@N, and Fmeasure@N (N=10,
13, 15, 17, 20). Figure 6 shows that on the sonar signal dataset, our method is consistently better than the
baseline algorithms in terms Precision@N, Recall@N, and Fmeasure@N (N=10, 13, 15, 17, 20).
Results on Efficiency Comparison. Figure 9 jointly show that on the cardiac arrhythmia dataset
and the sonar signal dataset, Mahalanobis distance (MD) takes the least time; Angle-based outlier detection
(ABOD) takes the most time as angle is expensive to compute; REMIX falls in the middle of this cost
spectrum.
11
1.3
1.2
1.1
0.35
0.25
EE
RSR
RS1
RS1R
NMFE
0.2
EE
RSR
RS1
RS1R
NMFE
0.3
EE
RSR
RS1
RS1R
NMFE
0.25
1
0.15
0.9
0.2
0.8
0.15
0.1
0.7
0.1
0.6
0.05
0.05
0.5
0.4
@10
@13
@15
@17
0
@20
0
@10
(a) Precision@N
@13
@15
@17
@20
@10
(b) Recall@N
@13
@15
@17
@20
(c) Fmeasure@N
Figure 8: Comparison of REMIX with other exploration and ensemble strategies on the sonar signals dataset
10 2
10 2
46.02
31.769
31.369
23.429
10 1
10
1
4.398
2.8
10 0
0.859
0.478
10 0
0.481
0.158
10 -1
10 -1
10 -2
0.043
0.00184
10 -2
LOF
MD
ABOD
FBOD
SOD
Our
10 -3
LOF
MD
ABOD
FBOD
SOD
Our
(a) Times costs on the cardiac (b) Times costs on the sonar
arrhythmia dataset
signal dataset
Figure 9: Comparison of execution costs of REMIX and baseline algorithms
5.3.3
The Effectiveness and Efficiency of REMIX NMFE
We study the effectiveness of the REMIX NMFE ensembling strategy, in which we used the rank-1 NMF to
factorize the outlier score matrix ∆ ≈ ΛΩ> and treat Ω as the predicted ensemble outlier scores. Let t denote
the number of detectors selected in the REMIX MIP solution. Let C be the set of all candidate detectors
enumerated by REMIX. We compared REMIX with the following exploration and ensembling strategies:
(1) Exhaustive Ensemble (EE): we execute all the detectors in C and averaged all the outlier scores;
(2) Randomly select t detectors (RSR): we randomly select and execute t detectors from C, and then
average the outlier scores of these detectors; (3) Randomly select 1 detector (RS1): we randomly select
one detector and use its results; (4) Randomly select 1 detector in the MIP solution (RS1R): we
randomly select one detector in the MIP solution and use its results; In this experiment, we provide sufficient
time for all strategies complete their execution. We set a limited time budget of 0.5 seconds for REMIX.
Results on Effectiveness Comparison. Figure 7 shows that on the cardiac arrhythmia dataset our
strategy outperforms the other exploration and ensembling strategies in terms of Precision@N, Recall@N,
and Fmeasure@N (N=10, 13, 15, 17, 20). Figure 8 shows that on the sonar signal dataset, REMIX is
consistently better than the other strategies in terms of Precision@N, Recall@N, and Fmeasure@N (N=10,
13, 15, 17, 20).
Results on Efficiency Comparison. Figure 10(a) shows that on the cardiac arrhythmia dataset
REMIX takes 0.48 second, and EE, RSR, RS1 require much more time. While RS1R takes only 0.0037
second, its detection accuracy is lower than ours. Figure 10(b) shows on the sonar signal dataset, our
strategy takes only 0.49 second, which is much less than the time costs of EE, RSR, RS1. While RS1R takes
only 0.0033 second, our method outperforms RS1R with respect to detection accuracy.
6
Conclusions and Future Directions
In this paper, we presented REMIX, a modular framework for outlier exploration which is the first to
study the problem of outlier detection in an interactive setting. At the heart of REMIX is an optimization
12
10 4
10
3
10
2
10 4
5931.62
10 3
10
2
10 1
10 1
0
0
10
0.471545
0.313311
10
7097.053
1597.939
1336.612
-1
10
10
10 -2
11.02027
0.4784815
-1
10 -2
0.0037129
10 -3
EE
RSR
RS1
RS1R
0.003299873
NMFE
10
-3
EE
RSR
RS1
RS1R
NMFE
(a) Times costs on the cardiac (b) Times costs on the sonar
arrhythmia dataset
signal dataset
Figure 10: Comparison of execution costs of REMIX and other exploration strategies
approach which systematically steers the selection of base outlier detectors in a manner that is sensitive
to their execution costs, while maximizing an aggregate utility function of the solution and also ensuring
diversity across algorithmic choice points present in exploration. Data analysts are naturally interested
in extracting and interpreting outliers through multiple detection mechanisms since distinct outliers could
lead to distinct actionable insights within the application domain. REMIX facilitates this understanding
in a practical manner by shifting the burden of exploration away from the analyst through automation,
and by summarizing the results of automated exploration into a few coherent heatmap visualizations called
perspectives.
We believe many of the techniques presented in this paper could be of independent interest to other
machine learning problems. We are interested in extending the REMIX exploratory approach beyond outlier
detection to clustering of high-dimensional data. Another interesting direction of research is sequential
recommendations for visual outlier exploration: in particular, we are interested in approaches for presenting
outlier insights approximately through a sequence of visualizations (like heatmaps) such that both the length
of this sequence as well the perceptual error involved across the visualizations is minimized while coverage
across the various exploratory choice points is maximized. Also the study of outlier aspect mining in an
interactive setting – which deals with the inverse problem of finding explanatory features which characterize
a given set of points as outliers in a cost sensitive manner – presents a new and interesting direction of
research.
References
[1] Data robot. http://www.datarobot.com, 2017.
[2] Outlier detection datasets. http://homepage.tudelft.nl/n9d04/occ/index.html, 2017.
[3] Sytree. http://www.skytree.net, 2017.
[4] Charu C Aggarwal. Outlier analysis. Springer Science & Business Media, 2013.
[5] Charu C Aggarwal and Saket Sathe. Theoretical foundations and algorithms for outlier ensembles.
ACM SIGKDD Explorations Newsletter, 17(1):24–47, 2015.
[6] Alain Biem, Maria Butrico, Mark Feblowitz, Tim Klinger, Yuri Malitsky, Kenney Ng, Adam Perer,
Chandra Reddy, Anton Riabov, Horst Samulowitz, Daby M. Sow, Gerald Tesauro, and Deepak S.
Turaga. Towards cognitive automation of data science. In Proceedings of the Twenty-Ninth AAAI
Conference on Artificial Intelligence, January 25-30, 2015, Austin, Texas, USA., pages 4268–4269,
2015.
[7] Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. Lof: identifying density-based
local outliers. In ACM sigmod record, volume 29, pages 93–104. ACM, 2000.
13
[8] Klemens Bhm, Fabian Keller, Emmanuel Mller, Hoang Vu Nguyen, and Jilles Vreeken. Cmi: An
information-theoretic contrast measure for enhancing subspace cluster and outlier detection. In SDM,
pages 198–206. SIAM, 2013.
[9] Santanu Das, Bryan L Matthews, Ashok N Srivastava, and Nikunj C Oza. Multiple kernel learning for
heterogeneous anomaly detection: algorithm and aviation safety case study. In Proceedings of the 16th
ACM SIGKDD international conference on Knowledge discovery and data mining, pages 47–56. ACM,
2010.
[10] Kyriaki Dimitriadou, Olga Papaemmanouil, and Yanlei Diao. Explore-by-example: An automatic query
steering framework for interactive data exploration. In Proceedings of the 2014 ACM SIGMOD International Conference on Management of Data, SIGMOD ’14, pages 517–528, New York, NY, USA, 2014.
ACM.
[11] Chris H. Q. Ding, Tao Li, and Michael I. Jordan. Convex and semi-nonnegative matrix factorizations.
IEEE Trans. Pattern Anal. Mach. Intell., 32(1):45–55, January 2010.
[12] David Duvenaud, James Robert Lloyd, Roger Grosse, Joshua B. Tenenbaum, and Zoubin Ghahramani.
Structure discovery in nonparametric regression through compositional kernel search. In Proceedings of
the 30th International Conference on Machine Learning, June 2013.
[13] Jing Gao, Wei Fan, Deepak Turaga, Srinivasan Parthasarathy, and Jiawei Han. A spectral framework
for detecting inconsistency across multi-source object relationships. In Data Mining (ICDM), 2011
IEEE 11th International Conference on, pages 1050–1055. IEEE, 2011.
[14] R.B. Grosse, R. Salakhutdinov, W.T. Freeman, and J.B. Tenenbaum. Exploiting compositionality to
explore a large space of model structures. In Uncertainty in Artificial Intelligence, 2012.
[15] Xiaofei He, Deng Cai, and Partha Niyogi. Laplacian score for feature selection. In Advances in neural
information processing systems, pages 507–514, 2005.
[16] Victoria J Hodge and Jim Austin. A survey of outlier detection methodologies. Artificial intelligence
review, 22(2):85–126, 2004.
[17] Alexander Kalinin, Ugur Cetintemel, and Stan Zdonik. Interactive data exploration using semantic
windows. In SIGMOD Conference, 2014.
[18] Fabian Keller, Emmanuel Müller, and Klemens Böhm. Hics: high contrast subspaces for density-based
outlier ranking. In Data Engineering (ICDE), 2012 IEEE 28th International Conference on, pages
1037–1048. IEEE, 2012.
[19] Fabian Keller, Emmanuel Mller, Andreas Wixler, and Klemens Bhm. Flexible and adaptive subspace
search for outlier analysis. In Qi He, Arun Iyengar, Wolfgang Nejdl, Jian Pei, and Rajeev Rastogi,
editors, CIKM, pages 1381–1390. ACM, 2013.
[20] Udayan Khurana, Deepak Turaga, Horst Samulowitz, and Srinivasan Parthasarathy. Cognito: Automated feature engineering for supervised learning. IEEE, 2016.
[21] Hans-Peter Kriegel, Peer Kröger, Erich Schubert, and Arthur Zimek. Outlier detection in axis-parallel
subspaces of high dimensional data. In Pacific-Asia Conference on Knowledge Discovery and Data
Mining, pages 831–838. Springer, 2009.
[22] Hans-Peter Kriegel, Arthur Zimek, et al. Angle-based outlier detection in high-dimensional data. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining,
pages 444–452. ACM, 2008.
[23] Aleksandar Lazarevic and Vipin Kumar. Feature bagging for outlier detection. In Proceedings of the
eleventh ACM SIGKDD international conference on Knowledge discovery in data mining, pages 157–
166. ACM, 2005.
14
[24] Lei Li, Ding-Ding Wang, Shun-Zhi Zhu, and Tao Li. Personalized news recommendation: a review and
an experimental investigation. Journal of computer science and technology, 26(5):754–766, 2011.
[25] James Robert Lloyd, David Duvenaud, Roger Grosse, Joshua B. Tenenbaum, and Zoubin Ghahramani.
Automatic construction and Natural-Language description of nonparametric regression models. In
Association for the Advancement of Artificial Intelligence (AAAI), 2014.
[26] Tyler Lu and Craig Boutilier. Budgeted social choice: From consensus to personalized decision making.
In IJCAI, volume 11, pages 280–286, 2011.
[27] Alejandro Marcos Alvarez, Makoto Yamada, Akisato Kimura, and Tomoharu Iwata. Clustering-based
anomaly detection in multi-view data. In Proceedings of the 22nd ACM international conference on
Conference on information & knowledge management, pages 1545–1548. ACM, 2013.
[28] Alex Memory and Ted Senator. Towards robust anomaly detection ensembles using explanations. In In
the ODDx3 workshop of KDD2015, 2015.
[29] Emmanuel Muller, Ira Assent, Patricia Iglesias, Yvonne Mulle, and Klemens Bohm. Outlier ranking via
subspace analysis in multiple views of the data. In Data Mining (ICDM), 2012 IEEE 12th International
Conference on, pages 529–538. IEEE, 2012.
[30] Emmanuel Müller, Matthias Schiffer, and Thomas Seidl. Statistical selection of relevant subspace
projections for outlier ranking. In Data Engineering (ICDE), 2011 IEEE 27th International Conference
on, pages 434–445. IEEE, 2011.
[31] Shebuti Rayana and Leman Akoglu. Less is more: Building selective anomaly ensembles. arXiv preprint
arXiv:1501.01924, 2015.
[32] Ashish Sabharwal, Horst Samulowitz, and Gerald Tesauro. Selecting near-optimal learners via incremental data allocation. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence,
February 12-17, 2016, Phoenix, Arizona, USA., pages 2007–2015, 2016.
[33] Tarique Siddiqui, Albert Kim, John Lee, Karrie Karahalios, and Aditya Parameswaran. Effortless data
exploration with zenvisage: An expressive and interactive visual analytics system. Proc. VLDB Endow.,
10(4):457–468, November 2016.
[34] Abdul Wasay, Manos Athanassoulis, and Stratos Idreos. Queriosity: Automated data exploration. In
Barbara Carminati and Latifur Khan, editors, BigData Congress, pages 716–719. IEEE, 2015.
[35] Arthur Zimek, Ricardo J.G.B. Campello, and Jörg Sander. Ensembles for unsupervised outlier detection:
Challenges and research questions a position paper. SIGKDD Explor. Newsl., 15(1):11–22, March 2014.
15
| 2cs.AI
|
Functional Units for Natural Numbers
arXiv:0911.1851v3 [cs.PL] 17 Oct 2010
J.A. Bergstra and C.A. Middelburg
Informatics Institute, Faculty of Science, University of Amsterdam,
Science Park 107, 1098 XG Amsterdam, the Netherlands
[email protected],[email protected]
Abstract. Interaction with services provided by an execution environment forms part of the behaviours exhibited by instruction sequences
under execution. Mechanisms related to the kind of interaction in question have been proposed in the setting of thread algebra. Like thread,
service is an abstract behavioural concept. The concept of a functional
unit is similar to the concept of a service, but more concrete. A state
space is inherent in the concept of a functional unit, whereas it is not
inherent in the concept of a service. In this paper, we establish the existence of a universal computable functional unit for natural numbers and
related results.
Keywords: functional unit, instruction sequence.
1998 ACM Computing Classification: F.1.1, F.4.1.
1
Introduction
We take the view that sequential programs are in essence sequences of instructions, and that interaction with services provided by an execution environment
forms part of the behaviours exhibited by instruction sequences under execution
(see e.g. [1,7]). The interaction in question is concerned with the processing of
instructions. In earlier work, mechanisms that have a direct bearing on this kind
of interaction have been proposed in the setting of basic thread algebra (see
e.g. [3,4]). Both thread and service are abstract behavioural concepts.
We experienced recently limitations of the concept of a service because a
state space is not inherent in this concept. This forms the greater part of our
motivation for introducing and studying the concept of a functional unit in this
paper. This concept is similar to the concept of a service, but it is at a lower
level of abstraction. In the concept of a functional unit, a state space is inherent.
Rather than first considering functional units in general for an arbitrary state
space, we first consider the special case where the state space is the set of natural
numbers. This case is arguably the simplest significant case. We establish general
results concerning functional units for natural numbers. The main result is the
existence of a universal computable functional unit for natural numbers. Results
like this one are outside the scope of the concept of a service.
The work presented in this paper belongs to a line of research whose working
hypothesis is that instruction sequence is a central notion of computer science.
In this line of research, program algebra [1] is the setting used for investigating
issues in which instruction sequences are involved. Instruction sequences are also
involved in the issues concerning functional units investigated in this paper. The
starting-point of program algebra is the perception of a program as a single-pass
instruction sequence, i.e. a finite or infinite sequence of instructions of which
each instruction is executed at most once and can be dropped after it has been
executed or jumped over. This perception is simple, appealing, and links up with
practice. Moreover, basic thread algebra [1] is the setting used for modelling the
behaviours exhibited by instruction sequences under execution.1 In this paper,
we use a program notation rooted in program algebra, instead of program algebra
itself.
This paper is organized as follows. First, we give a survey of the program
notation used in this paper (Section 2) and define its semantics using basic
thread algebra (Section 3). Next, we extend basic thread algebra with operators
that are related to the processing of instructions by services (Section 4). Then,
we introduce the concept of a functional unit and related concepts (Section 5).
After that, we investigate functional units for natural numbers (Section 6). We
also make some remarks about functional units for finite state spaces (Section 7).
Finally, we make some concluding remarks (Section 8).
2
PGLB with Boolean Termination
In this section, we give a survey of the program notation PGLBbt . This program
notation is a variant of the program notation PGLB, which belongs to a hierarchy
of program notations rooted in program algebra presented in [1]. PGLBbt is
PGLB with the Boolean termination instructions !t and !f from [3] instead of
the termination instruction ! from [1]. PGLB and PGLBbt are close to existing
assembly languages and have relative jump instructions.
In PGLBbt , it is assumed that a fixed but arbitrary non-empty finite set A of
basic instructions has been given. The intuition is that the execution of a basic
instruction may modify a state and produces t or f at its completion.
PGLBbt has the following primitive instructions:
–
–
–
–
–
–
–
for each a ∈ A, a plain basic instruction a;
for each a ∈ A, a positive test instruction +a;
for each a ∈ A, a negative test instruction −a;
for each l ∈ N, a forward jump instruction #l;
for each l ∈ N, a backward jump instruction \#l;
a positive termination instruction !t;
a negative termination instruction !f.
PGLBbt instruction sequences have the form u1 ; . . . ; uk , where u1 , . . . , uk are
primitive instructions of PGLBbt .
On execution of a PGLBbt instruction sequence, these primitive instructions
have the following effects:
1
In [1], basic thread algebra is introduced under the name basic polarized process
algebra.
2
– the effect of a positive test instruction +a is that basic instruction a is
executed and execution proceeds with the next primitive instruction if t
is produced and otherwise the next primitive instruction is skipped and
execution proceeds with the primitive instruction following the skipped one
– if there is no primitive instruction to proceed with, deadlock occurs;
– the effect of a negative test instruction −a is the same as the effect of +a,
but with the role of the value produced reversed;
– the effect of a plain basic instruction a is the same as the effect of +a,
but execution always proceeds as if t is produced;
– the effect of a forward jump instruction #l is that execution proceeds with
the lth next primitive instruction – if l equals 0 or there is no primitive
instruction to proceed with, deadlock occurs;
– the effect of a backward jump instruction \#l is that execution proceeds with
the lth previous primitive instruction – if l equals 0 or there is no primitive
instruction to proceed with, deadlock occurs;
– the effect of the positive termination instruction !t is that execution terminates and in doing so delivers the Boolean value t;
– the effect of the negative termination instruction !t is that execution terminates and in doing so delivers the Boolean value f.
3
Thread Extraction
In this section, we make precise in the setting of BTAbt (Basic Thread Algebra with Boolean termination) which behaviours are exhibited on execution by
PGLBbt instruction sequences. We start by reviewing BTAbt .
In BTAbt , it is assumed that a fixed but arbitrary non-empty finite set A of
basic actions, with tau 6∈ A, has been given. We write Atau for A ∪ {tau}. The
members of Atau are referred to as actions.
A thread is a behaviour which consists of performing actions in a sequential
fashion. Upon each basic action performed, a reply from an execution environment determines how it proceeds. The possible replies are the Boolean values t
(standing for true) and f (standing for false). Performing the action tau leads
always to the reply t.
BTAbt has one sort: the sort T of threads. We make this sort explicit because
we will extend BTAbt with additional sorts in Section 4. To build terms of sort
T, BTAbt has the following constants and operators:
–
–
–
–
the deadlock constant D : T;
the positive termination constant S+ : T;
the negative termination constant S− : T;
for each a ∈ Atau , the binary postconditional composition operator E a D :
T × T → T.
We assume that there is a countably infinite set of variables of sort T which
includes x, y, z. Terms of sort T are built as usual. We use infix notation for
3
Table 1. Axiom of BTAbt
x E tau D y = x E tau D x T1
Table 2. Approximation induction principle
V
n≥0
πn (x) = πn (y) ⇒ x = y
π0 (x) = D
AIP
P0
πn+1 (S+) = S+
P1a
πn+1 (S−) = S−
P1b
πn+1 (D) = D
P2
πn+1 (x E a D y) = πn (x) E a D πn (y) P3
postconditional composition. We introduce action prefixing as an abbreviation:
a ◦ p, where p is a term of sort T, abbreviates p E a D p.
The thread denoted by a closed term of the form p E a D q will first perform
a, and then proceed as the thread denoted by p if the reply from the execution
environment is t and proceed as the thread denoted by q if the reply from the
execution environment is f. The threads denoted by D, S+ and S− will become
inactive, terminate with Boolean value t and terminate with Boolean value f,
respectively.
BTAbt has only one axiom. This axiom is given in Table 1.
Each closed BTAbt term of sort T denotes a thread that will become inactive
or terminate after it has performed finitely many actions. Infinite threads can be
described by linear recursion. A linear recursive specification over BTAbt is a set
of recursion equations E = {x = tx | x ∈ V }, where V is a set of variables of sort
T and each tx is a BTAbt term of the form D, S+, S− or y E a D z with y, z ∈ V .
We are only interested in models of BTAbt in which linear recursive specifications
have unique solutions. Regular threads, i.e. threads that can only be in a finite
number of states, are solutions of finite linear recursive specifications.
To reason about infinite threads, we assume the infinitary conditional equation AIP (Approximation Induction Principle). AIP is based on the view that
two threads are identical if their approximations up to any finite depth are identical. The approximation up to depth n of a thread is obtained by cutting it
off after it has performed n actions. In AIP, the approximation up to depth n
is phrased in terms of the unary projection operator πn : T → T. AIP and the
axioms for the projection operators are given in Table 2. In this table, a stands
for an arbitrary action from Atau and n stands for an arbitrary natural number.
The behaviours exhibited on execution by PGLBbt instruction sequences
are considered to be regular threads, with the basic instructions taken for basic
actions. The thread extraction operation | | defines, for each PGLBbt instruction
sequence, the behaviour exhibited on execution by that PGLBbt instruction
sequence. The thread extraction operation is defined by |u1 ; . . . ; uk | = |1, u1 ;
. . . ; uk |, where the auxiliary operation | , | is defined by the equations given in
4
Table 3. Defining equations for thread extraction operation
|i, u1 ; . . . ; uk | = D
if not 1 ≤ i ≤ k
|i, u1 ; . . . ; uk | = a ◦ |i + 1, u1 ; . . . ; uk |
if ui = a
|i, u1 ; . . . ; uk | = |i + 1, u1 ; . . . ; uk | E a D |i + 2, u1 ; . . . ; uk | if ui = +a
|i, u1 ; . . . ; uk | = |i + 2, u1 ; . . . ; uk | E a D |i + 1, u1 ; . . . ; uk | if ui = −a
|i, u1 ; . . . ; uk | = |i + l, u1 ; . . . ; uk |
if ui = #l
.
|i, u1 ; . . . ; uk | = |i − l, u1 ; . . . ; uk |
if ui = \#l
|i, u1 ; . . . ; uk | = S+
if ui = !t
|i, u1 ; . . . ; uk | = S−
if ui = !f
Table 3 (for a ∈ A and l, i ∈ N) and the rule that |i, u1 ; . . . ; uk | = D if ui is the
beginning of an infinite jump chain.2
4
Interaction between Threads and Services
A thread may perform a basic action for the purpose of requesting a named
service to process a method and to return a reply value at completion of the
processing of the method. In this section, we extend BTAbt such that this kind
of interaction between threads and services can be dealt with, resulting in TAtsi
bt .
This involves the introduction of service families: collections of named services.
It is assumed that a fixed but arbitrary non-empty finite set M of methods
has been given. Methods play the role of commands. A service is able to process
certain methods. The processing of a method by a service may involve a change
of state of the service and at completion of the processing of the method the
service produces a reply value. The set R of reply values is the set {t, f, d}.
In SF, the algebraic theory of service families introduced below, the following
is assumed with respect to services:
– a set S of services has been given together with:
∂
: S → S;
• for each m ∈ M, a total function ∂m
• for each m ∈ M, a total function ̺m : S → R;
∂
(S) = S
satisfying the condition that there exists a unique S ∈ S with ∂m
and ̺m (S) = d for all m ∈ M;
– a signature ΣS has been given that includes the following sort:
• the sort S of services;
and the following constant and operators:
• the empty service constant δ : S;
∂
• for each m ∈ M, the derived service operator ∂m
: S → S;
– S and ΣS are such that:
• each service in S can be denoted by a closed term of sort S;
2
This rule can be formalized, cf. [2].
5
∂
• the constant δ denotes the unique S ∈ S such that ∂m
(S) = S and
̺m (S) = d for all m ∈ M;
∂
∂
(t) denotes service ∂m
(S).
• if closed term t denotes service S, then ∂m
When a request is made to service S to process method m:
– if ̺m (S) 6= d, then S processes m, produces the reply ̺m (S), and next
∂
(S);
proceeds as ∂m
– if ̺m (S) = d, then S rejects the request to process method m.
∂
The unique service S such that ∂m
(S) = S and ̺m (S) = d for all m ∈ M is
called the empty service. It is the service that is unable to process any method.
It is also assumed that a fixed but arbitrary non-empty finite set F of foci has
been given. Foci play the role of names of services in the service family offered
by an execution environment. A service family is a set of named services where
each name occurs only once.
SF has the sorts, constants and operators in ΣS and in addition the following
sort:
– the sort SF of service families;
and the following constant and operators:
–
–
–
–
the empty service family constant ∅ : SF;
for each f ∈ F , the unary singleton service family operator f. : S → SF;
the binary service family composition operator ⊕ : SF × SF → SF;
for each F ⊆ F , the unary encapsulation operator ∂F : SF → SF.
We assume that there is a countably infinite set of variables of sort SF which
includes u, v, w. Terms are built as usual in the many-sorted case (see e.g. [8,10]).
We use prefix notation for the singleton service family operators and infix notation for the service family composition operator.
The service family denoted by ∅ is the empty service family. The service
family denoted by a closed term of the form f.H consists of one named service
only, the service concerned is the service denoted by H, and the name of this
service is f . The service family denoted by a closed term of the form C ⊕ D
consists of all named services that belong to either the service family denoted by
C or the service family denoted by D. In the case where a named service from
the service family denoted by C and a named service from the service family
denoted by D have the same name, they collapse to an empty service with the
name concerned. The service family denoted by a closed term of the form ∂F (C)
consists of all named services with a name not in F that belong to the service
family denoted by C.
The service family composition operator takes the place of the non-interfering
combination operator from [4]. As suggested by the name, service family composition is composition of service families. Non-interfering combination is composition of services, which has the disadvantage that its usefulness is rather limited
without an additional renaming mechanism.
6
Table 4. Axioms of SF
u⊕∅=u
SFC1
u⊕v =v⊕u
SFC2
∂F (∅) = ∅
SFE1
∂F (f.H) = ∅
if f ∈ F SFE2
(u ⊕ v) ⊕ w = u ⊕ (v ⊕ w) SFC3
∂F (f.H) = f.H
if f ∈
/ F SFE3
f.H ⊕ f.H ′ = f.δ
∂F (u ⊕ v) = ∂F (u) ⊕ ∂F (v)
SFC4
SFE4
The axioms of SF are given in Table 4. In this table, f stands for an arbitrary
focus from F and H and H ′ stand for arbitrary closed terms of sort S. The
axioms of SF simply formalize the informal explanation given above.
Below we will introduce two operators related to the interaction between
threads and services. They are called the apply operator and the reply operator.
The apply operator is concerned with the effects of threads on service families
and therefore produces service families. The reply operator is concerned with the
effects of service families on the Boolean values that threads deliver at their termination. The reply operator does not only produce Boolean values: it produces
a special value in cases where no termination takes place.
For the set A of basic actions, we take the set {f.m | f ∈ F , m ∈ M}. Both
operators mentioned above relate to the processing of methods by services from
a service family in pursuance of basic actions performed by a thread. The service
involved in the processing of a method is the service whose name is the focus of
the basic action in question.
TAtsi
bt has the sorts, constants and operators of both BTAbt and SF, and in
addition the following sort:
– the sort R of replies;
and the following constants and operators:
– the reply constants t, f, d : R;
– the binary apply operator • : T × SF → SF;
– the binary reply operator ! : T × SF → R.
We use infix notation for the apply and reply operators.
The service family denoted by a closed term of the form p • C and the reply
denoted by a closed term of the form p ! C are the service family and reply,
respectively, that result from processing the method of each basic action with a
focus of the service family denoted by C that the thread denoted by p performs,
where the processing is done by the service in that service family with the focus
of the basic action as its name. When the method of a basic action performed
by a thread is processed by a service, the service changes in accordance with the
method concerned, and affects the thread as follows: the two ways to proceed
reduces to one on the basis of the reply value produced by the service. The reply
is the Boolean value that the thread denoted by p delivers at termination if it
terminates and the value d (standing for divergent) if it does not terminate.
7
Table 5. Axioms for apply operator
S+ • u = u
A1
S− • u = u
A2
D•u=∅
A3
(tau ◦ x) • u = x • u
A4
(x E f.m D y) • ∂{f } (u) = ∅
(x E f.m D y) • (f.H ⊕ ∂{f } (u)) =
(x E f.m D y) • (f.H ⊕ ∂{f } (u)) =
A5
∂
x • (f. ∂m
H
∂
y • (f. ∂m H
⊕ ∂{f } (u)) if ̺m (H) = t A6
⊕ ∂{f } (u)) if ̺m (H) = f A7
(x E f.m D y) • (f.H ⊕ ∂{f } (u)) = ∅
V
n≥0 πn (x) • u = πn (y) • v ⇒ x • u = y • v
if ̺m (H) = d A8
A9
Table 6. Axioms for reply operator
S+ ! u = t
R1
S− ! u = f
R2
D!u=d
R3
(tau ◦ x) ! u = x ! u
R4
(x E f.m D y) ! ∂{f } (u) = d
R5
∂
H ⊕ ∂{f } (u)) if ̺m (H) = t R6
(x E f.m D y) ! (f.H ⊕ ∂{f } (u)) = x ! (f. ∂m
∂
H ⊕ ∂{f } (u)) if ̺m (H) = f R7
(x E f.m D y) ! (f.H ⊕ ∂{f } (u)) = y ! (f. ∂m
(x E f.m D y) ! (f.H ⊕ ∂{f } (u)) = d
V
n≥0 πn (x) ! u = πn (y) ! v ⇒ x ! u = y ! v
if ̺m (H) = d R8
R9
The axioms of TAtsi
bt are the axioms of BTAbt , the axioms of SF, and the
axioms given in Tables 5 and 6. In these tables, f stands for an arbitrary focus
from F , m stands for an arbitrary method from M, H stands for an arbitrary
term of sort S, and n stands for an arbitrary natural number. The axioms simply
formalize the informal explanation given above and in addition stipulate what
is the result of apply and reply if inappropriate foci or methods are involved.
Axioms A9 and R9 allow for reasoning about infinite threads in the contexts of
apply and reply, respectively.
5
Functional Units
In this section, we introduce the concept of a functional unit and related concepts
such as a functional unit degree.
It is assumed that a non-empty set S of states has been given. As before, it is
assumed that a non-empty finite set M of methods has been given. However, in
the setting of functional units, methods serve as names of operations on a state
space. For that reason, the members of M will henceforth be called method
names.
8
A method operation on S is a total function from S to B×S . A partial method
operation on S is a partial function from S to B × S . We write MO(S ) for the
set of all method operations on S . We write M r and M e , where M ∈ MO(S ),
for the unique functions R : S → B and E : S → S , respectively, such that
M (s) = (R(s), E(s)) for all s ∈ S .
A functional unit for S is a finite subset H of M × MO(S ) such that
(m, M ) ∈ H and (m, M ′ ) ∈ H implies M = M ′ . We write F U(S ) for the set
of all functional units for S . We write I(H), where H ∈ F U(S ), for the set
{m ∈ M | ∃M ∈ MO(S ) • (m, M ) ∈ H}. We write mH , where H ∈ F U(S ) and
m ∈ I(H), for the unique M ∈ MO(S ) such that (m, M ) ∈ H.
We look upon the set I(H), where H ∈ F U (S ), as the interface of H. It looks
to be convenient to have a notation for the restriction of a functional unit to a
subset of its interface. We write (I, H), where H ∈ F U(S ) and I ⊆ I(H), for
the functional unit {(m, M ) ∈ H | m ∈ I}.
Let H ∈ F U(S ). Then an extension of H is an H′ ∈ F U (S ) such that H ⊆ H′ .
The following is a simple illustration of the use of functional units. An unbounded counter can be modelled by a functional unit for N with method operations for set to zero, increment by one, decrement by one, and test on zero.
According to the definition of a functional unit, ∅ ∈ F U(S ). By that we have
a unique functional unit with an empty interface, which is not very interesting
in itself. However, when considering services that behave according to functional
units, ∅ is exactly the functional unit according to which the empty service δ
(the service that is not able to process any method) behaves.
The method names attached to method operations in functional units should
not be confused with the names used to denote specific method operations in
describing functional units. Therefore, we will comply with the convention to use
names beginning with a lower-case letter in the former case and names beginning
with an upper-case letter in the latter case.
We will use PGLBbt instruction sequences to derive partial method operations from the method operations of a functional unit. We write L(f.I),
where I ⊆ M, for the set of all PGLBbt instruction sequences, taking the set
{f.m | m ∈ I} as the set A of basic instructions.
The derivation of partial method operations from the method operations
of a functional unit involves services whose processing of methods amounts to
replies and service changes according to corresponding method operations of
the functional unit concerned. These services can be viewed as the behaviours
of a machine, on which the processing in question takes place, in its different
states. We take the set F U(S ) × S as the set S of services. We write H(s), where
∂
H ∈ F U(S ) and s ∈ S , for the service (H, s). The functions ∂m
and ̺m are
defined as follows:
(
H(meH (s)) if m ∈ I(H)
∂
(H(s))
=
∂m
∅(s′ )
if m ∈
/ I(H) ,
(
mrH (s)
if m ∈ I(H)
̺m (H(s)) =
d
if m ∈
/ I(H) ,
9
where s′ is a fixed but arbitrary state in S. We assume that each H(s) ∈ S can
be denoted by a closed term of sort S. In this connection, we use the following
notational convention: for each H(s) ∈ S, we write H(s) for an arbitrary closed
term of sort T that denotes H(s). The ambiguity thus introduced could be
obviated by decorating H(s) wherever it stands for a closed term. However, in
this paper, it is always immediately clear from the context whether it stands for
a closed term. Moreover, we believe that the decorations are more often than
not distracting. Therefore, we leave it to the reader to make the decorations
mentally wherever appropriate.
Let H ∈ F U(S ), and let I ⊆ I(H). Then an instruction sequence x ∈ L(f.I)
produces a partial method operation |x|H as follows:
|x|H (s) = (|x|rH (s), |x|eH (s)) if |x|rH (s) = t ∨ |x|rH (s) = f ,
|x|H (s) is undefined
if |x|rH (s) = d ,
where
|x|rH (s) = x ! f.H(s) ,
|x|eH (s) = the unique s′ ∈ S such that x • f.H(s) = f.H(s′ ) .
If |x|H is total, then it is called a derived method operation of H.
The binary relation ≤ on F U(S ) is defined by H ≤ H′ iff for all (m, M ) ∈ H,
M is a derived method operation of H′ . The binary relation ≡ on F U(S ) is
defined by H ≡ H′ iff H ≤ H′ and H′ ≤ H.
Theorem 1.
1. ≤ is transitive;
2. ≡ is an equivalence relation.
Proof. Property 1: We have to prove that H ≤ H′ and H′ ≤ H′′ implies H ≤ H′′ .
It is sufficient to show that we can obtain instruction sequences in L(f.I(H′′ ))
that produce the method operations of H from the instruction sequences in
L(f.I(H′ )) that produce the method operations of H and the instruction sequences in L(f.I(H′′ )) that produce the method operations of H′ . Without loss
of generality, we may assume that all instruction sequences are of the form
u1 ; . . . ; uk ; !t ; !f, where, for each i ∈ [1, k], ui is a positive test instruction, a
forward jump instruction or a backward jump instruction. Let m ∈ I(H), let M
be such that (m, M ) ∈ H, and let xm ∈ L(f.I(H′ )) be such that M = |xm |H′ .
Suppose that I(H′ ) = {m′1 , . . . , m′n }. For each i ∈ [1, n], let Mi′ be such that
(m′i , Mi′ ) ∈ H′ and let xm′i = ui1 ; . . . ; uiki ; !t ; !f ∈ L(f.I(H′′ )) be such that
Mi′ = |xm′i |H′′ . Consider the x′m ∈ L(f.I(H′′ )) obtained from xm as follows: for
each i ∈ [1, n], (i) first increase each jump over the leftmost occurrence of +f.m′i
in xm with ki + 1, and next replace this instruction by ui1 ; . . . ; uiki ; (ii) repeat
the previous step as long as their are occurrences of +f.m′i . It is easy to see that
M = |x′m |H′′ .
10
Property 2: It follows immediately from the definition of ≡ that ≡ is symmetric and from the definition of ≤ that ≤ is reflexive. From these properties,
Property 1 and the definition of ≡, it follows immediately that ≡ is symmetric,
reflexive and transitive.
⊓
⊔
The members of the quotient set F U(S )/≡ are called functional unit degrees.
Let H ∈ F U(S ) and D ∈ F U(S )/≡. Then D is a functional unit degree below H
if there exists an H′ ∈ D such that H′ ≤ H.
6
Functional Units for Natural Numbers
In this section, we investigate functional units for natural numbers. The main
consequences of considering the special case where the state space is N are the
following: (i) N is infinite, (ii) there is a notion of computability known which
can be used without further preparations.
An example of a functional unit in F U(N) is an unbounded counter. The
method names involved are setzero, succ, pred, and iszero. The method operations
involved are the functions Setzero, Succ, Pred , Iszero : N → B × N defined as
follows:
Setzero(x) = (t, 0) ,
= (t, x + 1) ,
(
(t, x − 1)
Pred (x) =
(f, 0)
(
(t, x)
Iszero(x) =
(f, x)
Succ(x)
if x > 0 ,
if x = 0 ,
if x = 0 ,
if x > 0 .
The functional unit Counter is defined as follows:
Counter = {(setzero, Setzero), (succ, Succ), (pred, Pred), (iszero, Iszero)} .
Proposition 1. There are infinitely many functional unit degrees below
({pred, iszero} , Counter ).
Proof. For each n ∈ N, we define a functional unit Hn ∈ F U(N) such that
Hn ≤ ({pred, iszero} , Counter ) as follows:
Hn = {(pred:n, Pred :n), (iszero, Iszero)} ,
where
Pred :n(x) =
(t, x − n) if x ≥ n
(f, 0)
if x < n .
Let n, m ∈ N be such that n < m. Then Pred :n(m) = (t, m − n). However, there
does not exist an x ∈ L(f.I(Hm )) such that |x|Hm (m) = (t, m − n) because
Pred :m(m) = (t, 0). Hence, Hn 6≤ Hm for all n, m ∈ N with n < m.
⊓
⊔
11
A method operation M ∈ MO(N) is computable if there exist computable
functions F, G : N → N such that M (n) = (β(F (n)), G(n)) for all n ∈ N, where
β : N → B is inductively defined by β(0) = t and β(n + 1) = f. A functional unit
H ∈ F U (N) is computable if, for each (m, M ) ∈ H, M is computable.
Theorem 2. Let H, H′ ∈ F U(N) be such that H ≤ H′ . Then H is computable
if H′ is computable.
Proof. We will show that all derived method operations of H′ are computable.
Take an arbitrary P ∈ L(f.I(H′ )) such that |P |H′ is a derived method operations of H′ . It follows immediately from the definition of thread extraction
that |P | is the solution of a finite linear recursive specification over BTAbt , i.e.
a finite guarded recursive specification over BTAbt in which the right-hand side
of each equation is a BTAbt term of the form D, S+, S− or x E a D y where
x and y are variables of sort T. Let E be a finite linear recursive specification
over BTAbt of which the solution for x1 is |P |. Because |P |H′ is total, it may be
assumed without loss of generality that D does not occur as the right-hand side
of an equation in E. Suppose that
E = xi = xl(i) E f.mi D xr(i) | i ∈ [1, n] ∪ {xn+1 = S+, xn+2 = S−} .
From this set of equations, using the relevant axioms and definitions, we obtain
a set of equations of which the solution for F1 is |P |eH′ :
Fi (s) = Fl(i) (mi eH′ (s)) · sg(χi (s)) + Fr(i) (mi eH′ (s)) · sg(χi (s)) | i ∈ [1, n]
∪ {Fn+1 (s) = s, Fn+2 (s) = s} ,
where, for every i ∈ [1, n], the function χi : N → N is such that for all s ∈ N:
χi (s) = 0 ⇔ mi rH′ (s) = t ,
and the functions sg, sg : N → N are defined as usual:
sg(0)
=0,
sg(n + 1) = 1 ,
sg(0)
=1,
sg(n + 1) = 0 .
It follows from the way in which this set of equations is obtained from E, the
fact that mi eH′ and χi are computable for each i ∈ [1, n], and the fact that sg and
sg are computable, that this set of equations is equivalent to a set of equations
by which |P |eH′ is defined recursively in the sense of Kleene (see [5]). This means
that |P |eH′ is general recursive, and hence computable.
In a similar way, it is proved that |P |rH′ is computable.
⊓
⊔
A computable H ∈ F U(N) is universal if for each computable L ∈ F U(N),
we have L ≤ H. There exists a universal computable functional unit for natural
numbers.
Theorem 3. There exists a computable H ∈ F U(N) that is universal.
12
Proof. We will show that there exists a computable H ∈ F U(N) with the property that each computable M ∈ MO(N) is a derived method operation of H.
As a corollary of Theorem 10.3 from [9],3 we have that each computable
M ∈ MO(N) can be computed by means of a register machine with six registers, say r0, r1, r2, r3, r4, and r5. The registers are used as follows: r0 as input
register; r1 as output register for the output in B; r2 as output register for the
output in N; r3, r4 and r5 as auxiliary registers. The content of r1 represents the
Boolean output as follows: 0 represents t and all other natural numbers represent f. For each i ∈ [0, 5], register ri can be incremented by one, decremented
by one, and tested for zero by means of instructions ri.succ, ri.pred and ri.iszero,
respectively. We write L(RM6 ) for the set of all PGLBbt instruction sequences,
taking the set {ri.succ, ri.pred, ri.iszero | i ∈ [0, 5]} as the set A of basic instructions. Clearly, L(RM6 ) is adequate to represent all register machine programs
using six registers.
We define a computable functional unit U ∈ F U(N) whose method operations can simulate the effects of the register machine instructions by encoding
the register machine states by natural numbers such that the contents of the registers can reconstructed by prime factorization. This functional unit is defined
as follows:
U = {(exp2, Exp2 ), (fact5, Fact5 )}
∪ {(ri:succ, Ri:succ), (ri:pred, Ri:pred ), (ri:iszero, Ri:iszero) | i ∈ [0, 5]} ,
where the method operations are defined as follows:
Exp2 (x) = (t, 2x ) ,
Fact5 (x) = (t, max {y | ∃z • x = 5y · z})
and, for each i ∈ [0, 5]:4
Ri:succ(x) = (t, pi · x) ,
(
(t, x/pi )
Ri:pred (x) =
(f, x)
(
(t, x)
Ri:iszero(x) =
(f, x)
if pi | x
if ¬(pi | x) ,
if ¬(pi | x)
if pi | x ,
where pi is the (i+1)th prime number, i.e. p0 = 2, p1 = 3, p2 = 5, . . . .
We define a function rml2ful from L(RM6 ) to L(f.I(U)), which gives, for
each instruction sequence P in L(RM6 ), the instruction sequence in L(f.I(U))
by which the effect produced by P on a register machine with six registers can
be simulated on U. This function is defined as follows:
rml2ful(u1 ; . . . ; uk )
= f.exp2 ; φ(u1 ) ; . . . ; φ(uk ) ; −f.r1:iszero ; #3 ; f.fact5 ; !t ; f.fact5 ; !f ,
3
4
That theorem can be looked upon as a corollary of Theorem Ia from [6].
As usual, we write x | y for y is divisible by x.
13
where
φ(a)
= ψ(a) ,
φ(+a) = +ψ(a) ,
φ(−a) = −ψ(a) ,
φ(u)
=u
if u is a jump or termination instruction ,
where, for each i ∈ [0, 5]:
ψ(ri.succ) = f.ri:succ ,
ψ(ri.pred) = f.ri:pred ,
ψ(ri.iszero) = f.ri:iszero .
Take an arbitrary computable M ∈ MO(N). Then there exist an instruction
sequence in L(RM6 ) that computes M . Take an arbitrary P ∈ L(RM6 ) that
computes M . Then |rml2ful(P )|U = M . Hence, M is a derived method operation of U.
⊓
⊔
The universal computable functional unit U defined in the proof of Theorem 3
has 20 method operations. However, three method operations suffice.
Theorem 4. There exists a computable H ∈ F U(N) with only three method
operations that is universal.
Proof. We know from the proof of Theorem 3 that there exists a computable
H ∈ F U(N) with 20 method operations, say M0 , . . . , M19 . We will show that
there exists a computable H′ ∈ F U(N) with only three method operations such
that H ≤ H′ .
We define a computable functional unit U ′ ∈ F U(N) with only three method
operations such that U ≤ U ′ as follows:
U ′ = {(g1, G1), (g2, G2), (g3, G3)} ,
where the method operations are defined as follows:
G1(x) = (t, 2x ) ,
19
•
(t, 3 · x) if ¬(3 | x) ∧ ∀y (y | x ⇒ (y = 2 ∨ y = 3))
(t, x/319 ) if 319 | x ∧ ¬(320 | x) ∧ ∀y • (y | x ⇒ (y = 2 ∨ y = 3))
G2(x) =
(f, 0)
if 320 | x ∨ ¬∀y • (y | x ⇒ (y = 2 ∨ y = 3)) ,
G3(x) = Mfact3 (x) (fact2 (x)) ,
where
fact2 (x) = max {y | ∃z • x = 2y · z} ,
fact3 (x) = max {y | ∃z • x = 3y · z} .
14
We have that, for each i ∈ [0, 19], |f.g1 ; f.g2 i ; +f.g3 ; !t ; !f|U ′ = Mi .5 Hence,
M0 , . . . , M19 are derived method operations of U ′ .
⊓
⊔
The universal computable functional unit U ′ defined in the proof of Theorem 4
has three method operations. We can show that one method operation does not
suffice.
Theorem 5. There does not exist a computable H ∈ F U (N) with only one
method operation that is universal.
Proof. We will show that there does not exist a computable H ∈ F U(N) with
one method operation such that Counter ≤ H. Here, Counter is the functional
unit introduced at the beginning of this section.
Assume that there exists a computable H ∈ F U(N) with one method operation such that Counter ≤ H. Let H′ ∈ F U(N) be such that H′ has one method
operation and Counter ≤ H′ , and let m be the unique method name such that
I(H′ ) = {m}. Take arbitrary P1 , P2 ∈ L(f.I(H′ )) such that |P1 |H′ = Succ and
|P2 |H′ = Pred . Then |P1 |H′ (0) = (t, 1) and |P2 |H′ (1) = (t, 0). Instruction f.m is
processed at least once if P1 is applied to H′ (0) or P2 is applied to H′ (1). Let k0
be the number of times that instruction f.m is processed on application of P1
to H′ (0) and let k1 be the number of times that instruction f.m is processed on
application of P2 to H′ (1) (irrespective of replies). Then, from state 0, state 0 is
reached again after f.m is processed k0 +k1 times. Thus, by repeated application
of P1 to H′ (0) at most k0 + k1 different states can be reached. This contradicts
with |P1 |H′ = Succ. Hence, there does not exist a computable H ∈ F U(N) with
one method operation such that Counter ≤ H.
⊓
⊔
It is an open problem whether two method operations suffice.
7
Functional Units for Finite State Spaces
In this short section, we make some remarks about functional units for finite
state spaces.
In the special case where the state space is B, the state space consists of
only two states. Because there are four possible unary functions on B, there are
precisely 16 method operations in MO(B). There are in principle 216 different
functional units in F U(B), for it is useless to include the same method operation
more than once under different names in a functional unit. This means that 216 is
an upper bound of the number of functional unit degrees in F U(B)/≡. However,
it is straightforward to show that F U(B)/≡ has only 12 different functional unit
degrees.
In the more general case of a finite state space consisting of k states, say Sk ,
k k
there are in principle 22 ·k different functional units in F U(Sk ). Already with
k = 3, it becomes unclear whether the number of functional unit degrees in
F U(Sk ) can be determined manually. Actually, we do not know at the moment
whether it can be determined with computer support either.
5
For each primitive instruction u, the instruction sequence un is defined by induction
on n as follows: u0 = #1, u1 = u and un+2 = u ; un+1 .
15
8
Concluding Remarks
We have defined the concept of a functional unit for a state space and have
established general results concerning functional units for natural numbers. The
main result is the existence of a universal computable functional unit for natural
numbers. The case where the state space is the set of natural numbers is arguably
the simplest significant case. We have not yet investigated other significant cases.
An interesting case is the one where the state space is the set of all pairs of
sequences over some alphabet: the tape of a Turing machine can be modelled by
a functional unit for this state space. Each Turing machine can be simulated by
means of a functional unit that corresponds to the tape of the Turing machine
and a PGLBbt instruction sequence that corresponds to the finite control of the
Turing machine. Variations of the Turing machine theme can be dealt with in
this way as well. Thus, functional units allows for many computability issues to
be viewed as issues about programs rather than machines.
In [3], we introduce an extension of program algebra with Boolean termination instructions, called PGAbt , and define a thread extraction operation for it.
PGLBbt instruction sequences can be translated into closed PGAbt terms such
that thread extraction for PGLBbt yields the same behaviours as translation followed by thread extraction for PGAbt . In [3], we also introduce an extension of
basic thread algebra similar to TAtsi
bt . In addition to the constants and operators
of TAtsi
bt , that extension has a constant (S) for termination without delivery of a
Boolean value and an operator (/) which is concerned with the effects of service
families on threads and therefore produces threads.
References
1. Bergstra, J.A., Loots, M.E.: Program algebra for sequential code. Journal of Logic
and Algebraic Programming 51(2), 125–156 (2002)
2. Bergstra, J.A., Middelburg, C.A.: Program algebra with a jump-shift instruction.
Journal of Applied Logic 6(4), 553–563 (2008)
3. Bergstra, J.A., Middelburg, C.A.: Instruction sequence processing operators.
arXiv:0910.5564v2 [cs.LO] (2009)
4. Bergstra, J.A., Ponse, A.: Combining programs and state machines. Journal of
Logic and Algebraic Programming 51(2), 175–192 (2002)
5. Kleene, S.C.: General recursive functions of natural numbers. Mathematische Annalen 112, 727–742 (1936)
6. Minsky, M.L.: Recursive unsolvability of Post’s problem of “tag” and other topics
in theory of Turing machines. Annals of Mathematics 74(3), 437–455 (1961)
7. Ponse, A., van der Zwaag, M.B.: An introduction to program and thread algebra.
In: Beckmann, A., et al. (eds.) CiE 2006. Lecture Notes in Computer Science, vol.
3988, pp. 445–458. Springer-Verlag (2006)
8. Sannella, D., Tarlecki, A.: Algebraic preliminaries. In: Astesiano, E., Kreowski,
H.J., Krieg-Brückner, B. (eds.) Algebraic Foundations of Systems Specification,
pp. 13–30. Springer-Verlag, Berlin (1999)
9. Shepherdson, J.C., Sturgis, H.E.: Computability of recursive functions. Journal of
the ACM 10(2), 217–255 (1963)
16
10. Wirsing, M.: Algebraic specification. In: van Leeuwen, J. (ed.) Handbook of Theoretical Computer Science, vol. B, pp. 675–788. Elsevier, Amsterdam (1990)
17
| 6cs.PL
|
1
The evolution of representation in simple cognitive
networks
Lars Marstaller1, ? , Arend Hintze2, 3, 4, ? & Christoph Adami2, 3
arXiv:1206.5771v2 [q-bio.NC] 6 Aug 2013
1 Department of Cognitive Science, Macquarie University, Sydney, Australia
2 Microbiology & Molecular Genetics, Michigan State University, East Lansing, MI
3 BEACON Center for the Study of Evolution in Action, Michigan State University,
East Lansing, MI
4 Computer Science & Engineering, Michigan State University, East Lansing, MI ∗
Keywords: Information, Representation, Cognition, Evolution
Abstract
Representations are internal models of the environment that can provide guidance to a
behaving agent, even in the absence of sensory information. It is not clear how representations are developed and whether or not they are necessary or even essential for
intelligent behavior. We argue here that the ability to represent relevant features of the
environment is the expected consequence of an adaptive process, give a formal definition of representation based on information theory, and quantify it with a measure R.
To measure how R changes over time, we evolve two types of networks—an artificial
neural network and a network of hidden Markov gates—to solve a categorization task
using a genetic algorithm. We find that the capacity to represent increases during evolutionary adaptation, and that agents form representations of their environment during
their lifetime. This ability allows the agents to act on sensorial inputs in the context
of their acquired representations and enables complex and context-dependent behavior.
We examine which concepts (features of the environment) our networks are representing, how the representations are logically encoded in the networks, and how they form
as an agent behaves to solve a task. We conclude that R should be able to quantify
∗
These authors contributed equally.
the representations within any cognitive system, and should be predictive of an agent’s
long-term adaptive success.
1
Introduction
The notion of representation is as old as cognitive science itself (see, e.g., Chomsky,
1965; Newell and Simon, 1972; Fodor, 1975; Johnson-Laird and Wason, 1977; Marr,
1982; Pinker, 1989; Pitt, 2008), but its usefulness for Artificial Intelligence (AI) research has been doubted (Brooks, 1991). In his widely cited article “Intelligence without representation”, Brooks argued instead for a subsumption architecture where the
autonomous behavior producing components (or layers) of the cognitive system directly interface with the world and with each other rather than with a central symbol
processor dealing in explicit representations of the environment. In particular, inspired
by the biological path to intelligence, Brooks argued that AI research needs to be rooted
in mobile autonomous robotics and a direct interaction between action and perception.
Echoing Moravec (1984), he asserted that the necessary elements for the development
of intelligence are mobility, acute vision, and the ability to behave appropriately in a dynamic environment (Brooks, 1991). This architecture achieved insect-level intelligence
and Brooks argued that a path to higher level AI could be forged by incrementally increasing the complexity of subsumption architecture.
However, 20 years after advocating such a radical departure from the classical approach to AI, the subsumption approach seems to have stalled as well. We believe that
the reason for the lack of progress does not lie in the attempt to base AI research in
mobile autonomous robots, but that instead representations (also sometimes called “internal models”, (Craik, 1943; Wolpert et al., 1995; Kawato, 1999)) are key to complex
adaptive behavior. Indeed, while representation-free robotics has made some important strides (Nolfi, 2002), it is limited to problems that are not “representation hungry” (Clark, 1997), i.e., problems that do not require past information or additional
(external) knowledge about the current context. In addition, the technical difficulty
of developing a subsumption architecture increases with the number of layers or subsystems. This problem of subsumption architecture mirrors the difficulties of classic
representational AI approaches to build accurate and appropriate models of the world.
2
An alternative approach to engineering cognitive architectures and internal models
is evolutionary robotics (Nolfi and Floreano, 2000). Instead of designing the structure
or functions of a control architecture, principles of Darwinian evolution are used to
create complex networks that interface perception and action in non-obvious and often surprising ways. Such structures can give rise to complex representations of the
environment that are hard to engineer and equally hard to analyze (see, e.g. Floreano
and Mondada,1996). Evolved representations provide context, are flexible, and can be
readjusted given new stimuli that contradict the current assumptions. Representations
can be updated during the lifetime or over the course of evolution and thus are able to
handle even new sensory input (Bongard et al., 2006). We argue that as robots evolve
to behave appropriately (and survive) in a dynamic and noisy world, representations
of the environment emerge within the cognitive apparatus, and are integrated with the
perceived sensory data to create intelligent behavior–using not only the current state of
the environment but crucially taking into account historical data (memory) as well.
To test this hypothesis and make internal representations of evolved systems accessible to analysis, we propose a new information-theoretic measure of the degree to which
an embodied agent represents its environment within its internal states and show how
the capacity to represent environmental features emerges over thousands of generations
of simulated evolution. The main idea is that representations encode environmental features because of their relevance for the cognitive system in question (Clark and Toribio,
1994). Hence, for our purposes, representations can be symbolic or sub-symbolic (e.g.,
neural states) as long as they have a physical basis, i.e., as long as they are encoded
in measurable internal states. However, we distinguish representations from sensorial
input because sensor inputs cannot provide the same past or external context as internal states. We thus explicitly define representations as that information about relevant
features of the environment which is encoded in the internal states of an organism and
which goes beyond the information present in its sensors (Haugeland, 1991; Clark,
1997). In particular, this implies that representations can, at time, misrepresent (Haugeland, 1991)–unlike information present in sensors, which always truthfully correlates
with the environment. To illuminate the functioning of evolved cognitive systems, we
show how it is in principle possible to determine what a representation is about, and
how representations form during the lifetime of an agent. We argue that our measure
3
provides a valuable tool to investigate the organization of evolved cognitive systems
especially in cases where internal representations are “epistemically opaque”.
2
Methods
2.1
Information-theoretic measure of representation
Information theory has been used previously to quantify how context can modulate decisions based on sensory input (Phillipps et al., 1994; Phillips and Singer, 1997; Kay
and Phillips, 2011). Here, we present an information-theoretic construction that explicitly takes the entropy of environmental states into account. To quantify representation,
we first define the relationship between the representing system and the represented environment in terms of information (shared, or mutual, entropy). Information measures
the correlation between two random variables, while the entropy H is a measure of the
uncertainty we have about a random variable in the absence of information (uncertainty
is therefore potential information). For a random variable X that can take on the states
xi with probabilities p(xi ) = P (X = xi ), the entropy is given by (Shannon, 1948):
H(X) = −
N
X
p(xi ) log p(xi ) ,
(1)
i=1
where N is the number of possible states that X can take on.
The information between two random variables characterizes how much the degree
of order in one of the variables is predictive of the regularity in the other variable. It
can be defined using entropy as the difference between the sum of the entropies of two
random variables X and Y [written as H(X) and H(Y )] and the joint entropy of X and
Y , written as H(X, Y ):
I(X : Y ) = H(X) + H(Y ) − H(X, Y ) =
X
xy
p(x, y) log
p(x, y)
.
p(x)p(y)
(2)
In Eq. (2), p(x) and p(y) are the probability distributions for the random variables X
and Y respectively [that is, p(x) = P (X = x)], while p(x, y) is the joint probability
distribution of the (joint) random variable XY . The shared entropy I(X : Y ) can also
be written in terms of a difference between unconditional and conditional entropies, as
I(X : Y ) = H(X) − H(X|Y ) = H(Y ) − H(Y |X) .
4
(3)
This definition reminds us that information is that which reduces our uncertainty about
a system. In other words, it is that which allows us to make predictions about a system
with an accuracy that is higher than when we did not have that information. In Eq. (3),
we introduced the concept of a conditional entropy (Shannon, 1948). For example,
H(X|Y ) (read as “H of X given Y ”) is the entropy of X when the state of the variable
Y is known, and is calculated as
H(X|Y ) = −
X
p(x, y) log p(x|y) ,
(4)
xy
using the conditional probability p(x|y) = p(x, y)/p(y).
In general, information is able to detect arbitrary correlations between signals or
sets of events. We assume here that such correlations instantiate semiotic or information relationships between a representing and represented, and use mutual information
to measure the correlation between a network’s internal states and its environment [see
also Marstaller et al. (2010)]. So, for example, we could imagine that X stands for the
states of an environment, whereas Y is a variable that represents those states of the environment. We need to be careful, however, to exclude from possible representational
variables those that are mere images of the environment, such as the trace that the world
leaves in an agent’s sensors. Indeed, mere correlations between internal states and the
environment are not sufficient to be treated as representational because they could be
due to behavior that is entirely reactive (Clark, 1997). Haugeland, for example, understands representation as something that “stands in” for something in the environment,
but that is no longer reflected in the perceptual system of the agent (Haugeland, 1991).
Indeed, representation should be different from a mere translation: Consider a digital
camera’s relationship with its environment. The photo chip guarantees a one-to-one
mapping between the environment structure and the camera’s state patterns. But a
camera is not able to adapt to its environment. By taking a picture, the camera has
not ‘learned’ anything about its environment that will affect its future state. It simply
stores what it received through its inputs without extracting information from it, i.e., the
camera’s internal states are fully determined by its sensor inputs. Representation goes
beyond mere translation because the content, i.e. which feature of the environment is
represented, depends on the goals of the system. Not everything is represented in the
same way. A camera does not have this functional specification of its internal states.
5
To rule out trivial representations like a camera’s internal states, we define representation as the shared entropy between environment states and internal states, but given
the sensor states, i.e., conditioned on the sensors. Thus, representation is that part of the
shared entropy between environment states and internal states that goes beyond what is
seen in the sensors (see Fig. 1). For the following, we take E, given by its probability distribution p(ei ) = P (E = ei ), as the random variable to describe environmental
states, while S describes sensor states. If the internal states of the agent (hidden and
output states) are characterized by the random variable M with probability distribution
p(mj ) = P (M = mj ), then we define the representation R as (for an earlier version,
see Marstaller et al.,2010):
R = H(E : M |S) = I(E : M ) − I(E : M : S) = Hcorr − I(S : E) − I(S : M ) , (5)
where the correlation entropy Hcorr of the three variables E, S, and M [also called “total correlation” (Watanabe, 1960) or “multi-information” (McGill, 1954; Schneidman
et al., 2003)] is the amount of information they all three share:
Hcorr = H(E) + H(M ) + H(S) − H(E, M, S) .
(6)
In Eq. (5), we introduced the shared conditional entropy between three variables that is
defined as the difference between an information that is unshared and one that is shared
(with a third system), just as H(X|Y ) = H(X) − I(X : Y ), from Eq. (4). Thus, the
representation R of the world E within internal states M is the total correlation between
the three, but without what is reflected in S about E and M , respectively [measured by
I(S : E) and I(S : M )]. The relationship between R and the entropies of the three
variables S, E, and M is most conveniently summarized by an entropy Venn diagram,
as in Fig. 1. In these diagrams, a circle is a quantitative measure of the entropy of the
associated variable, and the shared entropy between two variables is represented by the
intersection of the variables, and so on (see, e.g., Cover and Thomas,1991).
Our information-theoretic definition of representation carries over from discrete
variables to continuous variables unchanged, as can be seen as follows. Let XE , XM ,
and XS be random variables defined with normalized probability density functions
fE (e), fM (m), and fS (s). The differential entropy h(X) (Cover and Thomas, 1991),
6
defined as
h(X) = −
Z
f (x) log f (x)dx
(7)
S
where S is the support of random variable X, is related to the discretized version
H(X ∆ ) by noting that
H(X ∆ ) + log ∆ → h(X)
(∆ → 0) ,
(8)
where we introduced the discretization
Z (i+1)∆
f (x)dx
pi =
(9)
i∆
in order to define the discretized Shannon entropy H(X ∆ ) = −
P
i
pi log pi . This
implies that an n-bit quantization of a continuous random variable is approximately
H(X) ≈ h(X) + n (Cover and Thomas, 1991). Let us now assume that the variables
XE , XM , and XS are each quantized by nE , nM , and nS bits respectively. Because
H(E, M, S) is then quantized by nE + nM + nS bits, it follows that Hcorr ≈ hcorr , that
is, the continuous and discrete variable correlation entropies are (in the limit of sufficiently small ∆) approximately the same because the discretization correction cancels.
The same is true for the informations I(S : E) and I(S : M ), as these are correlation
entropies between two variables. Thus, H(E : M |S) ≈ h(XE : XM |XS ), the differential entropy version of R. We stress that while an exact identity between discrete and
continuous variable definitions of R is only ensured in the limit of vanishing discretization, the cancellation of the correction terms log ∆ implies that the discrete version is
not biased with respect to the continuous version.
R defines a relation between a network’s activity patterns and its environment as
the result of information processing. R yields a positive quantity, measured in bits (if
logarithms are taken to base 2). In order to show that this measure of representation
reflects functional purpose (Clark, 1997), we evolve cognitive systems (networks) that
control the behavior of an embodied agent, and show that fitness, a measure for the
agent’s functional prowess, is correlated with R. In other words, we show that when the
environment (and task) is complex enough, agents react to this challenge by evolving
representations of that environment.
7
E
H(E|S, M )
M
R
I(E : S|M )
H(S|E, M )
H(M |E, S)
I(S : M |E)
S
I(E : M : S)
Figure 1: Venn diagram of entropies and informations for the three random variable E,
S, and M , describing the states of the environment, sensors, and agent internal degrees
of freedom. The representation R = H(E : M |S) is shaded.
2.2
Evolution of Active Categorical Perception
We study the evolution of an agent that solves an active categorical perception (ACP)
task (Beer, 1996, 2003), but with modifications suggested by van Dartel et al. (2005)
(see also van Dartel M.F.,2005). Categorization is thought to be one of the key elements
in cognition (see Harnad 1987; Cohen and Lefebvre 2005). In categorical perception,
an agent has to partition objects in the world into different discrete categories based
on their visual appearance. In active categorical perception, the agent takes an active
role in which aspects of the object to view, that is, perception is intimately linked with
action. Whether or not this task requires internal representations may depend on the
specific nature of the task, and it is in general not simple to determine whether the agent
uses internal states to represent the environment, and in particular which features of
the world are represented by which computational units (Beer, 2003; Ward and Ward,
2009). In the task studied by Beer (1996; 2003), the agent has to discriminate between
circles and diamonds that were falling vertically towards the agent, who could move
laterally to change its perception of the object. In the version we study here, the agent
8
A
B
1
2
0
3
15
14
0
1
2
3
Sensor States
= on
= off
Figure 2: A: In the simulation, large or small blocks fall diagonally towards the bottom
row of a 20×20 world, with the agent on the bottom row. For the purpose of illustrating
the task, a large brick (to be avoided) is falling to the left, while a small brick (to be
caught) is falling to the right. In simulations, only one block is falling at the time, and
both small and large bricks can fall either to the left or to the right. B: A depiction of the
agent’s neurons (bottom left: triangles depict sensors, circles illustrate brain neurons,
trapezoids denote actuators) and the sequence of activity patterns on the agent’s 4-bit
retina (right), as a large brick falls to the right.
9
has to discriminate between large and small blocks, by catching the small blocks while
avoiding the large ones. In order to create visual ambiguity, the agent’s four upwardlooking sensors have a blind spot (two non-sensing units) between them. Furthermore,
blocks are falling diagonally (from left to right or right to left), so that the agent not
only has to categorize, but also to predict (see Fig. 2A).
We evolve active categorical perception in two experiments, using two different
control architectures. One is an artificial neural network (ANN), and one is a Markov
network, specifically, a network of Hidden Markov Gates (HMGs) (Edlund et al., 2011),
described in more detail in section 2.3. The agent is located at the bottom row of a
20 × 20 world with periodic boundary conditions, and is able to move horizontally one
unit per time step (see Fig. 2A). Note that compared to van Dartel et al. (2005) we
doubled the vertical size of the arena in order to give the agent more time to assess
the direction of the falling block. The agent has four upward-looking sensors. The
state of a sensor is si = +1 when a block is detected, and si = 0 if no block is seen.
The arrangement of the agent’s sensors does not allow it to unambiguously identify the
falling block from three out of eight possible input states (counting the input where no
sensors are active). Only the input states S = (s0 , s1 , s2 , s3 ) = (0, 1, 1, 0), (0, 1, 0, 0),
and (0, 0, 1, 0) are unambiguous. The first classifies a large block that is positioned right
above the agent so that the second and the third sensor are in state s1 = s2 = 1 while the
remaining two units of the block are positioned above the two non-sensing units. The
other two patterns identify similarly positioned small blocks. All other five input states
can be created by either a small block, a big block, or no block at all (see Fig. 2B). At
each trial, a block of either small (two units) or large size (four units) falls from top to
bottom in 20 time steps. The blocks move continuously downwards and sideways one
unit per time step. Blocks either always move to the right or to the left. An object is
caught if the position of the block’s units and of the agent’s units at time step 20 overlap
in at least one unit.
For the information-theoretic characterization of correlations, we have to assign
probabilities to the possible states of the world. Theoretically, a falling block can be in
any of 20 different starting positions, large or small, and falling left or right, giving rise
to 80 possible experimental initial conditions. While the agent can be in any of 20 initial
positions, the periodic boundary conditions ensure that each of them is equivalent, given
10
the 20 initial positions of the falling block. Because there are 20 time steps before
the block reaches the bottom row, there are in total 1,600 possible different states the
world can be in. We do not expect that all of these states will be discriminated by
the agent, so instead we introduce a coarse-graining of the world by introducing four
bits that we believe capture salient aspects of the world. We define the environmental
(joint) variable E = E0 E1 E2 E3 to take on states as defined in Table 1. Of course, this
World state
World character
E0 = 0
no sensor activated
E0 = 1
at least one sensor activated
E1 = 0
block is to the left of agent
E1 = 1
block is to the right of agent
E2 = 0
block is two units (small)
E2 = 1
block is four units (large)
E3 = 0
block is moving left
E3 = 1
block is moving right
Table 1: Coarse-graining of world states into the four bits E0 , E1 , E2 , E3 . Note that
E1 could be ambiguous in case the block is centered over the agent or exactly 10 units
away. We resolve this ambiguity by setting E1 = 0 when the block is centered over the
agent, and E1 = 1 when it is exactly 10 units away.
encoding reveals a bias in what we, the experimenters, believe are salient states of the
world, and certainly underestimates the amount of “discoverable” entropy. However, in
hindsight this coarse-graining appears to be sufficient to capture the essential variations
in the world, and furthermore lends itself to study which aspects of the world are being
represented within the agent’s network controller, by defining representations about
different aspects i of the world as the representation Ri = H(Ei : M |S). Thus, we will
study the four representations
Rhit = H(E0 : M |S)
(10)
RLR = H(E1 : M |S)
(11)
R4/2 = H(E2 : M |S)
(12)
R+/− = H(E3 : M |S)
(13)
11
that represent whether the sensor has been activated [Eq. (10)], whether the block is
to the left or the right of the agent [Eq. (11)], if the block is large (size 4) or small
(size 2) [Eq. (12)], or whether the block is moving to the left or right [Eq. (13)]. We
can also measure how much (measured in bits) of each binary concept is represented in
any particular variable. For example, RLR (node 12) = H(E1 : M12 |S) measures how
much of the “block is to my left or to my right” concept is encoded in variable 12.
2.3
Two Architectures for Cognitive Systems
The agent is controlled by a cognitive system, composed of computational units (loosely
referred to as “neurons” from here on) that map sensor inputs into motor outputs. The
cognitive system also has neurons that are internal (a hidden layer), which are those
neurons that are not part of the input or of the output layer. We further define sensor
neurons as those neurons that directly process the input (the input layer) and we define output neurons as those units that do not map to other units in the network or to
themselves (the output layer).
Artificial Neural Networks (ANN) with evolvable topology. In our first experiment,
the robot’s movements are controlled by an artificial neural network that consists of 16
nodes: four input units (one for each sensor), two output units, and ten hidden units. The
states of the input units are discrete with values [+1, −1] specifying whether an object
is detected or not. The states of the output units (or actuators) are discrete with integer
values A = a ∈ [0, 1] encoding one of three possible actions: move one unit to the right
or left, or do not move (A1 A2 = 00 : stand still, A1 A2 = 01 : move right, A1 A2 = 10 :
move left, A1 A2 = 11 : stand still. While the hidden units’ states mi are continuous
with values [−1.0, 1.0], when evaluating these states to calculate R we discretize them
to binary (values below 0.0 become 0, every other value becomes 1). As discussed
earlier, this discretization does not introduce a bias in the value of R.
Usually, classic artificial neuronal networks have a fixed topology, i.e., one or more
layers and their connections are defined and associated with a weight. In a previous
experiment, we found that such fixed topologies lead to approximately constant R even
as the fitness of the agent increases (data not shown). One way to increase the complexity of the network and the information it represents is to evolve a network’s topology
12
as well as the connection weights. We make the network topology evolvable beyond
searching the connection weights by using neuronal gates (NG). A NG can arbitrarily
connect nodes of any type (input, hidden, and output nodes) without the fixed layered
topology of classic ANNs. Each connection is associated with a certain weight. A NG
(k)
calculates the sum of the values from a set of incoming nodes mj via gate k, multiplied
(k)
by the associated weight wj and applies a sigmoid function to calculate its output m
m(t + 1) = tanh
n X
X
!
(k)
(k)
wj mj (t)
,
(14)
k=1 j=1
where the sum over j runs over all the neurons that feed into gate k. This value m is
then propagated to every node this NG is connected to.
To apply a Genetic Algorithm to this system, each ANN is encoded in a genome as
follows: A start codon of two loci mark the beginning of a NG, the subsequent two loci
encode the NG’s number of inputs and outputs, while two further loci specify the origin
of the inputs (which neurons feed into the gate) and the outputs (where the NG writes
into). This information is then followed by an encoding of the n weights of an n-input
NG (see Fig. 3). The number of gates in the network can change as it evolves, and is
only determined by the number of start codons in the genome. The genomes encoding
these ANNs can undergo the same mutational changes as described later in the MB
section. In this respect, evolving open-topology ANNs is similar to using evolutionary
algorithms (such as NEAT, see, e.g., Stanley and Miikkulainen (2002)) to evolve neural
networks with augmented topology.
Markov Brains (MB). In our second experiment, the agent is controlled by a network
of 16 nodes (four input, two output, and ten internal nodes, i.e., with the same types and
number of nodes as the ANNs) which are connected via Hidden Markov Gates (HMGs,
see Edlund et al. 2011). Networks of HMGs (Markov brains or MBs for short) are a
type of stochastic Markov network (see, e.g., Koller and Friedman 2009) and are related to the hierarchical temporal memory model of neocortical function (Hawkins and
Blakeslee, 2004; George and Hawkins, 2005, 2009) and the HMAX algorithm (Riesenhuber and Poggio, 1999), except that Markov brains need not be organized in a strictly
hierarchical manner as their connectivity is evolved rather than designed top-down.
Each HMG can be understood as a finite-state machine that is defined by its in-
13
in
out
hidden
t
W1
W2
W3
W4
W1
W2
W3
W4
neuronal
gate 2
neuronal
gate 1
t+1
∑
Figure 3: Open-topology Artificial Neural Network, with four input neurons (orange),
ten hidden neurons (blue), and two motor neurons (green). The nodes are connected
via two neuronal gates (NG). Each NG connects four arbitrary input nodes with weight
W to four output nodes. This figure illustrates how the nodes become updated from
time point t to t + 1. When two NGs write into the same node, their outputs are added
P
(indicated by ) before the sigmoid function is applied.
14
put/output structure (Fig. 4A) and a state transition table (Fig. 4B). All nodes in Markov
brains are binary, and in principle the HMGs are stochastic, that is, the output nodes fire
(that is, are set to state ‘1’) with a probability determined by the state-to-state transition
table. Here, each HMG can receive up to four inputs, and distribute signals to up to
4 nodes, with a minimum of one input and one output node (these settings are configurable). For the evolution of the ACP task, we consider only deterministic HMGs (each
row of the transition table contains only one value of 1.0 and all other transitions have
a probability of 0.0), turning our hidden Markov gates into classical logic gates. In order to apply an evolutionary algorithm, each HMG is encoded in a similar way as the
NGs using a genome that specifies the network as a whole. Each locus of the genome
is an integer variable ∈ [0, 255]. Following a start codon (marking the beginning of a
gene, where each gene encodes a single HMG), the next two loci encode the number of
inputs and outputs of the gate respectively, followed by a specification of the origin of
the inputs, and the identity of the nodes being written to.
For example, for the HMG depicted in Fig. 4, the loci following the start codon
would specify ‘3 inputs’, ‘2 outputs’, ’read from 1,2,3’, ‘write to 3,4’. This information
is then followed by an encoding of the 2n+m probabilities of an n-input and m-output
state transition table (see Supplementary Fig. S1 in Edlund et al. 2011 for more details.)
For the example given in Fig. 4, the particular HMG is specified by a circular genome
with 39 loci (not counting the start). The start codon is universally (but arbitrarily) chosen as the consecutive loci (42,213). Because this combination only occurs by chance
once every 65,536 pairs of loci (making start codons rare), we insert four start codons at
arbitrary positions into a 5,000 loci initial genome to jump start evolution. Thus, the ancestral genomes of all experiments with Markov brains encode at least 4 HMGs. A set
of HMGs encoded in this manner uniquely specifies the Markov brain. The encoding
is robust in the sense that mutations that change the input-output structure of an HMG
leave the probability table intact, while either adding or removing parts of the table.
This flexibility also implies that there is considerable neutrality in the genome, as each
gene has 256 loci reserved for the probability table even if many fewer loci are used.
MBs and ANNs differ with respect to the gates connecting the nodes in each network. ANNs use weights, sums, and a tanh function together with continuous variables to compute their actions. In contrast MBs use discrete states and boolean logic
15
to perform their computation. Using a very similar encoding of the topology means
that mutations will have a similar effect on the topology of both systems, but different
effects on the computations each system performs.
A
B
...0110100110...
...0100110001...
1
2
3
4
....010100100.....
3
p00 p01 p02 p03
p10 p11 p12 p13
p20 p21 p22 p23
p30 p31 p32 p33
p40 p41 p42 p43
p50 p51 p52 p53
p60 p61 p62 p63
p70 p71 p72 p73
1
2
3
4
Figure 4: A: A single HMG with three inputs and two outputs reads from nodes 1, 2,
and 3, and writes to nodes 3 and 4, updating the states of these nodes in the process.
B: The output states are determined by a set of 23+2 probabilities, here denoted as pxy ,
where x and y are the decimal equivalent of the binary pattern of the input and output,
respectively. For example, p73 is the probability for the pattern ‘11’ to fire if the input
was ‘111’, that is, P (11|111) = p73 .
2.4
Evolutionary Algorithm
We evolve the two types of networks (ANNs and MBs) using a Genetic Algorithm
(GA). A GA can find solutions to problems by using evolutionary search [see, e.g.,
Michalewicz (1996)]. The GA operates on the specific genetic encoding of the networks’ structure (the genotype), by iterating through a cycle of assessing each network’s
fitness in a population of 100 candidates, selecting the successful ones for differential
replication, and finally mutating the new candidate pool. When testing a network’s
performance in controlling the agent, each network is faced with all 80 possible initial
conditions that the world can take on. The fitness w is calculated as the fraction of
successful actions (the number of large blocks avoided plus the number of small blocks
caught) out of 80 tests (a number between zero and 1). For the purpose of selection,
16
we use an exponential fitness measure that multiplies the score by a factor 1.1 for every successful action, but divides the score by 1.1 for every unsuccessful action, or
S = 1.180(2w−1) . After the fitness assessment, the genotypes are ranked according to
S and placed into the next generation with a probability that is proportional to the fitness (roulette wheel selection without elite). After replication, genotypes are mutated.
We implemented three different mutational mechanisms that all occur after replication,
with different probabiities. A point mutation happens with a probability of µ = 0.005
per locus, and causes the value at that locus to be replaced by a uniform random number drawn from the interval [0 · · · 255]. There is a 2% chance that we delete a sequence
of adjacent loci ranging from 256-512 in size, and a 5% chance that a stretch of 256512 adjacent loci is duplicated (the size of the sequence to be deleted or duplicated is
unformly distributed in the range given). The duplicated stretch is randomly inserted
between any two loci in the genome. Duplications and deletions are contrained so that
the genome is not allowed to shrink below 1024 sites, and genomes cannot grow beyond
20,000 sites. Because insertions are more likely than deletions, there is a tendency for
genomes to grow in size during evolution.
We evolve networks through 10,000 generations, and run 200 replicates of each experiment. Note that the type of gates is different between ANNs (neuronal) and MBs
(logic), so the rate of evolution of the two networks cannot be compared directly, because mutations will have vastly different effects with respect to the function of the
gates. Thus, the optimal mutation rate differs among networks (Orr, 2000). At the end
of each evolutionary run, we reconstruct the evolutionary line of descent (Lenski et al.,
2003) of the experiment, by following the lineage of the most successful agent at the
end of 10,000 generations backwards all the way to the random ancestor that was used
to seed the experiment. This is possible because we do not use cross-over between
genotypes in our GA. This line of descent, given by a temporally ordered sequence of
genotypes, recapitulates the unfolding of the evolutionary process, mutation by mutation, from ancestor to the evolved agent with high fitness, and captures the essence of
that particular evolutionary history. For each of the organisms on each of the 200 lines
of descent of any particular experiment, we calculate a number of information-theoretic
quantities, among which is how much of the world the agent represents in its brain,
using Equation (5).
17
2.5
Extracting probabilities from behavior
The inputs to the information-theoretic measure of representation R are the probabilities
to observe a particular state x, p(x), as well as the joint probabilities p(x, y) describing
the probability to observe a state x when at the same time another variable Y takes
on the state Y = y. For the representation R defined by Eq. (5), sensor, internal,
and environment variables are distinguished. For any particular organism (an agent
that performs the ACP task with an evolved controller), R is measured at any point
during the evolution by placing the organism into the simulated world and concurrently
recording time series data of the states of all 16 controller nodes and the states of the
environment. The recordings are then used to calculate the frequency of states. Based
on the frequencies of states, all probabilities relevant for the information-theoretical
quantities can be calculated, including those that take into account the temporal order of
events (for example, the probability p(xt , yt+1 ) that variable Xt takes on state xt while
variable Yt+1 takes on the state yt+1 ). If a particular state (or combination of states)
never occurs, a probability of zero is recorded for that entry (even though in principle
the state or combination of states could occur). R (and the other information-theoretic
quantities introduced in section 3) is calculated for organisms on the evolutionary line
of descent making it possible to follow the evolutionary trajectory of R from random
ancestor to adapted agent. For ANNs that have internal nodes with continuous rather
than binary states, a mapping of intervals [−1, 0] → 0 and ]0, 1] → 1 is applied before
calculation of probabilities.
3
Results
To establish a baseline, 100,000 random controllers for each of the two network types
were created and the distributions of R and fitness values were obtained. This baseline
served two purposes: it shows how well randomly generated (unevolved) networks perform, and how much information about the world they represent by chance as well as
providing information about the distribution of these values. Random ANNs and MBs
were created in the same way by randomly drawing values from a uniform probability
distribution of the integers ∈ [0, 255] for each of the genome’s loci. Each genome was
then sprinkled with 4 start codons at arbitrary positions within the genome.
18
Fig. 5A shows the distribution of fitness scores w for 100,000 random ANNs, and
Fig. 5B shows the distribution of their representation R. The respecitve distributions for
MBs are shown in Fig. 5C for fitness scores and in Fig. 5D for representation. While
both systems use different types of gates, ANNs and MBs do not differ with respect
to their initial fitness or representation distributions. This shows that random genomes
with a high fitness are–as expected–very rare, and need to evolve their functionality in
order to perform optimally.
0.7
0.6
0.5
A
B
ANN
ANN
0.4
0.4
p(r) dr
p(w) dw
0.5
0.3
0.3
0.2
0.2
0.1
0.1
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
w
0.7
0.6
0
0.2
0.4
r
0.6
0.8
1
0.5
C
D
MB
MB
0.4
0.4
p(r) dr
p(w) dw
0.5
0.3
0.3
0.2
0.2
0.1
0.1
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
w
0
0
0.2
0.4
r
0.6
0.8
1
Figure 5: Probability distribution of fitnesses and representation scores of random
machines.
A: Probability distribution of fitnesses (fraction of successful actions)
p(w)dw = P (w < W < w + dw) for 100,000 random ANNs (dw = 1/80) B:
Probability distribution of the representation variable R for the same random ANNs
p(r)dr = P (r < R < r + dr), with dr = 0.02. C: Distribution of fitnesses for 100,000
random Markov brains (dw = 1/80). D: Distribution of representation R in the same
MB networks (dr = 0.02).
In order to compare the two network architectures, the evolutionary trajectories for
fitness and representation were analyzed for the evolutionary line of descent (LOD)
19
as described in section 2.4. The different LODs obtained from following back any
other member of the final population quickly coalesce to a single line. Hence, the
LOD effectively recapitulates the genetic changes that led from random networks to
proficient ones. The development of fitness and representation over evolutionary time in
Fig. 6 is averaged over 200 independent replicates. While both ANNs and MBs have on
average low fitness at the begin of an evolutionary run (as seen in Fig. 5), MBs become
significantly more fit than ANNs, and after 10,000 generations we find that in 18 out of
200 runs MBs have evolved to perfect fitness, while none of the ANNs reached this level
(the best ANNs correctly make 77 out of 80 decisions). At the same time, we see that
the fitness of the ANNs after 10,000 generations is not stagnating, which suggests that
more runtime will allow for further improvement. As previously mentioned, the rate at
which fitness is achieved in evolutionary time cannot be compared across architectures
because mutations affect the function of the gates differently. While we tentatively
explain the difference in performance between ANNs and MBs by their difference in
representing the world below, we anticipate that the different network architectures also
solve the categorization task very differently. To understand the information dynamics
and the strategies employed in more detail, we measured a number of other informationtheoretic measures (besides R) (see sections below).
The evolutionary trajectory for representation R (see Fig. 6B) is similar to the evolution of fitness (see Fig. 6A), but MBs evolve to a significantly higher value of R. We
attribute this difference to the difference in fitness between the two types of networks,
as the discretization of the continuous ANN variables cannot introduce a bias in R.
Thus, it appears that an increased representation of the world within an agent’s network
controller correlates with fitness. We can test this correlation between fitness and representation at the end of a run for the 200 replicates of MBs and ANNs, and find that
fitness and R are significantly correlated (Spearman’s r = 0.55, P = 2.5 × 10−17 ) for
MBs, but not for ANNs (r = −0.10, P = 0.15). We speculate that because ANNs are
forced to compute using a sigmoid function only (effectively implementing a multipleAND gate) while MBs can use arbitrary logic operations to process data, ANNs struggle
to internalize (that is, represent) environmental states. In other words, it appears that
the ease of memory-formation is crucial in forming representations, which are then
efficiently transformed into fit decisions in MBs (Edlund et al., 2011).
20
1
0.8
A
0.9
B
0.7
0.6
0.5
0.7
R
w
0.8
0.4
0.3
0.6
0.2
0.5
0.4
0.1
0
2000
4000 6000
generations
8000
10000
0
0
2000
4000 6000
generations
8000
10000
Figure 6: A: Fitness w and B: representation R (in bits) along the line of descent as a
function of evolutionary generations, averaged over 200 independent evolutionary lines,
for evolved networks (ANNs: black, MBs: blue)
21
3.1
Analysis of Network Structures and Strategies
In order to be successful at the task described, an agent has to perform active categorical
perception followed by prediction. In the implementation of the ACP task by Beer
(1996; 2003), prediction can be achieved without memory, because once the network
has entered the attractor representing a category, the prediction (to move away or to
stay) can be directly coupled to the attractor. The task used here can only be achieved
using memory (data not shown) and requires the agent to perform categorical perception
by comparing sensory inputs from at least two different time points (which also allows
a prediction of where the object is going to land).
In order to analyze how information is processed, we calculated the predictive information (Bialek et al., 2001) of the evolved networks, given by the mutual Shannon
information between the network’s inputs at time t and its outputs at time t + 1. Predictive information, defined this way (Ay et al., 2008), measures how much of the entropy
of outputs (the firings of motor neurons that control the agent) can be understood in
terms of the signals that have appeared in the agent’s sensors just prior to the action.
Indirectly, predictive information therefore also indicates the contribution of the hidden
nodes of the network. A high predictive information would show that the hidden nodes
do not contribute much and that computations are performed mainly by input and output
neurons. Using the variable S for sensor states and A for actuator states, the predictive
information can be written in terms of the shared entropy between sensor states at time
t and motor states at time t + 1 as
Ipred = I(St : At+1 ) = −
X
st ,at+1
p(st , at+1 ) log
p(st , at+1 )
,
p(st )p(at+1 )
(15)
where p(st ) = P (St = st ) is the probability to observe variable St in state st , p(at+1 ) =
P (At+1 = at+1 ) is the probability observe variable At+1 in state at+1 , etc. Note that St
and At+1 are joint random variables created from the variables of each node, implying
that St can take on 16 different states while At+1 can take on 4 possible states. The
probabilities are extracted from time series data as described in section 2.5. Figure
7A shows that over the course of evolution, the predictive information Ipred decreases
for MBs after an initial increase, but increases slightly overall for ANNs. The drop in
predictive information for the MBs indicates that their actions become less dependent
22
2
A
Ipred
1.5
1
0.5
0
0
Hunpred
2
2000
4000
6000
8000
10000
2000
4000
6000
8000
10000
4000
6000
generations
8000
10000
B
1.5
1
0.5
SIatom
0
6
5
4
3
2
1
0
0
C
0
2000
Figure 7: Different measures of information processing and integration along the LOD
for both types of network architectures: ANNs (black) and Markov brains (blue). A:
Predictive information, Eq. (15). B: Unpredicted entropy [Eq. (16)] of the network’s
motor variables. C: Information integration SIatom based on Eq. (17).
23
on sensor inputs and are driven more by the hidden neurons, while the ANNs actions
remain to be predictable by sensor inputs.
To test whether it is the internal states that increasingly guide the agent, the predictive information was subtracted from the entropy of the output states (maximally two
bits) to calculate the unpredicted entropy of the outputs, i.e., how much of the motor
outputs are uncorrelated to signals from the input:
Hunpred = H(At+1 ) − Ipred = H(At+1 |St ) .
(16)
Figure 7B shows that Hunpred increases over the course of evolution, suggesting
that indeed signals other than the sensor readings are guiding the motors. In principle,
this increase could be due to an increase in the motor neuron entropy, however, as the
latter stays fairly constant we can conclude that the more a network adapts to its environment, the less its outputs are determined by its inputs and the more by its internal
states. Again, this effect is stronger for MBs than for ANNs, and suggests that it is
indeed the internal states that encode representations that drive the network’s behavior. It is also possible that the motors evolve to react to sensor signals further back in
time. Because sensor neurons cannot store information, such a delayed response also
has to be processed via internal states. While the absolute value of the predicted information and unpredicted entropy can depend on this time delay, we expect the overall
trend of a decreasing Ipred coupled with an increasing Hunpred to be the same as for the
one-step predictive information, because the sensorial signal stream itself has temporal
correlations, that is, it is non-random.
To quantify the synergy of the network we calculated a measure of information
integration called synergistic information. Roughly speaking, synergistic information
measures the amount of information that is processed by the network as a whole that
cannot be understood in terms of the information-processing of each individual node,
i.e., it measures the extent to which the whole network is–informationally–more than
the sum of its parts (Edlund et al., 2011):
SIatom = I(Xt : Xt+1 ) −
n
X
i
I(Xti : Xt+1
).
(17)
i=1
In Eq. (17), I(Xt : Xt+1 ) measures the amount of information that is processed (across
24
time) by the whole network X (the joint random variable composed of each of the node
i
variables), whereas I(Xti : Xt+1
) measures how much is processed by node i. The
negative of Eq. (17) has been used before, to quantify the redundancy of information
processing in a neural network (Atick, 1992; Nadal and Parga, 1994; Schneidman et al.,
2003). SIatom is a special case of the information integration measure Φ (Tononi, 2008;
Balduzzi and Tononi, 2008), which is computationally far more complex than SIatom
because it relies on computing information integration across all possible partitions of
a network. SIatom , instead, calculates information integration across the “atomic” partition only, that is, the partition where each node is its own part. Figure 7C shows that
SIatom increases for Markov brains as well as for ANNs, which indicates that both architectures evolve the ability to integrate information to perform the task at hand. We
only see a marginal difference between MBs and ANNs in their ability to integrate
information, while at the same time MBs are more dependent on internal states and ultimately perform better. This suggests that measuring integrated information in terms of
Eq. 17 does not allow inferences about a system’s capacity do memorize. In summary,
we observe that MBs evolve to become less dependent on sensorial inputs than ANNs,
and in addition, the actions of the MBs become more dependent on internal states than
in ANNs. Thus, we conclude that the network properties that R measures are indeed
representations that the networks create as an adaptive strategy. But how do the networks represent their environments? Which features of the world are represented and
form the successful task-solving strategies?
3.1
Epistemically Opaque Strategies
In order to analyze a Markov brain’s function, a number of different tools are available.
First, a causal diagram can be generated by drawing an edge between any two neurons
that are connected via an HMG. The edges are directed, but note that each edge can
in principal perform a different computation. When creating the causal diagram, nodes
that are never written into by any other nodes are removed, as they are computationally
inert (they remain in their default ‘off’ state). Such nodes can also be identified via a
“knock-out” procedure, where the input of each node is forced to either the “0” or “1”
state individually. If such a procedure has no effect on fitness, the node is inert.
Fig. 8 shows the causal diagram of an evolved Markov brain that solves the clas25
0
2
3
1
5
10
12
6
7
14
4
15
Figure 8: Causal diagram of a Markov brain with perfect fitness, correctly catching all
small blocks and avoiding all large ones. Nodes colored in red are sensors, while motor
variables are green. Double arrows represent two causal connections, one each way.
Nodes with arrows that point to themselves write their output back into their input,
and may work as memory (all nodes return to a default “0” if not set otherwise by
each update), so state information can only be maintained via such self-connections).
Internal nodes can read from motors, giving rise to proprioception, or more precisely,
kinesthesia: the ability to sense one’s own motion. The motors themselves can be used
as memory.
26
sification task perfectly. One can see that this network uses inputs from the sensors,
motors, and memory simultaneously for decisions, fusing the different modalities intelligently (Murphy, 1996).
The causal diagram by itself, however, does not reveal how function is achieved in
this network. As each HMG in the present instantiation represents a deterministic logic
gate (generally they are stochastic), it is possible to determine the logical rules by which
the network transitions from state to state by feeding the state-to-state transition table
into a logical analyzer (Rickmann, 2011). The analyzer converts the state transition
table into the minimal description of functions in Boolean logic using only NOT, AND
(∧), and OR (∨). With these functions we can exactly describe each node’s logical
influence on other nodes (and possibly itself). For the network depicted in Fig. 8, the
logic is given by (here, the numeral represents the node, and its index the state at time
t0 or the subsequent time point t1, while an overbar stands for NOT)
4t1 = (2t0 ∧ 5t0 ) ∨ (2t0 ∧ 6t0 )
5t1 = (2t0 ∧ 10t0 ∧ 12t0 ) ∨ (10t0 ∧ 12t0 ∧ 14t0 ) ∨ (2t0 ∧ 10t0 ∧ 12t0 ∧ 14t0 )
6t1 = 4t0 ∧ (0t0 ∨ 6t0 )
7t1 = (2t0 ∧ 12t0 ) ∨ (10t0 ∧ 12t0 ∧ 14t0 )
10t1 = (2t0 ∧ 6t0 ) ∨ (2t0 ∧ 10t0 ∧ 14t0 ) ∨ (10t0 ∧ 12t0 ∧ 14t0 ) ∨ (2t0 ∧ 14t0 )
∨(3t0 ∧ 7t0 ∧ 10t0 ∧ 12t0 ∧ 14t0 )
12t1 = (1t0 ∧ 7t0 ) ∨ (3t0 ∧ 7t0 ) ∨ (3t0 ∧ 7t0 )
14t1 = (2t0 ∧ 10t0 ∧ 14t0 ) ∨ (2t0 ∧ 10t0 ∧ 12t0 ) ∨ (2t0 ∧ 10t0 ∧ 12t0 )
15t1 = (3t0 ∧ 7t0 ) ∨ (0t0 ∧ 6t0 ) ∨ (4t0 ∧ 6t0 ) ∨ (0t0 ∧ 4t0 ∧ 6t0 ) .
Note that while this logical representation of the network’s dynamics is optimized (and
the contribution of inert nodes is removed), it is in general not possible to determine the
minimal logic network based on state-to-state transition information only, as finding the
minimal logic is believed to be a computationally intractable problem (Kabanets and
Cai, 2000). As a consequence, while it is possible to capture the network’s function in
terms of a set of logical rules, we should not be surprised that evolution delivers epistemically opaque designs (Humphreys, 2009), that is, designs that we do not understand
on a fundamental level. However, the strength of measuring representations with R is
27
that our measure is able to capture representations even if they are highly distributed
and take part in complex computations. As such, R provides a valuable tool to analyze
evolved neural networks as can be seen below.
0.4
node 4
node 5
node 6
node 7
node 10
node 12
Rhit
0.3
0.2
0.1
0
0.4
RLR
0.3
0.2
0.1
0
0.4
R+/-
0.3
0.2
0.1
0
0.4
R4/2
0.3
0.2
0.1
0
0 2 4 6 83
gener. [10 ]
0 2 4 6 83
gener. [10 ]
0 2 4 6 83
gener. [10 ]
0 2 4 6 83
gener. [10 ]
0 2 4 6 83
gener. [10 ]
0 2 4 6 83
gener. [10 ]
Figure 9: Representation of each of the 4 environmental properties (concepts) defined
in Eqs. (10-13) as a function of time, within each of the nodes of a network that evolved
to become the one depicted in Fig. 8. Representation is measured in bits, along the
temporal (genetic) line of descent (measured in generations).
3.2
Concepts and Memory
To understand what representations are acquired (representations about which concepts), we calculated R for each property of the environment defined in Eq. (10-13),
within each of the key nodes in our example network shown in Fig. 8. For this network,
Fig. 9 shows that some nodes prefer to represent given single features or concepts, while
others represent several features at the same time. In addition, the degree to which a
node represents a certain property changes during the course of evolution. Looking at
28
representation within each individual node, however, only tells part of the story as it
is clear that representations are generally “smeared” over several nodes. If this is the
case, a pair of nodes (for example) can represent more about a feature than the sum of
the representations in each node, i.e., variables can represent synergistically. In order to
discover which combination of nodes represents which feature most accurately, a search
over all partitions of the network would have to be performed, much like in the search
for the partition with minimum information processing in the calculation of a network’s
synergistic information processing (Balduzzi and Tononi, 2008).
One can also ask whether brain states represent the environment as it is at the time
it is being represented, or whether it represents the environment in its past state, or
in other words, we can ask whether representations are about more distant or more
proximal events. To answer this, we define temporal representations by including the
temporal index of the Markov variables. For example, a representation at the same time
point t is defined [as implicit in Eq. (5)] as
Rt = H(Et : Mt |St ) ,
(18)
while a representation of events one update prior is defined as
Rt−1 = H(Et−1 : Mt |St ) ,
(19)
i.e., the shared entropy between the internal variables at time t and the environmental
states at time t−1, given the sensor’s states at time t. Naturally, one can define temporal
representations about more distant events in the same manner. Rt , Rt−1 , and Rt−2 were
calculated and averaged over all 80 experiments in each generation (for both ANNs and
MBs) over the course of evolution. Figure 10 shows that in both systems Rt−1 is larger
than Rt , and Rt−2 is larger still (but note that Rt−3 is smaller, data not shown), and all
values increase over evolutionary time similar to R (see Fig. 6B). This suggest that both
networks evolve a form of memory that reaches further back than just one update.
We suggest that the peak in representation at a time difference of two updates implied by Fig. 10 can be explained by the hierarchical structure of the networks, which
have to process the sensorial information through at least two time steps to reach a decision (it takes at least two time steps in order to assess the direction of motion of the
block). Decisions have to be made shortly thereafter, however, in order to move the
29
1
1
A
B
0.8
0.6
0.6
R
R
0.8
0.4
0.4
0.2
0.2
0
0
0
2000
4000 6000
generations
8000
10000
0
2000
4000 6000
generations
8000
10000
Figure 10: Representation as a function of evolutionary time, for three different time
intervals. A: Representation in ANNs (red: Rt , green: Rt−1 , blue: Rt−2 . B: Representation in Markov brains (colors as in A). The red curves are the same as in Fig. 6B, and
are shown here for the purpose of comparison.
agent to the correct location in time. This further strengthens our view that representations are evolved, and furthermore that they build up during an agent’s lifetime as
memory of past events shape the agent’s decisions.
4
Conclusions
We defined a quantitative measure of representation R in terms of information theory,
as the shared entropy between the states of the environment and internal “brain” states,
given the states of the sensors. While internal states are necessary in order to encode
internal models, not all internal states are representations. Indeed, representational information (which is about the environment) is a subset of the information stored in
internal states. Information about the state of other past internal states or sub-states do
not count towards R, and neither would information about imagined worlds, for example. Testing which nodes of the system contain representations about what aspect of
the environment helps us to distinguish between information present in internal states
and information that is specifically used as a representation. We applied this measure
to two types of networks that were evolved to control a simulated agent in an active
categorical peception task. Our experiments showed that the achieved R increases with
fitness during evolution independently of the system used. We also showed that while
30
the (algorithmic) function of both artificial neural networks and Markov networks is difficult to understand, deterministic Markov networks can be reduced to sets of Boolean
logic functions. This logic, however, may be epistemically opaque. While representation R increases in networks over evolutionary time, each neuron can represent parts
of individual concepts (features of the environment). However, most often concepts are
distributed over several neurons and represent synergistically. In addition, representations also form over the lifetime of the agent, increasing as the agent integrates information about the different concepts to reach a decision. Thus, what evolves in Markov
and artificial neural networks via Darwinian processes are not the representations themselves, but rather what evolves is the capacity to represent the environment, while the
representations themselves are formed as the agent observes and interacts with the environment. We argued that R can be measured in continuous (artificial neural networks)
as well as discrete systems (Markov networks), which suggests that this measure can be
used in more complex and more natural systems. We found that in the implementation
used here, Markov networks were able to evolve their ability to form representations
more easily than artificial neural networks. Future investigations will show what kind
of system is more powerful to make intelligent decisions using representations.
Acknowledgements We thank C. Koch and G. Tononi for extensive discussions about
representations, information integration, and qualia. This research was supported in part
by the German Federal Ministry of Education and Research, by the Paul G. Allen Family Foundation, by the National Science Foundation’s Frontiers in Integrative Biological
Research grant FIBR-0527023, NSF’s BEACON Center for the Study of Evolution in
Action under contract No. DBI-0939454, as well as by the Agriculture and Food Research Initiative Competitive Grant no. 2010-65205-20361 from the USDA National
Institute of Food and Agriculture. We wish to acknowledge the support of the BerlinBrandenburg Academy of Sciences, the Michigan State University High Performance
Computing Center and the Institute for Cyber Enabled Research.
References
Atick, J. J. (1992). Could information theory provide an ecological theory of sensory
processing. Network, 3(2):213–251.
31
Ay, N., Bertschinger, N., Der, R., Guettler, F., and Olbrich, E. (2008). Predictive information and explorative behavior of autonomous robots. Eur Phys J B, 63(3):329–339.
Balduzzi, D. and Tononi, G. (2008). Integrated Information in Discrete Dynamical Systems: Motivation and Theoretical Framework. PLoS Comput Biol, 4(6):e1000091.
Beer, R. (1996). Toward the evolution of dynamical neural networks for minimally
cognitive behavior. In Maes, P., Mataric, M., Meyer, J., Pollack, J., and Wilson,
S., editors, From Animals to Animats: Proceedings of the Fourth International Conference on Simulation of Adaptive Behavior, pages 421–429, Cambridge, MA. MIT
Press.
Beer, R. (2003). The dynamics of active categorical perception in an evolved model
agent. Adaptive Behavior, 11:209–243.
Bialek, W., Nemenman, I., and Tishby, N. (2001). Predictability, complexity and learning. Neural Computation, 13(11):2409–63.
Bongard, J., Zykov, V., and Lipson, H. (2006). Resilient machines through continuous
self-modeling. Science, 314(5802):1118–1121.
Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence,
47:139–159.
Chomsky, N. (1965). Aspects of the Theory of Syntax. Cambridge, Mass.: The MIT
Press.
Clark, A. (1997). The dynamical challenge. Cognitive Science, 21:461–481.
Clark, A. and Toribio, J. (1994). Doing without representing? Synthese, 101:401–431.
Cohen, H. and Lefebvre, C., editors (2005). Handbook of Categorization in Cognitive
Science, Amsterdam, The Netherlands. Elsevier.
Cover, T. M. and Thomas, J. A. (1991). Elements of Information Theory. John Wiley,
New York, NY.
Craik, K. (1943). The Nature of Explanation. Cambridge University Press, Cambridge
(UK).
32
Edlund, J. A., Chaumont, N., Hintze, A., Koch, C., Tononi, G., and Adami, C.
(2011). Integrated information increases with fitness in the simulated evolution of
autonomous agents. PLoS Comput Biol, 7:e1002236.
Floreano, D. and Mondada, F. (1996). Evolution of homing navigation in a real mobile
robot. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on,
26(3):396–407.
Fodor, J. (1975). The Language of Thought. New York: Crowell.
George, D. and Hawkins, J. (2005). A hierarchical Bayesian model of invariant pattern
recognition in the visual cortex. In Prokhorov, D., editor, Proceedings of the International Joint Conference on Neural Networks (IJCNN), volume 3, pages 1812–1817.
IEEE.
George, D. and Hawkins, J. (2009). Towards a mathematical theory of cortical microcircuits. PLoS Comput Biol, 5(10):e1000532.
Harnad, S., editor (1987). Categorical Perception: The Groundwork of Cognition,
Cambridge, UK. Cambridge University Press.
Haugeland, J. (1991). Representational genera. In Ramsey, W., Stich, S. P., and Rumelhart, D. E., editors, Philosophy and connectionist theory, pages 61–89, Hillsdale, NJ.
Lawrence Erlbaum.
Hawkins, J. and Blakeslee, S. (2004). On Intelligence. Henry Holt and Co., New York,
NY.
Humphreys, P. (2009). The philosophical novelty of computer simulation methods.
Synthese, 169(3):615–626.
Johnson-Laird, P. and Wason, P. (1977). Thinking: Readings in Cognitive Science.
Cambridge University Press.
Kabanets, V. and Cai, J.-Y. (2000). Circuit minimization problem. In Yao, F. and Luks,
E., editors, Proc. 32nd Symposium on Theory of Computing, pages 73–79.
33
Kawato, M. (1999). Internal models for motor control and trajectory planning. Current
Opinion In Neurobiology, 9(6):718–727.
Kay, J. W. and Phillips, W. A. (2011). Coherent infomax as a computational goal for
neural systems. Bull Math Biol, 73(2):344–72.
Koller, D. and Friedman, N. (2009). Probabilistic Graphical Models. MIT Press,
Cambridge, MA.
Lenski, R. E., Ofria, C., Pennock, R. T., and Adami, C. (2003). The evolutionary origin
of complex features. Nature, 423(6936):139–144.
Marr, D. (1982). Vision: A Computational Investigation into the Human Representation
and Processing of Visual Information. Henry Holt and Co., Inc., New York, NY,
USA.
Marstaller, L., Hintze, A., and Adami, C. (2010). Measuring representation. In Christensen, W., Schier, E., and Sutton, J., editors, ASCS09: Proceedings of the 9th Conference of the Australasian Society for Cognitive Science, pages 232–237. Sydney:
Macquarie Centre for Cognitive Science.
McGill, W. J. (1954). Multivariate information transmission. Psychometrika, 19(2):97–
116.
Michalewicz, Z. (1996). Genetic Algorithms + Data Strucures = Evolution Programs.
Springer Verlag, New York.
Moravec, H. P. (1984). Locomotion, vision and intelligence. In Brady, M. and Paul, R.,
editors, Robotics Research 1, pages 215–224.
Murphy, R. (1996). Biological and cognitive foundations of intelligent sensor fusion.
Ieee Transactions On Systems Man and Cybernetics Part A-Systems and Humans,
26(1):42–51.
Nadal, J. P. and Parga, N. (1994). Nonlinear neurons in the low noise limit: A factorial
code maximizes information transfer. Network, 5(4):565–581.
34
Newell, A. and Simon, H. (1972). Human Problem Solving. Englewood Cliffs, NJ:
Prentice Hall.
Nolfi, S. (2002). Power and the limits of reactive agents. Neurocomputing, 42(1-4):119–
145.
Nolfi, S. and Floreano, D. (2000). Evolutionary Robotics. MIT Press, Cambridge, MA.
Orr, H. (2000). The rate of adaptation in asexuals. Genetics, 155(2):961–968.
Phillipps, W. A., Kay, J., and Smyth, D. M. (1994). How local cortical processors
that maximize coherent variation could lay foundations for representation proper. In
Smith, L. S. and Hancock, P. J. B., editors, Neural Computation and Psychology,
pages 117–136, New York. Springer Verlag.
Phillips, W. and Singer, W. (1997). In search of common foundations for cortical computation. Behavioral and Brain Sciences, 20(4):657–+.
Pinker, S. (1989). Learnability and Cognition. Cambridge, Mass.: The MIT Press.
Pitt, D. (2008). Mental representation. In Zalta, E. N., editor, The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, fall 2008 edition.
Rickmann, S. (2011). Logic friday (version 1.1.3) [computer software]. Retrieved from:
http://www.sontrak.com/downloads.html.
Riesenhuber, M. and Poggio, T. (1999). Hierarchical models of object recognition in
cortex. Nature Neuroscience, 2(11):1019–1025.
Schneidman, E., Still, S., Berry II, M. J., and Bialek, W. (2003). Network information
and connected correlations. Phys Rev Lett, 91(23):238701.
Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical Journal, 27:379– 423, 623–656.
Stanley, K. O. and Miikkulainen, R. (2002). Evolving neural networks through augmenting topologies. Evol Comput, 10(2):99–127.
35
Tononi, G. (2008). Consciousness as Integrated Information: a Provisional Manifesto.
Biol Bull, 215(3):216–242.
van Dartel, M., Sprinkhuizen-Kuyper, I., Postma, E., and van den Herik, J. (2005).
Reactive agents and perceptual ambiguity. Adaptive Behavior, 13:227–42.
van Dartel M.F. (2005). Situated Representation. PhD thesis, Maastricht University.
Ward, R. and Ward, R. (2009). Representation in dynamical agents. Neural Networks,
22:258–266.
Watanabe, S. (1960). Information theoretical analysis of multivariate correlation. IBM
Journal of Research and Development, 4:66–82.
Wolpert, D. M., Ghahramani, Z., and Jordan, M. I. (1995). An internal model for
sensorimotor integration. Science, 269(5232):1880–1882.
36
| 9cs.NE
|
FUSING VIDEO AND INERTIAL SENSOR DATA FOR WALKING PERSON
IDENTIFICATION
Yuehong Huang, Yu-Chee Tseng
arXiv:1802.07021v1 [cs.CV] 20 Feb 2018
Department of Computer Science, National Chiao Tung University, Taiwan
Emails:{huang516, yctseng}@cs.nctu.edu.tw
ABSTRACT
An autonomous computer system (such as a robot) typically
needs to identify, locate, and track persons appearing in its
sight. However, most solutions have their limitations regarding efficiency, practicability, or environmental constraints. In
this paper, we propose an effective and practical system which
combines video and inertial sensors for person identification
(PID). Persons who do different activities are easy to identify.
To show the robustness and potential of our system, we propose a walking person identification (WPID) method to identify persons walking at the same time. By comparing features
derived from both video and inertial sensor data, we can associate sensors in smartphones with human objects in videos.
Results show that the correctly identified rate of our WPID
method can up to 76% in 2 seconds.
Index Terms— artificial intelligence, computer vision,
gait analysis, inertial sensor, walking person identification.
1. INTRODUCTION
Human navigates the world through five senses, including
taste, touch, smell, hearing, and sight. We sometimes rely
on one sense while sometimes on multiple senses. For computer systems, the optical sensor is perhaps the most essential
sensor which captures information like human eyes. Cameras are widely used for public safety and services in hospitals, shopping malls, streets, etc. On the other hand, booming
use of other sensors is seen in many IoT applications due to
the advances in wireless communications and MEMS. In this
work, we like to raise one fundamental question: how can
we improve the perceptivity of computer systems by integrating multiple sensors? More specifically, we are interested in
fusing video and inertial sensor data to achieve person identification (PID), as is shown in Fig. 1(b).
Efficient PID is the first step toward surveillance, home
security, person tracking, no checkout supermarkets, and
human-robot conversation. Traditional PID technologies are
usually based on capturing biological features like face, voice,
tooth, fingerprint, DNA, and iris [1–3]. However, these techniques require intimate information of users, cumbersome
registration, training process, and user cooperation. Also,
(a)
(b)
Fig. 1: Scenes where biological features are difficult to extract.
relying on optical sensors implies high environmental dependency (such as lighting, obstacle, resolution, view angle,
etc.), thus not suitable for public sites. A scene captured in a
construction site is shown in Fig. 1(a), where workers must
wear helmets and masks to protect themselves from falling
objects and toxic gases. A top view of a courtyard is shown
in Fig. 1(b). Clearly, recognizing biological features is difficult in such scenarios. Some other recognition approaches
are based on wireless signals, but require active participation
by users [4, 5]. The ID-Match method proposed in [6] integrates computer vision via depth camera and UHF RFID. It
is capable of recognizing individuals walking in groups while
wearing RFID tags, thus enabling human-robot interaction.
However, this method is handicapped by short range, and all
the users need to carry extra RFID tags.
In this work, we propose a practical, effective and convenient PID system by combining computer vision and inertial
sensor data. Because almost everyone carries a smartphone
and almost every smartphone has inertial sensors inside. The
main workflow of our PID system is shown in Fig. 2. From
video data, a set O = {o1 , o2 , ...} of human objects and their
comparable features are retrieved. Similarly, from inertial
sensors, a set S = {s1 , s2 , ...} of inertial data and their comparable features are retrieved. Then, the similarity score of
each oi and each sj is calculated. By analyzing all the similarity scores, the pairing between O and S is derived, which
leads to PID result. Inertial sensors are widely used to de-
o1
s1
s3
s2
s4
data collector
o2
s
s12
s3
s4
o3
Video Camera
Inertial Sensor
Video Feature Extraction
Acc Feature Extraction
Human Object Retrieval
Remove Direction
Trace-Finding
Butter Worth Filter
Step Feature Extraction
Down Sampling
o4
o1 o2 o3 o4
matching result
Fig. 2: Data fusion workflow.
rive carrier’s motions, paths, and physical activities. They are
standard modules for current smartphones. On the other hand,
we can get motions, traces, and physical activities of people
from videos. When persons do different types of activities, it
is easy to pair an object with a sensor. But when people do
the same activity at the same time, it is difficult to identify
persons. So this work only discusses the situation that all the
people under camera are walking.
The contributions of this work are as follows. First, we
develop a practical, low-cost, and robust PID system. Second,
our solution integrates two types of popular sensors. Third, in
this work, our matching method focuses on WPID to show
the robust of our PID system that combines video and inertial
sensor data together.
The rest of this paper is structured as follows. Section 2
introduces our PID system and WPID method. Performance
evaluation results are presented in Section 3. Conclusions are
drawn in Section 4.
2. PROPOSED WALKING PERSON
IDENTIFICATION
We consider an environment in Fig. 2 with a video camera
and multiple users. The data collected from both camera and
smartphones is sent to a server for PID purpose. Our PID system has four software modules as shown in Fig. 3. The video
feature extraction module retrieves human objects and walking traces from a sequence of video frames. The acceleration
(Acc) feature extraction module retrieves walking information from acceleration data. The similarity scoring module
compares the walking features from both data sources and assigns them similarity scores. The object-ID pairing module
couples human objects with smartphones based on the similarity scores.
2.1. Video Feature Extraction Module
The human object retrieval sub-module processes each frame
to extract the objects that are recognized as human. It is directly realized by YOLO [7–9]. For each frame, YOLO outputs a set O of human objects represented by bounding boxes,
and each bounding box is a rectangle inside where YOLO
Features
Features
Similarity Scoring
Matching Matrix
Object-ID pairing
Person IDs
Video objects
(s4 , o1 ), (s2 , o2 )
(s1 , o3 ), (s3 , o4 )
Human object-ID pairs
Fig. 3: Our PID architecture.
Frame Number
2688
2690
2697
2699
2706
2709
Person 1
Person 2
Fig. 4: The changes of bounding boxes during walking.
recognizes a human object. The ith bounding box of O is
denoted by oi and its center, width, and height are denoted
by oi .c, oi .w, and oi .h, respectively. Examples are shown in
Fig. 4.
The trace-finding sub-module is to connect the human objects of adjacent video frames and form continuous traces,
where a trace is a sequence of human objects that are regarded as the same person. Efficient object tracking algorithms are available in [10–14], but we design a lightweight
tracing method based on movement limitation. Generally, human’s running speed is less than 15 km/h. Assuming a frame
rate of 30 frames per second (fps), in most cases, a person cannot move over 0.1 of his height between two frames. Based
on this assumption, each trace has a search range to find its
human object in the next frame. The results are some traces
connecting human objects in continuous frames.
After trace-finding, the step feature extraction sub-module
retrieves walking-related features from each trace. Fig. 4
shows two sequences of frames of two human objects. Suppose our camera has a downward viewing angle. Person
1 walks along a vertical line. When he steps forward, his
bounding box becomes taller. When he closes feet, his bounding box becomes shorter. Person 2 walks along a horizontal
line. When he steps forward, his bounding box becomes
as 0. This process transforms ti and aj into ternary sequences:
tei and aej . The similarity score between tei and aej is defined
as:
n
Sim(tei , aej ) = P
.
(1)
e
x dif (ti [x], aej )
4.5
Stride 1
Stride 2
Person 1
Person 2
Stride 3
Step Patterns
4
Stride 4
Stride 5
3.5
3
2.5
Stride 1
2
0
Stride 2
20
Stride 3
Stride 4
40
60
Stride 5
Stride 6
80
100
Frame Number
Fig. 5: Ratio-features of walking traces.
wider. When he closes feet, his bounding box narrows down.
As a result, the changes of oi .h/oi .w over time are regarded
as step patterns. We use ti to denote the ratio-feature of ith
trace. Fig. 5 shows the ratio-features extracted from two persons, who make 6 and 5 strides in 100 frames, respectively.
We also mark the ground truth of strides in the graph. As
can be seen, the ratio-feature can well present human step
patterns.
2.2. Acc Feature Extraction Module
In this work, each user carries a smartphone which has installed our application (app), and they can put them in pockets or just hand them. Our software only collects acceleration
from the inertial sensor. Since activity recognition from inertial sensor data has been intensively studied, we simply adopt
existing solutions [15, 16]. The sensor data aˆi from the ith
device is a sequence of acceleration magnitudes after removing direction. Since most energy captured by accelerometer
associated with human movements is below 15 Hz [17], we
remove the high-frequency components from aˆi . aˆi is lowpass filtered by a 10th order Butterworth filter with a 15 Hz
cut-off frequency [18]. Further, since our frame rate is 30
fps, the simple frequency of aˆi is decreased to 30 per second.
After these steps, we get ai as step feature.
2.3. Similarity Scoring Module
After retrieving step features from video data and sensor data,
we want to answer the following question: How similar is
ratio-feature sequence ti to Acc sequence aj ? The similarity
between ti and aj is denoted by Sim. In this work, we try to
match the extremum positions of two sequences by ignoring
their exact values. First, we conduct an extremum detection to
find local maximum/minimum points with a window of length
d. For example, when we set d = 10, we traverse all the
points and their most adjacent 10 points. If the value of a point
is bigger/smaller than all the other 10 most adjacent points,
this point is recorded as a maximum/minimum point. In our
experiments, we set d = 10. A maximum point is marked as
1, a minimum point is marked as −1, and the rest are marked
n is the number of extremums in tei and dif (tei [x], aej ) is defined as:
tei [x] = 0;
0,
e
e
dif (ti [x], aej ) =
|y − x|, ti [x] 6= 0 and y exists;
1.5 × d,
otherwise.
(2)
Here, we scan each binary value tei [x] of tei . If tei [x] = 0, then
dif (tei [x], aej ) returns 0. If tei [x] 6= 0, dif (tei [x], aej ) traverse
aej in the range x−d to x+d. y, a position in the search range,
is the nearest position from x and has aej [y] = tei [x]. If such y
exists in the search range, dif (tei [x], aej ) returns |y −x|; if not,
1.5×d is returned. Dividing n by the sum of these differences
gives the similarity score between tei and aej .
2.4. Object-ID Pairing Module
After similarity scoring, we get Sim. Sim is a two-dimensional
array recording all the similarity scores until frame f . Let
Pf be the Object-ID pairing result until frame f . The pairing
problem is now formulated as a different expression of linear
sum assignment problem (LSAP) [19]:
XX
max
simij pij ,
(3)
i∈O j∈S
simij is the similarity score between ith human object in O
and jth sensor in S, and the assignment constraints are:
X
pij ≤ 1
∀j ∈ S,
i∈O
X
pij ≤ 1
∀i ∈ O,
j∈S
pij ∈ {0, 1}
∀i ∈ O, j ∈ S.
We use hungarian algorithm to solve this problem. pij = 1
means that human object i is paired to sensor j; pij = 0
means that human object i cannot be paired to sensor j. In
our work, each frame f can have a pairing result Pf , and we
call the pairing result at this stage as Raw Pair stage result.
However, in practice, the identification result is unstable
if we base on our Raw Pair stage result. For example, we
may identify one person as Sansa when one frame comes in,
but we may identify this person as Jack when next frame
comes in, and this person may be identified as Lucy when
the frame after the next frame comes in. This problem, which
is especially serious when the trace of a person is still short,
makes the result rough and hard to see. For this consideration,
we propose a Refined Pair stage. In Refined Pair
stage, the identification result of a trace not just depends on
Pf , but P1 to Pf . Let RPf be the result generated in the
Refined Pair stage for frame f . Let RSim be a twodimensional array, and the value of rsimij is the number of
times that object i has been paired to sensor j. The refined
pairing problem can be formulated as a LSAP:
XX
max
rpij log2 (1 + rsimij ),
(4)
(a)
i∈O j∈S
subject to:
X
rpij ≤ 1
∀j ∈ S,
rpij ≤ 1
∀i ∈ O,
i∈O
X
(b)
Fig. 6: Some correctly identified results under different viewing angles and different spaces.
Table 1: The correctly identified rates with different T S.
j∈S
rpij ∈ {0, 1}
∀i ∈ O, j ∈ S.
Different from simij , rsimij is the number of pairing times.
simij is small, but rsimij can be very large if the trace of
human object i is long. The logarithmic function, shown in
Eq. 4, is used to weaken the impact of the length of traces on
pairing. As pij , rpij = 1 means that human object i is paired
to sensor j; rpij = 0 means that human object i is not paired
to sensor j.
3. PERFORMANCE EVALUATION
We have developed a prototype system with one video camera
and multiple mobile devices. The camera is Logitech webcam
with the resolution of 640 × 480. To prove that our solution is not device-dependent, we have tried different models
of smartphones, including Redmi Note 4X, ASUS ZenFone
3, HTC 10 Evo. The server is a personal computer with an Intel(R) Core(TM) i7-3770 CPU and an NVIDIA GeForce GT
620 graphics card. All devices used in our system are synchronized by the same network time server. We conduct a
number of experiments on our WPID method. The average
speed of our tracing and WPID method on different pairing
stages is around 120 fps. Apparently, our WPID method on
two different stages only consumes a few server resources.
To show the robustness of our WPID method, experiments
are carried out under different areas and viewing angles. A
downward viewing angle and outdoor area is set up as shown
in Fig. 6(a). The horizontal viewing angle and indoor area
is set up as shown in Fig. 6(b). During our experiments, all
the persons carry smartphones in their pockets or hands and
wander around freely in their styles. As shown in Fig. 6, our
WPID method can work under different areas, different view
angles, different ways of carrying the smartphones, and different walking styles. The following statistics are all the cases
of two persons, and the result of each condition is generated
from at least 2000 continuous frames.
To measure the accuracy of our WPID method, let O be
the number of persons having shown in front of the camera
PP
PP TS/s 0.33
Stage PPP
P
Raw
0.69
Refined
0.71
1
2
3
4
0.74
0.74
0.75
0.76
0.75
0.76
0.75
0.76
until the latest frame. Let NiID be the number of frames that
the ith person is identified by our program, and NiCD be the
number of frames that the ith person is correctly identified by
our program. We define our correctly identification rate Rcd
as:
PO
NiCD
Rcd = Pi=1
.
(5)
O
ID
i=1 Ni
Let T L be the length of time that a person is continuously detected by YOLO. If T L is too small, the sequences
extracted is too short to be considered as a step pattern. As a
result, we set a threshold T S for T L. Only when the lengths
of two sequences are both larger than T S, we do our matching processes. By setting T S from 0.33 to 4 seconds, Table 1
shows the Rcd on two stages. From Table 1, the increase
of T S leads to the increase of Rcd in most cases. However,
the increase of Rcd is not obvious. Also, Refined stage
achieves better performance than Raw stage, especially in visual performance.
4. CONCLUSIONS
We propose a new PID system by combining optical and inertial sensors. We design a light tracking algorithm, a WPID
method, and two pairing stages. When people do different activities, it is easy to identify persons by comparing behaviors
extracted from video and inertial sensor data. So the most
complex part of our system is to identify persons who do the
same activities at the same time. In this work, we design a
WPID method to identify walking persons to show the robustness and potential of our PID system. We conduct extensive
experiments and do a lot of discussions to validate the above
claims. Results show that the correct identification rate of our
WPID method can up to 76% in 2 seconds.
5. REFERENCES
[1] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “DeepFace: Closing the Gap to Human-Level Performance in
Face Verification,” in IEEE Conf. on Comput. Vision
and Pattern Recognition, 2014, pp. 1701–1708.
[12] S. H. Bae and K. J. Yoon, “Robust online multi-object
tracking based on tracklet confidence and online discriminative appearance learning,” in IEEE Conf. on
Comput. Vision and Pattern Recognition, June 2014, pp.
1218–1225.
[2] Dorothy A. Lunt, “Identification and tooth morphology,” J. of the Forensic Sci. Soc., vol. 14, pp. 203–207,
Apr. 2017.
[13] Min Yang and Yunde Jia, “Temporal dynamic appearance modeling for online multi-person tracking,” Comput. Vis. Image Underst., vol. 153, no. C, pp. 16–28,
Dec. 2016.
[3] F. Alonso-Fernandez, P. Tome-Gonzalez, V. RuizAlbacete, and J. Ortega-Garcia, “Iris recognition based
on sift features,” in First IEEE Int. Conf. on Biometrics,
Identity and Security, Sept 2009, pp. 1–8.
[14] Y. Xiang, A. Alahi, and S. Savarese, “Learning to track:
Online multi-object tracking by decision making,” in
IEEE Int. Conf. on Comput. Vision (ICCV), Dec 2015,
pp. 4705–4713.
[4] Mahsan Rofouei, Andrew Wilson, A.J. Brush, and Stewart Tansley, “Your Phone or Mine?: Fusing Body, Touch
and Device Sensing for Multi-user Device-display Interaction,” in Proc. of the SIGCHI Conf. on Human Factors
in Computing Syst., New York, NY, USA, 2012, CHI
’12, pp. 1915–1918, ACM.
[15] Shane Colton, “The balance filter: a simple solution for
integrating accelerometer and gyroscope measurements
for a balancing platform,” 2007.
[5] Sherry Hsi and Holly Fait, “RFID Enhances Visitors’
Museum Experience at the Exploratorium,” Commun.
ACM, vol. 48, no. 9, pp. 60–65, Sept. 2005.
[6] Hanchuan Li, Peijin Zhang, Samer Al Moubayed,
Shwetak N. Patel, and Alanson P. Sample, “ID-Match:
A Hybrid Computer Vision and RFID System for Recognizing Individuals in Groups,” in CHI Conf. on Extended Abstracts on Human Factors in Comput. Syst.
2016, pp. 7–7, ACM.
[7] Joseph Redmon, Santosh Divvala, Ross Girshick, and
Ali Farhadi, “You only look once: Unified, real-time
object detection,” in IEEE Conf. on Comput. Vision and
Pattern Recognition (CVPR), 2016.
[8] Joseph Redmon and Ali Farhadi, “YOLO9000: Better,
Faster, Stronger,” in IEEE Conf. on Comput. Vision and
Pattern Recognition (CVPR), 2017.
[9] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik, “Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation,” in
IEEE Conf. on Comput. Vision and Pattern Recognition
(CVPR), Washington, DC, USA, 2014, pp. 580–587.
[10] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos,
and Ben Upcroft, “Simple online and realtime tracking,”
in IEEE Int. Conf. on Image Process., 2016, pp. 3464–
3468.
[11] J. H. Yoon, M. H. Yang, J. Lim, and K. J. Yoon,
“Bayesian multi-object tracking using motion context
from multiple objects,” in IEEE Winter Conf. on Applicat. of Comput. Vision, Jan 2015, pp. 33–40.
[16] Muhammad Shoaib, Stephan Bosch, Ozlem Durmaz Incel, Hans Scholten, and Paul J. M. Havinga, “Fusion of
smartphone motion sensors for physical activity recognition,” in Sensors, 2014, vol. 14, pp. 10146–10176.
[17] M. Mathie, Monitoring and Interpreting Human Movement Patterns Using a Triaxial Accelerometer, University of New South Wales, 2003.
[18] Melania Susi, Valrie Renaudin, and Grard Lachapelle,
“Motion Mode Recognition and Step Detection Algorithms for Mobile Phone Users,” Sensors, vol. 13, no. 2,
pp. 1539–1562, 2013.
[19] Rainer E. Burkard and Eranda Çela, Linear Assignment
Problems and Extensions, pp. 75–149, Springer US,
Boston, MA, 1999.
| 1cs.CV
|
1
Pricing the Ramping Reserve and Capacity Reserve
in Real Time Markets
arXiv:1512.06050v1 [math.OC] 18 Dec 2015
Hongxing Ye, Student Member, IEEE, and Zuyi Li, Senior Member, IEEE
Abstract—The increasing penetration of renewable energy in
recent years has led to more uncertainties in power systems.
In order to maintain system reliability and security, electricity
market operators need to keep certain reserves in the SecurityConstrained Economic Dispatch (SCED) problems. A new concept, deliverable generation ramping reserve, is proposed in
this paper. The prices of generation ramping reserves and
generation capacity reserves are derived in the Affine Adjustable
Robust Optimization framework. With the help of these prices,
the valuable reserves can be identified among the available
reserves. These prices provide crucial information on the values
of reserve resources, which are critical for the long-term flexibility
investment. The market equilibrium based on these prices is
analyzed. Simulations on a 3-bus system and the IEEE 118-bus
system are performed to illustrate the concept of ramping reserve
price and capacity reserve price. The impacts of the reserve credit
on market participants are discussed.
Index Terms—Ramping Reserve, Capacity Reserve, Marginal
Price, Uncertainties, Affinely Adjustable Robust Optimization
N OMENCLATURE
Indices
i, k, l, t index for unit, uncertainty, line, and time
m, n
index for bus
Functions and sets
E(·)
expectation function
Tr(·), (·)> , (·)j · trace, transpose, j th row of a matrix
Λ(·)
cost as a function of adjustment matrix
Φ(·)
Lagrangian function
(·)∗
optimal value of a variable
G(m)
set of units located at bus m
Constants
ND , NG , NK number of buses, units and uncertainty constraints
NL , NT
number of lines and time intervals
qi , bi
cost coefficients for unit i
Qi
Qi = qi I NT ∈ RNT ×NT , I NT is unity matrix
Bi
Bi = [bi · · · bi ]> ∈ RNT
fi
cost related to unit commitment decision
dm,t , d
aggregated equivalent load demand on bus m
at t, d = [d1,t · · · dND ,1 · · · dND ,NT ]>
F̄l , F
branch flow limit, abstract vector F ∈
R2NL NT
Γl,m
shift factor for line l and bus m
Pimin , Pimax minimum and maximum generation outputs
Iˆi,t , ŷi,t , ẑi,t unit on/off, start-up and shutdown indicators
The authors are with the Robert W. Galvin Center for Electricity Innovation at Illinois Institute of Technology, Chicago, IL 60616, USA. (e-mail:
[email protected]; [email protected]).
RiU , RiD
D
A, Ri
Γi , Γd
S, h
Variables
Gi
π
unit ramping up/down limits
time-load incidence matrix, D ∈ RNT ×ND NT
abstract matrix and vector for (1-3), A ∈
R4NT ×NT , Ri ∈ R4NT .
abstract shift factor matrix for unit i and load
polyhedron uncertainty set matrix and vector
generation adjustment matrix, Gi ∈ RNT ×ND NT
Prices. π ei is the energy LMP for unit i, π ei ∈
RNT ; π ri is the reserve price vector for unit i,
π ri ∈ R4NT ; π ai is the marginal price for affine
adjustment, π ai ∈ RNT ×ND NT .
Pi,t
generation output, Pi,t ∈ R
Pi
generation vector, Pi = [Pi,1 · · · Pi,NT ]> ∈ RNT
P̂ i
generation re-dispatch vector, P̂i ∈ RNT
ru
rd
rd
Qi,t , Qi,t generation ramping reserve, Qru
i,t ∈ R, Qi,t ∈ R
cu
cd
cu
cd
Qi,t , Qi,t generation capacity reserve, Qi,t ∈ R, Qi,t ∈ R
uncertainty vector, ∈ RNT ×ND
λ, αi , η Lagrangian multipliers for constraints (15, 17,
19)
γ, βi , τ Lagrangian multipliers for constraints (16,18,20)
Θi
credits to uncertainty mitigators i
T
I. I NTRODUCTION
HE renewable energy sources (RES), such as wind power
generation, and price-sensitive demand response (DR)
have attracted a lot of attentions recently. The total installed
capacity of wind power in the U.S. reached 47 GW at the
end of 2011 [1]. Several ISOs/RTOs, such as PJM, ISO New
England, NYISO, and CAISO have initiated DR programs in
their markets [2]. The essential objective to use renewable
energy and initiate DR programs in electricity markets is
to maximize the total social warfare as well as to protect
the environment. However, they also pose new challenges
to the system operators in electricity markets. Due to its
intrinsic characteristics, the amount of available renewable
energy is sometimes hard to predict. For instance, large-scale
wind production varies from -20% to 20% of the installed
wind capacity in Denmark on an hourly basis [3]. Prediction
error for wind farms aggregated output by existing state-of-art
method may fall into a range of 5%-20% of the total installed
capacity [4]. In the meantime, the amount of un-predictable
loads also increases in the wholesale market as the forecasting
of price-sensitive loads relies on the forecasted price as input,
which itself has significant error.
Surviving uncertainties is fundamentally important for the
reliable and secure operation of a power system. If the system
2
with pre-planned schedule cannot accommodate the deviation
of the wind power and load from their forecasted values, the
system operator may have to curtail wind energy or shed load
in the real-time market (RTM). In order to keep a certain
level of reliability or security, the ISO/RTO has to increase the
ramping capability of the system to compensate the variations
of wind energy and load in a short time [5], [6]. More efficient
and reliable methods are required to determine the optimal
reserves when the uncertainty level is high. Recently stochastic
and robust approaches have been successfully applied by
researchers to address the issues related to uncertainties in
electricity markets. At the same time, market designers are
also seeking effective market mechanism to address the uncertainty issues. For instance, intraday market (IDM) is now
established between day-ahead market (DAM) and RTM in
European countries since uncertainties on the intraday level
are significantly smaller compared to those on the day-ahead
level [7]. In the U.S., hour-ahead scheduling process (HASP)
is employed by the California ISO [8].
Typical approaches to solving stochastic SCUC are scenario based [9], [10]. The basic idea is to generate enough
samples for uncertain parameters with an assumption that
their probability distribution function (PDF) is known. Those
samples are then modeled in a Mixed Integer Linear/Quadratic
Programming (MILP/MIQP) problem. The two main drawbacks of the scenario-based approaches are that the PDF
is hard to obtain in some circumstances, and uncertainty
accommodation is not guaranteed. In fact, the MILP/MIQP
problem becomes intractable when the sample size is large.
Comparing with stochastic optimization, two largest merits of
robust optimization are that the solution can be immunized
against all uncertainties and PDF is not required. In [11], [12],
robust SCUC problem is solved in two stages. The first stage
is to determine the unit commitment (UC) solution which is
immunized against the worst case with the lowest cost. In the
second stage, a feasible solution to SCED is obtained. Affinely
adjustable robust optimization (AARO) models are proposed
recently [13]–[15]. They employ an affine function to adjust
the generation output following the load deviation. Recently,
we propose an economic-efficient robust SCUC model with a
fast solution approach [16].
Although applying robust techniques in SCUC/SCED receives a lot of attentions from researchers, it still remains a
big challenge on how to credit the flexibilities in the U.S. electricity markets. In the existing Ancillary Service (AS) Market,
the reserves are determined in advance [17]. The amount of
required reserves is generally extracted from larger number
of Monte Carlo simulations for the contingencies [18]. With
the help of the AARO, those reserves are determined in one
shot based on uncertainty information. Then a critical issue is
how to price those reserves when there are no explicit reserve
requirement constraints. On the other hand, some reserves are
free byproducts in the co-optimization approach [17]. They are
kept because the market participants want to get the energy
profits. Moreover, some reserves are scarce resources due to
their deliverability. These observations indicate that not all the
available reserves in the system are valuable from the system
operator’s point of view.
In many countries, electricity markets are still evolving with
the challenges of uncertain energy resource and load [19].
The ramping products are proposed by California ISO [20] to
accommodate the uncertainties. It should be emphasized that
the bus-level delivery is not considered for ramping capacity
in [20]. Before applying the robust optimization SCED in the
real market, the corresponding pricing theory is imperative.
This paper tries to propose some new ideas to clear this
obstacle. The three major contributions of this paper are listed
as follows.
1) A new concept, deliverable generation ramping reserve,
is proposed within the AARO SCED framework. The
generation ramping reserve is the additional ramping
capability of the generator when part of it is “locked”
in the SCED schedule.
2) The prices for the generation ramping reserves as well
generation capacity reserves are derived within the robust co-optimization framework. With the help of the
price information, the valuable reserves can be easily
identified among the available reserves.
3) The market equilibrium is characterized by the proposed
prices and dispatch instructions. The market participants
can get the maximal profit by following ISO’s the
dispatch instruction and price signals.
The rest of this paper is organized as follows. The derivation
of the reserve prices is presented in Section II based on AARO
SCED, and then the market mechanism to credit flexibilities is
discussed. Case studies for 3-Bus and IEEE 118-Bus systems
are presented in Section III. Section IV concludes this paper.
II. AARO SCED AND P RICES
In electricity markets, RTOs/ISOs normally operate two
markets including DAM and RTM (or balancing market) [21].
The majority of the trades is cleared in DAM via SCUC [21],
and SCED is normally performed periodically in RTM. This
paper mainly focuses on the RTM.
In standard SCED problem, the unit generation output is
subject to the following constraints which include unit capacity
limits (1) and unit ramping up/down limits (2) (3).
Iˆi,t Pimin ≤ Pi,t ≤ Iˆi,t Pimax , ∀i, t
Pi,t − Pi,(t−1) ≤ RiU (1 − ŷi,t ) + Pimin ŷi,t , ∀i, t
−Pi,t + Pi,(t−1) ≤ RiD (1 − ẑi,t ) + Pimin ẑi,t , ∀i, t
(1)
(2)
(3)
Equations (2) and (3) show that a unit has to operate at its
minimum capacity in two cases: right after it is turned on or
right before it is turned off, which implies that the unit cannot
provide reserve in those two cases. For notation brevity, we
use matrix and vector to replace the formulations above. Then
the SCED problem can be formulated as
X
min
Pi> Qi Pi + Bi> Pi + fi
(4)
Pi
s.t.
(λ)
i
X
Pi = Dd,
(5)
i
(αi ≥ 0)
(η ≥ 0)
APi ≤ Ri , ∀i
X
Γi Pi − Γd d ≤ F ,
i
(6)
(7)
3
where (4) stands the operation cost. (5) denotes the load
balance constraints. (6) is the compact form of (1)-(3). (7)
represent the transmission constraints. Γi ∈ R2NL NT ×NT , and
Γd ∈ R2NL NT ×ND NT .
Due to the forecasting errors of renewable power output
and load, ISOs/RTOs need to run SCED on a rolling basis in
real time to balance the system. Recently, some approaches
have been successfully applied in the SCUC/SCED problem
to address the uncertainty issues caused by variations of
load and renewables. Both stochastic SCUC/SCED and robust
SCUC/SCED are studied intensively when considering the
uncertainties.
To our best knowledge, this paper represents the first work
on pricing the reserves in robust optimization framework.
Hence, the following assumptions are made so that we can
focus on the concept.
•
•
•
•
•
•
Transmission loss is ignored in the SCED problem.
The proposed approach is for ex ante dispatch and ex ante
pricing. It is assumed that units are dispatched according
to these instructions during the scheduling intervals.
Units bid only energy price. The reserve bid is zero.
Uncertainty comes from loads. Renewables are treated as
negative loads. Other uncertainties such as contingencies
are not discussed in this paper.
Uncertainty set information is available to the ISO/RTO.
Expectation and covariance of uncertainties can be obtained from historical data. However, PDF information is
not specified for uncertainty as it is hard to obtain for
uncertain sources such as RES.
as (ROP), can be formulated as
i
X h
(ROP) : min
E P̂i> Qi P̂i + Bi> P̂i + fi
Pi ,Gi
i
P̂i = Pi + Gi , ∀i,
(8)
where Gi ∈ RNT ×ND NT is the affine adjustment matrix.
Pi ∈ RNT and P̂i ∈ RNT are base dispatch and adjusted
dispatch respectively, and ∈ RND NT is the uncertainty vector
(i.e., deviation of loads from forecasted values). The new unit
dispatch can be regulated based on the load deviation. It is
noted that ∈ U, and
U := { : S ≤ h},
(9)
where S ∈ RNK ×ND NT and h ∈ RNK . U is a polyhedron
which includes more than just the lower and upper bounds
for uncertainty. It is noted that entry in h is considered as
uncertainty level which is positive. The AARO SCED, denoted
(11)
i
(8),
where Qi is a semi-definite matrix with only diagonal entry.
The objective function (10) is to minimize the expected cost.
Without loss of generality, assume E() = 0. The objective
function (10) can be rewritten as
X
Pi> Qi Pi + Bi> Pi + Λ(Gi ) ,
i
>
where Λ(Gi ) = Tr[G>
i Qi Gi E( )]+fi . It is assumed that
>
covariance matrix E( ) is available, but PDF is unavailable.
The problem (ROP) is converted to a computationally tractable
problem (P) as follows, where the constraints including uncertain parameters are exactly reformulated.
X
(P) : min
Pi> Qi Pi + Bi> Pi + Λ(Gi )
(14)
Pi ,Gi ,ρi ,ζ
s.t. (λ)
i
X
Pi = Dd
(15)
Gi = D
(16)
i
(γ)
X
i
(αi ≥ 0) −Ri + APi + ρ>
i h ≤ 0, ∀i
(η ≥ 0)
The basic idea of AARO optimization is originally from
paper [15], in which a linear “feedback” in control theory is
used to adjust dispatch with the realization of load. Authors
in [13], [14], [22] applied it to solve SCED problem. In this
section, the generation output is affinely adjusted according to
the uncertainties,
P̂i = D(d + ), ∀ ∈ U
AP̂i ≤ Ri , ∀i, ∀ ∈ U
(12)
X
Γi P̂i − Γd (d + ) ≤ F , ∀ ∈ U (13)
(βi )
A. Affinely Adjustable Robust SCED
X
s.t.
(10)
i
(τ )
(17)
ρ>
i S
= 0, ∀i
(18)
X
>
−F − Γd d +
Γi Pi + ζ h ≤ 0(19)
AGi −
X
i
i
Γi Gi − Γd − ζ > S = 0
(20)
ρi ≥ 0, ζ ≥ 0,
where ρi ∈ R
and ζ > ∈ R2NL NT ×NK are also
variables. (15) and (16) are derived from (8) and (11). (17)(20) are obtained from strong duality. Problem (P) is convex
and can be solved efficiently by commercial solvers such
as CPLEX and GUROBI. It can be observed that no PDF
information is required to solve (P).
Different from the standard SCED (4)-(7), extra terms ρ>
i h
and ζ > h are added in inequality constraints (17) and (19),
>
respectively, in problem (P). As ρ>
≥ 0 and h ≥ 0,
i ≥ 0, ζ
>
>
ρi h and ζ h are non-negative. It indicates that certain unit
and transmission constraints in standard SCED are replaced
with stronger constraints in the robust framework. Thus,
the system actually keeps certain flexibilities for uncertainty
accommodation. In the following sections, we will analyze
these flexibilities, which are also called reserves.
>
4NT ×NK
B. Generation Ramping Reserve and Capacity Reserve
An Uncertainty Mitigator (UM) refers to a flexible resource
provider that participates in the management of uncertainties.
4
C. Marginal Prices
Power
(MW)
RR
Ramping
Rate Limit
Pi,t+1
LR
P i,t
t
t+1
Time Interval
Fig. 1. Illustration of Upward Ramping Reserve (LR: Locked Ramping, RR:
Ramping Reserve)
The flexible resources include generators with available ramping capabilities and adjustable loads. UMs have to keep certain
reserves in order to accommodate the uncertainties. Compared
to (6), constraint (17) may be binding even when the scheduled
dispatch does not reach the capacity limits (1) or ramping
limits (2)-(3). Based on the optimal solution to (P), the reserves
can be calculated.
The generation ramping reserve is defined as the unused
unit ramping capability or the value of the slack variable for
(2-3), i.e.
∗
∗
Pimin ŷi,t − Pi,t
− Pi,(t−1)
U
, ∀i, t,
Qru
i,t := Ri (1 − ŷi,t ) +
t − (t − 1)
(21)
∗
∗
Pimin ẑi,t − Pi,(t−1)
− Pi,t
D
, ∀i, t,
Qrd
:=
R
(1
−
ẑ
)
+
i,t
i,t
i
t − (t − 1)
(22)
rd
where Qru
and
Q
are
the
upward
and
downward
ramping
i,t
i,t
reserves, respectively. Fig. 1 illustrate the concept of the
upward ramping reserve. With scheduled generation outputs
at t and t + 1, the locked ramping or the used unit ramping
capability is (Pi,t+1 −Pi,t )/1. The generation ramping reserve
is the available ramping rate less the locked ramping. The
ramping reserve guarantees that the uncertainty mitigator still
has additional ramping capability during the scheduled ramping process. In the existing AS market, the “locked” ramping
is normally ignored for the spinning reserves [23], [24]. In
the RTM, the time resolution is 5 minutes or 15 minutes, if
we ignore the ramping process, there is a chance that system
cannot provide enough ramping capability to accommodate
the uncertainties. In contrast, the time resolution is one hour
in DAM, and the units have enough time to redispatch even
if the locked ramping is ignored [17], [23].
The generation capacity reserve is defined as unused unit
generation capacity or the constraint slack for (1), i.e.
∗
ˆ max − Pi,t
Qcu
,
i,t := Ii,t Pi
cd
∗
min
Qi,t := Pi,t − Iˆi,t Pi , ∀i, t
(23)
(24)
cu
where Qcd
i,t and Qi,t are the lower and upper generation
capacity reserves, respectively. From the system’s point of
view, the total capacity reserves are fixed when the unit
commitment and load level are determined.
In most generation and reserve co-optimization approaches
in AS market, the explicit reserve requirement constraint is
modeled [17], [18], [25]. The shadow prices of this type
of constraint is employed to derive the reserve price which
reflects the coupled effects of the generation and reserve.
Instead of setting the reserve manually and heuristically based
on Monte Carlo simulations, the reserves in the AARO SCED
are determined automatically in one shot. Although it has
obvious advantages over the traditional reserve determination,
it also poses new challenges on reserve price derivation. The
existing pricing approaches cannot be used directly [17], [18]
due to the lack of explicit reserve requirement constraints.
While the amount of reserves can be calculated according
to (21)-(24), the question is how to set the prices for them.
On the one hand, it is well known that not all the reserves
are deliverable for uncertainty accommodation due to network
constraints. On the other hand, the generation and reserve are
coupled together in the RTM even if the reserve bid price
is zero [26], and the market clearing price for reserve has
certain relations with that for energy. In this paper, we call
the reserve obtained by (21)-(24) available reserve. If a small
increment or decrement of the reserve amount causes change
of the expected operation cost in (P), then this type of reserve
is called valuable reserve. To determine the exact value a
reserve has, we derive the marginal prices for the reserve
according to the Lagrangian function as follows.
The Lagrangian function for (P) is
Φ(Pi , Gi , ρi , ζ, λ, γ, αi , βi , η, τ )
X
=
Pi> Qi Pi + Bi> Pi + Λ(Gi )
i
+ λ> Dd −
+
X
X
i
>
αi (APi
X
Pi + Tr[γ > (D −
Gi )]
+
ρ>
i h
i
+ η > − F − Γd d +
+ Tr[τ
>
X
i
− Ri ) +
X
i
i
Tr[βi> (AGi
Γi Pi + ζ > h
(25)
− ρ>
i S)]
Γi Gi − Γd − ζ > S ]
The ramping reserve price for UM i is defined as the marginal
expected cost due to a unit decrement of the generation
ramping rate of UM i. The capacity reserve price for UM i is
defined as the marginal expected cost due to a unit decrement
of the generation capacity of UM i. They can be obtained from
the Lagrangian function as
π ri = −
∂Φ
∂Ri
= α∗i .
(26)
Hence, the reserve provided by i is valuable only when π ri is
non-zero. Note that α∗i consists of Lagrangian multipliers for
unit ramping limits and unit capacity limits.
The LMP for AARO SCED can be obtained based on
its definition. It is the marginal expected cost due to a unit
increment of the load. For generator i, it is formulated as
∗
π ei = λ∗ − Γ>
i η ,
(27)
5
where π ei ∈ RNT . π ei also consists of energy component and
congestion component. As shown in [13], the affine adjustment
price can be obtained as
∗
> ∗
π ai = γ ∗ − Γ>
i τ − A βi ,
(28)
π ai ∈ RNT ×ND NT represents the marginal value of the
a
adjustment coefficient in Gi . As stated in [13], Tr[G>
i πi ]
a
is the payment to i. It is noted that the unit of π i is $ as Gi
is participant factor.
D. Credit to Uncertainty Mitigators
Within the AARO framework, UMs help the system withstand the “load deviation” in the future. As the generation and
the reserve are coupled together, the credit to reserve should
reflect the coupling effect. Only the UMs who provide the
valuable reserves are entitled to credits.
The price of the flexible resources are π ri for UM i. The
total credit allocated to i is
>
Θi = Ri − APi∗ π ri ,
(29)
which is the product of the reserve price and the reserve
amount. In fact, the reserve price π ri reflects how much “value”
the reserve has. Ri − APi∗ reflects the available reserve,
which is the reserve quantity at each time interval. Only when
the reserve is a valuable reserve (i.e. π ri 6= 0), UM i get the
reserve credit. Otherwise, the credit entitled to UM i is zero
even if the available reserve it provides is non-zero. There
are similar phenomenons in the traditional zonal-based reserve
market. For example, the reserve price $0/MWh at a specific
zone occurs when the cleared system reserve is higher than
the required amount in Case I of Section VI [17].
E. Market Equilibrium
source i can obtain the maximum profit by following the ISO’s
dispatch instruction (Pi∗ , G∗i ). If the objective function (14) is
strictly convex, then (Pi∗ , G∗i ) is the unique choice for i to get
the maximal profit. Price signal π ri (associated with credit Θi )
and π ei provide the incentives for i to dispatch power output
to Pi∗ , which supplies the load and maintains the ramping and
capacity reserves. In addition, the price signal π ai provides the
incentive for i to follow the adjustment instruction Gi .
Given that the optimality condition of problem (P) is satisfied, dispatch signal (Pi∗ , G∗i ) and price signal (π ei , π ai , α∗i )
constitute a competitive partial equilibrium [27]. The assumption made for the equilibrium model is that the market
participant uses affine policy to adjust the generation, so the
dispatch obtained from AARO SCED is near-optimal solution.
Consequently, the equilibrium conditions are also near-optimal
for market participants.
An alternative way to get the partial market equilibrium is
to lump the reserve price into the LMP. However, it has serious
incentive issues. The integrated LMP can be written as
∗
>
π̂ie = λ∗ − Γ>
i η − A αi .
(31)
(π̂ie )> Pi = (π ei )> Pi − (αi )> APi .
(32)
The new credit UM i receives on the energy and reserve is
It can be observed that the reserve credit is negative when
the shadow price of the upper bound constraint is nonzero.
It means that the more flexible resources UM provides, the
fewer profit it gets. The UM is provided negative incentives
for the flexible resources. In contrast, the reserve credit defined
in (29) is always nonnegative.
III. C ASE S TUDY
A 3-Bus system and the modified IEEE 118-bus system
are studied in this section to illustrate the concepts of available/valuable ramping/capacity reserves and the associated
prices, as well as their impacts on market participants. The
simulations are carried out by CPLEX 12.5 on PC with Intel
i7-3770 3.40GHZ 8GB RAM.
In the partial market equilibrium model, it is assumed that
market participants are price takers [27]. This assumption is
popular in the electricity market [25], [28]. The expected profit
maximization problem for source i can be formulated as
(
)
a
>
Pi > π ei + Tr[G>
π
]
−
B
(P
+
G
)
i
i
i
i
i
(PMPi ) : max E
Pi ,Gi
−(Pi + Gi )> Qi (Pi + Gi ) − fi + Θi A. 3-Bus System
The 3-Bus system is consisted of two units, one wind farm,
s.t.
(8)(12)
two loads, and three lines. Please refer http://motor.ece.iit.edu/
where the decision variables for source i are Pi and Gi , data/rscuc/3 bus data.pdf for the detailed data. For simplicity,
a
given the price signal π ei ,π ai , and π ri . Credit Tr[G>
three time intervals are studied and a single-segment increi π i ] is
for the affine adjustment matrix as stated in [13]. It should mental cost (IC) is employed to represent the fuel cost. The
be noted that the constraint for mitigator i is (12), which is time resolution is 15 minutes. It is assumed that both units in
why incentive Θi is necessary for i to follow the dispatch the system are committed, which is determined in the DAM.
instruction. The objective function can be converted to
Load in the current interval (i.e., Time 1) is assumed to be
> e
> a
> r
accurate. Loads from Time 2 and Time 3 are forecasted based
Pi π i + Tr[Gi π i ] + (Ri − APi ) π i
(30)
on current available information, and forecasting errors may
>
−Pi> QPi − Bi> Pi − fi − Tr G>
i Qi Gi E( )
exist. Assume that the expectation of uncertainties at Time 2
It can be observed that (30) is a portion of the Lagrangian and Time 3 are 0, and their probability distribution is unknown.
The total expected cost calculated based on the AARO
function (25). The optimal solution (Pi , Gi ) to (PMPi ) is a
function of (π ei , π ai , π ri ). Since the problem (P) is convex and SCED is $1380.9. It is higher than the standard SCED cost
Slater’s condition is satisfied, the strong duality holds. There- $1333.5, which cannot be immuned against uncertainties. It
fore, the saddle point (Pi∗ , G∗i ), which is the optimal solution indicates that the expensive G2 supplies more loads within
to (P), is also the optimal solution to (PMPi ). Consequently, AARO SCED than it does in standard SCED. The reason
6
TABLE I
BASE ED AND LMP FOR 3-B US S YSTEM IN C ASE 2
G1 (MW)
G2 (MW)
1
2
3
126.5
131.65
143.6
13.5
23.35
23.4
Load (MW)
Time
6,280
LMP ($/MWh)
Bus1
Bus2
Bus3
10
10
10
10
32.5
23.5
10
32.5
23.5
6,260
6,240
6,220
1
2
3
4
5
Time Interval
6
7
8
Fig. 2. System-wide Net Load
TABLE II
U PWARD R ESERVES AND P RICES FOR 3-B US S YSTEM IN C ASE 2
Time
1
2
3
G1
G2
RR
RRP
CR
CRP
RR
RRP
CR
CRP
18.5
19.85
13.05
0
0
0
53.5
48.35
36.4
0
0
0
6.5
0.15
9.95
0
15
7.5
66.5
56.65
56.6
0
0
0
RR: Ramping Reserve, MW/15min; RRP: Ramping Reserve Price, $/(MWh/15min)
CR: Capacity Reserve, MW; CRP: Capacity Reserve Price, $/MWh
TABLE III
C REDITS E NTITLED TO U NITS BASED ON R ESERVE P RICES IN C ASE 2 ($)
Time
1
2
3
G1
0
0
0
G2
0
15*0.15*0.25=0.5625
7.5*9.95*0.25=18.6563
of dispatching the extra generation from G2 is that operator
needs additional deliverable reserves to accommodate the
uncertainties while minimizing the total cost. The non-zero
1
2
1
1
entries in G1 ∈ R3×6 are g11
= g12
= 1, g23
= 1, g24
=
1
3×6
0.98, g36 = 0.4, and the non-zero entries in G2 ∈ R
are
2
2
2
g24
= 0.02, g35
= 1, g36
= 0.6. If the loads on bus 1 and bus 3
at interval 2 are increased to 97MW and 67.5MW, respectively,
then the units can be re-dispatched based on the adjustment
matrix to 131.65 + 1 × 7 + 0.98 × 2.5 = 141.025MW and
23.35 + 0 × 7 + 0.02 × 2.5 = 23.475MW.
The question is whether the units have incentives to maintain the reserves. First, we consider the scenario without
reserve credits. The LMP at bus 2 at Time 2 of $32.5/MWh is
larger than G2’s marginal cost of $10/MWh. Therefore, G2 is
inclined to supply more loads at Time 2. The increase of the
output from G2 will shrink the reserve it can provide. The fact
that G2 has negative profit at Time 1 is the uplift issue [29],
which is beyond the scope of this paper. In this scenario, the
market participants would ignore the uncertainties and game
the market.
Now, we consider the second scenario with reserve credits.
The upward reserves provided by UMs are shown in Table II.
It is observed that the capacity reserve is not scarce resource,
i.e. the online capacity is adequate. They are just available
reserves. In contrast, the prices of upward ramping reserve at
Time 2 and Time 3 are non-zeros. According to the definition
in this paper, they are valuable reserves. In this scenario,
UMs are entitled certain credits based on the contribution. As
shown in Table III, uncertainty mitigator G2 is entitled $0.56
at Time 2 and $18.66 at Time 3, where $15/(MWh/15min)
and $7.5/(MWh/15min) are the entries in π r2 for ramping
constraints at Time 2 and Time 3, respectively; and 0.15
MW/15min and 9.95 MW/15min are the ramping reserves of
G2 at Times 2 and 3, respectively. The credit is calculated
based on (29). It is noted that the credits are time-coupled, as
the ramping reserve is calculated according to the difference
of the power outputs between two intervals.
Next we give an example to show whether G2 is inclined
to deviate from the base dispatch instruction with the credits
from LMPs and π ri . If G2 follows the instruction P2∗ as shown
in Table I, then the total credit associated with LMPs and
π ri are [13.5, 23.35, 23.4] · [10, 32.5, 32.5]> · 0.25 + 0.5625 +
18.6563 = $432.8125. Consider a possible base dispatch
[12.5, 22.35, 23.4], where G2 generates 1MW less at Time 1 as
its profit is negative at Time 1, and it also generates 1MW less
at Time 2 to maintain the same ramping reserve. Then the new
credit is [12.5, 22.35, 23.4] · [10, 32.5, 32.5]> · 0.25 + 0.5625 +
7.5·8.95·0.25 = $420.3125, where ramping reserve at Time 3
is decreased to 8.95MW/15min. The credit is $12.5 lower than
that by following P2∗ , and the saved fuel cost by supplying
smaller load is 2 · 25 · 0.25 = $12.5. Therefore, G2 does not
get more profit by using this dispatch. This example illustrates
how G2 gets credit associated with LMPs and π ri , and why
G2 is not inclined to deviate the dispatch instruction P2∗ . The
more rigorous mathematical analysis on market equilibrium is
shown in Section II-E.
B. Modified IEEE 118-Bus System
There are 54 traditional units and 186 branches in the
modified IEEE 118-Bus system. The scheduling period is 2
hours, and the time interval is 15 minutes. The loads are
depicted in Fig. 2. The UCs are determined in advance by
the solution to robust SCUC problem with 5% reserves. Five
wind farms are introduced in the system, and they are located
at buses 11, 49, 60, 78, and 90, respectively. We denote the
set of buses with uncertainty as M. It is assumed that the
forecasted power output (i.e. nominal output) and installed
capacity for each wind farm are 100 MW and 200 MW,
respectively. The uncertainties in this case are from the RES
only. The uncertainty m,t satisfies
|m,t | ≤ 100 · r1 1 + 0.01 · (t − 1) , m ∈ M, ∀t
P m,t ≤ 500 · r1 · r2 1 + 0.01 · (t − 1) , m ∈ M, ∀t
m
where r1 reflects the forecast error confidence interval for a
single wind farm [4], [14], and r2 reflects the forecast error
confidence interval for the aggregated wind output. When r2 <
1, it indicates that the aggregated forecast error confidence
interval is smaller than the sum of five single intervals. In the
experiment, the forecast error increases with the time intervals.
7
TABLE IV
R ESERVES W.R.T I NCREASING U NCERTAINTY L EVELS AND F IXED
N OMINAL W IND P OWER (r2 = 1, t = 2)
4,000
r1
0
0.3
0.4
r1
0.5
0.6
0.7
Fig. 3. Total Reserve Credit W.R.T. Uncertainty Levels and Fixed Nominal
Wind Power (r2 = 1)
The detailed data including unit parameters, uncertainty correlation matrix, line reactance and ratings, and net load profiles
can be found at http://motor.ece.iit.edu/Data/118 UMP.xls.
We consider the interval bounds for the uncertainties, and
perform the sensitivity analysis with respect to r1 . Fig. 3 shows
the reserve credits (RC) the UMs receive with the change of
the r1 . The reserve credit (RC) is the sum of the ramping
reserve credit and capacity reserve credit UMs are entitled to.
When r1 is high, the UMs are also entitled to high credits. As
shown in (29), the reserve credits are the sum of the products
of the amount of valuable reserve and the price of the valuable
reserve. They are analyzed as follows.
Table IV presents the upward/downward available reserves
and valuable reserves at Time 2 with increasing forecast errors,
fixed normal wind power output (100MW each), and fixed
r2 . The r1 ∈ [0.1, 0.7], so the error in percentage of the
installed capacity (200MW each) is from 5% to 35%. It is
observed that the available reserves remain the same while
valuable reserve change dramatically with the forecast errors.
As shown in Table IV, upward available ramping reserve
remains 886.70MW/15min, and the capacity reserve remains
661.43MW. The main reason is that the unit commitment and
load demand is fixed at Time 2 in the system. In contrast, the
upward valuable reserve is 0 when r1 = 0.1. It indicates that
the opportunity cost of keeping the ramping reserve for UM is
zero as it can recover the profit from the energy credit. When
r1 is 0.2, 0.3 and 0.4, the valuable ramping reserve is around
180MW/15min, and the valuable capacity reserve is around
11MW, 61MW, and 112MW respectively. UMs are entitled to
credits by keeping the reserves, which is also shown in Fig. 3.
It suggests that opportunity cost of keeping the reserve is nonzero, i.e., UMs can get more profits by deviating from the
dispatch instruction if they are not entitled to reserve credits.
When r1 is further increased to 0.5, the amount of valuable
ramping reserve jumps to 479MW/15min. It means that more
available reserves become valuable when the uncertainty level
is high. The valuable capacity reserve also increases to 163
MW in this case. A similar tendency can also be observed
for the downward reserves shown in Table IV. It should be
emphasized that the amount of valuable reserve doesn’t change
monotonically with the uncertainty level. Instead, what we
revealed in this paper is a trend.
The upward available ramping reserve and the price of unit
42 are depicted for different time intervals in Fig. 4 with
r1 = 0.7, r2 = 1. As shown in Fig. 4, at Time 7, the unit
only provides available reserve but not valuable reserve as its
0.1
0.2
0.3
0.4
0.5
0.6
0.7
886.79
886.79
886.79
886.79
886.79
886.79
886.79
Downward Reserve
0
180.79
180.29
179.79
479.29
513.33
684.04
661.43
661.43
661.43
661.43
661.43
661.43
661.43
0
10.61
61.36
112.11
162.86
215.49
316.72
938.46
938.46
938.46
938.46
938.46
938.46
938.46
0
10.37
0
0
361.61
337.5
680.33
4203.57
0
4203.57
0
4203.57
0
4203.57
0
4203.57
0
4203.57 4.52
4203.57 12.14
AvaRamp: Available Ramping Reserve (MW/15min); ValRamp: Valuable Ramping Reserve (MW/15min);
AvaCap: Available Capacity Reserve (MW); ValCap: Valuable Capacity Reserve (MW);
Ramping Reserve
Ramping Reserve (MW/15min)
0.2
Ramping Reserve Price
8
4
6
4
2
2
0
0
1
2
3
4
5
6
7
8
Time Interval
Fig. 4. Upward Ramping Reserve and Price for Unit 42 (r1 = 0.7, r2 = 1)
price is zero. It can be observed that the ramping reserve price
reaches its highest point at Time 5, which is also the peak load
interval. In contrast, the ramping reserve price is low at Time
6 although the load demand at Time 6 is still relatively high
compared to those at other intervals. It is observed the load
climbs from Time 4 to Time 5, but falls from Time 5 to Time
6 as shown in Fig. 2. It indicates that the ramping reserve
price is not only related with the load demand but also with
the load change. In this case, the upward ramping reserve is
scarce resource at Time 5, and the opportunity cost of keeping
them is also high. In contrast, the upward ramping reserve is
relatively cheap when the load demand is falling at Time 6,
7, and 8.
Fig. 5 depicts the available capacity reserve and the price of
capacity reserve for unit 24 with r1 = 0.7, r2 = 1. Although
the reserve amount unit 24 keeps is the same at each time
interval, the price is different. It is observed that the capacity
reserve price in this case has a similar trend to the system
load level shown in Fig. 2. For example, the capacity reserve
Capacity Reserve
Capacity Reserve (MW)
0.1
Upward Reserve
AvaRamp ValRamp AvaCap ValCap AvaRamp ValRamp AvaCap ValCap
Ramping Reserve Price ($/(MW/15min))
2,000
Capacity Reserve Price
14
10
12
5
10
1
2
3
4
5
6
7
Capacity Reserve Price ($/MW)
Reserve Credit ($)
6,000
8
Time Interval
Fig. 5. Upward Capacity Reserve and Price for Unit 24 (r1 = 0.7, r2 = 1)
8
2,000
RC
EOC
1.94
1.93
1,500
EOC ($)
RC ($)
·105
1,000
0.8
0.9
1
r2
Fig. 6. Reserve Credit (RC) and Expected Operation Cost (EOC) with
Different r2 (r1 = 0.6)
is the most expensive at Time 5 when the peak load occurs.
The reason is that the system-wide upward capacity reserve
is the online installed capacity, which is fixed, less the load
level. When the load level is high, the upward capacity reserve
is small which becomes a scarce resource in the system.
Reserve credits with respect to different r2 are presented in
Fig. 6. It shows that the decrease of r2 (from 1 to 0.9 to 0.8)
also leads to lower payments related to reserve. For example,
when r2 decreases from 1 to 0.8, the total RC decreases from
around $2,300 to $1,000. The expected operation cost also
decreases from $194,000 to $192,000. It indicates that the
shrinking uncertainty set actually increases the feasible set for
the robust dispatches.
The numerical results in this part indicate that the reserve
payment proposed in this paper helps maximize the social warfare. When the uncertainty level is high, the payment related
to reserve is also high, It may attract long-term investment
of flexible resources. More flexible resources also mean the
system has more capabilities to handle the uncertainties, and
the system can accommodate higher RES penetrations.
IV. C ONCLUSIONS
This paper proposes a new concept, ramping reserve, within
the AARO SCED framework. AARO SCED is an effective tool
in RTM to address the uncertainty issue although its solution
may only be near-optimal. The flexibilities in this paper
include the generation ramping reserve and the generation
capacity reserve. The prices for the ramping reserve and
capacity reserve are also derived based on the Lagrangian
function. They are the opportunity costs of the uncertainty
mitigators to keep the reserves or flexibilities. With the help
of these prices, the reserves are classified into two categories,
which are available reserves and valuable reserves. The case
studies explain the concept of these reserves and the impacts
on the behaviors of market participants.
Many researches on this topic are open in the future.
With increasing RES penetration in the power system, the
flexibilities play a crucial role in uncertainty accommodation.
The prices derived in this paper provide an option on how
to provide the reserve signals within the robust optimization
framework. Those reserve credits to UMs may attract the
investment of flexible resources in the long term. In return,
new flexibility investment allows the system to accommodate
higher RES level.
It should be pointed out that the reserve prices are unitspecified. It is admitted that it is not perfect. It is ideal that
all the resources at the same bus have the same price. An
extension of the proposed reserve prices is to set the maximum
ramping and capacity reserve prices for the units located at a
node as the nodal price. In this way, the nodal reserve prices
are determined by the most expensive opportunity cost for the
reserve in the node. However, due to the affine policy and the
non-zero optimality of the AARO SCED, it may deteriorate
the partial market equilibrium to a certain extent.
R EFERENCES
[1] R. Wiser and M. Bolinger, “2011 wind technologies market report,”
Lawrence Berkeley National Laboratory, Tech. Rep., 2012.
[2] “2012 economic demand response performance report,” PJM Interconnection, Tech. Rep., 2012. [Online]. Available: http://www.pjm.com
[3] H. Holttinen, “Impact of hourly wind power variations on the system
operation in the nordic countries,” Wind Energy, vol. 8, no. 2, pp. 197–
218, 2005.
[4] B.-M. Hodge and M. Milligan, “Wind power forecasting error distributions over multiple timescales,” in Power and Energy Society General
Meeting, 2011 IEEE. IEEE, 2011, pp. 1–8.
[5] “Integration of renewable resources,” California ISO, Tech. Rep., 2010.
[6] “Integration of wind into system dispatch,” New York ISO, Tech. Rep.,
2008.
[7] F. Borggrefe and K. Neuhoff, “Balancing and intraday market design:
Options for wind integration,” Discussion Papers, German Institute for
Economic Research, DIW Berlin, Tech. Rep., 2011.
[8] California ISO, “CAISO hour-ahead scheduling process.”
[9] S. Takriti, J. Birge, and E. Long, “A stochastic model for the unit
commitment problem,” IEEE Trans. Power Sys., vol. 11, no. 3, pp. 1497–
1508, 1996.
[10] L. Wu, M. Shahidehpour, and Z. Li, “Comparison of scenario-based
and interval optimization approaches to stochastic SCUC,” IEEE Trans.
Power Sys., vol. 27, no. 2, pp. 913–921, 2012.
[11] R. Jiang, M. Zhang, G. Li, and Y. Guan, “Two-stage network constrained
robust unit commitment problem,” European Journal of Operational
Research, vol. 234, no. 3, pp. 751 – 762, 2014.
[12] D. Bertsimas, E. Litvinov, X. Sun, J. Zhao, and T. Zheng, “Adaptive
robust optimization for the security constrained unit commitment problem,” IEEE Trans. Power Syst., vol. 28, no. 1, pp. 52–63, 2013.
[13] J. Warrington, P. Goulart, S. Mariethoz, and M. Morari, “Policy-based
reserves for power systems,” IEEE Trans. Power Sys., vol. 28, no. 4, pp.
4427–4437, 2013.
[14] R. A. Jabr, “Adjustable robust opf with renewable energy sources,” IEEE
Trans. Power Sys., vol. 28, no. 4, pp. 4742–4751, 2013.
[15] A. Ben-Tal, A. Goryashko, E. Guslitzer, and A. Nemirovski, “Adjustable
robust solutions of uncertain linear programs,” Math. Program., Ser. A,
vol. 99, no. 2, pp. 351–376, Mar. 2004.
[16] H. Ye and Z. Li, “Robust security-constrained unit commitment with
recourse cost requirement,” in Power Energy Society General Meeting,
2015 IEEE, July 2015, pp. 1–5.
[17] T. Zheng and E. Litvinov, “Contingency-Based Zonal Reserve Modeling
and Pricing in a Co-Optimized Energy and Reserve Market,” IEEE
Transactions on Power Systems, vol. 23, no. 2, pp. 277–286, May 2008.
[18] J. Wang, M. Shahidehpour, and Z. Li, “Contingency-constrained reserve
requirements in joint energy and ancillary services auction,” Power
Systems, IEEE Transactions on, vol. 24, no. 3, pp. 1457–1468, 2009.
[19] C. Weber, “Adequate intraday market design to enable the integration of
wind energy into the european power systems,” Energy Policy, vol. 38,
no. 7, pp. 3155–3163, 2010.
[20] C. ISO, “Flexible ramping product-draft technical appendix,”
http://www.caiso.com/Documents/DraftTechnicalAppendix
FlexibleRampingProduct.pdf, 2015, [Online; accessed 1-July-2015].
[21] A. L. Ott, “Experience with PJM market operation, system design, and
implementation,” IEEE Trans. Power Sys., vol. 18, no. 2, pp. 528–534,
2003.
[22] Z. Li, W. Wu, B. Zhang, and B. Wang, “Adjustable robust realtime power dispatch with large-scale wind power integration,” IEEE
Transactions on Sustainable Energy, vol. PP, no. 99, pp. 1–12, 2015.
[23] Z. Li and M. Shahidehpour, “Security-constrained unit commitment for
simultaneous clearing of energy and ancillary services markets,” IEEE
Trans. Power Sys., vol. 20, no. 2, pp. 1079–1088, 2005.
[24] J. F. Ellison, L. S. Tesfatsion, V. W. Loose, and R. H. Byrne, “Project
report: A survey of operating reserve markets in us iso/rto-managed
electric energy regions,” Sandia Natl Labs Publications, 2012.
[25] T. Zheng and E. Litvinov, “On ex post pricing in the real-time electricity
market,” IEEE Trans. Power Sys., vol. 26, no. 1, pp. 153–164, 2011.
9
[26] ISO New England Manual for Market Operations Manual M-11 Revision
44, ISO New England Inc., access:May 19, 2015. [Online]. Available:
http://www.iso-ne.com/static-assets/documents/rules proceds/
isone mnls/M11/m 11 market operations revision 44 05 23 13.doc
[27] A. Mas-Colell, M. D. Whinston, J. R. Green et al., Microeconomic
theory. Oxford university press New York, 1995, vol. 1.
[28] M. Shahidehpour, H. Yamin, and Z. Li, Market Operations in Electric
Power Systems: Forecasting, Scheduling, and Risk Management, 1st ed.
Wiley-IEEE Press, 2002.
[29] W. W. Hogan and B. J. Ring, “On minimum-uplift pricing for electricity
markets,” Electricity Policy Group.
| 3cs.SY
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.